Summary:
- The PR filters persisted OpenAI-compatible `reasoning_content` thinking placeholders from direct Anthropic replay payloads and updates the focused Anthropic provider test.
- PR surface: Source +1, Tests -4. Total -3 across 2 files.
- Reproducibility: yes. from source: current main serializes `thinkingSignature: "reasoning_content"` as a nat ... rror. The PR body also provides after-fix captured outbound payload proof for the production provider path.
Automerge notes:
- PR branch already contained follow-up commit before automerge: fix(anthropic): drop reasoning_content replay signatures
Validation:
- ClawSweeper review passed for head 6eaa72f3a3.
- Required merge gates passed before the squash merge.
Prepared head SHA: 6eaa72f3a3
Review: https://github.com/openclaw/openclaw/pull/91231#issuecomment-4643786130
Co-authored-by: 宇宙熊Yzx <53250620+849261680@users.noreply.github.com>
Co-authored-by: clawsweeper <274271284+clawsweeper[bot]@users.noreply.github.com>
Co-authored-by: clawsweeper[bot] <274271284+clawsweeper[bot]@users.noreply.github.com>
Approved-by: takhoffman
Co-authored-by: takhoffman <781889+takhoffman@users.noreply.github.com>
Preserve streamed Responses tool-call argument deltas when the final done event omits or sends empty arguments, fixing LM Studio argument-bearing tools from arriving as `{}`.
Fixes#90585.
Thanks @849261680.
Fix MiniMax-M3 Anthropic-compatible requests so OpenClaw no longer sends the disabled-thinking payload that makes M3 return empty content. M3 defaults now stay on MiniMax's omitted/adaptive thinking path, explicit `/think off` is still respected, and MiniMax-M2.x keeps the disabled-thinking default that prevents reasoning_content leaks.
Also wires the MiniMax thinking policy through bundled provider-policy loading so pre-runtime and configless embedded-agent paths resolve the same defaults.
Thanks @IamVNIE for the live MiniMax API repro and initial patch.
Summary:
- The branch moves Anthropic `start` emission into `message_start` handling for the provider and transport stream paths and adds focused ordering/error tests.
- PR surface: Source +5, Tests +149. Total +154 across 4 files.
- Reproducibility: Do we have a high-confidence way to reproduce the issue? Yes from source: current main emit ... ecovery intentionally refuses to retry after any non-error output; no live expired-cache run was performed.
Automerge notes:
- PR branch already contained follow-up commit before automerge: fix(agents): defer Anthropic transport stream start event until after…
Validation:
- ClawSweeper review passed for head 399a243c64.
- Required merge gates passed before the squash merge.
Prepared head SHA: 399a243c64
Review: https://github.com/openclaw/openclaw/pull/90697#issuecomment-4632866448
Co-authored-by: openperf <16864032@qq.com>
Co-authored-by: clawsweeper[bot] <274271284+clawsweeper[bot]@users.noreply.github.com>
Approved-by: takhoffman
Co-authored-by: takhoffman <781889+takhoffman@users.noreply.github.com>
Split Anthropic system prompts at the cache boundary so only stable prefixes get cache_control, strip the internal marker when cache control is disabled, and keep OpenAI-compatible Anthropic cache-control routes from caching dynamic suffixes.\n\nFixes #89386.
Enable Mistral prompt cache keys without long-retention forwarding. Update cached-read pricing and doctor migration for existing Mistral provider config. Fixes#83709.
Repairs a batch of narrow model/provider edge cases:
- honor OpenAI and Anthropic base URL environment overrides when provider config does not set an explicit base URL
- preserve OpenRouter Anthropic cache retention while stripping unsupported transport options
- allow apply_patch for non-OpenAI providers when the tool config otherwise permits it
- prune stale same-provider model selections from configure/model picker state
- expose GitHub Copilot bundled thinking policy metadata to offline/provider-policy lookups
- repair additive SQLite shared-state upgrades for existing databases
- keep same-size rotated log readers from reusing stale content in CI tooling
Proof:
- GitHub PR checks green on exact head 46514909b0
- Crabbox delegated Blacksmith Testbox tbx_01kt3em5r9vd7g0bnykrff6jdk exited 0
- Focused local Vitest/oxlint/format proof recorded in PR body and land-ready comment
Fixes#80347.
Fixes#88357.
Fixes#45269.
Supersedes #74427, #74432, #79370, #79894, #80366, and #88359.
Fix live model inference edge cases across provider streaming, model switching, outbound delivery, and gateway tool resolution.
Includes live/provider issue fixes and leaves #89100 explicitly partial for the remaining FM-2 group routing case.
Extract shared normalization/coercion helpers into private @openclaw/normalization-core workspace package while preserving existing plugin SDK helper subpaths.\n\nAlso keeps direct normalization-core imports internal, wires UI/build/loader resolution, and replaces the slow PR network CodeQL lane with a fast added-line boundary scan while retaining full CodeQL for scheduled/manual runs.\n\nVerification: local moved tests, plugin SDK boundary tests, extension loader tests, agents-support shard, UI build/test, build artifacts, lint, workflow guards, autoreview, and GitHub CI passed on PR head 963d893715.
Treat OpenAI-compatible streaming tool deltas as executable only when the final finish reason is `tool_calls`. This prevents malformed provider streams from triggering spurious tool execution while preserving normal tool-call responses.
Fixes#85161.
Verification:
- Local OpenAI-compatible SSE replay: spurious stop stream `finalToolCalls: 0`; valid tool-call stream `finalToolCalls: 1`.
- `pnpm test src/agents/openai-transport-stream.test.ts src/llm/providers/openai-completions.test.ts -- --reporter=verbose`
- PR CI green on `cdc2fc34753492c862cae99b37f8cf3761d9bbed`.
Co-authored-by: 忻役 <xinyi@mininglamp.com>
Co-authored-by: Jerry-Xin <jerryxin0@gmail.com>
* fix(responses): drop orphaned assistant msg_* id when reasoning is dropped (#88019)
When an Azure/OpenAI Responses session falls back to a non-Responses model
and later resumes a Responses model, sanitizeSessionHistory drops the
replayable reasoning (rs_*) item via downgradeOpenAIReasoningBlocks. The
paired assistant text block still carried its textSignature (the msg_* id),
so the transport replayed an assistant message item referencing msg_* with
no accompanying rs_* reasoning item. Azure Responses then rejected the next
turn with:
400 Item 'msg_...' provided without its required 'reasoning' item: 'rs_...'
permanently poisoning the session.
Fix:
- downgradeOpenAIReasoningBlocks now strips the textSignature from a turn's
text blocks whenever it drops a replayable reasoning item, so the msg_* id
and its rs_* reasoning are removed together. The transport then falls back
to a synthetic, unpaired id that Azure accepts.
- Because the synthetic fallback id is derived from the per-message msgIndex,
multiple id-less text blocks in one assistant turn (e.g. commentary +
final_answer) would collide on the same id. Make the fallback unique per
text block in both Responses conversion sites
(openai-transport-stream.ts and the shared llm provider
openai-responses-shared.ts).
Tests:
- sanitize-session-history: model-switch path drops the paired msg_* id.
- embedded-agent-helpers: downgrade strips paired text signature(s).
- reasoning-replay: multiple id-less text blocks get distinct item ids.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(responses): preserve phase metadata and guard malformed blocks (#88019)
Address PR review feedback on the orphaned msg_* replay fix:
- Preserve Responses phase metadata: dropping the paired msg_* id when its
rs_* reasoning is removed previously stripped the entire textSignature,
which also discarded the phase (commentary/final_answer). Phased text now
keeps a phase-only signature ({v:1,phase}) so commentary is not replayed
as user-visible output. Both parseTextSignature copies (shared provider and
embedded transport) now accept id-less phase-only signatures and fall back
to a synthetic id while preserving the phase.
- Guard malformed content blocks: the post-drop map no longer dereferences
contentBlock.type unconditionally, so a corrupted transcript with a
null/primitive block can still sanitize through a model switch.
Tests:
- sanitize-session-history: phase metadata is preserved while the paired id
is dropped on a model switch.
- reasoning-replay: id-less phase-only signatures get distinct synthetic ids
and retain their phase.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Repair invalid \u escapes during streaming JSON parsing without changing valid Unicode escapes. Split oversized node CI doctor/infra shards and fix the restart test mock deadlock so PR CI stays under the no-output threshold.\n\nCo-authored-by: Coder <83845889+coder999999999@users.noreply.github.com>
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
Refactor OpenAI provider identity so OpenAI remains the canonical provider for API-key and OAuth-backed flows while legacy openai-codex state is doctor/migration-only.
Keeps OpenAI Codex Responses as an API/transport class rather than a provider identity, moves auth aliases through providerAuthAliases, updates doctor repair sequencing for old auth/profile state, and refreshes tests/docs around the canonical OpenAI behavior.
