* fix(agents): keep session model selection canonical
Preserve configured session model identity across response aliases and turn-local fallbacks.
* fix(gateway): align session model projection consumers
* test(gateway): mark session event model selection
* fix(agents): resolve session-stable reply mode for direct resolver turns
Heartbeat wakes and other direct getReplyFromConfig callers bypass
dispatch, so no session-stable delivery mode reached their CLI binding
facts and prepare computed no messageToolPolicyHash, while dispatched
chat turns hashed the stable mode. resolveCliSessionReuse treats that
one-sided hash as a policy change, hard-invalidating the binding on
every chat<->heartbeat transition and leaving heartbeats to run in
fresh CLI sessions with no conversation history (historyPrompt=none).
Extract the synthetic-turn stable-mode resolution that
prepareAgentCommandExecution already used into
resolveSessionStableReplyMode, and apply it in prepareReplyRunContext
whenever a synthetic turn arrives without dispatch's injected mode, so
every turn kind on a session derives the same binding facts and hash.
Fixes#121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): derive synthetic binding facts from the dispatch policy owner
Consolidate the synthetic-turn stable-mode resolution onto dispatch's own
policy pieces instead of a simplified parallel resolver:
- resolveVisibleRepliesPolicy is extracted from dispatch prepare-context
(configured visible-replies guard + harness default chain) and shared, so
the fallback and dispatch cannot drift on harness defaults, live-vs-entry
chat facts, or ExplicitDeliverRoute.
- The stable fact now applies dispatch's messageToolAvailable downgrade via
the canonical tool-policy resolvers: tool-only delivery with a policy-denied
message tool records automatic, matching dispatch.
- Synthetic turns no longer fall back to their effective turn mode: a
response-tool heartbeat's message_tool_only is per-turn enforcement, not
session policy, and previously bypassed the stable fallback entirely.
Regression coverage: response-tool heartbeat facts match dispatched turns,
and a message-tool-denied config downgrades the synthetic stable fact.
Refs #121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): keep the session-stable reply mode sender-independent
Dispatch's stable-mode downgrade used the turn's sender-aware message-tool
availability, so a sender-scoped message denial hashed the stable policy as
automatic on chat turns while sender-less synthetic turns hashed tool-only —
the same binding reset loop on another axis.
resolveStableMessageToolAvailability becomes the one sender-independent
owner: dispatch passes it as sessionStableMessageToolAvailable for the
stable-mode resolution only (effective per-turn enforcement keeps the
sender-aware verdict), and synthetic binding facts already consume it.
Computed only when the visible-replies candidate is message_tool.
Refs #121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): derive stable reply facts from session surfaces, not wake plumbing
Third-pass review fixes on the #121485 consolidation:
- Strip system-event wake providers ("heartbeat", "cron-event") from the
stable context and resolve origin-less entries as internal, so synthetic
facts take the same internal-channel branch dispatch's live webchat turns
do instead of falling through to harness defaults.
- Fall back to persisted session facts (delivery channel/origin, groupId,
groupChannel/subject, accountId) for the sender-independent availability
stack, so bare-ctx callers like command prepare resolve the same
group/account-scoped policies as dispatched turns; dispatch now passes its
session entry too.
- Collapse the redundant second mode resolution into the availability
downgrade, surface dispatch's injected stable mode separately from
resolvePromptSourceReplyMode so the synthetic slot has one owner, and
reuse resolveTurnModelOverride instead of an inline copy.
- Refresh docs/.generated/plugin-sdk-api-baseline.jsonl: closure hashes only
(17 exports, zero declaration changes, verified) — plugin-sdk reply/meeting
entrypoints transitively bundle the reply pipeline this branch refactors.
Regression: origin-less entry heartbeat resolves internal-automatic facts.
Refs #121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: adapt to main module moves and split API baseline
Mechanical rebase refresh onto d3696f8d76: subagent-capabilities moved to
src/agents/subagents/spawn/, and the plugin-sdk API baseline is now
per-entrypoint content hashes (#122082) — regenerated for the reply-pipeline
closure this branch touches.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* refactor: consolidate coercion ownership
Centralize four canonical coercion helpers, migrate exact core and plugin duplicates through narrow Plugin SDK facades, and enforce declaration and plugin-normalization ownership boundaries.
The sweep adds eight focused SDK exports while deleting more production and tooling code than it adds. User-visible behavior is unchanged except for safer equivalent object and UI parsing at existing boundaries.
* fix: guard integer option ownership
Register resolveIntegerOption with the canonical function owner and extend the declaration-guard fixture so future local duplicates fail validation.
* fix: keep integer helpers on numeric facade
Remove the unshipped duplicate string-coerce exports and route every affected plugin consumer through the existing number-runtime contract.
* fix: point numeric coercion to number runtime
Make boundary and declaration diagnostics recommend the canonical numeric facade, with failing-before coverage for both guidance paths.
Close the barrel gap (SessionEntryListScope), migrate the 14 production
type-bypass imports and the sessions-rewind raw-op imports through the
accessor barrel, and document session-accessor.sqlite-scope as the
sanctioned low-level entry point for doctor/migrations/infra. Net +8
production LOC: the barrel gains its missing surface. Part 3 (final) of
the session-accessor dual-layer collapse (#121316, #121536).
* refactor(skills): split workspace loading into concept modules
* refactor(skills): delete dead seams and duplicate helpers
* chore(lint): ratchet max-lines baseline after workspace split
* chore(plugin-sdk): refresh API baseline
* refactor(skills): internalize discovery result type
* test(tooling): drop stale unit-fast fixtures
Re-lands the reverted #121507 on the current custody contracts. The reply dispatcher now owns direct-send pending-final custody: claim before provider I/O, terminal settlement for delivered/suppressed/failed outcomes, proven no-send stays replayable, ambiguous evidence fails closed — so Gateway restarts can no longer duplicate an already-accepted final reply.
Proof: ClawSweeper local review clean, exact-head ci-gate green, live Telegram E2E (one turn, one final, no duplicates).
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* refactor(agents): split claude-live-session into concept modules
* refactor(agents): delete duplicated live-session helpers
* fix(agents): fence live-session close during pending spawn
* chore(lint): ratchet max-lines baseline after live-session split
* style(agents): satisfy lint on live-session split modules
* fix(agents): fence live session close state
* refactor(agents): extract Claude live turn timeouts
* fix(agents): preserve exec policy after Claude live split
* test(agents): complete Claude live policy session fixture
* refactor(agents): move announce family into subagents/announce/
* refactor(agents): move completion and main-session recovery families
* refactor(agents): move registry family into subagents/registry
* refactor(agents): update registry state type imports
* style(agents): format moved subagent imports
* test(agents): preserve isolated registry test routing
* fix(scripts): recognize relocated subagent announce seams
Keep configured engines selected before the current turn is persisted, then durably advance accepted turns across embedded and CLI paths. Preserve runtime plugin ownership and canonical transcript anchors through prepared execution.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
Co-authored-by: Peter <noreply@anthropic.com>
* fix(codex): report harness context window as session contextTokens
Codex app-server reports model_context_window per turn. Carry it through the projector into the run result meta so session rows show the real window instead of the catalog's standard-tier input cap (272k vs 1M for gpt-5.6 models).
* improve(ui): compact chat context popover
Inline stat rows replace boxed tiles; zero-value cost rows and the whole cost section when empty are omitted; provider/model provenance lines are removed because the footer already shows the model; and the popover is narrowed to 300px.
* refactor(codex): split attempt-result assembly out of event projector
* fix(codex): seed attempt context window from startup binding
App-server v2 turn/started omits the core model_context_window, so thread/tokenUsage/updated is the only live carrier. Seed usage-less attempts from the retained startup binding rollout/session window so session metadata cannot regress to the catalog fallback.
* fix(codex): prefer native startup context window
Persisted session contextTokens has no source provenance and may contain the catalog fallback. Keep the minimum window for the conservative rotation fuse, but seed the projector from the native rollout when it is available.
* chore(plugin-sdk): regenerate api baseline (new format)
* revert(gateway): "prevent restart replay after final delivery" (broke 5 CI jobs)
Record channel delivery custody before recipient-visible sends so restart recovery cannot duplicate accepted finals.
Ambiguous legacy markers now fail closed with a visible interruption instead of blind replay.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* perf(agents): keep turn-path model catalog reads off the full live build
First agent turns (embedded and cron) resolved thinking capability through
loadPreparedModelCatalogSnapshot without readOnly, which materialized the
full live model-runtime catalog: ambient synthetic-auth discovery fanned out
to every registered provider and loaded plugin discovery modules through
jiti source transform (3,172 TS modules, 36s event-loop block, +600MB heap,
58.7s model-selection on a cold gateway).
- add loadProviderScopedThinkingCatalog: manifest metadata first, then a
provider-scoped read-only static catalog, then scoped live discovery only
for runtime-discovery providers (preserves #116584 Ollama semantics)
- route scopedLiveProviderDiscovery through the scoped read-only loader
- scope live-mode ambient synthetic-auth refs to the requested providers
- bound the last-resort synthetic-auth sweep to discovery entry modules
- memoize per-turn plugin skill dir resolution/republish (single-slot,
lifecycle-cleared; was a full walk + symlink republish every turn)
Cold first turn 72.7s -> ~22s wall (remaining cost is provider prefill of
the ~19.5k-token default prompt); model-selection 58,726ms -> 124ms.
* test(agents): align model-catalog.runtime mocks with scoped thinking catalog seam
Explicit vi.mock factories must export every binding prod touches; the new
loadProviderScopedThinkingCatalog export is now mocked everywhere the module
is stubbed, and the live-model-switch Ollama hydration test asserts the new
provider-scoped seam instead of the retired unscoped snapshot call shape.
* test(agents): export scoped thinking catalog from every prepared-catalog mock; split synthetic-auth helpers
- add loadProviderScopedThinkingCatalog to all explicit prepared-model-catalog
and model-catalog.runtime mock factories (vi.mock factories must export every
binding prod touches)
- move synthetic-auth ref scoping/resolution into
prepared-model-runtime.synthetic-auth.ts; keeps facts under the max-lines cap
* test(agents): prove scoped thinking hydration for runtime-only models
Boundary proof for the ClawSweeper review gap: the three-tier helper stops at
manifest or scoped-static when they resolve, and runs provider-scoped live
discovery (no broad fanout) only for runtime-only models; cron selection
hydrates through the same scoped helper and skips it entirely for thinking=off.
* test(agents): accept rest args in scoped thinking catalog mocks
Josh Lehman