* fix(llm-core): settle result() when a stream ends without a terminal event
EventStream.end() with no argument and no prior terminal event left the final
result promise pending forever; any consumer awaiting result() dead-ended
silently for the whole run budget (the agent loop's fall-through in
streamAssistantResponse awaits exactly that promise). All in-repo pumps push
terminals or pass end(result), so this only fires on producer contract
violations — which previously hung and now reject loudly with a named error.
A pre-attached catch keeps iterate-only consumers free of unhandled
rejections. Also dedupe the agent loop's two identical final-message blocks
into one finalizeAssistantMessage helper (net-negative production delta).
Follow-up from #120426's turn-liveness investigation.
* fix(llm-core): void the intentional rejection pre-catch
ClawSweeper P1: the pre-attached no-op catch returns a Promise that would
float under typescript/no-floating-promises.
* fix(ai): preserve Responses server compaction state
Persist opaque Responses compaction items as fenced provider replay state so long stateless sessions can resume authoritative compressed history without exposing it in display or diagnostics. Carry state through worker transcripts and prune replay prefixes without splitting tool pairs.
Release note: Preserve long OpenAI Responses sessions across server-side compaction and worker restarts.
Related: #95788
* test(ai): align long-context fixtures with CI contracts
Make tool-result fixtures type-complete, use the canonical model selector helper, remove unused test-helper exports, and route the paid long-context live probe through the dedicated Gateway profile shard.
* test(ai): type mocked Responses terminal events
Give the mock SSE event collection an explicit open event shape so terminal response events coexist with output-item events under the root test typecheck.
* fix(ai): suppress rejected compaction replay
Persist a route-fenced suppression tombstone when encrypted-content recovery rejects a compaction item, so later turns do not retry the same opaque state. Preserve the tombstone through transcript redaction and cover successful fallback followed by the next turn.
* fix(ai): keep compaction suppression transport-private
Keep the suppression contract local to its sole Responses transport owner and make the regression fixture satisfy root type and lint checks without widening the Plugin SDK surface.
* refactor(ai): remove compaction suppression re-export
* fix(ai): scope compaction suppression to replay route
Keep foreign-route rejection tombstones from hiding the newest compatible Responses compaction while preserving same-route suppression.
* fix(ai): harden Responses replay recovery
Stage encrypted replay recovery so compaction is only suppressed after an attributable rejection. Preserve terminal ordering and keep provider replay within worker frame budgets without truncating opaque state.
* refactor(ai): centralize Responses output indexes
Keep normalized output identity tracking in the stream-slot owner, move response failure state to its diagnostic owner, and remove the obsolete replay clone export so exact-head static gates remain shrink-only.
* fix(ai): retain idless terminal tool identity
Use the canonical empty identity only when a provider supplies neither call nor item id, preventing terminal recovery from duplicating a done-only tool call while preserving stronger identities when available.
* fix(sessions): hide provider replay from public events
* fix(ai): stage encrypted replay recovery
* fix(ai): keep replay attempt kind internal
* fix(ai): route Azure through replay recovery
Use the shared encrypted-content retry owner for Azure Responses so compaction suppression and prompt-observer variants stay coherent across transports.
* fix(ai): harden replay persistence boundaries
Fence Azure replay by the resolved request endpoint, drop invalid replay during transcript sanitization, and surface worker-launch replay omissions through the existing redacted diagnostic path.
* feat(anthropic): add Claude Opus 5 model support
Wire claude-opus-5 through the Claude 5 contract seams: adaptive-by-default
thinking with the full low..max effort range, default-sampling and prefill
stripping, streaming refusal contract, model-bound thinking replay, and 1M/128k
catalog metadata across anthropic, claude-cli, Vertex, Bedrock, and Mantle.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(anthropic-vertex): normalize explicit Vertex Opus 5 model rows
Extend normalizeAnthropicVertexResolvedModel so user-configured Vertex Opus 5
rows regain reasoning, image input, 1M/128k limits, and the native thinking
map; update the live-model priority expectation and regenerate docs_map.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(ai): track Responses output items per index and require terminal stream events
* test(ai): add live Responses stream coverage
* test(infra): mirror packages live glob in live-config test
* feat(tooling): extend strict-ratchet lane to all remaining leaf packages
* fix(packages): burn down indexed-access debt in ai and agent-core
All 196 noUncheckedIndexedAccess errors fixed behavior-identically
(iteration, .at()/slices, DataView byte math, guarded queue peeks) plus
cleanup finds: a sparse-tool diagnostic bug, deduplicated Responses
call/item ID parsing with an honest string|undefined splitter, and a
clearer local failure for redacted-thinking blocks missing signatures.
Removes all 17 remaining non-null assertions across ai, gateway-client,
and llm-core so the assertion ban aligns 1:1 with the ratchet lane.
speech-core joins memory-host-sdk as structurally deferred (imports
src/** via plugin-sdk path mappings).
The active user turn was decorated at request-build time with an inbound
metadata block (Conversation info / reply-target / sender / forwarded /
chat-history), but the session store persists the bare text and the LLM boundary
strips that block from historical replay. So a user message's serialized bytes
changed retroactively — decorated as the active turn, bare as history — on every
turn, invalidating any prefix/exact-match provider prompt cache from that point
onward, every turn. (#90811 fixed the timestamp half of this same asymmetry.)
Rework of the earlier "strip the active turn" approach (which lost model-visible
context — ClawSweeper's P1) into one that keeps the context AND recovers the
cache cost:
- Route current-turn inbound metadata out of the user prompt and into the hidden
runtime-context custom message (the mechanism from openclaw#89428), and
relocate that carrier to the ABSOLUTE TAIL of the wire request (after the
active user turn and any tool-call scaffolding). The user turn stays bare and
byte-identical in both the active and historical positions; the volatile
carrier vanishes next turn exactly where the assistant reply begins anyway, so
the request is an append-only prefix-extension through the active user turn.
- A durable marker (UserMessage.runtimeContextCarrier, set by convertToLlm from
the carrier's details) lets the Anthropic SDK transport, the managed Anthropic
transport (anthropic-transport-stream / anthropic-payload-policy), and
OpenAI-completions skip the carrier when selecting the deepest cache_control
breakpoint, keeping the anchor on the last stable user turn.
Storage stays BARE (invariant preserved); runtime-only turns (room events) keep
their existing inline behavior, which is byte-stable because room-event/system
context is not strip-eligible; legacy bare-stored sessions are unchanged.
Reviewed with adversarial gpt-5.5/codex sweeps (correctness, byte-identity,
compatibility, test adequacy): found + fixed the managed Anthropic transport
cache-anchor path and added the missing OpenAI-completions/managed-transport
breakpoint tests; a theoretical runtime-only mutation was verified unreachable
and pinned with a test. Final sweep clean.
Co-authored-by: openclaw#89428 (runtime-context handoff approach)
* fix(llm): coerce stringified JSON arrays/objects in tool argument validation
When LLMs serialize array or object tool parameters as JSON strings
(e.g. tags: '["test","debug"]' instead of tags: ["test","debug"]),
validateToolArguments now attempts JSON.parse coercion before
rejecting the value. This mirrors the existing numeric string
coercion path and fixes MCP tool calls from providers like MiMo,
Ollama, and others that stringify complex parameters.
Fixes#96916
* fix(llm): bound JSON.parse size for schema-gated array/object coercion
Add MAX_JSON_COERCE_LENGTH (64KB) guard before JSON.parse in the array
and object coercion branches. Oversized stringified arguments are left
for normal validation to reject rather than synchronously parsed.
Addresses Codex review finding: unbounded JSON.parse on model-controlled
tool arguments could block the event loop or spike memory.
* fix(validation): preserve null in anyOf unions instead of coercing to empty string
Fixes#96716
* fix(validation): preserve null in anyOf unions instead of coercing to empty string
* fix(validation): preserve null in anyOf unions instead of coercing to empty string
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Trim dense plain text-delta stream snapshots for OpenAI-compatible, Responses, and Ollama providers while preserving full snapshots on stream checkpoints and terminal events.
Reconstruct partial-less text deltas in the agent loop so live message updates continue to advance for immutable snapshot providers, and document the optional text_delta.partial contract.
Fixes#86599.
Enable Mistral prompt cache keys without long-retention forwarding. Update cached-read pricing and doctor migration for existing Mistral provider config. Fixes#83709.
Adds broad inline comments and JSDoc for CLI, cron, outbound/channel, plugin SDK, ACP, shared helpers, net policy, and related utility contracts. Proof: git diff --check on latest exact head plus focused cron tests passed; CI had no failing checks observed before merge attempt.
Forward OpenAI-compatible stop sequences from gateway chat completions through the agent runner into provider transports.
The gateway now normalizes stop into sampling extras, agent transports pass it into the shared stream options, and OpenAI, Anthropic, Mistral, Google, and Vertex-backed simple providers map it to their native request fields. Provider/gateway/agent coverage plus Crabbox live gateway proof verify valid stop dispatch and invalid stop rejection.
Refs #87920
Peter Steinberger