chore: bump version to 1.5.6

* feat(providers): api_surface toggle + mistral medium reasoning fix

Mistral medium open-weights served by vLLM expects reasoning_effort via
the Responses API (`reasoning.effort`), not as a `chat_template_kwargs`
entry on Chat Completions.  The session was unconditionally injecting
`{"reasoning_effort": ...}` into `chat_template_kwargs` for every
openai-compatible request, which corrupted the prompt rendering for any
backend whose chat template didn't consume that key (Mistral medium,
Mistral cloud, Groq, OpenRouter).

Changes:
- Add `api_surface` ("chat" | "responses") to `ModelConfig.server_compat`
  and thread it through `create_provider` / `model_registry.get_provider`.
  `openai-compatible` defaults to Chat Completions; operators can flip
  individual aliases to Responses for endpoints that support it.
- New `vllm-mistral-medium` profile that pre-fills api_surface=responses
  on Detect for known Mistral medium model ids.
- Drop the unconditional `reasoning_effort` injection into
  `chat_template_kwargs`.  Operators running gpt-oss-style local
  templates that consume `reasoning_effort` from the chat template now
  opt in via `server_compat.extra_body.chat_template_kwargs`.
- New "API Surface" select in the Models admin tab; allowlist-validated
  server-side at create/update time; pre-filled by Detect via the
  profile suggestion.
- Evict the cached provider singleton in `ModelRegistry.reload()` when
  api_surface changes (previously only cfg.provider triggered eviction).
- Fix `_run_agent` fallback path to inherit the session's primary alias
  for capability and server_compat resolution; previously the fallback
  passed `alias=None`, which silently dropped per-model caps on the
  agent path.

Tests: 5117 passed (-m "not live"); ruff + mypy clean.

* fix(providers): don't auto-suggest Responses for Mistral medium

vLLM's Responses API surface for Mistral medium open-weights doesn't
wire up the Mistral tool-call parser as of vLLM 0.x — tool calls leak
into the response as ``[TOOL_CALLS]<name>{...}`` text instead of
structured tool_calls.  Chat Completions on the same engine handles
tools cleanly via ``--tool-call-parser mistral``, and reasoning can be
turned on via the vLLM CLI ``--reasoning-parser`` flag.

Drop the auto-suggest mapping so Detect falls back to the generic
``vllm`` profile.  Keep the ``vllm-mistral-medium`` profile definition
in place so an operator who specifically wants per-request effort and
accepts the tool-calling limitation can still pick "Responses API"
manually in the admin UI.

* fix(providers): address Copilot review on PR #469

- providers/__init__.py: drop the redundant *_responses_provider /
  *_chat_provider names; have create_provider use _openai_provider and
  _openai_compat_provider directly so they're not flagged as unused
  globals.
- console/server.py: tighten _validate_api_surface to a strict equality
  match against the canonical {"chat", "responses"} set.  The previous
  strip().lower() membership check accepted ' Responses '/'CHAT' but
  stored the raw string verbatim, which then failed to round-trip
  through the admin <select>.
- console/static/admin.js: gate the entire server_compat block (server
  type, api_surface, extra_body) on provider == "openai-compatible" at
  save time so toggling provider away can't leave a stale hidden surface
  selection in the persisted capabilities JSON.
- tests/test_session.py: splat the bad kwarg via **dict so CodeQL no
  longer flags the call as a wrong-name keyword (the point of the test
  is the runtime contract, not the static type).
- tests/test_admin_model_registry_refresh.py: add endpoint-level tests
  for the api_surface validation on both create and update — covers the
  bogus-value rejection, non-canonical-string rejection, and the happy
  path persisting through to the refreshed registry.

docs/skills/import-conversation-history/SKILL.md

@@ -0,0 +1,233 @@
+---
+name: import-conversation-history
+description: Use this skill when the user wants to import or migrate conversation history from another LLM chat or coding tool (e.g. ChatGPT, Claude.ai, Cursor, Copilot Chat, Aider, Gemini, a custom JSON export) into Turnstone. The skill teaches Turnstone's destination contracts — workstream identity, the OpenAI-shaped message rows, tool-call/result pairing, provider-fidelity blobs, attachments, and archive-vs-resumable choice — so the agent can map any source format onto them. Trigger phrases: "import my chats", "migrate this transcript into Turnstone", "bring my Claude.ai history over", "load this export as a workstream".
+version: 1.0.0
+---
+
+# Importing Conversation History into Turnstone
+
+## Overview
+
+Source formats vary; the destination does not. Your job is to translate whatever the user hands you (JSON dump, ZIP export, scraped HTML, screenshot OCR, raw transcript) into Turnstone's internal shape: **one workstream row** plus an ordered sequence of **conversation rows** in OpenAI message format. This skill documents the destination so you can write a correct mapper for any source.
+
+Two questions to settle with the user before writing anything:
+
+1. **Archive or resumable?** An archive ("saved" workstream — `state="closed"`) is read-only history. A resumable workstream (`state="idle"`) lets the user continue the conversation; this only works cleanly when the source LLM matches a Turnstone-supported provider/model and tool definitions still resolve.
+2. **One workstream per source thread, or merge?** Default to one-to-one unless the user explicitly asks to merge.
+
+Default to **archive** when in doubt — resuming a foreign transcript with mismatched tool schemas or stale provider signatures will fail at the next turn.
+
+## Turnstone Data Model (the destination)
+
+Two tables carry the conversation:
+
+### `workstreams` (one row per imported thread)
+
+| Column | Required | Notes |
+|---|---|---|
+| `ws_id` | yes | 32-char lowercase hex. Auto-generate with `secrets.token_hex(16)` if you don't already have one. **First 4 hex chars are the routing bucket** — see "Identity & Routing" below. |
+| `name` | yes | Short title. Pull from source thread title; fall back to first ~60 chars of first user message. |
+| `state` | yes | `"closed"` for archive, `"idle"` for resumable. Never set `"running"` on import. |
+| `kind` | yes | `"interactive"` for normal threads. Do NOT use `"coordinator"` for imports — that's reserved for cluster-spawned coordinator workstreams. |
+| `parent_ws_id` | no | Leave NULL. Only set if you're importing a coordinator-spawned subtree and re-parenting it; rare. |
+| `user_id` | yes | Owner. Must exist in `users`; importer must know which Turnstone user owns the imported history. |
+| `node_id` | yes (multi-node) | Denormalized cache of the node that owns this `ws_id`'s bucket. Single-node deployments can leave it NULL or set it to the only node. |
+| `alias` | no | Human-typeable short name. Optional; must be unique cluster-wide if set. |
+| `title` | no | Auto-titled later by the LLM; safe to leave NULL on import. |
+| `skill_id`, `skill_version` | yes | Default `""` and `0` unless the source thread was scoped to a Turnstone skill. |
+| `created`, `updated` | yes | ISO8601 strings. Use the source's first/last message timestamps when available. |
+
+### `conversations` (many rows per thread, ordered by `id`/`timestamp`)
+
+| Column | Notes |
+|---|---|
+| `ws_id` | The workstream this row belongs to. |
+| `timestamp` | ISO8601 string. Preserve source timestamps; fall back to monotonically increasing values if unknown. **Order is canonical via `id` (autoincrement), not `timestamp`** — but always insert in conversational order so both agree. |
+| `role` | One of `system`, `user`, `assistant`, `tool`, `developer`. See role mapping below. |
+| `content` | Text. May be NULL for assistant rows that are *only* tool calls. |
+| `tool_name` | Set on `role="tool"` rows (the tool whose result this is). NULL otherwise. |
+| `tool_call_id` | Set on `role="tool"` rows (matches the assistant row's `tool_calls[].id`). NULL otherwise. |
+| `tool_calls` | JSON-encoded list, on `role="assistant"` rows that issued tool calls. OpenAI shape — see "Tool Calls" below. |
+| `provider_data` | JSON blob preserving provider-native content blocks (Anthropic `signature`, Gemini `thought_signature`, etc.). Optional; only matters for **resumable** imports against the same provider. Skip for archives. |
+
+The internal format is **OpenAI-shaped**, even when the source was Anthropic or Gemini. Providers translate at their own API boundary; storage stays uniform.
+
+## Identity & Routing (`ws_id`)
+
+- `ws_id` is **32-char lowercase hex** (i.e. `secrets.token_hex(16)`).
+- The **routing bucket** is `int(ws_id[:4], 16)` — the first 4 hex chars place this workstream on a specific node via the consistent hash ring.
+- For multi-node imports: either insert through the console's routing proxy (which forwards to the owning node), or generate `ws_id`s and write directly to each node's database in batches grouped by bucket.
+- For single-node imports: bucket math is irrelevant; any `ws_id` works.
+- **Do not reuse the source platform's IDs as `ws_id`** unless they happen to be 32-char hex. Generate fresh; if you need the old ID for traceability, store it in `workstream_config` under a key like `import.source_id`.
+
+## Recommended Import Path
+
+Three options, in order of preference:
+
+### 1. Storage protocol (recommended for full history)
+
+Use `turnstone.core.storage.Storage.save_messages_bulk(rows)`. This is the canonical bulk-insert primitive and bypasses the LLM round-trip entirely.
+
+```python
+from turnstone.core.storage import get_storage  # construct via the same path the server uses
+
+storage = get_storage(...)  # see turnstone.core.storage.__init__ for the project's wiring
+
+storage.create_workstream(  # or whatever the project's exposed creator is — check turnstone/core/storage/_protocol.py
+    ws_id=ws_id,
+    user_id=user_id,
+    name=name,
+    state="closed",
+    kind="interactive",
+    ...
+)
+
+storage.save_messages_bulk([
+    {"ws_id": ws_id, "role": "user", "content": "Hello"},
+    {"ws_id": ws_id, "role": "assistant", "content": "Hi! What can I help with?"},
+    {"ws_id": ws_id, "role": "assistant", "content": None,
+     "tool_calls": json.dumps([{"id": "call_1", "type": "function",
+                                "function": {"name": "search", "arguments": "{\"q\":\"x\"}"}}])},
+    {"ws_id": ws_id, "role": "tool", "tool_name": "search", "tool_call_id": "call_1",
+     "content": "result text"},
+    # ...
+])
+```
+
+`save_messages_bulk` handles `timestamp` and the workstream's `updated` column internally, so you don't need to compute them per row. **Verify the exact creator signature** by reading `turnstone/core/storage/_protocol.py` — table layout has shifted across migrations and the Storage protocol is the source of truth.
+
+### 2. SDK `create_workstream(resume_ws=...)` (when the source is already a Turnstone workstream)
+
+Only useful for *Turnstone → Turnstone* re-parenting. Not relevant for foreign sources.
+
+### 3. SDK `create_workstream(initial_message=...)` + `send()` per turn (last resort)
+
+Only fits archives where the source had **no tool calls** and you don't care about preserving assistant turns verbatim. Each `send()` triggers a real LLM round-trip, which is expensive and rewrites assistant content. Don't use this for full history.
+
+## Role Mapping
+
+Common source-role conventions and how they map to Turnstone:
+
+| Source role | Turnstone `role` | Notes |
+|---|---|---|
+| `user`, `human` | `user` | Direct map. |
+| `assistant`, `ai`, `model`, `bot` | `assistant` | Direct map. |
+| `system` | `system` | Preserve only if it's content the user wrote (custom instructions). Drop boilerplate provider preambles — Turnstone composes its own system message. |
+| `developer` (OpenAI o-series) | `developer` | Preserve. |
+| `tool`, `function`, `tool_result` | `tool` | Must carry `tool_name` and `tool_call_id` matching the prior assistant row's `tool_calls[].id`. |
+| `tool_use` (Anthropic) | `assistant` with `tool_calls` | Anthropic emits tool calls *inside* an assistant message; flatten to OpenAI shape. |
+| `human_feedback`, `revision` | `user` | Treat as a follow-up user turn. |
+
+## Tool Calls (the most error-prone part)
+
+Turnstone stores tool calls in OpenAI's nested-function shape on the assistant row, and matches them with `role="tool"` result rows by `tool_call_id`.
+
+### Assistant row with tool calls
+
+```json
+{
+  "role": "assistant",
+  "content": null,
+  "tool_calls": [
+    {
+      "id": "call_abc123",
+      "type": "function",
+      "function": {
+        "name": "search_web",
+        "arguments": "{\"query\":\"turnstone import\"}"
+      }
+    }
+  ]
+}
+```
+
+`tool_calls[].function.arguments` is **a JSON-encoded string**, not an object. Source formats commonly get this wrong — Anthropic stores arguments as a parsed object, Gemini as a struct. Always re-serialize to a string.
+
+### Tool result row
+
+```json
+{
+  "role": "tool",
+  "tool_name": "search_web",
+  "tool_call_id": "call_abc123",
+  "content": "..."
+}
+```
+
+Pairing rules:
+- Every assistant `tool_calls[].id` MUST be followed by exactly one `role="tool"` row with the matching `tool_call_id`, before the next user/assistant turn.
+- If the source dropped the tool result (cut-off transcript), insert a synthetic `role="tool"` row with `content="[tool result missing in source]"` to keep the chain valid. An assistant row with an unanswered `tool_calls[].id` will break replay and any LLM round-trip.
+- Multi-tool assistant turns: one `role="tool"` row per call, in any order, all before the next non-tool row.
+
+### Tool ID generation
+
+If the source used opaque tool IDs that aren't unique within a thread (some platforms reuse them), regenerate with a stable scheme like `f"call_{i}"` where `i` is a per-thread counter. Update both the assistant and tool rows together.
+
+## Provider Fidelity (`provider_data`)
+
+Skip this entirely for **archive** imports.
+
+For **resumable** imports against the same provider, populate `provider_data` to preserve provider-specific tool-call metadata that the next API round-trip will require:
+
+- **Anthropic**: `signature` field on thinking blocks; required for round-tripping extended-thinking responses.
+- **Gemini**: `thought_signature` on tool calls; required for fidelity.
+- **OpenAI**: typically nothing to preserve.
+
+The runtime-side dict key is `_provider_content` (a list of provider-native blocks); the persisted column is `provider_data` (the same list, JSON-encoded). If you don't have provider-native blocks from the source — and you usually won't, because a foreign export won't include them — leave `provider_data` NULL. The first new turn will succeed without it, but the previous assistant turn's reasoning won't replay back to the model.
+
+## Attachments
+
+If the source thread had image or file attachments:
+
+- **Size limits**: images ≤ 4 MiB, text documents ≤ 512 KiB. Reject or downsample anything bigger.
+- **Allowed types**: server validates magic bytes for images and UTF-8-decodes for text. Binary blobs that aren't images won't pass.
+- **Lifecycle**: pending → reserved → consumed. For imports, the cleanest path is to upload as pending and immediately consume by attaching to the relevant `conversations.id`.
+
+Two import paths:
+
+1. **Bulk-insert + post-attach**: insert messages first, get back the assistant/user `conversations.id`, then write `workstream_attachments` rows linking the file to `message_id`.
+2. **SDK multipart create**: `create_workstream(attachments=[...], initial_message=...)` for the *first* turn only — the server reserves and consumes them onto that turn. Doesn't help for mid-thread attachments.
+
+For full-history imports with multiple attachments at different turns, path (1) is the only option.
+
+## Validation Checklist
+
+Before declaring success, verify:
+
+- [ ] `ws_id` is 32-char lowercase hex.
+- [ ] `workstreams` row exists with the right `user_id`, `state`, `kind`.
+- [ ] Conversation rows are inserted **in order** (autoincrement `id` will reflect insert order).
+- [ ] Every assistant `tool_calls[].id` has a matching `role="tool"` row with the same `tool_call_id`.
+- [ ] `tool_calls[].function.arguments` is a JSON-encoded **string**, not a parsed object.
+- [ ] First message is typically `role="user"` (not `system`) — Turnstone composes its own system prompt at runtime.
+- [ ] No empty assistant rows (`content=NULL` AND `tool_calls=NULL` is invalid).
+- [ ] If multi-node: the `ws_id`'s bucket maps to a node that exists; `workstreams.node_id` matches.
+- [ ] Round-trip test: run `Storage.load_messages(ws_id)` and confirm the reconstructed list matches what you inserted (modulo timestamps).
+
+## Anti-patterns
+
+- **Don't import the source provider's system prompt verbatim.** Provider boilerplate ("You are Claude...", "You are ChatGPT...") will conflict with Turnstone's composed system message and confuse the model on resume. Drop it; preserve only user-authored custom instructions.
+- **Don't preserve foreign tool definitions as Turnstone tools.** If the source had custom tools that don't exist in Turnstone, the assistant rows that called them are still valid history (archive), but the workstream is **not resumable** — mark `state="closed"`.
+- **Don't fabricate `tool_call_id`s without re-pairing.** Mismatched ids silently break the replay chain on the next turn.
+- **Don't skip the `tool_name` field on `role="tool"` rows.** Some load paths use it for display and audit; NULL there will render as "unknown tool".
+- **Don't write through the LLM (`send()` per turn) for full history.** It's expensive, rewrites assistant turns, and rate-limits will bite long imports.
+
+## Quick Reference
+
+| Task | Path |
+|---|---|
+| Generate ws_id | `secrets.token_hex(16)` |
+| Bulk insert messages | `Storage.save_messages_bulk(rows)` |
+| Archive (read-only) | `state="closed"`, skip `provider_data` |
+| Resumable | `state="idle"`, populate `provider_data` if same provider |
+| Tool call id | OpenAI shape: `{"id": ..., "type": "function", "function": {"name": ..., "arguments": "<json string>"}}` |
+| Tool result row | `role="tool"`, `tool_name`, `tool_call_id`, `content` |
+| Source role → Turnstone role | See "Role Mapping" table |
+| Per-thread metadata | Store source IDs in `workstream_config` under `import.*` keys |
+
+## Files to read before writing the importer
+
+- `turnstone/core/storage/_schema.py` — authoritative table definitions.
+- `turnstone/core/storage/_protocol.py` — `save_message`, `save_messages_bulk`, `load_messages` signatures.
+- `turnstone/core/session.py` (around the message-save section) — how the runtime constructs in-memory message dicts; mirror this shape on import to round-trip cleanly.
+- `turnstone/api/server_schemas.py` — Pydantic shapes for the SDK paths if you go through HTTP.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "turnstone"
-version = "1.5.4"
+version = "1.5.6"
 description = "Multi-node AI orchestration platform with tool use, agent routing, and cluster simulation."
 readme = "README.md"
 license = "BUSL-1.1"

tests/_helpers.py

@@ -0,0 +1,28 @@
+"""Shared test helpers — kept out of conftest.py since these are factories,
+not fixtures, and several test files want to import them directly."""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import MagicMock
+
+
+def make_chat_session(**overrides: Any) -> Any:
+    """Build a minimal ``ChatSession`` with sane test defaults.
+
+    Caller passes any constructor arg as a kwarg to override the default —
+    e.g. ``make_chat_session(memory_config=MemoryConfig(fetch_limit=5))``.
+    """
+    from turnstone.core.session import ChatSession
+
+    defaults: dict[str, Any] = {
+        "client": MagicMock(),
+        "model": "test-model",
+        "ui": MagicMock(),
+        "instructions": None,
+        "temperature": 0.5,
+        "max_tokens": 4096,
+        "tool_timeout": 30,
+    }
+    defaults.update(overrides)
+    return ChatSession(**defaults)

tests/test_admin_model_registry_refresh.py

@@ -352,6 +352,90 @@ def test_update_endpoint_skips_refresh_on_empty_body(
     assert calls == []  # gate held: empty body did not trigger a refresh
 
 
+def test_create_rejects_invalid_api_surface(storage: SQLiteBackend) -> None:
+    """POST with a bogus server_compat.api_surface returns 400 rather than
+    persisting a value that would make get_provider() raise on every later
+    ChatSession init for the alias."""
+    _seed_model_def(storage, definition_id="m1", alias="local", model="m")
+    registry = _make_registry(alias="local", model="m")
+    client = _make_client(storage, registry)
+
+    resp = client.post(
+        "/v1/api/admin/model-definitions",
+        json={
+            "alias": "bad",
+            "model": "x",
+            "provider": "openai-compatible",
+            "base_url": "http://localhost:9000/v1",
+            "api_key": "sk-x",
+            "capabilities": {"server_compat": {"api_surface": "BOGUS"}},
+        },
+    )
+    assert resp.status_code == 400, resp.text
+    assert "api_surface" in resp.json()["error"]
+    # And the alias is not persisted
+    assert not registry.has_alias("bad")
+
+
+def test_create_rejects_non_canonical_api_surface(storage: SQLiteBackend) -> None:
+    """Strict validation: ' Responses ' / 'CHAT' don't round-trip through the
+    admin <select>, so they're rejected even though they'd survive a
+    case-insensitive membership check."""
+    _seed_model_def(storage, definition_id="m1", alias="local", model="m")
+    registry = _make_registry(alias="local", model="m")
+    client = _make_client(storage, registry)
+
+    for bad in (" responses ", "RESPONSES", "Chat"):
+        resp = client.post(
+            "/v1/api/admin/model-definitions",
+            json={
+                "alias": "noncanon",
+                "model": "x",
+                "provider": "openai-compatible",
+                "base_url": "http://localhost:9000/v1",
+                "api_key": "sk-x",
+                "capabilities": {"server_compat": {"api_surface": bad}},
+            },
+        )
+        assert resp.status_code == 400, f"{bad!r}: {resp.text}"
+
+
+def test_create_accepts_valid_api_surface(storage: SQLiteBackend) -> None:
+    """Canonical 'chat' / 'responses' / unset are all accepted and persisted."""
+    _seed_model_def(storage, definition_id="m1", alias="local", model="m")
+    registry = _make_registry(alias="local", model="m")
+    client = _make_client(storage, registry)
+
+    resp = client.post(
+        "/v1/api/admin/model-definitions",
+        json={
+            "alias": "responses-alias",
+            "model": "x",
+            "provider": "openai-compatible",
+            "base_url": "http://localhost:9000/v1",
+            "api_key": "sk-x",
+            "capabilities": {"server_compat": {"api_surface": "responses"}},
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    assert registry.has_alias("responses-alias")
+
+
+def test_update_rejects_invalid_api_surface(storage: SQLiteBackend) -> None:
+    """PUT path also gates the validation, so an admin can't smuggle a bad
+    value into an existing alias."""
+    _seed_model_def(storage, definition_id="m1", alias="local", model="m")
+    registry = _make_registry(alias="local", model="m")
+    client = _make_client(storage, registry)
+
+    resp = client.put(
+        "/v1/api/admin/model-definitions/m1",
+        json={"capabilities": {"server_compat": {"api_surface": "junk"}}},
+    )
+    assert resp.status_code == 400, resp.text
+    assert "api_surface" in resp.json()["error"]
+
+
 def test_delete_endpoint_refreshes_registry(storage: SQLiteBackend) -> None:
     """DELETE drops the alias from the in-process registry too — a
     coord session that tried to resolve the deleted alias would

tests/test_app_js.py

@@ -92,3 +92,71 @@ def test_tool_error_does_not_overwrite_approval_badge() -> None:
         "badge instead so the approval verdict stays visible alongside "
         "the error."
     )
+
+
+def test_replay_history_renders_content_before_tool_block() -> None:
+    """In ``replayHistory``'s ``role === "assistant"`` branch, the
+    ``msg.content`` render must precede the ``msg.tool_calls`` render.
+
+    Two reasons, both load-bearing:
+
+    1. **Structural** — the next loop iteration's ``role === "tool"``
+       message anchors to ``lastToolBlock``. The tool-block branch sets
+       that anchor; the content branch clears it. If content runs after
+       the tool block, the clear silently drops the upcoming tool
+       result. Pre-fix, every interactive tool result was missing from
+       saved-workstream replays whenever the assistant turn carried
+       both narration and tool calls (very common output shape).
+
+    2. **Visual** — the live SSE path renders content first
+       (``stream_text`` streams before ``tool_info`` /
+       ``approve_request``), so replay should match.
+
+    The test pins the order via the offsets of the ``msg.content`` and
+    ``msg.tool_calls`` branch headers inside the function body."""
+    body = _APP_JS.read_text(encoding="utf-8")
+    start = body.index("Pane.prototype.replayHistory = function")
+    end = body.index("Pane.prototype._attachRetryToLastAssistant", start)
+    fn = body[start:end]
+    # Locate the assistant branch and bound the search to its body —
+    # the function also handles user / tool roles which would otherwise
+    # confuse the offset comparison.
+    asst_start = fn.index('msg.role === "assistant"')
+    asst_end = fn.index('msg.role === "tool"', asst_start)
+    asst = fn[asst_start:asst_end]
+    content_idx = asst.index("if (msg.content)")
+    tool_calls_idx = asst.index("if (msg.tool_calls && msg.tool_calls.length)")
+    assert content_idx < tool_calls_idx, (
+        "replayHistory must render msg.content BEFORE msg.tool_calls "
+        "inside the assistant branch — otherwise the lastToolBlock "
+        "anchor is clobbered before the next iteration's tool result "
+        "can attach to it (and the visual order also drifts from the "
+        "live SSE flow)."
+    )
+
+
+def test_replay_history_renders_persisted_verdict_badge() -> None:
+    """Saved-workstream replays must paint the persisted intent verdict
+    next to each tool div, using the same ``renderVerdictBadge`` helper
+    the live ``showInlineToolBlock`` path uses. Pre-fix the audit trail
+    was complete in storage (``intent_verdicts`` table) but never
+    surfaced on replay — operators reviewing a saved workstream
+    couldn't see what the heuristic / LLM judge thought of any tool
+    call. This test pins the call site so a refactor that drops the
+    decoration regresses the audit surface."""
+    body = _APP_JS.read_text(encoding="utf-8")
+    start = body.index("Pane.prototype.replayHistory = function")
+    end = body.index("Pane.prototype._attachRetryToLastAssistant", start)
+    fn = body[start:end]
+    # Match a `renderVerdictBadge(<something>.verdict, ...)` call inside
+    # the replay loop.  Loose on whitespace + identifier so a future
+    # rename of the iteration variable doesn't trip CI.
+    badge_call_re = re.compile(
+        r"renderVerdictBadge\(\s*\w+\.verdict\b",
+    )
+    assert badge_call_re.search(fn), (
+        "replayHistory must call renderVerdictBadge(tc.verdict, ...) "
+        "when a persisted verdict is attached to a tool_call entry — "
+        "otherwise the audit-trail data persisted to intent_verdicts "
+        "doesn't surface on saved-workstream replays."
+    )

tests/test_close_reason_persistence.py

@@ -30,9 +30,25 @@ def _full_hdr() -> dict[str, str]:
     }
 
 
+@pytest.fixture(autouse=True)
+def _isolate_metrics(monkeypatch):
+    """Swap ``turnstone.server._metrics`` for a fresh collector
+    per-test, with auto-restore.
+
+    Bare ``srv_mod._metrics = MetricsCollector()`` (the prior
+    pattern) leaks into any test file that already bound the name
+    via ``from turnstone.server import _metrics`` at import time —
+    those tests' patches then operate on a different instance from
+    the one the live ``_publish_models_metadata`` reads, and the
+    monkeypatch silently no-ops.  ``monkeypatch.setattr`` restores
+    after the test, so the leak is contained.
+    """
+    fresh = MetricsCollector()
+    fresh.model = "test-model"
+    monkeypatch.setattr(srv_mod, "_metrics", fresh)
+
+
 def _make_app(storage: Any) -> TestClient:
-    srv_mod._metrics = MetricsCollector()
-    srv_mod._metrics.model = "test-model"
     mock_session = MagicMock()
     mock_ws = MagicMock()
     mock_ws.id = "ws-target"

tests/test_coordinator_client.py

@@ -949,6 +949,137 @@ def test_list_nodes_empty_on_no_matching_filters(storage_with_nodes):
     assert result["truncated"] is False
 
 
+def test_list_nodes_surfaces_healthy_model_aliases(tmp_path):
+    """The node's heartbeat loop projects its registry into a ``models``
+    metadata entry shaped like ``[{alias, provider, healthy}, ...]``.
+    ``list_nodes`` flattens that to the healthy-alias list at the top
+    level (under ``model_aliases``) so a coordinator can pass aliases
+    straight to ``spawn_workstream(model=)`` without having to
+    introspect the metadata blob.  The provider-side model identifier
+    (``cfg.model``) is intentionally NOT in the payload — coords kept
+    reaching for it when they should pass the local alias."""
+    st = SQLiteBackend(str(tmp_path / "nodes.db"))
+    _set_meta(
+        st,
+        "node-x",
+        [
+            ("arch", "x86_64", "auto"),
+            (
+                "models",
+                [
+                    {"alias": "gpt5", "provider": "openai", "healthy": True},
+                    {"alias": "claude-opus-47", "provider": "anthropic", "healthy": True},
+                    {"alias": "broken", "provider": "openai", "healthy": False},
+                ],
+                "auto",
+            ),
+        ],
+    )
+    _register_service(st, "node-x")
+    client = _make_read_client(st)
+    result = client.list_nodes()
+    node = result["nodes"][0]
+    assert node["model_aliases"] == ["gpt5", "claude-opus-47"]
+    # Full per-alias info still available under metadata for callers
+    # that want provider / healthy detail (e.g. surfacing degraded
+    # aliases in a UI).
+    full = node["metadata"]["models"]["value"]
+    assert {row["alias"] for row in full} == {"gpt5", "claude-opus-47", "broken"}
+    # ``model`` (the provider-side identifier) is intentionally absent
+    # — keep the payload to the three values a coord actually uses.
+    for row in full:
+        assert "model" not in row
+
+
+def test_list_nodes_model_aliases_distinct_from_metadata_models(tmp_path):
+    """Pin the naming distinction explicitly: the top-level shortlist
+    (``model_aliases``, list of strings) and the rich metadata blob
+    (``metadata.models.value``, list of dicts) live under different
+    keys so a caller that confuses them gets a clear KeyError rather
+    than a silent shape mismatch."""
+    st = SQLiteBackend(str(tmp_path / "nodes.db"))
+    _set_meta(
+        st,
+        "node-x",
+        [
+            (
+                "models",
+                [{"alias": "a", "provider": "openai", "healthy": True}],
+                "auto",
+            ),
+        ],
+    )
+    _register_service(st, "node-x")
+    client = _make_read_client(st)
+    node = client.list_nodes()["nodes"][0]
+    # No top-level ``models`` field — only ``model_aliases``.
+    assert "models" not in node
+    assert node["model_aliases"] == ["a"]
+    # Rich shape stays under metadata.
+    assert isinstance(node["metadata"]["models"]["value"], list)
+    assert isinstance(node["metadata"]["models"]["value"][0], dict)
+
+
+def test_list_nodes_model_aliases_empty_when_node_has_not_published(tmp_path):
+    """Nodes from older builds — or a node mid-startup before its first
+    metadata write — won't have a ``models`` entry.  The top-level
+    ``model_aliases`` field defaults to ``[]`` rather than being
+    omitted so coordinators can rely on the key being present."""
+    st = SQLiteBackend(str(tmp_path / "nodes.db"))
+    _set_meta(st, "node-y", [("arch", "x86_64", "auto")])
+    _register_service(st, "node-y")
+    client = _make_read_client(st)
+    result = client.list_nodes()
+    assert result["nodes"][0]["model_aliases"] == []
+
+
+def test_list_nodes_models_tolerates_malformed_entries(tmp_path):
+    """If a node ever stores a malformed ``models`` entry (wrong outer
+    type, missing alias, non-bool healthy), the projection drops the
+    bad rows rather than raising — the rest of the response should
+    still be useful."""
+    st = SQLiteBackend(str(tmp_path / "nodes.db"))
+    _set_meta(
+        st,
+        "node-z",
+        [
+            (
+                "models",
+                [
+                    {"alias": "ok", "provider": "p", "healthy": True},
+                    "not-a-dict",
+                    {"provider": "p", "healthy": True},  # missing alias
+                    {"alias": "", "healthy": True},  # empty alias
+                    {"alias": "degraded", "healthy": False},
+                    {"alias": 42, "healthy": True},  # non-string alias
+                ],
+                "auto",
+            ),
+        ],
+    )
+    _register_service(st, "node-z")
+    client = _make_read_client(st)
+    result = client.list_nodes()
+    assert result["nodes"][0]["model_aliases"] == ["ok"]
+
+
+def test_list_nodes_models_handles_non_list_payload(tmp_path):
+    """A node with a corrupted models entry (dict, scalar, null) shouldn't
+    blow up the whole list_nodes call.  ``model_aliases`` falls back to ``[]``."""
+    st = SQLiteBackend(str(tmp_path / "nodes.db"))
+    _set_meta(
+        st,
+        "node-w",
+        [
+            ("models", {"oops": "not a list"}, "auto"),
+        ],
+    )
+    _register_service(st, "node-w")
+    client = _make_read_client(st)
+    result = client.list_nodes()
+    assert result["nodes"][0]["model_aliases"] == []
+
+
 # ---------------------------------------------------------------------------
 # list_skills
 # ---------------------------------------------------------------------------

tests/test_coordinator_page.py

@@ -157,6 +157,16 @@ def test_coordinator_js_exposes_inline_approval_helpers():
     # any prior denial.  bug-1 / bug-3 from the second /review pass.
     assert "Denied by user" in body
     assert "callOutcomes" in body
+    # User-message attachment pills — both live send (coordSend) and
+    # history replay route through appendUserMessageWithAttachments.
+    # Renaming or dropping the helper would silently regress the
+    # attachment affordance to the pre-fix plain-text bubble, which
+    # would only surface in manual testing of an attached-file flow.
+    # The CSS class is the visual anchor (coordinator.css) — keeping
+    # both literals in the smoke layer covers JS↔CSS drift in either
+    # direction.
+    assert "function appendUserMessageWithAttachments" in body
+    assert "msg-user-attach" in body
 
 
 def test_coordinator_js_handle_child_state_no_longer_reads_sse_pending_approval_detail():

tests/test_history_decoration.py

@@ -0,0 +1,232 @@
+"""Unit tests for ``turnstone.core.history_decoration``.
+
+The decoration helpers are shared between two surfaces — interactive's
+SSE replay (``_build_history``) and the lifted ``/history`` REST
+endpoint (``make_history_handler``, used by both interactive and
+coord). Pinning the wire shape here lets a future schema/projection
+change land in one file rather than spread across the two surfaces.
+"""
+
+from __future__ import annotations
+
+from turnstone.core.history_decoration import (
+    build_output_assessment_payload,
+    build_verdict_payload,
+    decorate_history_messages,
+    decorate_tool_call,
+)
+
+
+class TestBuildVerdictPayload:
+    """The wire-shape projection that's the single source of truth for
+    what intent_verdict fields ship to the client."""
+
+    def test_skips_unflagged_baseline(self) -> None:
+        """``risk_level`` "none" is the unflagged-tool baseline; the
+        client filters those anyway, so projecting None at the wire
+        layer keeps the payload tight on long workstreams."""
+        row = {"risk_level": "none", "recommendation": "approve", "tier": "heuristic"}
+        assert build_verdict_payload(row) is None
+
+    def test_drops_call_id_and_func_name(self) -> None:
+        """The client already has these on ``tc.id`` / ``tc.name``;
+        re-shipping them per-tool_call would balloon long replays."""
+        row = {
+            "call_id": "call_abc",
+            "func_name": "bash",
+            "risk_level": "medium",
+            "recommendation": "review",
+            "confidence": 0.8,
+            "intent_summary": "summary",
+            "tier": "heuristic",
+        }
+        out = build_verdict_payload(row)
+        assert out is not None
+        assert "call_id" not in out
+        assert "func_name" not in out
+        # Sanity — the kept fields are the ones renderVerdictBadge reads.
+        assert out["risk_level"] == "medium"
+        assert out["recommendation"] == "review"
+        assert out["confidence"] == 0.8
+        assert out["intent_summary"] == "summary"
+        assert out["tier"] == "heuristic"
+
+    def test_includes_reasoning_for_either_tier_when_present(self) -> None:
+        """Heuristic verdicts in this project emit structured
+        rationales (one per matched pattern) — e.g.
+        ``policy.py`` writes a reasoning string per heuristic hit.
+        Ship the field for either tier when it has content; only
+        omit when the row didn't write one."""
+        for tier in ("heuristic", "llm"):
+            row = {
+                "risk_level": "high",
+                "tier": tier,
+                "reasoning": "The command exfiltrates ~/.ssh/id_rsa over an external connection.",
+            }
+            out = build_verdict_payload(row)
+            assert out is not None
+            assert "id_rsa" in out["reasoning"]
+
+    def test_omits_reasoning_when_empty(self) -> None:
+        """An absent / empty reasoning string shouldn't ship as
+        ``reasoning: ""`` — the rationale ``<details>`` block on the
+        client renders an empty disclosure when the field is present
+        but empty."""
+        row = {"risk_level": "high", "tier": "heuristic", "reasoning": ""}
+        out = build_verdict_payload(row)
+        assert out is not None
+        assert "reasoning" not in out
+
+    def test_includes_judge_model_when_present(self) -> None:
+        """``judge_model`` rides through so the batch tier badge can
+        render ``⚖ llm:claude-haiku-4`` on history-only replays
+        rather than the bare ``⚖ llm`` label."""
+        row = {"risk_level": "high", "tier": "llm", "judge_model": "claude-haiku-4"}
+        out = build_verdict_payload(row)
+        assert out is not None
+        assert out["judge_model"] == "claude-haiku-4"
+
+    def test_omits_judge_model_when_empty(self) -> None:
+        row = {"risk_level": "medium", "tier": "heuristic", "judge_model": ""}
+        out = build_verdict_payload(row)
+        assert out is not None
+        assert "judge_model" not in out
+
+
+class TestBuildOutputAssessmentPayload:
+    """Output-guard wire shape — flags decoded from JSON string at
+    this layer so the client never has to parse twice."""
+
+    def test_skips_unflagged_baseline(self) -> None:
+        row = {"risk_level": "none", "flags": "[]"}
+        assert build_output_assessment_payload(row) is None
+
+    def test_decodes_flags_from_json(self) -> None:
+        row = {"risk_level": "high", "flags": '["api_key","email"]', "redacted": 1}
+        out = build_output_assessment_payload(row)
+        assert out is not None
+        assert out["flags"] == ["api_key", "email"]
+        assert out["redacted"] is True
+        assert out["risk_level"] == "high"
+
+    def test_handles_malformed_flags_json(self) -> None:
+        """Bad JSON in ``flags`` must not block the rest of the
+        assessment from rendering — degrade to empty list."""
+        row = {"risk_level": "medium", "flags": "not-json", "redacted": 0}
+        out = build_output_assessment_payload(row)
+        assert out is not None
+        assert out["flags"] == []
+        assert out["redacted"] is False
+
+
+class TestDecorateToolCall:
+    """In-place mutation of either OpenAI-format or flattened tool_call
+    entries — both shapes carry ``id`` at the top level."""
+
+    def test_attaches_verdict_when_present(self) -> None:
+        tc: dict[str, object] = {"id": "call_1", "function": {"name": "bash", "arguments": "{}"}}
+        verdicts = {
+            "call_1": {
+                "risk_level": "medium",
+                "recommendation": "review",
+                "confidence": 0.7,
+                "intent_summary": "summary",
+                "tier": "heuristic",
+            }
+        }
+        decorate_tool_call(tc, verdicts, {})
+        assert "verdict" in tc
+        assert tc["verdict"]["risk_level"] == "medium"  # type: ignore[index]
+
+    def test_skips_when_no_call_id_match(self) -> None:
+        tc: dict[str, object] = {"id": "call_other", "name": "bash"}
+        verdicts = {
+            "call_1": {"risk_level": "medium", "tier": "heuristic"},
+        }
+        decorate_tool_call(tc, verdicts, {})
+        assert "verdict" not in tc
+
+    def test_skips_unflagged_verdict(self) -> None:
+        """``build_verdict_payload`` returns None for unflagged rows;
+        decorate_tool_call must not stamp ``verdict`` in that case."""
+        tc: dict[str, object] = {"id": "call_1", "name": "bash"}
+        verdicts = {"call_1": {"risk_level": "none", "tier": "heuristic"}}
+        decorate_tool_call(tc, verdicts, {})
+        assert "verdict" not in tc
+
+    def test_handles_empty_id(self) -> None:
+        """A tool_call with no id can't be paired against the lookup
+        table — must not raise (or stamp the wrong row's verdict)."""
+        tc: dict[str, object] = {"id": "", "name": "bash"}
+        verdicts = {"call_1": {"risk_level": "high", "tier": "heuristic"}}
+        decorate_tool_call(tc, verdicts, {})
+        assert "verdict" not in tc
+
+
+class TestDecorateHistoryMessages:
+    """End-to-end mutation of a /history-shaped message list — covers
+    the full transform applied by ``make_history_handler``."""
+
+    def test_decorates_tool_calls_and_marks_truncated(self) -> None:
+        verdicts = {
+            "call_a": {
+                "risk_level": "high",
+                "recommendation": "deny",
+                "confidence": 0.95,
+                "intent_summary": "exfil",
+                "tier": "llm",
+                "reasoning": "ssh key access",
+            }
+        }
+        assessments = {
+            "call_a": {"risk_level": "high", "flags": '["secret"]', "redacted": 1},
+        }
+        # Tool result content of exactly TOOL_RESULT_STORAGE_CAP chars
+        # hits the storage cap (longer is impossible — storage clamps
+        # at the cap).  Reference the constant rather than a literal so
+        # this test stays correct if the cap moves again.
+        from turnstone.core.history_decoration import TOOL_RESULT_STORAGE_CAP
+
+        truncated_content = "x" * TOOL_RESULT_STORAGE_CAP
+        messages: list[dict[str, object]] = [
+            {"role": "user", "content": "hi"},
+            {
+                "role": "assistant",
+                "content": "running",
+                "tool_calls": [
+                    {
+                        "id": "call_a",
+                        "function": {"name": "bash", "arguments": "{}"},
+                    }
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_a", "content": truncated_content},
+            {"role": "tool", "tool_call_id": "call_b", "content": "short"},
+        ]
+        decorate_history_messages(messages, verdicts, assessments)
+        # Assistant tool_calls got both decorations.
+        tc = messages[1]["tool_calls"][0]  # type: ignore[index]
+        assert tc["verdict"]["risk_level"] == "high"
+        assert tc["verdict"]["tier"] == "llm"
+        assert "reasoning" in tc["verdict"]
+        assert tc["output_assessment"]["flags"] == ["secret"]
+        assert tc["output_assessment"]["redacted"] is True
+        # Truncated tool message got the flag; the short one did not.
+        assert messages[2].get("truncated") is True
+        assert "truncated" not in messages[3]
+
+    def test_no_op_on_empty_indexes(self) -> None:
+        """When neither table has rows for the workstream, the wire
+        shape passes through unchanged — replay must degrade
+        gracefully when verdict storage is empty / unavailable."""
+        messages: list[dict[str, object]] = [
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [{"id": "call_a", "function": {"name": "bash", "arguments": "{}"}}],
+            },
+        ]
+        decorate_history_messages(messages, {}, {})
+        tc = messages[0]["tool_calls"][0]  # type: ignore[index]
+        assert "verdict" not in tc
+        assert "output_assessment" not in tc

tests/test_memory_relevance.py

@@ -1,6 +1,9 @@
 """Tests for turnstone.core.memory_relevance — scoring, formatting, context extraction."""
 
+from unittest.mock import patch
+
 from turnstone.core.memory_relevance import (
+    MemoryConfig,
     build_memory_context,
     extract_recent_context,
     score_memories,
@@ -192,3 +195,265 @@ class TestExtractRecentContext:
 
     def test_empty_messages(self):
         assert extract_recent_context([]) == ""
+
+
+# ---------------------------------------------------------------------------
+# Composition candidate-selection (_init_system_messages)
+# ---------------------------------------------------------------------------
+
+
+def _make_mem(name: str, content: str = "", memory_id: str | None = None) -> dict[str, str]:
+    return {
+        "name": name,
+        "memory_id": memory_id or f"mid_{name}",
+        "type": "project",
+        "scope": "global",
+        "scope_id": "",
+        "description": "",
+        "content": content or name,
+        "updated": "2024-01-01T00:00:00",
+    }
+
+
+def _make_session(fetch_limit: int = 5, relevance_k: int = 3, **kwargs: object):
+    """Composition tests need a real ChatSession (constructor calls
+    ``_init_system_messages`` once, unpatched, before the test gets a chance
+    to install patches).  ``tmp_db`` initializes the storage singleton that
+    constructor needs; tests then patch the visibility helpers and call
+    ``_init_system_messages`` a second time to exercise the new logic.
+    """
+    from tests._helpers import make_chat_session
+
+    return make_chat_session(
+        memory_config=MemoryConfig(fetch_limit=fetch_limit, relevance_k=relevance_k),
+        **kwargs,
+    )
+
+
+class TestCompositionCandidateSelection:
+    """Verify the query-aware candidate set in _init_system_messages."""
+
+    def test_recency_ceiling_regression(self, tmp_db):
+        """Old relevant memory not in recency top-N still injected via search path."""
+        session = _make_session(fetch_limit=5, relevance_k=3)
+        session.messages = [{"role": "user", "content": "postgres database configuration"}]
+
+        old_mem = _make_mem(
+            "ancient_db_config",
+            content="postgres database configuration connection host port",
+            memory_id="m_old",
+        )
+        # Recency top-5 do not include old_mem
+        recent = [_make_mem(f"recent_{i}", memory_id=f"mr{i}") for i in range(5)]
+
+        with (
+            patch.object(session, "_search_visible_memories", return_value=[old_mem]),
+            patch.object(session, "_list_visible_memories", return_value=recent),
+        ):
+            session._init_system_messages()
+
+        joined = "\n".join(m["content"] for m in session.system_messages if m["role"] == "system")
+        # With the fix, old_mem enters the candidate pool via search and wins BM25
+        assert "ancient_db_config" in joined
+
+    def test_empty_query_falls_back_to_recency(self, tmp_db):
+        """No user messages → empty context → recency path, search never called."""
+        session = _make_session()
+        session.messages = []  # extract_recent_context returns ""
+
+        recency = [_make_mem("note_alpha"), _make_mem("note_beta")]
+
+        with (
+            patch.object(session, "_list_visible_memories", return_value=recency),
+            patch.object(session, "_search_visible_memories") as search_mock,
+        ):
+            session._init_system_messages()
+
+        search_mock.assert_not_called()
+        joined = "\n".join(m["content"] for m in session.system_messages if m["role"] == "system")
+        assert "note_alpha" in joined
+
+    def test_sparse_match_union_fills_candidate_pool(self, tmp_db):
+        """Search returning < fetch_limit results unions with recency fillers."""
+        session = _make_session(fetch_limit=5, relevance_k=4)
+        session.messages = [{"role": "user", "content": "unique_term xyzzy"}]
+
+        hit_a = _make_mem("hit_alpha", content="unique_term xyzzy alpha", memory_id="m_ha")
+        hit_b = _make_mem("hit_beta", content="unique_term xyzzy beta", memory_id="m_hb")
+        search_hits = [hit_a, hit_b]  # 2 < fetch_limit=5 → triggers union
+
+        # Recency overlaps on hit_a/hit_b and adds 3 fillers
+        filler = [_make_mem(f"filler_{i}", memory_id=f"mf{i}") for i in range(3)]
+        recency = [hit_a, hit_b] + filler
+
+        with (
+            patch.object(session, "_search_visible_memories", return_value=search_hits),
+            patch.object(session, "_list_visible_memories", return_value=recency),
+        ):
+            session._init_system_messages()
+
+        joined = "\n".join(m["content"] for m in session.system_messages if m["role"] == "system")
+        # Both hits match "unique_term xyzzy" well → appear after BM25 ranking
+        assert "hit_alpha" in joined
+        assert "hit_beta" in joined
+
+    def test_recency_preserved_when_search_returns_noise_above_relevance_k(self, tmp_db):
+        """Pool guarantee: recency-50 always reaches BM25, even when search
+        returns enough noise hits to clear ``relevance_k``.
+
+        Closes the narrow regression vs. the original bug — without the
+        ``fetch_limit`` threshold, a stopword-dominated cap-search that
+        returned >= relevance_k irrelevant hits would short-circuit and
+        evict the recency-only memory the bug had been surfacing.
+        """
+        session = _make_session(fetch_limit=10, relevance_k=3)
+        session.messages = [{"role": "user", "content": "configure host"}]
+
+        # Search returns relevance_k=3 noise hits — enough to skip recency
+        # under the OLD threshold, not enough to fill fetch_limit=10.
+        noise = [
+            _make_mem(f"noise_{i}", content="generic content", memory_id=f"mn{i}") for i in range(3)
+        ]
+        # The memory the user actually wants — distinctive, in recency,
+        # but its content doesn't share any token with the noise hits.
+        wanted = _make_mem(
+            "host_config_v2",
+            content="host=localhost port=5432 db=production",
+            memory_id="m_wanted",
+        )
+        recency = [wanted] + [_make_mem(f"recent_{i}", memory_id=f"mr{i}") for i in range(5)]
+
+        with (
+            patch.object(session, "_search_visible_memories", return_value=noise),
+            patch.object(session, "_list_visible_memories", return_value=recency),
+        ):
+            session._init_system_messages()
+
+        joined = "\n".join(m["content"] for m in session.system_messages if m["role"] == "system")
+        # ``wanted`` reached BM25 via the union and matched "host" → injected.
+        assert "host_config_v2" in joined
+
+    def test_recency_tail_preserved_when_search_adds_distinct_hits(self, tmp_db):
+        """SUPERSET invariant: every recency item is in the candidate pool
+        when search adds hits, even if the resulting union exceeds
+        fetch_limit.  Truncating the union at fetch_limit (the prior
+        behavior) evicted the recency tail — which is exactly where
+        ancient-but-recently-touched memories live, the recall this PR
+        sets out to improve.
+        """
+        session = _make_session(fetch_limit=10, relevance_k=3)
+        session.messages = [{"role": "user", "content": "alpha"}]
+
+        # 5 search hits, none of which appear in recency.
+        search_hits = [
+            _make_mem(f"search_{i}", content="alpha", memory_id=f"ms{i}") for i in range(5)
+        ]
+        # 10 recency items; without the union uncap, the 5 oldest of these
+        # would be displaced by the 5 search hits.
+        recency = [_make_mem(f"recency_{i}", memory_id=f"mr{i}") for i in range(10)]
+
+        with (
+            patch.object(session, "_search_visible_memories", return_value=search_hits),
+            patch.object(session, "_list_visible_memories", return_value=recency),
+        ):
+            candidates, source = session._select_memory_candidates("alpha")
+
+        candidate_ids = {c["memory_id"] for c in candidates}
+        # Pool is search_hits ∪ recency — 15 items, no truncation.
+        assert len(candidates) == 15
+        assert source == "union"
+        # Every recency item present (no tail eviction).
+        for i in range(10):
+            assert f"mr{i}" in candidate_ids, f"recency item {i} evicted"
+        # And every search hit is also in the pool.
+        for i in range(5):
+            assert f"ms{i}" in candidate_ids, f"search hit {i} missing"
+
+    def test_coord_scope_isolated_visibility(self, tmp_db):
+        """Coord composition queries the coord scope alone, never the
+        global/workstream/user union."""
+        from turnstone.core.workstream import WorkstreamKind
+
+        coord = _make_session(
+            fetch_limit=5,
+            relevance_k=3,
+            ws_id="coord-1",
+            user_id="user-1",
+            kind=WorkstreamKind.COORDINATOR,
+        )
+        scopes = coord._visible_scopes()
+        assert scopes == [("coordinator", "coord-1")]
+        # And: search uses those same scopes (no global/user fan-in)
+        coord.messages = [{"role": "user", "content": "anything"}]
+        with patch(
+            "turnstone.core.session.search_visible_structured_memories",
+            return_value=[],
+        ) as search_mock:
+            coord._search_visible_memories("anything", limit=5)
+        search_mock.assert_called_once()
+        # Second positional arg is the scopes list
+        assert search_mock.call_args.args[1] == [("coordinator", "coord-1")]
+
+
+class TestMemorySearchToolExecution:
+    """End-to-end test of ``memory(action='search')`` through _exec_memory.
+
+    Drives the actual tool dispatch (not just the storage facade) so the
+    OR-of-terms fix and the coalesced ``memory.search`` log get exercised
+    together.
+    """
+
+    def test_search_action_returns_or_of_terms_results(self, tmp_db):
+        """Multi-word query returns rows where ANY term matches — not all."""
+        from turnstone.core.memory import save_structured_memory
+
+        save_structured_memory("postgres_notes", "host=localhost port=5432")
+        save_structured_memory("redis_notes", "host=redis port=6379")
+        save_structured_memory("unrelated", "completely different")
+
+        session = _make_session()
+        item = session._prepare_memory(
+            "call-1",
+            {"action": "search", "query": "postgres no_such_word_a no_such_word_b"},
+        )
+        # Sanity: prepare returned a search-ready dispatch (not an error item)
+        assert item.get("action") == "search"
+
+        call_id, msg = session._exec_memory(item)
+        assert call_id == "call-1"
+        assert "postgres_notes" in msg
+        # Other memories don't match any query term
+        assert "unrelated" not in msg
+
+
+class TestPerTurnSearchCache:
+    """The per-turn cache spares redundant SQL across mid-turn rebuilds."""
+
+    def test_repeated_search_in_same_turn_hits_cache(self, tmp_db):
+        from turnstone.core.memory import save_structured_memory
+
+        save_structured_memory("hello_mem", "alpha beta gamma")
+        session = _make_session()
+        with patch(
+            "turnstone.core.session.search_visible_structured_memories",
+            return_value=[],
+        ) as backend_mock:
+            session._search_visible_memories("alpha beta", limit=5)
+            session._search_visible_memories("alpha beta", limit=5)
+            session._search_visible_memories("alpha beta", limit=5)
+        # 3 calls but only 1 backend hit — cache absorbed the rest
+        assert backend_mock.call_count == 1
+
+    def test_user_turn_invalidates_cache(self, tmp_db):
+        from turnstone.core.memory import save_structured_memory
+
+        save_structured_memory("hello_mem", "alpha")
+        session = _make_session()
+        with patch(
+            "turnstone.core.session.search_visible_structured_memories",
+            return_value=[],
+        ) as backend_mock:
+            session._search_visible_memories("alpha", limit=5)
+            session._invalidate_memory_cache()  # simulates new user turn
+            session._search_visible_memories("alpha", limit=5)
+        assert backend_mock.call_count == 2

tests/test_model_registry.py

@@ -1126,24 +1126,57 @@ class TestSessionAgentModel:
     def _captured_effort(captured: dict[str, Any]) -> str | None:
         """Pull reasoning_effort out of provider-specific shapes.
 
-        openai-compatible servers receive it via extra_body.chat_template_kwargs;
-        commercial providers receive it as a top-level kwarg.
+        Chat Completions delivers it as a top-level ``reasoning_effort`` kwarg
+        (when the model's caps permit it).  Operators who route reasoning_effort
+        through ``chat_template_kwargs`` (gpt-oss-style local templates) get
+        it inside ``extra_body.chat_template_kwargs``.
         """
+        if "reasoning_effort" in captured:
+            return captured["reasoning_effort"]
         eb = captured.get("extra_body") or {}
         ctk = eb.get("chat_template_kwargs") or {}
-        return ctk.get("reasoning_effort") or captured.get("reasoning_effort")
+        return ctk.get("reasoning_effort")
+
+    @staticmethod
+    def _effort_caps() -> dict[str, Any]:
+        """Capabilities that allow Chat-Completions reasoning_effort to flow."""
+        return {
+            "reasoning_effort_values": [
+                "minimal",
+                "low",
+                "medium",
+                "high",
+                "max",
+            ],
+        }
 
     def _three_model_registry(self, **kwargs: Any) -> ModelRegistry:
+        caps = self._effort_caps()
         return ModelRegistry(
             models={
                 "main": ModelConfig(
-                    "main", "http://m/v1", "k", "main-model", provider="openai-compatible"
+                    "main",
+                    "http://m/v1",
+                    "k",
+                    "main-model",
+                    provider="openai-compatible",
+                    capabilities=dict(caps),
                 ),
                 "smart": ModelConfig(
-                    "smart", "http://s/v1", "k", "smart-model", provider="openai-compatible"
+                    "smart",
+                    "http://s/v1",
+                    "k",
+                    "smart-model",
+                    provider="openai-compatible",
+                    capabilities=dict(caps),
                 ),
                 "fast": ModelConfig(
-                    "fast", "http://f/v1", "k", "fast-model", provider="openai-compatible"
+                    "fast",
+                    "http://f/v1",
+                    "k",
+                    "fast-model",
+                    provider="openai-compatible",
+                    capabilities=dict(caps),
                 ),
             },
             default="main",
@@ -1249,6 +1282,45 @@ class TestSessionAgentModel:
         session._run_agent([{"role": "user", "content": "x"}], label="plan", agent_alias="fast")
         assert captured["model"] == "fast-model"
 
+    def test_session_fallback_inherits_primary_alias_for_caps(self) -> None:
+        """When _run_agent has no registry agent route, it must fall back to
+        the session's primary alias for capability and server_compat lookup —
+        otherwise per-model caps (reasoning_effort_values, server_compat) get
+        silently dropped on the agent path."""
+        reg = self._three_model_registry()  # no agent_model / plan_model set
+        session = _make_session(registry=reg, model_alias="main")
+        # Probe what _run_agent passes to _provider_extra_params and
+        # _resolve_capabilities by recording the model_alias on each call.
+        captured_extra_alias: list[str | None] = []
+        captured_resolve_alias: list[str | None] = []
+        original_extra = session._provider_extra_params
+        original_resolve = session._resolve_capabilities
+
+        def spy_extra(*args: Any, **kwargs: Any) -> Any:
+            captured_extra_alias.append(kwargs.get("model_alias"))
+            return original_extra(*args, **kwargs)
+
+        def spy_resolve(*args: Any, **kwargs: Any) -> Any:
+            # _resolve_capabilities(provider, model, alias)
+            alias = args[2] if len(args) >= 3 else kwargs.get("alias")
+            captured_resolve_alias.append(alias)
+            return original_resolve(*args, **kwargs)
+
+        session._provider_extra_params = spy_extra  # type: ignore[method-assign]
+        session._resolve_capabilities = spy_resolve  # type: ignore[method-assign]
+
+        self._capture_on(session.client)  # patch client.chat.completions.create
+        session._run_agent([{"role": "user", "content": "x"}], label="plan")
+
+        assert captured_extra_alias and captured_extra_alias[-1] == "main", (
+            f"agent fallback path did not inherit primary alias for extra_params: "
+            f"{captured_extra_alias!r}"
+        )
+        assert captured_resolve_alias and captured_resolve_alias[-1] == "main", (
+            f"agent fallback path did not inherit primary alias for caps: "
+            f"{captured_resolve_alias!r}"
+        )
+
     def test_invalid_alias_raises_in_run_agent(self) -> None:
         """Defence-in-depth: _prepare_* validates first, but _run_agent
         rejects unknown aliases too rather than silently falling back."""

tests/test_models_changed_event.py

@@ -0,0 +1,274 @@
+"""``models_changed`` SSE fanout coverage.
+
+The console pushes a ``models_changed`` cluster event whenever a model
+definition is created / updated / deleted / reloaded, or whenever a
+setting in :data:`turnstone.console.server._MODEL_AFFECTING_SETTING_KEYS`
+is updated or reset.  Connected browsers refetch ``/v1/api/models`` on
+receipt so the home composer dropdown + admin Models → Roles sub-tab
+reflect alias edits without a manual reload.
+
+These tests pin two contracts:
+
+- every model-definition CRUD path emits exactly one ``models_changed``
+  fanout (so the browser stays in sync with the DB);
+- settings PUT / DELETE only emit the fanout when the key is
+  model-affecting — unrelated keys (e.g. ``session.retention_days``)
+  must not trigger spurious dropdown re-renders across the cluster.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+from starlette.applications import Starlette
+from starlette.middleware import Middleware
+from starlette.routing import Route
+from starlette.testclient import TestClient
+
+from tests._coord_test_helpers import _AuthMiddleware
+from turnstone.console.server import (
+    _MODEL_AFFECTING_SETTING_KEYS,
+    admin_create_model_definition,
+    admin_delete_model_definition,
+    admin_delete_setting,
+    admin_model_reload,
+    admin_update_model_definition,
+    admin_update_setting,
+)
+from turnstone.core.storage._sqlite import SQLiteBackend
+
+
+@pytest.fixture
+def storage(tmp_path: Any) -> SQLiteBackend:
+    return SQLiteBackend(str(tmp_path / "models_changed.db"))
+
+
+def _seed(storage: SQLiteBackend, *, definition_id: str, alias: str) -> None:
+    storage.create_model_definition(
+        definition_id=definition_id,
+        alias=alias,
+        model="model-x",
+        provider="openai-compatible",
+        base_url="http://localhost:8000/v1",
+        api_key="sk-test",
+        context_window=8192,
+        capabilities="{}",
+        enabled=True,
+        created_by="admin",
+    )
+
+
+def _make_client(storage: SQLiteBackend) -> tuple[TestClient, MagicMock]:
+    """Build a TestClient + return the stub collector for assertion.
+
+    Wires the four model-definition CRUD/reload routes plus the two
+    settings mutation routes.  Collector is a MagicMock so each
+    ``emit_models_changed`` call lands as a recorded call without
+    spinning up the full SSE listener queue.
+    """
+    app = Starlette(
+        routes=[
+            Route(
+                "/v1/api/admin/model-definitions",
+                admin_create_model_definition,
+                methods=["POST"],
+            ),
+            Route(
+                "/v1/api/admin/model-definitions/reload",
+                admin_model_reload,
+                methods=["POST"],
+            ),
+            Route(
+                "/v1/api/admin/model-definitions/{definition_id}",
+                admin_update_model_definition,
+                methods=["PUT"],
+            ),
+            Route(
+                "/v1/api/admin/model-definitions/{definition_id}",
+                admin_delete_model_definition,
+                methods=["DELETE"],
+            ),
+            Route(
+                "/v1/api/admin/settings/{key:path}",
+                admin_update_setting,
+                methods=["PUT"],
+            ),
+            Route(
+                "/v1/api/admin/settings/{key:path}",
+                admin_delete_setting,
+                methods=["DELETE"],
+            ),
+        ],
+        middleware=[Middleware(_AuthMiddleware)],
+    )
+    app.state.auth_storage = storage
+    app.state.coord_registry = None  # CRUD endpoints handle this gracefully
+    collector = MagicMock()
+    collector.get_all_nodes.return_value = []
+    app.state.collector = collector
+    app.state.proxy_client = MagicMock()
+    app.state.config_store = MagicMock()
+    client = TestClient(app)
+    client.headers.update(
+        {
+            "X-Test-User": "admin",
+            "X-Test-Perms": "admin.models,admin.settings",
+        }
+    )
+    return client, collector
+
+
+# ---------------------------------------------------------------------------
+# Model-definition CRUD endpoints fan out ``models_changed``
+# ---------------------------------------------------------------------------
+
+
+def test_create_emits_models_changed(storage: SQLiteBackend) -> None:
+    client, collector = _make_client(storage)
+    resp = client.post(
+        "/v1/api/admin/model-definitions",
+        json={
+            "alias": "fast",
+            "model": "fast-model",
+            "provider": "openai-compatible",
+            "base_url": "http://localhost:9000/v1",
+            "api_key": "sk-x",
+            "context_window": 4096,
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 1
+
+
+def test_update_emits_models_changed(storage: SQLiteBackend) -> None:
+    _seed(storage, definition_id="m1", alias="local")
+    client, collector = _make_client(storage)
+    resp = client.put(
+        "/v1/api/admin/model-definitions/m1",
+        json={"model": "swapped-model"},
+    )
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 1
+
+
+def test_update_with_empty_body_does_not_emit(storage: SQLiteBackend) -> None:
+    """Empty-body PUT writes no rows + skips the registry refresh — no
+    SSE fanout either, since nothing actually changed."""
+    _seed(storage, definition_id="m1", alias="local")
+    client, collector = _make_client(storage)
+    resp = client.put("/v1/api/admin/model-definitions/m1", json={})
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 0
+
+
+def test_delete_emits_models_changed(storage: SQLiteBackend) -> None:
+    _seed(storage, definition_id="m1", alias="local")
+    client, collector = _make_client(storage)
+    resp = client.delete("/v1/api/admin/model-definitions/m1")
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 1
+
+
+def test_reload_emits_models_changed(storage: SQLiteBackend) -> None:
+    _seed(storage, definition_id="m1", alias="local")
+    client, collector = _make_client(storage)
+    resp = client.post("/v1/api/admin/model-definitions/reload")
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# Settings PUT / DELETE only emit for model-affecting keys
+# ---------------------------------------------------------------------------
+
+
+# Pinned snapshot of the role-related keys we expect the allowlist to
+# cover today.  The frozenset itself is asserted further down so a
+# stray addition doesn't silently bypass coverage.
+_EXPECTED_AFFECTING_KEYS = frozenset(
+    {
+        "model.default_alias",
+        "model.plan_alias",
+        "model.plan_effort",
+        "model.task_alias",
+        "model.task_effort",
+        "coordinator.model_alias",
+        "coordinator.reasoning_effort",
+        "judge.model",
+    }
+)
+
+
+def _value_for_key(key: str) -> str:
+    """Return a registry-valid value for ``key``.
+
+    ``reasoning_effort`` keys have a fixed choice list; alias-shaped
+    keys accept arbitrary strings.  Avoids per-key custom payloads.
+    """
+    if (
+        key.endswith("reasoning_effort")
+        or key.endswith("plan_effort")
+        or key.endswith("task_effort")
+    ):
+        return "low"
+    return "anything"
+
+
+def test_affecting_keys_set_matches_expected() -> None:
+    """Lock in the allowlist so an unintentional removal is caught."""
+    assert _MODEL_AFFECTING_SETTING_KEYS == _EXPECTED_AFFECTING_KEYS
+
+
+@pytest.mark.parametrize("key", sorted(_EXPECTED_AFFECTING_KEYS))
+def test_settings_put_emits_for_model_affecting_key(storage: SQLiteBackend, key: str) -> None:
+    client, collector = _make_client(storage)
+    resp = client.put(
+        f"/v1/api/admin/settings/{key}",
+        json={"value": _value_for_key(key)},
+    )
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 1
+
+
+@pytest.mark.parametrize("key", sorted(_EXPECTED_AFFECTING_KEYS))
+def test_settings_delete_emits_for_model_affecting_key(storage: SQLiteBackend, key: str) -> None:
+    client, collector = _make_client(storage)
+    # Seed a row so DELETE has something to remove (otherwise 404).
+    client.put(
+        f"/v1/api/admin/settings/{key}",
+        json={"value": _value_for_key(key)},
+    )
+    collector.emit_models_changed.reset_mock()
+    resp = client.delete(f"/v1/api/admin/settings/{key}")
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 1
+
+
+def test_settings_put_does_not_emit_for_unrelated_key(
+    storage: SQLiteBackend,
+) -> None:
+    """Updating a non-model setting (here: a session retention knob)
+    must not trigger a cluster-wide dropdown refresh."""
+    client, collector = _make_client(storage)
+    resp = client.put(
+        "/v1/api/admin/settings/session.retention_days",
+        json={"value": 30},
+    )
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 0
+
+
+def test_settings_delete_does_not_emit_for_unrelated_key(
+    storage: SQLiteBackend,
+) -> None:
+    client, collector = _make_client(storage)
+    client.put(
+        "/v1/api/admin/settings/session.retention_days",
+        json={"value": 30},
+    )
+    collector.emit_models_changed.reset_mock()
+    resp = client.delete("/v1/api/admin/settings/session.retention_days")
+    assert resp.status_code == 200, resp.text
+    assert collector.emit_models_changed.call_count == 0

tests/test_providers.py

@@ -1383,6 +1383,34 @@ class TestProviderFactory:
         p2 = create_provider("openai")
         assert p1 is p2
 
+    def test_create_provider_compat_responses_surface(self) -> None:
+        """openai-compatible + api_surface=responses returns the Responses provider."""
+        from turnstone.core.providers import OpenAIResponsesProvider, create_provider
+
+        provider = create_provider("openai-compatible", api_surface="responses")
+        assert isinstance(provider, OpenAIResponsesProvider)
+
+    def test_create_provider_compat_chat_surface_default(self) -> None:
+        """openai-compatible defaults to Chat Completions."""
+        from turnstone.core.providers import create_provider
+
+        for surface in (None, "", "chat"):
+            provider = create_provider("openai-compatible", api_surface=surface)
+            assert isinstance(provider, OpenAIChatCompletionsProvider)
+
+    def test_create_provider_invalid_api_surface(self) -> None:
+        from turnstone.core.providers import create_provider
+
+        with pytest.raises(ValueError, match="Unknown api_surface"):
+            create_provider("openai-compatible", api_surface="bogus")
+
+    def test_create_provider_openai_ignores_api_surface(self) -> None:
+        """Cloud OpenAI is always Responses regardless of api_surface."""
+        from turnstone.core.providers import OpenAIResponsesProvider, create_provider
+
+        provider = create_provider("openai", api_surface="chat")
+        assert isinstance(provider, OpenAIResponsesProvider)
+
     # -- Google provider -------------------------------------------------------
 
     def test_create_provider_google(self) -> None:

tests/test_server_compat.py

@@ -89,6 +89,27 @@ class TestSuggestProfile:
         p = suggest_profile("vllm", "Google/GEMMA-4-31B-IT")
         assert p["capabilities"]["thinking_mode"] == "manual"
 
+    def test_vllm_mistral_medium_not_auto_suggested(self) -> None:
+        """Mistral medium falls back to the generic vLLM profile.
+
+        We don't auto-suggest the Responses surface for Mistral medium because
+        vLLM's Responses API tool-call parser isn't wired up for it yet —
+        operators who want per-request reasoning effort must pick "Responses
+        API" manually in the admin UI and accept the tool-calling limitation.
+        """
+        p = suggest_profile("vllm", "mistralai/Mistral-Medium-3-Instruct")
+        assert p["server_compat"]["server_type"] == "vllm"
+        assert "api_surface" not in p["server_compat"]
+        assert "capabilities" not in p
+
+    def test_vllm_mistral_medium_profile_still_available(self) -> None:
+        """The vllm-mistral-medium profile remains in _PROFILES so an operator
+        who explicitly opts in via the admin UI gets the Responses surface."""
+        from turnstone.core.server_compat import _PROFILES
+
+        assert "vllm-mistral-medium" in _PROFILES
+        assert _PROFILES["vllm-mistral-medium"]["server_compat"]["api_surface"] == "responses"
+
     def test_holo_requires_holo2(self) -> None:
         """Short 'holo' prefix shouldn't false-match; 'holo2' should match."""
         p_short = suggest_profile("vllm", "some-org/hologram-7b")
@@ -110,32 +131,47 @@ class TestSuggestProfile:
 
 
 class TestMergeServerCompat:
-    def test_empty_compat_returns_base_only(self) -> None:
+    def test_empty_base_and_compat_is_empty(self) -> None:
+        """No base, no compat → no extra_body needed."""
+        assert merge_server_compat(None, {}) == {}
+        assert merge_server_compat({}, {}) == {}
+
+    def test_explicit_base_passes_through(self) -> None:
+        """Explicit chat_template_kwargs base is forwarded as-is."""
         base = {"reasoning_effort": "medium"}
         result = merge_server_compat(base, {})
         assert result == {"chat_template_kwargs": {"reasoning_effort": "medium"}}
 
-    def test_extra_body_merged_top_level(self) -> None:
-        base = {"reasoning_effort": "medium"}
-        compat = {"extra_body": {"skip_special_tokens": False}}
-        result = merge_server_compat(base, compat)
-        assert result["skip_special_tokens"] is False
-        assert "chat_template_kwargs" in result
+    def test_extra_body_merged_top_level_no_base(self) -> None:
+        """Server-level overrides forward without a chat_template_kwargs wrapper."""
+        result = merge_server_compat(None, {"extra_body": {"skip_special_tokens": False}})
+        assert result == {"skip_special_tokens": False}
 
-    def test_full_vllm_gemma_compat(self) -> None:
-        base = {"reasoning_effort": "medium"}
+    def test_full_vllm_gemma_compat_no_base(self) -> None:
+        """vLLM workaround forwards on its own."""
         compat = {
             "server_type": "vllm",
             "extra_body": {"skip_special_tokens": False},
         }
-        result = merge_server_compat(base, compat)
+        result = merge_server_compat(None, compat)
+        assert result == {"skip_special_tokens": False}
+
+    def test_operator_chat_template_kwargs_only(self) -> None:
+        """Operator can set chat_template_kwargs explicitly without seeding the base."""
+        compat = {
+            "extra_body": {
+                "chat_template_kwargs": {"reasoning_effort": "high"},
+                "skip_special_tokens": False,
+            },
+        }
+        result = merge_server_compat(None, compat)
         assert result == {
-            "chat_template_kwargs": {"reasoning_effort": "medium"},
+            "chat_template_kwargs": {"reasoning_effort": "high"},
             "skip_special_tokens": False,
         }
 
-    def test_extra_body_chat_template_kwargs_deep_merged(self) -> None:
-        """chat_template_kwargs in extra_body is deep-merged, operator wins."""
+    def test_extra_body_chat_template_kwargs_deep_merged_with_base(self) -> None:
+        """Operator chat_template_kwargs deep-merges over the seeded base."""
         base = {"reasoning_effort": "medium"}
         compat = {
             "extra_body": {
@@ -144,17 +180,15 @@ class TestMergeServerCompat:
             },
         }
         result = merge_server_compat(base, compat)
-        # Operator values win over base
         assert result["chat_template_kwargs"]["custom_flag"] is True
+        # Operator value wins over seeded base
         assert result["chat_template_kwargs"]["reasoning_effort"] == "high"
         assert result["skip_special_tokens"] is False
 
     def test_extra_body_chat_template_kwargs_non_dict_ignored(self) -> None:
         """Non-dict chat_template_kwargs in extra_body is safely ignored."""
-        base = {"reasoning_effort": "medium"}
         compat = {"extra_body": {"chat_template_kwargs": "bad"}}
-        result = merge_server_compat(base, compat)
-        assert result["chat_template_kwargs"] == {"reasoning_effort": "medium"}
+        assert merge_server_compat(None, compat) == {}
 
     def test_base_not_mutated(self) -> None:
         base = {"reasoning_effort": "medium"}
@@ -164,9 +198,7 @@ class TestMergeServerCompat:
 
     def test_non_dict_extra_body_ignored(self) -> None:
         """Gracefully handle malformed server_compat."""
-        base = {"reasoning_effort": "medium"}
-        result = merge_server_compat(base, {"extra_body": 42})
-        assert result == {"chat_template_kwargs": {"reasoning_effort": "medium"}}
+        assert merge_server_compat(None, {"extra_body": 42}) == {}
 
 
 # ---------------------------------------------------------------------------
@@ -178,45 +210,58 @@ class TestEndToEndRequestShaping:
     """Compose both layers — session builds extra_params, provider applies thinking."""
 
     def test_vllm_gemma_full_flow(self) -> None:
-        """Session merges server workarounds, provider adds thinking param."""
+        """Session forwards server workarounds, provider adds thinking param."""
         caps = ModelCapabilities(thinking_mode="manual", thinking_param="enable_thinking")
-        base_ctk = {"reasoning_effort": "medium"}
         server_compat = {
             "server_type": "vllm",
             "extra_body": {"skip_special_tokens": False},
         }
-        # Step 1: session merges
-        extra_params = merge_server_compat(base_ctk, server_compat)
-        # Step 2: provider finalises
+        # Step 1: session forwards (no auto-injection of reasoning_effort).
+        extra_params = merge_server_compat(None, server_compat)
+        # Step 2: provider injects thinking param into chat_template_kwargs.
         extra_body = dict(extra_params)
         OpenAIChatCompletionsProvider._apply_thinking_mode(extra_body, caps)
 
         assert extra_body == {
-            "chat_template_kwargs": {
-                "reasoning_effort": "medium",
-                "enable_thinking": True,
-            },
+            "chat_template_kwargs": {"enable_thinking": True},
             "skip_special_tokens": False,
         }
 
     def test_granite_thinking_key(self) -> None:
         """Granite uses 'thinking' instead of 'enable_thinking'."""
         caps = ModelCapabilities(thinking_mode="manual", thinking_param="thinking")
-        extra_params = merge_server_compat({"reasoning_effort": "low"}, {})
+        extra_params = merge_server_compat(None, {})
         extra_body = dict(extra_params)
         OpenAIChatCompletionsProvider._apply_thinking_mode(extra_body, caps)
 
-        assert extra_body["chat_template_kwargs"]["thinking"] is True
-        assert "enable_thinking" not in extra_body["chat_template_kwargs"]
+        assert extra_body == {"chat_template_kwargs": {"thinking": True}}
 
     def test_non_thinking_model_no_injection(self) -> None:
-        """Non-thinking model gets no thinking params."""
+        """Non-thinking model gets no chat_template_kwargs at all."""
         caps = ModelCapabilities()  # thinking_mode="none"
-        extra_params = merge_server_compat({"reasoning_effort": "medium"}, {})
+        extra_params = merge_server_compat(None, {})
         extra_body = dict(extra_params)
         OpenAIChatCompletionsProvider._apply_thinking_mode(extra_body, caps)
 
-        assert extra_body == {"chat_template_kwargs": {"reasoning_effort": "medium"}}
+        assert extra_body == {}
+
+    def test_operator_reasoning_effort_passthrough(self) -> None:
+        """Operator-supplied reasoning_effort under chat_template_kwargs is preserved."""
+        caps = ModelCapabilities(thinking_mode="manual", thinking_param="enable_thinking")
+        compat = {
+            "server_type": "vllm",
+            "extra_body": {"chat_template_kwargs": {"reasoning_effort": "high"}},
+        }
+        extra_params = merge_server_compat(None, compat)
+        extra_body = dict(extra_params)
+        OpenAIChatCompletionsProvider._apply_thinking_mode(extra_body, caps)
+
+        assert extra_body == {
+            "chat_template_kwargs": {
+                "reasoning_effort": "high",
+                "enable_thinking": True,
+            },
+        }
 
 
 # ---------------------------------------------------------------------------

tests/test_server_node_models_metadata.py

@@ -0,0 +1,346 @@
+"""Tests for the per-node ``models`` metadata pipeline.
+
+Two helpers in ``server.py`` carry the load:
+
+- ``_collect_node_models_metadata`` projects the live ``ModelRegistry``
+  into the node_metadata row shape ``[{alias, provider, healthy}, ...]``.
+- ``_publish_models_metadata`` short-circuits redundant writes via a
+  payload cache on ``app_state`` and is the helper called from both
+  the heartbeat loop and ``internal_model_reload``.
+
+These tests pin the projection shape, the health-flag wiring, the
+cache short-circuit, and the model-reload integration.
+"""
+
+from __future__ import annotations
+
+import json
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+from turnstone.core.healthcheck import HealthTrackerRegistry
+from turnstone.core.model_registry import ModelConfig, ModelRegistry
+from turnstone.server import (
+    _collect_node_models_metadata,
+    _publish_models_metadata,
+)
+
+
+def _registry(*aliases_with_url: tuple[str, str]) -> ModelRegistry:
+    """Build a registry from ``(alias, base_url)`` pairs.
+
+    Two aliases sharing a ``base_url`` deliberately share a tracker —
+    that's the contract the cluster-level health surface needs to
+    preserve, and it's worth pinning in a test.
+    """
+    models = {
+        alias: ModelConfig(alias=alias, base_url=url, api_key="k", model=alias, provider="openai")
+        for alias, url in aliases_with_url
+    }
+    default = aliases_with_url[0][0]
+    return ModelRegistry(models, default=default)
+
+
+def test_returns_none_when_registry_missing():
+    state = SimpleNamespace()
+    assert _collect_node_models_metadata(state) is None
+
+
+def test_projects_all_aliases_with_default_healthy_when_no_tracker():
+    """Without a ``health_registry`` (or before any request has flowed
+    through a backend), every alias surfaces as ``healthy=True`` —
+    operators shouldn't get an empty ``models`` list on a freshly
+    started node just because the backends haven't been exercised."""
+    reg = _registry(("a", "http://x"), ("b", "http://y"))
+    state = SimpleNamespace(registry=reg)
+    entry = _collect_node_models_metadata(state)
+    assert entry is not None
+    key, value, source = entry
+    assert key == "models"
+    assert source == "auto"
+    rows = json.loads(value)
+    assert len(rows) == 2
+    aliases = {r["alias"] for r in rows}
+    assert aliases == {"a", "b"}
+    assert all(r["healthy"] is True for r in rows)
+    assert all(r["provider"] == "openai" for r in rows)
+    # Provider-side model identifier intentionally omitted — coords
+    # kept passing it as ``spawn_workstream(model=...)`` when they
+    # should have passed the local alias.  Lock the projected keys
+    # so a future contributor doesn't reintroduce the footgun.
+    for row in rows:
+        assert set(row.keys()) == {"alias", "provider", "healthy"}
+
+
+def test_health_flag_reflects_tracker_state():
+    reg = _registry(("a", "http://x"), ("b", "http://y"))
+    health_reg = HealthTrackerRegistry(failure_threshold=2)
+    # Seed the tracker for "a"'s backend and drive it into the degraded
+    # state — two consecutive failures cross the threshold.
+    bad_tracker = health_reg.get_tracker(provider="openai", base_url="http://x")
+    bad_tracker.record_failure()
+    bad_tracker.record_failure()
+    assert bad_tracker.is_degraded
+    # "b" gets a tracker that has only seen successes.
+    good_tracker = health_reg.get_tracker(provider="openai", base_url="http://y")
+    good_tracker.record_success()
+    state = SimpleNamespace(registry=reg, health_registry=health_reg)
+    rows = json.loads(_collect_node_models_metadata(state)[1])
+    by_alias = {r["alias"]: r for r in rows}
+    assert by_alias["a"]["healthy"] is False
+    assert by_alias["b"]["healthy"] is True
+
+
+def test_two_aliases_sharing_a_backend_share_a_tracker():
+    """Two aliases that point at the same ``(provider, base_url)``
+    share a single :class:`BackendHealthTracker` — degrading one is
+    expected to surface as degraded on the other.  The list_nodes
+    projection should respect that, otherwise a coord could see
+    ``alias-a`` healthy and ``alias-b`` degraded for the same
+    backend."""
+    reg = _registry(("alpha", "http://shared"), ("beta", "http://shared"))
+    health_reg = HealthTrackerRegistry(failure_threshold=1)
+    tracker = health_reg.get_tracker(provider="openai", base_url="http://shared")
+    tracker.record_failure()  # threshold=1 — degraded immediately
+    state = SimpleNamespace(registry=reg, health_registry=health_reg)
+    rows = json.loads(_collect_node_models_metadata(state)[1])
+    assert {r["alias"]: r["healthy"] for r in rows} == {"alpha": False, "beta": False}
+
+
+def test_alias_with_no_tracker_yet_defaults_to_healthy():
+    """An alias the registry knows about but whose backend hasn't been
+    invoked yet has no tracker.  Default to healthy so a brand-new
+    alias is immediately visible to coordinators rather than waiting
+    for the first request to seed a tracker.
+
+    The collector calls ``health_reg.get_tracker(...)`` which mints a
+    fresh tracker on first lookup — that's the path under test here.
+    The freshly minted tracker reports ``is_healthy=True`` (default
+    state), so the projection labels the alias healthy.
+    """
+    reg = _registry(("a", "http://x"))
+    health_reg = HealthTrackerRegistry()  # empty — no trackers seeded
+    state = SimpleNamespace(registry=reg, health_registry=health_reg)
+    rows = json.loads(_collect_node_models_metadata(state)[1])
+    assert rows[0]["healthy"] is True
+
+
+# ---------------------------------------------------------------------------
+# _publish_models_metadata — cache short-circuit + projection wiring
+# ---------------------------------------------------------------------------
+
+
+def _publish_state() -> SimpleNamespace:
+    """Build an ``app_state`` with a minimal registry + health surface."""
+    reg = _registry(("a", "http://x"))
+    return SimpleNamespace(registry=reg, health_registry=HealthTrackerRegistry())
+
+
+def test_publish_writes_when_payload_changes():
+    """First publish has nothing in the cache — write happens; cache
+    fills.  Second publish on the same unchanged registry skips the
+    write entirely."""
+    state = _publish_state()
+    storage = MagicMock()
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 1
+    cached = state._last_models_payload
+    assert isinstance(cached, str) and "alias" in cached
+    # Second call, same registry, same health: cached payload matches
+    # — write must be skipped to avoid the per-30s UPSERT churn.
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 1
+
+
+def test_publish_records_metric_outcome(monkeypatch):
+    """The publish helper feeds ``record_node_models_publish`` so
+    Prometheus can expose the hit-rate.  Storage failures must NOT
+    record either outcome — counters should reflect actual cache
+    decisions, not transient DB errors that will retry.
+
+    Replaces the module-level ``turnstone.server._metrics`` binding
+    via string-form monkeypatch (with auto-restore) rather than
+    patching an instance attribute on the imported singleton.  Other
+    tests in the suite reassign ``srv_mod._metrics`` (some without
+    using monkeypatch), so an instance captured at import time can
+    diverge from the binding the live ``_publish_models_metadata``
+    reads on each call.
+    """
+    state = _publish_state()
+    storage = MagicMock()
+    calls: list[bool] = []
+
+    class _FakeMetrics:
+        def record_node_models_publish(self, *, written: bool) -> None:
+            calls.append(written)
+
+    monkeypatch.setattr("turnstone.server._metrics", _FakeMetrics())
+
+    _publish_models_metadata(state, storage, "node-a")  # first → write
+    _publish_models_metadata(state, storage, "node-a")  # second → skip
+    assert calls == [True, False]
+
+    # Storage error: no metric recorded.
+    storage.set_node_metadata_bulk.side_effect = RuntimeError("db down")
+    state._last_models_payload = None  # invalidate cache to force a write attempt
+    _publish_models_metadata(state, storage, "node-a")
+    assert calls == [True, False]  # unchanged
+
+
+def test_publish_rewrites_when_health_flips():
+    """A health-tracker state change must invalidate the cache and
+    drive a fresh write — otherwise the discovery surface would lag
+    a flip indefinitely."""
+    state = _publish_state()
+    storage = MagicMock()
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 1
+    # Drive the only tracker to degraded.
+    tracker = state.health_registry.get_tracker(provider="openai", base_url="http://x")
+    for _ in range(10):
+        tracker.record_failure()
+    assert tracker.is_degraded
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 2
+
+
+def test_publish_swallows_storage_error_without_updating_cache():
+    """A storage failure must NOT poison the cache — the next call
+    should retry the write rather than think it succeeded."""
+    state = _publish_state()
+    storage = MagicMock()
+    storage.set_node_metadata_bulk.side_effect = RuntimeError("db down")
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 1
+    assert getattr(state, "_last_models_payload", None) is None
+    # Recover: a subsequent successful call writes again.
+    storage.set_node_metadata_bulk.side_effect = None
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 2
+    assert state._last_models_payload is not None
+
+
+def test_publish_skips_when_registry_missing():
+    """Without a registry there's nothing to project; nothing should
+    be written and the cache must not be set."""
+    state = SimpleNamespace()
+    storage = MagicMock()
+    _publish_models_metadata(state, storage, "node-a")
+    assert storage.set_node_metadata_bulk.call_count == 0
+    assert getattr(state, "_last_models_payload", None) is None
+
+
+# ---------------------------------------------------------------------------
+# internal_model_reload — integration: registry change must rewrite the row
+# ---------------------------------------------------------------------------
+
+
+def test_model_reload_endpoint_rewrites_models_metadata(monkeypatch, tmp_path):
+    """A successful ``internal_model_reload`` must refresh
+    ``node_metadata.models`` so a coordinator sees the new alias on
+    its next ``list_nodes`` without waiting up to 30s for the
+    heartbeat tick.
+
+    The endpoint pulls a fresh registry from
+    ``load_model_registry(...)`` and reloads in-place — we stub the
+    loader to return a registry with a different alias set so the
+    publish-cache invalidation is exercised end-to-end.
+    """
+    from turnstone.core.storage._sqlite import SQLiteBackend
+    from turnstone.server import internal_model_reload
+
+    storage = SQLiteBackend(str(tmp_path / "reload.db"))
+
+    # Old registry — single alias "a".
+    old_reg = _registry(("a", "http://x"))
+    # New registry that ``load_model_registry`` will return — adds "b".
+    new_reg = ModelRegistry(
+        {
+            "a": ModelConfig(
+                alias="a", base_url="http://x", api_key="k", model="a", provider="openai"
+            ),
+            "b": ModelConfig(
+                alias="b", base_url="http://y", api_key="k", model="b", provider="openai"
+            ),
+        },
+        default="a",
+    )
+    health_reg = HealthTrackerRegistry()
+
+    app_state = SimpleNamespace(
+        registry=old_reg,
+        health_registry=health_reg,
+        cli_model_args={
+            "base_url": "",
+            "api_key": "",
+            "model": "",
+            "context_window": 0,
+            "provider": "openai",
+        },
+        config_store=None,
+        node_id="node-a",
+    )
+    request = SimpleNamespace(app=SimpleNamespace(state=app_state))
+
+    # Patch the loader and storage accessors used inside the endpoint.
+    # ``internal_model_reload`` does ``from turnstone.core.storage._registry
+    # import get_storage`` inline, so patching the symbol on that module
+    # is what intercepts the call.
+    monkeypatch.setattr("turnstone.core.model_registry.load_model_registry", lambda **_kw: new_reg)
+    monkeypatch.setattr("turnstone.core.storage._registry.get_storage", lambda: storage)
+    # The endpoint also broadcasts schema refreshes to active sessions
+    # — stub this out, it's irrelevant to the metadata-write path.
+    monkeypatch.setattr("turnstone.server._broadcast_agent_tool_schema_refresh", lambda _s: None)
+
+    response = internal_model_reload(request)  # type: ignore[arg-type]
+    assert response.status_code == 200
+
+    rows = storage.get_node_metadata("node-a")
+    by_key = {r["key"]: r for r in rows}
+    assert "models" in by_key
+    payload = json.loads(by_key["models"]["value"])
+    assert {r["alias"] for r in payload} == {"a", "b"}
+
+
+# ---------------------------------------------------------------------------
+# Shutdown race: heartbeat write must NOT resurrect post-shutdown delete
+# ---------------------------------------------------------------------------
+
+
+def test_heartbeat_write_awaits_before_shutdown_delete():
+    """Pin the shutdown-race fix.
+
+    Before the fix, the lifespan shutdown sequence was:
+
+    1. ``_heartbeat_task.cancel()`` — fire-and-forget
+    2. ``delete_node_metadata_by_source(node_id, "auto")``
+
+    A heartbeat tick already inside ``asyncio.to_thread(...)`` for
+    the ``set_node_metadata_bulk`` call would complete AFTER step 2,
+    resurrecting the deleted ``models`` row.  The fix awaits the
+    cancelled task with ``contextlib.suppress(...)`` between (1) and
+    (2), so the in-flight write lands first.
+
+    We verify the fix by introspecting ``server.py`` source — the
+    real lifespan is hard to test deterministically without a full
+    Starlette app, but the textual ordering between
+    ``_heartbeat_task.cancel()`` and the delete is a stable contract
+    that catches the regression cheaply.
+    """
+    import inspect
+    import sys
+
+    src = inspect.getsource(sys.modules[_collect_node_models_metadata.__module__])
+    cancel_idx = src.find("_heartbeat_task.cancel()")
+    delete_idx = src.find('delete_node_metadata_by_source, _svc_node_id, "auto"')
+    await_idx = src.find("await _heartbeat_task", cancel_idx)
+    assert cancel_idx != -1
+    assert delete_idx != -1
+    assert await_idx != -1
+    # The fix-line must sit BETWEEN the cancel and the delete.
+    assert cancel_idx < await_idx < delete_idx, (
+        "Shutdown race regression: "
+        "_heartbeat_task.cancel() must be followed by `await _heartbeat_task` "
+        "BEFORE delete_node_metadata_by_source(..., 'auto') so an in-flight "
+        "set_node_metadata_bulk lands before the delete."
+    )

tests/test_session.py

@@ -1182,7 +1182,7 @@ class TestAgentOutputGuard:
 
 
 class TestProviderExtraParams:
-    """Tests for _provider_extra_params — local-only chat_template_kwargs."""
+    """Tests for _provider_extra_params — server_compat passthrough only."""
 
     def _session_with_provider(self, provider_name: str, tmp_db) -> ChatSession:
         from turnstone.core.providers import create_provider
@@ -1191,68 +1191,36 @@ class TestProviderExtraParams:
         session._provider = create_provider(provider_name)
         return session
 
-    def test_openai_compatible_returns_chat_template_kwargs(self, tmp_db):
+    def test_openai_compatible_no_compat_returns_none(self, tmp_db):
+        """No server_compat → no extra_body needed (no auto-injection)."""
         session = self._session_with_provider("openai-compatible", tmp_db)
-        result = session._provider_extra_params()
-        assert result is not None
-        assert "chat_template_kwargs" in result
-        assert result["chat_template_kwargs"]["reasoning_effort"] == "medium"
+        assert session._provider_extra_params() is None
 
-    def test_openai_commercial_returns_none(self, tmp_db):
+    def test_openai_commercial_no_compat_returns_none(self, tmp_db):
+        """Cloud OpenAI without server_compat → None."""
         session = self._session_with_provider("openai", tmp_db)
-        result = session._provider_extra_params()
-        assert result is None
+        assert session._provider_extra_params() is None
 
     def test_anthropic_returns_none(self, tmp_db):
         session = self._session_with_provider("anthropic", tmp_db)
-        result = session._provider_extra_params()
-        assert result is None
+        assert session._provider_extra_params() is None
 
-    def test_reasoning_effort_override(self, tmp_db):
+    def test_no_reasoning_effort_kwarg(self, tmp_db):
+        """reasoning_effort is not part of the surface; passing it should TypeError.
+
+        Splatted via ``**kwargs`` so static analyzers (CodeQL "wrong-name
+        argument" / mypy) don't flag the call — the point of this test is the
+        runtime contract, not the static type.
+        """
+        import pytest
+
+        bad_kwargs = {"reasoning_effort": "high"}
         session = self._session_with_provider("openai-compatible", tmp_db)
-        result = session._provider_extra_params(reasoning_effort="high")
-        assert result is not None
-        assert result["chat_template_kwargs"]["reasoning_effort"] == "high"
+        with pytest.raises(TypeError):
+            session._provider_extra_params(**bad_kwargs)
 
-    def test_explicit_openai_provider_overrides_session(self, tmp_db):
-        """Passing an explicit commercial OpenAI provider returns None even
-        when the session's own provider is openai-compatible."""
-        from turnstone.core.providers import create_provider
-
-        session = self._session_with_provider("openai-compatible", tmp_db)
-        openai_prov = create_provider("openai")
-        result = session._provider_extra_params(provider=openai_prov)
-        assert result is None
-
-    def test_server_compat_extra_body_merged(self, tmp_db):
-        """server_compat.extra_body workarounds are merged into extra_params."""
-        from turnstone.core.model_registry import ModelConfig, ModelRegistry
-
-        session = self._session_with_provider("openai-compatible", tmp_db)
-        cfg = ModelConfig(
-            alias="test",
-            base_url="http://localhost:8000/v1",
-            api_key="none",
-            model="google/gemma-4-31B-it",
-            server_compat={
-                "extra_body": {"skip_special_tokens": False},
-            },
-        )
-        session._registry = ModelRegistry(models={"test": cfg}, default="test")
-        session._model_alias = "test"
-        result = session._provider_extra_params()
-        assert result is not None
-        assert result["chat_template_kwargs"]["reasoning_effort"] == "medium"
-        assert result["skip_special_tokens"] is False
-
-    def test_empty_server_compat_backwards_compatible(self, tmp_db):
-        """Empty server_compat produces same output as before."""
-        session = self._session_with_provider("openai-compatible", tmp_db)
-        result = session._provider_extra_params()
-        assert result == {"chat_template_kwargs": {"reasoning_effort": "medium"}}
-
-    def test_server_compat_with_reasoning_effort_override(self, tmp_db):
-        """reasoning_effort override works alongside server_compat."""
+    def test_server_compat_extra_body_passes_through(self, tmp_db):
+        """server_compat.extra_body workarounds forward as extra_params."""
         from turnstone.core.model_registry import ModelConfig, ModelRegistry
 
         session = self._session_with_provider("openai-compatible", tmp_db)
@@ -1265,10 +1233,25 @@ class TestProviderExtraParams:
         )
         session._registry = ModelRegistry(models={"test": cfg}, default="test")
         session._model_alias = "test"
-        result = session._provider_extra_params(reasoning_effort="high")
-        assert result is not None
-        assert result["chat_template_kwargs"]["reasoning_effort"] == "high"
-        assert result["skip_special_tokens"] is False
+        result = session._provider_extra_params()
+        assert result == {"skip_special_tokens": False}
+
+    def test_operator_chat_template_kwargs_pass_through(self, tmp_db):
+        """Operator-set chat_template_kwargs (e.g. for gpt-oss) forwards verbatim."""
+        from turnstone.core.model_registry import ModelConfig, ModelRegistry
+
+        session = self._session_with_provider("openai-compatible", tmp_db)
+        cfg = ModelConfig(
+            alias="test",
+            base_url="http://localhost:8000/v1",
+            api_key="none",
+            model="openai/gpt-oss-120b",
+            server_compat={"extra_body": {"chat_template_kwargs": {"reasoning_effort": "high"}}},
+        )
+        session._registry = ModelRegistry(models={"test": cfg}, default="test")
+        session._model_alias = "test"
+        result = session._provider_extra_params()
+        assert result == {"chat_template_kwargs": {"reasoning_effort": "high"}}
 
     def test_model_alias_resolves_target_compat(self, tmp_db):
         """model_alias parameter selects compat from the target, not the primary."""
@@ -1297,14 +1280,9 @@ class TestProviderExtraParams:
         session._model_alias = "primary"
 
         # Primary alias → gets Gemma workaround
-        result_primary = session._provider_extra_params()
-        assert result_primary is not None
-        assert result_primary["skip_special_tokens"] is False
-
-        # Fallback alias → no compat, just base kwargs
-        result_fallback = session._provider_extra_params(model_alias="fallback")
-        assert result_fallback == {"chat_template_kwargs": {"reasoning_effort": "medium"}}
-        assert "skip_special_tokens" not in result_fallback
+        assert session._provider_extra_params() == {"skip_special_tokens": False}
+        # Fallback alias → no compat at all
+        assert session._provider_extra_params(model_alias="fallback") is None
 
 
 class TestSafePrepareTool:

tests/test_session_manager.py

@@ -101,6 +101,7 @@ class FakeAdapter:
         self.cleaned_up: list[str] = []
         self.build_session_calls = 0
         self.build_session_raises = build_session_raises
+        self.last_build_model: object | None = None
         # Slow down session build so concurrent tests can race.
         self.build_session_delay = 0.0
 
@@ -144,8 +145,13 @@ class FakeAdapter:
     def build_ui(self, ws: Workstream) -> Any:
         return FakeUI()
 
-    def build_session(self, ws: Workstream, **_: object) -> Any:
+    def build_session(self, ws: Workstream, **kwargs: object) -> Any:
         self.build_session_calls += 1
+        # Record the ``model`` kwarg (None on fresh-create, the saved
+        # alias on rehydrate) so tests can assert SessionManager.open()
+        # threads the persisted alias through to construction instead
+        # of letting the adapter resolve the *current* default alias.
+        self.last_build_model = kwargs.get("model")
         if self.build_session_delay:
             time.sleep(self.build_session_delay)
         if self.build_session_raises:
@@ -184,6 +190,13 @@ class FakeStorage:
         # "no peers alive" (every row unprotected by liveness).
         self.live_services: dict[str, list[str]] = {}
         self.list_services_raises = False
+        # Per-ws config (model_alias, temperature, …).  Populated by
+        # tests that exercise the rehydrate-preserves-config path; the
+        # SessionManager.open() rehydrate path reads this through
+        # ``self._storage.load_workstream_config`` so it can pass the
+        # saved alias into ``build_session`` and avoid clobbering the
+        # original on construction.
+        self.ws_config: dict[str, dict[str, str]] = {}
 
     @staticmethod
     def _now_iso() -> str:
@@ -294,6 +307,18 @@ class FakeStorage:
     def count_skill_versions(self, template_id: str) -> int:
         return 0
 
+    def load_workstream_config(self, ws_id: str) -> dict[str, str]:
+        with self.lock:
+            return dict(self.ws_config.get(ws_id, {}))
+
+    def save_workstream_config(self, ws_id: str, config: dict[str, str]) -> None:
+        # Mirrors the real backend's INSERT OR REPLACE per-key semantics
+        # — callers expect a partial save to overwrite only the keys
+        # they pass, not the whole row.
+        with self.lock:
+            row = self.ws_config.setdefault(ws_id, {})
+            row.update(config)
+
 
 _EMITTER_DEFAULT = object()
 
@@ -305,6 +330,7 @@ def _make_manager(
     storage: FakeStorage | None = None,
     event_emitter: Any = _EMITTER_DEFAULT,
     node_id: str | None = None,
+    model_validator: Callable[[str], bool] | None = None,
 ) -> tuple[SessionManager, FakeAdapter, FakeStorage]:
     """Build a SessionManager wired to a FakeAdapter for both Protocols.
 
@@ -324,6 +350,7 @@ def _make_manager(
         max_active=max_active,
         event_emitter=emitter,
         node_id=node_id,
+        model_validator=model_validator,
     )
     return mgr, adapter, storage
 
@@ -657,6 +684,102 @@ def test_open_resurrects_closed_state() -> None:
     assert ws_id in [e.ws_id for e in adapter.events_of("rehydrated")]
 
 
+def test_open_threads_saved_model_alias_into_build_session() -> None:
+    """Reopening a closed ws must build the session with the *original*
+    model alias, not the current registry default.
+
+    Without this, ``build_session(ws)`` is called with ``model=None`` →
+    the production session_factory resolves ``_effective_default_alias()``
+    → ChatSession's ``__init__`` writes those defaults to
+    ``workstream_config`` (INSERT OR REPLACE) → the subsequent
+    ``resume()`` restores what is now the default. Net effect: every
+    persisted knob (model, temperature, reasoning_effort, max_tokens,
+    skill, creative_mode, instructions, …) silently resets on every
+    reopen and on every service restart.
+    """
+    mgr, adapter, storage = _make_manager()
+    ws = mgr.create(user_id="u1")
+    ws_id = ws.id
+    # Pretend the user set a non-default alias when the ws was created;
+    # the real path goes through ChatSession._save_config but the
+    # FakeSession in this suite doesn't model that, so seed directly.
+    storage.ws_config[ws_id] = {"model_alias": "gpt-5-pro"}
+    mgr.close(ws_id)
+    adapter.last_build_model = "<unset>"  # sentinel — must be overwritten
+
+    reopened = mgr.open(ws_id)
+
+    assert reopened is not None
+    assert adapter.last_build_model == "gpt-5-pro"
+
+
+def test_open_drops_saved_alias_when_validator_rejects() -> None:
+    """When the persisted alias is no longer in the registry, the
+    manager must drop it before reaching ``build_session``. The
+    factory still raises on unknown aliases on the fresh-create path
+    (so a typo in body.model surfaces as 503), so the rehydrate path
+    has to filter the alias here rather than relying on factory-side
+    fallback. Without this filter, every reopen of a workstream pinned
+    to a since-removed alias 500s."""
+    mgr, adapter, storage = _make_manager(
+        # Validator says "alias is no longer in the registry".
+        model_validator=lambda alias: False,
+    )
+    ws = mgr.create(user_id="u1")
+    ws_id = ws.id
+    storage.ws_config[ws_id] = {"model_alias": "since-removed-alias"}
+    mgr.close(ws_id)
+    adapter.last_build_model = "<unset>"
+
+    reopened = mgr.open(ws_id)
+
+    assert reopened is not None
+    assert adapter.last_build_model is None  # alias dropped before reaching build_session
+
+
+def test_open_keeps_saved_alias_when_validator_accepts() -> None:
+    """Sanity: an alias that still resolves must be passed through
+    unchanged. Filter only fires for stale aliases."""
+    accepted: list[str] = []
+
+    def validator(alias: str) -> bool:
+        accepted.append(alias)
+        return True
+
+    mgr, adapter, storage = _make_manager(model_validator=validator)
+    ws = mgr.create(user_id="u1")
+    ws_id = ws.id
+    storage.ws_config[ws_id] = {"model_alias": "still-live"}
+    mgr.close(ws_id)
+    adapter.last_build_model = "<unset>"
+
+    reopened = mgr.open(ws_id)
+
+    assert reopened is not None
+    assert accepted == ["still-live"]
+    assert adapter.last_build_model == "still-live"
+
+
+def test_open_falls_back_to_none_when_no_saved_alias() -> None:
+    """Reopening a ws with no saved alias must pass ``model=None`` to
+    ``build_session`` so the adapter's session_factory can fall back to
+    the current default — matching the user's intent: best effort
+    restore, default when the original is gone."""
+    mgr, adapter, storage = _make_manager()
+    ws = mgr.create(user_id="u1")
+    ws_id = ws.id
+    # No ws_config row — simulates "alias was never saved" or "saved
+    # alias was empty string".
+    assert ws_id not in storage.ws_config
+    mgr.close(ws_id)
+    adapter.last_build_model = "<unset>"
+
+    reopened = mgr.open(ws_id)
+
+    assert reopened is not None
+    assert adapter.last_build_model is None
+
+
 def test_open_touches_workstream_on_rehydrate() -> None:
     """Rehydrating a workstream must bump its ``updated`` so a concurrent
     close_idle pass-2 in this same process can't clobber the freshly-loaded

tests/test_sessions.py

@@ -523,8 +523,15 @@ class TestWorkstreamConfig:
         assert session.instructions == "be concise"
         assert session.creative_mode is True
 
-    def test_resume_restores_model(self, tmp_db):
-        """ChatSession.resume() should restore the model from workstream config."""
+    def test_resume_keeps_defaults_when_alias_unresolvable(self, tmp_db):
+        """When the saved alias is empty or no longer in the registry,
+        ``resume()`` must NOT copy ``saved_model`` onto the constructor's
+        default provider.  Pairing a removed model name with a default
+        provider that doesn't know about it produces a broken session
+        whose next API call fails — the exact regression Copilot flagged
+        on PR #465.  The constructor already resolved a coherent default
+        (provider + model + capabilities); resume should leave it intact
+        and just log the unreachable saved values."""
         client = MagicMock()
         client.models.list.return_value.data = [MagicMock(id="test-model")]
         ui = MagicMock()
@@ -533,13 +540,14 @@ class TestWorkstreamConfig:
         ui.on_state_change = MagicMock()
         ui.on_rename = MagicMock()
 
-        # Create a workstream that was using a specific model
         register_workstream("model_ws")
         save_message("model_ws", "user", "hello")
         save_message("model_ws", "assistant", "hi")
+        # Empty alias + an orphan model name — same shape resume sees
+        # when an operator removes an alias from the registry that the
+        # workstream was originally pinned to.
         save_workstream_config("model_ws", {"model": "gpt-5", "model_alias": ""})
 
-        # Resume into a session that was created with a different model
         session = ChatSession(
             client=client,
             model="gpt-5-nano",
@@ -552,7 +560,102 @@ class TestWorkstreamConfig:
         assert session.model == "gpt-5-nano"
         result = session.resume("model_ws")
         assert result is True
-        assert session.model == "gpt-5"
+        # Constructor's coherent default is preserved — saved orphan
+        # model name is NOT copied over.
+        assert session.model == "gpt-5-nano"
+
+    def test_init_does_not_clobber_existing_config(self, tmp_db):
+        """ChatSession.__init__ must NOT overwrite existing
+        ``workstream_config`` keys when constructing for an already-
+        persisted ws_id.
+
+        This is the fix for the rehydrate bug: ``SessionManager.open()``
+        builds a ChatSession with the persisted ws_id; the legacy
+        ``__init__`` unconditionally called ``_save_config()`` which is
+        ``INSERT OR REPLACE`` per-key — silently resetting model_alias,
+        temperature, reasoning_effort, max_tokens, skill, creative_mode,
+        and instructions to the constructor defaults *before*
+        ``resume()`` got a chance to read them back.
+        """
+        client = MagicMock()
+        client.models.list.return_value.data = [MagicMock(id="test-model")]
+        ui = MagicMock()
+        ui.on_info = MagicMock()
+        ui.on_error = MagicMock()
+        ui.on_state_change = MagicMock()
+        ui.on_rename = MagicMock()
+
+        register_workstream("rehydrate_ws")
+        save_workstream_config(
+            "rehydrate_ws",
+            {
+                "model": "gpt-5-pro",
+                "model_alias": "gpt-5-pro",
+                "temperature": "0.2",
+                "reasoning_effort": "high",
+                "max_tokens": "8192",
+                "creative_mode": "True",
+                "instructions": "preserve me",
+            },
+        )
+
+        ChatSession(
+            client=client,
+            model="some-default-model",
+            ui=ui,
+            instructions=None,
+            temperature=0.7,
+            max_tokens=4096,
+            tool_timeout=30,
+            reasoning_effort="medium",
+            ws_id="rehydrate_ws",
+        )
+
+        loaded = load_workstream_config("rehydrate_ws")
+        assert loaded["model"] == "gpt-5-pro"
+        assert loaded["model_alias"] == "gpt-5-pro"
+        assert loaded["temperature"] == "0.2"
+        assert loaded["reasoning_effort"] == "high"
+        assert loaded["max_tokens"] == "8192"
+        assert loaded["creative_mode"] == "True"
+        assert loaded["instructions"] == "preserve me"
+
+    def test_init_writes_config_on_fresh_create(self, tmp_db):
+        """The opposite half of the contract: when no config row exists
+        yet, ``__init__`` must still persist the constructor's values so
+        a later resume can find them. This is the path that previously
+        worked — the fix must not break it."""
+        client = MagicMock()
+        client.models.list.return_value.data = [MagicMock(id="test-model")]
+        ui = MagicMock()
+        ui.on_info = MagicMock()
+        ui.on_error = MagicMock()
+        ui.on_state_change = MagicMock()
+        ui.on_rename = MagicMock()
+
+        # No save_workstream_config() before ChatSession() — this is
+        # the fresh-create path the SessionManager.create() flow takes.
+        register_workstream("fresh_ws")
+        assert load_workstream_config("fresh_ws") == {}
+
+        ChatSession(
+            client=client,
+            model="gpt-5-mini",
+            ui=ui,
+            instructions="be terse",
+            temperature=0.4,
+            max_tokens=2048,
+            tool_timeout=30,
+            reasoning_effort="low",
+            ws_id="fresh_ws",
+        )
+
+        loaded = load_workstream_config("fresh_ws")
+        assert loaded["model"] == "gpt-5-mini"
+        assert loaded["temperature"] == "0.4"
+        assert loaded["reasoning_effort"] == "low"
+        assert loaded["max_tokens"] == "2048"
+        assert loaded["instructions"] == "be terse"
 
 
 # ── Prune workstreams ─────────────────────────────────────────────────

tests/test_structured_memory.py

@@ -67,6 +67,42 @@ class TestSearchStructuredMemories:
         assert len(results) >= 1
         assert any(r["name"] == "db_host" for r in results)
 
+    def test_multiword_or_matches_partial(self, tmp_db):
+        """OR-of-terms: memory matching only 1 of 3 query terms is returned."""
+        save_structured_memory("postgres_config", "host=localhost port=5432")
+        save_structured_memory("redis_config", "host=redis port=6379")
+        save_structured_memory("unrelated", "nothing relevant here")
+
+        # "postgres missing_word_a missing_word_b": only postgres_config matches "postgres"
+        results = search_structured_memories("postgres missing_word_a missing_word_b")
+        names = {r["name"] for r in results}
+        assert "postgres_config" in names
+        assert "unrelated" not in names
+
+    def test_multiword_or_multiple_partial_matches(self, tmp_db):
+        """Multiple memories each matching different terms are all returned."""
+        save_structured_memory("key_alpha", "alpha content here")
+        save_structured_memory("key_beta", "beta content here")
+        save_structured_memory("key_other", "completely different")
+
+        results = search_structured_memories("alpha beta")
+        names = {r["name"] for r in results}
+        assert "key_alpha" in names
+        assert "key_beta" in names
+        assert "key_other" not in names
+
+    def test_search_scope_filtering_preserved(self, tmp_db):
+        """Search with scope filter only returns memories in that scope."""
+        save_structured_memory("ws1_fact", "alpha info", scope="workstream", scope_id="ws1")
+        save_structured_memory("ws2_fact", "alpha info", scope="workstream", scope_id="ws2")
+        save_structured_memory("global_fact", "alpha info", scope="global")
+
+        results = search_structured_memories("alpha", scope="workstream", scope_id="ws1")
+        names = {r["name"] for r in results}
+        assert "ws1_fact" in names
+        assert "ws2_fact" not in names
+        assert "global_fact" not in names
+
 
 class TestGetStructuredMemoryByName:
     def test_get_existing(self, tmp_db):

tests/test_structured_memory_storage.py

@@ -126,3 +126,148 @@ class TestCount:
         backend.create_structured_memory("m2", "b", "", "project", "workstream", "ws1", "2")
         assert backend.count_structured_memories(scope="global") == 1
         assert backend.count_structured_memories(scope="workstream") == 1
+
+
+class TestSearchOrOfTerms:
+    """Verify that multi-word search uses OR-of-terms (any term matches → row included)."""
+
+    def test_single_matching_term_in_multi_word_query(self, backend):
+        """Memory with content 'apple' found when query is 'apple banana cherry'."""
+        backend.create_structured_memory("m1", "apple_mem", "", "project", "global", "", "apple")
+        backend.create_structured_memory("m2", "other_mem", "", "project", "global", "", "grape")
+
+        results = backend.search_structured_memories("apple banana cherry")
+        names = {r["name"] for r in results}
+        assert "apple_mem" in names  # matches "apple" — OR-of-terms keeps it
+        assert "other_mem" not in names  # "grape" matches nothing in the query
+
+    def test_partial_overlap_across_memories(self, backend):
+        """Each memory matches one of three terms; all three are returned."""
+        backend.create_structured_memory("m1", "alpha_doc", "", "project", "global", "", "alpha")
+        backend.create_structured_memory("m2", "beta_doc", "", "project", "global", "", "beta")
+        backend.create_structured_memory("m3", "gamma_doc", "", "project", "global", "", "gamma")
+        backend.create_structured_memory("m4", "unrelated", "", "project", "global", "", "delta")
+
+        results = backend.search_structured_memories("alpha beta gamma")
+        names = {r["name"] for r in results}
+        assert "alpha_doc" in names
+        assert "beta_doc" in names
+        assert "gamma_doc" in names
+        assert "unrelated" not in names  # "delta" doesn't appear in the query
+
+    def test_scope_filter_preserved(self, backend):
+        """OR-of-terms search still respects scope / scope_id filters."""
+        backend.create_structured_memory(
+            "m1", "ws1_note", "", "project", "workstream", "ws1", "info"
+        )
+        backend.create_structured_memory(
+            "m2", "ws2_note", "", "project", "workstream", "ws2", "info"
+        )
+        backend.create_structured_memory("m3", "global_note", "", "project", "global", "", "info")
+
+        results = backend.search_structured_memories("info", scope="workstream", scope_id="ws1")
+        names = {r["name"] for r in results}
+        assert "ws1_note" in names
+        assert "ws2_note" not in names
+        assert "global_note" not in names
+
+    def test_term_cap_normalizes_unbounded_query(self, backend):
+        """A multi-KB query collapses to <= MAX terms (de-dupe + length filter)."""
+        backend.create_structured_memory("m1", "alpha_doc", "", "project", "global", "", "alpha")
+        backend.create_structured_memory(
+            "m2", "other_doc", "", "project", "global", "", "irrelevant"
+        )
+
+        # Build a noisy query: same word repeated, plus 1-char tokens that
+        # the normalizer drops, plus the actual signal "alpha".
+        noisy = " ".join(["x"] * 100 + ["alpha"] * 50)
+        results = backend.search_structured_memories(noisy)
+        names = {r["name"] for r in results}
+        assert "alpha_doc" in names
+
+
+class TestVisibleStructuredMemories:
+    """Single-query union helpers used by the composition path."""
+
+    def test_list_visible_unions_global_workstream_user(self, backend):
+        backend.create_structured_memory("m1", "g_note", "", "project", "global", "", "g")
+        backend.create_structured_memory("m2", "ws_note", "", "project", "workstream", "ws1", "w")
+        backend.create_structured_memory("m3", "u_note", "", "project", "user", "u1", "u")
+        backend.create_structured_memory("m4", "other_ws", "", "project", "workstream", "ws2", "x")
+
+        scopes = [("global", ""), ("workstream", "ws1"), ("user", "u1")]
+        rows = backend.list_visible_structured_memories(scopes)
+        names = {r["name"] for r in rows}
+        assert names == {"g_note", "ws_note", "u_note"}  # ws2 excluded
+
+    def test_search_visible_unions_scopes_and_terms(self, backend):
+        backend.create_structured_memory("m1", "g_alpha", "", "project", "global", "", "alpha")
+        backend.create_structured_memory(
+            "m2", "ws_beta", "", "project", "workstream", "ws1", "beta"
+        )
+        backend.create_structured_memory(
+            "m3", "ws_other", "", "project", "workstream", "ws2", "alpha"
+        )
+
+        scopes = [("global", ""), ("workstream", "ws1")]
+        rows = backend.search_visible_structured_memories("alpha beta", scopes)
+        names = {r["name"] for r in rows}
+        assert "g_alpha" in names  # global, matches "alpha"
+        assert "ws_beta" in names  # ws1, matches "beta"
+        assert "ws_other" not in names  # ws2 -> outside visibility
+
+    def test_visible_helpers_handle_empty_scopes(self, backend):
+        backend.create_structured_memory("m1", "anything", "", "project", "global", "", "x")
+        assert backend.list_visible_structured_memories([]) == []
+        assert backend.search_visible_structured_memories("x", []) == []
+
+
+class TestStableOrderingOnTimestampTies:
+    """When two memories share an `updated` timestamp, secondary sort on
+    memory_id keeps the order deterministic across calls.
+
+    `updated` is second-precision, and touch_structured_memories() can bump
+    a batch to identical timestamps — without a tie-breaker BM25 input
+    order shuffles run-to-run, busting the LLM-side prompt cache.
+    """
+
+    def _seed_with_shared_timestamp(self, backend):
+        # Create three memories then force their `updated` columns equal —
+        # mirrors the real-world case where a touch_structured_memories
+        # batch lands them in the same second.
+        for mid in ("zebra_id", "apple_id", "mango_id"):
+            backend.create_structured_memory(
+                mid, f"name_{mid}", "", "project", "global", "", "shared content"
+            )
+        import sqlalchemy as sa
+
+        with backend._conn() as conn:
+            conn.execute(sa.text("UPDATE structured_memories SET updated = '2024-01-01T00:00:00'"))
+            conn.commit()
+
+    def test_list_stable_order_under_tied_updated(self, backend):
+        self._seed_with_shared_timestamp(backend)
+        first = [r["memory_id"] for r in backend.list_structured_memories()]
+        second = [r["memory_id"] for r in backend.list_structured_memories()]
+        # Deterministic across calls AND sorted by memory_id ASC for ties
+        assert first == second
+        assert first == ["apple_id", "mango_id", "zebra_id"]
+
+    def test_search_stable_order_under_tied_updated(self, backend):
+        self._seed_with_shared_timestamp(backend)
+        first = [r["memory_id"] for r in backend.search_structured_memories("shared")]
+        second = [r["memory_id"] for r in backend.search_structured_memories("shared")]
+        assert first == second
+        assert first == ["apple_id", "mango_id", "zebra_id"]
+
+    def test_visible_search_stable_order_under_tied_updated(self, backend):
+        self._seed_with_shared_timestamp(backend)
+        scopes = [("global", "")]
+        first = [
+            r["memory_id"] for r in backend.search_visible_structured_memories("shared", scopes)
+        ]
+        second = [
+            r["memory_id"] for r in backend.search_visible_structured_memories("shared", scopes)
+        ]
+        assert first == second
+        assert first == ["apple_id", "mango_id", "zebra_id"]

turnstone/__init__.py

@@ -1,3 +1,3 @@
 """turnstone - Multi-node AI orchestration platform with tool use, agent routing, and cluster simulation."""
 
-__version__ = "1.5.4"
+__version__ = "1.5.6"

turnstone/console/collector.py

@@ -1258,6 +1258,17 @@ class ClusterCollector:
             entry["name"] = name
         self._fanout({"type": "ws_rename", "ws_id": ws_id, "name": name})
 
+    def emit_models_changed(self) -> None:
+        """Fan out a ``models_changed`` notice to all SSE listeners.
+
+        Browsers re-fetch :http:get:`/v1/api/models` on receipt so the
+        coordinator-composer model dropdown + the admin Models tab
+        reflect alias / underlying-model edits without a manual reload.
+        Body is intentionally empty — listeners refetch authoritative
+        state rather than diffing the event payload.
+        """
+        self._fanout({"type": "models_changed"})
+
     def emit_console_ws_intent_verdict(self, ws_id: str, verdict: dict[str, Any]) -> None:
         """Fan an LLM intent-judge verdict for a console-pseudo-node ws.
 

turnstone/console/coordinator_client.py

@@ -1081,7 +1081,38 @@ class CoordinatorClient:
                     "value": decoded,
                     "source": str(r.get("source", "")),
                 }
-            nodes.append({"node_id": nid, "metadata": meta})
+            # Project ``metadata.models`` (a list of
+            # ``{alias, provider, healthy}`` written by the node's
+            # heartbeat loop — see ``_collect_node_models_metadata``
+            # in ``turnstone/server.py``) down to the healthy-alias
+            # shortlist the coordinator passes back as ``model=`` to
+            # ``spawn_workstream`` / ``spawn_batch``.  The top-level
+            # field is named ``model_aliases`` (not ``models``) so it
+            # doesn't collide with ``metadata.models`` — the two
+            # carry different shapes (list of strings vs list of
+            # dicts) and a coord that conflates them gets a runtime
+            # error.  Empty list when the node hasn't published a
+            # models entry yet — older nodes without the heartbeat-
+            # side projection, or a brand new node mid-startup before
+            # the first metadata write.
+            models_entry = meta.get("models", {}).get("value")
+            healthy_aliases: list[str] = []
+            if isinstance(models_entry, list):
+                for row in models_entry:
+                    if not isinstance(row, dict):
+                        continue
+                    if not row.get("healthy", False):
+                        continue
+                    alias = row.get("alias")
+                    if isinstance(alias, str) and alias:
+                        healthy_aliases.append(alias)
+            nodes.append(
+                {
+                    "node_id": nid,
+                    "metadata": meta,
+                    "model_aliases": healthy_aliases,
+                }
+            )
         return {"nodes": nodes, "truncated": truncated}
 
     def list_skills(

turnstone/console/server.py

@@ -2413,7 +2413,7 @@ def _require_coord_mgr(request: Request) -> tuple[Any, JSONResponse | None]:
     if coord_mgr is None:
         registry_err = getattr(request.app.state, "coord_registry_error", "") or ""
         msg = "Coordinator subsystem not initialized. " + (
-            registry_err or "Check coordinator.model_alias and Models tab configuration."
+            registry_err or "Add a model definition in the admin Models tab."
         )
         return None, JSONResponse({"error": msg}, status_code=503)
     if config_store is None:
@@ -2444,8 +2444,7 @@ def _require_coord_mgr(request: Request) -> tuple[Any, JSONResponse | None]:
             {
                 "error": (
                     f"{hint} does not resolve: {exc}. "
-                    "Configure a model in the admin Models tab, or set "
-                    "``coordinator.model_alias`` in Settings to an existing alias."
+                    "Add or enable a model in the admin Models tab."
                 )
             },
             status_code=503,
@@ -4026,6 +4025,11 @@ async def _lifespan(app: Starlette) -> AsyncGenerator[None, None]:
                     # the cluster collector's pseudo-node so the
                     # dashboard tree mirrors child state.
                     event_emitter=coord_adapter,
+                    # Filter out persisted aliases that no longer resolve
+                    # so a coordinator pinned to a since-removed alias
+                    # still rehydrates (on the registry default) instead
+                    # of 500-ing on every reopen.
+                    model_validator=coord_registry.has_alias,
                 )
                 # Late-bind the manager onto the adapter so
                 # ``_rebuild_children_registry`` / ``send`` /
@@ -6976,6 +6980,35 @@ async def admin_delete_memory(request: Request) -> JSONResponse:
 # ---------------------------------------------------------------------------
 
 
+def _emit_models_changed(request: Request) -> None:
+    """Fan a ``models_changed`` SSE notice to connected browsers, if any.
+
+    Best-effort: silently no-ops when the collector isn't attached
+    (e.g. test fixtures that bypass the cluster collector).
+    """
+    collector = getattr(request.app.state, "collector", None)
+    if collector is not None:
+        collector.emit_models_changed()
+
+
+# Settings whose change should refresh the model dropdown / Roles UI in
+# every connected browser — covers the global default plus the per-role
+# overrides surfaced in the admin Models → Roles sub-tab.  Additions
+# here are purely additive (e.g. future ``perception.*.model`` keys).
+_MODEL_AFFECTING_SETTING_KEYS: frozenset[str] = frozenset(
+    {
+        "model.default_alias",
+        "model.plan_alias",
+        "model.plan_effort",
+        "model.task_alias",
+        "model.task_effort",
+        "coordinator.model_alias",
+        "coordinator.reasoning_effort",
+        "judge.model",
+    }
+)
+
+
 async def _publish_config_change(request: Request) -> None:
     """Fan out config-reload to all known server nodes (best-effort, async).
 
@@ -7178,6 +7211,8 @@ async def admin_update_setting(request: Request) -> JSONResponse:
     )
 
     await _publish_config_change(request)
+    if key in _MODEL_AFFECTING_SETTING_KEYS:
+        _emit_models_changed(request)
 
     return JSONResponse(
         {
@@ -7233,6 +7268,8 @@ async def admin_delete_setting(request: Request) -> JSONResponse:
     )
 
     await _publish_config_change(request)
+    if key in _MODEL_AFFECTING_SETTING_KEYS:
+        _emit_models_changed(request)
 
     return JSONResponse({"status": "ok", "key": key, "default": defn.default})
 
@@ -8053,6 +8090,33 @@ _MODEL_PROVIDERS = frozenset({"openai", "anthropic", "openai-compatible", "googl
 _REASONING_EFFORT_CHOICES = frozenset(
     {"", "none", "minimal", "low", "medium", "high", "xhigh", "max"}
 )
+# Keep in sync with turnstone.core.providers._VALID_API_SURFACES.
+_API_SURFACE_CHOICES = frozenset({"chat", "responses"})
+
+
+def _validate_api_surface(caps: Any) -> str | None:
+    """Return an error message if ``caps["server_compat"]["api_surface"]`` is invalid.
+
+    Strict equality match (no strip/lower normalisation): the persisted value
+    is bound directly to the admin ``<select>`` whose options are the canonical
+    ``"chat"`` / ``"responses"`` strings, so anything else fails to round-trip
+    through edit/save.  The provider factory raises ``ValueError`` at request
+    time for an unknown surface; validating here turns that into a 400 at
+    write time so an admin can't poison a model alias via direct API calls.
+    """
+    if not isinstance(caps, dict):
+        return None
+    sc = caps.get("server_compat")
+    if not isinstance(sc, dict):
+        return None
+    raw = sc.get("api_surface")
+    if raw is None or raw == "":
+        return None
+    if not isinstance(raw, str) or raw not in _API_SURFACE_CHOICES:
+        return f"Invalid server_compat.api_surface: {raw!r}"
+    return None
+
+
 # Keep in sync with turnstone.core.providers._google.GOOGLE_DEFAULT_BASE_URL
 _PROVIDER_DEFAULT_URLS: dict[str, str] = {
     "openai": "https://api.openai.com/v1",
@@ -8354,6 +8418,9 @@ async def admin_create_model_definition(request: Request) -> JSONResponse:
     ctx_raw = body.get("context_window", 32768)
     context_window = max(0, int(ctx_raw)) if isinstance(ctx_raw, (int, float)) else 0
     caps = body.get("capabilities", {})
+    err_msg = _validate_api_surface(caps)
+    if err_msg:
+        return JSONResponse({"error": err_msg}, status_code=400)
     capabilities = json.dumps(caps) if isinstance(caps, dict) else "{}"
     enabled = bool(body.get("enabled", True))
 
@@ -8414,6 +8481,7 @@ async def admin_create_model_definition(request: Request) -> JSONResponse:
     )
 
     await asyncio.to_thread(_refresh_coord_registry, request.app.state, storage)
+    _emit_models_changed(request)
 
     created = storage.get_model_definition(definition_id)
     if created is None:
@@ -8507,6 +8575,9 @@ async def admin_update_model_definition(request: Request) -> JSONResponse:
         updates["context_window"] = max(0, int(ctx_raw)) if isinstance(ctx_raw, (int, float)) else 0
     if "capabilities" in body:
         caps = body["capabilities"]
+        err_msg = _validate_api_surface(caps)
+        if err_msg:
+            return JSONResponse({"error": err_msg}, status_code=400)
         updates["capabilities"] = json.dumps(caps) if isinstance(caps, dict) else "{}"
     if "enabled" in body:
         updates["enabled"] = bool(body["enabled"])
@@ -8574,6 +8645,7 @@ async def admin_update_model_definition(request: Request) -> JSONResponse:
 
     if updates:
         await asyncio.to_thread(_refresh_coord_registry, request.app.state, storage)
+        _emit_models_changed(request)
 
     model_def = storage.get_model_definition(definition_id)
     return JSONResponse(_mask_model_secrets(model_def or {}))
@@ -8611,6 +8683,7 @@ async def admin_delete_model_definition(request: Request) -> JSONResponse:
     )
 
     await asyncio.to_thread(_refresh_coord_registry, request.app.state, storage)
+    _emit_models_changed(request)
 
     return JSONResponse({"status": "ok", "definition_id": definition_id})
 
@@ -8637,6 +8710,7 @@ async def admin_model_reload(request: Request) -> JSONResponse:
     # otherwise the coord LLM keeps calling the prior model name even
     # after a successful reload.
     await asyncio.to_thread(_refresh_coord_registry, request.app.state, storage)
+    _emit_models_changed(request)
 
     results = await _notify_nodes_model_reload(request)
     return JSONResponse({"status": "ok", "results": results})
@@ -9127,6 +9201,8 @@ async def admin_update_judge_setting(request: Request) -> JSONResponse:
         ip,
     )
     await _publish_config_change(request)
+    if key in _MODEL_AFFECTING_SETTING_KEYS:
+        _emit_models_changed(request)
 
     effective = config_store.get(key, defn.default)
     return JSONResponse(
@@ -9163,6 +9239,8 @@ async def admin_delete_judge_setting(request: Request) -> JSONResponse:
     audit_uid, ip = _audit_context(request)
     record_audit(storage, audit_uid, "setting.delete", "setting", key, {}, ip)
     await _publish_config_change(request)
+    if key in _MODEL_AFFECTING_SETTING_KEYS:
+        _emit_models_changed(request)
 
     return JSONResponse({"status": "ok", "key": key, "default": defn.default})
 

turnstone/console/static/admin.js

@@ -2625,11 +2625,22 @@ function loadSettings() {
         schemaMap[schemaArr[i].key] = schemaArr[i];
       }
 
-      // Merge values + schema
+      // Merge values + schema.  Skip role-assignment settings owned by
+      // the Models → Roles sub-tab (judge.* settings still live on the
+      // Judge tab; the four model-tab roles render only there).
       var merged = {};
+      var roleKeys = {
+        "coordinator.model_alias": 1,
+        "coordinator.reasoning_effort": 1,
+        "model.plan_alias": 1,
+        "model.plan_effort": 1,
+        "model.task_alias": 1,
+        "model.task_effort": 1,
+      };
       for (var j = 0; j < valuesArr.length; j++) {
         var v = valuesArr[j];
         if (v.key.startsWith("judge.")) continue;
+        if (roleKeys[v.key]) continue;
         var s = schemaMap[v.key] || {};
         merged[v.key] = {
           key: v.key,
@@ -4440,7 +4451,69 @@ var _modelDefaultAlias = "";
 var _modelCreateTrap = null;
 var _modelCreateTrigger = null;
 
+// Roles surfaced in the Models → Roles sub-tab.  Each entry maps a
+// settings-registry key onto a UX label.  ``effortKey`` is optional —
+// roles whose registry entry has a paired ``*.reasoning_effort``
+// setting render a second selector inline.  Adding a new role (e.g.
+// ``perception.audio.model``) is purely additive: drop a row here once
+// the SettingDef lands in turnstone/core/settings_registry.py.
+var MODEL_ROLES = [
+  {
+    label: "Coordinator",
+    description:
+      "Console-hosted coordinator sessions that drive child workstreams.",
+    aliasKey: "coordinator.model_alias",
+    effortKey: "coordinator.reasoning_effort",
+  },
+  {
+    label: "Judge",
+    description:
+      "Intent-validation judge that scores tool calls before approval.",
+    aliasKey: "judge.model",
+  },
+  {
+    label: "Plan agent",
+    description:
+      "plan_agent sub-agent — produces high-level plans before task dispatch.",
+    aliasKey: "model.plan_alias",
+    effortKey: "model.plan_effort",
+  },
+  {
+    label: "Task agent",
+    description:
+      "task_agent sub-agent — runs autonomous subtasks dispatched by the parent.",
+    aliasKey: "model.task_alias",
+    effortKey: "model.task_effort",
+  },
+];
+
+// Roles sub-tab reads/writes via ``/v1/api/admin/settings`` which
+// requires ``admin.settings`` — different from the ``admin.models``
+// permission gating the Models tab itself.  When the user has Models
+// access but not Settings, hide the sub-tab button + force the
+// Definitions panel visible so they don't see a perpetual 403 loader.
+function _modelRolesAccessible() {
+  var perms = sessionStorage.getItem("turnstone_permissions") || "";
+  return perms.split(",").indexOf("admin.settings") !== -1;
+}
+
+function _applyModelRolesPermission() {
+  var btn = document.getElementById("models-tab-roles");
+  if (!btn) return;
+  if (_modelRolesAccessible()) {
+    btn.style.display = "";
+    return;
+  }
+  btn.style.display = "none";
+  // If Roles was the active sub-tab, snap back to Definitions so the
+  // user isn't staring at a hidden panel.
+  if (btn.classList.contains("active")) {
+    switchModelsSection("models-list");
+  }
+}
+
 function loadAdminModels() {
+  _applyModelRolesPermission();
   authFetch("/v1/api/admin/model-definitions")
     .then(function (r) {
       if (!r.ok) throw new Error("Failed");
@@ -4450,6 +4523,12 @@ function loadAdminModels() {
       _modelDefs = data.models || [];
       _modelDefaultAlias = data.default_alias || "";
       _renderModels(_modelDefs);
+      // Roles sub-tab piggybacks on the model list; skip it when the
+      // user has no settings permission since the underlying API will
+      // 403 anyway.
+      if (_modelRolesAccessible()) {
+        loadAdminModelRoles();
+      }
     })
     .catch(function () {
       var el = document.getElementById("admin-models-table");
@@ -4461,6 +4540,223 @@ function loadAdminModels() {
     });
 }
 
+function switchModelsSection(section) {
+  var sections = document.querySelectorAll("#admin-models .models-section");
+  for (var i = 0; i < sections.length; i++) sections[i].style.display = "none";
+  var switcher = document.querySelector("#admin-models .admin-subtab-switcher");
+  var btns = switcher ? switcher.querySelectorAll(".admin-subtab-btn") : [];
+  for (var k = 0; k < btns.length; k++) {
+    var isActive = btns[k].getAttribute("data-section") === section;
+    btns[k].classList.toggle("active", isActive);
+    btns[k].setAttribute("aria-selected", isActive ? "true" : "false");
+    btns[k].setAttribute("tabindex", isActive ? "0" : "-1");
+  }
+  var target = document.getElementById(section + "-section");
+  if (target) target.style.display = "";
+}
+
+// Arrow key navigation for Models sub-tabs (matches the Judge tab).
+(function () {
+  var switcher = document.querySelector("#admin-models .admin-subtab-switcher");
+  if (!switcher) return;
+  switcher.addEventListener("keydown", function (e) {
+    if (e.key !== "ArrowLeft" && e.key !== "ArrowRight") return;
+    var btns = switcher.querySelectorAll(".admin-subtab-btn");
+    var secs = [];
+    for (var i = 0; i < btns.length; i++)
+      secs.push(btns[i].getAttribute("data-section"));
+    var current = switcher.querySelector(".admin-subtab-btn.active");
+    var idx = secs.indexOf(current ? current.getAttribute("data-section") : "");
+    if (e.key === "ArrowRight") idx = (idx + 1) % secs.length;
+    else idx = (idx - 1 + secs.length) % secs.length;
+    e.preventDefault();
+    switchModelsSection(secs[idx]);
+    btns[idx].focus();
+  });
+})();
+
+function _modelRolesError(container, msg) {
+  while (container.firstChild) container.removeChild(container.firstChild);
+  var d = document.createElement("div");
+  d.className = "dashboard-empty";
+  d.textContent = msg;
+  container.appendChild(d);
+}
+
+function loadAdminModelRoles() {
+  var c = document.getElementById("admin-models-roles-container");
+  if (!c) return;
+  // Reads ``_modelDefs`` / ``_modelDefaultAlias`` populated by the most
+  // recent ``loadAdminModels`` — both entry points into the Models tab
+  // (initial open + ``models_changed`` SSE refresh) go through
+  // ``loadAdminModels`` first, so the cached snapshot is fresh.  Role
+  // saves don't change model definitions, so the snapshot stays
+  // accurate after ``_saveModelRole`` chains back here.
+  Promise.all([
+    authFetch("/v1/api/admin/settings").then(function (r) {
+      if (!r.ok) throw new Error("settings " + r.status);
+      return r.json();
+    }),
+    authFetch("/v1/api/admin/settings/schema").then(function (r) {
+      if (!r.ok) throw new Error("schema " + r.status);
+      return r.json();
+    }),
+  ])
+    .then(function (results) {
+      var values = {};
+      var arr = results[0].settings || [];
+      for (var i = 0; i < arr.length; i++) values[arr[i].key] = arr[i];
+      var schema = {};
+      var sa = results[1].schema || [];
+      for (var j = 0; j < sa.length; j++) schema[sa[j].key] = sa[j];
+      _renderModelRoles(c, values, schema);
+    })
+    .catch(function () {
+      _modelRolesError(c, "Failed to load roles");
+    });
+}
+
+function _renderModelRoles(container, values, schema) {
+  var enabledAliases = [];
+  for (var i = 0; i < _modelDefs.length; i++) {
+    if (_modelDefs[i].enabled) enabledAliases.push(_modelDefs[i]);
+  }
+  container.textContent = "";
+  for (var r = 0; r < MODEL_ROLES.length; r++) {
+    var role = MODEL_ROLES[r];
+    var aliasInfo = values[role.aliasKey];
+    if (!aliasInfo) continue; // setting not registered (e.g. older server)
+
+    var row = document.createElement("div");
+    row.className = "model-role-row";
+
+    // The dropdown's selected-option text is the single source of
+    // truth for default vs override — when nothing is set it shows
+    // "(default — <alias>)", otherwise it shows the chosen alias.  No
+    // separate badge: redundant with the select, and prone to
+    // confusing color contrasts on freshly-rendered rows.
+    var head = document.createElement("div");
+    head.className = "model-role-head";
+    var nameEl = document.createElement("span");
+    nameEl.className = "model-role-label";
+    nameEl.textContent = role.label;
+    head.appendChild(nameEl);
+    row.appendChild(head);
+
+    if (role.description) {
+      var desc = document.createElement("div");
+      desc.className = "model-role-desc";
+      desc.textContent = role.description;
+      row.appendChild(desc);
+    }
+
+    var controls = document.createElement("div");
+    controls.className = "model-role-controls";
+
+    // Alias dropdown
+    var aliasWrap = document.createElement("label");
+    aliasWrap.className = "model-role-control";
+    var aliasLabel = document.createElement("span");
+    aliasLabel.className = "model-role-control-label";
+    aliasLabel.textContent = "Model";
+    aliasWrap.appendChild(aliasLabel);
+    var aliasSel = document.createElement("select");
+    aliasSel.setAttribute("data-role-key", role.aliasKey);
+    aliasSel.setAttribute(
+      "aria-label",
+      role.label + " model (empty = default)",
+    );
+    var blank = document.createElement("option");
+    blank.value = "";
+    blank.textContent = _modelDefaultAlias
+      ? "(default — " + _modelDefaultAlias + ")"
+      : "(default)";
+    aliasSel.appendChild(blank);
+    var currentAlias = aliasInfo.value || "";
+    var matched = false;
+    for (var m = 0; m < enabledAliases.length; m++) {
+      var md = enabledAliases[m];
+      var opt = document.createElement("option");
+      opt.value = md.alias;
+      opt.textContent =
+        md.alias === md.model ? md.alias : md.alias + " (" + md.model + ")";
+      if (currentAlias && currentAlias === md.alias) {
+        opt.selected = true;
+        matched = true;
+      }
+      aliasSel.appendChild(opt);
+    }
+    if (currentAlias && !matched) {
+      var manual = document.createElement("option");
+      manual.value = currentAlias;
+      manual.textContent = currentAlias + " (manual)";
+      manual.selected = true;
+      aliasSel.appendChild(manual);
+    }
+    aliasSel.addEventListener("change", function () {
+      _saveModelRole(this.getAttribute("data-role-key"), this.value);
+    });
+    aliasWrap.appendChild(aliasSel);
+    controls.appendChild(aliasWrap);
+
+    // Optional reasoning effort dropdown
+    if (role.effortKey && values[role.effortKey] && schema[role.effortKey]) {
+      var effortWrap = document.createElement("label");
+      effortWrap.className = "model-role-control";
+      var effortLabel = document.createElement("span");
+      effortLabel.className = "model-role-control-label";
+      effortLabel.textContent = "Reasoning effort";
+      effortWrap.appendChild(effortLabel);
+      var effortSel = document.createElement("select");
+      effortSel.setAttribute("data-role-key", role.effortKey);
+      effortSel.setAttribute("aria-label", role.label + " reasoning effort");
+      var choices = schema[role.effortKey].choices || [];
+      var currentEffort = values[role.effortKey].value;
+      for (var c2 = 0; c2 < choices.length; c2++) {
+        var eo = document.createElement("option");
+        eo.value = choices[c2];
+        eo.textContent = choices[c2] === "" ? "(inherit)" : choices[c2];
+        if (currentEffort === choices[c2]) eo.selected = true;
+        effortSel.appendChild(eo);
+      }
+      effortSel.addEventListener("change", function () {
+        _saveModelRole(this.getAttribute("data-role-key"), this.value);
+      });
+      effortWrap.appendChild(effortSel);
+      controls.appendChild(effortWrap);
+    }
+
+    row.appendChild(controls);
+    container.appendChild(row);
+  }
+
+  if (!container.children.length) {
+    _modelRolesError(container, "No model roles configured");
+  }
+}
+
+function _saveModelRole(key, value) {
+  authFetch("/v1/api/admin/settings/" + encodeURIComponent(key), {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ value: value }),
+  })
+    .then(function (r) {
+      if (!r.ok)
+        return r.json().then(function (d) {
+          throw new Error(d.error || "Failed");
+        });
+      return r.json();
+    })
+    .then(function () {
+      showToast("Saved");
+      loadAdminModelRoles();
+    })
+    .catch(function (e) {
+      showToast("Error: " + (e && e.message ? e.message : "save failed"));
+    });
+}
+
 function _renderModels(items) {
   var el = document.getElementById("admin-models-table");
   // Clear previous content
@@ -4705,6 +5001,7 @@ function showCreateModelModal() {
   document.getElementById("model-max-tokens").value = "";
   document.getElementById("model-reasoning-effort").value = "";
   document.getElementById("model-server-type").value = "";
+  document.getElementById("model-api-surface").value = "";
   document.getElementById("model-thinking-mode").value = "";
   document.getElementById("model-thinking-param").value = "";
   document.getElementById("model-thinking-param-row").style.display = "none";
@@ -4781,8 +5078,9 @@ function showEditModelModal(definitionId) {
         document.getElementById("model-thinking-param").value = "";
       }
       _toggleThinkingParam();
-      // Server compat: server_type and extra_body workarounds
+      // Server compat: server_type, api_surface, and extra_body workarounds
       document.getElementById("model-server-type").value = sc.server_type || "";
+      document.getElementById("model-api-surface").value = sc.api_surface || "";
       var eb = sc.extra_body || {};
       var ebText = JSON.stringify(eb, null, 2);
       document.getElementById("model-extra-body").value =
@@ -4864,26 +5162,34 @@ function submitCreateModel() {
     if (savedParam) caps.thinking_param = savedParam;
   }
 
-  // Build server_compat from structured fields
+  // Build server_compat from structured fields.  Only meaningful for
+  // openai-compatible aliases — for other providers the section is hidden
+  // but the form values can linger after a provider switch, so gate the
+  // whole block on the active provider to keep persisted state honest.
   var serverCompat = {};
-  var serverType = document.getElementById("model-server-type").value;
-  if (serverType) serverCompat.server_type = serverType;
+  var providerVal = document.getElementById("model-provider").value;
   var ebEl = document.getElementById("model-extra-body");
-  var ebText = ebEl.value.trim();
   ebEl.removeAttribute("aria-invalid");
   ebEl.style.borderColor = "";
-  if (ebText) {
-    try {
-      var ebParsed = JSON.parse(ebText);
-      if (!_isPlainObject(ebParsed)) {
-        throw new Error("not an object");
+  if (providerVal === "openai-compatible") {
+    var serverType = document.getElementById("model-server-type").value;
+    if (serverType) serverCompat.server_type = serverType;
+    var apiSurface = document.getElementById("model-api-surface").value;
+    if (apiSurface) serverCompat.api_surface = apiSurface;
+    var ebText = ebEl.value.trim();
+    if (ebText) {
+      try {
+        var ebParsed = JSON.parse(ebText);
+        if (!_isPlainObject(ebParsed)) {
+          throw new Error("not an object");
+        }
+        serverCompat.extra_body = ebParsed;
+      } catch (e) {
+        ebEl.setAttribute("aria-invalid", "true");
+        ebEl.style.borderColor = "var(--red)";
+        _showModelError("Extra body params must be a JSON object");
+        return;
       }
-      serverCompat.extra_body = ebParsed;
-    } catch (e) {
-      ebEl.setAttribute("aria-invalid", "true");
-      ebEl.style.borderColor = "var(--red)";
-      _showModelError("Extra body params must be a JSON object");
-      return;
     }
   }
   if (Object.keys(serverCompat).length > 0) {
@@ -5114,6 +5420,13 @@ function detectModel() {
           stOpts2.indexOf(ssc.server_type) !== -1
         )
           stEl2.value = ssc.server_type;
+        // Restrict to the known set so a hostile detect response can't
+        // smuggle a non-listed value into the form.
+        var _SURFACE_SUGGESTABLE = { chat: 1, responses: 1 };
+        if (ssc.api_surface && _SURFACE_SUGGESTABLE[ssc.api_surface]) {
+          var asEl = document.getElementById("model-api-surface");
+          if (!asEl.value) asEl.value = ssc.api_surface;
+        }
         if (ssc.extra_body) {
           var ebEl2 = document.getElementById("model-extra-body");
           if (!ebEl2.value.trim()) {

turnstone/console/static/app.js

@@ -5,17 +5,13 @@ window.onLoginSuccess = function () {
   if (typeof _refreshHomeComposerVisibility === "function") {
     _refreshHomeComposerVisibility();
   }
-  // Re-populate the home-composer skill dropdown and re-probe the
-  // coordinator subsystem now that auth has landed.  The initial
-  // page-load pass runs before login completes, so /v1/api/skills
-  // and /v1/api/workstreams both 401; without this re-run the
-  // dropdown stays empty and the 503 banner never flips correctly.
+  // Re-populate the home-composer skill dropdown now that auth has
+  // landed.  The initial page-load pass runs before login completes,
+  // so /v1/api/skills 401s; without this re-run the dropdown stays
+  // empty.
   if (typeof _populateHomeSkillDropdown === "function") {
     _populateHomeSkillDropdown();
   }
-  if (typeof _probeCoordSubsystem === "function") {
-    _probeCoordSubsystem();
-  }
   // Active-coordinators list is SSE-driven via the console pseudo-node
   // (#9) — no poller to restart after login.  The home-view renderer
   // reads from clusterState.nodes["console"].workstreams on every SSE
@@ -427,6 +423,22 @@ function handleClusterEvent(data) {
   if (data.type === "ws_closed" && data.reason === "evicted") {
     showToast("Evicted" + (data.name ? ": " + data.name : "") + " (capacity)");
   }
+  if (data.type === "models_changed") {
+    // Server emits this when a model definition or a role-assignment
+    // setting (model.default_alias, judge.model, coordinator.model_alias,
+    // coordinator.reasoning_effort) changes.  Refresh anything that
+    // renders model aliases so labels stay accurate without a reload.
+    if (typeof _populateHomeModelDropdowns === "function") {
+      _populateHomeModelDropdowns();
+    }
+    if (
+      typeof _adminTab !== "undefined" &&
+      _adminTab === "models" &&
+      typeof loadAdminModels === "function"
+    ) {
+      loadAdminModels();
+    }
+  }
 }
 
 // --- Home View ---
@@ -1445,8 +1457,8 @@ function _hasCoordPermission() {
 // POST /v1/api/workstreams/new.  Accepts the three request fields
 // directly + an errEl / setBusy callback so the caller owns the
 // loading-state UX (button label swap, composer disabled flag, etc.).
-// On success redirects to /coordinator/{ws_id}; on 503 invokes on503
-// so the caller can surface the "subsystem not configured" banner.
+// On success redirects to /coordinator/{ws_id}; on failure surfaces
+// the server's error text inline through errEl.
 function _createCoordinator(opts) {
   var name = (opts.name || "").trim();
   var skill = opts.skill || "";
@@ -1455,10 +1467,12 @@ function _createCoordinator(opts) {
   var task = (opts.task || "").trim();
   var errEl = opts.errEl;
   var setBusy = opts.setBusy || function () {};
-  var on503 = opts.on503 || function () {};
   var onSuccess = opts.onSuccess || function () {};
 
-  errEl.style.display = "none";
+  // Error region is always rendered with reserved min-height (see
+  // .home-composer-error in style.css) so toggling validation messages
+  // doesn't reflow the active-coordinators list below — clear the
+  // textContent only, no display toggle.
   errEl.textContent = "";
   setBusy(true);
 
@@ -1469,11 +1483,30 @@ function _createCoordinator(opts) {
   if (judgeModel) body.judge_model = judgeModel;
   if (task) body.initial_message = task;
 
-  authFetch("/v1/api/workstreams/new", {
-    method: "POST",
-    headers: { "Content-Type": "application/json" },
-    body: JSON.stringify(body),
-  })
+  // Multipart when files are staged — the coord create endpoint
+  // accepts a `meta` JSON field plus zero-or-more `file` parts and
+  // reserves attachments for the very first turn (same flow the
+  // interactive UI's new-ws modal uses against the server).  Plain
+  // JSON stays the default when no files are attached.
+  var files = Array.isArray(opts.files) ? opts.files : [];
+  var fetchOpts;
+  if (files.length > 0) {
+    var form = new FormData();
+    form.append("meta", JSON.stringify(body));
+    for (var i = 0; i < files.length; i++) {
+      form.append("file", files[i], files[i].name);
+    }
+    // Don't set Content-Type — the browser adds the correct boundary.
+    fetchOpts = { method: "POST", body: form };
+  } else {
+    fetchOpts = {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    };
+  }
+
+  authFetch("/v1/api/workstreams/new", fetchOpts)
     .then(function (r) {
       return r.json().then(function (data) {
         return { ok: r.ok, status: r.status, data: data };
@@ -1481,14 +1514,9 @@ function _createCoordinator(opts) {
     })
     .then(function (res) {
       setBusy(false);
-      if (res.status === 503) {
-        on503(res);
-        return;
-      }
       if (!res.ok || !res.data || !res.data.ws_id) {
         errEl.textContent =
           (res.data && res.data.error) || "HTTP " + res.status;
-        errEl.style.display = "block";
         return;
       }
       onSuccess(res);
@@ -1498,7 +1526,6 @@ function _createCoordinator(opts) {
     .catch(function () {
       setBusy(false);
       errEl.textContent = "Request failed";
-      errEl.style.display = "block";
     });
 }
 
@@ -1511,19 +1538,166 @@ function _createCoordinator(opts) {
 // ---------------------------------------------------------------------------
 
 var _homeComposerInit = false;
-var _homeCoordReady = null; // tri-state: null = unknown, true = ready, false = 503
 var _homeCoordComposer = null; // shared Composer instance
 var _homeCoordBusy = false;
 
-// Single owner for sendBtn.disabled: disabled if EITHER busy OR the
-// subsystem probe flipped to 503.  Every setter for _homeCoordBusy /
-// _homeCoordReady ends with a call here so the two inputs can't drift
-// out of sync (and a probe resolving mid-submit can't re-enable the
-// button under an in-flight request).
+// Attachment staging for the home coord composer.  The coord ws_id
+// doesn't exist until the create POST resolves, so we hold File
+// objects in memory and ship them as multipart parts on submit (same
+// pattern interactive uses for its new-ws modal + dashboard composer).
+var _homeStagedFiles = [];
+
+// Per-kind size caps + allowlist mirrored from turnstone/core/attachments.py
+// so the browser can fail fast.  Keep in sync with the interactive
+// UI's _ATTACH_* constants in turnstone/ui/static/app.js.
+var _HOME_IMAGE_CAP = 4 * 1024 * 1024;
+var _HOME_TEXT_CAP = 512 * 1024;
+var _HOME_MAX_FILES = 10;
+var _HOME_IMAGE_MIMES = ["image/png", "image/jpeg", "image/gif", "image/webp"];
+var _HOME_TEXT_APP_MIMES = [
+  "application/json",
+  "application/xml",
+  "application/x-yaml",
+  "application/yaml",
+  "application/toml",
+];
+var _HOME_TEXT_EXTENSIONS = [
+  ".c",
+  ".conf",
+  ".cpp",
+  ".css",
+  ".go",
+  ".h",
+  ".hpp",
+  ".html",
+  ".ini",
+  ".java",
+  ".js",
+  ".json",
+  ".jsx",
+  ".md",
+  ".py",
+  ".rs",
+  ".sh",
+  ".sql",
+  ".toml",
+  ".ts",
+  ".tsx",
+  ".txt",
+  ".xml",
+  ".yaml",
+  ".yml",
+];
+
+function _homeFormatSize(n) {
+  if (n < 1024) return n + " B";
+  if (n < 1024 * 1024) return (n / 1024).toFixed(1) + " KB";
+  return (n / (1024 * 1024)).toFixed(1) + " MB";
+}
+
+function _homeIsAttachmentAllowed(file) {
+  var mime = (file.type || "").toLowerCase();
+  if (_HOME_IMAGE_MIMES.indexOf(mime) !== -1) return true;
+  if (mime.indexOf("text/") === 0) return true;
+  if (_HOME_TEXT_APP_MIMES.indexOf(mime) !== -1) return true;
+  var name = (file.name || "").toLowerCase();
+  var dot = name.lastIndexOf(".");
+  if (dot >= 0 && _HOME_TEXT_EXTENSIONS.indexOf(name.substr(dot)) !== -1) {
+    return true;
+  }
+  return false;
+}
+
+function _homeShowError(msg) {
+  var errEl = document.getElementById("home-coord-error");
+  if (!errEl) return;
+  // Element is always rendered (min-height reserves the row); just
+  // toggle the message text so layout doesn't shift on validation.
+  errEl.textContent = msg || "";
+}
+
+function _homeRenderChips() {
+  if (!_homeCoordComposer || !_homeCoordComposer.chipsEl) return;
+  var chipsEl = _homeCoordComposer.chipsEl;
+  chipsEl.textContent = "";
+  for (var i = 0; i < _homeStagedFiles.length; i++) {
+    (function (idx) {
+      var f = _homeStagedFiles[idx];
+      var isImage = (f.type || "").indexOf("image/") === 0;
+      var chip = document.createElement("span");
+      chip.className =
+        "composer-chip composer-chip-" + (isImage ? "image" : "text");
+      chip.setAttribute("role", "listitem");
+
+      var icon = document.createElement("span");
+      icon.className = "composer-chip-icon";
+      icon.setAttribute("aria-hidden", "true");
+      icon.textContent = isImage ? "🖼" : "📄";
+      chip.appendChild(icon);
+
+      var name = document.createElement("span");
+      name.className = "composer-chip-name";
+      name.textContent = f.name;
+      name.title = f.name + " (" + f.size + " bytes)";
+      chip.appendChild(name);
+
+      var size = document.createElement("span");
+      size.className = "composer-chip-size";
+      size.textContent = _homeFormatSize(f.size);
+      chip.appendChild(size);
+
+      var rm = document.createElement("button");
+      rm.type = "button";
+      rm.className = "composer-chip-remove";
+      rm.setAttribute("aria-label", "Remove " + f.name);
+      rm.title = "Remove";
+      rm.textContent = "×";
+      rm.onclick = function () {
+        _homeStagedFiles.splice(idx, 1);
+        _homeRenderChips();
+      };
+      chip.appendChild(rm);
+      chipsEl.appendChild(chip);
+    })(i);
+  }
+}
+
+function _homeStageFile(file) {
+  if (!file) return;
+  if (_homeStagedFiles.length >= _HOME_MAX_FILES) {
+    _homeShowError(
+      "At most " + _HOME_MAX_FILES + " attachments per coordinator",
+    );
+    return;
+  }
+  if (!_homeIsAttachmentAllowed(file)) {
+    _homeShowError(
+      "Unsupported file type: " +
+        file.name +
+        " (allowed: png/jpeg/gif/webp images, text)",
+    );
+    return;
+  }
+  var isImage = (file.type || "").indexOf("image/") === 0;
+  var cap = isImage ? _HOME_IMAGE_CAP : _HOME_TEXT_CAP;
+  if (file.size > cap) {
+    _homeShowError(file.name + " exceeds the " + _homeFormatSize(cap) + " cap");
+    return;
+  }
+  _homeShowError("");
+  _homeStagedFiles.push(file);
+  _homeRenderChips();
+}
+
+function _homeClearStagedFiles() {
+  _homeStagedFiles = [];
+  _homeRenderChips();
+}
+
+// Sole owner of sendBtn.disabled: disables while a submit is in flight.
 function _refreshHomeCoordSubmitEnabled() {
   if (!_homeCoordComposer) return;
-  _homeCoordComposer.sendBtn.disabled =
-    _homeCoordBusy || _homeCoordReady === false;
+  _homeCoordComposer.sendBtn.disabled = _homeCoordBusy;
 }
 
 function _ensureHomeComposerInit() {
@@ -1532,7 +1706,6 @@ function _ensureHomeComposerInit() {
   _mountHomeCoordComposer();
   _populateHomeSkillDropdown();
   _populateHomeModelDropdowns();
-  _probeCoordSubsystem();
   _refreshHomeComposerVisibility();
 }
 
@@ -1596,6 +1769,12 @@ function _mountHomeCoordComposer() {
         },
       ],
     },
+    attachments: {
+      onAttach: function (file) {
+        _homeStageFile(file);
+      },
+    },
+    dragDrop: { targetEl: mount, dropClass: "home-coord-drop" },
     onSend: function (text) {
       submitHomeCoord(text);
     },
@@ -1646,40 +1825,6 @@ function _populateHomeModelDropdowns() {
     });
 }
 
-// Probe GET /v1/api/workstreams — 200 = subsystem ready; 503 = no model
-// alias resolvable, show remediation banner.  4xx (auth / permission) is
-// treated as "unknown, don't flip the banner" because the probe cannot
-// actually tell us anything about subsystem readiness in that case —
-// the caller is expected to re-invoke this after login lands so a real
-// answer can arrive.  Leaving the submit button enabled on unknown
-// keeps first-paint usable; a subsequent 503 from the actual submit
-// flips the banner via _createCoordinator's on503 hook.
-//
-// Skip the probe entirely for users without admin.coordinator — they
-// can't see the composer anyway (see _refreshHomeComposerVisibility),
-// and the endpoint returns 403 for them, producing a useless network
-// round-trip on every login.
-function _probeCoordSubsystem() {
-  if (!_hasCoordPermission()) return;
-  authFetch("/v1/api/workstreams")
-    .then(function (r) {
-      if (r.status === 503) {
-        _homeCoordReady = false;
-      } else if (r.ok) {
-        _homeCoordReady = true;
-      } else {
-        _homeCoordReady = null;
-        return;
-      }
-      var banner = document.getElementById("coord-composer-503");
-      if (banner) banner.style.display = _homeCoordReady ? "none" : "";
-      _refreshHomeCoordSubmitEnabled();
-    })
-    .catch(function () {
-      /* network error — leave banner hidden; submit will surface a retryable error */
-    });
-}
-
 function _refreshHomeComposerVisibility() {
   var panel = document.getElementById("coord-composer-panel");
   if (!panel) return;
@@ -1698,23 +1843,36 @@ function submitHomeCoord(textFromComposer) {
   var task =
     textFromComposer != null ? textFromComposer : _homeCoordComposer.value;
   var opts = _homeCoordComposer.getOptionValues();
+  // Snapshot at submit time so a chip remove mid-request can't race
+  // the multipart payload (the actual reset only fires on the success
+  // branch, after the response lands).
+  var files = _homeStagedFiles.slice();
+  // Files-without-text would upload pending attachment rows but the
+  // server's _coord_create_post_install only reserves+dispatches when
+  // initial_message is non-empty — uploaded files would orphan as
+  // pending storage rows until the GC sweep.  Require text whenever
+  // attachments are staged so the first turn always picks them up.
+  if (files.length > 0 && !(task || "").trim()) {
+    _homeShowError(
+      "Add a task message — attachments need an initial turn to dispatch on.",
+    );
+    return;
+  }
   _createCoordinator({
     name: opts.name || "",
     skill: opts.skill || "",
     model: opts.model || "",
     judge_model: opts.judge_model || "",
     task: task,
+    files: files,
     errEl: document.getElementById("home-coord-error"),
     setBusy: function (b) {
       _homeCoordBusy = b;
       if (_homeCoordComposer) _homeCoordComposer.setBusy(b);
       _refreshHomeCoordSubmitEnabled();
     },
-    on503: function () {
-      _homeCoordReady = false;
-      var banner = document.getElementById("coord-composer-503");
-      if (banner) banner.style.display = "";
-      _refreshHomeCoordSubmitEnabled();
+    onSuccess: function () {
+      _homeClearStagedFiles();
     },
   });
 }
@@ -1730,9 +1888,6 @@ document.addEventListener("keydown", function (e) {
   var mount = document.getElementById("home-coord-composer-mount");
   if (!mount || !mount.contains(e.target)) return;
   e.preventDefault();
-  // sendBtn.disabled is the single reconciler of busy + 503-ready —
-  // checking it here is enough to avoid double-submits or submits
-  // while the subsystem is down.
   if (!_homeCoordComposer.sendBtn.disabled) submitHomeCoord();
 });
 

turnstone/console/static/coordinator/coordinator.css

@@ -396,6 +396,87 @@
   background: color-mix(in srgb, var(--err) 12%, var(--panel-2));
 }
 
+/* Output-guard finding rendered under its specific .coord-tool-row.
+   Stays anchored to the call that tripped the guard rather than
+   floating into the chat log as a generic "[output guard]" line —
+   matches interactive's `.output-warning` placement convention.
+   Severity drives the hue (matches .verdict-badge.verdict-* palette
+   so an operator scanning a workstream reads risk consistently
+   across both surfaces). */
+.coord-tool-row-warning {
+  display: inline-flex;
+  align-items: center;
+  gap: 6px;
+  margin-top: 4px;
+  padding: 2px 8px;
+  font-size: 11px;
+  font-family: var(--font-mono);
+  border: 1px solid var(--hair);
+  border-radius: 3px;
+  border-left-width: 3px;
+  background: var(--panel-2);
+  color: var(--ink-2);
+  max-width: max-content;
+}
+.coord-tool-row-warning--low {
+  color: color-mix(in srgb, var(--ok) 70%, var(--ink-2));
+  border-left-color: var(--ok);
+  background: color-mix(in srgb, var(--ok) 12%, var(--panel-2));
+}
+.coord-tool-row-warning--medium {
+  color: color-mix(in srgb, var(--warn) 70%, var(--ink-2));
+  border-left-color: var(--warn);
+  background: var(--warn-tint);
+}
+.coord-tool-row-warning--high,
+.coord-tool-row-warning--critical {
+  color: color-mix(in srgb, var(--err) 70%, var(--ink-2));
+  border-left-color: var(--err);
+  background: color-mix(in srgb, var(--err) 12%, var(--panel-2));
+}
+.coord-tool-row-warning-redacted {
+  color: var(--ink-3);
+  font-size: 10px;
+}
+
+/* Storage-truncation indicator — same convention as the interactive
+   UI's `.tool-output-truncated` pill (transparent bg, dim border,
+   small font) so the operator reads the affordance the same way on
+   both surfaces.  Sibling node next to .coord-tool-row-result rather
+   than text-in-content so a future "best-effort JSON repair" pass
+   on the result body doesn't have to strip a marker string. */
+.coord-tool-truncated {
+  display: inline-block;
+  margin-top: 4px;
+  margin-left: 6px;
+  padding: 1px 6px;
+  font-size: 10px;
+  font-family: var(--font-mono);
+  color: var(--ink-3);
+  background: transparent;
+  border: 1px solid var(--ink-3);
+  border-radius: 3px;
+}
+
+/* memory/recall calls are background metadata — the audit trail is
+   useful but they crowd the tree on workstreams with heavy memory
+   usage.  Dim the row by default; full opacity on hover so they
+   stay inspectable.  Mirrors the interactive UI's metacog dim rule
+   (style.css `.ts-approval-tool[data-func-name="memory"]`).  The
+   row stamps `data-tool-name` from item.func_name in coordinator.js
+   so this selector has something to match. */
+.coord-tool-row[data-tool-name="memory"],
+.coord-tool-row[data-tool-name="recall"] {
+  opacity: 0.55;
+  transition: opacity 120ms ease-out;
+}
+.coord-tool-row[data-tool-name="memory"]:hover,
+.coord-tool-row[data-tool-name="memory"]:focus-within,
+.coord-tool-row[data-tool-name="recall"]:hover,
+.coord-tool-row[data-tool-name="recall"]:focus-within {
+  opacity: 1;
+}
+
 /* Tool result block — paired under its row, mono pre-block.  Capped
    at 240px with internal scroll so a long tool output doesn't push
    the rest of the chat off-screen.  The interactive UI uses a
@@ -537,6 +618,43 @@
   color: var(--ink-3);
 }
 
+/* User-message attachment pills — rendered beneath the bubble for
+   live sends and on history replay.  Mirrors the .msg-user-attach*
+   rules in turnstone/ui/static/style.css so coord and interactive
+   surfaces show the same affordance for attached files. */
+.msg-user-attach {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 4px;
+  margin-top: 6px;
+}
+.msg-user-attach-pill {
+  display: inline-flex;
+  align-items: center;
+  gap: 4px;
+  /* --panel (not --panel-2) — the .msg bubble is already --panel-2,
+     so pulling the pill onto the alternate surface keeps it visible
+     against the bubble in both themes (WCAG 1.4.11 non-text contrast).
+     Mirrors the interactive UI's --bg-surface vs .msg --panel-2 split. */
+  background: var(--panel);
+  color: var(--ink-2);
+  border: 1px solid var(--hair);
+  border-radius: 4px;
+  padding: 2px 6px;
+  font-family: var(--font-mono);
+  font-size: 10px;
+}
+.msg-user-attach-icon {
+  font-size: 11px;
+  opacity: 0.7;
+}
+.msg-user-attach-name {
+  max-width: 200px;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
 /* Mobile (<700px) — keep action targets ≥44px for WCAG 2.5.5. */
 @media (max-width: 700px) {
   .coord-tool-actions {

turnstone/console/static/coordinator/coordinator.js

@@ -335,6 +335,39 @@
     return appendMsg(role, esc(text), opts);
   }
 
+  // User-message bubble with attachment-pill cluster appended below
+  // the text.  Mirrors Pane.prototype.addUserMessage in the
+  // interactive UI so live-send and history-replay both render the
+  // same chip strip the composer staged on submit.  Attachments is a
+  // list of {kind, filename}; falsy/empty falls through to plain text.
+  function appendUserMessageWithAttachments(text, attachments, opts) {
+    const el = appendText("user", text, opts);
+    if (!Array.isArray(attachments) || attachments.length === 0) return el;
+    const pills = document.createElement("div");
+    pills.className = "msg-user-attach";
+    pills.setAttribute("role", "list");
+    attachments.forEach((a) => {
+      const kind = (a && a.kind) || "other";
+      const pill = document.createElement("span");
+      pill.className = "msg-user-attach-pill msg-user-attach-pill-" + kind;
+      pill.setAttribute("role", "listitem");
+      const icon = document.createElement("span");
+      icon.className = "msg-user-attach-icon";
+      icon.setAttribute("aria-hidden", "true");
+      icon.textContent = kind === "image" ? "🖼" : "📄";
+      pill.appendChild(icon);
+      const name = document.createElement("span");
+      name.className = "msg-user-attach-name";
+      name.textContent =
+        (a && a.filename) || (kind === "image" ? "image" : "document");
+      pill.appendChild(name);
+      pills.appendChild(pill);
+    });
+    el.appendChild(pills);
+    _scheduleScroll();
+    return el;
+  }
+
   // Metacognitive reminder bubble (user-channel correction / denial /
   // resume / start / completion AND tool-channel tool_error / repeat).
   // Mirrors Pane.prototype.addUserReminder / addToolReminder in the
@@ -459,10 +492,10 @@
     };
   }
 
-  function appendToolResult(name, callId, output, isError) {
+  function appendToolResult(name, callId, output, isError, opts) {
     if (callId && toolRows.has(callId)) {
       const entry = toolRows.get(callId);
-      _appendResultToRow(entry.row, output, isError);
+      _appendResultToRow(entry.row, output, isError, opts);
       // Result blocks grow scrollHeight; without this the user pinned
       // at the bottom loses their pin when the row inflates.  appendMsg
       // already routes through _scheduleScroll on the legacy path; this
@@ -655,6 +688,11 @@
     const row = document.createElement("div");
     row.className = "coord-tool-row";
     if (item.call_id) row.dataset.callId = item.call_id;
+    // Stamp the function name so the metacog dim rule for memory /
+    // recall (coordinator.css `.coord-tool-row[data-tool-name=...]`)
+    // has something to match.  Cheap; lets long workstreams with
+    // heavy memory usage stay readable without per-call JS work.
+    if (item.func_name) row.dataset.toolName = item.func_name;
 
     const callLine = document.createElement("div");
     callLine.className = "coord-tool-row-call";
@@ -683,6 +721,35 @@
     return row;
   }
 
+  // Attach an output-guard finding chip to a specific .coord-tool-row.
+  // Idempotent — replacing an existing chip in place lets the live
+  // SSE handler upgrade severity without stacking duplicates when a
+  // late event arrives after replay seeded an initial chip.
+  function _attachOutputWarningChip(row, oa) {
+    if (!row || !oa) return;
+    const risk = String(oa.risk_level || "medium");
+    const flags = oa.flags || [];
+    const existing = row.querySelector(".coord-tool-row-warning");
+    const chip = existing || document.createElement("div");
+    chip.className = "coord-tool-row-warning coord-tool-row-warning--" + risk;
+    chip.setAttribute("role", "status");
+    chip.textContent = "";
+    const label = document.createElement("span");
+    label.className = "coord-tool-row-warning-label";
+    label.textContent = "⚠ " + risk.toUpperCase();
+    chip.appendChild(label);
+    if (flags.length) {
+      chip.appendChild(document.createTextNode(" " + flags.join(", ")));
+    }
+    if (oa.redacted) {
+      const redacted = document.createElement("span");
+      redacted.className = "coord-tool-row-warning-redacted";
+      redacted.textContent = " (credentials redacted)";
+      chip.appendChild(redacted);
+    }
+    if (!existing) row.appendChild(chip);
+  }
+
   // Stable signature for a verdict — used to skip the DOM rebuild when
   // an SSE replay (or duplicate intent_verdict event) carries the same
   // verdict body we already painted.  Any field change (rec/risk/conf
@@ -795,10 +862,14 @@
     line.appendChild(chip);
   }
 
-  function _appendResultToRow(row, output, isError) {
+  function _appendResultToRow(row, output, isError, opts) {
     if (!row) return;
     const existing = row.querySelector(".coord-tool-row-result");
     if (existing) existing.remove();
+    // Re-fires (cancel + rerun, error + retry) clear any prior
+    // truncation pill so it doesn't stack on the new result.
+    const existingTrunc = row.querySelector(".coord-tool-truncated");
+    if (existingTrunc) existingTrunc.remove();
     if (isError) {
       row.classList.add("error");
       // Lift the row's error onto the enclosing batch so the left
@@ -854,6 +925,19 @@
     body.textContent = pretty;
     block.appendChild(body);
     row.appendChild(block);
+    // Storage-truncation indicator — sibling pill (not text inside
+    // the result body) so renderers / parsers / copy-as-text paths
+    // see the unmodified output.  Same convention as interactive's
+    // .tool-output-truncated; styled by .coord-tool-truncated in
+    // coordinator.css.
+    if (opts && opts.truncated) {
+      const pill = document.createElement("span");
+      pill.className = "coord-tool-truncated";
+      pill.textContent = "… truncated in storage";
+      pill.title =
+        "Full tool output was sent to the model live; only the first 10000 characters are persisted to the conversation row.";
+      row.appendChild(pill);
+    }
   }
 
   function _makeActionButton(label, role, kbdHint, ariaLabel) {
@@ -1463,7 +1547,13 @@
       queuedEl = queue.addQueuedMessage(displayText, priority);
     } else {
       setBusy(true);
-      appendText("user", trimmed, { label: "you" });
+      // snap.attachments carries the chip metadata (kind + filename)
+      // for every stable chip the composer holds; pass it through so
+      // the optimistic user bubble shows the same pill cluster the
+      // history-replay path renders below.
+      appendUserMessageWithAttachments(trimmed, snap.attachments, {
+        label: "you",
+      });
     }
     composer.clear();
 
@@ -1891,14 +1981,27 @@
         );
         break;
       case "output_warning":
-        appendText(
-          "error",
-          "[output guard] " +
-            (ev.risk_level || "?") +
-            ": " +
-            (ev.flags || []).join(","),
-          { label: "warning" },
-        );
+        // Anchor the finding to the specific .coord-tool-row that
+        // tripped the guard so the operator reads call → finding
+        // adjacency on both live and replay surfaces.  Falls back
+        // to a chat line only when the call_id no longer maps to a
+        // row (e.g. event arrived after the row was evicted).
+        if (ev.call_id && toolRows.has(ev.call_id)) {
+          _attachOutputWarningChip(toolRows.get(ev.call_id).row, {
+            risk_level: ev.risk_level,
+            flags: ev.flags,
+            redacted: ev.redacted,
+          });
+        } else {
+          appendText(
+            "info",
+            "[output guard] " +
+              (ev.risk_level || "?") +
+              ": " +
+              (ev.flags || []).join(","),
+            { label: "warning" },
+          );
+        }
         break;
       case "error":
         appendText("error", ev.message || "(unknown error)", {
@@ -3759,6 +3862,33 @@
               parsedArgs,
               argsRaw,
             );
+            // Server attaches the persisted intent_verdict to each
+            // tc on /history (newest-wins per call_id; LLM upgrade
+            // beats heuristic when both exist).  Stamp on the item
+            // under the field name the render path already consumes
+            // (judge_verdict for LLM tier, heuristic_verdict
+            // otherwise) so the verdict pill paints on history rows
+            // without a render-path fork.  Also seed the
+            // judgeVerdicts cache so a later live SSE event for the
+            // same call_id reads "already painted" and skips the
+            // rebuild.
+            if (tc && tc.verdict) {
+              if (tc.verdict.tier === "llm") {
+                item.judge_verdict = tc.verdict;
+              } else {
+                item.heuristic_verdict = tc.verdict;
+              }
+              if (callId) _cacheJudgeVerdict(callId, tc.verdict);
+            }
+            // Output-guard finding — surface as the same
+            // "[output guard] ..." chat line the live handler emits
+            // (case "output_warning" above).  Stamp on the item so
+            // the post-batch loop below can read + emit; rendering
+            // anchored next to the call gives the operator the same
+            // adjacency they'd see live.
+            if (tc && tc.output_assessment) {
+              item.output_assessment = tc.output_assessment;
+            }
             // needs_approval is unknown at replay time (the
             // assistant.tool_calls history payload doesn't persist
             // the bit).  Leave it unset; the upgrade-in-place path
@@ -3786,17 +3916,32 @@
           } else {
             appendToolBatch(items, { resolved: { approved: true } });
           }
+          // Output-guard findings — render each one as a chip
+          // anchored to the .coord-tool-row that tripped the guard
+          // rather than a generic "[output guard]" chat line.
+          // Anchored placement preserves per-call adjacency on
+          // multi-tool batches (live + replay) and the chip's
+          // severity styling makes the visual weight match the
+          // verdict pill on the same row.
+          for (let oi = 0; oi < items.length; oi++) {
+            const oa = items[oi].output_assessment;
+            if (!oa || !oa.risk_level || oa.risk_level === "none") continue;
+            const cid = items[oi].call_id || "";
+            if (!cid) continue;
+            const entry = toolRows.get(cid);
+            if (!entry || !entry.row) continue;
+            _attachOutputWarningChip(entry.row, oa);
+          }
         }
 
         // User messages with attachments arrive as multipart list
         // content (text + image_url/document parts) and may carry an
-        // ``_attachments_meta`` side-channel with display metadata.
-        // Extract the text portion + attachment count for a readable
-        // history replay; chip-rendering parity with the interactive
-        // pane is deferred (the coord dashboard is diagnostic-leaning
-        // — primary use is monitoring, not authoring).
+        // ``_attachments_meta`` side-channel with display metadata
+        // (kind + filename + mime_type).  Extract the text portion and
+        // build a structured attachment list so the user bubble can
+        // render the same pill cluster the interactive pane shows.
         let content;
-        let attachmentCount = 0;
+        const userAttachments = [];
         if (typeof m.content === "string") {
           content = m.content;
         } else if (Array.isArray(m.content)) {
@@ -3805,8 +3950,14 @@
             if (!part || typeof part !== "object") continue;
             if (part.type === "text") {
               textParts.push(String(part.text || ""));
-            } else if (part.type === "image_url" || part.type === "document") {
-              attachmentCount += 1;
+            } else if (part.type === "image_url") {
+              userAttachments.push({ kind: "image", filename: "" });
+            } else if (part.type === "document") {
+              const doc = part.document || {};
+              userAttachments.push({
+                kind: "text",
+                filename: String(doc.name || ""),
+              });
             }
           }
           content = textParts.join("\n");
@@ -3816,19 +3967,18 @@
         const meta = Array.isArray(m._attachments_meta)
           ? m._attachments_meta
           : null;
-        if (meta && meta.length > attachmentCount) {
-          // Prefer the side-channel count when present — it covers
-          // attachments whose multipart parts couldn't be reconstructed.
-          attachmentCount = meta.length;
-        }
-        if (attachmentCount > 0) {
-          const noun = attachmentCount === 1 ? "attachment" : "attachments";
-          content =
-            (content ? content + "\n\n" : "") +
-            "📎 " +
-            attachmentCount +
-            " " +
-            noun;
+        if (meta && meta.length) {
+          // Side-channel is authoritative — it carries filenames that
+          // image_url data URIs can't express, and covers attachments
+          // whose multipart parts couldn't be reconstructed.
+          userAttachments.length = 0;
+          for (const a of meta) {
+            if (!a || typeof a !== "object") continue;
+            userAttachments.push({
+              kind: String(a.kind || "other"),
+              filename: String(a.filename || ""),
+            });
+          }
         }
         if (role === "tool") {
           // Tool result content can legitimately be empty (e.g. a
@@ -3846,7 +3996,14 @@
           const toolName =
             (callId && toolNameByCallId.get(callId)) || m.tool_name || "tool";
           const isError = callOutcomes.get(callId) === "error";
-          appendToolResult(toolName, callId, content || "", isError);
+          // Storage truncation surfaces as a sibling pill next to
+          // the result (see _appendResultToRow's opts.truncated
+          // branch) rather than as text inside the result body — a
+          // future "best-effort JSON repair" pass would otherwise
+          // need to strip a marker string before parsing.
+          appendToolResult(toolName, callId, content || "", isError, {
+            truncated: !!m.truncated,
+          });
           // Tool-channel metacog reminders ride the same _reminders
           // side-channel as the user channel; surface as a themed
           // bubble below the .coord-tool-batch construct.
@@ -3877,20 +4034,26 @@
             body.textContent = content;
           }
         } else {
-          if (!content) return;
           // user / reasoning / system / other roles render as plain
           // text on history replay — matches the live-streaming paths
           // (appendReasoningToken uses textContent; user/system are
-          // typed verbatim and don't carry markdown structure).
-          appendText(role, content, { label: role });
-          // User-channel metacog reminders attach to the just-appended
-          // user bubble (the most recent .msg.user in messagesEl).
-          if (
-            role === "user" &&
-            Array.isArray(m.reminders) &&
-            m.reminders.length
-          ) {
-            appendUserReminderLive(m.reminders);
+          // typed verbatim and don't carry markdown structure).  User
+          // bubbles additionally render the pill strip beneath the
+          // text when the message carried attachments — even when the
+          // text portion is empty (image-only sends).
+          if (role === "user") {
+            if (!content && userAttachments.length === 0) return;
+            appendUserMessageWithAttachments(content, userAttachments, {
+              label: role,
+            });
+            // User-channel metacog reminders attach to the just-appended
+            // user bubble (the most recent .msg.user in messagesEl).
+            if (Array.isArray(m.reminders) && m.reminders.length) {
+              appendUserReminderLive(m.reminders);
+            }
+          } else {
+            if (!content) return;
+            appendText(role, content, { label: role });
           }
         }
       });

turnstone/console/static/governance.js

@@ -2791,7 +2791,8 @@ var _eogpTriggerEl = null;
 function switchJudgeSection(section) {
   var sections = document.querySelectorAll(".judge-section");
   for (var i = 0; i < sections.length; i++) sections[i].style.display = "none";
-  var btns = document.querySelectorAll(".judge-section-btn");
+  var switcher = document.querySelector("#admin-judge .admin-subtab-switcher");
+  var btns = switcher ? switcher.querySelectorAll(".admin-subtab-btn") : [];
   for (var i = 0; i < btns.length; i++) {
     var isActive = btns[i].getAttribute("data-section") === section;
     btns[i].classList.toggle("active", isActive);
@@ -2804,15 +2805,15 @@ function switchJudgeSection(section) {
 
 // Arrow key navigation for judge sub-section tabs
 (function () {
-  var switcher = document.querySelector(".judge-section-switcher");
+  var switcher = document.querySelector("#admin-judge .admin-subtab-switcher");
   if (!switcher) return;
   switcher.addEventListener("keydown", function (e) {
     if (e.key !== "ArrowLeft" && e.key !== "ArrowRight") return;
-    var btns = switcher.querySelectorAll(".judge-section-btn");
+    var btns = switcher.querySelectorAll(".admin-subtab-btn");
     var secs = [];
     for (var i = 0; i < btns.length; i++)
       secs.push(btns[i].getAttribute("data-section"));
-    var current = switcher.querySelector(".judge-section-btn.active");
+    var current = switcher.querySelector(".admin-subtab-btn.active");
     var idx = secs.indexOf(current ? current.getAttribute("data-section") : "");
     if (e.key === "ArrowRight") idx = (idx + 1) % secs.length;
     else idx = (idx - 1 + secs.length) % secs.length;
@@ -2874,6 +2875,7 @@ function renderJudgeSettings() {
   var html = "";
   for (var i = 0; i < _judgeSettings.length; i++) {
     var s = _judgeSettings[i];
+    if (s.key === "judge.model") continue;
     var shortKey = s.key.replace("judge.", "");
     var inputHtml = "";
     var currentVal = s.value;

turnstone/console/static/index.html

@@ -87,33 +87,16 @@
       <div id="view-home">
         <!-- Persistent "start a new coordinator task" composer.  Visibility is
          gated on the admin.coordinator permission (same rule the existing
-         +coordinator header button + modal use).  Shows a remediation
-         banner when the create endpoint would return 503 (no coordinator
-         model alias resolves). -->
+         +coordinator header button + modal use).  Submission errors surface
+         inline via #home-coord-error; the create endpoint falls back to
+         the registry default model when ``coordinator.model_alias`` is
+         unset, so no proactive readiness probe is needed. -->
         <section
           id="coord-composer-panel"
           class="home-panel"
           style="display: none"
           aria-label="Start a new orchestration task"
         >
-          <div
-            id="coord-composer-503"
-            class="home-composer-banner"
-            role="status"
-            style="display: none"
-          >
-            Coordinator subsystem not configured —
-            <a
-              href="#"
-              onclick="
-                showAdmin();
-                switchAdminTab('models');
-                return false;
-              "
-              >open Admin → Models</a
-            >
-            to set <code>coordinator.model_alias</code> or a registry default.
-          </div>
           <!-- Composer DOM is built by shared_static/composer.js into this
            mount — stacked layout (textarea above, options toggle +
            Start button below) with Name + Skill in an Options
@@ -122,9 +105,8 @@
           <div
             id="home-coord-error"
             class="home-composer-error"
-            role="alert"
-            aria-live="assertive"
-            style="display: none"
+            role="status"
+            aria-live="polite"
           ></div>
         </section>
 
@@ -883,13 +865,13 @@
 
               <!-- Sub-panel switcher -->
               <div
-                class="judge-section-switcher"
+                class="admin-subtab-switcher"
                 role="tablist"
                 aria-label="Judge sections"
               >
                 <button
                   id="judge-tab-settings"
-                  class="judge-section-btn active"
+                  class="admin-subtab-btn active"
                   role="tab"
                   aria-selected="true"
                   aria-controls="judge-settings-section"
@@ -901,7 +883,7 @@
                 </button>
                 <button
                   id="judge-tab-heuristic"
-                  class="judge-section-btn"
+                  class="admin-subtab-btn"
                   role="tab"
                   aria-selected="false"
                   aria-controls="judge-heuristic-section"
@@ -913,7 +895,7 @@
                 </button>
                 <button
                   id="judge-tab-output-guard"
-                  class="judge-section-btn"
+                  class="admin-subtab-btn"
                   role="tab"
                   aria-selected="false"
                   aria-controls="judge-output-guard-section"
@@ -1796,37 +1778,106 @@
               style="display: none"
             >
               <div class="admin-toolbar">
-                <span class="section-header">Models</span>
-                <button
-                  id="model-sync-btn"
-                  class="admin-action-btn admin-action-btn-ghost"
-                  onclick="reloadModelNodes()"
-                  title="Push model config to all cluster nodes"
-                >
-                  Sync to Nodes
-                </button>
-                <button
-                  class="admin-action-btn"
-                  onclick="showCreateModelModal()"
-                >
-                  + Add Model
-                </button>
-              </div>
-              <div class="admin-colheaders models-grid" aria-hidden="true">
-                <span class="admin-col">ALIAS</span>
-                <span class="admin-col">MODEL</span>
-                <span class="admin-col">PROVIDER</span>
-                <span class="admin-col">CTX WINDOW</span>
-                <span class="admin-col">STATUS</span>
-                <span class="admin-col">ACTIONS</span>
+                <span class="section-header" style="margin: 0">MODELS</span>
               </div>
+
+              <!-- Sub-panel switcher -->
               <div
-                id="admin-models-table"
-                role="list"
-                aria-label="Model definitions"
-                aria-live="polite"
+                class="admin-subtab-switcher"
+                role="tablist"
+                aria-label="Models sections"
               >
-                <div class="dashboard-empty">Loading...</div>
+                <button
+                  id="models-tab-list"
+                  class="admin-subtab-btn active"
+                  role="tab"
+                  aria-selected="true"
+                  aria-controls="models-list-section"
+                  tabindex="0"
+                  data-section="models-list"
+                  onclick="switchModelsSection('models-list')"
+                >
+                  Definitions
+                </button>
+                <button
+                  id="models-tab-roles"
+                  class="admin-subtab-btn"
+                  role="tab"
+                  aria-selected="false"
+                  aria-controls="models-roles-section"
+                  tabindex="-1"
+                  data-section="models-roles"
+                  onclick="switchModelsSection('models-roles')"
+                >
+                  Roles
+                </button>
+              </div>
+
+              <!-- Models list section -->
+              <div
+                id="models-list-section"
+                class="models-section"
+                role="tabpanel"
+                aria-labelledby="models-tab-list"
+              >
+                <div class="admin-toolbar" style="margin-bottom: 12px">
+                  <span style="font-size: 13px; color: var(--fg-dim)"
+                    >Model definitions used by sessions across the cluster</span
+                  >
+                  <button
+                    id="model-sync-btn"
+                    class="admin-action-btn admin-action-btn-ghost"
+                    onclick="reloadModelNodes()"
+                    title="Push model config to all cluster nodes"
+                  >
+                    Sync to Nodes
+                  </button>
+                  <button
+                    class="admin-action-btn"
+                    onclick="showCreateModelModal()"
+                  >
+                    + Add Model
+                  </button>
+                </div>
+                <div class="admin-colheaders models-grid" aria-hidden="true">
+                  <span class="admin-col">ALIAS</span>
+                  <span class="admin-col">MODEL</span>
+                  <span class="admin-col">PROVIDER</span>
+                  <span class="admin-col">CTX WINDOW</span>
+                  <span class="admin-col">STATUS</span>
+                  <span class="admin-col">ACTIONS</span>
+                </div>
+                <div
+                  id="admin-models-table"
+                  role="list"
+                  aria-label="Model definitions"
+                  aria-live="polite"
+                >
+                  <div class="dashboard-empty">Loading...</div>
+                </div>
+              </div>
+
+              <!-- Roles section (Coordinator / Judge / future perception roles) -->
+              <div
+                id="models-roles-section"
+                class="models-section"
+                role="tabpanel"
+                aria-labelledby="models-tab-roles"
+                style="display: none"
+              >
+                <div style="margin-bottom: 12px">
+                  <span style="font-size: 13px; color: var(--fg-dim)"
+                    >Per-role model assignments. Empty = use the default
+                    model.</span
+                  >
+                </div>
+                <div
+                  id="admin-models-roles-container"
+                  aria-live="polite"
+                  style="max-width: 720px"
+                >
+                  <div class="dashboard-empty">Loading&hellip;</div>
+                </div>
               </div>
             </div>
 
@@ -3869,6 +3920,17 @@
             <option value="llama.cpp">llama.cpp</option>
             <option value="openai-compatible">Other OpenAI-compatible</option>
           </select>
+          <label for="model-api-surface"
+            >API Surface
+            <span style="font-weight: 400; text-transform: none"
+              >(Chat Completions vs Responses)</span
+            ></label
+          >
+          <select id="model-api-surface">
+            <option value="">Inherit (Chat Completions)</option>
+            <option value="chat">Chat Completions (pinned)</option>
+            <option value="responses">Responses API</option>
+          </select>
           <label for="model-thinking-mode"
             >Thinking Mode
             <span style="font-weight: 400; text-transform: none"

turnstone/console/static/style.css

@@ -88,27 +88,45 @@
   text-transform: uppercase;
 }
 
-.home-composer-banner {
-  background: var(--bg-surface);
-  border: 1px solid var(--yellow);
-  border-left-width: 3px;
-  border-radius: var(--radius-sm);
-  padding: 8px 10px;
-  color: var(--fg-bright);
-  font-size: 12px;
-}
-.home-composer-banner a {
-  color: var(--accent);
-  text-decoration: underline;
-  text-decoration-thickness: 2px;
-}
-
 .home-composer-error {
+  /* Always rendered (no display toggle in JS) so toggling validation
+     messages doesn't reflow the active-coordinators list below.  The
+     min-height holds a single 12px line + padding so an empty state
+     reserves the same space the rendered error will occupy. */
+  min-height: 20px;
   color: var(--red);
   font-size: 12px;
   padding: 4px 2px 0;
 }
 
+/* Drag-over feedback for the home coord composer — wired by Composer's
+   dragDrop option (dropClass: home-coord-drop) on the composer mount.
+   Mirrors the dashed-outline affordance on coord-main so users see the
+   same drop visual the in-coord composer uses. */
+#home-coord-composer-mount {
+  position: relative;
+}
+#home-coord-composer-mount.home-coord-drop {
+  outline: 2px dashed var(--accent);
+  outline-offset: -6px;
+  border-radius: var(--radius);
+}
+
+/* Cap chip filename width inside the home composer so a long-name
+   attachment doesn't push the strip wider than the textarea or wrap
+   unpredictably across multiple rows.  shared_static/chat.css defines
+   .composer-chip / .composer-chip-size / .composer-chip-remove but
+   leaves .composer-chip-name unstyled — the span just inherits the
+   .composer-chip font with no width cap, fine for chat-pane width but
+   too loose for the narrower home column.  Apply the same ellipsis
+   cap the .msg-user-attach-pill rule uses on the user bubble. */
+#home-coord-composer-mount .composer-chip-name {
+  max-width: 200px;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
 .home-section {
   display: flex;
   flex-direction: column;
@@ -2317,15 +2335,15 @@ textarea.skill-content-area {
 }
 
 /* ==========================================================================
-   Judge sub-section tabs
+   Admin sub-section tabs (used by Judge + Models tabs)
    ========================================================================== */
-.judge-section-switcher {
+.admin-subtab-switcher {
   display: flex;
   gap: 8px;
   margin: 12px 0 16px;
   border-bottom: 1px solid var(--border-strong);
 }
-.judge-section-btn {
+.admin-subtab-btn {
   padding: 6px 14px;
   background: none;
   border: none;
@@ -2338,14 +2356,14 @@ textarea.skill-content-area {
     color 0.15s,
     border-color 0.15s;
 }
-.judge-section-btn:hover {
+.admin-subtab-btn:hover {
   color: var(--fg);
 }
-.judge-section-btn.active {
+.admin-subtab-btn.active {
   border-bottom-color: var(--accent);
   color: var(--fg);
 }
-.judge-section-btn:focus-visible {
+.admin-subtab-btn:focus-visible {
   outline: 2px solid var(--accent);
   outline-offset: -2px;
 }
@@ -3808,6 +3826,64 @@ textarea.skill-content-area {
   letter-spacing: 0.02em;
 }
 
+/* Models → Roles sub-tab rows */
+.model-role-row {
+  padding: 12px 0;
+  border-bottom: 1px solid var(--border-strong);
+}
+.model-role-row:last-child {
+  border-bottom: none;
+}
+.model-role-head {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  margin-bottom: 4px;
+}
+.model-role-label {
+  font-size: 13px;
+  font-weight: 600;
+  color: var(--fg);
+}
+.model-role-desc {
+  font-size: 11px;
+  color: var(--fg-dim);
+  margin-bottom: 8px;
+}
+.model-role-controls {
+  display: grid;
+  grid-template-columns: minmax(240px, 1fr) minmax(180px, auto);
+  column-gap: 16px;
+  row-gap: 8px;
+  align-items: end;
+}
+.model-role-control {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  min-width: 0;
+}
+.model-role-control-label {
+  font-size: 10px;
+  text-transform: uppercase;
+  letter-spacing: 0.06em;
+  color: var(--fg-dim);
+}
+.model-role-control select {
+  width: 100%;
+  padding: 4px 8px;
+  background: var(--bg);
+  border: 1px solid var(--border-strong);
+  color: var(--fg);
+  border-radius: 3px;
+  font-family: var(--font-ui);
+  font-size: 12px;
+}
+.model-role-control select:focus-visible {
+  outline: 2px solid var(--accent);
+  outline-offset: 1px;
+}
+
 /* Modal section divider for field groups */
 .modal-section-divider {
   font-family: var(--font-ui);
@@ -3855,7 +3931,7 @@ textarea.skill-content-area {
   .admin-btn-danger,
   .admin-btn-caution,
   .admin-btn-action,
-  .judge-section-btn {
+  .admin-subtab-btn {
     transition: none;
   }
   .settings-toggle-slider,

turnstone/core/history_decoration.py

@@ -0,0 +1,205 @@
+"""Shared history-replay decoration helpers.
+
+Both surfaces that build a history wire payload — interactive's SSE
+``_build_history`` and the lifted ``make_history_handler`` REST
+endpoint — need the same audit-trail data attached to each
+``tool_calls`` entry: the persisted intent verdict (``intent_verdicts``
+table) and the output-guard assessment (``output_assessments`` table).
+
+Centralising the lookup + decoration here keeps the two surfaces from
+drifting on which fields ship to the client and how they're shaped.
+The shared helpers also let us project only the fields the UI actually
+renders, dropping redundant ones (``call_id``/``func_name`` already
+carried on ``tc.id``/``tc.name``) so the wire payload stays tight.
+
+All functions are pure I/O or pure transforms — safe to call from
+either an async caller (via ``asyncio.to_thread``) or a sync hook.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from turnstone.core.log import get_logger
+
+log = get_logger(__name__)
+
+# Tool results are clamped at this length per row at storage time
+# (see ``session.py``'s ``store_text = raw_output[:TOOL_RESULT_STORAGE_CAP]``).
+# Keeping the constant here lets the truncation flag detection in
+# ``decorate_history_messages`` stay in sync without a magic number
+# duplicated across server.py / session.py.
+#
+# Raised from 2000 → 10000 because a 2000-char clip routinely cut
+# the body of a single grep / file read mid-line, leaving the
+# historical record useless for retrospective debugging.  FTS5
+# index + row size grow proportionally; the per-tool upper bound is
+# still bounded upstream by ``_truncate_output``'s context-budget
+# clamp (so a single huge result can't blow past the live context
+# window).
+TOOL_RESULT_STORAGE_CAP = 10000
+
+
+def load_verdict_indexes(
+    ws_id: str,
+) -> tuple[dict[str, dict[str, Any]], dict[str, dict[str, Any]]]:
+    """Bulk-load intent verdicts and output assessments for a workstream.
+
+    Returns ``(verdicts_by_call_id, assessments_by_call_id)``.  Both
+    tables are indexed by ws_id so the queries are O(rows-for-ws); the
+    DESC ordering plus first-seen-wins dedupe leaves the newest
+    verdict per call_id (LLM upgrade beats heuristic when both exist).
+
+    Pure storage I/O — safe to run in ``asyncio.to_thread`` from an
+    async caller.  Returns empty dicts when storage is unavailable or
+    the lookup raises (best-effort: replay must never block on
+    audit-trail decoration).
+    """
+    verdicts_by_call_id: dict[str, dict[str, Any]] = {}
+    assessments_by_call_id: dict[str, dict[str, Any]] = {}
+    if not ws_id:
+        return verdicts_by_call_id, assessments_by_call_id
+    try:
+        from turnstone.core.storage._registry import get_storage
+
+        storage = get_storage()
+        if storage is None:
+            return verdicts_by_call_id, assessments_by_call_id
+        for v in storage.list_intent_verdicts(ws_id=ws_id, limit=10000):
+            cid = v.get("call_id") or ""
+            if cid and cid not in verdicts_by_call_id:
+                verdicts_by_call_id[cid] = v
+        for a in storage.list_output_assessments(ws_id=ws_id, limit=10000):
+            cid = a.get("call_id") or ""
+            if cid and cid not in assessments_by_call_id:
+                assessments_by_call_id[cid] = a
+    except Exception:
+        # Missing storage / migration drift / driver error must not
+        # block replay — degrade to an unannotated history.
+        log.debug(
+            "verdict/assessment lookup failed; replay continues unannotated",
+            exc_info=True,
+        )
+    return verdicts_by_call_id, assessments_by_call_id
+
+
+def build_verdict_payload(vrow: dict[str, Any]) -> dict[str, Any] | None:
+    """Project a stored ``intent_verdicts`` row into the wire shape.
+
+    Returns ``None`` when the verdict is the unflagged baseline
+    (``risk_level == "none"``) — the client's ``renderVerdictBadge``
+    helper would suppress those anyway, so skipping at the wire layer
+    keeps the payload tight on long workstreams.
+
+    Drops ``call_id`` and ``func_name`` from the wire payload — they're
+    already carried on the parent ``tc.id`` / ``tc.name`` fields.
+    Ships ``reasoning`` for either tier when the row has non-empty
+    prose (heuristic rules in this project DO write meaningful
+    rationales — e.g. ``policy.py`` emits structured reasoning per
+    matched pattern). ``judge_model`` rides through so the batch tier
+    badge can render ``⚖ llm:claude-haiku-4`` on history-only batches
+    rather than the bare ``⚖ llm`` label.
+    """
+    if (vrow.get("risk_level") or "none") == "none":
+        return None
+    payload: dict[str, Any] = {
+        "risk_level": vrow.get("risk_level", "medium"),
+        "recommendation": vrow.get("recommendation", "review"),
+        "confidence": vrow.get("confidence", 0.0),
+        "intent_summary": vrow.get("intent_summary", ""),
+        "tier": vrow.get("tier", "heuristic"),
+    }
+    if vrow.get("reasoning"):
+        payload["reasoning"] = vrow.get("reasoning", "")
+    judge_model = vrow.get("judge_model") or ""
+    if judge_model:
+        payload["judge_model"] = judge_model
+    return payload
+
+
+def build_output_assessment_payload(arow: dict[str, Any]) -> dict[str, Any] | None:
+    """Project a stored ``output_assessments`` row into the wire shape.
+
+    Returns ``None`` when the assessment is the unflagged baseline
+    (``risk_level == "none"``) — same skip-on-clean pattern as
+    :func:`build_verdict_payload`.
+
+    Decodes ``flags`` from its JSON string form here so the client
+    never has to parse twice.  Falls back to an empty list on bad JSON
+    rather than raising — the rest of the assessment is still useful.
+    """
+    if (arow.get("risk_level") or "none") == "none":
+        return None
+    flags_raw = arow.get("flags") or "[]"
+    try:
+        flags = json.loads(flags_raw) if isinstance(flags_raw, str) else flags_raw
+    except (ValueError, TypeError):
+        flags = []
+    return {
+        "risk_level": arow.get("risk_level", "none"),
+        "flags": flags if isinstance(flags, list) else [],
+        "redacted": bool(arow.get("redacted", 0)),
+    }
+
+
+def decorate_tool_call(
+    tc: dict[str, Any],
+    verdicts_by_call_id: dict[str, dict[str, Any]],
+    assessments_by_call_id: dict[str, dict[str, Any]],
+) -> None:
+    """Mutate ``tc`` in place, attaching ``verdict`` / ``output_assessment``.
+
+    Works on either tool_call shape:
+    - OpenAI format (``{id, function: {name, arguments}}``) — used by
+      ``/history`` REST.
+    - Flattened format (``{id, name, arguments}``) — used by SSE replay.
+
+    Both carry ``id`` at the top level, which is the only field this
+    helper reads.  No-ops cleanly when the call_id has no matching
+    row (unflagged tools stay clean).
+    """
+    call_id = tc.get("id", "") or ""
+    if not call_id:
+        return
+    vrow = verdicts_by_call_id.get(call_id)
+    if vrow is not None:
+        verdict = build_verdict_payload(vrow)
+        if verdict is not None:
+            tc["verdict"] = verdict
+    arow = assessments_by_call_id.get(call_id)
+    if arow is not None:
+        assessment = build_output_assessment_payload(arow)
+        if assessment is not None:
+            tc["output_assessment"] = assessment
+
+
+def decorate_history_messages(
+    messages: list[dict[str, Any]],
+    verdicts_by_call_id: dict[str, dict[str, Any]],
+    assessments_by_call_id: dict[str, dict[str, Any]],
+) -> None:
+    """Mutate a list of OpenAI-format messages, decorating tool_calls.
+
+    Used by the ``/history`` REST endpoint after ``load_messages``
+    returns.  For each assistant message with ``tool_calls``, runs
+    :func:`decorate_tool_call` on every entry.  For each tool message
+    whose content hits the storage cap, sets ``truncated: True`` so
+    the client can render the "… truncated in storage" pill.
+
+    Pure transform — no I/O.  Async callers should pre-load the
+    indexes via :func:`load_verdict_indexes` (in ``to_thread``) and
+    pass them in.
+    """
+    for msg in messages:
+        role = msg.get("role")
+        if role == "assistant":
+            tcs = msg.get("tool_calls")
+            if isinstance(tcs, list):
+                for tc in tcs:
+                    if isinstance(tc, dict):
+                        decorate_tool_call(tc, verdicts_by_call_id, assessments_by_call_id)
+        elif role == "tool":
+            content = msg.get("content")
+            if isinstance(content, str) and len(content) >= TOOL_RESULT_STORAGE_CAP:
+                msg["truncated"] = True

turnstone/core/memory.py

@@ -757,6 +757,37 @@ def search_structured_memories(
         return []
 
 
+def list_visible_structured_memories(
+    scopes: list[tuple[str, str]],
+    mem_type: str = "",
+    limit: int = 100,
+) -> list[dict[str, str]]:
+    """Single-query union across visible (scope, scope_id) pairs."""
+    try:
+        return get_storage().list_visible_structured_memories(
+            scopes, mem_type=mem_type, limit=limit
+        )
+    except Exception:
+        log.warning("Failed to list visible structured memories", exc_info=True)
+        return []
+
+
+def search_visible_structured_memories(
+    query: str,
+    scopes: list[tuple[str, str]],
+    mem_type: str = "",
+    limit: int = 20,
+) -> list[dict[str, str]]:
+    """OR-of-terms search joined with a single visibility OR-group."""
+    try:
+        return get_storage().search_visible_structured_memories(
+            query, scopes, mem_type=mem_type, limit=limit
+        )
+    except Exception:
+        log.warning("Failed to search visible structured memories", exc_info=True)
+        return []
+
+
 def touch_structured_memories(keys: list[tuple[str, str, str]]) -> int:
     """Batch-touch memories (bump last_accessed, increment access_count).
 

turnstone/core/metrics.py

@@ -32,6 +32,9 @@ class MetricsCollector:
         # counters (continued)
         self._ratelimit_rejects: int = 0  # counter: total 429 responses
         self._evictions: int = 0  # counter: workstreams evicted
+        # node_models publish (heartbeat-loop refresh of node_metadata.models)
+        self._node_models_publish_written: int = 0
+        self._node_models_publish_skipped: int = 0
         # judge metrics
         self._judge_verdicts: dict[tuple[str, str], int] = defaultdict(int)
         self._judge_latency: dict[str, Any] = {
@@ -104,6 +107,22 @@ class MetricsCollector:
         with self._lock:
             self._evictions += 1
 
+    def record_node_models_publish(self, *, written: bool) -> None:
+        """Record one heartbeat-loop attempt to refresh ``node_metadata.models``.
+
+        ``written=True`` means the projected payload differed from the
+        cached one and we ran an UPSERT.  ``written=False`` means the
+        cache short-circuited the call.  In a stable cluster the
+        skipped:written ratio runs ~100:1 — a sustained drop in that
+        ratio is the signal an operator wants (backend health flapping
+        or a runaway model-reload loop).
+        """
+        with self._lock:
+            if written:
+                self._node_models_publish_written += 1
+            else:
+                self._node_models_publish_skipped += 1
+
     def set_judge_enabled(self, enabled: bool) -> None:
         with self._lock:
             self._judge_enabled = enabled
@@ -170,6 +189,8 @@ class MetricsCollector:
             judge_verdicts = dict(self._judge_verdicts)
             judge_latency = dict(self._judge_latency)
             judge_enabled = self._judge_enabled
+            node_models_publish_written = self._node_models_publish_written
+            node_models_publish_skipped = self._node_models_publish_skipped
 
         # turnstone_build_info
         lines.append("# HELP turnstone_build_info Server version and model info")
@@ -277,6 +298,25 @@ class MetricsCollector:
             evictions,
         )
 
+        # turnstone_node_models_publish_total — split by outcome so an
+        # operator can compute hit-rate as
+        # ``rate(skipped) / (rate(skipped) + rate(written))``.  In a
+        # stable cluster this ratio sits very close to 1.0; sustained
+        # dips signal backend health flapping or reload churn.
+        lines.append(
+            "# HELP turnstone_node_models_publish_total "
+            "node_metadata.models refresh attempts by outcome"
+        )
+        lines.append("# TYPE turnstone_node_models_publish_total counter")
+        lines.append(
+            f'turnstone_node_models_publish_total{{outcome="written"}} '
+            f"{node_models_publish_written}"
+        )
+        lines.append(
+            f'turnstone_node_models_publish_total{{outcome="skipped"}} '
+            f"{node_models_publish_skipped}"
+        )
+
         # turnstone_judge_enabled
         gauge(
             "turnstone_judge_enabled",

turnstone/core/model_registry.py

@@ -44,6 +44,19 @@ class ModelConfig:
     server_compat: dict[str, Any] = field(default_factory=dict)
 
 
+def _api_surface_of(cfg: ModelConfig) -> str | None:
+    """Extract the operator-pinned api_surface from *cfg*, or ``None``.
+
+    Used both at provider-cache lookup time and at reload-eviction time so the
+    two sites stay in sync.  Returns ``None`` when the field is absent, blank,
+    or not a string — matching the "inherit provider default" semantics.
+    """
+    raw = cfg.server_compat.get("api_surface") if isinstance(cfg.server_compat, dict) else None
+    if isinstance(raw, str) and raw.strip():
+        return raw
+    return None
+
+
 # ---------------------------------------------------------------------------
 # Registry
 # ---------------------------------------------------------------------------
@@ -127,7 +140,9 @@ class ModelRegistry:
                 raise ValueError(f"Unknown model alias: {alias}")
             if alias not in self._providers:
                 cfg = self._models[alias]
-                self._providers[alias] = create_provider(cfg.provider)
+                self._providers[alias] = create_provider(
+                    cfg.provider, api_surface=_api_surface_of(cfg)
+                )
             return self._providers[alias]
 
     def get_config(self, alias: str) -> ModelConfig:
@@ -257,13 +272,18 @@ class ModelRegistry:
                     if hasattr(client, "close"):
                         client.close()
                     del self._clients[alias]
-            # Providers are keyed on alias but only depend on
-            # ``cfg.provider`` — drop only when the provider string
-            # changed or the alias was removed.
+            # Providers are keyed on alias and depend on (cfg.provider,
+            # cfg.server_compat["api_surface"]) — drop when either changes
+            # or the alias was removed.
             for alias in list(self._providers.keys()):
                 old_cfg = old_models.get(alias)
                 new_cfg = self._models.get(alias)
-                if new_cfg is None or old_cfg is None or old_cfg.provider != new_cfg.provider:
+                if (
+                    new_cfg is None
+                    or old_cfg is None
+                    or old_cfg.provider != new_cfg.provider
+                    or _api_surface_of(old_cfg) != _api_surface_of(new_cfg)
+                ):
                     del self._providers[alias]
 
     def shutdown(self) -> None:

turnstone/core/providers/__init__.py

@@ -33,7 +33,9 @@ __all__ = [
     "lookup_model_capabilities",
 ]
 
-# Singleton instances (stateless, safe to share)
+# Singleton instances (stateless, safe to share).  ``_openai_provider``
+# is reused for both cloud OpenAI and ``openai-compatible`` with
+# ``api_surface="responses"`` — see the ``create_provider`` docstring.
 _provider_lock = threading.Lock()
 _openai_provider = OpenAIResponsesProvider()
 _openai_compat_provider = OpenAIChatCompletionsProvider()
@@ -41,12 +43,45 @@ _anthropic_provider: LLMProvider | None = None
 _google_provider: LLMProvider | None = None
 
 
-def create_provider(provider_name: str) -> LLMProvider:
-    """Return a provider adapter for the given provider name. Thread-safe."""
+_VALID_API_SURFACES = ("chat", "responses")
+
+
+def create_provider(
+    provider_name: str,
+    *,
+    api_surface: str | None = None,
+) -> LLMProvider:
+    """Return a provider adapter for the given provider name. Thread-safe.
+
+    *api_surface* selects the OpenAI-compatible API surface for
+    ``provider_name="openai-compatible"``:
+
+    - ``"chat"`` (default) → Chat Completions (vLLM, llama.cpp, SGLang).
+    - ``"responses"``      → Responses API (commercial OpenAI-compat
+      endpoints like Mistral cloud, or local servers that expose the
+      Responses surface).
+
+    Ignored for non-OpenAI providers.  ``provider_name="openai"`` always
+    uses the Responses API regardless of *api_surface*.
+
+    Note: the ``OpenAIResponsesProvider`` singleton is reused for both
+    cloud OpenAI and ``openai-compatible`` + responses, so its
+    ``provider_name`` reports ``"openai"`` even when serving an
+    openai-compatible config.  Code that needs to distinguish the two
+    must read ``ModelConfig.provider`` and ``server_compat["api_surface"]``
+    rather than ``provider.provider_name``.
+    """
     global _anthropic_provider, _google_provider  # noqa: PLW0603
     if provider_name == "openai":
         return _openai_provider
     if provider_name == "openai-compatible":
+        normalised = (api_surface or "").strip().lower()
+        if normalised and normalised not in _VALID_API_SURFACES:
+            raise ValueError(
+                f"Unknown api_surface: {api_surface!r}. Supported: {', '.join(_VALID_API_SURFACES)}"
+            )
+        if normalised == "responses":
+            return _openai_provider
         return _openai_compat_provider
     if provider_name == "anthropic":
         with _provider_lock:

turnstone/core/server_compat.py

@@ -1,14 +1,20 @@
 """Server compatibility profiles for OpenAI-compatible backends.
 
 Different local model servers (vLLM, llama.cpp, SGLang) need different
-request shaping.  This module separates two concerns:
+request shaping.  This module separates three concerns:
 
 1. **Model capabilities** — ``thinking_mode`` and ``thinking_param`` are
    properties of the *model* (Gemma thinks, Llama doesn't).  These go
    into the ``capabilities`` dict and flow through ``ModelCapabilities``
    so the provider can act on them (just like Anthropic's thinking mode).
 
-2. **Server workarounds** — ``extra_body`` overrides like
+2. **API surface** — ``api_surface`` selects which OpenAI-compatible
+   API surface the provider talks to: ``"chat"`` (Chat Completions,
+   the default) or ``"responses"`` (Responses API, native reasoning).
+   Stored under ``server_compat`` because it's an endpoint property,
+   not a model property.
+
+3. **Server workarounds** — ``extra_body`` overrides like
    ``skip_special_tokens=false`` are properties of the *server* (vLLM
    bug workaround).  These stay in ``server_compat`` and get merged
    into the request's ``extra_body`` at call time.
@@ -82,6 +88,23 @@ _PROFILES: dict[str, dict[str, Any]] = {
             "server_type": "vllm",
         },
     },
+    "vllm-mistral-medium": {
+        # Mistral medium open-weights served by vLLM can deliver reasoning
+        # via either surface, but the trade-off is asymmetric:
+        #   * Chat Completions — tool calling works (``--tool-call-parser
+        #     mistral``); reasoning is enabled via the vLLM CLI
+        #     (``--reasoning-parser``) rather than per-request.
+        #   * Responses API   — reasoning effort is per-request and clean,
+        #     but as of vLLM 0.x the tool-call parser is not wired up on
+        #     this surface so tool calls leak as ``[TOOL_CALLS]`` text.
+        # We do **not** auto-suggest this profile from Detect; an operator
+        # who needs per-request effort and accepts the tool-calling
+        # limitation can pick "Responses API" manually in the admin UI.
+        "server_compat": {
+            "server_type": "vllm",
+            "api_surface": "responses",
+        },
+    },
     "vllm": {
         "server_compat": {
             "server_type": "vllm",
@@ -125,6 +148,9 @@ _VLLM_MODEL_PROFILES: list[tuple[str, str]] = [
     ("granite3", "vllm-granite-thinking"),
     ("deepseek-r1", "vllm-deepseek-thinking"),
     ("holo2", "vllm-holo-thinking"),
+    # Mistral medium intentionally omitted — see ``vllm-mistral-medium``
+    # profile docstring for the Chat-vs-Responses trade-off; operator
+    # picks manually rather than letting Detect auto-suggest Responses.
 ]
 
 # llama.cpp model-family → profile key mapping.
@@ -175,31 +201,39 @@ def suggest_profile(server_type: str, model_id: str) -> dict[str, Any]:
 
 
 def merge_server_compat(
-    base_chat_template_kwargs: dict[str, Any],
+    base_chat_template_kwargs: dict[str, Any] | None,
     server_compat: dict[str, Any],
 ) -> dict[str, Any]:
     """Build the ``extra_body`` dict by merging server compat into base kwargs.
 
-    *base_chat_template_kwargs* always contains at least ``reasoning_effort``.
-    *server_compat* comes from ``ModelConfig.server_compat``.
+    *base_chat_template_kwargs* is an explicit ``chat_template_kwargs`` dict
+    to seed the request with, or ``None``/empty to skip seeding.  Operator-
+    supplied entries in ``server_compat["extra_body"]["chat_template_kwargs"]``
+    are deep-merged on top.  Top-level ``extra_body`` keys (``skip_special_tokens``,
+    ``reasoning_format``, etc.) are forwarded as-is.
 
     Note: thinking-mode params (``enable_thinking``, ``thinking``) are **not**
     merged here — the provider handles those via ``ModelCapabilities``.
-    This function only merges server workarounds from ``extra_body``.
+    This function only merges what the operator stored in ``server_compat``.
 
     Returns the complete dict to pass as ``extra_body`` to the OpenAI client.
+    May be empty when there is nothing to send.
     """
-    extra: dict[str, Any] = {"chat_template_kwargs": dict(base_chat_template_kwargs)}
+    extra: dict[str, Any] = {}
+    if base_chat_template_kwargs:
+        extra["chat_template_kwargs"] = dict(base_chat_template_kwargs)
 
     # Merge top-level extra_body overrides (skip_special_tokens, etc.)
     compat_eb = server_compat.get("extra_body")
     if isinstance(compat_eb, dict):
         for key, value in compat_eb.items():
             if key == "chat_template_kwargs":
-                # Deep-merge: operator values in extra_body win over the
-                # base dict (which has reasoning_effort).  This lets
-                # operators intentionally extend chat_template_kwargs.
+                # Deep-merge with operator values winning so an operator
+                # can intentionally extend chat_template_kwargs (e.g. set
+                # ``reasoning_effort`` for gpt-oss-style local templates).
                 if isinstance(value, dict):
+                    if "chat_template_kwargs" not in extra:
+                        extra["chat_template_kwargs"] = {}
                     extra["chat_template_kwargs"].update(value)
                 continue
             extra[key] = value

turnstone/core/session.py

@@ -44,6 +44,7 @@ from turnstone.core.attachments import (
 )
 from turnstone.core.config import get_tavily_key
 from turnstone.core.edit import find_occurrences, pick_nearest
+from turnstone.core.history_decoration import TOOL_RESULT_STORAGE_CAP
 from turnstone.core.log import get_logger
 from turnstone.core.memory import (
     count_structured_memories,
@@ -57,6 +58,7 @@ from turnstone.core.memory import (
     list_default_skills,
     list_skills_by_activation,
     list_structured_memories,
+    list_visible_structured_memories,
     list_workstreams_with_history,
     load_messages,
     load_workstream_config,
@@ -70,6 +72,7 @@ from turnstone.core.memory import (
     search_history,
     search_history_recent,
     search_structured_memories,
+    search_visible_structured_memories,
     set_workstream_alias,
     unreserve_attachments,
     update_workstream_title,
@@ -91,6 +94,7 @@ from turnstone.core.providers import create_provider
 from turnstone.core.safety import is_command_blocked, sanitize_command
 from turnstone.core.sandbox import execute_math_sandboxed
 from turnstone.core.storage._registry import get_storage
+from turnstone.core.storage._utils import normalize_search_terms
 from turnstone.core.tool_advisory import escape_wrapper_tags, render_system_reminder
 from turnstone.core.tool_search import ToolSearchManager
 from turnstone.core.tools import (
@@ -427,6 +431,11 @@ class ChatSession:
             except Exception:
                 log.debug("rule_registry.init_failed", exc_info=True)
         self._memory_config = memory_config or MemoryConfig()
+        # Per-turn cache for _search_visible_memories — _init_system_messages
+        # fires many times within one turn (state transitions, MCP refresh,
+        # tool results) and the recent-context string is identical across
+        # them.  Invalidated on user-turn append and on memory write/delete.
+        self._mem_search_cache: dict[tuple[str, str, int], list[dict[str, str]]] = {}
         self._ws_id = ws_id or uuid.uuid4().hex
         self._title_generated = False
         self._read_files: set[str] = set()
@@ -578,7 +587,15 @@ class ChatSession:
         self._skill_resources_dir: str | None = None
         self._load_skills()
         self._init_system_messages()
-        self._save_config()
+        # Skip on rehydrate — ``_save_config`` is ``INSERT OR
+        # REPLACE`` per-key, and the persisted row is what
+        # ``ChatSession.resume`` is about to read back.  Pairs with
+        # ``SessionManager.open``'s saved-alias threading; together
+        # they keep reopened workstreams on their original model and
+        # settings instead of silently resetting to constructor
+        # defaults.
+        if not load_workstream_config(self._ws_id):
+            self._save_config()
 
     @property
     def ws_id(self) -> str:
@@ -1339,16 +1356,23 @@ class ChatSession:
                     model_name,
                     cfg.context_window,
                 )
-            elif saved_model and saved_model != self.model:
-                # No alias or alias no longer in registry — at least set the model name
-                self.model = saved_model
-                self._model_alias = None
-                self._cached_capabilities = None
+            elif saved_alias or saved_model:
+                # Saved alias is unset or no longer in the registry.
+                # Don't copy ``saved_model`` onto the constructor's
+                # default provider/client — pairing a removed model
+                # name with the default provider produces an API call
+                # the default provider can't service, which is exactly
+                # the broken state operators see today on the reopen
+                # path.  The constructor already resolved a coherent
+                # default; keep it intact and warn so the missing
+                # alias is auditable.
                 log.warning(
-                    "Resume: alias %r not in registry, keeping default provider=%s for model=%s",
+                    "Resume: saved alias=%r model=%r unreachable; "
+                    "keeping default provider=%s model=%s",
                     saved_alias,
-                    type(self._provider).__name__,
                     saved_model,
+                    type(self._provider).__name__,
+                    self.model,
                 )
             if "temperature" in config:
                 self.temperature = float(config["temperature"])
@@ -1434,11 +1458,6 @@ class ChatSession:
         """
         new_system_messages: list[dict[str, Any]] = []
 
-        # -- Chat template kwargs --
-        self._chat_template_kwargs_base: dict[str, Any] = {
-            "reasoning_effort": self.reasoning_effort,
-        }
-
         # -- Developer message --
         if self.creative_mode:
             dev_parts = [
@@ -1617,10 +1636,16 @@ class ChatSession:
         if self.instructions:
             dev_parts.append("")
             dev_parts.append(self.instructions)
-        visible_mems = self._list_visible_memories(limit=self._mem_cfg.fetch_limit)
+        context = extract_recent_context(self.messages)
+        visible_mems, candidate_source = self._select_memory_candidates(context)
         if visible_mems:
-            context = extract_recent_context(self.messages)
             relevant = score_memories(visible_mems, context, k=self._mem_cfg.relevance_k)
+            log.info(
+                "memory.composition",
+                source=candidate_source,
+                candidates=len(visible_mems),
+                injected=len(relevant),
+            )
             if relevant:
                 dev_parts.append("")
                 dev_parts.append(build_memory_context(relevant))
@@ -1814,48 +1839,48 @@ class ChatSession:
 
     def _provider_extra_params(
         self,
-        reasoning_effort: str | None = None,
         provider: LLMProvider | None = None,
         model_alias: str | None = None,
     ) -> dict[str, Any] | None:
         """Build provider-specific extra parameters.
 
-        ``chat_template_kwargs`` is only meaningful for local model servers
-        (``openai-compatible``).  Commercial OpenAI rejects it as an unknown
-        parameter, and handles ``reasoning_effort`` natively.
+        Forwards operator-supplied ``server_compat["extra_body"]`` overrides
+        (``skip_special_tokens``, ``reasoning_format``, or explicit
+        ``chat_template_kwargs``) to the OpenAI SDK ``extra_body``.  Operators
+        running gpt-oss-style local templates that consume ``reasoning_effort``
+        from ``chat_template_kwargs`` should set it explicitly under
+        ``server_compat["extra_body"]["chat_template_kwargs"]``.
 
-        Merges server workarounds (``skip_special_tokens``, etc.) from
-        ``ModelConfig.server_compat`` into the request's ``extra_body``.
-        Thinking-mode params (``enable_thinking``) are handled separately
-        by the provider based on ``ModelCapabilities.thinking_mode``.
+        Thinking-mode params (``enable_thinking``, ``thinking``) are added
+        separately by ``OpenAIChatCompletionsProvider._apply_thinking_mode``
+        based on ``ModelCapabilities.thinking_mode`` — the Responses API
+        surface handles reasoning natively and ignores ``extra_body``.
 
-        *model_alias* controls which model config supplies server compat
+        *model_alias* selects which stored config supplies server compat
         settings.  When ``None``, defaults to the session's primary alias.
         """
         from turnstone.core.server_compat import merge_server_compat
 
         prov = provider or self._provider
-        if prov.provider_name == "openai-compatible":
-            ctk_base = dict(self._chat_template_kwargs_base)
-            if reasoning_effort:
-                ctk_base["reasoning_effort"] = reasoning_effort
-            return merge_server_compat(
-                ctk_base,
-                self._get_server_compat(model_alias),
-            )
-        return None
+        # Only OpenAI-shaped providers consume extra_body.  Anthropic/Google
+        # have their own param paths handled inside their providers.
+        if prov.provider_name not in ("openai", "openai-compatible"):
+            return None
+        extra = merge_server_compat(None, self._get_server_compat(model_alias))
+        return extra or None
 
     def _get_server_compat(self, model_alias: str | None = None) -> dict[str, Any]:
         """Get server compatibility settings from a model config.
 
         *model_alias* selects the config to read.  Falls back to the
-        session's primary alias when ``None``.
+        session's primary alias when ``None``.  The returned dict is the
+        live ``ModelConfig.server_compat`` reference — callers must not
+        mutate it.  ``merge_server_compat`` reads only.
         """
         alias = model_alias or self._model_alias
         if self._registry and alias:
             try:
-                cfg = self._registry.get_config(alias)
-                return dict(cfg.server_compat)
+                return self._registry.get_config(alias).server_compat
             except (ValueError, KeyError):
                 pass
         return {}
@@ -1884,7 +1909,7 @@ class ChatSession:
             max_tokens=clamped,
             temperature=temperature,
             reasoning_effort=reasoning_effort,
-            extra_params=self._provider_extra_params(reasoning_effort=reasoning_effort),
+            extra_params=self._provider_extra_params(),
             capabilities=caps,
         )
 
@@ -2178,6 +2203,9 @@ class ChatSession:
         consume step adds it to the WHERE clause so a stale send can't
         steal rows reserved to a different one.
         """
+        # New user content invalidates the per-turn memory-search cache
+        # (composition will see a different recent-context string).
+        self._invalidate_memory_cache()
         user_content: str | list[dict[str, Any]]
         if attachments:
             parts: list[dict[str, Any]] = [{"type": "text", "text": user_input}]
@@ -2414,16 +2442,7 @@ class ChatSession:
                 if assistant_msg.get("_provider_content"):
                     provider_data = json.dumps(assistant_msg["_provider_content"])
 
-                # Build tool_calls JSON (excluding memory tools)
-                tool_calls_json: str | None = None
-                if tc:
-                    filtered_tc = [
-                        call
-                        for call in tc
-                        if call.get("function", {}).get("name", "") not in ("memory", "recall")
-                    ]
-                    if filtered_tc:
-                        tool_calls_json = json.dumps(filtered_tc)
+                tool_calls_json: str | None = json.dumps(tc) if tc else None
 
                 # Save assistant message atomically (content + tool_calls in one row)
                 if content or provider_data is not None or tool_calls_json:
@@ -2571,27 +2590,27 @@ class ChatSession:
                         tok_est = max(1, int(len(output) / self._chars_per_token))
                     self._msg_tokens.append(tok_est)
 
-                    # Log tool result (skip memory tools to avoid noise).
-                    # Use raw_output (pre-advisory-wrap) so DB stores clean
-                    # tool output without ephemeral advisory XML.
+                    # Log tool result.  Use raw_output (pre-advisory-wrap)
+                    # so the DB stores clean tool output without ephemeral
+                    # advisory XML.  memory/recall persist alongside every
+                    # other tool: replays show the full audit trail, and
+                    # output already passes through _truncate_output above
+                    # so size is bounded by the same budget every other
+                    # tool uses.
                     _tname = _tc_names.get(tc_id, "")
-                    if _tname not in (
-                        "memory",
-                        "recall",
-                    ):
-                        if isinstance(raw_output, list):
-                            store_text = " ".join(
-                                p.get("text", "") for p in raw_output if p.get("type") == "text"
-                            )[:2000]
-                        else:
-                            store_text = raw_output[:2000]
-                        save_message(
-                            self._ws_id,
-                            "tool",
-                            store_text,
-                            _tname,
-                            tool_call_id=tc_id,
-                        )
+                    if isinstance(raw_output, list):
+                        store_text = " ".join(
+                            p.get("text", "") for p in raw_output if p.get("type") == "text"
+                        )[:TOOL_RESULT_STORAGE_CAP]
+                    else:
+                        store_text = raw_output[:TOOL_RESULT_STORAGE_CAP]
+                    save_message(
+                        self._ws_id,
+                        "tool",
+                        store_text,
+                        _tname,
+                        tool_call_id=tc_id,
+                    )
                 # Inject user feedback from approval prompt (e.g. "y, use full path")
                 if user_feedback:
                     self.messages.append({"role": "user", "content": user_feedback})
@@ -5406,60 +5425,90 @@ class ChatSession:
             n += count_structured_memories(scope="user", scope_id=self._user_id)
         return n
 
+    def _visible_scopes(self) -> list[tuple[str, str]]:
+        """Return the (scope, scope_id) pairs visible to this session.
+
+        Coord sessions see ONLY their coord-scope; interactive sessions see
+        global + their workstream + their user (when uid present).  Drives
+        the single-query visibility helpers.
+        """
+        if self._kind == WorkstreamKind.COORDINATOR:
+            return [("coordinator", self._ws_id)]
+        scopes: list[tuple[str, str]] = [("global", ""), ("workstream", self._ws_id)]
+        if self._user_id:
+            scopes.append(("user", self._user_id))
+        return scopes
+
     def _list_visible_memories(self, mem_type: str = "", limit: int = 50) -> list[dict[str, str]]:
         """List memories visible to this session with optional type filter.
 
+        Single SQL round-trip — collapses the prior per-scope fan-out.
         See :meth:`_visible_memory_count` for the coord-isolation rule.
         """
-        if self._kind == WorkstreamKind.COORDINATOR:
-            return list_structured_memories(
-                mem_type=mem_type,
-                scope="coordinator",
-                scope_id=self._ws_id,
-                limit=limit,
-            )
-        global_mems = list_structured_memories(mem_type=mem_type, scope="global", limit=limit)
-        ws_mems = list_structured_memories(
-            mem_type=mem_type, scope="workstream", scope_id=self._ws_id, limit=limit
+        return list_visible_structured_memories(
+            self._visible_scopes(), mem_type=mem_type, limit=limit
         )
-        user_mems: list[dict[str, str]] = []
-        if self._user_id:
-            user_mems = list_structured_memories(
-                mem_type=mem_type, scope="user", scope_id=self._user_id, limit=limit
-            )
-        combined = global_mems + ws_mems + user_mems
-        combined.sort(key=lambda m: m.get("updated", ""), reverse=True)
-        return combined[:limit]
 
     def _search_visible_memories(
         self, query: str, mem_type: str = "", limit: int = 20
     ) -> list[dict[str, str]]:
         """Search memories visible to this session (scope-filtered).
 
+        Single SQL round-trip with a per-turn cache: ``_init_system_messages``
+        is invoked many times within a turn (state transitions, MCP refresh,
+        tool results) and the recent-context query is identical across them.
+        Cache is cleared on each new user turn and after memory writes/deletes.
         See :meth:`_visible_memory_count` for the coord-isolation rule.
         """
-        if self._kind == WorkstreamKind.COORDINATOR:
-            return search_structured_memories(
-                query,
-                mem_type=mem_type,
-                scope="coordinator",
-                scope_id=self._ws_id,
-                limit=limit,
-            )
-        global_mems = search_structured_memories(
-            query, mem_type=mem_type, scope="global", limit=limit
+        cache_key = (query, mem_type, limit)
+        cached = self._mem_search_cache.get(cache_key)
+        if cached is not None:
+            return cached
+        rows = search_visible_structured_memories(
+            query, self._visible_scopes(), mem_type=mem_type, limit=limit
         )
-        ws_mems = search_structured_memories(
-            query, mem_type=mem_type, scope="workstream", scope_id=self._ws_id, limit=limit
-        )
-        user_mems: list[dict[str, str]] = []
-        if self._user_id:
-            user_mems = search_structured_memories(
-                query, mem_type=mem_type, scope="user", scope_id=self._user_id, limit=limit
-            )
-        combined = global_mems + ws_mems + user_mems
-        combined.sort(key=lambda m: m.get("updated", ""), reverse=True)
-        return combined[:limit]
+        self._mem_search_cache[cache_key] = rows
+        return rows
+
+    def _invalidate_memory_cache(self) -> None:
+        """Drop the per-turn search cache; call on user-turn append + memory writes."""
+        self._mem_search_cache.clear()
+
+    def _select_memory_candidates(self, context: str) -> tuple[list[dict[str, str]], str]:
+        """Pick the candidate set fed into BM25 ranking.
+
+        Returns ``(memories, source_label)`` where source is one of:
+        ``recency`` (no context, or search returned nothing),
+        ``search`` (search saturated the fetch_limit budget alone), or
+        ``union`` (search hits ∪ recency, deduped by memory_id).
+
+        Invariant: the candidate pool is always a SUPERSET of the
+        recency-only pool the original bug used — recency is fully
+        preserved (not truncated) whenever it gets unioned.  Worst
+        case the union is 2 × fetch_limit candidates (~100 with
+        defaults), which BM25 ranks in pure Python in well under a
+        millisecond.  BM25's score>0 cutoff in bm25.py drops anything
+        that doesn't match the query, so unranked recency tail items
+        cost nothing on irrelevant candidates while saving the
+        relevant ones.
+
+        Capping the union at fetch_limit (the prior behavior) would
+        evict the recency tail when search added distinct hits — and
+        the recency tail is exactly where ancient-but-recently-touched
+        memories live, which is the recall the PR sets out to improve.
+        """
+        fetch_limit = self._mem_cfg.fetch_limit
+        if not context:
+            return self._list_visible_memories(limit=fetch_limit), "recency"
+        search_hits = self._search_visible_memories(context, limit=fetch_limit)
+        if len(search_hits) >= fetch_limit:
+            return search_hits, "search"
+        recency = self._list_visible_memories(limit=fetch_limit)
+        seen = {m["memory_id"] for m in search_hits}
+        extra = [m for m in recency if m["memory_id"] not in seen]
+        if not search_hits:
+            return extra, "recency"
+        return search_hits + extra, ("union" if extra else "search")
 
     def _check_metacognitive_nudge(self, user_message: str) -> tuple[str, str] | None:
         """Check if a metacognitive nudge should fire for *user_message*.
@@ -7983,6 +8032,10 @@ class ChatSession:
             agent_client = self.client
             agent_model = self.model
             agent_provider = self._provider
+            # When falling through to the session's primary model, use the
+            # session's primary alias for capability and server_compat
+            # resolution so the agent sees the same caps as the main loop.
+            agent_alias = self._model_alias
 
         # Per-kind reasoning effort.  Explicit caller arg wins; otherwise
         # delegate to the registry which knows the per-kind default (plan
@@ -8005,7 +8058,6 @@ class ChatSession:
         # Build extra params for agent calls — resolve server compat from the
         # agent's own model alias, not the session's primary model.
         agent_extra = self._provider_extra_params(
-            reasoning_effort=reasoning_effort,
             provider=agent_provider,
             model_alias=agent_alias,
         )
@@ -8444,6 +8496,7 @@ class ChatSession:
                     msg = f"Error: failed to save memory '{item['name']}'"
                     self._report_tool_result(call_id, "memory", msg, is_error=True)
                     return call_id, msg
+                self._invalidate_memory_cache()
                 self._init_system_messages()
                 if old is not None:
                     msg = f"Updated memory '{item['name']}' (type={item['mem_type']}, scope={item['scope']})"
@@ -8489,6 +8542,7 @@ class ChatSession:
                     msg = f"Error: memory '{item['name']}' not found (searched scopes: {tried})"
                     self._report_tool_result(call_id, "memory", msg, is_error=True)
                 else:
+                    self._invalidate_memory_cache()
                     self._init_system_messages()
                     msg = f"Deleted memory '{item['name']}' (scope={deleted_scope})"
                     self._report_tool_result(call_id, "memory", msg)
@@ -8516,6 +8570,12 @@ class ChatSession:
                         mem_type=item.get("mem_type", ""),
                         limit=item["limit"],
                     )
+                log.info(
+                    "memory.search",
+                    term_count=len(normalize_search_terms(item["query"])),
+                    result_count=len(rows),
+                    query=item["query"][:120],
+                )
                 if rows:
                     lines = []
                     for m in rows:

turnstone/core/session_manager.py

@@ -180,6 +180,7 @@ class SessionManager:
         node_id: str | None = None,
         state_writer: StateWriter | None = None,
         event_emitter: SessionEventEmitter | None = None,
+        model_validator: Callable[[str], bool] | None = None,
     ) -> None:
         if max_active < 1:
             raise ValueError(f"max_active must be >= 1, got {max_active}")
@@ -199,6 +200,15 @@ class SessionManager:
         # effects, and reserved for future kinds whose lifecycle
         # transitions don't fan out anywhere.
         self._event_emitter = event_emitter
+        # Optional registry-membership check applied to the persisted
+        # ``model_alias`` on the rehydrate path before threading it
+        # into ``build_session``.  Production wiring passes
+        # ``registry.has_alias``; an alias that has been removed from
+        # the registry since the workstream was created is filtered
+        # out so the session_factory falls back to its default rather
+        # than raising.  Restricted to the rehydrate path — fresh
+        # creates still want unknown aliases to surface as 503.
+        self._model_validator = model_validator
         self._node_id = node_id
         self._workstreams: dict[str, Workstream] = {}
         self._order: list[str] = []
@@ -603,8 +613,34 @@ class SessionManager:
                             evicted.id, reason="evicted", name=evicted.name
                         )
 
+                # Thread the persisted ``model_alias`` into
+                # ``build_session`` so reopened workstreams keep the
+                # model they were created with.  Pairs with the
+                # ``ChatSession.__init__`` skip-save guard: without
+                # both halves, ``_save_config`` clobbers persisted
+                # config with constructor defaults before
+                # ``ChatSession.resume`` reads them back.  When
+                # ``model_validator`` is wired and the saved alias is
+                # no longer in the registry, drop it so the factory
+                # falls back to its default — the session_factory
+                # itself still raises on unknown aliases, since
+                # fresh-create paths want that to surface as a 503.
+                saved_cfg = self._storage.load_workstream_config(ws_id)
+                saved_alias = (saved_cfg.get("model_alias") or None) if saved_cfg else None
+                if (
+                    saved_alias
+                    and self._model_validator is not None
+                    and not self._model_validator(saved_alias)
+                ):
+                    log.warning(
+                        "session_mgr.stale_alias_dropped ws=%s alias=%s",
+                        ws_id[:8],
+                        saved_alias,
+                    )
+                    saved_alias = None
+
                 try:
-                    ws.session = self._adapter.build_session(ws)
+                    ws.session = self._adapter.build_session(ws, model=saved_alias)
                 except Exception:
                     # Clean up the UI the adapter built before re-raising
                     # so any listener/lock resources are released.

turnstone/core/session_routes.py

@@ -394,6 +394,15 @@ class SessionEndpointConfig:
     # separate ``/history`` endpoint and doesn't render the per-tab
     # status bar). Kinds that don't need pre-replay wire ``None``.
     events_replay: EventsReplay | None = None
+    # async (ws, ui, request) -> None. Kind-specific async pre-step
+    # the lifted ``events`` body awaits BEFORE iterating
+    # ``events_replay``. Lets a kind move blocking storage I/O off
+    # the event loop (via ``asyncio.to_thread``) and stash results
+    # on ``request.state`` for the sync replay generator to read.
+    # Interactive uses it to pre-load intent_verdicts +
+    # output_assessments so ``_build_history``'s decoration stays
+    # off the hot path. Coord wires ``None``.
+    events_replay_prepare: Callable[..., Any] | None = None
     # (request) -> Executor for the SSE live-loop's blocking
     # ``queue.get`` wait. Interactive returns the dedicated
     # ``request.app.state.sse_executor`` (200-thread pool) so SSE
@@ -1203,6 +1212,8 @@ def make_open_handler(
     """
 
     async def open_ws(request: Request) -> Response:
+        import asyncio
+
         if cfg.permission_gate is not None:
             err = cfg.permission_gate(request)
             if err is not None:
@@ -1304,7 +1315,14 @@ def make_open_handler(
         # emit_rehydrated path).
         if cfg.open_post_load is not None:
             try:
-                cfg.open_post_load(request, ws)
+                # Off-loop: interactive's post_load runs the sync
+                # ``_build_history`` (storage I/O for verdict
+                # indexes + message reconstruction) — without the
+                # to_thread wrap this blocks the event loop on every
+                # workstream open, mirroring the SSE replay path
+                # that's already protected via
+                # ``events_replay_prepare``.
+                await asyncio.to_thread(cfg.open_post_load, request, ws)
             except Exception:
                 # Post-load is observational — never let a hook bug
                 # block the open. Log + continue.
@@ -1454,6 +1472,20 @@ def make_events_handler(cfg: SessionEndpointConfig) -> Handler:
                 # 500-slot cap on a chatty mid-generation workstream)
                 # while replay was being built.
                 if replay_cb is not None:
+                    # Kind-specific async prep — runs before the sync
+                    # replay generator iterates so blocking storage
+                    # I/O lands in the executor pool rather than the
+                    # event loop's hot path. Interactive uses this
+                    # to pre-load verdict indexes; coord skips.
+                    if cfg.events_replay_prepare is not None:
+                        try:
+                            await cfg.events_replay_prepare(ws, ui, request)
+                        except Exception:
+                            log.debug(
+                                "ws.events.replay_prepare_failed ws=%s",
+                                ws_id[:8],
+                                exc_info=True,
+                            )
                     try:
                         for ev in replay_cb(ws, ui, request):
                             yield {"data": json.dumps(ev)}
@@ -2240,6 +2272,34 @@ def make_history_handler(cfg: SessionEndpointConfig) -> Handler:
                 messages = await asyncio.to_thread(storage.load_messages, ws_id, limit=limit)
             except Exception:
                 log.debug("ws.history.load_failed ws=%s", ws_id[:8], exc_info=True)
+        # Audit-trail decoration — attach persisted intent_verdict and
+        # output_assessment data to each assistant.tool_calls entry so
+        # the dashboard's history replay paints the same verdict pills
+        # / output-warning bubbles the live SSE path shows.  Both
+        # storage queries are off-loop via ``to_thread``.  Best-effort:
+        # any failure leaves messages undecorated — replay degrades to
+        # the pre-decoration shape rather than 500-ing.
+        if messages:
+            try:
+                from turnstone.core.history_decoration import (
+                    decorate_history_messages,
+                    load_verdict_indexes,
+                )
+
+                indexes = await asyncio.to_thread(load_verdict_indexes, ws_id)
+                decorate_history_messages(messages, indexes[0], indexes[1])
+            except Exception:
+                # Operationally interesting: a persistent decoration
+                # failure (missing migration, driver mismatch, schema
+                # drift) silently strips verdict pills + output
+                # warnings from every reload of every workstream.
+                # Log at warning so it surfaces in normal log review
+                # rather than only when DEBUG is on.
+                log.warning(
+                    "ws.history.decoration_failed ws=%s",
+                    ws_id[:8],
+                    exc_info=True,
+                )
         return JSONResponse({"ws_id": ws_id, "messages": messages})
 
     return history

turnstone/core/storage/_postgresql.py

@@ -87,6 +87,9 @@ from turnstone.core.storage._utils import (
 from turnstone.core.storage._utils import (
     VERDICT_MUTABLE as _VERDICT_MUTABLE,
 )
+from turnstone.core.storage._utils import (
+    normalize_search_terms as _normalize_search_terms,
+)
 from turnstone.core.storage._utils import (
     reconstruct_messages as _reconstruct_messages,
 )
@@ -3315,7 +3318,10 @@ class PostgreSQLBackend:
         limit: int = 100,
     ) -> list[dict[str, str]]:
         with self._conn() as conn:
-            q = sa.select(structured_memories).order_by(structured_memories.c.updated.desc())
+            q = sa.select(structured_memories).order_by(
+                structured_memories.c.updated.desc(),
+                structured_memories.c.memory_id.asc(),
+            )
             if mem_type:
                 q = q.where(structured_memories.c.type == mem_type)
             if scope:
@@ -3334,11 +3340,16 @@ class PostgreSQLBackend:
         scope_id: str = "",
         limit: int = 20,
     ) -> list[dict[str, str]]:
+        """OR-of-terms ILIKE search; ranking is the caller's job (BM25 downstream)."""
         if not query or not query.strip():
             return self.list_structured_memories(
                 mem_type=mem_type, scope=scope, scope_id=scope_id, limit=limit
             )
-        terms = query.split()
+        terms = _normalize_search_terms(query)
+        if not terms:
+            return self.list_structured_memories(
+                mem_type=mem_type, scope=scope, scope_id=scope_id, limit=limit
+            )
         with self._conn() as conn:
             clauses = []
             params: dict[str, str] = {}
@@ -3352,25 +3363,116 @@ class PostgreSQLBackend:
                 params[f"n{i}"] = f"%{escaped}%"
                 params[f"d{i}"] = f"%{escaped}%"
                 params[f"c{i}"] = f"%{escaped}%"
-            where = " AND ".join(clauses)
+            term_clause = " OR ".join(clauses)
+            scope_filters = ""
             if mem_type:
-                where += " AND type = :type_filter"
+                scope_filters += " AND type = :type_filter"
                 params["type_filter"] = mem_type
             if scope:
-                where += " AND scope = :scope_filter"
+                scope_filters += " AND scope = :scope_filter"
                 params["scope_filter"] = scope
             if scope_id and scope:
-                where += " AND scope_id = :scope_id_filter"
+                scope_filters += " AND scope_id = :scope_id_filter"
                 params["scope_id_filter"] = scope_id
             rows = conn.execute(
                 sa.text(
-                    f"SELECT * FROM structured_memories WHERE {where} "
-                    f"ORDER BY updated DESC LIMIT :lim"
+                    f"SELECT * FROM structured_memories WHERE ({term_clause}){scope_filters} "
+                    f"ORDER BY updated DESC, memory_id ASC LIMIT :lim"
                 ),
                 {**params, "lim": limit},
             ).fetchall()
             return [dict(r._mapping) for r in rows]
 
+    def list_visible_structured_memories(
+        self,
+        scopes: list[tuple[str, str]],
+        mem_type: str = "",
+        limit: int = 100,
+    ) -> list[dict[str, str]]:
+        """Single-query union across visible (scope, scope_id) pairs.
+
+        Replaces the per-scope fan-out (one query per visible scope) so the
+        composition path issues 1 round-trip instead of 3.
+        """
+        if not scopes:
+            return []
+        with self._conn() as conn:
+            scope_clauses, params = self._build_scope_or_clause(scopes)
+            extra = ""
+            if mem_type:
+                extra = " AND type = :type_filter"
+                params["type_filter"] = mem_type
+            rows = conn.execute(
+                sa.text(
+                    f"SELECT * FROM structured_memories WHERE ({scope_clauses}){extra} "
+                    f"ORDER BY updated DESC, memory_id ASC LIMIT :lim"
+                ),
+                {**params, "lim": limit},
+            ).fetchall()
+            return [dict(r._mapping) for r in rows]
+
+    def search_visible_structured_memories(
+        self,
+        query: str,
+        scopes: list[tuple[str, str]],
+        mem_type: str = "",
+        limit: int = 20,
+    ) -> list[dict[str, str]]:
+        """OR-of-terms search joined with a single visibility OR-group.
+
+        Replaces the per-scope search fan-out; ranking is the caller's job.
+        """
+        if not scopes:
+            return []
+        if not query or not query.strip():
+            return self.list_visible_structured_memories(scopes, mem_type=mem_type, limit=limit)
+        terms = _normalize_search_terms(query)
+        if not terms:
+            return self.list_visible_structured_memories(scopes, mem_type=mem_type, limit=limit)
+        with self._conn() as conn:
+            scope_clauses, params = self._build_scope_or_clause(scopes)
+            term_clauses = []
+            for i, t in enumerate(terms):
+                escaped = _escape_ilike(t)
+                term_clauses.append(
+                    f"(name ILIKE :n{i} ESCAPE '\\' "
+                    f"OR description ILIKE :d{i} ESCAPE '\\' "
+                    f"OR content ILIKE :c{i} ESCAPE '\\')"
+                )
+                params[f"n{i}"] = f"%{escaped}%"
+                params[f"d{i}"] = f"%{escaped}%"
+                params[f"c{i}"] = f"%{escaped}%"
+            term_clause = " OR ".join(term_clauses)
+            extra = ""
+            if mem_type:
+                extra = " AND type = :type_filter"
+                params["type_filter"] = mem_type
+            rows = conn.execute(
+                sa.text(
+                    f"SELECT * FROM structured_memories "
+                    f"WHERE ({scope_clauses}) AND ({term_clause}){extra} "
+                    f"ORDER BY updated DESC, memory_id ASC LIMIT :lim"
+                ),
+                {**params, "lim": limit},
+            ).fetchall()
+            return [dict(r._mapping) for r in rows]
+
+    @staticmethod
+    def _build_scope_or_clause(
+        scopes: list[tuple[str, str]],
+    ) -> tuple[str, dict[str, str]]:
+        """Build a parameterized OR-group of (scope[, scope_id]) predicates."""
+        params: dict[str, str] = {}
+        clauses: list[str] = []
+        for i, (s, sid) in enumerate(scopes):
+            params[f"sc{i}"] = s
+            if sid:
+                params[f"sid{i}"] = sid
+                clauses.append(f"(scope = :sc{i} AND scope_id = :sid{i})")
+            else:
+                clauses.append(f"scope = :sc{i}")
+        return " OR ".join(clauses), params
+
     def touch_structured_memories(self, keys: list[tuple[str, str, str]]) -> int:
         """Batch-touch multiple memories by (name, scope, scope_id)."""
         if not keys:

turnstone/core/storage/_protocol.py

@@ -386,6 +386,34 @@ class StorageBackend(Protocol):
         """Search structured memories by query. Returns matching memory dicts."""
         ...
 
+    def list_visible_structured_memories(
+        self,
+        scopes: list[tuple[str, str]],
+        mem_type: str = "",
+        limit: int = 100,
+    ) -> list[dict[str, str]]:
+        """List memories matching ANY of the (scope, scope_id) pairs in *scopes*.
+
+        A pair with an empty ``scope_id`` matches the scope alone (used for
+        ``("global", "")``).  Single SQL query — replaces the per-scope fan-out
+        pattern that issued one query per visible scope.
+        """
+        ...
+
+    def search_visible_structured_memories(
+        self,
+        query: str,
+        scopes: list[tuple[str, str]],
+        mem_type: str = "",
+        limit: int = 20,
+    ) -> list[dict[str, str]]:
+        """OR-of-terms search across memories visible under *scopes*.
+
+        Single SQL query joining the scope OR-group with the term OR-group.
+        Ranking is the caller's job (BM25 downstream).
+        """
+        ...
+
     def touch_structured_memories(self, keys: list[tuple[str, str, str]]) -> int:
         """Batch-touch multiple memories.
 

turnstone/core/storage/_sqlite.py

@@ -87,6 +87,9 @@ from turnstone.core.storage._utils import (
 from turnstone.core.storage._utils import (
     VERDICT_MUTABLE as _VERDICT_MUTABLE,
 )
+from turnstone.core.storage._utils import (
+    normalize_search_terms as _normalize_search_terms,
+)
 from turnstone.core.storage._utils import (
     reconstruct_messages as _reconstruct_messages,
 )
@@ -3454,7 +3457,10 @@ class SQLiteBackend:
         limit: int = 100,
     ) -> list[dict[str, str]]:
         with self._conn() as conn:
-            q = sa.select(structured_memories).order_by(structured_memories.c.updated.desc())
+            q = sa.select(structured_memories).order_by(
+                structured_memories.c.updated.desc(),
+                structured_memories.c.memory_id.asc(),
+            )
             if mem_type:
                 q = q.where(structured_memories.c.type == mem_type)
             if scope:
@@ -3473,11 +3479,16 @@ class SQLiteBackend:
         scope_id: str = "",
         limit: int = 20,
     ) -> list[dict[str, str]]:
+        """OR-of-terms LIKE search; ranking is the caller's job (BM25 downstream)."""
         if not query or not query.strip():
             return self.list_structured_memories(
                 mem_type=mem_type, scope=scope, scope_id=scope_id, limit=limit
             )
-        terms = query.split()
+        terms = _normalize_search_terms(query)
+        if not terms:
+            return self.list_structured_memories(
+                mem_type=mem_type, scope=scope, scope_id=scope_id, limit=limit
+            )
         with self._conn() as conn:
             clauses = []
             params: dict[str, str] = {}
@@ -3491,25 +3502,109 @@ class SQLiteBackend:
                 params[f"n{i}"] = f"%{escaped}%"
                 params[f"d{i}"] = f"%{escaped}%"
                 params[f"c{i}"] = f"%{escaped}%"
-            where = " AND ".join(clauses)
+            term_clause = " OR ".join(clauses)
+            scope_filters = ""
             if mem_type:
-                where += " AND type = :type_filter"
+                scope_filters += " AND type = :type_filter"
                 params["type_filter"] = mem_type
             if scope:
-                where += " AND scope = :scope_filter"
+                scope_filters += " AND scope = :scope_filter"
                 params["scope_filter"] = scope
             if scope_id and scope:
-                where += " AND scope_id = :scope_id_filter"
+                scope_filters += " AND scope_id = :scope_id_filter"
                 params["scope_id_filter"] = scope_id
             rows = conn.execute(
                 sa.text(
-                    f"SELECT * FROM structured_memories WHERE {where} "
-                    f"ORDER BY updated DESC LIMIT :lim"
+                    f"SELECT * FROM structured_memories WHERE ({term_clause}){scope_filters} "
+                    f"ORDER BY updated DESC, memory_id ASC LIMIT :lim"
                 ),
                 {**params, "lim": limit},
             ).fetchall()
             return [dict(r._mapping) for r in rows]
 
+    def list_visible_structured_memories(
+        self,
+        scopes: list[tuple[str, str]],
+        mem_type: str = "",
+        limit: int = 100,
+    ) -> list[dict[str, str]]:
+        """Single-query union across visible (scope, scope_id) pairs."""
+        if not scopes:
+            return []
+        with self._conn() as conn:
+            scope_clauses, params = self._build_scope_or_clause(scopes)
+            extra = ""
+            if mem_type:
+                extra = " AND type = :type_filter"
+                params["type_filter"] = mem_type
+            rows = conn.execute(
+                sa.text(
+                    f"SELECT * FROM structured_memories WHERE ({scope_clauses}){extra} "
+                    f"ORDER BY updated DESC, memory_id ASC LIMIT :lim"
+                ),
+                {**params, "lim": limit},
+            ).fetchall()
+            return [dict(r._mapping) for r in rows]
+
+    def search_visible_structured_memories(
+        self,
+        query: str,
+        scopes: list[tuple[str, str]],
+        mem_type: str = "",
+        limit: int = 20,
+    ) -> list[dict[str, str]]:
+        """OR-of-terms search joined with a single visibility OR-group."""
+        if not scopes:
+            return []
+        if not query or not query.strip():
+            return self.list_visible_structured_memories(scopes, mem_type=mem_type, limit=limit)
+        terms = _normalize_search_terms(query)
+        if not terms:
+            return self.list_visible_structured_memories(scopes, mem_type=mem_type, limit=limit)
+        with self._conn() as conn:
+            scope_clauses, params = self._build_scope_or_clause(scopes)
+            term_clauses = []
+            for i, t in enumerate(terms):
+                escaped = _escape_like(t)
+                term_clauses.append(
+                    f"(name LIKE :n{i} ESCAPE '\\' "
+                    f"OR description LIKE :d{i} ESCAPE '\\' "
+                    f"OR content LIKE :c{i} ESCAPE '\\')"
+                )
+                params[f"n{i}"] = f"%{escaped}%"
+                params[f"d{i}"] = f"%{escaped}%"
+                params[f"c{i}"] = f"%{escaped}%"
+            term_clause = " OR ".join(term_clauses)
+            extra = ""
+            if mem_type:
+                extra = " AND type = :type_filter"
+                params["type_filter"] = mem_type
+            rows = conn.execute(
+                sa.text(
+                    f"SELECT * FROM structured_memories "
+                    f"WHERE ({scope_clauses}) AND ({term_clause}){extra} "
+                    f"ORDER BY updated DESC, memory_id ASC LIMIT :lim"
+                ),
+                {**params, "lim": limit},
+            ).fetchall()
+            return [dict(r._mapping) for r in rows]
+
+    @staticmethod
+    def _build_scope_or_clause(
+        scopes: list[tuple[str, str]],
+    ) -> tuple[str, dict[str, str]]:
+        """Build a parameterized OR-group of (scope[, scope_id]) predicates."""
+        params: dict[str, str] = {}
+        clauses: list[str] = []
+        for i, (s, sid) in enumerate(scopes):
+            params[f"sc{i}"] = s
+            if sid:
+                params[f"sid{i}"] = sid
+                clauses.append(f"(scope = :sc{i} AND scope_id = :sid{i})")
+            else:
+                clauses.append(f"scope = :sc{i}")
+        return " OR ".join(clauses), params
+
     def touch_structured_memories(self, keys: list[tuple[str, str, str]]) -> int:
         """Batch-touch multiple memories by (name, scope, scope_id)."""
         if not keys:

turnstone/core/storage/_utils.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import base64
 import contextlib
 import json
+import re
 from typing import Any
 
 from turnstone.core.attachments import unreadable_placeholder
@@ -48,6 +49,39 @@ def _attachment_to_content_part(att: dict[str, Any]) -> dict[str, Any] | None:
     return None
 
 
+# ---------------------------------------------------------------------------
+# Search-term normalization
+# ---------------------------------------------------------------------------
+
+# Composition can hand a multi-KB pasted user message to ILIKE-based search;
+# without a cap, every distinct token would emit one unindexable predicate
+# per scope-fanned query, producing hundreds of seq-scan clauses on a single
+# rebuild.  Cap + dedupe + length filter keeps the SQL bounded.
+_MAX_SEARCH_TERMS = 16
+_MIN_TERM_LEN = 2
+
+# Streaming tokenizer — finditer doesn't allocate a full list up front,
+# so a multi-KB pasted query stops being scanned the moment the cap is
+# hit instead of after splitting every token.
+_TOKEN_RE = re.compile(r"\S+")
+
+
+def normalize_search_terms(query: str) -> list[str]:
+    """De-dupe (case-insensitive), drop short tokens, and cap at MAX terms."""
+    seen: set[str] = set()
+    terms: list[str] = []
+    for match in _TOKEN_RE.finditer(query):
+        raw = match.group()
+        lowered = raw.lower()
+        if len(lowered) < _MIN_TERM_LEN or lowered in seen:
+            continue
+        seen.add(lowered)
+        terms.append(raw)
+        if len(terms) >= _MAX_SEARCH_TERMS:
+            break
+    return terms
+
+
 # ---------------------------------------------------------------------------
 # Text sanitization
 # ---------------------------------------------------------------------------

turnstone/server.py

@@ -53,6 +53,15 @@ from turnstone.core.auth import (
     _DenyFilter,
     jwt_version_slot,
 )
+from turnstone.core.history_decoration import (
+    TOOL_RESULT_STORAGE_CAP,
+)
+from turnstone.core.history_decoration import (
+    decorate_tool_call as _decorate_tool_call,
+)
+from turnstone.core.history_decoration import (
+    load_verdict_indexes as _load_verdict_indexes,
+)
 from turnstone.core.log import get_logger
 from turnstone.core.metrics import metrics as _metrics
 from turnstone.core.ratelimit import resolve_client_ip
@@ -408,8 +417,20 @@ class WebUI(SessionUIBase):
 # ---------------------------------------------------------------------------
 
 
+# Verdict + output-assessment decoration helpers (``_decorate_tool_call``,
+# ``_load_verdict_indexes``) are imported at module top alongside the
+# rest of ``turnstone.core.*``.  Both this builder and
+# :func:`make_history_handler` (the /history REST endpoint coord uses
+# as its primary history loader) share them so the two surfaces don't
+# drift on the wire shape they emit.
+
+
 def _build_history(
-    session: ChatSession, has_pending_approval: bool = False
+    session: ChatSession,
+    has_pending_approval: bool = False,
+    *,
+    verdicts: dict[str, dict[str, Any]] | None = None,
+    assessments: dict[str, dict[str, Any]] | None = None,
 ) -> list[dict[str, Any]]:
     """Build a history replay list from ChatSession messages.
 
@@ -421,6 +442,12 @@ def _build_history(
     ``"denied": True``, and the corresponding assistant entry that
     issued the tool calls is also marked ``"denied": True`` so the
     client can render the correct badge.
+
+    ``verdicts`` and ``assessments`` are optional pre-loaded
+    ``{call_id → row}`` dicts (see :func:`_load_verdict_indexes`).
+    Async callers should pre-load via ``asyncio.to_thread`` and pass
+    them in to avoid blocking the event loop on storage I/O. When
+    omitted, the storage call runs inline (sync call sites).
     """
     # Metacognitive nudges live on the message dict's ``_reminders``
     # side-channel — user messages carry user-channel nudges
@@ -432,6 +459,18 @@ def _build_history(
     # ``content`` never carries the ``<system-reminder>`` envelope —
     # that splice is transient, applied to a wire-bound copy in
     # ``ChatSession._apply_reminders_for_provider``.
+    #
+    # Verdict + output-assessment lookup tables — populated either
+    # inline (sync call sites) or pre-loaded by an async caller via
+    # asyncio.to_thread (see _load_verdict_indexes).  Pre-loading is
+    # what keeps _build_history off the event loop's hot path on the
+    # SSE replay generator path.
+    if verdicts is not None and assessments is not None:
+        verdicts_by_call_id = verdicts
+        assessments_by_call_id = assessments
+    else:
+        ws_id = getattr(session, "_ws_id", "") or ""
+        verdicts_by_call_id, assessments_by_call_id = _load_verdict_indexes(ws_id)
     history = []
     for msg in session.messages:
         content = msg.get("content")
@@ -496,17 +535,45 @@ def _build_history(
             if clean_reminders:
                 entry["reminders"] = clean_reminders
         if msg.get("tool_calls"):
-            entry["tool_calls"] = [
-                {
-                    "id": tc.get("id", ""),
+            tc_entries: list[dict[str, Any]] = []
+            for tc in msg["tool_calls"]:
+                tc_entry: dict[str, Any] = {
+                    "id": tc.get("id", "") or "",
                     "name": tc["function"]["name"],
                     "arguments": tc["function"].get("arguments", ""),
                 }
-                for tc in msg["tool_calls"]
-            ]
+                # Decorate with persisted verdict + output_assessment
+                # via the shared helper (also used by
+                # ``make_history_handler``).  Skips unflagged
+                # ("risk_level == 'none'") rows so the wire stays
+                # tight; ships only the fields the UI renders.
+                _decorate_tool_call(
+                    tc_entry,
+                    verdicts_by_call_id,
+                    assessments_by_call_id,
+                )
+                tc_entries.append(tc_entry)
+            entry["tool_calls"] = tc_entries
         # Detect denied/blocked/errored tool results by their content prefix.
         if msg.get("role") == "tool":
             content = msg.get("content", "")
+            # Propagate tool_call_id so replayHistory can anchor the
+            # rendered output to the specific .ts-approval-tool element
+            # by data-call-id (mirrors the live appendToolOutput path).
+            # Without this, multi-tool batches render every result at
+            # the bottom of the block rather than under each header.
+            result_call_id = msg.get("tool_call_id")
+            if result_call_id:
+                entry["tool_call_id"] = str(result_call_id)
+            # Tool results are clamped to TOOL_RESULT_STORAGE_CAP
+            # chars per row at storage time (session.py).  Surface
+            # that on replay so the user knows the visible output is
+            # a clipped view of what the live session saw, rather
+            # than the full result.  Reference the shared constant
+            # rather than a literal so the UI pill logic can't
+            # silently desync if the cap ever changes.
+            if isinstance(content, str) and len(content) >= TOOL_RESULT_STORAGE_CAP:
+                entry["truncated"] = True
             if isinstance(content, str):
                 if content.startswith("Denied by user") or content.startswith("Blocked"):
                     entry["denied"] = True
@@ -747,6 +814,31 @@ def _audit_close_workstream(
     )
 
 
+async def _interactive_events_replay_prepare(ws: Workstream, ui: Any, request: Request) -> None:
+    """Async pre-step run before ``_interactive_events_replay`` iterates.
+
+    Loads ``intent_verdicts`` + ``output_assessments`` for the
+    workstream off the event loop (via ``asyncio.to_thread``) and
+    stashes the result on ``request.state.verdict_indexes``. The sync
+    replay generator reads from there and passes the dicts into
+    ``_build_history`` so the storage I/O never blocks the event loop
+    on the SSE replay path.
+
+    Best-effort: if the workstream has no session or no ws_id, leaves
+    ``request.state.verdict_indexes`` unset and ``_build_history``
+    falls back to the inline storage call (sync path).
+    """
+    del ui  # not needed; lookup is keyed on ws.session._ws_id
+    session = ws.session
+    if session is None:
+        return
+    ws_id = getattr(session, "_ws_id", "") or ""
+    if not ws_id:
+        return
+    indexes = await asyncio.to_thread(_load_verdict_indexes, ws_id)
+    request.state.verdict_indexes = indexes
+
+
 def _interactive_events_replay(
     ws: Workstream, ui: Any, request: Request
 ) -> Iterable[dict[str, Any]]:
@@ -764,7 +856,6 @@ def _interactive_events_replay(
 
     Pure read — never mutates ``ws`` / ``ui`` / ``session``.
     """
-    del request  # not needed; replay reads ws/ui/session state
     session = ws.session
     if session is None:
         # Defensive — the lifted body's UI presence check guarantees
@@ -779,9 +870,22 @@ def _interactive_events_replay(
 
     # History replay — pending-approval flag rides on the last
     # assistant entry's tool_calls so the client renders them as
-    # awaiting approval rather than already approved.
+    # awaiting approval rather than already approved.  Verdict /
+    # assessment indexes were pre-loaded off the event loop by
+    # _interactive_events_replay_prepare; passing them in here keeps
+    # _build_history's storage I/O out of the sync generator path.
     pending_approval = getattr(ui, "_pending_approval", None)
-    history = _build_history(session, has_pending_approval=pending_approval is not None)
+    cached_indexes = getattr(request.state, "verdict_indexes", None)
+    if isinstance(cached_indexes, tuple) and len(cached_indexes) == 2:
+        verdicts, assessments = cached_indexes
+    else:
+        verdicts, assessments = None, None
+    history = _build_history(
+        session,
+        has_pending_approval=pending_approval is not None,
+        verdicts=verdicts,
+        assessments=assessments,
+    )
     if history:
         yield {"type": "history", "messages": history}
 
@@ -1456,13 +1560,13 @@ async def command(request: Request) -> JSONResponse:
             ui._enqueue({"type": "clear_ui"})
         elif cmd_word == "/resume":
             ui._enqueue({"type": "clear_ui"})
-            history = _build_history(ws.session)
+            history = await asyncio.to_thread(_build_history, ws.session)
             if history:
                 ui._enqueue({"type": "history", "messages": history})
         elif cmd_word in ("/rewind", "/retry"):
             # Refresh frontend with truncated history
             ui._enqueue({"type": "clear_ui"})
-            history = _build_history(ws.session)
+            history = await asyncio.to_thread(_build_history, ws.session)
             if history:
                 ui._enqueue({"type": "history", "messages": history})
             # Audit trail
@@ -1937,7 +2041,7 @@ async def _interactive_create_post_install(
             ui = ws.ui
             if isinstance(ui, WebUI):
                 ui._enqueue({"type": "clear_ui"})
-                history = _build_history(ws.session)
+                history = await asyncio.to_thread(_build_history, ws.session)
                 if history:
                     ui._enqueue({"type": "history", "messages": history})
             with contextlib.suppress(queue.Full):
@@ -2863,13 +2967,14 @@ def internal_model_reload(request: Request) -> JSONResponse:
     if registry is None or cli_args is None:
         return JSONResponse({"status": "error", "reason": "no registry"}, status_code=503)
 
+    storage = get_storage()
     new_registry = load_model_registry(
         base_url=cli_args["base_url"],
         api_key=cli_args["api_key"],
         model=cli_args["model"],
         context_window=cli_args["context_window"],
         provider=cli_args["provider"],
-        storage=get_storage(),
+        storage=storage,
     )
     cs = getattr(request.app.state, "config_store", None)
     if cs is not None:
@@ -2935,6 +3040,13 @@ def internal_model_reload(request: Request) -> JSONResponse:
     # `model` parameter descriptions reflect the current registry.
     _broadcast_agent_tool_schema_refresh(request.app.state)
 
+    # Refresh the per-node ``models`` metadata entry the coord reads on
+    # ``list_nodes``.  Without this, the heartbeat loop's 30s tick would
+    # be the coord's first chance to see new aliases an admin just added.
+    node_id = getattr(request.app.state, "node_id", "")
+    if node_id:
+        _publish_models_metadata(request.app.state, storage, node_id)
+
     return JSONResponse({"status": "ok", "aliases": registry.list_aliases()})
 
 
@@ -2960,6 +3072,109 @@ def internal_model_status(request: Request) -> JSONResponse:
     return JSONResponse({"models": models})
 
 
+def _collect_node_models_metadata(app_state: Any) -> tuple[str, str, str] | None:
+    """Build the ``("models", json_value, "auto")`` node_metadata entry.
+
+    Each model alias on the live registry is projected to
+    ``{alias, provider, healthy}`` — the alias is what the coordinator
+    passes back as ``spawn_workstream(model=...)``, ``provider`` lets
+    coordinators classify or filter (e.g. "any anthropic node"), and
+    ``healthy`` reflects the backend's :class:`BackendHealthTracker`
+    state at call time.  The underlying model identifier (``cfg.model``)
+    is intentionally omitted — coordinators kept reaching for the
+    provider-side string when they should have been passing the local
+    alias, and dropping it removes the footgun.  Operators who need
+    the model string can hit ``/v1/api/_internal/model-status`` on the
+    node directly.
+
+    Trackers are eagerly seeded for every alias at server startup and
+    on every model-reload, so ``health_reg.get_tracker(...)`` returns
+    the existing tracker rather than minting a fresh one in steady
+    state.  In the unlikely race where a tracker hasn't been seeded
+    yet, the freshly created tracker reports ``is_healthy=True``
+    (default state) — which matches the prior "default to True when
+    no tracker" behavior, just routed through the tracker object.
+
+    Returns ``None`` when the registry has not yet been built (caller
+    should skip the write rather than zero out a previous snapshot).
+    """
+    registry = getattr(app_state, "registry", None)
+    if registry is None:
+        return None
+    health_reg = getattr(app_state, "health_registry", None)
+    aliases_info: list[dict[str, Any]] = []
+    # Iterate aliases in a stable order — ``list_aliases`` returns dict
+    # insertion order, so two structurally identical registries built
+    # from different sources (config.toml vs. DB rows in different
+    # commit order) would otherwise serialize to different JSON and
+    # defeat the publish-cache hit-rate that the
+    # ``turnstone_node_models_publish_total`` metric tracks.
+    for alias in sorted(registry.list_aliases()):
+        try:
+            cfg = registry.get_config(alias)
+        except (ValueError, KeyError):
+            continue
+        healthy = True
+        if health_reg is not None:
+            # Direct keyed lookup — ``get_tracker_for_alias`` would
+            # do a second ``registry.get_config(alias)`` internally,
+            # but ``cfg`` is already in hand here.
+            tracker = health_reg.get_tracker(provider=cfg.provider, base_url=cfg.base_url)
+            healthy = tracker.is_healthy
+        aliases_info.append(
+            {
+                "alias": alias,
+                "provider": cfg.provider,
+                "healthy": healthy,
+            }
+        )
+    return ("models", json.dumps(aliases_info), "auto")
+
+
+def _publish_models_metadata(app_state: Any, storage: Any, node_id: str) -> None:
+    """Refresh the per-node ``models`` row when the projection changed.
+
+    Caches the last-written JSON on ``app_state._last_models_payload``
+    so back-to-back heartbeat ticks with no health flip don't churn
+    the row — without this, the ``updated`` timestamp on every node's
+    ``models`` row advances every 30s across the whole cluster.
+
+    Records the cache outcome on the metrics collector so
+    ``turnstone_node_models_publish_total{outcome=...}`` exposes the
+    hit/miss ratio to Prometheus.  Storage-error attempts don't
+    record either outcome — the next call will retry and the
+    counters reflect actual cache decisions, not transient DB
+    failures.
+
+    Sync — callers on the asyncio loop wrap with ``asyncio.to_thread``.
+    Concurrent callers (heartbeat tick vs. ``internal_model_reload``)
+    can race on the cache attribute; the worst case is a redundant
+    write, never a stale row, so we skip the lock.
+    """
+    from turnstone.core.storage._registry import StorageUnavailableError
+
+    try:
+        entry = _collect_node_models_metadata(app_state)
+    except Exception:
+        log.warning("server.node_models_projection_failed", exc_info=True)
+        return
+    if entry is None:
+        return
+    payload = entry[1]
+    if payload == getattr(app_state, "_last_models_payload", None):
+        _metrics.record_node_models_publish(written=False)
+        return
+    try:
+        storage.set_node_metadata_bulk(node_id, [entry])
+    except StorageUnavailableError:
+        return  # storage layer already logged
+    except Exception:
+        log.exception("server.node_models_publish_failed")
+        return
+    app_state._last_models_payload = payload
+    _metrics.record_node_models_publish(written=True)
+
+
 # ---------------------------------------------------------------------------
 # Global SSE fan-out
 # ---------------------------------------------------------------------------
@@ -3236,6 +3451,22 @@ async def _lifespan(app: Starlette) -> AsyncGenerator[None, None]:
             ]
             _cfg_meta = _load_meta_config("metadata")
             _meta_entries.extend((k, json.dumps(v), "config") for k, v in _cfg_meta.items())
+            # Project the live model registry into a ``models`` entry so
+            # coord-side ``list_nodes`` can surface healthy aliases per
+            # node without a fan-out HTTP probe.  Re-collected on each
+            # heartbeat tick so health flips converge within ~30s.
+            # Wrapped in its own try/except so a projection failure
+            # doesn't take out the auto+config metadata write — losing
+            # the discovery surface is recoverable on the next heartbeat
+            # tick, but losing ``arch`` / ``os`` / ``cpu_count`` blinds
+            # the cluster's capability filters until the next restart.
+            try:
+                _models_entry = _collect_node_models_metadata(app.state)
+            except Exception:
+                log.warning("server.node_models_projection_failed", exc_info=True)
+                _models_entry = None
+            if _models_entry is not None:
+                _meta_entries.append(_models_entry)
             if _meta_entries:
                 # Clear stale auto/config rows from a prior run before upserting
                 _svc_storage.delete_node_metadata_by_source(_svc_node_id, "auto")
@@ -3246,11 +3477,26 @@ async def _lifespan(app: Starlette) -> AsyncGenerator[None, None]:
                     node_id=_svc_node_id,
                     count=len(_meta_entries),
                 )
+                # Seed the publish-cache so the first heartbeat tick
+                # doesn't redundant-write the same payload we just put
+                # in the bulk above.
+                if _models_entry is not None:
+                    app.state._last_models_payload = _models_entry[1]
         except Exception:
             log.warning("server.node_metadata_failed", node_id=_svc_node_id, exc_info=True)
 
         async def _heartbeat_loop() -> None:
-            """Periodically update service heartbeat."""
+            """Periodically update service heartbeat and refresh models metadata.
+
+            The ``models`` entry on ``node_metadata`` doubles as the
+            coord-side discovery surface for healthy model aliases per
+            node — refreshed every 30s so health flips and registry
+            reloads converge promptly without a fan-out HTTP probe on
+            the coord's ``list_nodes`` path.  The publish step short-
+            circuits when the projection is byte-identical to the
+            last write (cache lives on ``app.state``), so a stable
+            cluster doesn't pay UPSERT churn here.
+            """
             from turnstone.core.storage._registry import StorageUnavailableError
 
             while True:
@@ -3261,6 +3507,12 @@ async def _lifespan(app: Starlette) -> AsyncGenerator[None, None]:
                     pass  # already logged by storage layer
                 except Exception:
                     log.exception("server.heartbeat_failed")
+                # Both projection and write happen in the worker thread
+                # — keeps the registry-lock acquisition off the loop and
+                # bundles the round-trip into a single offload.
+                await asyncio.to_thread(
+                    _publish_models_metadata, app.state, _svc_storage, _svc_node_id
+                )
 
         _heartbeat_task = asyncio.create_task(_heartbeat_loop())
 
@@ -3268,6 +3520,13 @@ async def _lifespan(app: Starlette) -> AsyncGenerator[None, None]:
     # Shutdown
     if _heartbeat_task is not None:
         _heartbeat_task.cancel()
+        # Wait for the cancel to land before we run the metadata
+        # delete below — a heartbeat tick mid-write would otherwise
+        # complete its ``set_node_metadata_bulk`` AFTER our
+        # ``delete_node_metadata_by_source(..., "auto")`` and
+        # resurrect the row we just cleared.
+        with contextlib.suppress(asyncio.CancelledError, Exception):
+            await _heartbeat_task
     if _svc_node_id and _svc_url:
         from turnstone.core.storage import get_storage as _get_svc_dereg
 
@@ -3433,6 +3692,7 @@ def create_app(
         open_resolve_alias=_resolve_workstream_alias,
         open_post_load=_interactive_open_post_load,
         events_replay=_interactive_events_replay,
+        events_replay_prepare=_interactive_events_replay_prepare,
         # Pre-lift ``events_sse`` used the dedicated 200-thread
         # ``sse_executor`` so SSE polling stayed isolated from
         # every other ``asyncio.to_thread`` caller in the process
@@ -3914,6 +4174,13 @@ def main() -> None:
         assert ui is not None
         # Resolve the effective alias once and use it consistently
         # for both client resolution and ChatSession.model_alias.
+        # Unknown aliases here raise ValueError — the create handler
+        # maps that to a 503 with operator-friendly text so a typo or
+        # removed alias in body.model surfaces instead of silently
+        # starting on the default. SessionManager.open's rehydrate
+        # path is the one place where unknown aliases must NOT fail
+        # loud; the manager filters those out via its model_validator
+        # before the alias reaches this factory.
         model_alias = model_alias or _effective_default_alias()
         r_client, r_model, r_cfg = registry.resolve(model_alias)
         # Read MCP client from shared ref — may have been replaced after startup
@@ -4043,6 +4310,10 @@ def main() -> None:
         # emit_rehydrated are no-ops because those events fire from
         # out-of-band paths (create handler + WebUI._broadcast_state).
         event_emitter=interactive_adapter,
+        # Filter out persisted aliases that no longer resolve so a
+        # workstream pinned to a since-removed alias still rehydrates
+        # (on the registry default) instead of 500-ing on every reopen.
+        model_validator=registry.has_alias,
     )
     interactive_adapter.attach(manager)
     WebUI._workstream_mgr = manager

turnstone/tools/list_nodes.json

@@ -1,6 +1,6 @@
 {
   "name": "list_nodes",
-  "description": "List active cluster nodes with their metadata. By default only nodes with a fresh service-registry heartbeat (within 120s) are returned. Pass arbitrary `key=value` filters to narrow; all filters must match (AND). Pair with `target_node` on spawn_workstream to pin a child to a node that matches a capability. The 120s heartbeat is a sliding window, so a node returned here can drop out before a follow-up spawn lands — the race produces `\"No available node for routing\"`; omit `target_node` to let rendezvous pick from the still-healthy set, or retry after re-listing if a specific node is required. The `interfaces` key (container IPs, interface names) is stripped by default — routing should use capability/region tags, not IPs; pass `include_network_detail=true` only for debugging. Pass `include_inactive=true` to surface stale registrations (those nodes will reject `target_node` pinning).",
+  "description": "List active cluster nodes with their metadata. By default only nodes with a fresh service-registry heartbeat (within 120s) are returned. Each row carries `node_id`, `metadata`, and `model_aliases` — the latter being a list of healthy model aliases the node will accept on `spawn_workstream(model=...)` / `spawn_batch` (refreshed every 30s by the node's heartbeat). Pass arbitrary `key=value` filters to narrow; all filters must match (AND). Pair with `target_node` on spawn_workstream to pin a child to a node that matches a capability. The 120s heartbeat is a sliding window, so a node returned here can drop out before a follow-up spawn lands — the race produces `\"No available node for routing\"`; omit `target_node` to let rendezvous pick from the still-healthy set, or retry after re-listing if a specific node is required. The `interfaces` key (container IPs, interface names) is stripped by default — routing should use capability/region tags, not IPs; pass `include_network_detail=true` only for debugging. Pass `include_inactive=true` to surface stale registrations (those nodes will reject `target_node` pinning).",
   "parameters": {
     "type": "object",
     "properties": {

turnstone/tools/spawn_batch.json

@@ -24,7 +24,7 @@
             },
             "model": {
               "type": "string",
-              "description": "Optional model alias."
+              "description": "Optional model alias. Discover available aliases per node via `list_nodes.model_aliases`."
             },
             "target_node": {
               "type": "string",

turnstone/tools/spawn_workstream.json

@@ -18,7 +18,7 @@
       },
       "model": {
         "type": "string",
-        "description": "Optional model alias from the registry. Omit to use the coordinator's default model (or the one the skill prescribes)."
+        "description": "Optional model alias from the registry. Discover available aliases per node via `list_nodes` (the `model_aliases` field on each row lists the healthy aliases that node will accept). Omit to use the coordinator's default model (or the one the skill prescribes)."
       },
       "target_node": {
         "type": "string",

turnstone/ui/static/app.js

@@ -1036,6 +1036,19 @@ Pane.prototype.replayHistory = function (messages) {
     this.showEmptyState();
     return;
   }
+  // Suppress the polite live region while we batch-build the replay
+  // — messagesEl is aria-live="polite" so a fresh replay would otherwise
+  // queue an announcement for every approved/denied/verdict pill we
+  // insert.  Restored after the loop so live SSE updates announce
+  // normally.  WCAG 4.1.3 — historical content should not behave like
+  // real-time updates.
+  this.messagesEl.setAttribute("aria-busy", "true");
+  // pendingAssessments[call_id] = output_assessment dict.  Populated
+  // from the assistant branch, consumed by the role==="tool" branch
+  // (or after the loop, for legacy rows missing tool_call_id).
+  // Replaces a JSON.stringify→dataset→JSON.parse round-trip with an
+  // in-memory map keyed by call_id.
+  var pendingAssessments = {};
   var lastToolBlock = null;
   for (var i = 0; i < messages.length; i++) {
     var msg = messages[i];
@@ -1052,6 +1065,25 @@ Pane.prototype.replayHistory = function (messages) {
       }
       lastToolBlock = null;
     } else if (msg.role === "assistant") {
+      // Render content BEFORE the tool block so the visual order
+      // matches the live SSE flow (stream_text streams content first,
+      // then tool_info / approve_request paints the tool block, then
+      // tool_result fills it in).  Order also matters structurally:
+      // the tool-result message in the NEXT iteration anchors via
+      // lastToolBlock, which the tool-block branch sets last — so
+      // content must run first to avoid clobbering that anchor.
+      if (msg.content) {
+        var el = document.createElement("div");
+        el.className = "msg assistant";
+        var bodyEl = document.createElement("div");
+        bodyEl.className = "msg-body";
+        var rendered = renderMarkdown(msg.content);
+        bodyEl.innerHTML = rendered;
+        el.appendChild(bodyEl);
+        postRenderMarkdown(el);
+        self.messagesEl.appendChild(el);
+        lastToolBlock = null;
+      }
       if (msg.tool_calls && msg.tool_calls.length) {
         if (msg.pending) {
           lastToolBlock = null;
@@ -1096,7 +1128,35 @@ Pane.prototype.replayHistory = function (messages) {
               cmd.textContent = tc.arguments.substring(0, 100);
             }
             div.appendChild(cmd);
+            // Verdict badge — anchor to THIS tool's row (div) rather
+            // than the whole block, so a multi-tool batch with one
+            // flagged call doesn't drift the badge above unrelated
+            // calls.  Same renderVerdictBadge helper as live; pass
+            // judgePending=false because any verdict on replay is
+            // final — no spinner.
+            if (tc.verdict) {
+              div.insertAdjacentHTML(
+                "beforeend",
+                renderVerdictBadge(tc.verdict, false),
+              );
+            }
             block.appendChild(div);
+            // Output-guard finding — defer insertion until the tool
+            // result lands so the warning anchors under the output
+            // (mirrors live showOutputWarning placement).  Stash in
+            // a function-local map keyed by call_id so the
+            // role==="tool" branch below can pick it up; legacy rows
+            // missing tool_call_id are flushed at end-of-replay.
+            if (
+              tc.output_assessment &&
+              tc.output_assessment.risk_level &&
+              tc.output_assessment.risk_level !== "none"
+            ) {
+              pendingAssessments[tc.id || ""] = {
+                assessment: tc.output_assessment,
+                toolDiv: div,
+              };
+            }
           });
           var badge = document.createElement("div");
           badge.setAttribute("role", "status");
@@ -1112,18 +1172,6 @@ Pane.prototype.replayHistory = function (messages) {
           lastToolBlock = block;
         }
       }
-      if (msg.content) {
-        var el = document.createElement("div");
-        el.className = "msg assistant";
-        var bodyEl = document.createElement("div");
-        bodyEl.className = "msg-body";
-        var rendered = renderMarkdown(msg.content);
-        bodyEl.innerHTML = rendered;
-        el.appendChild(bodyEl);
-        postRenderMarkdown(el);
-        self.messagesEl.appendChild(el);
-        lastToolBlock = null;
-      }
     } else if (msg.role === "tool") {
       if (lastToolBlock) {
         var stripped = stripAnsi(msg.content || "").trim();
@@ -1132,27 +1180,76 @@ Pane.prototype.replayHistory = function (messages) {
           /^Denied by user/.test(stripped) ||
           /^Blocked/.test(stripped);
         var isToolError = !!msg.is_error;
+        // Anchor the rendered output to the specific .ts-approval-tool
+        // element matching this result's tool_call_id — mirrors the
+        // live appendToolOutput path so multi-tool batches show
+        // [hdr A][out A][hdr B][out B] rather than [A][B][out A][out B].
+        // Falls back to "before badge" when tool_call_id is absent
+        // (legacy rows pre-dating the wire-format addition).
+        var resultTarget = null;
+        if (msg.tool_call_id) {
+          resultTarget = lastToolBlock.querySelector(
+            '.ts-approval-tool[data-call-id="' +
+              CSS.escape(msg.tool_call_id) +
+              '"]',
+          );
+        }
+        // Cursor-style append: cursor advances after each insert so
+        // the next sibling lands AFTER the previous one.  Fixes the
+        // bug where calling resultTarget.after(node) twice put the
+        // second node BETWEEN resultTarget and the first (the second
+        // .after call was always relative to the same anchor).
+        // Resulting order with all three present:
+        //   [tool div][output][truncation pill][output-warning]
+        var insertCursor = resultTarget;
+        var insertChained = function (node) {
+          if (insertCursor) {
+            insertCursor.after(node);
+            insertCursor = node;
+          } else {
+            var bdg = lastToolBlock.querySelector(".ts-approval-badge");
+            if (bdg) lastToolBlock.insertBefore(node, bdg);
+            else lastToolBlock.appendChild(node);
+          }
+        };
         if (stripped && !isDenied) {
           var media = !isToolError ? tryParseMedia(stripped) : null;
           if (media) {
-            var embed = buildMediaEmbed(media, stripped);
-            var bdg = lastToolBlock.querySelector(".ts-approval-badge");
-            if (bdg) lastToolBlock.insertBefore(embed, bdg);
-            else lastToolBlock.appendChild(embed);
+            insertChained(buildMediaEmbed(media, stripped));
           } else {
             var out = renderToolOutput(stripped, isToolError);
             if (out.textContent.split("\n").length > 10) {
               makeCollapsible(out);
             }
-            var bdg = lastToolBlock.querySelector(".ts-approval-badge");
-            if (bdg) lastToolBlock.insertBefore(out, bdg);
-            else lastToolBlock.appendChild(out);
+            insertChained(out);
+          }
+          // Truncation pill — server marks this when the stored row
+          // hit the 2000-char cap.  Live tool_result events carry full
+          // output so they don't need the indicator.
+          if (msg.truncated) {
+            var pill = document.createElement("span");
+            pill.className = "tool-output-truncated";
+            pill.textContent = "… truncated in storage";
+            pill.title =
+              "The full tool output was sent to the model live; only the first 10000 characters are persisted to the conversation row.";
+            insertChained(pill);
           }
         }
         if (isToolError && !lastToolBlock.classList.contains("denied")) {
           lastToolBlock.classList.add("error");
           appendToolErrorBadge(lastToolBlock);
         }
+        // Output-guard warning — pull the assessment out of the
+        // function-local pendingAssessments map (populated in the
+        // assistant branch).  Skip when the tool result was denied —
+        // the ✗ denied badge already signals the deny path.
+        if (!isDenied && msg.tool_call_id) {
+          var pending = pendingAssessments[msg.tool_call_id];
+          if (pending) {
+            insertChained(_buildOutputWarningEl(pending.assessment));
+            delete pendingAssessments[msg.tool_call_id];
+          }
+        }
       }
       // Tool-channel metacog reminders (tool_error / repeat) attach
       // to the LAST tool message in a batch; on replay we render the
@@ -1165,10 +1262,74 @@ Pane.prototype.replayHistory = function (messages) {
       }
     }
   }
+  // Flush any output_assessments left in the map — these correspond
+  // to assistant tool_calls whose tool result row didn't carry a
+  // tool_call_id (legacy / migrated rows pre-dating the wire-format
+  // addition).  Render the warning under the tool div itself rather
+  // than dropping the safety information silently.
+  var leftoverIds = Object.keys(pendingAssessments);
+  for (var p = 0; p < leftoverIds.length; p++) {
+    var leftover = pendingAssessments[leftoverIds[p]];
+    if (!leftover) continue;
+    leftover.toolDiv.insertAdjacentElement(
+      "afterend",
+      _buildOutputWarningEl(leftover.assessment),
+    );
+  }
   this._attachRetryToLastAssistant();
   this.scrollToBottom();
+  // Focus the input so keyboard users land on the next-action target
+  // after replay finishes — but only when this is the focused pane,
+  // there's no pending approval competing for focus, and an input
+  // element actually exists.  Skipping when not the focused pane
+  // avoids stealing focus from another tab the user is interacting
+  // with while a background replay completes.
+  if (
+    this.id === focusedPaneId &&
+    !this.pendingApproval &&
+    this.inputEl &&
+    !this.busy
+  ) {
+    try {
+      this.inputEl.focus({ preventScroll: true });
+    } catch (_) {
+      this.inputEl.focus();
+    }
+  }
+  // Restore live-region semantics now that the batch build is done.
+  this.messagesEl.removeAttribute("aria-busy");
 };
 
+// Shared output-warning DOM builder — used by both replayHistory
+// (saved-workstream rendering) and the live appendToolOutput path
+// via showOutputWarning.  Single source of truth keeps the two
+// surfaces from drifting on role / class / escape semantics.
+function _buildOutputWarningEl(assessment) {
+  var risk = (assessment && assessment.risk_level) || "medium";
+  var flags = (assessment && assessment.flags) || [];
+  var warning = document.createElement("div");
+  warning.className = "output-warning output-warning-" + risk;
+  // role="status" (polite) rather than "alert" (assertive) — these
+  // are findings, not emergencies; the assertive announcement live
+  // would interrupt the user mid-typing on a high-risk match, which
+  // is more disruptive than informative.
+  warning.setAttribute("role", "status");
+  var labelEl = document.createElement("span");
+  labelEl.className = "output-warning-label";
+  labelEl.textContent = "⚠ " + String(risk).toUpperCase();
+  warning.appendChild(labelEl);
+  if (flags.length) {
+    warning.appendChild(document.createTextNode(" " + flags.join(", ")));
+  }
+  if (assessment && assessment.redacted) {
+    var redacted = document.createElement("span");
+    redacted.className = "output-warning-redacted";
+    redacted.textContent = " (credentials redacted)";
+    warning.appendChild(redacted);
+  }
+  return warning;
+}
+
 Pane.prototype._attachRetryToLastAssistant = function () {
   // Remove any previous retry buttons
   var old = this.messagesEl.querySelectorAll(".msg.assistant .msg-actions");
@@ -1176,6 +1337,21 @@ Pane.prototype._attachRetryToLastAssistant = function () {
   // Find the last assistant message with content and add retry.
   // Reasoning blocks emit as .msg.reasoning (distinct modifier) so the
   // .msg.assistant selector already excludes them — no extra guard needed.
+  //
+  // Skip retry attachment when the most recent semantic turn is
+  // tool-only — last DOM child is a .ts-approval block.  Walk back
+  // past .user-reminder bubbles (added via addToolReminder /
+  // addUserReminder AFTER the .ts-approval block they advise) so the
+  // guard fires correctly even when the tool turn carried a metacog
+  // reminder.  Without this skip, retry lands on a stale prior
+  // assistant content bubble belonging to an earlier turn.
+  var lastChild = this.messagesEl.lastElementChild;
+  while (lastChild && lastChild.classList.contains("user-reminder")) {
+    lastChild = lastChild.previousElementSibling;
+  }
+  if (lastChild && lastChild.classList.contains("ts-approval")) {
+    return;
+  }
   var assistants = this.messagesEl.querySelectorAll(".msg.assistant");
   if (assistants.length) {
     this._addRetryAction(assistants[assistants.length - 1]);
@@ -1512,20 +1688,14 @@ Pane.prototype.showOutputWarning = function (evt) {
     '.ts-approval-tool[data-call-id="' + escapedId + '"]',
   );
   if (!toolDiv) return;
-  var risk = evt.risk_level || "medium";
-  var flags = evt.flags || [];
-  var warning = document.createElement("div");
-  warning.className = "output-warning output-warning-" + risk;
-  warning.setAttribute("role", "alert");
-  warning.innerHTML =
-    '<span class="output-warning-label">\u26a0 ' +
-    escapeHtml(risk.toUpperCase()) +
-    "</span> " +
-    flags.map(escapeHtml).join(", ");
-  if (evt.redacted) {
-    warning.innerHTML +=
-      ' <span class="output-warning-redacted">(credentials redacted)</span>';
-  }
+  // Shared DOM-builder with replayHistory \u2014 single source of truth for
+  // role / class / escape semantics.  Argument shape mirrors the
+  // server-side output_assessment dict (risk_level / flags / redacted).
+  var warning = _buildOutputWarningEl({
+    risk_level: evt.risk_level,
+    flags: evt.flags,
+    redacted: evt.redacted,
+  });
   var nextEl = toolDiv.nextElementSibling;
   if (nextEl && nextEl.classList.contains("tool-output")) {
     nextEl.insertAdjacentElement("afterend", warning);

turnstone/ui/static/style.css

@@ -1631,6 +1631,77 @@ body {
 .ts-approval-tool .tool-diff .diff-warn {
   color: var(--yellow);
 }
+/* memory/recall calls are background metadata — the audit trail is
+   valuable but they crowd the narrative when a workstream contains
+   dozens of them.  Dim by default; full opacity on hover/focus so
+   they remain inspectable without permanently competing for
+   attention.  General-sibling combinator (~) extends the fade past
+   any verdict-badge or output-warning sitting between the tool row
+   and its output, so the whole sub-tree fades together rather than
+   leaving a full-opacity badge stranded next to a dim row. */
+.ts-approval-tool[data-func-name="memory"],
+.ts-approval-tool[data-func-name="recall"] {
+  opacity: 0.55;
+  transition: opacity 120ms ease-out;
+}
+.ts-approval-tool[data-func-name="memory"]:hover,
+.ts-approval-tool[data-func-name="memory"]:focus-within,
+.ts-approval-tool[data-func-name="recall"]:hover,
+.ts-approval-tool[data-func-name="recall"]:focus-within {
+  opacity: 1;
+}
+.ts-approval-tool[data-func-name="memory"] ~ .tool-output,
+.ts-approval-tool[data-func-name="recall"] ~ .tool-output,
+.ts-approval-tool[data-func-name="memory"] ~ .media-embed,
+.ts-approval-tool[data-func-name="recall"] ~ .media-embed,
+.ts-approval-tool[data-func-name="memory"] ~ .output-warning,
+.ts-approval-tool[data-func-name="recall"] ~ .output-warning,
+.ts-approval-tool[data-func-name="memory"] ~ .tool-output-truncated,
+.ts-approval-tool[data-func-name="recall"] ~ .tool-output-truncated {
+  opacity: 0.55;
+  transition: opacity 120ms ease-out;
+}
+/* Reveal on hover OR focus-within across the entire dimmed
+   subtree.  Without :focus-within on the siblings, a keyboard user
+   tabbing into a link or collapsible toggle inside .tool-output
+   sees the content remain dimmed — a11y regression.  Cover the
+   warning + truncation pills too so they fully reveal alongside
+   the result they decorate. */
+.ts-approval-tool[data-func-name="memory"] ~ .tool-output:hover,
+.ts-approval-tool[data-func-name="recall"] ~ .tool-output:hover,
+.ts-approval-tool[data-func-name="memory"] ~ .tool-output:focus-within,
+.ts-approval-tool[data-func-name="recall"] ~ .tool-output:focus-within,
+.ts-approval-tool[data-func-name="memory"] ~ .media-embed:hover,
+.ts-approval-tool[data-func-name="recall"] ~ .media-embed:hover,
+.ts-approval-tool[data-func-name="memory"] ~ .media-embed:focus-within,
+.ts-approval-tool[data-func-name="recall"] ~ .media-embed:focus-within,
+.ts-approval-tool[data-func-name="memory"] ~ .output-warning:hover,
+.ts-approval-tool[data-func-name="recall"] ~ .output-warning:hover,
+.ts-approval-tool[data-func-name="memory"] ~ .output-warning:focus-within,
+.ts-approval-tool[data-func-name="recall"] ~ .output-warning:focus-within,
+.ts-approval-tool[data-func-name="memory"] ~ .tool-output-truncated:hover,
+.ts-approval-tool[data-func-name="recall"] ~ .tool-output-truncated:hover,
+.ts-approval-tool[data-func-name="memory"] ~ .tool-output-truncated:focus-within,
+.ts-approval-tool[data-func-name="recall"] ~ .tool-output-truncated:focus-within {
+  opacity: 1;
+}
+/* Truncation indicator — the persisted tool result is clamped at
+   2000 chars per row in storage; surface that on replay so users
+   know they're seeing a clipped view rather than the full output
+   the live session saw.  Aligns with .output-warning's left gutter
+   (margin-left: 16px) and uses transparent background + dim border
+   so it reads as quiet metadata rather than a foreign element. */
+.tool-output-truncated {
+  display: inline-block;
+  margin-top: 4px;
+  margin-left: 16px;
+  padding: 1px 6px;
+  font-size: 10px;
+  color: var(--fg-dim);
+  background: transparent;
+  border-radius: 3px;
+  border: 1px solid var(--fg-dim);
+}
 /* .ts-approval (chat.css) stacks its children with flex gap, so a
    border-top on the body would float above a strip of container
    background instead of sitting flush against the previous tool row.
@@ -2475,38 +2546,13 @@ audio.media-player {
   }
 }
 
-/* ==========================================================================
-   Verdict badges (intent judge)
-   ========================================================================== */
-.verdict-badge {
-  padding: 4px 10px;
-  font-size: 11px;
-  font-weight: 600;
-  display: flex;
-  align-items: center;
-  gap: 8px;
-  margin-top: 2px;
-  /* No top separator — .ts-verdict-badge (chat.css) shrinks to
-     max-content width, and a 1px border-top would extend only under
-     the badge text and read as a truncated line. */
-}
-.verdict-low {
-  color: var(--green);
-  border-left: 3px solid var(--green);
-}
-.verdict-medium {
-  color: var(--yellow);
-  border-left: 3px solid var(--yellow);
-}
-.verdict-high {
-  color: var(--red);
-  border-left: 3px solid var(--red);
-}
-.verdict-critical {
-  color: var(--red);
-  border-left: 3px solid var(--red);
-  background: rgba(255, 80, 80, 0.05);
-}
+/* Verdict-badge styling lives in the color-mix block further down
+   in this file (.verdict-badge.verdict-{low,medium,high,critical}).
+   The earlier flat-palette duplicate that lived here was removed —
+   two competing .verdict-badge rule sets caused subtle cascade drift
+   (the color-mix block won for backgrounds, the flat one won for the
+   bare .verdict-low/medium/high/critical class names) which made
+   tweaks fragile. Single source of truth now. */
 
 .verdict-detail {
   padding: 6px 12px;

uv.lock

@@ -2533,7 +2533,7 @@ wheels = [
 
 [[package]]
 name = "turnstone"
-version = "1.5.4"
+version = "1.5.6"
 source = { editable = "." }
 dependencies = [
     { name = "alembic" },