chore: bump version to 1.6.9

test: address Copilot review on the leaked-thread guard
- The guard snapshotted live threads by `Thread.ident`, but idents are recycled after a thread exits — a new leaked thread reusing an exited thread's ident would be mistaken for pre-existing and missed (false negative). Snapshot the Thread OBJECTS and compare by identity instead. - Fix the `serve` fixture docstring: the factory returns the ephemeral port, not the server.
2026-08-14 07:52:25 -06:00 · 2026-06-17 18:21:01 -07:00 · 2026-06-17 18:16:34 -07:00 · 2026-06-17 18:16:34 -07:00 · 2026-06-17 18:16:34 -07:00 · 2026-06-17 18:16:34 -07:00
48 changed files with 1893 additions and 502 deletions
@@ -35,6 +35,9 @@ jobs:

  test:
    runs-on: ubuntu-latest
+    # Cap a hung run at 20 min instead of riding GitHub's 6-hour default
+    # (a flaky-hang run otherwise streams -v output for hours).
+    timeout-minutes: 20
    strategy:
      matrix:
        python-version: ["3.11", "3.12", "3.13"]
@@ -51,7 +54,10 @@ jobs:
        with:
          node-version: "24"
      - run: pip install -e ".[test]"
-      - run: pytest tests/ -m "not live" --cov=turnstone --cov-report=term-missing --cov-report=xml -q
+      # -v lists each test id as it starts (pytest prints the nodeid at
+      # logstart), so a hang names the culprit on the last line instead of
+      # riding the job timeout with only a trail of "..." dots.
+      - run: pytest tests/ -m "not live" --cov=turnstone --cov-report=term-missing --cov-report=xml -v
      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
        if: always()
        with:
@@ -60,6 +66,7 @@ jobs:

  test-postgres:
    runs-on: ubuntu-latest
+    timeout-minutes: 20
    services:
      postgres:
        image: postgres:18
@@ -83,7 +90,7 @@ jobs:
        with:
          node-version: "24"
      - run: pip install -e ".[test]"
-      - run: pytest tests/ -m "not live" --storage-backend=postgresql -q
+      - run: pytest tests/ -m "not live" --storage-backend=postgresql -v
        env:
          TURNSTONE_TEST_PG_URL: postgresql+psycopg://postgres:postgres@localhost:5432/turnstone_test

@@ -17,8 +17,10 @@ RUN rm -f /etc/dpkg/dpkg.cfg.d/docker
 # ripgrep is the preferred backend for the search tool — natively bounds
 # per-line, per-file, and per-filesize so pathological inputs (minified
 # bundles, training-data JSONL with multi-MB single records) can't OOM us.
+# ffmpeg transcodes omni STT uploads (browser webm/opus) to the 16 kHz mono
+# WAV the omni chat-audio lane decodes.
 RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-recommends \
-    libpq5 git curl jq man-db manpages procps file ripgrep \
+    libpq5 git curl jq man-db manpages procps file ripgrep ffmpeg \
    && rm -rf /var/lib/apt/lists/*

 # Node.js LTS (for npx-based MCP servers like @modelcontextprotocol/server-github)
@@ -40,7 +40,7 @@ api_key = ""
 smart_approvals = false       # auto-approve high-confidence "approve" LLM verdicts (opt-in)
 confidence_threshold = 0.95   # Smart Approvals auto-approve bar (LLM recommendation=approve)
 max_context_ratio = 0.5       # max % of judge context window for history
-timeout = 60.0                # seconds (generous for local models)
+timeout = 120.0               # seconds (generous for local models)
 read_only_tools = true        # judge can use read_file/list_directory
 cancel_on_approval = false    # stop judging remaining tool calls once user decides
 ```
@@ -71,7 +71,7 @@ All fields are optional. The judge is enabled by default; use `enabled = false`
 --judge / --no-judge           Enable/disable (default: enabled)
 --judge-model MODEL            Model for judge
 --judge-provider PROVIDER      Provider for judge
--judge-timeout SECONDS        LLM judge timeout (default: 60)
+--judge-timeout SECONDS        LLM judge timeout (default: 120)
 --judge-confidence FLOAT       Confidence threshold, 0-1 (default: 0.95)
 ```

@@ -193,9 +193,10 @@ Security hardening blocks access to sensitive paths:

 ### Timeout

-The `timeout` setting (default 60 seconds) is a total budget across all judge
-turns. Time is decremented after each LLM call. If the budget expires mid-turn,
-the judge attempts to parse whatever partial response is available.
+The `timeout` setting (default 120 seconds) applies **per turn**, not as a total
+budget across turns — each of the up to 5 turns gets a fresh budget, so a slow
+earlier turn doesn't starve later ones. If a turn's budget expires, the judge
+attempts to parse whatever partial response is available.

 ---

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "turnstone"
-version = "1.7.0a2"
+version = "1.6.9"
 description = "Multi-node AI orchestration platform with tool use, agent routing, and cluster simulation."
 readme = "README.md"
 license = "Apache-2.0"
@@ -95,7 +95,10 @@ include = [

 [tool.pytest.ini_options]
 testpaths = ["tests"]
-markers = ["live: requires a running LLM backend"]
+markers = [
+    "live: requires a running LLM backend",
+    "allow_thread_leak: test intentionally leaves a background thread running (opts out of the leaked-thread guard)",
+]
 filterwarnings = [
    # mcp v1 deprecates streamablehttp_client for an entry point whose call
    # shape only settles in v2 — adoption rides the deliberate v2 migration
@@ -55,14 +55,14 @@
      "license": "MIT"
    },
    "node_modules/@napi-rs/wasm-runtime": {
-      "version": "1.1.5",
-      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.5.tgz",
-      "integrity": "sha512-AWPoBRJ9tsnVhor4sjO7rkni+7p+2IAEFj6cx06UgP10jkQHqay/36uRV/bFkgrh18D9vb4cr8Q0Pthskgzy+Q==",
+      "version": "1.1.4",
+      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.4.tgz",
+      "integrity": "sha512-3NQNNgA1YSlJb/kMH1ildASP9HW7/7kYnRI2szWJaofaS1hWmbGI4H+d3+22aGzXXN9IJ+n+GiFVcGipJP18ow==",
      "dev": true,
      "license": "MIT",
      "optional": true,
      "dependencies": {
-        "@tybys/wasm-util": "^0.10.2"
+        "@tybys/wasm-util": "^0.10.1"
      },
      "funding": {
        "type": "github",
@@ -409,16 +409,16 @@
      "license": "MIT"
    },
    "node_modules/@vitest/expect": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/expect/-/expect-4.1.9.tgz",
-      "integrity": "sha512-vl/rYsUKcBr3SnQn166+XR5ZQcgMx3DQhFWdfli/cWpLnLUmbxZvyrJZotLFUryib+LtArYMSTJ5RbQ57ZqrlA==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/expect/-/expect-4.1.8.tgz",
+      "integrity": "sha512-h3nDO677RDLEGlBxyQ5CW8RlMThSKSRLUePLOx09gNIWRL40edgA1GCZSZgf1W55MFAG6/Sw14KeaAnqv0NKdQ==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
        "@standard-schema/spec": "^1.1.0",
        "@types/chai": "^5.2.2",
-        "@vitest/spy": "4.1.9",
-        "@vitest/utils": "4.1.9",
+        "@vitest/spy": "4.1.8",
+        "@vitest/utils": "4.1.8",
        "chai": "^6.2.2",
        "tinyrainbow": "^3.1.0"
      },
@@ -427,13 +427,13 @@
      }
    },
    "node_modules/@vitest/mocker": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/mocker/-/mocker-4.1.9.tgz",
-      "integrity": "sha512-EVkXzBjrPGM+cK8/ANWgBrkUCfJfb38/EfTSO8h7pWvKkyPkpWxvR7BkD2MyItMF62C97zAEoqdpUixwR/e+Rw==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/mocker/-/mocker-4.1.8.tgz",
+      "integrity": "sha512-LEiN/xe4OSIbKe9HQIp5OC24agGD9J5CnmMgsLohVVoOPWL9a2sBoR6VBx43jQZb7Kr1l4RCuyCJzcAa0+dojw==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
-        "@vitest/spy": "4.1.9",
+        "@vitest/spy": "4.1.8",
        "estree-walker": "^3.0.3",
        "magic-string": "^0.30.21"
      },
@@ -454,9 +454,9 @@
      }
    },
    "node_modules/@vitest/pretty-format": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/pretty-format/-/pretty-format-4.1.9.tgz",
-      "integrity": "sha512-s0iufns3iIFitdgm+YR7g1whCAaGtXz459VS9/PqyKDEEFgYIhsHOQmXgIgDuYCt7DeQmiZT0Qe2OA2p4ZPu5A==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/pretty-format/-/pretty-format-4.1.8.tgz",
+      "integrity": "sha512-9GasEBxpZ1VYIpqHf/0+YGg121uSNwCKOJqIrTwWP/TB7DmFCiaBpNl3aPZzoLWfWkuqhbH8vJIVobZkvdo2cA==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
@@ -467,13 +467,13 @@
      }
    },
    "node_modules/@vitest/runner": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/runner/-/runner-4.1.9.tgz",
-      "integrity": "sha512-KXLMDtc7oe70+3mJfGrPUWPesswH+3sTxAMAMl8DG7I8IUQT4XW718dY5ID3vPUcmlu27CcKfY4P3h3I29SLJg==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/runner/-/runner-4.1.8.tgz",
+      "integrity": "sha512-EmVxeBAfMJvycdjd6Hm+RbFBbA9fKvo0Kx37hNpBYoYeavH3RNsBXWDooR1mgD52dCrxIIuP7UotpfiwOikvcg==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
-        "@vitest/utils": "4.1.9",
+        "@vitest/utils": "4.1.8",
        "pathe": "^2.0.3"
      },
      "funding": {
@@ -481,14 +481,14 @@
      }
    },
    "node_modules/@vitest/snapshot": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/snapshot/-/snapshot-4.1.9.tgz",
-      "integrity": "sha512-Jc7RKGNBo8Z28WYIm0Niej4xdSPByRf6mU58VpHQkd6Zh05rlnA+twjbK5HyeIGHxrzsc3mJgS43uM0CZKzaIA==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/snapshot/-/snapshot-4.1.8.tgz",
+      "integrity": "sha512-acfZboRmAIf05DEKcBQy33VXojFJjtUdLyo7oOmV9kebb2xdU01UknNiPuPZoJZQyO7DF0gZdTGTpeAzET9QPQ==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
-        "@vitest/pretty-format": "4.1.9",
-        "@vitest/utils": "4.1.9",
+        "@vitest/pretty-format": "4.1.8",
+        "@vitest/utils": "4.1.8",
        "magic-string": "^0.30.21",
        "pathe": "^2.0.3"
      },
@@ -497,9 +497,9 @@
      }
    },
    "node_modules/@vitest/spy": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/spy/-/spy-4.1.9.tgz",
-      "integrity": "sha512-fHpsS6mIi+PiEW+vcRVOMkX1oSaPKne3VOclSFICPcGOmfKgXPU5iAah+wcNcj2xPrCCmfq99IDGf+EojhhvhA==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/spy/-/spy-4.1.8.tgz",
+      "integrity": "sha512-6EevtBp6OZOPF7bmz36HrGMeP3txgVSrgebWxHOafDXGkhIzfXK14f8KF6MuFfgXXUeHxmpD3BQxkV00/3s5mA==",
      "dev": true,
      "license": "MIT",
      "funding": {
@@ -507,13 +507,13 @@
      }
    },
    "node_modules/@vitest/utils": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/@vitest/utils/-/utils-4.1.9.tgz",
-      "integrity": "sha512-A51o8ymO5PpqlWNnBP9ZHPXDIpuMtTLlGSjN7la4US+LJzoUMyhwjA5QXlm39JexgwHKW4Xjs8Z2d3dLCXOeuA==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/@vitest/utils/-/utils-4.1.8.tgz",
+      "integrity": "sha512-uOJamYALNhfJ6iolExyQM40yIQwDqYnkKtQ5VCiSe17E33H0aQ/u+1GlRuz4LZBk6Mm3sg90G9hEbmEt37C1Zg==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
-        "@vitest/pretty-format": "4.1.9",
+        "@vitest/pretty-format": "4.1.8",
        "convert-source-map": "^2.0.0",
        "tinyrainbow": "^3.1.0"
      },
@@ -921,9 +921,9 @@
      }
    },
    "node_modules/obug": {
-      "version": "2.1.3",
-      "resolved": "https://registry.npmjs.org/obug/-/obug-2.1.3.tgz",
-      "integrity": "sha512-9miFgM2OFba7hB+pRgvtV84pYTBaoTHohvmIgiRt6dRIzbwEOIaNaP+dIlGs2fNFoB0SeISs0Jz5WFVRid6Xyg==",
+      "version": "2.1.2",
+      "resolved": "https://registry.npmjs.org/obug/-/obug-2.1.2.tgz",
+      "integrity": "sha512-AWGB9WFcRXOQs48Z/udjI5ZcZMHXwX8XPByNpOydgcGsDLIzjGizhoMWJyKAWze7AVW/2W1i+/gPX4YtKe5cyg==",
      "dev": true,
      "funding": [
        "https://github.com/sponsors/sxzz",
@@ -1200,19 +1200,19 @@
      }
    },
    "node_modules/vitest": {
-      "version": "4.1.9",
-      "resolved": "https://registry.npmjs.org/vitest/-/vitest-4.1.9.tgz",
-      "integrity": "sha512-nE3/LEyc0z87uHYLZebqCUOaJr2hdtuPp7BQ4BosVFnfltxgAvMG08NyrSGlPpOUWvR27c5flSmYFTNr78L9GQ==",
+      "version": "4.1.8",
+      "resolved": "https://registry.npmjs.org/vitest/-/vitest-4.1.8.tgz",
+      "integrity": "sha512-flY6ScbCIt9HThs+C5HS7jvGOB560DJtk/Z15IQROTA6zEy49Nh8T/dofWTQL+n3vswqn87sbJNiuqw1SDp5Ig==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
-        "@vitest/expect": "4.1.9",
-        "@vitest/mocker": "4.1.9",
-        "@vitest/pretty-format": "4.1.9",
-        "@vitest/runner": "4.1.9",
-        "@vitest/snapshot": "4.1.9",
-        "@vitest/spy": "4.1.9",
-        "@vitest/utils": "4.1.9",
+        "@vitest/expect": "4.1.8",
+        "@vitest/mocker": "4.1.8",
+        "@vitest/pretty-format": "4.1.8",
+        "@vitest/runner": "4.1.8",
+        "@vitest/snapshot": "4.1.8",
+        "@vitest/spy": "4.1.8",
+        "@vitest/utils": "4.1.8",
        "es-module-lexer": "^2.0.0",
        "expect-type": "^1.3.0",
        "magic-string": "^0.30.21",
@@ -1240,12 +1240,12 @@
        "@edge-runtime/vm": "*",
        "@opentelemetry/api": "^1.9.0",
        "@types/node": "^20.0.0 || ^22.0.0 || >=24.0.0",
-        "@vitest/browser-playwright": "4.1.9",
-        "@vitest/browser-preview": "4.1.9",
-        "@vitest/browser-webdriverio": "4.1.9",
-        "@vitest/coverage-istanbul": "4.1.9",
-        "@vitest/coverage-v8": "4.1.9",
-        "@vitest/ui": "4.1.9",
+        "@vitest/browser-playwright": "4.1.8",
+        "@vitest/browser-preview": "4.1.8",
+        "@vitest/browser-webdriverio": "4.1.8",
+        "@vitest/coverage-istanbul": "4.1.8",
+        "@vitest/coverage-v8": "4.1.8",
+        "@vitest/ui": "4.1.8",
        "happy-dom": "*",
        "jsdom": "*",
        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0"
@@ -1,17 +1,118 @@
 from __future__ import annotations

+import asyncio
+import contextlib
+import logging
 import os
+import threading
+import time
 from typing import TYPE_CHECKING, Any
 from unittest.mock import MagicMock

 import pytest

+
+def stop_loop_thread(loop: asyncio.AbstractEventLoop, thread: threading.Thread) -> None:
+    """Fully tear down a ``loop.run_forever``-in-a-thread test loop.
+
+    Shuts the loop's default executor down ON the loop (joining its worker
+    threads — the ``asyncio_N`` threads that otherwise leak past the test),
+    then stops the loop, joins the thread, and closes the loop. Use in the
+    ``finally`` of a background-loop fixture so nothing outlives the test.
+    """
+    with contextlib.suppress(Exception):
+        asyncio.run_coroutine_threadsafe(loop.shutdown_default_executor(), loop).result(timeout=5)
+    loop.call_soon_threadsafe(loop.stop)
+    thread.join(timeout=5)
+    with contextlib.suppress(Exception):
+        loop.close()
+
+
+def serve_until_exit(server: Any) -> None:
+    """Run a uvicorn ``Server`` on a fresh event loop until it exits.
+
+    The thread target for an in-thread test upstream: when ``server.serve()``
+    returns (the fixture set ``server.should_exit`` / ``force_exit``), the loop
+    is closed so it doesn't leak past the fixture.
+    """
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
+    try:
+        loop.run_until_complete(server.serve())
+    finally:
+        # Cancel + drain anything the app left pending (e.g. sse_starlette's
+        # shutdown watcher) so loop.close() doesn't warn "Task was destroyed
+        # but it is pending".
+        pending = asyncio.all_tasks(loop)
+        for task in pending:
+            task.cancel()
+        if pending:
+            with contextlib.suppress(Exception):
+                loop.run_until_complete(asyncio.gather(*pending, return_exceptions=True))
+        loop.close()
+
+
 if TYPE_CHECKING:
+    from collections.abc import Iterator
+
    from turnstone.core.mcp_client import MCPClientManager, StaticServerState
    from turnstone.core.mcp_crypto import MCPTokenCipher
    from turnstone.core.oidc import OIDCConfig


+# A background daemon (e.g. title generation) can log into pytest's per-test
+# capture as it is torn down — a benign "I/O operation on closed file" handler
+# error. Don't let the logging module turn that race into noisy stderr
+# tracebacks. (Process-global, test-only — product runtime keeps the default.)
+logging.raiseExceptions = False
+
+
+# Threads a test leaves running after teardown bleed into LATER tests' captured
+# output (the "I/O operation on closed file" heisenbug) and, worse, can wedge
+# the whole run (a leaked event loop / server that never stops). This grace
+# lets a legitimately-finishing quick daemon settle before we judge a leak.
+_THREAD_LEAK_GRACE = 5.0
+
+
+@pytest.fixture(autouse=True)
+def _no_leaked_threads(request: pytest.FixtureRequest) -> Iterator[None]:
+    """Fail a test that leaves a background thread running past teardown.
+
+    Snapshots the live threads at setup; at teardown, gives any NEW thread a
+    short grace to finish, then fails listing those still alive — so a leak is
+    caught here instead of as a heisenbug days later. Opt out with
+    ``@pytest.mark.allow_thread_leak`` (e.g. module-scoped servers in the live
+    suite).
+    """
+    if request.node.get_closest_marker("allow_thread_leak"):
+        yield
+        return
+    # Snapshot the Thread OBJECTS, not their idents: Thread.ident is recycled
+    # after a thread exits, so an ident-based snapshot could mistake a new
+    # leaked thread (reusing an exited thread's ident) for a pre-existing one.
+    before = set(threading.enumerate())
+    yield
+    main = threading.main_thread()
+    current = threading.current_thread()
+    # One deadline shared across all joined threads — a deliberate TOTAL
+    # teardown budget (not per-thread), so a pathological test can't stall
+    # teardown by N×grace. A genuine never-stopping leak exhausts it and fails.
+    deadline = time.monotonic() + _THREAD_LEAK_GRACE
+    leaked = []
+    for t in threading.enumerate():
+        if t in before or t is main or t is current or not t.is_alive():
+            continue
+        t.join(timeout=max(0.0, deadline - time.monotonic()))
+        if t.is_alive():
+            leaked.append(t.name)
+    if leaked:
+        pytest.fail(
+            f"test left background threads running after teardown: {leaked}. "
+            "Stop them in teardown (shut down servers / close event loops / join "
+            "threads), or mark @pytest.mark.allow_thread_leak if intentional."
+        )
+
+
 def make_mcp_token_cipher() -> MCPTokenCipher:
    """Build a single-key MCP token cipher for tests.

@@ -7,6 +7,7 @@ helper code runs end-to-end without a network call.

 from __future__ import annotations

+import shutil
 from unittest.mock import MagicMock

 import pytest
@@ -18,11 +19,16 @@ class _Cfg:
    """Stand-in for ModelConfig — only the fields audio.py reads."""

    def __init__(
-        self, model: str, capabilities: dict | None = None, provider: str = "openai"
+        self,
+        model: str,
+        capabilities: dict | None = None,
+        provider: str = "openai",
+        server_compat: dict | None = None,
    ) -> None:
        self.model = model
        self.capabilities = capabilities or {}
        self.provider = provider
+        self.server_compat = server_compat or {}


 class _FakeConfigStore:
@@ -191,7 +197,8 @@ class TestTranscribe:
        with pytest.raises(audio.AudioBackendError):
            audio.transcribe(registry=reg, alias="voice", data=b"x", filename="a.wav")

-    def test_omni_model_transcribes_via_chat(self):
+    def test_omni_model_transcribes_via_chat(self, monkeypatch):
+        monkeypatch.setattr(audio, "_to_wav_16k_mono", lambda data: data)
        client = MagicMock()
        msg = MagicMock(content="  the transcript  ")
        client.chat.completions.create.return_value = MagicMock(choices=[MagicMock(message=msg)])
@@ -202,15 +209,18 @@ class TestTranscribe:
        assert res.transcript == "the transcript"
        # The dedicated transcription endpoint is NOT used for an omni model.
        client.audio.transcriptions.create.assert_not_called()
-        # Audio rides as an input_audio chat part; format comes from the filename.
        parts = client.chat.completions.create.call_args.kwargs["messages"][0]["content"]
+        # Prompt precedes the audio part — the order Gemma documents for transcription.
+        assert [p["type"] for p in parts] == ["text", "input_audio"]
+        # The clip is transcoded to wav regardless of the upload container.
        audio_part = next(p for p in parts if p["type"] == "input_audio")
-        assert audio_part["input_audio"]["format"] == "webm"
+        assert audio_part["input_audio"]["format"] == "wav"
        # A blank prompt falls back to the omni STT default instruction.
        text_part = next(p for p in parts if p["type"] == "text")
        assert "Only output the transcription" in text_part["text"]

-    def test_omni_prompt_override_used(self):
+    def test_omni_prompt_override_used(self, monkeypatch):
+        monkeypatch.setattr(audio, "_to_wav_16k_mono", lambda data: data)
        client = MagicMock()
        client.chat.completions.create.return_value = MagicMock(
            choices=[MagicMock(message=MagicMock(content="x"))]
@@ -338,3 +348,190 @@ class TestTranscribeCached:
        assert audio.transcribe_cached(**kw) == ""
        audio.transcribe_cached(**kw)
        assert len(calls) == 2  # failure not cached -> retried
+
+
+# ---------------------------------------------------------------------------
+# Omni chat request shaping — transcode + thinking-off + token cap
+# ---------------------------------------------------------------------------
+
+
+class TestOmniChatExtraBody:
+    """``_omni_chat_extra_body`` re-applies what the raw-client STT path skips."""
+
+    _THINKING = {"thinking_mode": "manual", "thinking_param": "enable_thinking"}
+
+    def test_disables_thinking_via_model_param(self):
+        cfg = _Cfg("gemma", dict(self._THINKING))
+        assert audio._omni_chat_extra_body(cfg) == {
+            "chat_template_kwargs": {"enable_thinking": False}
+        }
+
+    def test_thinking_off_wins_over_operator_flag(self):
+        cfg = _Cfg(
+            "gemma",
+            dict(self._THINKING),
+            server_compat={"extra_body": {"chat_template_kwargs": {"enable_thinking": True}}},
+        )
+        # STT never wants reasoning, even if an operator stored thinking on.
+        assert audio._omni_chat_extra_body(cfg)["chat_template_kwargs"]["enable_thinking"] is False
+
+    def test_forwards_operator_server_compat_extra_body(self):
+        cfg = _Cfg(
+            "model",
+            dict(self._THINKING),
+            server_compat={"extra_body": {"reasoning_format": "auto"}},
+        )
+        extra = audio._omni_chat_extra_body(cfg)
+        assert extra["reasoning_format"] == "auto"
+        assert extra["chat_template_kwargs"] == {"enable_thinking": False}
+
+    def test_empty_for_non_thinking_model(self):
+        cfg = _Cfg("omni", {"supports_audio_input": True})
+        assert audio._omni_chat_extra_body(cfg) == {}
+
+
+class TestOmniChatCall:
+    """The omni chat call carries the thinking-off extra_body and a token cap."""
+
+    def test_sends_thinking_off_and_token_cap(self, monkeypatch):
+        monkeypatch.setattr(audio, "_to_wav_16k_mono", lambda data: data)
+        client = MagicMock()
+        client.chat.completions.create.return_value = MagicMock(
+            choices=[MagicMock(message=MagicMock(content="hi"))]
+        )
+        cfg = _Cfg(
+            "gemma-omni",
+            {
+                "supports_audio_input": True,
+                "thinking_mode": "manual",
+                "thinking_param": "enable_thinking",
+            },
+        )
+        audio.transcribe(
+            registry=_FakeRegistry("omni", cfg, client),
+            alias="omni",
+            data=b"webmbytes",
+            filename="speech.webm",
+        )
+        kwargs = client.chat.completions.create.call_args.kwargs
+        assert kwargs["extra_body"]["chat_template_kwargs"]["enable_thinking"] is False
+        assert kwargs["max_tokens"] == audio._OMNI_STT_MAX_TOKENS
+
+
+class TestTranscode:
+    """``_to_wav_16k_mono`` normalizes any container to 16 kHz mono WAV via ffmpeg."""
+
+    def _stereo_wav_44k(self) -> bytes:
+        import io
+        import wave
+
+        buf = io.BytesIO()
+        with wave.open(buf, "wb") as w:
+            w.setnchannels(2)
+            w.setsampwidth(2)
+            w.setframerate(44100)
+            w.writeframes(b"\x00\x01\x00\x01" * 4410)  # 0.1 s of stereo
+        return buf.getvalue()
+
+    @pytest.mark.skipif(shutil.which("ffmpeg") is None, reason="ffmpeg not installed")
+    def test_transcodes_to_16k_mono(self):
+        import io
+        import wave
+
+        out = audio._to_wav_16k_mono(self._stereo_wav_44k())
+        with wave.open(io.BytesIO(out), "rb") as w:
+            assert w.getnchannels() == 1
+            assert w.getframerate() == 16000
+
+    @pytest.mark.skipif(shutil.which("ffmpeg") is None, reason="ffmpeg not installed")
+    def test_undecodable_bytes_raise_backend_error(self):
+        with pytest.raises(audio.AudioBackendError):
+            audio._to_wav_16k_mono(b"this is not audio at all")
+
+    def test_missing_ffmpeg_raises_backend_error(self, monkeypatch):
+        def _no_ffmpeg(*a, **k):
+            raise FileNotFoundError("ffmpeg")
+
+        monkeypatch.setattr(audio.subprocess, "run", _no_ffmpeg)
+        with pytest.raises(audio.AudioBackendError, match="ffmpeg is not installed"):
+            audio._to_wav_16k_mono(b"x")
+
+    def test_invokes_ffmpeg_with_hardened_argv(self, monkeypatch):
+        # Covers the argv shaping even on a CI image without ffmpeg installed.
+        captured = {}
+
+        def _fake_run(cmd, **kwargs):
+            captured["cmd"] = cmd
+            captured["input"] = kwargs.get("input")
+            return MagicMock(returncode=0, stdout=b"RIFF....WAVE", stderr=b"")
+
+        monkeypatch.setattr(audio.subprocess, "run", _fake_run)
+        assert audio._to_wav_16k_mono(b"rawclip") == b"RIFF....WAVE"
+        cmd = captured["cmd"]
+        assert cmd[0] == "ffmpeg"
+        assert captured["input"] == b"rawclip"
+        # SSRF/decompression-bomb hardening + the 16 kHz mono normalization.
+        assert cmd[cmd.index("-protocol_whitelist") + 1] == "pipe"
+        assert "-vn" in cmd
+        assert cmd[cmd.index("-ac") + 1] == "1"
+        assert cmd[cmd.index("-ar") + 1] == "16000"
+        assert cmd[cmd.index("-f") + 1] == "wav"
+
+    def test_nonzero_returncode_raises_backend_error(self, monkeypatch):
+        monkeypatch.setattr(
+            audio.subprocess,
+            "run",
+            lambda *a, **k: MagicMock(returncode=1, stdout=b"", stderr=b"boom"),
+        )
+        with pytest.raises(audio.AudioBackendError, match="Audio transcode failed"):
+            audio._to_wav_16k_mono(b"x")
+
+
+def _stream_chunk(content):
+    return MagicMock(choices=[MagicMock(delta=MagicMock(content=content))])
+
+
+class TestTranscribeStream:
+    """``transcribe_stream`` yields content deltas; resolve/transcode are eager."""
+
+    def test_streams_chat_deltas_with_thinking_off(self, monkeypatch):
+        monkeypatch.setattr(audio, "_to_wav_16k_mono", lambda data: data)
+        client = MagicMock()
+        client.chat.completions.create.return_value = iter(
+            [_stream_chunk("and so"), _stream_chunk(None), _stream_chunk(" my fellow americans")]
+        )
+        cfg = _Cfg(
+            "gemma-omni",
+            {
+                "supports_audio_input": True,
+                "thinking_mode": "manual",
+                "thinking_param": "enable_thinking",
+            },
+        )
+        gen = audio.transcribe_stream(
+            registry=_FakeRegistry("omni", cfg, client), alias="omni", data=b"webmbytes"
+        )
+        # Empty/None deltas are skipped; the rest stream through in order.
+        assert list(gen) == ["and so", " my fellow americans"]
+        kwargs = client.chat.completions.create.call_args.kwargs
+        assert kwargs["stream"] is True
+        assert kwargs["extra_body"]["chat_template_kwargs"]["enable_thinking"] is False
+
+    def test_non_audio_provider_raises_before_streaming(self):
+        client = MagicMock()
+        cfg = _Cfg("gemma", {"supports_audio_input": True}, provider="anthropic-compatible")
+        with pytest.raises(audio.AudioUnavailableError, match="OpenAI-compatible provider"):
+            audio.transcribe_stream(
+                registry=_FakeRegistry("omni", cfg, client), alias="omni", data=b"x"
+            )
+        client.chat.completions.create.assert_not_called()
+
+    def test_whisper_alias_emits_single_chunk(self):
+        client = MagicMock()
+        client.audio.transcriptions.create.return_value = MagicMock(text="  full transcript  ")
+        cfg = _Cfg("whisper-1")  # name inference -> dedicated endpoint, no chat stream
+        gen = audio.transcribe_stream(
+            registry=_FakeRegistry("w", cfg, client), alias="w", data=b"x"
+        )
+        assert list(gen) == ["full transcript"]
+        client.chat.completions.create.assert_not_called()
@@ -85,6 +85,59 @@ def test_emit_created_calls_collector_with_coord_fields() -> None:
    )


+def test_emit_created_seeds_resolved_display_name(tmp_path: Any) -> None:
+    """The collector seed uses the resolved display name (alias > title >
+    name), not the synthetic ``ws.name``.  A coordinator carrying a
+    persisted LLM auto-title (written by ``update_workstream_title``) then
+    shows that title in the live cluster tree instead of reverting to
+    ``ws-xxxx``.  Regression guard for the adapter half of the
+    coordinator-title-persistence fix — the server-side ``_coordinator_rows``
+    half is pinned in test_coordinator_endpoints.py."""
+    from turnstone.core.storage import init_storage, reset_storage
+
+    reset_storage()
+    backend = init_storage("sqlite", path=str(tmp_path / "adapter.db"), run_migrations=False)
+    try:
+        # Titled coordinator → the title surfaces over the placeholder name.
+        backend.register_workstream(
+            "coord-1",
+            node_id="console",
+            user_id="u1",
+            name="ws-c0c0",
+            kind=WorkstreamKind.COORDINATOR,
+        )
+        backend.update_workstream_title("coord-1", "Investigate the title bug")
+        adapter, collector = _make_adapter()
+        adapter.emit_created(_make_ws(name="ws-c0c0"))
+        assert (
+            collector.emit_console_ws_created.call_args.kwargs["name"]
+            == "Investigate the title bug"
+        )
+
+        # A user alias outranks the auto-title (alias > title > name).
+        assert backend.set_workstream_alias("coord-1", "Pinned name")
+        collector.emit_console_ws_created.reset_mock()
+        adapter._fanout_console_ws_created(_make_ws(name="ws-c0c0"))
+        assert collector.emit_console_ws_created.call_args.kwargs["name"] == "Pinned name"
+    finally:
+        reset_storage()
+
+
+def test_coord_display_name_skips_uninitialized_storage() -> None:
+    """_coord_display_name runs on a lifecycle-event path and must NOT trip
+    get_storage()'s SQLite auto-init (a stray .turnstone.db in the CWD) when
+    storage isn't initialized — it falls back to the placeholder ws.name and
+    leaves storage untouched."""
+    from turnstone.console.coordinator_adapter import _coord_display_name
+    from turnstone.core.storage import is_storage_initialized, reset_storage
+
+    reset_storage()
+    assert not is_storage_initialized()
+    assert _coord_display_name(_make_ws(name="ws-abcd")) == "ws-abcd"
+    # The resolution did not auto-initialize storage as a side effect.
+    assert not is_storage_initialized()
+
+
 def test_emit_state_calls_collector_state() -> None:
    """Post-rich-payload, emit_state passes tokens / context_ratio /
    activity / activity_state / content kwargs read from ws.ui's
@@ -65,8 +65,10 @@ from turnstone.core.session_routes import (
    make_history_handler,
    make_list_handler,
    make_open_handler,
+    make_refresh_title_handler,
    make_saved_handler,
    make_send_handler,
+    make_set_title_handler,
 )
 from turnstone.core.workstream import WorkstreamKind

@@ -204,6 +206,16 @@ def _make_client(
                ),
                methods=["POST"],
            ),
+            Route(
+                "/v1/api/workstreams/{ws_id}/refresh-title",
+                make_refresh_title_handler(_coord_endpoint_config),
+                methods=["POST"],
+            ),
+            Route(
+                "/v1/api/workstreams/{ws_id}/title",
+                make_set_title_handler(_coord_endpoint_config),
+                methods=["POST"],
+            ),
            Route(
                "/v1/api/workstreams/{ws_id}/history",
                make_history_handler(_coord_endpoint_config),
@@ -370,6 +382,114 @@ def test_unresolvable_alias_returns_503(storage):
 _COORD_HEADERS = {"X-Test-User": "user-1", "X-Test-Perms": "admin.coordinator"}


+# ---------------------------------------------------------------------------
+# Title verbs — refresh-title (LLM regenerate) + set title (manual alias),
+# ported to coordinators via the lifted make_refresh_title_handler /
+# make_set_title_handler factories so both kinds share one body.
+# ---------------------------------------------------------------------------
+
+
+def test_coord_refresh_title_triggers_regeneration(storage):
+    mgr = _build_mgr(storage)
+    ws = mgr.create(user_id="user-1", name="c1")
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(f"/v1/api/workstreams/{ws.id}/refresh-title", headers=_COORD_HEADERS)
+    assert resp.status_code == 200
+    # The lifted handler resolves the current display name and asks the
+    # live session to regenerate a (different) title in the background.
+    ws.session.request_title_refresh.assert_called_once_with("c1")
+
+
+def test_coord_refresh_title_requires_operator_permission(storage):
+    mgr = _build_mgr(storage)
+    ws = mgr.create(user_id="user-1", name="c1")
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(
+        f"/v1/api/workstreams/{ws.id}/refresh-title",
+        headers={"X-Test-User": "user-1", "X-Test-Perms": "read"},
+    )
+    assert resp.status_code == 403
+    ws.session.request_title_refresh.assert_not_called()
+
+
+def test_coord_refresh_title_unknown_ws_404(storage):
+    mgr = _build_mgr(storage)
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(
+        "/v1/api/workstreams/" + ("0" * 32) + "/refresh-title", headers=_COORD_HEADERS
+    )
+    assert resp.status_code == 404
+
+
+def test_coord_set_title_stores_alias_and_broadcasts(storage):
+    from turnstone.core.memory import get_workstream_display_name
+
+    mgr = _build_mgr(storage)
+    ws = mgr.create(user_id="user-1", name="c1")
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(
+        f"/v1/api/workstreams/{ws.id}/title",
+        json={"title": "Nightly migration sweep"},
+        headers=_COORD_HEADERS,
+    )
+    assert resp.status_code == 200
+    assert resp.json()["title"] == "Nightly migration sweep"
+    # Stored as the alias (outranks the auto-title) ...
+    assert get_workstream_display_name(ws.id) == "Nightly migration sweep"
+    # ... and broadcast live to the dashboard via the session UI.
+    ws.session.ui.on_rename.assert_called_once_with("Nightly migration sweep")
+
+
+def test_coord_set_title_empty_400(storage):
+    mgr = _build_mgr(storage)
+    ws = mgr.create(user_id="user-1", name="c1")
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(
+        f"/v1/api/workstreams/{ws.id}/title", json={"title": "   "}, headers=_COORD_HEADERS
+    )
+    assert resp.status_code == 400
+
+
+def test_coord_set_title_alias_conflict_409(storage):
+    mgr = _build_mgr(storage)
+    first = mgr.create(user_id="user-1", name="c1")
+    second = mgr.create(user_id="user-1", name="c2")
+    storage.set_workstream_alias(first.id, "taken")
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(
+        f"/v1/api/workstreams/{second.id}/title", json={"title": "taken"}, headers=_COORD_HEADERS
+    )
+    assert resp.status_code == 409
+
+
+def test_coord_set_title_rejects_unowned_ws_404(storage):
+    """An admin.coordinator operator can't rename a workstream the coord
+    manager doesn't own (here a cross-kind interactive row) via the coord
+    /title route: set_workstream_alias is a global kind-unscoped UPDATE, so
+    the handler 404s on the in-memory coord lookup BEFORE writing — no
+    silent 200, no cross-kind alias write."""
+    from turnstone.core.memory import get_workstream_display_name
+
+    mgr = _build_mgr(storage)
+    # An interactive-kind row in storage, NOT held by coord_mgr.
+    storage.register_workstream(
+        "i" * 32,
+        node_id="node-1",
+        user_id="user-1",
+        name="interactive-ws",
+        kind=WorkstreamKind.INTERACTIVE,
+    )
+    client = _make_client(storage, coord_mgr=mgr, registry=_fake_registry())
+    resp = client.post(
+        f"/v1/api/workstreams/{'i' * 32}/title",
+        json={"title": "hijacked"},
+        headers=_COORD_HEADERS,
+    )
+    assert resp.status_code == 404
+    # The interactive ws's display name is untouched — the alias write never fired.
+    assert get_workstream_display_name("i" * 32) == "interactive-ws"
+
+
 def test_active_list_row_shape_includes_unified_fields(storage):
    """Stage 2 list-verb-lift parity regression — coord active-list row
    carries the always-include fields (ws_id, name, state, kind,
@@ -2433,6 +2553,54 @@ def test_coordinator_rows_persisted_cluster_wide(storage):
        assert {r["name"] for r in rows} == {"alice-closed", "bob-closed", "orphan-closed"}


+def test_coordinator_rows_surface_persisted_title(storage):
+    """Regression for the coordinator-title-persistence bug.
+
+    The LLM auto-title (``update_workstream_title``) and the user alias
+    (``set_workstream_alias``) live only in ``workstreams.title`` /
+    ``workstreams.alias``.  ``_coordinator_rows`` must resolve the
+    display name ``alias > title > name`` from the persisted row for BOTH
+    lanes — the in-memory ``ws.name`` is the synthetic ``ws-xxxx``
+    placeholder.  Before the fix the read path hardcoded ``title=""`` and
+    used ``ws.name`` / the ``name`` column, so a generated title was
+    written but never read back: it reverted to ``ws-xxxx`` on every
+    dashboard refresh."""
+    from turnstone.console.server import _coordinator_rows
+    from turnstone.core.workstream import WorkstreamKind
+
+    mgr = _build_mgr(storage)
+
+    # In-memory lane: a LIVE coordinator titled after creation.  The
+    # manager assigned the placeholder ``ws.name``; the title is in the DB.
+    live = mgr.create(user_id="alice", name="ws-abcd")
+    storage.update_workstream_title(live.id, "Refactor the auth layer")
+
+    # Persisted lane: a closed coordinator (evicted from the manager)
+    # carrying BOTH a title and a user alias — the alias must win.
+    storage.register_workstream(
+        "f" * 32,
+        node_id="console",
+        user_id="bob",
+        name="ws-f0f0",
+        state="closed",
+        kind=WorkstreamKind.COORDINATOR,
+        parent_ws_id=None,
+    )
+    storage.update_workstream_title("f" * 32, "auto-generated title")
+    assert storage.set_workstream_alias("f" * 32, "Bob's pinned name")
+
+    request = _persisted_rows_request(storage, mgr, "alice", frozenset({"read"}))
+    rows = {r["id"]: r for r in _coordinator_rows(request)}
+
+    live_row = rows[live.id]
+    assert live_row["name"] == "Refactor the auth layer"
+    assert live_row["title"] == "Refactor the auth layer"
+
+    closed_row = rows["f" * 32]
+    assert closed_row["name"] == "Bob's pinned name"  # alias > title > name
+    assert closed_row["title"] == "auto-generated title"
+
+
 # ---------------------------------------------------------------------------
 # Stage 2 P1.5 — coord attachment surface parity with interactive
 # ---------------------------------------------------------------------------
@@ -0,0 +1,66 @@
+"""Tests for turnstone.core.deadline.run_with_deadline.
+
+The load-bearing property is the daemon worker: on timeout or cancel the call
+is abandoned, and the abandoned thread must be a daemon so it can never block
+interpreter exit (the bug that motivated the helper — a non-daemon
+ThreadPoolExecutor worker is joined by concurrent.futures' atexit hook).
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+
+import pytest
+
+from turnstone.core.deadline import (
+    DeadlineCancelledError,
+    DeadlineExceededError,
+    run_with_deadline,
+)
+
+
+def test_returns_result_on_success() -> None:
+    assert run_with_deadline(lambda: 42, timeout=1.0) == 42
+
+
+def test_reraises_callable_exception() -> None:
+    def boom() -> None:
+        raise ValueError("upstream failed")
+
+    with pytest.raises(ValueError, match="upstream failed"):
+        run_with_deadline(boom, timeout=1.0)
+
+
+def test_timeout_returns_promptly_and_abandons_a_daemon_worker() -> None:
+    # The worker sleeps far past the deadline; the call must return promptly
+    # via DeadlineExceededError, and the abandoned worker must be a daemon so
+    # it cannot pin interpreter exit.
+    start = time.monotonic()
+    with pytest.raises(DeadlineExceededError):
+        run_with_deadline(lambda: time.sleep(2.0), timeout=0.2, poll=0.05, thread_name="dl-timeout")
+    assert time.monotonic() - start < 1.0
+    stragglers = [t for t in threading.enumerate() if t.name == "dl-timeout" and not t.daemon]
+    assert stragglers == [], f"non-daemon worker survived: {stragglers}"
+
+
+def test_cancel_returns_promptly() -> None:
+    cancel = threading.Event()
+
+    def _fire() -> None:
+        time.sleep(0.1)
+        cancel.set()
+
+    threading.Thread(target=_fire, daemon=True).start()
+    start = time.monotonic()
+    with pytest.raises(DeadlineCancelledError):
+        run_with_deadline(
+            lambda: time.sleep(2.0),
+            timeout=10.0,
+            cancel_event=cancel,
+            poll=0.05,
+            thread_name="dl-cancel",
+        )
+    assert time.monotonic() - start < 1.0
+    stragglers = [t for t in threading.enumerate() if t.name == "dl-cancel" and not t.daemon]
+    assert stragglers == [], f"non-daemon worker survived: {stragglers}"
@@ -52,13 +52,32 @@ class _Handler(http.server.BaseHTTPRequestHandler):
        pass


-def _serve(handler_cls, ssl_context: ssl.SSLContext | None = None) -> int:
-    """Start a daemon-thread HTTP(S) server on an ephemeral port."""
-    httpd = http.server.HTTPServer(("127.0.0.1", 0), handler_cls)
-    if ssl_context is not None:
-        httpd.socket = ssl_context.wrap_socket(httpd.socket, server_side=True)
-    threading.Thread(target=httpd.serve_forever, daemon=True).start()
-    return httpd.server_address[1]
+@pytest.fixture
+def serve():
+    """Factory that starts an HTTP(S) server on an ephemeral port and returns
+    that port.
+
+    Every server it starts is shut down + its serve_forever thread joined at
+    teardown, so the thread never outlives the test (which would otherwise bleed
+    into a later test's captured output / leak the listener).
+    """
+    started: list[tuple[http.server.HTTPServer, threading.Thread]] = []
+
+    def _factory(handler_cls, ssl_context: ssl.SSLContext | None = None) -> int:
+        httpd = http.server.HTTPServer(("127.0.0.1", 0), handler_cls)
+        if ssl_context is not None:
+            httpd.socket = ssl_context.wrap_socket(httpd.socket, server_side=True)
+        thread = threading.Thread(target=httpd.serve_forever, daemon=True)
+        thread.start()
+        started.append((httpd, thread))
+        return httpd.server_address[1]
+
+    yield _factory
+
+    for httpd, thread in started:
+        httpd.shutdown()  # break the serve_forever loop
+        httpd.server_close()  # release the listening socket
+        thread.join(timeout=5)


@pytest.fixture
@@ -90,29 +109,29 @@ def mtls_setup(tmp_path):
 # ── Plain HTTP (mTLS disabled — the default deployment) ─────────────────────


-def test_plain_http_ok():
+def test_plain_http_ok(serve):
    """Default path: plain probe succeeds, PEM dir never consulted."""
-    port = _serve(_Handler)
+    port = serve(_Handler)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health")
    assert result.returncode == 0, result.stderr


-def test_plain_http_degraded_is_healthy():
+def test_plain_http_degraded_is_healthy(serve):
    """'degraded' (backend down, server up) still counts as container-healthy."""

    class Degraded(_Handler):
        payload = {"status": "degraded"}

-    port = _serve(Degraded)
+    port = serve(Degraded)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health")
    assert result.returncode == 0, result.stderr


-def test_plain_http_bad_status_fails():
+def test_plain_http_bad_status_fails(serve):
    class Bad(_Handler):
        payload = {"status": "error"}

-    port = _serve(Bad)
+    port = serve(Bad)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health")
    assert result.returncode == 1
    assert "unhealthy payload" in result.stderr
@@ -128,40 +147,40 @@ def test_server_down_fails():
 # ── mTLS (tls.enabled) ───────────────────────────────────────────────────────


-def test_mtls_probe_with_pem_dir(mtls_setup):
+def test_mtls_probe_with_pem_dir(mtls_setup, serve):
    """The regression case: mTLS node + plain-HTTP probe URL.

    The plain attempt is rejected at the socket; the script must fall back
    to HTTPS with the node cert as client cert and report healthy."""
    pem_root, server_ctx = mtls_setup
-    port = _serve(_Handler, ssl_context=server_ctx)
+    port = serve(_Handler, ssl_context=server_ctx)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health", pem_root=pem_root)
    assert result.returncode == 0, result.stderr


-def test_mtls_probe_without_pems_fails(mtls_setup):
+def test_mtls_probe_without_pems_fails(mtls_setup, serve):
    """mTLS node but no PEM material on disk: the probe must fail."""
    _, server_ctx = mtls_setup
-    port = _serve(_Handler, ssl_context=server_ctx)
+    port = serve(_Handler, ssl_context=server_ctx)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health", pem_root=None)
    assert result.returncode == 1
    assert "Health check failed" in result.stderr


-def test_mtls_unhealthy_payload_fails(mtls_setup):
+def test_mtls_unhealthy_payload_fails(mtls_setup, serve):
    """A reachable mTLS server with a bad payload is still unhealthy."""
    pem_root, server_ctx = mtls_setup

    class Bad(_Handler):
        payload = {"status": "error"}

-    port = _serve(Bad, ssl_context=server_ctx)
+    port = serve(Bad, ssl_context=server_ctx)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health", pem_root=pem_root)
    assert result.returncode == 1
    assert "unhealthy payload" in result.stderr


-def test_mtls_incomplete_pem_dir_fails(mtls_setup, tmp_path):
+def test_mtls_incomplete_pem_dir_fails(mtls_setup, tmp_path, serve):
    """A PEM dir missing the key is skipped, not half-used."""
    _, server_ctx = mtls_setup
    incomplete = tmp_path / "incomplete-root"
@@ -170,7 +189,7 @@ def test_mtls_incomplete_pem_dir_fails(mtls_setup, tmp_path):
    (d / "fullchain.pem").write_text("not a cert")
    (d / "ca.pem").write_text("not a cert")

-    port = _serve(_Handler, ssl_context=server_ctx)
+    port = serve(_Handler, ssl_context=server_ctx)
    result = run_healthcheck(f"http://127.0.0.1:{port}/health", pem_root=incomplete)
    assert result.returncode == 1

@@ -5,7 +5,6 @@ from __future__ import annotations
 import json
 import threading
 import time
-from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 from typing import Any
 from unittest.mock import MagicMock
@@ -120,8 +119,7 @@ class TestVerdictParsing:
            [{"role": "user", "content": "Run echo hello"}],
            callback_results.append,
        )
-        # Wait for daemon thread
-        time.sleep(0.5)
+        _wait_for(callback_results, 1)

        assert len(heuristics) == 1
        assert heuristics[0].tier == "heuristic"
@@ -184,14 +182,12 @@ class TestErrorHandling:
        provider = _make_mock_provider(side_effect=RuntimeError("API error"))
        judge = _make_judge(provider)

-        with ThreadPoolExecutor(max_workers=1) as pool:
-            result = judge._evaluate_single(
-                _make_item(),
-                [{"role": "user", "content": "test"}],
-                cancel_event=None,
-                executor=pool,
-                client=MagicMock(),
-            )
+        result = judge._evaluate_single(
+            _make_item(),
+            [{"role": "user", "content": "test"}],
+            cancel_event=None,
+            client=MagicMock(),
+        )
        assert result is None

    def test_provider_error_heuristic_still_returned(self):
@@ -209,7 +205,7 @@ class TestErrorHandling:
            [{"role": "user", "content": "test"}],
            callback_results.append,
        )
-        time.sleep(0.5)
+        _wait_for(callback_results, 1)

        assert len(heuristics) == 1
        assert heuristics[0].tier == "heuristic"
@@ -233,20 +229,19 @@ class TestErrorHandling:
            [{"role": "user", "content": "test"}],
            callback_results.append,
        )
-        time.sleep(0.5)
+        _wait_for(callback_results, 1)
        assert len(callback_results) == 1
        assert callback_results[0].tier == "llm_fallback"

-    def test_executor_poison_delivers_fallback(self):
-        """An _ExecutorPoisonedError (a judge-call timeout poisoning the
-        single-worker executor) restarts the executor AND still delivers one
-        fallback for the interrupted item — the twin of the generic-exception
-        path, and load-bearing for Smart Approvals' batch-completeness wait."""
-        from turnstone.core.judge import _ExecutorPoisonedError
-
+    def test_evaluate_single_none_delivers_fallback(self):
+        """A judge-call timeout now surfaces as ``_evaluate_single`` returning
+        None (the executor-poison restart dance is gone); the daemon must still
+        deliver exactly one fallback for that item — Smart Approvals waits on
+        the full verdict set before gating, so a silently-skipped item would
+        block that wait until its timeout."""
        judge = _make_judge()
        judge._evaluate_single = MagicMock(  # type: ignore[method-assign]
-            side_effect=_ExecutorPoisonedError()
+            return_value=None
        )
        callback_results: list[IntentVerdict] = []
        judge.evaluate(
@@ -254,7 +249,7 @@ class TestErrorHandling:
            [{"role": "user", "content": "test"}],
            callback_results.append,
        )
-        time.sleep(0.5)
+        _wait_for(callback_results, 1)
        assert len(callback_results) == 1
        assert callback_results[0].tier == "llm_fallback"

@@ -266,14 +261,12 @@ class TestErrorHandling:
        result_mock.content = ""

        judge = _make_judge(provider)
-        with ThreadPoolExecutor(max_workers=1) as pool:
-            result = judge._evaluate_single(
-                _make_item(),
-                [{"role": "user", "content": "test"}],
-                cancel_event=None,
-                executor=pool,
-                client=MagicMock(),
-            )
+        result = judge._evaluate_single(
+            _make_item(),
+            [{"role": "user", "content": "test"}],
+            cancel_event=None,
+            client=MagicMock(),
+        )
        assert result is None

    def test_empty_content_length_stop_no_retry(self):
@@ -285,14 +278,12 @@ class TestErrorHandling:
        result_mock.finish_reason = "length"

        judge = _make_judge(provider)
-        with ThreadPoolExecutor(max_workers=1) as pool:
-            result = judge._evaluate_single(
-                _make_item(),
-                [{"role": "user", "content": "test"}],
-                cancel_event=None,
-                executor=pool,
-                client=MagicMock(),
-            )
+        result = judge._evaluate_single(
+            _make_item(),
+            [{"role": "user", "content": "test"}],
+            cancel_event=None,
+            client=MagicMock(),
+        )
        assert result is None
        # Should have been called exactly once — no retries
        assert provider.create_completion.call_count == 1
@@ -404,14 +395,12 @@ class TestMultiTurnToolUse:
        provider.create_completion.side_effect = [turn1, turn2]

        judge = _make_judge(provider)
-        with ThreadPoolExecutor(max_workers=1) as pool:
-            verdict = judge._evaluate_single(
-                _make_item(),
-                [{"role": "user", "content": "test"}],
-                cancel_event=None,
-                executor=pool,
-                client=MagicMock(),
-            )
+        verdict = judge._evaluate_single(
+            _make_item(),
+            [{"role": "user", "content": "test"}],
+            cancel_event=None,
+            client=MagicMock(),
+        )
        assert verdict is not None
        assert verdict.tier == "llm"
        assert provider.create_completion.call_count == 2
@@ -454,14 +443,12 @@ class TestMultiTurnToolUse:
        ]

        judge = _make_judge(provider)
-        with ThreadPoolExecutor(max_workers=1) as pool:
-            judge._evaluate_single(
-                _make_item(),
-                [{"role": "user", "content": "test"}],
-                cancel_event=None,
-                executor=pool,
-                client=MagicMock(),
-            )
+        judge._evaluate_single(
+            _make_item(),
+            [{"role": "user", "content": "test"}],
+            cancel_event=None,
+            client=MagicMock(),
+        )
        # Should have called create_completion exactly _JUDGE_MAX_TURNS times
        assert provider.create_completion.call_count == 5

@@ -507,7 +494,7 @@ class TestConfidenceArbitration:
            [{"role": "user", "content": "Run echo hello"}],
            callback_results.append,
        )
-        time.sleep(0.5)
+        _wait_for(callback_results, 1)

        assert len(heuristics) == 1
        assert heuristics[0].confidence == 0.85
@@ -527,7 +514,7 @@ class TestConfidenceArbitration:
            [{"role": "user", "content": "Run echo hello"}],
            callback_results.append,
        )
-        time.sleep(0.5)
+        _wait_for(callback_results, 1)

        assert len(heuristics) == 1
        # LLM verdict is always delivered regardless of confidence comparison
@@ -966,11 +953,7 @@ class TestModelAliasResolution:
            [{"role": "user", "content": "delegate the audit"}],
            callback_results.append,
        )
-        # Wait for daemon thread.
-        for _ in range(20):
-            if callback_results:
-                break
-            time.sleep(0.1)
+        _wait_for(callback_results, 1)

        assert callback_results, "judge never delivered a verdict"
        assert callback_results[0].tier == "llm"
@@ -38,7 +38,7 @@ import uvicorn
 from mcp.server.fastmcp import FastMCP
 from starlette.middleware.base import BaseHTTPMiddleware

-from tests.conftest import make_mcp_token_cipher
+from tests.conftest import make_mcp_token_cipher, serve_until_exit, stop_loop_thread
 from turnstone.core.mcp_client import MCPClientManager
 from turnstone.core.mcp_crypto import MCPTokenStore
 from turnstone.core.mcp_oauth import TokenLookupResult
@@ -187,7 +187,14 @@ def _build_server(port: int, behaviour: dict[str, Any]) -> uvicorn.Server:

    app = mcp.streamable_http_app()
    app.add_middleware(BehaviorMiddleware, behaviour=behaviour)
-    config = uvicorn.Config(app, host="127.0.0.1", port=port, log_level="warning", access_log=False)
+    config = uvicorn.Config(
+        app,
+        host="127.0.0.1",
+        port=port,
+        log_level="warning",
+        access_log=False,
+        timeout_graceful_shutdown=0,  # don't block teardown on a held-open stream
+    )
    return uvicorn.Server(config)


@@ -215,9 +222,7 @@ def upstream():
    server = _build_server(port, behaviour)

    def _run() -> None:
-        loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(loop)
-        loop.run_until_complete(server.serve())
+        serve_until_exit(server)

    t = threading.Thread(target=_run, daemon=True, name="phase6-upstream")
    t.start()
@@ -225,7 +230,11 @@ def upstream():
        _wait_ready(port)
        yield f"http://127.0.0.1:{port}/mcp", behaviour
    finally:
+        # should_exit alone triggers a GRACEFUL shutdown that can wait forever
+        # on a held-open streamable-http stream; force_exit skips that wait so
+        # serve() returns and the upstream thread doesn't leak past the test.
        server.should_exit = True
+        server.force_exit = True
        t.join(timeout=5)


@@ -311,8 +320,7 @@ def running_loop_mgr():

        with contextlib.suppress(Exception):
            asyncio.run_coroutine_threadsafe(_drain(mgr), loop).result(timeout=2)
-        loop.call_soon_threadsafe(loop.stop)
-        thread.join(timeout=2)
+        stop_loop_thread(loop, thread)


 # ---------------------------------------------------------------------------
@@ -34,7 +34,7 @@ from unittest.mock import AsyncMock, MagicMock
 import httpx
 import pytest

-from tests.conftest import make_mcp_token_cipher
+from tests.conftest import make_mcp_token_cipher, stop_loop_thread
 from turnstone.core.mcp_client import (
    MCPClientManager,
    _AuthCapture,
@@ -131,8 +131,7 @@ def running_loop_mgr():

        with contextlib.suppress(Exception):
            asyncio.run_coroutine_threadsafe(_drain(mgr), loop).result(timeout=2)
-        loop.call_soon_threadsafe(loop.stop)
-        thread.join(timeout=2)
+        stop_loop_thread(loop, thread)


 def _run_on_loop(loop: asyncio.AbstractEventLoop, coro: Any) -> Any:
@@ -26,7 +26,7 @@ import uvicorn
 from mcp.server.fastmcp import FastMCP
 from starlette.middleware.base import BaseHTTPMiddleware

-from tests.conftest import make_mcp_token_cipher
+from tests.conftest import make_mcp_token_cipher, serve_until_exit, stop_loop_thread
 from turnstone.core.mcp_client import MCPClientManager
 from turnstone.core.mcp_crypto import MCPTokenStore
 from turnstone.core.mcp_oauth import TokenLookupResult
@@ -130,7 +130,14 @@ def _build_server(port: int, behaviour: dict[str, Any]) -> uvicorn.Server:

    app = mcp.streamable_http_app()
    app.add_middleware(BehaviorMiddleware, behaviour=behaviour)
-    config = uvicorn.Config(app, host="127.0.0.1", port=port, log_level="warning", access_log=False)
+    config = uvicorn.Config(
+        app,
+        host="127.0.0.1",
+        port=port,
+        log_level="warning",
+        access_log=False,
+        timeout_graceful_shutdown=0,  # don't block teardown on a held-open stream
+    )
    return uvicorn.Server(config)


@@ -152,9 +159,7 @@ def upstream():
    server = _build_server(port, behaviour)

    def _run() -> None:
-        loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(loop)
-        loop.run_until_complete(server.serve())
+        serve_until_exit(server)

    t = threading.Thread(target=_run, daemon=True, name="phase7b-prompt-upstream")
    t.start()
@@ -162,7 +167,11 @@ def upstream():
        _wait_ready(port)
        yield f"http://127.0.0.1:{port}/mcp", behaviour
    finally:
+        # should_exit alone triggers a GRACEFUL shutdown that can wait forever
+        # on a held-open streamable-http stream; force_exit skips that wait so
+        # serve() returns and the upstream thread doesn't leak past the test.
        server.should_exit = True
+        server.force_exit = True
        t.join(timeout=5)


@@ -248,8 +257,7 @@ def running_loop_mgr():

        with contextlib.suppress(Exception):
            asyncio.run_coroutine_threadsafe(_drain(mgr), loop).result(timeout=2)
-        loop.call_soon_threadsafe(loop.stop)
-        thread.join(timeout=2)
+        stop_loop_thread(loop, thread)


 def _seed_pool_prompt_map(
@@ -26,7 +26,7 @@ import uvicorn
 from mcp.server.fastmcp import FastMCP
 from starlette.middleware.base import BaseHTTPMiddleware

-from tests.conftest import make_mcp_token_cipher
+from tests.conftest import make_mcp_token_cipher, serve_until_exit, stop_loop_thread
 from turnstone.core.mcp_client import MCPClientManager
 from turnstone.core.mcp_crypto import MCPTokenStore
 from turnstone.core.mcp_oauth import TokenLookupResult
@@ -139,7 +139,14 @@ def _build_server(port: int, behaviour: dict[str, Any]) -> uvicorn.Server:

    app = mcp.streamable_http_app()
    app.add_middleware(BehaviorMiddleware, behaviour=behaviour)
-    config = uvicorn.Config(app, host="127.0.0.1", port=port, log_level="warning", access_log=False)
+    config = uvicorn.Config(
+        app,
+        host="127.0.0.1",
+        port=port,
+        log_level="warning",
+        access_log=False,
+        timeout_graceful_shutdown=0,  # don't block teardown on a held-open stream
+    )
    return uvicorn.Server(config)


@@ -161,9 +168,7 @@ def upstream():
    server = _build_server(port, behaviour)

    def _run() -> None:
-        loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(loop)
-        loop.run_until_complete(server.serve())
+        serve_until_exit(server)

    t = threading.Thread(target=_run, daemon=True, name="phase7b-resource-upstream")
    t.start()
@@ -171,7 +176,11 @@ def upstream():
        _wait_ready(port)
        yield f"http://127.0.0.1:{port}/mcp", behaviour
    finally:
+        # should_exit alone triggers a GRACEFUL shutdown that can wait forever
+        # on a held-open streamable-http stream; force_exit skips that wait so
+        # serve() returns and the upstream thread doesn't leak past the test.
        server.should_exit = True
+        server.force_exit = True
        t.join(timeout=5)


@@ -257,8 +266,7 @@ def running_loop_mgr():

        with contextlib.suppress(Exception):
            asyncio.run_coroutine_threadsafe(_drain(mgr), loop).result(timeout=2)
-        loop.call_soon_threadsafe(loop.stop)
-        thread.join(timeout=2)
+        stop_loop_thread(loop, thread)


 def _seed_pool_resource_map(
@@ -25,7 +25,7 @@ from unittest.mock import AsyncMock, MagicMock

 import pytest

-from tests.conftest import make_mcp_token_cipher
+from tests.conftest import make_mcp_token_cipher, stop_loop_thread
 from turnstone.core.mcp_client import MCPClientManager, PoolEntryState
 from turnstone.core.mcp_crypto import MCPTokenStore
 from turnstone.core.storage._sqlite import SQLiteBackend
@@ -123,8 +123,7 @@ def running_loop_mgr():

        with contextlib.suppress(Exception):
            asyncio.run_coroutine_threadsafe(_drain(mgr), loop).result(timeout=2)
-        loop.call_soon_threadsafe(loop.stop)
-        thread.join(timeout=2)
+        stop_loop_thread(loop, thread)


 def _run_on_loop(loop: asyncio.AbstractEventLoop, coro: Any) -> Any:
@@ -220,6 +220,26 @@ class TestEvaluateFailurePaths:
        # Cancel should return promptly, well below the 10s timeout.
        assert elapsed < 2.0, f"cancel returned in {elapsed:.2f}s, expected < 2.0s"

+    def test_timeout_leaves_no_nondaemon_straggler(self) -> None:
+        # Regression: evaluate() abandons a slow upstream call on timeout, but
+        # the worker must be a *daemon* so it can never pin interpreter exit.
+        # The old ThreadPoolExecutor worker was non-daemon and got joined by
+        # concurrent.futures' atexit hook, hanging the whole test run at
+        # shutdown.  See turnstone/core/deadline.py.
+        judge = _make_judge(
+            content='{"risk_level":"medium","flags":[],"reasoning":""}',
+            timeout=1.0,
+            delay=5.0,
+        )
+        v = judge.evaluate("payload", call_id="c1")
+        assert v.error == "timeout"
+        stragglers = [
+            t
+            for t in threading.enumerate()
+            if t.name.startswith("output-guard-judge") and not t.daemon
+        ]
+        assert stragglers == [], f"non-daemon worker survived evaluate(): {stragglers}"
+

 class TestAliasResolution:
    def test_unknown_alias_falls_back_to_session_model(self) -> None:
@@ -19,7 +19,8 @@ class TestSuggestProfile:
        p = suggest_profile("vllm", "google/gemma-4-31B-it")
        assert p["capabilities"]["thinking_mode"] == "manual"
        assert p["capabilities"]["thinking_param"] == "enable_thinking"
-        assert p["server_compat"]["extra_body"]["skip_special_tokens"] is False
+        # No bug-workaround extra_body — gemma-4 needs only the thinking param.
+        assert "extra_body" not in p["server_compat"]

    def test_vllm_gemma3(self) -> None:
        p = suggest_profile("vllm", "google/gemma-3-27b-it")
@@ -147,14 +148,14 @@ class TestMergeServerCompat:
        result = merge_server_compat(None, {"extra_body": {"skip_special_tokens": False}})
        assert result == {"skip_special_tokens": False}

-    def test_full_vllm_gemma_compat_no_base(self) -> None:
-        """vLLM workaround forwards on its own."""
+    def test_full_server_compat_extra_body_no_base(self) -> None:
+        """A server workaround (e.g. llama.cpp reasoning_format) forwards on its own."""
        compat = {
-            "server_type": "vllm",
-            "extra_body": {"skip_special_tokens": False},
+            "server_type": "llama.cpp",
+            "extra_body": {"reasoning_format": "auto"},
        }
        result = merge_server_compat(None, compat)
-        assert result == {"skip_special_tokens": False}
+        assert result == {"reasoning_format": "auto"}

    def test_operator_chat_template_kwargs_only(self) -> None:
        """Operator can set chat_template_kwargs explicitly without seeding the base."""
@@ -210,21 +211,27 @@ class TestEndToEndRequestShaping:
    """Compose both layers — session builds extra_params, provider applies thinking."""

    def test_vllm_gemma_full_flow(self) -> None:
-        """Session forwards server workarounds, provider adds thinking param."""
+        """Gemma now needs only the thinking param — no server workaround."""
        caps = ModelCapabilities(thinking_mode="manual", thinking_param="enable_thinking")
-        server_compat = {
-            "server_type": "vllm",
-            "extra_body": {"skip_special_tokens": False},
-        }
+        server_compat = {"server_type": "vllm"}
        # 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": {"enable_thinking": True}}
+
+    def test_server_workaround_composes_with_thinking(self) -> None:
+        """A top-level server workaround forwards alongside the injected thinking param."""
+        caps = ModelCapabilities(thinking_mode="manual", thinking_param="enable_thinking")
+        compat = {"server_type": "llama.cpp", "extra_body": {"reasoning_format": "auto"}}
+        extra_body = dict(merge_server_compat(None, compat))
+        OpenAIChatCompletionsProvider._apply_thinking_mode(extra_body, caps)
+
        assert extra_body == {
            "chat_template_kwargs": {"enable_thinking": True},
-            "skip_special_tokens": False,
+            "reasoning_format": "auto",
        }

    def test_granite_thinking_key(self) -> None:
@@ -292,7 +299,7 @@ class TestProbeIntegration:
        assert result["server_type"] == "vllm"
        assert result["suggested_capabilities"]["thinking_mode"] == "manual"
        assert result["suggested_capabilities"]["thinking_param"] == "enable_thinking"
-        assert result["suggested_server_compat"]["extra_body"]["skip_special_tokens"] is False
+        assert "extra_body" not in result["suggested_server_compat"]

    def test_detect_non_thinking_no_suggested_capabilities(self) -> None:
        """Non-thinking vLLM model gets server_compat but no capabilities suggestion."""
@@ -150,6 +150,27 @@ def _send_with_mocks(session, responses, mock_execute, **extra_patches):
        yield save_msg


+def _capturing_thread_cls():
+    """Return a no-op ``threading.Thread`` stand-in plus the list it records
+    each constructed thread's ``target`` into.
+
+    Patched over ``session.threading.Thread`` so a test can assert WHICH
+    callable was scheduled (e.g. ``_generate_title``) without the thread
+    actually running — ``start()`` is a no-op, so no background LLM call
+    fires.
+    """
+    started: list = []
+
+    class _CaptureThread:
+        def __init__(self, *a, target=None, **kw):
+            started.append(target)
+
+        def start(self):
+            pass
+
+    return _CaptureThread, started
+
+
 def _user_pending(session) -> list[tuple[str, str]]:
    """Return user-channel queued nudges as ``(type, text)`` tuples.

@@ -1010,6 +1031,66 @@ class TestTitleRetry:
        # Restore for cleanup
        session._ws_id = original_ws_id

+    def test_title_fires_after_send_not_after_tool_free_turn(self, tmp_db):
+        """Auto-title fires right after the user turn is recorded, BEFORE
+        tools run — it no longer waits for a tool-call-free assistant
+        turn.  Coordinators spend nearly every turn in tool calls and may
+        never reach that terminal text turn, so the old end-of-turn
+        trigger almost never fired for them (the timing half of the
+        coordinator-title bug)."""
+        session = _make_session()
+        assert session._title_generated is False
+        # The assistant's opening turn is ALL tool calls — under the old
+        # trigger no title would generate until a later text-only turn.
+        responses = [
+            {
+                "role": "assistant",
+                "content": "working",
+                "tool_calls": [
+                    {
+                        "id": "c1",
+                        "type": "function",
+                        "function": {"name": "echo", "arguments": "{}"},
+                    }
+                ],
+            },
+            {"role": "assistant", "content": "done"},
+        ]
+        capture_cls, started = _capturing_thread_cls()
+
+        def mock_execute(_tool_calls):
+            # The title must already be scheduled by the time tools run.
+            assert session._title_generated is True
+            return [("c1", "ok")], None
+
+        with (
+            _send_with_mocks(session, responses, mock_execute),
+            patch("turnstone.core.session.threading.Thread", capture_cls),
+        ):
+            session.send("refactor the auth layer")
+
+        assert session._title_generated is True
+        assert session._generate_title in started
+
+    def test_title_not_generated_for_blank_or_wake_send(self, tmp_db):
+        """Blank input and synthetic wake sends don't burn the one-shot
+        auto-title — ``_generate_title`` needs first-user-message text,
+        and a wake carries none."""
+        capture_cls, started = _capturing_thread_cls()
+
+        def mock_execute(_tool_calls):
+            return [], None
+
+        for user_input, kwargs in (("   ", {}), ("a real message", {"from_wake": True})):
+            session = _make_session()
+            with (
+                _send_with_mocks(session, [{"role": "assistant", "content": "ok"}], mock_execute),
+                patch("turnstone.core.session.threading.Thread", capture_cls),
+            ):
+                session.send(user_input, **kwargs)
+            assert session._generate_title not in started
+            assert session._title_generated is False
+

 class TestLiveConfigUpdate:
    """ConfigStore-backed sessions pick up settings changes at point-of-use."""
@@ -602,6 +602,27 @@ def test_step7_tab_menu_wired_per_persona() -> None:
    )


+def test_coordinator_tab_menu_enables_title_verbs() -> None:
+    """Coordinators carry LLM/auto titles like interactive workstreams, so
+    their tab dropdown must surface Refresh/Edit title — convTabMenu's
+    ``titleVerbs`` block, POSTed to the console-origin coord
+    ``refresh-title`` / ``title`` routes via the base-aware lane (default
+    base ""). Scoped to the coordinator registerType block so it can't
+    pass on the interactive pane's long-standing ``titleVerbs``."""
+    shell = _SHELL_JS.read_text(encoding="utf-8")
+    start = shell.index('registerType("coordinator"')
+    tail = shell[start:]
+    nxt = tail.find("registerType(", 1)  # bound at the next pane registration
+    coord_block = tail[:nxt] if nxt != -1 else tail
+    assert "pane._ctl.closeSession()" in coord_block, (
+        "sanity: the extracted block is the coordinator pane"
+    )
+    assert "convTabMenu(" in coord_block, "the coordinator pane must wire a tab menu"
+    assert "titleVerbs: true" in coord_block, (
+        "the coordinator tab menu must enable titleVerbs (Refresh/Edit title)"
+    )
+
+
 def test_tab_menu_base_aware_verb_lane() -> None:
    """Lifecycle round 2: a proxied interactive pane's tab menu must act on the
    pane's OWN transport base, not the console origin — the globals lane only
@@ -96,10 +96,8 @@ def _make_flaky_client(monkeypatch, failures: int):
    """TLSClient whose CA fetch fails ``failures`` times, then succeeds.

    Returns (client, calls, sleeps) — mutable lists recording each CA-fetch
-    attempt and each backoff delay (asyncio.sleep is stubbed out).
+    attempt and each backoff delay (the client's backoff sleep is stubbed).
    """
-    import asyncio
-
    from turnstone.core.tls import TLSClient

    client = TLSClient(
@@ -119,11 +117,14 @@ def _make_flaky_client(monkeypatch, failures: int):
        pass

    async def fake_sleep(delay):
+        # Stub the client's own _sleep seam, NOT the global asyncio.sleep:
+        # patching the global also intercepts any concurrent task sharing the
+        # event loop, which corrupted a background poller and hung CI.
        sleeps.append(delay)

    monkeypatch.setattr(client, "_fetch_ca_cert", flaky_fetch)
    monkeypatch.setattr(client, "_request_cert", ok_request)
-    monkeypatch.setattr(asyncio, "sleep", fake_sleep)
+    monkeypatch.setattr(client, "_sleep", fake_sleep)
    return client, calls, sleeps


@@ -175,8 +176,6 @@ async def test_init_retries_exhausted_raises(monkeypatch):
@pytest.mark.anyio
 async def test_init_retries_discovery_failure(monkeypatch):
    """Console discovery (not-yet-registered console) is retried too."""
-    import asyncio
-
    from turnstone.core.tls import TLSClient

    client = TLSClient(storage=get_storage(), hostnames=["node-1"])
@@ -191,10 +190,13 @@ async def test_init_retries_discovery_failure(monkeypatch):
    async def ok():
        pass

+    async def fake_sleep(_delay):
+        pass
+
    monkeypatch.setattr(client, "_discover_console_url", flaky_discover)
    monkeypatch.setattr(client, "_fetch_ca_cert", ok)
    monkeypatch.setattr(client, "_request_cert", ok)
-    monkeypatch.setattr(asyncio, "sleep", lambda _: ok())
+    monkeypatch.setattr(client, "_sleep", fake_sleep)

    await client.init(attempts=2)
    assert attempts == [1, 2]
@@ -0,0 +1,40 @@
+"""Tests for turnstone.console.server._validate_regex_pattern.
+
+The catastrophic-backtracking branch is verified by simulating the deadline
+firing rather than running a real ReDoS regex — a genuine runaway pattern would
+leave a CPU-pinned daemon worker for the rest of the suite.  The daemon-abandon
+mechanism itself is covered in tests/test_deadline.py.
+"""
+
+from __future__ import annotations
+
+from turnstone.console.server import _validate_regex_pattern
+from turnstone.core.deadline import DeadlineExceededError
+
+
+def test_valid_pattern_returns_none() -> None:
+    assert _validate_regex_pattern(r"\d{3}-\d{4}") is None
+
+
+def test_invalid_pattern_returns_error() -> None:
+    msg = _validate_regex_pattern(r"(unclosed")
+    assert msg is not None
+    assert msg.startswith("Invalid regex")
+
+
+def test_catastrophic_backtracking_returns_message(monkeypatch) -> None:
+    def _deadline(*_args, **_kwargs):
+        raise DeadlineExceededError
+
+    monkeypatch.setattr("turnstone.console.server.run_with_deadline", _deadline)
+    # The pattern is arbitrary — run_with_deadline is stubbed to raise, so the
+    # probe never runs; a real backtracking literal here would only trip CodeQL.
+    assert _validate_regex_pattern(r"\w+") == "Regex appears to have catastrophic backtracking"
+
+
+def test_probe_error_returns_generic_message(monkeypatch) -> None:
+    def _err(*_args, **_kwargs):
+        raise RuntimeError("boom")
+
+    monkeypatch.setattr("turnstone.console.server.run_with_deadline", _err)
+    assert _validate_regex_pattern(r"abc") == "Regex caused an error during test"
@@ -31,16 +31,17 @@ from turnstone.core.session_routes import (
    make_export_handler,
    make_history_handler,
    make_open_handler,
+    make_refresh_title_handler,
    make_retry_handler,
    make_rewind_handler,
+    make_set_title_handler,
 )
 from turnstone.core.storage._sqlite import SQLiteBackend
 from turnstone.core.workstream import WorkstreamKind
 from turnstone.server import (
+    _interactive_tenant_check,
    delete_workstream_endpoint,
    list_interface_settings,
-    refresh_workstream_title,
-    set_workstream_title,
    update_interface_setting,
 )

@@ -113,6 +114,18 @@ def delete_client(_inject_storage):

@pytest.fixture
 def title_client(_inject_storage):
+    # Build the lifted refresh/set-title handlers the same way server.py
+    # wires the interactive bundle — same SessionEndpointConfig
+    # (manager_lookup + _interactive_tenant_check) so the tests exercise
+    # the production resolution path (mgr fast-path → storage ownership).
+    mock_mgr = MagicMock()
+    cfg = SessionEndpointConfig(
+        permission_gate=None,
+        manager_lookup=lambda _r: (mock_mgr, None),
+        tenant_check=_interactive_tenant_check,
+        not_found_label="Workstream not found",
+        audit_action_prefix="workstream",
+    )
    app = Starlette(
        routes=[
            Mount(
@@ -120,12 +133,12 @@ def title_client(_inject_storage):
                routes=[
                    Route(
                        "/api/workstreams/{ws_id}/title",
-                        set_workstream_title,
+                        make_set_title_handler(cfg),
                        methods=["POST"],
                    ),
                    Route(
                        "/api/workstreams/{ws_id}/refresh-title",
-                        refresh_workstream_title,
+                        make_refresh_title_handler(cfg),
                        methods=["POST"],
                    ),
                ],
@@ -133,7 +146,6 @@ def title_client(_inject_storage):
        ],
        middleware=[Middleware(_InjectAuthMiddleware)],
    )
-    mock_mgr = MagicMock()
    app.state.workstreams = mock_mgr
    return TestClient(app), mock_mgr

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

-__version__ = "1.7.0a2"
+__version__ = "1.6.9"
@@ -1022,8 +1022,8 @@ def main() -> None:
        "--judge-timeout",
        dest="judge_timeout",
        type=float,
-        default=60.0,
-        help="LLM judge timeout in seconds (default: 60)",
+        default=120.0,
+        help="LLM judge timeout in seconds (default: 120)",
    )
    judge_group.add_argument(
        "--judge-confidence",
@@ -94,6 +94,11 @@ class ClusterCollector:
        self._nodes: dict[str, NodeSnapshot] = {}
        self._running = False
        self._threads: list[threading.Thread] = []
+        # Wakes the discovery loop out of its inter-scan sleep so ``stop()``
+        # can join it promptly instead of blocking up to ``discovery_interval``
+        # (a long interval would otherwise leave the thread sleeping past
+        # join's timeout — a leaked background thread).
+        self._discovery_wake = threading.Event()

        # SSE fan-out to browser clients
        self._listeners: list[queue.Queue[dict[str, Any]]] = []
@@ -135,6 +140,9 @@ class ClusterCollector:
    def start(self) -> None:
        """Start background threads."""
        self._running = True
+        # Clear the shutdown wake so a restarted collector (stop() set it) sleeps
+        # the full interval again instead of busy-spinning the discovery loop.
+        self._discovery_wake.clear()
        # Subscribe to the ``services`` channel for reactive node discovery.
        # NOTIFY-driven wake-ups bring new-node visibility from up-to-60 s
        # (next discovery tick) down to ~500 ms on Postgres; the 60 s
@@ -161,6 +169,7 @@ class ClusterCollector:
        its ``finally`` cleanup (cancel tasks, close AsyncClient).
        """
        self._running = False
+        self._discovery_wake.set()  # wake the discovery loop out of its sleep
        if self._notify_unsubscribe is not None:
            with contextlib.suppress(Exception):
                self._notify_unsubscribe()
@@ -417,7 +426,9 @@ class ClusterCollector:
                pass  # already logged by storage layer
            except Exception:
                log.exception("Node discovery error")
-            time.sleep(self._discovery_interval)
+            # Interruptible inter-scan sleep — ``stop()`` sets the event to
+            # wake us immediately instead of blocking out the full interval.
+            self._discovery_wake.wait(self._discovery_interval)

    def _discover_nodes(self) -> None:
        """Query the service registry and update the node map."""
@@ -23,6 +23,8 @@ from turnstone.core.child_event_bus import ChildEventBus
 from turnstone.core.child_source import ClusterChildSource
 from turnstone.core.children_registry import ChildrenRegistry
 from turnstone.core.log import get_logger
+from turnstone.core.memory import get_workstream_display_name, get_workstream_display_names
+from turnstone.core.storage import is_storage_initialized
 from turnstone.core.workstream import Workstream, WorkstreamKind, WorkstreamState

 if TYPE_CHECKING:
@@ -37,6 +39,28 @@ if TYPE_CHECKING:
 log = get_logger(__name__)


+def _coord_display_name(ws: Workstream) -> str:
+    """Resolve a coordinator's display name (``alias > title > name``).
+
+    ``ws.name`` is the synthetic ``ws-xxxx`` placeholder; the persisted
+    auto-title (``update_workstream_title``) and user alias live only in
+    the DB.  Seeding the collector with the resolved name means a
+    rehydrated coordinator shows its title in the live cluster tree
+    immediately, rather than reverting to ``ws-xxxx`` until a (for
+    coordinators, rarely-firing) ``on_rename`` event arrives.
+
+    Skips the DB read when storage isn't initialized: this runs on a
+    lifecycle-event path, and a display-name resolution must never trip
+    ``get_storage``'s SQLite auto-init side effect (a stray
+    ``.turnstone.db``) before the host has called ``init_storage`` (the
+    real cluster always does so at startup — this only bites early /
+    test call paths). The placeholder ``ws.name`` is the right fallback.
+    """
+    if not is_storage_initialized():
+        return ws.name
+    return get_workstream_display_name(ws.id) or ws.name
+
+
 class CoordinatorAdapter:
    """Bridges SessionManager to the console's coordinator transport."""

@@ -132,7 +156,7 @@ class CoordinatorAdapter:
        try:
            self._collector.emit_console_ws_created(
                ws.id,
-                name=ws.name,
+                name=_coord_display_name(ws),
                user_id=ws.user_id,
                kind=ws.kind.value,
                state=ws.state.value,
@@ -466,11 +490,16 @@ class CoordinatorAdapter:
        # creates happened before the collector was wired up and their
        # rows never showed on the snapshot. (Coord-specific — interactive
        # has no analogous pseudo-node.)
-        for ws in mgr.list_all():
+        coords = mgr.list_all()
+        # One round-trip for every coordinator's display name instead of a
+        # per-``ws`` ``_coord_display_name`` lookup (N+1); cold path, but
+        # the bulk helper is right there.
+        seed_names = get_workstream_display_names([ws.id for ws in coords])
+        for ws in coords:
            try:
                collector.emit_console_ws_created(
                    ws.id,
-                    name=ws.name,
+                    name=seed_names.get(ws.id) or ws.name,
                    user_id=ws.user_id or "",
                    kind=WorkstreamKind.COORDINATOR.value,
                    state=ws.state.value,
@@ -55,6 +55,8 @@ from turnstone.core.auth import (
    jwt_version_slot,
    require_permission,
 )
+from turnstone.core.deadline import DeadlineExceededError, run_with_deadline
+from turnstone.core.memory import get_workstream_display_names
 from turnstone.core.rendezvous import NoAvailableNodeError
 from turnstone.core.session_replay import session_replay_preamble
 from turnstone.core.session_routes import (
@@ -74,9 +76,11 @@ from turnstone.core.session_routes import (
    make_history_handler,
    make_list_handler,
    make_open_handler,
+    make_refresh_title_handler,
    make_retry_handler,
    make_rewind_handler,
    make_send_handler,
+    make_set_title_handler,
    make_unified_saved_handler,
    register_coord_verbs,
    register_session_routes,
@@ -852,6 +856,14 @@ def _coordinator_rows(request: Request) -> list[dict[str, Any]]:
    In-memory wins on ws_id conflict so live state stays authoritative
    for active sessions.

+    Display name resolves ``alias > title > name`` from the persisted
+    row for BOTH lanes.  ``ws.name`` on the in-memory Workstream is the
+    synthetic ``ws-xxxx`` placeholder; the LLM auto-title
+    (``update_workstream_title``) and the user alias
+    (``set_workstream_alias``) live only in the DB, so without the
+    persisted lookup the live lane would show ``ws-xxxx`` and the
+    auto-title would never survive a dashboard refresh.
+
    Trusted-team visibility (post-#400): the cluster dashboard shows
    every coordinator regardless of caller identity; ``user_id`` is
    surfaced on each row as display metadata.
@@ -869,6 +881,61 @@ def _coordinator_rows(request: Request) -> list[dict[str, Any]]:
        val = getattr(sess, name, "") if sess else ""
        return val if isinstance(val, str) else ""

+    # Persisted coordinator rows serve two purposes: (1) surface
+    # closed / error / deleted coordinators the manager has already
+    # evicted from ``self._workstreams``, and (2) supply the persisted
+    # display name (``alias > title > name``) for the LIVE coordinators
+    # too — ``ws.name`` is the synthetic placeholder.  Cluster-wide
+    # (trusted-team visibility).  Indexed by ws_id so both lanes resolve
+    # the same way.
+    storage = getattr(request.app.state, "auth_storage", None)
+    persisted: list[Any] = []
+    if storage is not None:
+        try:
+            persisted = storage.list_workstreams(
+                kind=WorkstreamKind.COORDINATOR,
+                user_id=None,
+                limit=200,
+            )
+        except Exception:
+            log.debug("cluster_workstreams.coord_persisted_failed", exc_info=True)
+            persisted = []
+    # SQLAlchemy Row — access via _mapping so future SELECT reorders /
+    # new columns don't silently corrupt the projection (per the
+    # storage-protocol guidance on list_workstreams).  Test doubles must
+    # expose the same ._mapping attribute.
+    meta: dict[str, Any] = {}
+    for row in persisted:
+        m = row._mapping
+        rid = m.get("ws_id") or ""
+        if rid:
+            meta[rid] = m
+
+    # Live coordinators resolve their display name through the bulk
+    # helper keyed on their EXACT ids (one round-trip, no row cap) rather
+    # than the ``limit=200`` ``meta`` map: a live coord that has dropped
+    # below the 200-row ``updated DESC`` window would otherwise revert to
+    # its synthetic ``ws.name``.  Closed/evicted rows (the persisted lane
+    # below) already carry alias/title in their own ``_mapping``.
+    live_display = get_workstream_display_names([ws.id for ws in wss]) if wss else {}
+
+    def _display_name(ws_id: str, fallback: str) -> str:
+        m = meta.get(ws_id)
+        if m is None:
+            return fallback
+        return m.get("alias") or m.get("title") or m.get("name") or fallback
+
+    def _title(ws_id: str) -> str:
+        # Best-effort: the secondary ``title`` field is sourced from the
+        # ``limit=200`` ``meta`` map, so a live coord outside that window
+        # reports ``""`` here. The user-visible ``name`` stays correct
+        # (resolved via the uncapped ``live_display`` above, and the UI
+        # renders ``title || name``); the empty title is harmless and the
+        # window is unreachable in practice (live coords are bounded by
+        # ``max_active`` and sort to the top of ``updated DESC``).
+        m = meta.get(ws_id)
+        return str(m.get("title") or "") if m is not None else ""
+
    rows: list[dict[str, Any]] = []
    seen: set[str] = set()
    for ws in wss:
@@ -876,9 +943,9 @@ def _coordinator_rows(request: Request) -> list[dict[str, Any]]:
        rows.append(
            {
                "id": ws.id,
-                "name": ws.name,
+                "name": live_display.get(ws.id) or ws.name,
                "state": ws.state.value,
-                "title": "",
+                "title": _title(ws.id),
                "node": "console",
                "server_url": "",
                "model": _str_sess_attr(sess, "model"),
@@ -895,30 +962,7 @@ def _coordinator_rows(request: Request) -> list[dict[str, Any]]:
        )
        seen.add(ws.id)

-    # Second lane — persisted coordinator rows, used to surface
-    # closed / error / deleted coordinators the manager has already
-    # evicted from ``self._workstreams``.  Cluster-wide (trusted-team
-    # visibility).
-    storage = getattr(request.app.state, "auth_storage", None)
-    if storage is None:
-        return rows
-    try:
-        persisted = storage.list_workstreams(
-            kind=WorkstreamKind.COORDINATOR,
-            user_id=None,
-            limit=200,
-        )
-    except Exception:
-        log.debug("cluster_workstreams.coord_persisted_failed", exc_info=True)
-        return rows
-
    for row in persisted:
-        # SQLAlchemy Row — access via _mapping so future SELECT reorders
-        # / new columns don't silently corrupt the projection (per the
-        # storage-protocol guidance on list_workstreams).  Test doubles
-        # must expose the same ._mapping attribute; positional indexing
-        # was removed because it hard-coded column offsets that drift
-        # with migrations.
        m = row._mapping
        row_id = m.get("ws_id") or ""
        if not row_id or row_id in seen:
@@ -927,9 +971,9 @@ def _coordinator_rows(request: Request) -> list[dict[str, Any]]:
        rows.append(
            {
                "id": row_id,
-                "name": m.get("name") or f"coord-{row_id[:4]}",
+                "name": _display_name(row_id, f"coord-{row_id[:4]}"),
                "state": str(m.get("state") or "idle"),
-                "title": "",
+                "title": _title(row_id),
                "node": "console",
                "server_url": "",
                "model": "",
@@ -11530,16 +11574,15 @@ def _validate_regex_pattern(pattern: str, flags: int = 0) -> str | None:
            compiled.search(s)

    try:
-        from concurrent.futures import ThreadPoolExecutor
-        from concurrent.futures import TimeoutError as FuturesTimeout
-
-        pool = ThreadPoolExecutor(max_workers=1)
-        try:
-            pool.submit(_probe).result(timeout=0.5)
-        except FuturesTimeout:
-            return "Regex appears to have catastrophic backtracking"
-        finally:
-            pool.shutdown(wait=False, cancel_futures=True)
+        # Daemon worker: a catastrophically-backtracking regex must be
+        # abandonable without pinning a non-daemon thread that would hang
+        # interpreter exit (a ThreadPoolExecutor worker is joined at exit).
+        # Budget is generous — a legitimately complex pattern can take a second
+        # or two on the probe strings; only exponential blowup (which sails past
+        # any few-second bound) should trip the catastrophic-backtracking guard.
+        run_with_deadline(_probe, timeout=3.0, poll=0.1, thread_name="regex-redos-probe")
+    except DeadlineExceededError:
+        return "Regex appears to have catastrophic backtracking"
    except Exception:
        return "Regex caused an error during test"
    return None
@@ -12979,6 +13022,8 @@ def create_app(
                audit_emit=_audit_close_coordinator,
                supports_close_reason=False,
            ),
+            refresh_title=make_refresh_title_handler(coord_endpoint_config),  # lifted: shared body
+            set_title=make_set_title_handler(coord_endpoint_config),  # lifted: shared body
            send=make_send_handler(coord_endpoint_config),  # lifted: shared body (P1.5)
            dequeue=make_dequeue_handler(coord_endpoint_config),  # lifted: shared body
            approve=make_approve_handler(coord_endpoint_config),  # lifted: shared body
@@ -15,11 +15,16 @@ backend is surfaced as a typed error the endpoint maps to 503 / 502.

 from __future__ import annotations

+import subprocess
 import threading
 from dataclasses import dataclass
-from typing import Any
+from typing import TYPE_CHECKING, Any

 from turnstone.core.log import get_logger
+from turnstone.core.server_compat import merge_server_compat
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator

 log = get_logger(__name__)

@@ -83,6 +88,22 @@ _OMNI_STT_PROMPT = (
    "seven, and write 3 instead of three"
 )

+# Bound the omni STT decode.  Gemma caps audio at 30 s and a 30 s transcript is
+# well under this, so the cap only catches a pathological runaway — it never
+# truncates a real transcript.
+_OMNI_STT_MAX_TOKENS = 1024
+
+# Hard limit on the ffmpeg transcode subprocess (seconds).
+_FFMPEG_TIMEOUT_S = 30
+
+# Cap the decoded audio duration so a crafted clip can't expand into an
+# unbounded decode (the upload itself is already size-capped at the endpoint).
+_MAX_AUDIO_SECONDS = 300
+
+# Per-request timeout for the streaming STT chat call — bounds a hung backend
+# (the whole transcription is ~1 s; this only catches a stalled stream).
+_OMNI_STT_TIMEOUT_S = 60
+

@dataclass(frozen=True)
 class TranscriptionResult:
@@ -185,30 +206,115 @@ def _serves_transcription_endpoint(cfg: Any, model: str) -> bool:
    return _infer_audio_capability(model, "stt")


-def _transcribe_via_chat(client: Any, model: str, data: bytes, filename: str, prompt: str) -> str:
+def _to_wav_16k_mono(data: bytes) -> bytes:
+    """Decode any ffmpeg-readable audio container to 16 kHz mono PCM WAV.
+
+    Browsers record webm/opus (or ogg/mp4); the omni chat lane — vLLM in
+    particular — only decodes wav/mp3 and sniffs the bytes, so the raw upload is
+    rejected as an "Invalid or unsupported audio file".  ffmpeg reads the
+    container from the byte stream (no reliance on the filename) and resamples to
+    the 16 kHz mono PCM the model documents.  Raises :class:`AudioBackendError`
+    (the endpoint maps it to 502) if ffmpeg is missing or the bytes don't decode.
+    """
+    # ffmpeg reads only the piped bytes (-protocol_whitelist pipe) so a crafted
+    # container can't open file:/http: references (SSRF / local file read); -vn
+    # drops video streams and -t bounds the decode against a decompression bomb.
+    try:
+        proc = subprocess.run(
+            [
+                "ffmpeg",
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-protocol_whitelist",
+                "pipe",
+                "-i",
+                "pipe:0",
+                "-vn",
+                "-t",
+                str(_MAX_AUDIO_SECONDS),
+                "-ac",
+                "1",
+                "-ar",
+                "16000",
+                "-f",
+                "wav",
+                "pipe:1",
+            ],
+            input=data,
+            capture_output=True,
+            timeout=_FFMPEG_TIMEOUT_S,
+        )
+    except FileNotFoundError as exc:
+        raise AudioBackendError("ffmpeg is not installed; cannot transcode audio") from exc
+    except subprocess.TimeoutExpired as exc:
+        raise AudioBackendError("Audio transcode timed out") from exc
+    if proc.returncode != 0 or not proc.stdout:
+        detail = proc.stderr.decode("utf-8", "replace").strip()
+        raise AudioBackendError(f"Audio transcode failed: {detail[-200:] or 'no output'}")
+    return proc.stdout
+
+
+def _omni_chat_extra_body(cfg: Any) -> dict[str, Any]:
+    """Build the chat ``extra_body`` for an omni STT call.
+
+    The STT path calls the raw client, so it bypasses the provider's request
+    shaping.  Reuse ``merge_server_compat`` to forward any operator-stored
+    ``server_compat["extra_body"]``, then force **thinking OFF** via the model's
+    own ``thinking_param``: transcription needs no reasoning, and leaving it on
+    multiplies latency ~10x and (on some chat templates) empties the content.
+    The override is applied last so it wins over any operator thinking flag.
+    """
+    server_compat = getattr(cfg, "server_compat", None)
+    extra = merge_server_compat(None, server_compat) if isinstance(server_compat, dict) else {}
+    caps = getattr(cfg, "capabilities", None) or {}
+    thinking_param = caps.get("thinking_param")
+    if thinking_param and caps.get("thinking_mode") in ("manual", "adaptive"):
+        extra.setdefault("chat_template_kwargs", {})[thinking_param] = False
+    return extra
+
+
+def _omni_chat_messages(prompt: str, audio_b64: str) -> list[dict[str, Any]]:
+    """The single user turn for an omni STT chat call: the prompt precedes the
+    audio part — the order Gemma documents for transcription."""
+    return [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": prompt},
+                {"type": "input_audio", "input_audio": {"data": audio_b64, "format": "wav"}},
+            ],
+        }
+    ]
+
+
+def _transcribe_via_chat(
+    client: Any,
+    model: str,
+    data: bytes,
+    prompt: str,
+    *,
+    extra_body: dict[str, Any] | None = None,
+    max_tokens: int = _OMNI_STT_MAX_TOKENS,
+) -> str:
    """Transcribe by handing the clip to an omni *chat* model as ``input_audio``.

    For models that accept audio in chat (``supports_audio_input``) but don't
-    serve ``/audio/transcriptions``.  The instruction ``prompt`` steers the model
-    to emit only the transcript.  The audio format is taken from the upload's
-    filename extension (the same shape the attachment wire path uses).
+    serve ``/audio/transcriptions``.  The clip is transcoded to 16 kHz mono WAV
+    first (browsers record webm/opus, which the chat lane can't decode).  The
+    instruction ``prompt`` precedes the audio part — the order Gemma documents
+    for transcription — and ``extra_body`` carries the thinking-off / server
+    compat params the raw-client path would otherwise skip.
    """
    import base64

-    name = filename or "speech.webm"
-    fmt = name.rsplit(".", 1)[-1].lower() if "." in name else "wav"
-    audio_b64 = base64.b64encode(data).decode("ascii")
+    wav = _to_wav_16k_mono(data)
+    audio_b64 = base64.b64encode(wav).decode("ascii")
    resp = client.chat.completions.create(
        model=model,
-        messages=[
-            {
-                "role": "user",
-                "content": [
-                    {"type": "text", "text": prompt},
-                    {"type": "input_audio", "input_audio": {"data": audio_b64, "format": fmt}},
-                ],
-            }
-        ],
+        messages=_omni_chat_messages(prompt, audio_b64),
+        max_tokens=max_tokens,
+        extra_body=extra_body or None,
    )
    choices = getattr(resp, "choices", None) or []
    if not choices:
@@ -261,13 +367,86 @@ def transcribe(
            transcript = (getattr(resp, "text", "") or "").strip()
        else:
            transcript = _transcribe_via_chat(
-                client, model, data, filename, prompt or _OMNI_STT_PROMPT
+                client,
+                model,
+                data,
+                prompt or _OMNI_STT_PROMPT,
+                extra_body=_omni_chat_extra_body(cfg),
            )
+    except AudioBackendError:
+        # Transcode errors already carry an actionable message — keep it.
+        raise
    except Exception as exc:
        raise AudioBackendError(f"Transcription backend failed: {exc}") from exc
    return TranscriptionResult(transcript=transcript, model_alias=alias, model=model)


+def _iter_stream_deltas(stream: Any) -> Iterator[str]:
+    """Yield non-empty content deltas from an OpenAI streaming chat response.
+
+    Owns the stream's lifecycle: exhausting or closing this generator releases
+    the underlying HTTP connection, so an abandoned stream can't leak it.
+    """
+    try:
+        for chunk in stream:
+            choices = getattr(chunk, "choices", None) or []
+            if not choices:
+                continue
+            delta = getattr(choices[0].delta, "content", None)
+            if delta:
+                yield delta
+    finally:
+        close = getattr(stream, "close", None)
+        if callable(close):
+            close()
+
+
+def transcribe_stream(*, registry: Any, alias: str, data: bytes, prompt: str = "") -> Iterator[str]:
+    """Stream transcript content deltas for the STT role alias.
+
+    Resolve, transcode, and opening the streaming-chat request all run eagerly
+    (before the returned generator yields its first delta) so the caller can
+    surface a clean 503 / 502; only the token iteration is deferred.  A
+    whisper-style endpoint alias has no chat stream, so it emits the whole
+    transcript as a single chunk.
+    """
+    try:
+        client, model, cfg = registry.resolve(alias)
+    except Exception as exc:  # unknown/removed alias
+        raise AudioUnavailableError(f"STT model alias {alias!r} is not available") from exc
+    if not _provider_carries_audio(cfg):
+        raise AudioUnavailableError(
+            f"STT model alias {alias!r} (provider {getattr(cfg, 'provider', 'unknown')!r}) "
+            "can't transcribe audio — audio roles require an OpenAI-compatible provider."
+        )
+    if _serves_transcription_endpoint(cfg, model):
+        # Whisper-style endpoint: no chat stream — emit the whole transcript once.
+        text = transcribe(
+            registry=registry, alias=alias, data=data, filename="speech.webm", prompt=prompt
+        ).transcript
+        return iter([text] if text else [])
+    caps = getattr(cfg, "capabilities", None) or {}
+    if not caps.get("supports_audio_input"):
+        raise AudioUnavailableError(f"STT model alias {alias!r} cannot transcribe audio")
+
+    import base64
+
+    wav = _to_wav_16k_mono(data)
+    audio_b64 = base64.b64encode(wav).decode("ascii")
+    try:
+        stream = client.chat.completions.create(
+            model=model,
+            messages=_omni_chat_messages(prompt or _OMNI_STT_PROMPT, audio_b64),
+            max_tokens=_OMNI_STT_MAX_TOKENS,
+            extra_body=_omni_chat_extra_body(cfg) or None,
+            stream=True,
+            timeout=_OMNI_STT_TIMEOUT_S,
+        )
+    except Exception as exc:
+        raise AudioBackendError(f"Transcription backend failed: {exc}") from exc
+    return _iter_stream_deltas(stream)
+
+
 # -- transcript memoization (no-native-audio wire fallback) -------------------
 # Caching an STT result is an audio-domain concern, so it lives here next to
 # ``transcribe``.  The wire resolver re-materializes every attachment on every
@@ -0,0 +1,94 @@
+"""Run a blocking call under a wall-clock deadline on a daemon thread.
+
+The motivating constraint comes from the judges (:mod:`turnstone.core.judge`,
+:mod:`turnstone.core.output_guard_judge`): an upstream LLM call must be
+*abandonable* the instant its timeout or cancel fires, without the abandoned
+call being able to block process or interpreter exit.
+
+A :class:`~concurrent.futures.ThreadPoolExecutor` worker is **non-daemon**, and
+``concurrent.futures`` joins every executor worker from an ``atexit`` hook
+(``_python_exit``) regardless of ``shutdown(wait=False)``.  So an upstream call
+wedged with no socket timeout hangs interpreter shutdown forever — which is
+exactly how a single slow judge call can deadlock a whole test run at exit.
+
+A **daemon** worker is never joined at exit, so abandoning one is always safe:
+the call keeps running until it returns or the process dies, whichever comes
+first, and never pins shutdown.
+"""
+
+from __future__ import annotations
+
+import queue
+import threading
+import time
+from typing import TYPE_CHECKING, TypeVar
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+_T = TypeVar("_T")
+
+
+class DeadlineExceededError(Exception):
+    """The call did not complete before its wall-clock deadline."""
+
+
+class DeadlineCancelledError(Exception):
+    """The cancel event fired before the call completed."""
+
+
+def run_with_deadline(
+    fn: Callable[[], _T],
+    *,
+    timeout: float,
+    cancel_event: threading.Event | None = None,
+    poll: float = 1.0,
+    thread_name: str = "deadline-worker",
+) -> _T:
+    """Run ``fn()`` on a daemon thread, bounded by ``timeout``/``cancel_event``.
+
+    Returns ``fn()``'s result, or re-raises whatever ``fn`` raised.  Raises
+    :class:`DeadlineExceededError` if ``timeout`` seconds elapse first, or
+    :class:`DeadlineCancelledError` if ``cancel_event`` fires first.  On either
+    abort the worker thread is abandoned; being a daemon it cannot block
+    process or interpreter exit.
+
+    ``poll`` bounds how often ``cancel_event`` is checked (and thus the worst-
+    case latency from a cancel to this function returning).
+    """
+    box: queue.Queue[tuple[bool, object]] = queue.Queue(maxsize=1)
+
+    def _runner() -> None:
+        try:
+            box.put((True, fn()))
+        except BaseException as exc:  # noqa: BLE001 - relayed to the caller verbatim
+            box.put((False, exc))
+
+    threading.Thread(target=_runner, name=thread_name, daemon=True).start()
+
+    deadline = time.monotonic() + timeout
+    while True:
+        # Prefer a result that has already arrived over a deadline or cancel
+        # firing in the same scheduling window — otherwise a completed call
+        # could be reported as a spurious timeout/cancel under jitter.
+        try:
+            ok, payload = box.get_nowait()
+        except queue.Empty:
+            pass
+        else:
+            if ok:
+                return payload  # type: ignore[return-value]  # ok=True ⇒ payload is _T
+            raise payload  # type: ignore[misc]  # ok=False ⇒ payload is the raised exc
+
+        if cancel_event is not None and cancel_event.is_set():
+            raise DeadlineCancelledError
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            raise DeadlineExceededError
+        try:
+            ok, payload = box.get(timeout=min(remaining, poll))
+        except queue.Empty:
+            continue
+        if ok:
+            return payload  # type: ignore[return-value]
+        raise payload  # type: ignore[misc]
@@ -15,11 +15,16 @@ import re
 import threading
 import time
 import uuid
-from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass, field
+from functools import partial
 from pathlib import Path
 from typing import TYPE_CHECKING, Any

+from turnstone.core.deadline import (
+    DeadlineCancelledError,
+    DeadlineExceededError,
+    run_with_deadline,
+)
 from turnstone.core.log import get_logger

 if TYPE_CHECKING:
@@ -76,8 +81,8 @@ class JudgeConfig:
    """Configuration for the intent validation judge.

    The *timeout* value applies **per turn**, not as a total budget across
-    all turns.  With the default of 60 s and a maximum of 5 turns, a
-    single tool-call evaluation can take up to 300 s in the worst case
+    all turns.  With the default of 120 s and a maximum of 5 turns, a
+    single tool-call evaluation can take up to 600 s in the worst case
    (e.g. a multi-turn tool-use exchange with a slow local model).
    """

@@ -86,13 +91,13 @@ class JudgeConfig:
    smart_approvals: bool = False  # auto-approve high-confidence "approve" LLM verdicts
    confidence_threshold: float = 0.95  # Smart Approvals auto-approve bar (recommendation=approve)
    max_context_ratio: float = 0.5
-    timeout: float = 60.0  # per-turn timeout in seconds (see class docstring)
+    timeout: float = 120.0  # per-turn timeout in seconds (see class docstring)
    read_only_tools: bool = True
    output_guard: bool = True
    output_guard_budget_seconds: float = 30.0  # wall-clock budget for output_guard regex scan
    output_guard_llm: bool = False  # enable LLM stage on tool output (issue #560 mitigation #1)
    output_guard_model: str = ""  # alias for the LLM stage; empty = inherit session model
-    output_guard_llm_timeout: float = 30.0  # wall-clock budget for the LLM stage
+    output_guard_llm_timeout: float = 60.0  # wall-clock budget for the LLM stage
    redact_secrets: bool = True
    # True = the approval gate's resolution aborts remaining evaluations
    # (saves inference; undone items degrade to ``llm_fallback`` verdicts
@@ -881,10 +886,6 @@ If you used read_file to check a target, cite what you found."""
 # ---------------------------------------------------------------------------


-class _ExecutorPoisonedError(Exception):
-    """Raised when a timeout leaves the executor's worker thread stuck."""
-
-
 class IntentJudge:
    """Session-scoped LLM judge for intent validation.

@@ -1054,7 +1055,6 @@ class IntentJudge:
        verdicts are delivered.
        """
        client = self._create_client()
-        executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="judge-api")
        try:
            for idx, (item, h_verdict) in enumerate(zip(items, heuristic_verdicts, strict=True)):
                if cancel_event and cancel_event.is_set():
@@ -1071,7 +1071,6 @@ class IntentJudge:
                        item,
                        messages,
                        cancel_event,
-                        executor,
                        client,
                    )
                    if llm_verdict:
@@ -1116,17 +1115,6 @@ class IntentJudge:
                            "judge cancelled before evaluating this call",
                        )
                        return
-                except _ExecutorPoisonedError:
-                    executor.shutdown(wait=False, cancel_futures=True)
-                    executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="judge-api")
-                    # Deliver a fallback for the interrupted item so every
-                    # call still gets exactly one verdict. Smart Approvals
-                    # waits on the full set before gating; a silently-
-                    # skipped item would otherwise block that wait until
-                    # its timeout (and the advisory UI would miss a chip).
-                    self._deliver_fallbacks(
-                        [item], [h_verdict], callback, "judge executor restarted"
-                    )
                except Exception:
                    log.exception(
                        "Judge evaluation failed for %s",
@@ -1134,7 +1122,6 @@ class IntentJudge:
                    )
                    self._deliver_fallbacks([item], [h_verdict], callback, "judge evaluation error")
        finally:
-            executor.shutdown(wait=False, cancel_futures=True)
            try:
                if hasattr(client, "close"):
                    client.close()
@@ -1172,7 +1159,6 @@ class IntentJudge:
        item: dict[str, Any],
        messages: list[dict[str, Any]],
        cancel_event: threading.Event | None,
-        executor: ThreadPoolExecutor,
        client: Any,
    ) -> IntentVerdict | None:
        """Run LLM judge for a single tool call. Returns verdict or None."""
@@ -1237,32 +1223,29 @@ class IntentJudge:
            # models aren't penalised for slow earlier turns.
            per_call_timeout = max(self._config.timeout, 5.0)  # at least 5s
            try:
-                future = executor.submit(
-                    self._provider.create_completion,
-                    client=client,
-                    model=self._model,
-                    messages=judge_messages,
-                    tools=None if is_last_turn else tools,
-                    max_tokens=2048,
-                    temperature=0.0,
-                    reasoning_effort="medium",
+                # Each turn runs on its own daemon worker (1s cancel polling).
+                # A timeout or cancel abandons the call without pinning a
+                # non-daemon thread that would block interpreter exit — the old
+                # single-slot ThreadPoolExecutor left a stuck worker that
+                # poisoned the pool, which is why the restart dance existed.
+                result = run_with_deadline(
+                    partial(
+                        self._provider.create_completion,
+                        client=client,
+                        model=self._model,
+                        messages=judge_messages,
+                        tools=None if is_last_turn else tools,
+                        max_tokens=2048,
+                        temperature=0.0,
+                        reasoning_effort="medium",
+                    ),
+                    timeout=per_call_timeout,
+                    cancel_event=cancel_event,
+                    thread_name="judge-api",
                )
-                # Poll in 1s increments so we notice cancellation promptly
-                # instead of blocking for the full per_call_timeout.
-                deadline = time.monotonic() + per_call_timeout
-                while True:
-                    remaining = deadline - time.monotonic()
-                    if cancel_event and cancel_event.is_set():
-                        future.cancel()
-                        return None
-                    if remaining <= 0:
-                        raise TimeoutError
-                    try:
-                        result = future.result(timeout=min(remaining, 1.0))
-                        break
-                    except TimeoutError:
-                        pass  # loop back to check remaining/cancel
-            except TimeoutError:
+            except DeadlineCancelledError:
+                return None
+            except DeadlineExceededError:
                log.info("judge.turn.timeout", turn=turn + 1, timeout=per_call_timeout)
                # Safety net: if we have a partial result from a previous turn,
                # try to parse a verdict from it before giving up.
@@ -1277,7 +1260,7 @@ class IntentJudge:
                    if verdict:
                        log.info("judge.verdict.from_partial", turn=turn + 1)
                        return verdict
-                raise _ExecutorPoisonedError from None
+                return None
            except Exception as e:
                log.info("judge.turn.failed", turn=turn + 1, error=str(e))
                return None
@@ -12,13 +12,13 @@ Design:
  a static tool result doesn't benefit from multi-turn — the text is
  already in hand.
 - JSON-in-content verdict.  4-strategy parser inlined from
-  :class:`IntentJudge` (``judge.py:1603-1659``).
- ``ThreadPoolExecutor`` + ``future.result(timeout=)`` with 1 s
-  cancel-event polling.  The executor is owned explicitly with
-  ``shutdown(wait=False, cancel_futures=True)`` so a timeout or
-  cancellation returns promptly even if the worker thread is still
-  blocked on the upstream LLM call.  This mirrors
-  :meth:`IntentJudge._run_judge`'s pattern at ``judge.py:1117-1118``.
+  :meth:`IntentJudge._parse_verdict`.
+- Wall-clock deadline via :func:`turnstone.core.deadline.run_with_deadline`,
+  which runs the call on a *daemon* worker and polls the cancel event each
+  second.  A timeout or cancel abandons the call rather than waiting it out,
+  and the daemon worker can never block process or interpreter exit — unlike
+  a ``ThreadPoolExecutor`` worker, which ``concurrent.futures`` joins from an
+  ``atexit`` hook regardless of ``shutdown(wait=False)``.
 - HTTP client is lazy-init + reused across evaluations on a single
  judge instance.  Session-side model swaps drop the entire
  :class:`OutputGuardJudge` (``session.py:1733``/``:2136``), which
@@ -39,11 +39,15 @@ import json
 import re
 import time
 import uuid
-from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any

 from turnstone.core import fence
+from turnstone.core.deadline import (
+    DeadlineCancelledError,
+    DeadlineExceededError,
+    run_with_deadline,
+)
 from turnstone.core.log import get_logger

 if TYPE_CHECKING:
@@ -370,9 +374,10 @@ class OutputGuardJudge:
        by :meth:`_user_prompt`.  Callers that don't have a particular
        field leave it at its default — the prompt skips empty sections.

-        Timeout enforcement is real wall-clock: the executor is shut
-        down with ``wait=False, cancel_futures=True`` on the timeout /
-        cancel path, so a hung upstream LLM call does not block return.
+        Timeout enforcement is real wall-clock: the upstream call runs on a
+        daemon worker via :func:`~turnstone.core.deadline.run_with_deadline`
+        and is abandoned on the timeout / cancel path, so a hung upstream LLM
+        call neither blocks return nor pins interpreter exit.
        """
        if not output:
            return OutputJudgeVerdict(
@@ -407,15 +412,16 @@ class OutputGuardJudge:
                verdict_id, call_id, start, f"client_create_failed: {type(e).__name__}"
            )

-        # Explicit executor lifetime — the `with ... as ex:` form's
-        # implicit shutdown(wait=True) would block return until the
-        # upstream call completed, defeating the wall-clock timeout.
-        # Mirror IntentJudge's pattern at judge.py:1117-1118.
-        ex = ThreadPoolExecutor(max_workers=1, thread_name_prefix="output-guard-judge")
+        # Run the upstream call on a *daemon* worker bounded by a real
+        # wall-clock deadline: a timeout or cancel abandons the call instead of
+        # waiting it out, and because the worker is a daemon an abandoned call
+        # can never block process or interpreter exit.  (A ThreadPoolExecutor
+        # worker is non-daemon, and concurrent.futures joins it from an atexit
+        # hook regardless of shutdown(wait=False) — so a wedged upstream call
+        # would otherwise hang shutdown.)
        try:
-            try:
-                future = ex.submit(
-                    self._provider.create_completion,
+            result = run_with_deadline(
+                lambda: self._provider.create_completion(
                    client=client,
                    model=self._model,
                    messages=judge_messages,
@@ -423,27 +429,19 @@ class OutputGuardJudge:
                    max_tokens=512,
                    temperature=0.0,
                    reasoning_effort="low",
-                )
-                deadline = time.monotonic() + timeout
-                while True:
-                    if cancel_event is not None and cancel_event.is_set():
-                        future.cancel()
-                        return self._error_verdict(verdict_id, call_id, start, "cancelled")
-                    remaining = deadline - time.monotonic()
-                    if remaining <= 0:
-                        future.cancel()
-                        return self._error_verdict(verdict_id, call_id, start, "timeout")
-                    try:
-                        result = future.result(timeout=min(remaining, 1.0))
-                        break
-                    except TimeoutError:
-                        continue
-            except Exception as e:
-                return self._error_verdict(
-                    verdict_id, call_id, start, f"provider_error: {type(e).__name__}"
-                )
-        finally:
-            ex.shutdown(wait=False, cancel_futures=True)
+                ),
+                timeout=timeout,
+                cancel_event=cancel_event,
+                thread_name="output-guard-judge",
+            )
+        except DeadlineCancelledError:
+            return self._error_verdict(verdict_id, call_id, start, "cancelled")
+        except DeadlineExceededError:
+            return self._error_verdict(verdict_id, call_id, start, "timeout")
+        except Exception as e:
+            return self._error_verdict(
+                verdict_id, call_id, start, f"provider_error: {type(e).__name__}"
+            )

        content = (getattr(result, "content", "") or "").strip()
        if not content:
@@ -14,10 +14,12 @@ request shaping.  This module separates three concerns:
   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.
+3. **Server workarounds** — ``extra_body`` overrides like llama.cpp's
+   ``reasoning_format`` are properties of the *server*, not the model.
+   These stay in ``server_compat`` and get merged into the request's
+   ``extra_body`` at call time.  Reserve these for stable server config:
+   bug-workaround flags for fast-moving open models go stale the moment
+   the upstream bug is fixed, so we don't carry them speculatively.

 Profiles are *suggestions* only.  The admin UI auto-fills them on
 Detect; the operator has final say, and the stored DB config is what
@@ -45,11 +47,6 @@ _PROFILES: dict[str, dict[str, Any]] = {
        },
        "server_compat": {
            "server_type": "vllm",
-            # Workaround: vLLM strips special tokens before the Gemma4
-            # reasoning parser sees them.  skip_special_tokens=false
-            # preserves <|channel> / <channel|> markers so reasoning
-            # content is extracted correctly.
-            "extra_body": {"skip_special_tokens": False},
        },
    },
    "vllm-qwen-thinking": {
@@ -2425,10 +2425,15 @@ class ChatSession:
        ws_id = self._ws_id  # Capture before async work
        log.info("ws.title.gen_start", ws_id=ws_id[:8])
        try:
-            # Gather first user message and first assistant reply
+            # Gather first user message and first assistant reply.
+            # Snapshot ``self.messages`` (C-level atomic copy under the
+            # GIL): this runs in a background thread that may now fire
+            # while the main ``send`` loop is still streaming and
+            # appending turns, so iterating the live list directly could
+            # raise "list changed size during iteration".
            user_msg = ""
            asst_msg = ""
-            for m in self.messages:
+            for m in list(self.messages):
                content = m.text  # joins text blocks; multipart attachments contribute none
                if m.role is Role.USER and not user_msg:
                    user_msg = content[:300]
@@ -4150,6 +4155,25 @@ class ChatSession:
        # legacy per-message ``_reminders`` side-channel splice.
        self._emit_pending_user_nudges()

+        # Auto-title from the opening user message — fire NOW rather than
+        # waiting for the assistant's final tool-call-free turn.  The old
+        # trigger sat in the ``not tool_calls`` branch of the loop below;
+        # coordinators spend nearly every turn in tool calls and may never
+        # reach that terminal text turn, so the title almost never
+        # generated for them.  Gate on a real user message: synthetic wake
+        # sends carry no content and ``_generate_title`` would no-op on the
+        # empty/attachment-only case anyway (it needs first-user-message
+        # text).  Concurrency: this background thread runs alongside the
+        # streaming turn started below, but safely — it snapshots
+        # ``self.messages`` for iteration, and the only UI it touches is
+        # ``on_aux_usage`` (storage/metrics, no ``_ws_lock`` state) and
+        # ``on_rename`` (queue/locked fan-out), both documented
+        # auxiliary-thread-safe on ``SessionUIBase``; the provider + client
+        # handle concurrent requests (the same path ``task_agent`` uses).
+        if not self._title_generated and user_input.strip() and not from_wake:
+            self._title_generated = True
+            threading.Thread(target=self._generate_title, daemon=True).start()
+
        # A fresh session composed its system prefix at __init__ with an empty
        # history, so memory selection fell back to recency (no query, no rerank).
        # Recompose once the first real user message exists so the opening turn
@@ -4292,10 +4316,6 @@ class ChatSession:
                        self._compact_messages(auto=True)
                        # Update status bar with post-compaction token counts
                        self._print_status_line()
-                    # Auto-title session after first exchange
-                    if not self._title_generated:
-                        self._title_generated = True
-                        threading.Thread(target=self._generate_title, daemon=True).start()
                    # Flush any queued messages that weren't injected
                    # (no tool calls → no advisory seam to inject at).
                    # If anything drained, the model hasn't seen those
@@ -493,9 +493,8 @@ class SharedSessionVerbHandlers:
    """Bundle of HTTP handler callables for verbs both kinds expose.

    All handlers are optional; ``None`` skips that route. One bundle
-    describes either kind — coord omits ``delete`` / ``refresh_title``
-    / ``set_title`` / attachments; interactive populates every
-    interaction verb post-Stage-2.
+    describes either kind — coord omits ``delete``; interactive
+    populates every interaction verb post-Stage-2.
    """

    # Listing
@@ -972,6 +971,122 @@ def make_close_handler(
    return close


+def make_refresh_title_handler(cfg: SessionEndpointConfig) -> Handler:
+    """Lifted body for ``POST {prefix}/{ws_id}/refresh-title``.
+
+    Regenerates the workstream title via a background LLM call
+    (:meth:`ChatSession.request_title_refresh`). Both kinds share the
+    auth → mgr → ws-lookup → request sequence; the session must be live
+    in memory (``mgr.get``, not ``open``) since the refresh runs on the
+    loaded :class:`ChatSession`. The current display name is passed so
+    the generator is steered toward a *different* title on a manual
+    refresh.
+    """
+
+    async def refresh_title(request: Request) -> Response:
+        import asyncio
+
+        from turnstone.core.memory import get_workstream_display_name
+
+        if cfg.permission_gate is not None:
+            err = cfg.permission_gate(request)
+            if err is not None:
+                return err
+        mgr_opt, err503 = cfg.manager_lookup(request)
+        if err503 is not None:
+            return err503
+        # See ``make_approve_handler`` for the cast rationale.
+        mgr = cast("SessionManager", mgr_opt)
+        ws_id = request.path_params.get("ws_id", "")
+
+        if cfg.tenant_check is not None:
+            err_tenant = await asyncio.to_thread(cfg.tenant_check, request, ws_id, mgr)
+            if err_tenant is not None:
+                return err_tenant
+
+        ws = mgr.get(ws_id)
+        if ws is None or ws.session is None:
+            return JSONResponse({"error": cfg.not_found_label}, status_code=404)
+
+        current_title = await asyncio.to_thread(get_workstream_display_name, ws_id) or ""
+        ws.session.request_title_refresh(current_title)
+        return JSONResponse({"status": "ok"})
+
+    return refresh_title
+
+
+def make_set_title_handler(cfg: SessionEndpointConfig) -> Handler:
+    """Lifted body for ``POST {prefix}/{ws_id}/title``.
+
+    Sets a user-chosen title manually. Stored as the workstream *alias*
+    so it outranks the LLM auto-title in the display fallback chain
+    (``alias > title > name``). Both kinds share the auth → validate →
+    ``set_workstream_alias`` → ``on_rename`` sequence. Returns 409 when
+    the name collides with another workstream's alias.
+
+    Behavior matches the pre-lift interactive handler: the alias is set
+    against storage regardless of whether the session is loaded (so a
+    saved/closed workstream can still be renamed), and the live
+    ``on_rename`` broadcast fires only when the session is in memory.
+    """
+
+    async def set_title(request: Request) -> Response:
+        import asyncio
+
+        from turnstone.core.memory import set_workstream_alias
+        from turnstone.core.web_helpers import read_json_or_400
+
+        if cfg.permission_gate is not None:
+            err = cfg.permission_gate(request)
+            if err is not None:
+                return err
+        mgr_opt, err503 = cfg.manager_lookup(request)
+        if err503 is not None:
+            return err503
+        mgr = cast("SessionManager", mgr_opt)
+        ws_id = request.path_params.get("ws_id", "")
+        if not ws_id:
+            return JSONResponse({"error": "ws_id is required"}, status_code=400)
+
+        if cfg.tenant_check is not None:
+            err_tenant = await asyncio.to_thread(cfg.tenant_check, request, ws_id, mgr)
+            if err_tenant is not None:
+                return err_tenant
+
+        # Resolve the workstream BEFORE writing the alias. ``set_workstream_alias``
+        # is a global, kind-unscoped UPDATE keyed on ``ws_id`` alone (it returns
+        # True even on a 0-row match), so a kind that has no ``tenant_check``
+        # storage gate (coord — the in-memory manager is its existence + kind
+        # authority) must 404 here, or an operator could rename a workstream this
+        # manager doesn't own (e.g. an interactive ws via the coord route) and a
+        # bogus id would silently 200. Interactive keeps ``tenant_check`` as its
+        # existence gate, so this stays skipped there and a non-loaded
+        # saved/closed ws still renames.
+        ws = mgr.get(ws_id)
+        if cfg.tenant_check is None and ws is None:
+            return JSONResponse({"error": cfg.not_found_label}, status_code=404)
+
+        body = await read_json_or_400(request)
+        if isinstance(body, JSONResponse):
+            return body
+        title = str(body.get("title", "")).strip()
+        if not title:
+            return JSONResponse({"error": "title is required"}, status_code=400)
+        title = title[:80]
+
+        if not await asyncio.to_thread(set_workstream_alias, ws_id, title):
+            return JSONResponse(
+                {"error": "That name is already used by another workstream"},
+                status_code=409,
+            )
+
+        if ws is not None and ws.session is not None and ws.session.ui is not None:
+            ws.session.ui.on_rename(title)
+        return JSONResponse({"status": "ok", "title": title})
+
+    return set_title
+
+
 CancelAuditEmitter = Callable[
    ["Request", str, "Workstream", bool],
    None,
@@ -201,8 +201,18 @@ class SessionUIBase:
    methods (and the approval blocking helpers that live on
    subclasses); HTTP handlers drive ``_register_listener`` /
    ``_unregister_listener`` / ``resolve_approval`` from the event
-    loop. All shared state is guarded by ``_listeners_lock`` or
-    ``threading.Event`` primitives.
+    loop. All shared state is guarded by ``_listeners_lock`` /
+    ``_ws_lock`` or ``threading.Event`` primitives.
+
+    Two ``on_*`` methods are additionally safe to call from a
+    *concurrent* auxiliary thread (e.g. background title generation in
+    ``ChatSession._generate_title``, or ``task_agent`` sub-agents), even
+    while the worker thread is mid-stream: :meth:`on_aux_usage` (a
+    storage ``usage_event`` write + thread-safe metric counters — it
+    touches none of the ``_ws_lock``-guarded inflight state
+    :meth:`on_status`/token writers mutate) and :meth:`on_rename` (a
+    queue / locked fan-out). Keep those two free of unguarded
+    ``_ws_*`` writes so the auxiliary-thread guarantee holds.
    """

    def __init__(self, ws_id: str = "", user_id: str = "") -> None:
@@ -577,7 +577,7 @@ def _build_registry() -> dict[str, SettingDef]:
        SettingDef(
            "judge.timeout",
            "float",
-            60.0,
+            120.0,
            "Judge evaluation timeout in seconds",
            "judge",
            min_value=5.0,
@@ -641,7 +641,7 @@ def _build_registry() -> dict[str, SettingDef]:
        SettingDef(
            "judge.output_guard_llm_timeout",
            "float",
-            30.0,
+            60.0,
            "Wall-clock budget for the output-guard LLM judge call",
            "judge",
            min_value=1.0,
@@ -8,6 +8,7 @@ from turnstone.core.storage._registry import (
    StorageUnavailableError,
    get_storage,
    init_storage,
+    is_storage_initialized,
    reset_storage,
 )

@@ -17,5 +18,6 @@ __all__ = [
    "StorageUnavailableError",
    "get_storage",
    "init_storage",
+    "is_storage_initialized",
    "reset_storage",
 ]
@@ -1052,6 +1052,12 @@ class PostgreSQLBackend:
                    workstreams.c.skill_id,
                    workstreams.c.skill_version,
                    workstreams.c.user_id,
+                    # Appended after ``user_id`` so positional fallbacks in
+                    # consumers (``_coord_children_row`` et al.) that index
+                    # up to row[9] stay valid; ``_coordinator_rows`` reads
+                    # these by name to surface the persisted display title.
+                    workstreams.c.title,
+                    workstreams.c.alias,
                )
                .order_by(workstreams.c.updated.desc())
                .limit(limit)
@@ -679,7 +679,9 @@ class StorageBackend(Protocol):
        Returns a list of SQLAlchemy ``Row`` objects.  **Prefer dict access
        via ``row._mapping[<col>]``**; positional indexing is brittle against
        future SELECT reorders and against new columns appearing in the
-        tail (the select currently ends with ``user_id``).
+        tail (the select currently ends with ``user_id, title, alias`` —
+        ``title``/``alias`` were appended after ``user_id`` so existing
+        positional fallbacks that index up to row[9] stay valid).
        """
        ...

@@ -124,6 +124,19 @@ def get_storage() -> StorageBackend:
    return _storage


+def is_storage_initialized() -> bool:
+    """Return True when the storage singleton has been initialized.
+
+    Lets callers on lifecycle / early-startup paths consult storage
+    without tripping :func:`get_storage`'s SQLite auto-init side effect
+    (which would create ``.turnstone.db`` in the CWD). Use this to guard
+    a best-effort read that should simply be skipped before the host has
+    called :func:`init_storage` — never as a substitute for the explicit
+    init the app's startup performs.
+    """
+    return _storage is not None
+
+
 def reset_storage() -> None:
    """Close and clear the storage backend singleton (for tests)."""
    global _storage
@@ -1209,6 +1209,12 @@ class SQLiteBackend:
                    workstreams.c.skill_id,
                    workstreams.c.skill_version,
                    workstreams.c.user_id,
+                    # Appended after ``user_id`` so positional fallbacks in
+                    # consumers (``_coord_children_row`` et al.) that index
+                    # up to row[9] stay valid; ``_coordinator_rows`` reads
+                    # these by name to surface the persisted display title.
+                    workstreams.c.title,
+                    workstreams.c.alias,
                )
                .order_by(workstreams.c.updated.desc())
                .limit(limit)
@@ -294,8 +294,6 @@ class TLSClient:
        Discovery, CA fetch, and cert request are all idempotent, so the whole
        sequence is retried as a unit.
        """
-        import asyncio
-
        if attempts < 1:
            # range(1, attempts + 1) would be empty: init() would return
            # "successfully" with no CA and no cert.
@@ -321,7 +319,18 @@ class TLSClient:
                    delay_seconds=delay,
                    error=f"{type(exc).__name__}: {exc}",
                )
-                await asyncio.sleep(delay)
+                await self._sleep(delay)
+
+    async def _sleep(self, delay: float) -> None:
+        """Backoff sleep behind a seam so tests can stub it in isolation.
+
+        Patching the module-global ``asyncio.sleep`` would also intercept it
+        for every other task sharing the event loop; routing the retry backoff
+        through a method keeps test stubs from corrupting concurrent tasks.
+        """
+        import asyncio
+
+        await asyncio.sleep(delay)

    def _discover_console_url(self) -> str:
        """Look up the console URL from the services table."""
@@ -39,7 +39,7 @@ from sse_starlette import EventSourceResponse
 from starlette.applications import Starlette
 from starlette.middleware import Middleware
 from starlette.requests import Request
-from starlette.responses import HTMLResponse, JSONResponse, Response
+from starlette.responses import HTMLResponse, JSONResponse, Response, StreamingResponse
 from starlette.routing import Mount, Route
 from starlette.staticfiles import StaticFiles

@@ -77,10 +77,12 @@ from turnstone.core.session_routes import (
    make_history_handler,
    make_list_handler,
    make_open_handler,
+    make_refresh_title_handler,
    make_retry_handler,
    make_rewind_handler,
    make_saved_handler,
    make_send_handler,
+    make_set_title_handler,
    register_session_routes,
 )
 from turnstone.core.session_ui_base import (
@@ -1290,6 +1292,100 @@ async def speech_to_text(request: Request) -> JSONResponse:
    )


+async def speech_to_text_stream(request: Request) -> Response:
+    """POST /v1/api/workstreams/{ws_id}/speech-to-text/stream — stream the
+    transcript as plain-text deltas for lower perceived latency than the JSON
+    ``speech-to-text`` endpoint.  Resolve/transcode failures surface as
+    503 / 502 before any bytes are sent; once streaming begins the body is
+    best-effort (a mid-stream backend error just ends the partial stream)."""
+    from turnstone.core.audio import (
+        AudioBackendError,
+        AudioUnavailableError,
+        resolve_role_alias,
+        transcribe_stream,
+    )
+    from turnstone.core.web_helpers import read_multipart_file_or_400
+
+    ws_id = request.path_params.get("ws_id", "")
+    if not ws_id:
+        return JSONResponse({"error": "ws_id is required"}, status_code=400)
+    _user_id, err = _require_ws_access(request, ws_id)
+    if err:
+        return err
+
+    registry = getattr(request.app.state, "registry", None)
+    config_store = getattr(request.app.state, "config_store", None)
+    alias = resolve_role_alias(config_store=config_store, registry=registry, role="stt")
+    if not alias:
+        return JSONResponse(
+            {
+                "error": (
+                    "Speech-to-text is not configured. Assign an STT model role in Models → Roles."
+                )
+            },
+            status_code=503,
+        )
+
+    got = await read_multipart_file_or_400(request, field="audio", max_bytes=_STT_UPLOAD_CAP)
+    if isinstance(got, JSONResponse):
+        return got
+    _filename, _claimed_mime, data = got
+    if not data:
+        return JSONResponse({"error": "Empty audio upload"}, status_code=400)
+
+    stt_prompt = ""
+    if config_store is not None:
+        stt_prompt = (config_store.get("audio.stt_prompt") or "").strip()
+
+    # Resolve + transcode + open the stream eagerly (off the event loop) so the
+    # common failures map to a clean status before any bytes are sent.
+    try:
+        deltas = await asyncio.to_thread(
+            transcribe_stream, registry=registry, alias=alias, data=data, prompt=stt_prompt
+        )
+    except AudioUnavailableError as exc:
+        return JSONResponse({"error": str(exc)}, status_code=503)
+    except AudioBackendError:
+        log.warning("speech_to_text_stream.backend_failed", exc_info=True)
+        return JSONResponse({"error": "Speech transcription backend failed"}, status_code=502)
+
+    # Drive the blocking stream from one worker thread that owns (and closes)
+    # the upstream connection, handing deltas to the loop via a queue.  A client
+    # disconnect sets ``stop`` so the thread releases the connection promptly
+    # instead of being pinned mid-``next()`` (which can't be cancelled).
+    async def _body() -> AsyncGenerator[bytes, None]:
+        loop = asyncio.get_running_loop()
+        queue: asyncio.Queue[bytes | None] = asyncio.Queue()
+        stop = threading.Event()
+
+        def _pump() -> None:
+            try:
+                for delta in deltas:
+                    if stop.is_set():
+                        break
+                    loop.call_soon_threadsafe(queue.put_nowait, delta.encode("utf-8"))
+            except Exception:
+                # Mid-stream backend failure: end the partial stream (logged).
+                log.warning("speech_to_text_stream.mid_stream_failed", exc_info=True)
+            finally:
+                close = getattr(deltas, "close", None)
+                if callable(close):
+                    close()
+                loop.call_soon_threadsafe(queue.put_nowait, None)
+
+        loop.run_in_executor(None, _pump)
+        try:
+            while True:
+                chunk = await queue.get()
+                if chunk is None:
+                    break
+                yield chunk
+        finally:
+            stop.set()
+
+    return StreamingResponse(_body(), media_type="text/plain; charset=utf-8")
+
+
 async def text_to_speech(request: Request) -> Response:
    """POST /v1/api/tts — synthesize assistant text into playable audio."""
    from turnstone.core.audio import (
@@ -2215,74 +2311,6 @@ async def delete_workstream_endpoint(request: Request) -> JSONResponse:
        return JSONResponse({"error": "Delete failed"}, status_code=500)


-async def refresh_workstream_title(request: Request, ws_id: str = "") -> JSONResponse:
-    """POST /v1/api/workstreams/{ws_id}/refresh-title — regenerate workstream title via LLM."""
-    from turnstone.core.log import get_logger
-    from turnstone.core.memory import get_workstream_display_name
-
-    log = get_logger(__name__)
-    ws_id = request.path_params.get("ws_id", "")
-    log.info("ws.title.refresh_requested", ws_id=ws_id[:8] if ws_id else "empty")
-    mgr = request.app.state.workstreams
-    _owner, err = _require_ws_access(request, ws_id, mgr=mgr)
-    if err:
-        return err
-    ws = mgr.get(ws_id)
-    if not ws or not ws.session:
-        log.warning(
-            "ws.title.refresh_failed",
-            ws_id=ws_id[:8] if ws_id else "empty",
-            reason="workstream_not_found",
-        )
-        return JSONResponse({"error": "Workstream not found or not active"}, status_code=404)
-    # Fetch current title so the LLM can generate something different
-    current_title = get_workstream_display_name(ws_id) or ""
-    log.info("ws.title.refresh_triggered", ws_id=ws_id[:8], current_title=current_title[:50])
-    ws.session.request_title_refresh(current_title)
-    return JSONResponse({"status": "ok"})
-
-
-async def set_workstream_title(request: Request, ws_id: str = "") -> JSONResponse:
-    """POST /v1/api/workstreams/{ws_id}/title — set workstream title manually.
-
-    Stores the user-chosen title as the workstream *alias* so it takes
-    priority over the LLM auto-generated title in the display name
-    fallback chain (alias -> title -> name).
-    """
-    from turnstone.core.log import get_logger
-    from turnstone.core.memory import set_workstream_alias
-    from turnstone.core.web_helpers import read_json_or_400
-
-    log = get_logger(__name__)
-    ws_id = request.path_params.get("ws_id", "")
-    log.info("ws.title.set_requested", ws_id=ws_id[:8] if ws_id else "empty")
-    if not ws_id:
-        return JSONResponse({"error": "ws_id is required"}, status_code=400)
-    mgr = request.app.state.workstreams
-    _owner, err = _require_ws_access(request, ws_id, mgr=mgr)
-    if err:
-        return err
-    body = await read_json_or_400(request)
-    if isinstance(body, JSONResponse):
-        return body
-    title = str(body.get("title", "")).strip()
-    if not title:
-        return JSONResponse({"error": "title is required"}, status_code=400)
-    title = title[:80]
-    if not set_workstream_alias(ws_id, title):
-        log.warning("ws.title.set_alias_conflict", ws_id=ws_id[:8], title=title[:50])
-        return JSONResponse(
-            {"error": "That name is already used by another workstream"},
-            status_code=409,
-        )
-    log.info("ws.title.set_alias_updated", ws_id=ws_id[:8])
-    ws = mgr.get(ws_id)
-    if ws and ws.session and ws.session.ui:
-        ws.session.ui.on_rename(title)
-    log.info("ws.title.set_success", ws_id=ws_id[:8], title=title)
-    return JSONResponse({"status": "ok", "title": title})
-
-
 def _auth_user_id(request: Request) -> str:
    """Return the authenticated user's id (empty string when absent).

@@ -3865,6 +3893,8 @@ def create_app(
    history_handler = make_history_handler(interactive_endpoint_config)
    export_handler = make_export_handler(interactive_endpoint_config)
    detail_handler = make_detail_handler(interactive_endpoint_config)
+    refresh_title_handler = make_refresh_title_handler(interactive_endpoint_config)
+    set_title_handler = make_set_title_handler(interactive_endpoint_config)
    v1_routes: list[Any] = [
        Route("/api/events/global", global_events_sse),
    ]
@@ -3879,8 +3909,8 @@ def create_app(
            detail=detail_handler,  # lifted: shared body (interactive feature gain)
            open=open_handler,  # lifted: shared body
            close=close_handler,  # lifted: shared body
-            refresh_title=refresh_workstream_title,
-            set_title=set_workstream_title,
+            refresh_title=refresh_title_handler,  # lifted: shared body
+            set_title=set_title_handler,  # lifted: shared body
            send=send_handler,  # lifted: shared body (P1.5)
            dequeue=dequeue_handler,  # lifted (P1.5) — DELETE /send
            approve=approve_handler,  # lifted: shared body
@@ -3901,6 +3931,13 @@ def create_app(
            methods=["POST"],
        )
    )
+    v1_routes.append(
+        Route(
+            "/api/workstreams/{ws_id}/speech-to-text/stream",
+            speech_to_text_stream,
+            methods=["POST"],
+        )
+    )
    v1_routes.append(Route("/api/tts", text_to_speech, methods=["POST"]))

    app = Starlette(
@@ -1580,32 +1580,62 @@ class Pane {
      this._micBtn.classList.add("is-busy");
    }
    voiceAnnounce("Transcribing…");
+    const resetMic = () => {
+      if (this._micBtn && !this._micDenied) {
+        this._micBtn.disabled = !!this.busy;
+        this._micBtn.classList.remove("is-busy");
+      }
+    };
    authFetch(
      this._base +
        "/v1/api/workstreams/" +
        encodeURIComponent(this.wsId) +
-        "/speech-to-text",
+        "/speech-to-text/stream",
      { method: "POST", body: fd },
    )
-      .then((r) => r.json().then((body) => ({ ok: r.ok, body })))
-      .then((res) => {
-        if (!res.ok) {
-          showToast(
-            (res.body && res.body.error) || "Transcription failed",
-            "error",
-          );
+      .then(async (r) => {
+        if (!r.ok) {
+          let msg = "Transcription failed";
+          try {
+            const body = await r.json();
+            if (body && body.error) msg = body.error;
+          } catch (_e) {
+            /* non-JSON error body */
+          }
+          showToast(msg, "error");
          return;
        }
-        const text = (res.body && res.body.transcript) || "";
-        if (text && this.inputEl) {
-          const cur = this.inputEl.value || "";
-          this.inputEl.value = cur
-            ? cur.replace(/\s*$/, "") + " " + text
-            : text;
+        // Stream transcript deltas into the composer as they arrive (first word
+        // in ~0.3s) instead of waiting for the whole transcript.
+        const reader = r.body.getReader();
+        const decoder = new TextDecoder();
+        let started = false;
+        let got = false;
+        while (true) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          const chunk = decoder.decode(value, { stream: true });
+          if (!chunk || !this.inputEl) continue;
+          got = true;
+          if (!started) {
+            // Read the composer's value now (not before the await) so text the
+            // user typed while transcribing isn't clobbered.
+            const cur = this.inputEl.value || "";
+            this.inputEl.value = cur
+              ? cur.replace(/\s*$/, "") + " " + chunk
+              : chunk;
+            started = true;
+          } else {
+            this.inputEl.value += chunk;
+          }
          // Drive the composer's auto-resize + send-enable listeners.
          this.inputEl.dispatchEvent(new Event("input", { bubbles: true }));
-          this.inputEl.focus();
+        }
+        if (got) {
+          if (this.inputEl) this.inputEl.focus();
          voiceAnnounce("Transcript added to message.");
+        } else {
+          showToast("No speech detected", "error");
        }
      })
      .catch((err) => {
@@ -1614,12 +1644,7 @@ class Pane {
          "error",
        );
      })
-      .finally(() => {
-        if (this._micBtn && !this._micDenied) {
-          this._micBtn.disabled = !!this.busy;
-          this._micBtn.classList.remove("is-busy");
-        }
-      });
+      .finally(resetMic);
  }

  _addTtsAction(el) {
@@ -837,6 +837,11 @@ async function mountShell() {
        });
        pane.tabMenu = () =>
          convTabMenu(pane, pm, id, {
+            // Coordinators carry titles like interactive workstreams now:
+            // surface Refresh/Edit title. The default base ("") targets the
+            // console origin, where the coord refresh-title / title routes
+            // are mounted (same base coordinator.js posts every verb to).
+            titleVerbs: true,
            closeSession: () => {
              if (pane._ctl && pane._ctl.closeSession) pane._ctl.closeSession();
            },
@@ -172,7 +172,7 @@ wheels = [

 [[package]]
 name = "anthropic"
-version = "0.109.1"
+version = "0.108.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "anyio" },
@@ -184,9 +184,9 @@ dependencies = [
    { name = "sniffio" },
    { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/54/0b/ce24a4f275573f5e436ca954faca60c759d58ed152b8fa36a1e3b888e261/anthropic-0.109.1.tar.gz", hash = "sha256:83e06b3d9d40ff5898f588020e0cc4e42187de954549a3b5fbe6e2685a09c785", size = 927569, upload-time = "2026-06-09T23:55:24.884Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/ae/c7/d7f6d2e3975893958081f0282751217757333a3830d0d95859023d7006d0/anthropic-0.108.0.tar.gz", hash = "sha256:91b70253debb477a99f7ca43dac3f71e52207db79d4b06f104080b8dd1693e3b", size = 909409, upload-time = "2026-06-09T16:37:43.584Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/91/0f/a6110d713370bc92f074a622f8a5ebdec7e92360149b1048dca258a07b2f/anthropic-0.109.1-py3-none-any.whl", hash = "sha256:ce7d94a7657f2aa29338cca448945eac621b4f62c1794cf461cb32847223e9b8", size = 923851, upload-time = "2026-06-09T23:55:23.348Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/40/75a937ddd8f230ec129d27de60df69ce8afcab1d0b15f7d651a5a95fac8a/anthropic-0.108.0-py3-none-any.whl", hash = "sha256:bdee7b14c13cf5a60b2c8ae0cf195720e0ea7fd8ab90df5a3899c50f1c91c4be", size = 870079, upload-time = "2026-06-09T16:37:44.895Z" },
 ]

 [[package]]
@@ -1419,7 +1419,7 @@ wheels = [

 [[package]]
 name = "openai"
-version = "2.41.1"
+version = "2.41.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "anyio" },
@@ -1431,9 +1431,9 @@ dependencies = [
    { name = "tqdm" },
    { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/40/36/4c926a91554483977608951360c18c2e911592785eb87a6437813f6123f7/openai-2.41.1.tar.gz", hash = "sha256:23d617a0432457ad844973bee8f540be9da90894f7c5686852d2d365da058f57", size = 783584, upload-time = "2026-06-10T16:10:37.667Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/3c/a6/5815fe2e2aca74b36c650d1bd43b69827cee568073d0d2d9b6fc5aaac80c/openai-2.41.0.tar.gz", hash = "sha256:db5c362acd6604b84f076abbefa66826ea4b46ecba2954ed866e6a149a1352c0", size = 783525, upload-time = "2026-06-03T22:39:40.719Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/20/74/925d7b3892927e9804aaf58d374a45dc28e4420ff90e992272b77286343e/openai-2.41.1-py3-none-any.whl", hash = "sha256:a939565f350cb7443cb843b801b88c716ac8024b492fb94ca269d5f6b1bbefd6", size = 1353380, upload-time = "2026-06-10T16:10:35.756Z" },
+    { url = "https://files.pythonhosted.org/packages/be/51/d82bb424e8aa372190c5233253a2ceb399a778747d18b42cff487411e663/openai-2.41.0-py3-none-any.whl", hash = "sha256:20cc7952e8501c7e5773dd2ef7be437bae9cb549044902e1041a83a54516e375", size = 1353378, upload-time = "2026-06-03T22:39:38.964Z" },
 ]

 [[package]]
@@ -1924,7 +1924,7 @@ wheels = [

 [[package]]
 name = "pytest"
-version = "9.1.0"
+version = "9.0.3"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "colorama", marker = "sys_platform == 'win32'" },
@@ -1933,9 +1933,9 @@ dependencies = [
    { name = "pluggy" },
    { name = "pygments" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/84/0e/b5858858d74958632c49b72cb25a3976ff9f632397626715be71c89d3971/pytest-9.1.0.tar.gz", hash = "sha256:41dd9148c08072446394cefd3d79701701335a9f4cae69ba92e39f6c7f5c061c", size = 1634181, upload-time = "2026-06-13T18:52:45.983Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/7d/0d/549bd94f1a0a402dc8cf64563a117c0f3765662e2e668477624baeec44d5/pytest-9.0.3.tar.gz", hash = "sha256:b86ada508af81d19edeb213c681b1d48246c1a91d304c6c81a427674c17eb91c", size = 1572165, upload-time = "2026-04-07T17:16:18.027Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/8b/5a/ba30a81239b909821b3153e303e7def45178bf353da4f72380e6c5e8793b/pytest-9.1.0-py3-none-any.whl", hash = "sha256:8ebb0e7888bdf2bdfc602ec51f8f62d50200af37356c74e503c79a94f5c81f32", size = 386453, upload-time = "2026-06-13T18:52:44.045Z" },
+    { url = "https://files.pythonhosted.org/packages/d4/24/a372aaf5c9b7208e7112038812994107bc65a84cd00e0354a88c2c77a617/pytest-9.0.3-py3-none-any.whl", hash = "sha256:2c5efc453d45394fdd706ade797c0a81091eccd1d6e4bccfcd476e2b8e0ab5d9", size = 375249, upload-time = "2026-04-07T17:16:16.13Z" },
 ]

 [[package]]
@@ -2224,27 +2224,27 @@ wheels = [

 [[package]]
 name = "ruff"
-version = "0.15.17"
+version = "0.15.16"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/8c/a9/3abdf488f1bf3d24c699415e454ed554a6350d5d89ce183be1ee0a3361ac/ruff-0.15.17.tar.gz", hash = "sha256:2ec446937fd16c8c4de2674a209cc5af64d9c6f17d21fbf1151054fa0bcf5219", size = 4743346, upload-time = "2026-06-11T17:54:47.663Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/a6/bd/5f7ec371001337d8fa61701c186ff8b613ecac1651848c5950f4c4d5f2e9/ruff-0.15.16.tar.gz", hash = "sha256:d05e78d38c78caf020b03789e25106c93017db5a0cb6e2819885018c61343b78", size = 4714267, upload-time = "2026-06-04T16:33:09.974Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/db/4d/e11259f5da07cb6afb2d074c31bf09da9671993f7329d4f15d2fdc458301/ruff-0.15.17-py3-none-linux_armv6l.whl", hash = "sha256:d9feddb927fc68bd295f5eebc587a7e42cfaf9b65f60ca4a2386febff575da8f", size = 10856677, upload-time = "2026-06-11T17:54:49.533Z" },
-    { url = "https://files.pythonhosted.org/packages/29/3e/772d679e1a0dc058e58875bd2c0cb713a0530877b4a76fee3c7966df0d49/ruff-0.15.17-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:25805a226d741c47d274a35ad5c10a7dde175fcddfa511d7cf3da0a21eb3eab7", size = 11223443, upload-time = "2026-06-11T17:55:00.573Z" },
-    { url = "https://files.pythonhosted.org/packages/68/58/bd41f7688b2fd5623012605130ed70e60aa7f2244baa3d5066bdd61530c8/ruff-0.15.17-py3-none-macosx_11_0_arm64.whl", hash = "sha256:f6ad73b14c2d18a3bf8ad7cb6974294d7f613a7898604826058e6ac64918ef4d", size = 10566458, upload-time = "2026-06-11T17:55:07.52Z" },
-    { url = "https://files.pythonhosted.org/packages/d8/5b/733371013fcf1ec339e477ece6ab42bfe10bdd9bba8ee88a9516aa56bfc0/ruff-0.15.17-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6ba0c1e4f95bcb3869d0d30cbd5917071ef2e28665abfec970cdab0492c713ed", size = 10914483, upload-time = "2026-06-11T17:55:05.501Z" },
-    { url = "https://files.pythonhosted.org/packages/bd/cc/6f24251cc0252f7239391ccb85833f320efad14ebe5b443943f37ced6332/ruff-0.15.17-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:81647960f10bff57d2e51cadd0c3950fe598400c852863a038720ef5b8cca91e", size = 10647497, upload-time = "2026-06-11T17:54:57.733Z" },
-    { url = "https://files.pythonhosted.org/packages/68/dd/0d10c17ce1a1624d6fc3156309c3f834fdb5dfaad026ec90c85684f3990e/ruff-0.15.17-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0e01a84ddbc8c16c23055ba3924476850f1bbc1917cebbb9376665a63e74260d", size = 11416967, upload-time = "2026-06-11T17:54:51.461Z" },
-    { url = "https://files.pythonhosted.org/packages/2f/91/556bfb156f6144f355e831c23db00b2fc4120f86b3ce81cc5f7fd2df51f3/ruff-0.15.17-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:84fe9f653152f8f294f9f7e03bf3a453d8b4a27f7a59c78c8666167f2b17b96c", size = 12335770, upload-time = "2026-06-11T17:54:45.793Z" },
-    { url = "https://files.pythonhosted.org/packages/88/82/8b5999aa13355e926f06d9f42a32dcca862f623bf0363785ff89d607dffd/ruff-0.15.17-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8c0fe88a7676e7a05b73174d4d4a59cb2ac21ff8263583f87a81a6018475a978", size = 11575441, upload-time = "2026-06-11T17:54:32.661Z" },
-    { url = "https://files.pythonhosted.org/packages/11/93/f10377bb04109ca0e8cbc483ff1982c54b6d418210041776f93e8cdc7fa9/ruff-0.15.17-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ecfc3c7878fff94633ab0348524e093f9ce3243080416dd7d14f8ba400174719", size = 11557614, upload-time = "2026-06-11T17:54:34.698Z" },
-    { url = "https://files.pythonhosted.org/packages/c7/a6/eeeae7f7d5493df41649ab3db92f086b2d0a30199e4efdf8e3dd7a033f24/ruff-0.15.17-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:b8461180b22420b1bdc289909410930761629fddf2a5aaf60fae1ab26cedc4c4", size = 11544450, upload-time = "2026-06-11T17:54:39.042Z" },
-    { url = "https://files.pythonhosted.org/packages/32/88/5991ce565129a24dd4a00db1254b3b5db2e53018cbe4018ea5a89738e727/ruff-0.15.17-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:6eccbe50a038b503e7140b441aa9c7fc8c1f36edf23ebef9f4165c2f28f568b7", size = 10892524, upload-time = "2026-06-11T17:55:09.432Z" },
-    { url = "https://files.pythonhosted.org/packages/f5/1d/0fdd248313425f55223968af04b0a42125466a8d88d21c1d99c6af0a51e8/ruff-0.15.17-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:382fc0521025f5a8ad447d8bdd523545d0d7646adb718eb1c2dac5065ec27c0f", size = 10659573, upload-time = "2026-06-11T17:54:36.824Z" },
-    { url = "https://files.pythonhosted.org/packages/9e/0e/072e8260deb9461062ce9311ced27a8e541229a6ffd483013dd37661e43e/ruff-0.15.17-py3-none-musllinux_1_2_i686.whl", hash = "sha256:456d41fcd1b2777ad63f09a6e7121d43f7b688bbc76a800c10f7f8fb1f912c3f", size = 11127818, upload-time = "2026-06-11T17:55:03.124Z" },
-    { url = "https://files.pythonhosted.org/packages/ab/b4/55060a34163121498014696b5f656db5b8c6963768f227dbf0d76b311073/ruff-0.15.17-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:b1a04bcc94ae6194e9db05d16ad31f298a7194bfbcb08258bbe589cee1d587b8", size = 11655901, upload-time = "2026-06-11T17:54:53.562Z" },
-    { url = "https://files.pythonhosted.org/packages/49/71/9b29d6b87cef468d697f43c6a91e3fae4a80185779d7d5a4ef27d173439f/ruff-0.15.17-py3-none-win32.whl", hash = "sha256:596065960ab1ff593f744220c9fe6580eda00a95003cffa9f4048bb5b1bf0392", size = 10925574, upload-time = "2026-06-11T17:54:55.723Z" },
-    { url = "https://files.pythonhosted.org/packages/3d/b2/8fc77f3723228836fa5d12497eb71c808f83782e10d058d2b15cfa14640b/ruff-0.15.17-py3-none-win_amd64.whl", hash = "sha256:6769e5fa1710b179b92e0bfa5a51735b35baea9013dadb06d5f44cbcf9547084", size = 12058788, upload-time = "2026-06-11T17:54:41.042Z" },
-    { url = "https://files.pythonhosted.org/packages/2d/c7/c53e8dbff9c9dc4b7928773421ae294a5d28fcb8dcda1a089579d3a7e510/ruff-0.15.17-py3-none-win_arm64.whl", hash = "sha256:f3be1fbb34bcdfd146240d8fb92a709d4c2c8191348580a3c044ec60fa0b4456", size = 11355275, upload-time = "2026-06-11T17:54:43.635Z" },
+    { url = "https://files.pythonhosted.org/packages/0c/42/53ef1c3953f157956db9bf7861e3bc50b9b887ce93300aa48cdba8336fe6/ruff-0.15.16-py3-none-linux_armv6l.whl", hash = "sha256:6ac3c0b3969cc6cf6b158c4e2f8f682acb58e7d700d8a44b65ecdc72d66ab0b2", size = 10709025, upload-time = "2026-06-04T16:32:51.935Z" },
+    { url = "https://files.pythonhosted.org/packages/93/9a/a79159346f19134a956607754e57d8d128f7a4c00f4ad2f7514d224c172c/ruff-0.15.16-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:197c207ed75ffba54a0dec23db4aa939a27a3053073e085e0042433cbdc58e4a", size = 11063550, upload-time = "2026-06-04T16:32:42.24Z" },
+    { url = "https://files.pythonhosted.org/packages/bc/72/3ce2ac000a5299ec238e01f51397b3b653c93b077d9b1bfe8715bb895f20/ruff-0.15.16-py3-none-macosx_11_0_arm64.whl", hash = "sha256:3a39fec45ab316cc23e7558f23fea4a70403ddb5648ea9a4a3854a16973d0071", size = 10421345, upload-time = "2026-06-04T16:32:37.251Z" },
+    { url = "https://files.pythonhosted.org/packages/b0/c2/cc7fad3ec9169373f5b6a18f1917b91080feec40c3f9658334a1d28e2f03/ruff-0.15.16-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ba93191d79003116b95128c9d306e045200fdbd0bccb782b110f3cd1d4abc5cf", size = 10757217, upload-time = "2026-06-04T16:32:54.722Z" },
+    { url = "https://files.pythonhosted.org/packages/69/d2/3474009eaa0a65b31fa7152a2fad5e2f050c640ceb1e6b02ee6922e94c82/ruff-0.15.16-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:c6ee4b90520630120ef032aa5cc10db483852dff950e78b1d717e2993a61ac8d", size = 10507035, upload-time = "2026-06-04T16:33:05.343Z" },
+    { url = "https://files.pythonhosted.org/packages/ca/81/b7ae6ccbd11f0c8dc3d5d67fc4be9b57ff57ca86ba56152021378e1277f2/ruff-0.15.16-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4e4215bc938bc3c8215c1472c1aa437e310fee20cd427335fec9d7e609563628", size = 11255291, upload-time = "2026-06-04T16:32:49.49Z" },
+    { url = "https://files.pythonhosted.org/packages/d9/e1/46e526f1a7cc90857ce6ddf25fbb77eb6568651ac38d71b033af07076dd5/ruff-0.15.16-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7c8d26be963b090f10e29abc8b3e74a2a321f6fa34e02424e30b5af89350ecbb", size = 12124922, upload-time = "2026-06-04T16:33:07.821Z" },
+    { url = "https://files.pythonhosted.org/packages/1a/da/5c791b088b596b24d0deb967fa28ae02ad751a140c0b9ea81c5ab915d6c0/ruff-0.15.16-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f198cf4123602a2280ed46c307bcbafe41758d6fee5b456b6b6058ca1514b3b4", size = 11332186, upload-time = "2026-06-04T16:33:02.971Z" },
+    { url = "https://files.pythonhosted.org/packages/72/11/5da87abe20047c8962361473923ebb2f62b595250126aadfad8c20649c1e/ruff-0.15.16-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bb27515fa6240fb586ae82b901a59e67d24acff86f2190b433dc542fe0435aeb", size = 11373541, upload-time = "2026-06-04T16:32:47.007Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/2a/8554754c23a854ae3fd6b507e36ad61ddb121e298c6d5d617dec94ed0f14/ruff-0.15.16-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:a267c46ba1593fc26b8eecbea050b39d40c0b6bb7781ee11c90a02cd10032951", size = 11353014, upload-time = "2026-06-04T16:32:34.795Z" },
+    { url = "https://files.pythonhosted.org/packages/62/25/62ea41529ec89f742ea3fed9cb1059c72877ec7cf9b9e99ac9cf3294d1d9/ruff-0.15.16-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:528c68f39a91498a8d50e91ff5985df3d105782bab49cc378e73ac26bff083e8", size = 10737467, upload-time = "2026-06-04T16:32:26.348Z" },
+    { url = "https://files.pythonhosted.org/packages/90/17/334d3ad9de4d40f9dd58fdd09e35ce64553bb501e2f19a839e2fb6be14fc/ruff-0.15.16-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:7ed55c58950df60589a9a7a5d2f8fa5f54ebd287163be805adfe6ee95a9de123", size = 10521910, upload-time = "2026-06-04T16:32:32.54Z" },
+    { url = "https://files.pythonhosted.org/packages/4d/bd/3ac7c6ae77a885c1004b3dda2446ea401768d24f851c14b4ad4b24f6639c/ruff-0.15.16-py3-none-musllinux_1_2_i686.whl", hash = "sha256:d482feaf51512b50f9790ceb417a56a61dd1e9d9bf967662b9ed27c01b34f53a", size = 10979190, upload-time = "2026-06-04T16:32:57.492Z" },
+    { url = "https://files.pythonhosted.org/packages/33/d7/609546e6a413c3f216fbf2a50c928f97c80939154f6a0503114094a86191/ruff-0.15.16-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:1e15bc8c94513dae2a40cc9ef07c94fdd4ecc9e29dabebeebe170f952322c9e3", size = 11477014, upload-time = "2026-06-04T16:32:44.687Z" },
+    { url = "https://files.pythonhosted.org/packages/74/0d/f2cd247ad32633a5c36e97141a2c21b11c6279f7957bc2ff360b1e08fddd/ruff-0.15.16-py3-none-win32.whl", hash = "sha256:580378f7bd4aa25f72e74aa54948a9622f142b1e509521dd10902e886681cc1e", size = 10735541, upload-time = "2026-06-04T16:32:30.145Z" },
+    { url = "https://files.pythonhosted.org/packages/8b/9e/02e845ef151b1dee585e55c4739f8e1734ae1d9f1221dff65761c162208b/ruff-0.15.16-py3-none-win_amd64.whl", hash = "sha256:408256017284eddf98fff77b29aa4fb30f586042d535b2d9befc6512f400aaec", size = 11843403, upload-time = "2026-06-04T16:32:39.76Z" },
+    { url = "https://files.pythonhosted.org/packages/15/19/016553f86f207450aebebc2b2b5088d086b901cc8186c02ac4284db3bd88/ruff-0.15.16-py3-none-win_arm64.whl", hash = "sha256:8cd61783afb39638a7133ef0d2dfb1e91277593962f81b5a8423eb0b888a6121", size = 11134555, upload-time = "2026-06-04T16:33:00.136Z" },
 ]

 [[package]]
@@ -2425,19 +2425,19 @@ wheels = [

 [[package]]
 name = "tqdm"
-version = "4.68.2"
+version = "4.68.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "colorama", marker = "sys_platform == 'win32'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/85/05/0d5260f1f1ca784f4a4a0def9cbe6affe587f5b4025328d446c3d67765f4/tqdm-4.68.2.tar.gz", hash = "sha256:89c230e8dbc67c7615c142487111222f878c77427ea09549960f62389e258add", size = 171923, upload-time = "2026-06-09T13:26:42.539Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/06/b3/36c8ecf72e8925200671613332db156d84b99b3aee742a41c1938ebb0808/tqdm-4.68.1.tar.gz", hash = "sha256:fc163d96b287bd031e1aa24421ce4411b25559bd0a1be4fe649bdaa4d2c02bf5", size = 171236, upload-time = "2026-06-05T17:23:15.267Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/eb/75/1a0392bcc21c44dcdf87b3cf2d137e7829be2c083a1e38d44efca3d57a16/tqdm-4.68.2-py3-none-any.whl", hash = "sha256:d4240441fb5353290b87d6a85968c9decc131a99b8c7faa28269d829de669ede", size = 78578, upload-time = "2026-06-09T13:26:40.731Z" },
+    { url = "https://files.pythonhosted.org/packages/47/aa/218a0eb34de1f753c83e4d0d1c8e7c4cef27f20dcb8342e024f63a80dc86/tqdm-4.68.1-py3-none-any.whl", hash = "sha256:fea4a90e4023f764914569f7802a297277c5ab1a66be5144143e142e1a4031d8", size = 78354, upload-time = "2026-06-05T17:23:13.654Z" },
 ]

 [[package]]
 name = "turnstone"
-version = "1.7.0a2"
+version = "1.6.9"
 source = { editable = "." }
 dependencies = [
    { name = "alembic" },