* fix(sse): surface OpenRouter mid-stream error chunks instead of a false empty success
OpenRouter (and similar OpenAI-compatible aggregators) can send an HTTP 200
SSE stream whose body carries a chat.completion.chunk with empty choices and
a top-level error object instead of any delta -- e.g. the underlying
provider hitting its own capacity limit mid-request. The Responses-API
response translator's `!chunk.choices?.length` branch treated this exactly
like a legitimate trailing-usage/no-op chunk, so the stream silently ended
with response.completed / error: null / output: [] -- a false "successful
but empty" response masking a real 502 provider_unavailable failure.
Live repro: request 1784726796287-a45bb3 (OpenClaw via the default combo,
nvidia/nemotron-3-ultra-550b-a55b:free via OpenRouter) got HTTP 200 with 0
tokens in/out after Nvidia returned "Worker local total request limit
reached (33/32)" mid-stream; the client saw an empty completed response
with no indication anything failed.
Mirrors the Gemini mid-stream error fix (#4177): set state.upstreamError
from the in-band error chunk so stream.ts's existing upstreamError handling
takes over (sendCompleted emits status: "failed" with a real error object).
Co-Authored-By: Markus Hartung <markus.hartream@gmail.com>
* chore(quality): file-size baseline for own-growth (#8210)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Markus Hartung <markus.hartream@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(sse): collapse single-text-part Responses-API content to a plain string
Every /v1/responses request — even the simplest single-string input —
got 500'd by AI Horde's Aphrodite-backed facade. Root cause:
normalizeResponsesInputForChat() always wraps a plain string input as
`content: [{ type: "input_text", text: value }]` (a one-element array),
and openaiResponsesToOpenAIRequest() mapped that straight through to
`content: [{ type: "text", text: value }]` on the Chat Completions side
— an array. That's spec-valid (OpenAI's own API accepts both shapes),
but strict/naive OpenAI-compatible backends like AI Horde's only
implement the plain-string form and reject the array form outright.
A single-text-part array and a plain string are semantically
identical, so collapse is safe. Real multi-part messages (text+image,
text+file) are left untouched.
Regression test: tests/unit/openai-responses-single-text-content-string.test.ts
(RED before the fix — every collapsed-content assertion failed with an
object instead of a string; GREEN after).
Also adds a deeper AI Horde load-test suite (sequential/concurrent/
cross-model/sustained-throughput/new-capable-model-candidates) that
surfaced this bug via real live traffic after Behemoth-X-123B was
temporarily added to the "default" combo for evaluation.
Co-Authored-By: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): unsupportedParams provider-level fallback for aihorde's live-discovered models
Real OpenClaw traffic against the newly-added Behemoth-X-123B combo
target kept 500ing on every attempt even after the Responses-API
content-array fix landed. The pipeline artifact showed why: `tools`
was still present, unstripped, in the request actually sent to AI
Horde's Aphrodite backend.
Root cause: `unsupportedParams: ["tools", "tool_choice",
"parallel_tool_calls"]` was only declared on the 3 models statically
listed in the aihorde registry entry (Cydonia-24B, Skyfall-31B,
google/gemma-4-31b). AI Horde uses `passthroughModels: true` — its
live worker roster changes constantly — so Behemoth-X-123B, like every
other dynamically-discovered aihorde model, had no model-specific
unsupportedParams entry, and getUnsupportedParams() returned [] for
it. But "the workers run raw text-completion backends" (no tool
calling) is true of every model AI Horde serves, not just the 3
catalogued ones.
Adds a provider-level `unsupportedParams` fallback on RegistryEntry,
checked by getUnsupportedParams() after the per-model lookup misses.
Set on the aihorde entry so it covers its entire live-discovered
roster, present and future, without needing a static per-model catalog
entry for each one.
Regression test: tests/unit/aihorde-tools-unsupported-provider-fallback.test.ts
(RED before the fix — Behemoth-X and deepseek-v4-flash both returned
[] instead of the stripped param list; GREEN after, with a control
case confirming the fallback doesn't leak to unrelated providers).
Co-Authored-By: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): flatten leftover tool-call history when stripping unsupported tools
Third bug in the same AI Horde/Behemoth-X saga: even after tools/
tool_choice were correctly stripped from the live request (previous
fix), real combo traffic still 500'd. The conversation history itself
carried a prior turn's role:"assistant" tool_calls and role:"tool"
result messages, left over from before the combo failed over from a
tool-capable model (Gemini) to a non-tool-capable one (AI Horde). Its
raw completion backend doesn't understand those message shapes at all,
independent of whether live `tools` is present — confirmed by
reproducing with a role:"tool" message and NO tools param at all.
flattenToolHistory() (open-sse/utils/flattenToolHistory.ts) already
existed for exactly this, fully unit-tested — it just had zero call
sites anywhere in the request pipeline. Extracts the unsupported-params
strip into a small testable module
(open-sse/handlers/chatCore/unsupportedParamsStrip.ts, following the
existing chatCore god-file decomposition pattern e.g.
executorClientHeaders.ts) that now also flattens tool-call history
whenever "tools" was among the stripped params.
Regression test: tests/unit/chatcore-unsupported-params-strip.test.ts
(RED before the fix — the flattening test failed with the raw
tool_calls array still present; GREEN after). All 434 existing
chatcore-*.test.ts tests still pass.
Co-Authored-By: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): gate tool-history flattening on unsupported, not on stripped-this-request
The previous commit's flattening only fired when "tools" was actually
present-and-stripped on THIS request. A second live reproduction
against AI Horde had no live `tools` param at all — only stale
tool_calls/tool-result messages inherited from before a combo
failover — and still 500'd, because that condition never triggered.
A model that can't do tool calling can't do it whether or not the
current request happens to carry a `tools` array. Gate on the
unsupported-params list itself (unsupported.includes("tools")) instead
of the subset that was actually present-and-deleted this time.
Regression test added to the same file (RED before — the no-live-tools
case left tool_calls/role:"tool" untouched; GREEN after). All 435
chatcore-*.test.ts still pass.
Co-Authored-By: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): skip tool-incapable combo targets, error clearly on direct requests
Two complementary fixes for a model that structurally can't do tool
calling at all (e.g. AI Horde's raw completion backends) rather than
silently degrading — following up on the earlier strip/flatten fix,
which stopped the crashes but let a tool-incapable target still get
selected and return a 200 that narrates a fake tool call in prose
instead of erroring or being skipped.
1. Root cause, combo routing: getResolvedModelCapabilities()'s
`supportsTools` resolution only checked per-model registry entries,
synced capabilities, and static specs — none of which exist for a
dynamically-discovered model (AI Horde's passthroughModels roster
changes as workers come and go). It fell through to
heuristicToolCalling(), which optimistically defaults to `true` for
any unrecognized model (TOOL_CALLING_UNSUPPORTED_PATTERNS is empty).
Added a provider-level fallback reusing the same unsupportedParams
signal the request-time strip already relies on. This makes the
EXISTING filterTargetsByRequestCompatibility (comboStructure.ts) —
which already correctly excludes non-tool-capable targets when a
request requires tools — actually work for these models; no combo.ts
changes were needed, it was only ever fed bad capability data.
2. Direct/pinned requests: filterTargetsByRequestCompatibility only
protects combo routing. A direct request naming an exact
tool-incapable model has no other target to fail over to — added
checkToolCallingRequiredButUnsupported (chatCore/toolCallingRequiredCheck.ts),
gated on isCombo: false, returning a clear 400 instead of a 200 that
silently can't do what was asked.
Regression tests (both RED before, GREEN after):
- tests/unit/model-capabilities-provider-unsupported-tools.test.ts
- tests/unit/chatcore-tool-calling-required-check.test.ts
All 463 chatcore-*/model-capabilities-*.test.ts and 31 combo
compatibility-filter tests still pass.
Co-Authored-By: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): correct handleChatCore return shape for the tool-calling-blocked error
handleChatCore's documented contract is `{ success, response, status,
error }`, not a raw Response — returning `new Response(...)` directly
(copied from a different early-return whose surrounding context turned
out not to share this function's top-level contract) produced "No
response is returned from route handler ... Expected a Response object
but received 'undefined'" and a bare 500 with an empty body, caught
immediately when verifying the previous commit live.
Uses createErrorResult() (already used by the adjacent
translation-failure branch a few lines up) instead of hand-building the
Response, matching the same pattern already established in this
function for early error returns.
Co-Authored-By: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): scope Responses single-text-content collapse to providers that need it
The single-text-part content array -> plain string collapse (added for AI
Horde's Aphrodite facade, which 500s on the array form) was applied
unconditionally to every provider, silently breaking the standard OpenAI
array-shaped content contract that other providers and existing tests
depend on. Added RegistryEntry.requiresPlainStringContent, gated the
collapse on it (true only for aihorde), and threaded modelInfo.provider
through responsesHandler -> responsesApiHelper -> the translator so the
real /v1/responses call site can identify the provider.
Co-Authored-By: Markus Hartung <markus.hartream@gmail.com>
---------
Co-authored-by: Markus Hartung <markus.hartung@gmail.com>
Co-authored-by: Markus Hartung <markus.hartream@gmail.com>
* fix(sse): synthesize tool_calls for Gemini's malformed function-call abort reasons
Live incident (dashboard log id 1784489701456-d8c0e9): Gemini terminates a
stream with finishReason MALFORMED_FUNCTION_CALL/UNEXPECTED_TOOL_CALL when its
own parser rejects an attempted tool call — there's no real functionCall part,
only a human-readable finishMessage. gemini-to-openai.ts passed this through
raw as finish_reason (9router#2462's fix, correctly keeping it off a clean
"stop"/Claude end_turn), but a raw "malformed_function_call" isn't one of
OpenAI's 5 documented finish_reason values, so a real OpenAI-format client
(OpenClaw) has no handling for it at all and silently never notices the turn
failed — confirmed live via tests/integration/live-gemini-workload.test.ts's
[28] streaming case after the Gemini TPM/rebase work on this branch.
Fix: synthesize a tool_calls entry (arguments carry the error code + Gemini's
finishMessage, valid JSON) and finish_reason: "tool_calls" instead, routing
the failure into the ordinary "tool call arguments didn't parse" path every
OpenAI-compatible agent loop already handles. Defers to a real tool call if
one already completed earlier in the same turn — the real call wins, no
synthetic entry piles on top of it.
Tests (TDD, each confirmed red-before-green):
- 5 new unit tests in the existing 9router#2462 regression file, covering the
synthesis itself, UNEXPECTED_TOOL_CALL, the real-tool-call-wins edge case,
and no-regression on a clean STOP.
- New fixture (tests/fixtures/translation/gemini-malformed-function-call-stream.json):
the real 6-chunk event series from the live incident, sanitized (personal
paths/URLs replaced with generic placeholders, structure preserved exactly).
- New integration test chains the real translator into the real Responses API
transformer using that same fixture, proving correct behavior on BOTH
/v1/chat/completions and /v1/responses from one shared ground-truth event
series.
Co-authored-by: Markus Hartung <markus.hartung@gmail.com>
* fix(sse): don't drop a malformed tool-call failure when it lands beside a real one
Live incident (dashboard log id 1784589106014-2a42f8), analyzing why the
prior malformed-function-call fix (3568c7259) still wasn't reaching the
client in this case: Gemini can emit a REAL, valid functionCall AND finish
the SAME candidate with MALFORMED_FUNCTION_CALL — the model attempted
multiple tool calls in one turn (here: a real status-check call plus a
malformed "exec"+"cron" multi-call attempt), one parsed cleanly and the
other didn't.
The first fix version skipped synthesizing a failure signal whenever a real
tool call already existed (state.toolCalls.size > 0), on the assumption
that meant the model was retrying a LATER, separate attempt after an
earlier one already succeeded. That's indistinguishable, from the
translator's state, from this same-turn case — so it silently discarded
the malformed attempt's information entirely: the client saw the real call
succeed and never learned the other tool calls were attempted and rejected.
Fix: always synthesize the failure entry when a malformed abort reason is
seen, appending it alongside any real tool call rather than skipping it.
Multiple tool_calls in one response is normal, well-supported OpenAI
behavior (parallel tool calls), so this adds the failure as an additional
entry instead of replacing or hiding the real one.
Tests (TDD, confirmed red-before-green):
- Rewrote the unit test that encoded the old (wrong) assumption to assert
both the real and synthesized calls are present.
- New fixture (gemini-malformed-function-call-parallel-real-call-stream.json):
the real event series from this incident, sanitized.
- New integration tests (same file as 3568c7259's) prove both
/v1/chat/completions and /v1/responses surface both tool calls correctly
from this fixture.
Co-authored-by: Markus Hartung <markus.hartung@gmail.com>
* feat(sse): honor tool_choice when translating OpenAI requests to Gemini
Investigating a live report that gemini-3.1-flash-lite frequently narrates
an intended tool call in plain text instead of actually emitting one
(dashboard log id 1784591483850-49c408 — 9 raw provider chunks, all plain
text, zero functionCall parts, clean finishReason STOP): body.tool_choice
was never read anywhere in the OpenAI->Gemini request translator.
result.toolConfig was unconditionally hardcoded to
{ functionCallingConfig: { mode: "VALIDATED" } } whenever tools were
present, regardless of what the caller sent. VALIDATED lets the model
respond with plain text OR a schema-validated function call at its own
discretion — it never forces a call the way OpenAI's tool_choice:
"required" (Gemini's ANY mode) does, so a caller had no way to compel a
tool call even when explicitly requesting one.
Added convertOpenAIToolChoiceToGemini(), mirroring the existing
convertOpenAIToolChoice() in openai-to-claude.ts for the same OpenAI
tool_choice shapes (string "auto"/"none"/"required", or
{type:"function",function:{name}} to force one specific tool):
- unset/"auto" -> VALIDATED (unchanged default, no regression)
- "required"/"any" -> ANY (forces a call)
- "none" -> NONE (disables function calling)
- {type:"function",...} -> ANY + allowedFunctionNames: [name]
Wired into both Gemini request paths: the direct/base translator
(openaiToGeminiBase) and the Antigravity/Cloud Code envelope
(wrapInCloudCodeEnvelope), which now reuses the base translator's already-
computed toolConfig instead of re-deriving its own hardcoded VALIDATED.
This unblocks (but does not itself resolve) the live question — a
tool_choice: "required" A/B test against gemini-3.1-flash-lite follows to
confirm ANY mode actually changes the narrate-vs-act behavior in practice.
Also updates the T11 any-budget allowlist for this file: the "any" string
comparisons (tool_choice value "any", not a TypeScript type) are the same
documented false-positive pattern already carved out for executors/base.ts.
Tests (TDD, confirmed red-before-green): 9 new unit tests covering all
tool_choice shapes on both the direct and Antigravity/Cloud Code paths,
plus the no-tools and unset-default no-regression cases.
Co-authored-by: Markus Hartung <markus.hartung@gmail.com>
* chore(quality): file-size baseline for own-growth (#8211)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Markus Hartung <markus.hartung@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
RubyLLM (and other OpenAI-convention clients) embed strict:true/false directly
inside a function tools parameters JSON schema. Gemini/Antigravity rejects the
unrecognized keyword with a 400 ("Unknown name strict ... Cannot find field"),
the same failure class as the existing multipleOf entry. Broke every Chatwit
Captain tool-calling call routed through witdev_antigravity/gemini-*.
Reconstructed onto current release/v3.8.49 tip (preserves #8231 CIVIC_INTEGRITY
exclusion; the author's stale base showed it as a spurious revert).
Co-authored-by: Witroch4 <wital@example.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix: strip internal reasoning placeholder from user-visible content (#8081)
The internal reasoning replay sentinel '(prior reasoning summary unavailable)'
can leak into user-visible assistant content when a model echoes it through
ordinary message.content / delta.content. Existing suppression only checked
reasoning_content fields and reasoning-specific events.
Changes:
- Add stripInternalReasoningPlaceholder() to reasoningPlaceholder.ts —
removes all occurrences of the sentinel and trims; returns '' when
nothing meaningful remains
- Streaming: strip in responsesTransformer.ts, openai-responses.ts, and
openai-to-claude.ts at the delta.content entry point; skip emission
entirely when only the placeholder was present
- Non-streaming: strip in responseSanitizer.ts sanitizeMessageContent()
and sanitizeResponsesMessageContent() (all three text paths)
translateText is unaffected (uses mode='translate' via plain newsClient).
The per-provider reasoning_content check remains as defense-in-depth.
* fix: skip only empty content block on reasoning-placeholder, keep finish_reason/tool_calls (#8081)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* chore(quality): rebaseline openai-responses.ts own-growth (#8081 guard)
---------
Co-authored-by: Austin Liu <austinliu@Austins-MacBook-Air-3.local>
Co-authored-by: Probe Test <probe@example.com>
Co-authored-by: Dingding-leo <Dingding-leo@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(responses): close namespace round-trip for Responses-Chat translation (#7936)
The #7905 custom-tool-call path landed in release/v3.8.49 but left #7936
open: Responses namespace sub-tools were flattened to a bare leaf on the
Chat wire with no response-side closure, so Codex's adjudicator rejected
every namespace sub-tool call with `unsupported call` -- it only has a
dispatch entry for the header bits (namespace+name), no entry for the
bare leaf.
This patch closes the round-trip without mutating the Chat wire name
(alignment with #7905's bare-leaf contract and with the issue author's
proposed fix):
* request side (openai-responses.ts): keep tool.function.name as the bare
leaf, populate a side-band namespaceToolIdentityMap keyed on that leaf,
and thread it through translatedBody._toolNameMap.
* request -> response seam (chatCore.ts): extract the identity map before
dispatch and pass it through to the non-stream completion path and to
all three stream pipelines (translate openai-responses, translate
other, passthrough).
* response translator (response/openai-responses.ts): in emitToolCall
(response.output_item.added) and closeToolCall (custom_tool_call /
function_call output_item.done), call resolveRequestToolIdentity() to
rewrite the bare leaf back to {namespace,name} and emit codex-compatible
independent fields.
* passthrough (utils/stream.ts): add a response passthrough rewriter
restoreResponsesPassthroughFunctionCallIdentity that intercepts
response.output_item.added, response.output_item.done, and
response.completed and stamps the same {namespace,name} tuple.
* helper (requestToolIdentity.ts): a 20-line stateless resolver; never
parses a name.
The wire-visible Chat tool.function.name stays the bare leaf -- non-OpenAI
providers (NVIDIA, GLM, Kimi, Gemini, ...) frequently truncate or rewrite
long __-dotted names; bare leaves avoid that failure mode entirely. The
codex ResponseItem::FunctionCall schema (models.rs) declares an
independent namespace: Option<String> field and has a
function_call_deserializes_optional_namespace round-trip test, so
emitting it separately matches the codex adjudicator dispatch.
Includes 19 new test cases across 4 files:
- request-side bare-leaf wire + side-band ledger construction
- response-side tuple emit + unmapped passthrough + apply_patch exclusion
- ambiguous-leaf collision safety (entry dropped, leaf emits verbatim)
- per-request isolation between concurrent streams
- a precompiled Atlassian-style nested namespace override
* fix(responses): skip tool_search_call input items instead of 400 (#7936 addendum)
Codex 0.42+ emits `tool_search_call` (and later `tool_search_result`) input
items when the model uses the dynamic tool-search optimization. They are
metadata-only: they record that the model queried a subset of the
available tools, and carry nothing that OpenAI Chat Completions can
represent.
Without an explicit skip in openai-responses.ts, the input loop threw
Unsupported Responses API feature: input item type 'tool_search_call'
cannot be represented in Chat Completions -- and because these items
stay in the Responses API `input` for every follow-up turn, the whole
server returned 400 on EVERY subsequent /v1/responses in the same
session until the user cleared history.
Observed in the wild:
/v1/responses 400 "Unsupported Responses API feature: input item
type 'tool_search_call' cannot be represented in Chat Completions
[longcat/LongCat-2.0 (400), longcat/LongCat-2.0 (400)]"
The meituan combo (longcat fallback) was the most visible victim, but
the underlying throw is source-format-side and hits any Responses-API
consumer whose upstream does not natively support Responses.
Fix: stop on the item type the same way `reasoning` is skipped --
display-only metadata, no chat side-effect. Covers both
`tool_search_call` and the follow-up `tool_search_result` shapes.
Adds 3 unit tests:
- tool_search_call is silently skipped (no 400)
- tool_search_result is silently skipped
- tool_search_call items interspersed with real messages are skipped
in order; real messages survive
* chore(quality): rebaseline openai-responses.ts + stream.ts own-growth (#7936 namespace round-trip)
---------
Co-authored-by: TonPro <hello@tonpro.fu>
Co-authored-by: RCrushMe <RCrushMe@users.noreply.github.com>
* fix(translators): normalize TitleCase tool names for non-Anthropic models
* fix(translator): parse XML <invoke> blocks in OpenAI→Claude response translator
* fix(gemini-web): mark gemini-3.1-flash-lite as toolCalling: false
* fix(gemini-web): mark all models as toolCalling: false, add to no-tools combo
* fix(web-providers): mark all web-cookie models as toolCalling: false
* fix(nvidia): disable tool calling on models that can't handle it
* fix(lint): drop explicit any annotation on extractXmlInvokeBlocks state param
Removes the no-explicit-any violation introduced alongside the TitleCase
tool-name normalization work so the merged branch stays lint-clean
(state stays implicitly typed like its sibling helpers in this file).
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix: hide internal reasoning replay placeholder
* fix: hide internal reasoning replay placeholder in OpenAI→Claude translation too
Mirror the isInternalReasoningPlaceholder() guard already applied to the
Responses-API and OpenAI-Responses reasoning paths in
openaiToClaudeResponse() (open-sse/translator/response/openai-to-claude.ts).
The internal reasoning-replay sentinel was still leaking into the Claude
"thinking" content block on this translation path.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* test: close both-add merge of #7912 and #7905 reasoning/tool tests
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(sse): convert OpenAI media parts to Claude blocks in the CC bridge
OpenAI-format clients (OpenCode/Kilo/Cline) reach the Claude-Code-compatible
bridge untranslated: chatCore skips the OpenAI->Claude translator when
sourceFormat is OPENAI, so image_url / AI-SDK image / file parts either went
upstream in OpenAI shape (silently ignored) or were dropped by the text-only
extraction, and media-only user turns were removed by hasValidContent().
- claudeCodeCompatible: convertOpenAiMediaBlock() converts image_url
(base64 + remote), AI-SDK string image and file parts (pdf->document,
image mime->image) to Claude blocks in both bridge paths; Claude-native
blocks pass through unchanged and the text-only wire image is preserved.
- claudeHelper: hasValidContent() now counts image/document blocks so
media-only user turns are not silently deleted.
Reported-by: beingshafin
Refs #7777
* refactor(sse): extract CC media-block conversion to ccOpenAiMediaBlocks.ts
claudeCodeCompatible.ts is frozen at 1202 lines by check:file-size; the #7777
helpers pushed it to 1291. Move them to a dedicated module, no behavior change.
* test(quality): register cc-bridge-openai-image-7777 test in stryker tap.testFiles
AntigravityExecutor.execute() overrides BaseExecutor.execute() and never calls
super.execute(), so the shared orphan-tool_use guard (fixToolPairs, #2382/#4714)
never ran on the Antigravity/Vertex Claude dispatch path. A client history with a
genuinely orphaned tool_use (no matching tool_result anywhere — e.g. left behind by
OpenCode's known abort/cancel bug) sailed straight through openaiToAntigravityRequest
into Google's Cloud Code envelope as an unpaired functionCall, which Vertex's Claude
backend rejects with HTTP 400. Mirror-image gap of #6026 (incoming direction).
Fix: run fixToolPairs on body.messages in openaiToGeminiBase before building the
tool_call_id map and the functionCall/functionResponse translation loop.
Codex CLI injects prompt_cache_key natively for its own prompt caching.
injectPromptCacheKey() only guards against the router injecting a NEW
key for nvidia/codex/xai — it never strips a key that arrived already
present in the inbound body. NVIDIA NIM's OpenAI-compatible wrapper
rejects the field with a 400, and NIM has no documented support for
prompt caching (providerSupportsCaching already treats nvidia as
non-cache-capable).
Adds a provider-wide STRIP_RULES entry in paramSupport.ts (match-all,
since prompt_cache_key rejection isn't model-specific) so
stripUnsupportedParams() drops it for every nvidia target before the
request reaches DefaultExecutor.
* feat(routing): add reasoning-based model and effort routing
* refactor(routing): modularize reasoning and auto-routing pipeline
* fix(routing): remove redundant DB re-export and prevent SQL scan false positives
* fix(routing): resolve reasoning routing review blockers
* fix(i18n): keep release ranking fallbacks outside reasoning
* fix(db): renumber reasoning-routing migration past release tip (124→125)
124_generic_session_affinity_ttl.sql (#7274) has since landed on
release/v3.8.49 at version 124, colliding with this PR's own
124_reasoning_routing_rules.sql. Renumbers to 125 (the next free slot
past the current release tip) and updates the one filename reference
in docs/routing/REASONING_ROUTING.md.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* chore(db): renumber reasoning-routing migration 125→126 (slot taken by #7360)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* refactor(api): compact temp-path decls in exportAll GET (complexity-ratchet lines budget)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* refactor(api): single-statement auth guard in exportAll GET (function under 80-line cap)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(translator): synthesize tool call chunks from response.completed output[]
When an upstream provider sends a batched response.completed event carrying
function_call items in its data.response.output[] array — without having
sent the individual response.output_item.added / .delta / .done events —
the state variables toolCallIndex and currentToolCallId were never set,
causing computeFinishReason to return 'stop' instead of 'tool_calls'.
This broke the agent loop for downstream Chat Completions clients
(OpenCode, Hermes, etc.) when routing through providers that batch their
output into the completed event.
Fix: parse data.response.output[] for function_call items in the
response.completed handler, synthesize the tool call header + arguments
delta chunks, advance state, and emit finish_reason: 'tool_calls'.
Also updates withAssistantRoleOnFirstDelta to handle array results.
Fixes#180, #3980
Refs: https://github.com/diegosouzapw/OmniRoute/issues/180
Refs: https://github.com/diegosouzapw/OmniRoute/issues/3980
* fix(translator): guard against double-emission for incrementally-streamed tool calls
Add a guard that skips response.completed synthesis for call_ids already
tracked via incremental output_item.added/.done events. Without this,
providers that stream incrementally AND echo function_call items in the
response.completed output[] snapshot get duplicate tool call chunks.
Also adds a regression test combining both incremental events and a
response.completed snapshot in the same turn.
Refs: diegosouzapw/OmniRoute#7613
* chore: add docker-compose.yml.bak to gitignore
* refactor(translator): extract response.completed synthesis, fix ratchets
Fixes the file-size and complexity/cognitive-complexity ratchet
regressions the dedup-guard commit (6bbff5ea) introduced, so the PR's
own validation block (typecheck, eslint, file-size, complexity,
cognitive-complexity, changelog-integrity, test-discovery) is fully
green, not just its own tests:
- Extract the response.completed batched-tool-call synthesis body into
a new leaf module
open-sse/translator/response/openai-responses/synthesizeCompletedToolCalls.ts,
mirroring this file's own established eventEmitter.ts/toolSchemas.ts/
pureHelpers.ts extraction pattern, further split internally
(buildToolCallChunks/buildFinalChunk/resolveArgsStr/baseChunk) to
keep synthesizeCompletedToolCalls() itself under the complexity/
cognitive-complexity/max-lines-per-function thresholds.
- computeFinishReason moves alongside it (not into the sibling
pureHelpers.ts) because it takes stream `state` — pureHelpers.ts is
guarded by tests/unit/response-openai-responses-purehelpers-split.test.ts
to have NO state coupling at all.
- DRY the withAssistantRoleOnFirstDelta array/single-result branches
into a shared setAssistantRoleIfEligible(state, delta) helper — the
array branch alone pushed this function's cyclomatic complexity to
16 (over the 15 threshold).
- Split the 5 new response.completed tests out of
tests/unit/translator-resp-openai-responses.test.ts (which would
have exceeded the 800-line test-file-size cap) into a new sibling
tests/unit/translator-resp-openai-responses-completed-synthesis.test.ts.
- Bump the frozen open-sse/translator/response/openai-responses.ts
file-size baseline for the small irreducible remainder (dedup-guard
state tracking + call-site wiring) with a justification entry in
config/quality/file-size-baseline.json.
Independently re-verified red-first (temporarily reintroducing the
pre-guard filter reproduces exactly one failure — the dedup test — no
collateral damage) and confirmed check:complexity-ratchets is back to
exactly baseline (2058/890) and check:file-size is fully green.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* refactor(translator): move completed-tool-call glue into module (file-size cap)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Erick Kinnee <erick@ekinnee.dev>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* feat(kimi): sync Code, Web, and Moonshot providers
* chore(quality): trim frozen file-size overflow in Kimi sync
The Kimi/Moonshot provider sync added a net +1 line to both
src/sse/services/auth.ts and ProviderDetailPageClient.tsx, pushing
each 1 line past its frozen cap in file-size-baseline.json. Drop one
optional blank line in each (prettier-neutral, no behavior change) to
land back at/under the frozen baseline.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(codex): preserve GPT-5.6 reasoning contract
* fix(vscode): expose Responses text models
* fix(codex): keep GPT-5.6 limits through discovery
* fix(ci): extract isUsableChatModel helpers to satisfy complexity ratchet
Splitting the supported_endpoints/output_modalities guard clauses into
excludesChatAndResponsesEndpoints() / excludesTextOutputModality() drops
isUsableChatModel's cyclomatic complexity from 16 to under the ratchet's
max of 15 (complexity-ratchets gate: 2057 -> 2056, back at baseline).
Behavior is unchanged; existing vscode/codex route tests cover it.
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(codex): merge capacity limits conservatively (smaller of live vs pinned wins)
Resolve the #7012 catalog-merge policy collision: instead of the pinned
GPT-5.6 contract always winning for a fixed set of model ids, capacity
limits (inputTokenLimit/outputTokenLimit) now merge via
mergeCapacityLimitConservatively — Math.min(pinned, live) when both are
present, so OmniRoute never promises more context than the account can
actually serve. All other overlapping fields still take the live value
unconditionally.
Guard tests cover both directions (pinned smaller wins / pinned larger
loses) at the route level and via an isolated helper-level unit test.
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
---------
Co-authored-by: Xiangzhe <xz-dev@users.noreply.github.com>
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(responses): map mid-conversation system turns to developer role (#6954)
* test(responses): add #6954 mid-conversation system -> developer regression
* fix(6954): keep bare-string content parts in buildResponsesTextParts
* test(6954): cover array-form system content with bare string
Validated in merge-train 2026-07-18 @ 9084b408b: 12198/12201 pass; single red = earlyStreamKeepalive timer test, confirmed load-flake (6/6 green isolated on release tip AND the merged tree; no boarded PR touches keepalive)
* fix(sse): stop dropping tool_search and stop leaking OpenAI-only params in Responses->Chat translation (#7532, #7533)
#7532: `openai-responses.ts` unconditionally dropped `tool_search` when
downgrading a Responses-shaped request to Chat Completions, hiding the tool
from the model and breaking Codex's deferred/lazy tool-discovery protocol for
any provider that gets downgraded (e.g. built-in providers like opencode-go).
tool_search carries `execution: "client"` — the client resolves the call
locally regardless of wire shape — so it is now mapped to a normal Chat
function tool, mirroring the existing local_shell -> shell pattern in the
same file, instead of being silently discarded.
#7533: the same translator unconditionally copied two GPT-5/OpenAI-only
fields (`verbosity`, `prompt_cache_key`) into the translated Chat body
regardless of destination provider. A strict-protocol non-OpenAI upstream
(NVIDIA confirmed by the reporter) 400s on unrecognized top-level parameters.
Both fields are now gated on `credentials.provider === "openai"`, stripped
otherwise; the existing OpenAI-destined behavior (needed for #517's
prompt-caching fix) is preserved byte-identical via a dedicated sanity test.
Regression tests: tests/unit/tool-search-filtered-responses-to-chat-7532.test.ts,
tests/unit/verbosity-prompt-cache-key-provider-gate-7533.test.ts. Two existing
tests that encoded the old buggy contract (unconditional tool_search drop /
unconditional field leak with no credentials) were aligned to the corrected
contract: tests/unit/translator-openai-responses-req.test.ts,
tests/unit/openai-responses-verbosity.test.ts.
Gates run green: file-size, complexity, cognitive-complexity, typecheck:core,
lint (scoped to changed files), and the full touched-area unit test suite
(329 tests, 0 failures).
* fix(sse): keep prompt_cache_key/verbosity for the codex destination (#7533)
The #7533 provider gate allowlisted only "openai", but /v1/responses routes
EVERY request through this downgrade (handleResponsesCore ->
convertResponsesApiFormat) regardless of provider, and codex is an
OpenAI-operated upstream (chatgpt.com/backend-api/codex). Gating it out
stripped prompt_cache_key for Codex and silently re-broke the prompt-cache
affinity #517 exists to protect — with no test covering it.
Allowlist is now {openai, codex} and carries two #517 regression guards.
Non-OpenAI upstreams (NVIDIA) still get both fields stripped, per #7533.
Add a per-connection cache capability override (supportsPromptCaching,
cacheControlPassthrough) stored in provider_specific_data.cache, consulted
first by providerSupportsCaching() / providerHonorsOpenAIFormatCacheControl()
before falling back to the hardcoded CACHING_PROVIDERS name sets. Unblocks
prompt_cache_key injection, the compression cache-aware guard, and
cache_control passthrough for openai-compatible-chat-<uuid>-style custom
connections that can never match the hardcoded provider-name sets. Default
(no override) is byte-identical to current behavior.
Codex Responses API strict mode forces every "optional" tool property into
`required`, so a model that intends to OMIT an optional enum property (no
declared `default`, e.g. Agent.isolation: enum["worktree","remote"]) must
still emit a concrete value. Neither of the two ops shipped in #6992
(drop-if-default, generalized drop-if-empty) can catch this: drop-if-default
needs a declared default (none exists); drop-if-empty needs an empty
string/array (the emitted value is non-empty).
Adds a paired request/response transform scoped strictly to
targetFormat === OPENAI_RESPONSES:
- Request side: injectOptionalEnumOmissionSentinel/-ForTools widen no-default
optional enum properties to accept `null` (OpenAI's own documented
nullable-union idiom for this exact strict-mode limitation).
- Response side: isDroppableNullEntry drops the key when the model emits
`null` for a non-required, schema-declared property, reusing the existing
#6992 toolSchemas plumbing (no new tracking structure needed).
* feat(mitm): add Antigravity reasoning-effort overrides
The Antigravity MITM alias mapping only ever swapped the destination model;
there was no way to override the reasoning effort Antigravity's own
thinkingConfig requested. Alias entries are now `{ model?, reasoningEffort? }`
(a legacy plain-string mapping still normalizes to `{ model }`, so no DB
migration is required). The standalone proxy (server.cjs) forwards the chosen
tier as a top-level `reasoningEffortOverride` on the intercepted request; the
antigravity->openai translator honors it ahead of its thinkingConfig-derived
guess, and an explicit "none" suppresses reasoning_effort entirely even when
Antigravity's own request asked for thinking. Reuses the existing canonical
5-tier reasoning vocabulary (`@/shared/reasoning/effortStandardization.ts`,
with max/extra aliasing to xhigh) instead of introducing a new one. The API
route validates the reasoning-effort value at the boundary and the Antigravity
tool card UI now exposes a per-model reasoning-effort selector alongside the
existing model-mapping input.
Co-authored-by: Truong Fiu <gnourtf@gmail.com>
Inspired-by: https://github.com/decolua/9router/pull/2584
* chore(changelog): fragment for #7228
---------
Co-authored-by: Truong Fiu <gnourtf@gmail.com>
* fix(translator): register openai response projection for gemini clients
The response-translator registry had an OpenAI -> Antigravity response
projection registered, but no OpenAI -> Gemini one. When a client request is
detected as Gemini format (body-shape match on `contents: [...]`, per
detectFormat()) and combo routing lands the request on an OpenAI-native
provider, translateResponse() fell through its hub-and-spoke path with no
`openai -> gemini` translator registered, so the raw OpenAI
`chat.completion.chunk` shape reached the client unchanged instead of the
shared Gemini `response.candidates[]` envelope.
Registers FORMATS.OPENAI -> FORMATS.GEMINI reusing the existing
openaiToAntigravityResponse projection — Gemini and Antigravity already
share the same wrapped `{ response: { candidates: [...] } }` envelope
elsewhere in the pipeline (see the unwrapGeminiChunk callers in
open-sse/utils/stream.ts, which treat FORMATS.GEMINI and FORMATS.ANTIGRAVITY
identically), so no new conversion logic is introduced.
Co-authored-by: W ARELIK <warelik@users.noreply.github.com>
Inspired-by: https://github.com/decolua/9router/pull/2399
* chore(changelog): fragment for #7207
---------
Co-authored-by: W ARELIK <warelik@users.noreply.github.com>
* fix(translator): preserve Gemini thought parts as reasoning_content on the OpenAI request bridge
Gemini thinking-mode output marks internal reasoning with `part.thought === true`
inside a content's `parts` array. geminiToOpenAIRequest() ran every part (thought
or not) through the same text-part branch, so a thought part was merged straight
into the message's visible `content` — leaking private reasoning into whatever the
OpenAI pivot forwarded downstream, and hiding it from Reasoning Replay Cache (which
only ever inspects `reasoning_content`).
Add convertGeminiContentWithReasoning(): split out `thought: true` parts before
delegating to the existing convertGeminiContent(), then re-attach the joined
thought text as `reasoning_content` on the resulting message (skipping tool/
functionResponse messages, whose schema has no such field). Non-strict-provider
stripping and reasoning-replay injection in translator/index.ts are untouched —
this only fixes what reasoning_content gets populated with on this one inbound
hop.
Co-authored-by: W ARELIK <warelik@users.noreply.github.com>
Inspired-by: https://github.com/decolua/9router/pull/2401
* chore(changelog): fragment for #7206
---------
Co-authored-by: W ARELIK <warelik@users.noreply.github.com>
* fix(sse): sanitize empty-signature thinking blocks + hoist strict-provider system messages (#6953, #7293)
#6953: prepareClaudeRequest's "preserve latest-assistant thinking verbatim"
guard (claudeHelper.ts, anti-400 for legitimate Anthropic replay) did not
distinguish a genuine Claude signature from an empty one fabricated by a
non-Anthropic leg (e.g. codex reasoning_content). It forwarded signature:""
verbatim to Anthropic, which always 400s ("Invalid signature in thinking
block"), permanently locking combo routing onto the non-Anthropic leg. The
response-side half of this bug (openai-to-claude.ts synthesizing the empty
signature in the first place) was already fixed by #6982/PR#6982; this PR
closes the remaining request-side half. Fix: the verbatim-preserve guard now
requires every thinking-ish block on the latest assistant message to carry a
non-empty signature/data; otherwise it falls through to the existing
sanitization path (redacted_thinking + DEFAULT_THINKING_CLAUDE_SIGNATURE)
already applied to older turns.
#7293: translateRequest() is the single outbound choke point every chat
request passes through, including same-format (OpenAI→OpenAI) passthrough
where none of the format-specific translators run. systemMessageMustBeFirst()
/ PROVIDERS_SYSTEM_MUST_BE_FIRST (src/lib/memory/injection.ts, #6135/PR#6225)
was only consulted by the memory injector, so a client-injected system
message landing mid-array (OpenCode/Kilo Code style clients, Discussion
#6129) reached strict providers (xiaomi-mimo) untouched and 400'd. Fix: a new
helper (open-sse/translator/helpers/strictSystemHoist.ts) hoists every system
message onto index 0 for strict providers, reusing systemMessageMustBeFirst()
as the single source of truth, merging (never dropping) multiple offenders in
original order, and no-op'ing (same array reference) for non-strict providers
and already-compliant requests to preserve prompt-cache prefix stability.
Both defects live in the same file cluster (openai-to-claude request-path
translator + its helpers), hence one PR for both issues per triage guidance.
Regression tests:
- tests/unit/repro-6953.test.ts — RED (actual signature:'' forwarded) → GREEN
- tests/unit/probe-7293-strict-system-hoist.test.ts — RED (system message
left at index 10 of 70) → GREEN, plus multi-offender merge, existing-leading
merge, non-strict-provider no-op, and already-compliant no-op cases.
Gates run: file-size, complexity, cognitive-complexity (both at/under
baseline), typecheck:core (clean), eslint on changed files (clean),
test:vitest (254/254 green), plus all directly relevant existing suites
(translator-claude-helper-thinking, translator-xiaomi-mimo-reasoning-replay,
memory-system-first-6135, dashscope-cache-control-openai-2069,
xiaomi-mimo-provider, role-normalizer, translation.golden,
translators.property, translator-helper-branches, translator-claude-to-openai,
translator-same-format-null-flush — all green).
Closes#6953Closes#7293
* chore(quality): prune the now-stale claudeHelper no-explicit-any suppression (#6953)
The #6953 fix removed the single `any` that config/quality/eslint-suppressions.json
still had frozen for open-sse/translator/helpers/claudeHelper.ts, so the entry became
stale and ESLint's stale-suppression enforcement failed the 'No new ESLint warnings'
gate — the gate went red because the code got better.
Pruned that one entry only (never a global --prune-suppressions: other entries are
other sessions' frozen debt).
Z.AI's glm-4.6v vision endpoint enforces a 32768 max_tokens ceiling
server-side and 400s when a client sends a larger explicit max_tokens (e.g.
a client defaulting to 65536). paramSupport.ts's STRIP_RULES already has a
working clampToModelMaxOutput/maxOutputCap mechanism (used today for a
VolcEngine Kimi rule) but had no entry for zai/glm + glm-4.6v.
Added two rules: "zai" uses a fixed maxOutputCap (glm-4.6v is only reachable
there as a custom model attached to the connection, so it is not in
PROVIDER_MODELS["zai"] and clampToModelMaxOutput would find no catalog
ceiling); "glm" uses clampToModelMaxOutput (glm-4.6v IS in the registry
catalog there, GLM_SHARED_MODELS, maxOutputTokens: 32768).
Also discovered and fixed a second, deeper bug the "glm" rule alone would
not have caught: GlmExecutor.execute() drives its own fetch flow
(executeTransport()/transformForTransport()) and never runs through
DefaultExecutor.execute()'s stripUnsupportedParams() call site — so a
STRIP_RULES clamp entry for provider "glm" was dead code until
transformForTransport() now calls stripUnsupportedParams() directly.
Regression tests: tests/unit/zai-glm-max-tokens-clamp-7364.test.ts (reused
from the triage plan-file's RED probe, sanity assertion updated to lock the
fix instead of the bug) and tests/unit/glm-executor-max-tokens-clamp-7364.test.ts
(proves the real GlmExecutor.transformForTransport wiring, not just the
STRIP_RULES entry in isolation).
Gates run: check-file-size, check-complexity, check-cognitive-complexity,
typecheck:core, eslint (suppressions), tests/unit/zai-glm-max-tokens-clamp-7364.test.ts,
tests/unit/glm-executor-max-tokens-clamp-7364.test.ts,
tests/unit/executors-strip-unsupported-params.test.ts,
tests/unit/nvidia-minimax-thinking-strip.test.ts, tests/unit/glm-executor.test.ts — all green.
Refs #7364
* fix(codex): strip regex lookaround from tool schema patterns (port from 9router#1556)
Codex/OpenAI's Responses API rejects JSON Schema pattern fields using regex lookaround (e.g. ^(?=.*@).+$) with a 400 'regex lookaround is not supported' error. The existing numeric-field sanitizer (coerceSchemaNumericFields) was only wired into the translated-request path (openai-to-claude.ts), not the native codex/openai passthrough path (normalizeCodexTools in open-sse/executors/codex/tools.ts), so lookahead/lookbehind patterns reached upstream unmodified and broke tool calls for clients that emit them (e.g. IDE agent harnesses validating an email field).
Reported-by: evin (@evinjohnn) (https://github.com/decolua/9router/issues/1556)
* chore(changelog): move #1556 entry to changelog.d fragment
Consistency with the repo's canonical changelog.d/fixes/ workflow (avoids
merge-storm re-conflicts from editing CHANGELOG.md directly).
* refactor(codex): table-drive the regex-strip recursion to keep the complexity ratchet at baseline
The #1556 lookaround strip walked every sub-schema field with its own
copy-pasted if-block (properties / patternProperties / definitions / $defs,
then prefixItems / anyOf / oneOf / allOf), pushing stripUnsupportedRegexPatterns
past the cyclomatic threshold and check:complexity to 2057 > baseline 2056.
Collapse the eight near-identical blocks into two loops over the field-name
constants, with the object-map recursion factored into a helper. Same fields,
same traversal order, same behavior — complexity is back at baseline 2056 and
the #1556 regression tests still pass.
* fix(6954,6953): preserve system role + strip empty-signature thinking blocks
#6954 — System turns misattributed as assistant (claude-to-openai.ts:352)
The ternary `msg.role === 'user' || msg.role === 'tool' ? 'user' : 'assistant'`
mapped any non-user/non-tool role (including 'system') to 'assistant'.
Mid-conversation system turns (Claude format) lost their role on
translation to OpenAI format, causing them to be treated as assistant
output. Fix: add explicit 'system' branch to the ternary.
#6953 — Empty-signature thinking blocks poison Anthropic leg (openai-to-claude.ts)
Non-Anthropic providers (codex/gpt-5.x) synthesize thinking blocks with
signature:''\. On replay, the old code fabricated a DEFAULT_THINKING_CLAUDE_SIGNATURE
to fill the empty signature — but Anthropic rejects foreign signatures with
HTTP 400, permanently degrading combo/blend routes to codex-only.
Fix: strip thinking blocks with empty/missing signatures and redacted_thinking
blocks with empty/missing data entirely. They carry no replayable value.
Tests: 8 new tests (4 per bug), all passing. Existing #5312 and #5945
regression tests still pass — no interference.
* fix(6953): strip only signature:"" thinking blocks, preserve undefined signature
CI caught a regression: translator-helper-branches test had a Claude-format
thinking block without signature field (undefined) that was being stripped
by the original fix. The fix was too aggressive — it stripped both
signature:"" (non-Anthropic synthesized) and signature: undefined
(legitimate Claude-format).
Correct behavior:
- signature === "" (empty string): strip — hallmark of codex/gpt-5.x block
- signature === undefined: preserve with DEFAULT_THINKING_CLAUDE_SIGNATURE fallback
- redacted_thinking data === "": strip
- redacted_thinking data === undefined: preserve with fallback
Added regression test for undefined-signature preservation.
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
* fix(stream): reconcile encrypted Codex reasoning visibility without mutating upstream item
Resolves the collision between two open PRs on ensureVisibleResponsesReasoningSummary:
#7095 (xz-dev) found that chat clients see nothing when Codex exposes reasoning
only as encrypted_content, and added a visible placeholder — but did so by
mutating item.summary in place. #7176 (JxnLexn) found that same mutation
corrupts the forwarded response item, discarding the encrypted_content shape
Codex needs for follow-up requests, and removed the mutation — but that also
silently dropped the placeholder, so chat clients went back to seeing nothing.
The mutation existed only so a later line could read the summary text back off
the same item. getVisibleResponsesReasoningSummaryText() computes that text
without touching the item, so:
- synthetic response.reasoning_summary_text.delta / .part.done events still
carry the placeholder for chat clients (#7095's goal), and
- the forwarded response.output_item.done payload keeps its original
encrypted_content intact with no fabricated summary field (#7176's goal).
Applied at both call sites #7095 identified: the native Responses passthrough
in stream.ts/passthroughTailProcessor.ts, and the Responses-to-Chat-Completions
translator in openai-responses.ts.
Closes#7095, closes#7176.
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
* test(stream): guard the encrypted-reasoning mutation via the completed backfill path
The output_item.done line is echoed verbatim on the wire, so a re-introduced
item.summary mutation does NOT surface in that event — verified by re-injecting
the mutation, which left the existing assertion green. The mutation does surface
in the response.completed snapshot, where the captured reasoning item is
re-serialized when upstream sends an empty output (store: false).
Adds that case, which fails as expected when the mutation is re-introduced,
making the #7176 half of the reconciliation an enforced regression guard rather
than an incidental property of the current code path.
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
---------
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
stripEmptyOptionalToolArgs was allowlist-only (Read/Subagent) and only
stripped empty-string/empty-array values, so Responses API strict mode
(every property forced into `required`) could forward a forced non-empty
value (e.g. Agent.isolation) or a schema-declared default value verbatim
to the client. Add schema-aware drop-if-default and generalized
drop-if-empty (any tool, gated on schema.required), and thread each
tool's JSON Schema from the request's tools[] into the two streaming
call sites (response.output_item.done handling).
Closes#6951
* fix(electron): bump electron 42→43 + build better-sqlite3 from source (ABI 148) (#6605)
fix(electron): bump electron 42→43 + rebuild better-sqlite3 from source against the Electron ABI (148).
Electron 43 raises NODE_MODULE_VERSION to 148; better-sqlite3@12.11.1 has no electron-v148 prebuild, so the packaged app died with 'Nenhum driver SQLite disponível'. prepare-electron-standalone now compiles better-sqlite3 from source against the electron headers into build/Release (where 'bindings' resolves it). Validated by Electron Package Smoke (green) + local (node_register_module_v148).
Supersedes #6378. (--admin: the only reds are SonarQube/SonarCloud failing on a coverage-report artifact digest-mismatch — a GitHub Actions infra flake, not this diff; Sonar is green on main and the diff touches only the electron build.)
* deps: bump the development group across 1 directory with 6 updates (#6588)
deps: bump the development group (6 updates). Rebased onto current main; all checks green after the electron-smoke fix (#6605).
* fix(proxy): force CONNECT tunnel for HTTP proxied requests (undici 8.7) + production deps bump (#6620)
fix(proxy): force CONNECT tunnel for HTTP proxied requests (undici 8.7) + production deps bump.
undici 8.6+ changed ProxyAgent to forward plain-HTTP via request-proxy instead of CONNECT, breaking OAuth refresh through a connection proxy (501). proxyDispatcher now passes proxyTunnel:true. Validated: Unit Tests 3/8 (the OAuth-proxy test) green, new regression test green (fails without the fix on undici 8.7), SonarQube green.
Supersedes #6380. (--admin: the only red is Electron Package Smoke failing on a next-build artifact 'digest-mismatch' — a GitHub Actions infra flake corrupting the asar ('file data stream has unexpected number of bytes'); the better-sqlite3 rebuild itself succeeded (gyp ok) and the electron path is unchanged from #6605 which passed the smoke. Not this diff.)
* fix(6813): fix thinking budget zero drop and default thinkingConfig injection
- Fix truthy check for budget_tokens to allow 0
- Stop injecting default thinkingConfig when no knobs present
- Add tests covering all scenarios
Related: #6813
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Real OpenAI-compatible upstreams with stream_options.include_usage=true
send finish_reason in one chunk (usage: null) and the actual token counts
in a separate, trailing usage-only chunk (choices: [], usage: {...}).
Both the live translator (openai-responses.ts) and the legacy transformer
(responsesTransformer.ts) fired response.completed as soon as they saw
finish_reason, so the trailing usage chunk's token counts were captured
into state but never emitted -- Codex CLI and other /v1/responses
consumers saw response.completed with no usage field (permanent 0%
context-used).
Both translators now defer response.completed via an
awaitingTrailingUsage state flag when finish_reason arrives without
usage already captured, and complete on the next usage-only chunk (or
at stream end via the existing flush fallback) instead. Extracted the
duplicated events/emit boilerplate into a new
openai-responses/eventEmitter.ts leaf to keep the frozen
openai-responses.ts file under its file-size baseline.
Fixes 3 existing tests that encoded the old chunk ordering and adds a
permanent regression test (tests/unit/responses-usage-trailing-6906.ts)
covering both translators.
A Chat-Completions client can only express reasoning via the top-level
reasoning_effort hint and has no way to request a reasoning summary. When
that hint is promoted to the Responses API's reasoning.effort, the
upstream returns an empty summary and downstream chat clients see no
thinking stream (encrypted reasoning only).
Default reasoning.summary "auto" plus include ["reasoning.encrypted_content"]
on the effort-only path so the summary actually streams back to the chat
client, mirroring the Codex executor's ensureCodexReasoningSummary. An
explicit reasoning object from a Responses-shaped client is preserved
untouched, and reasoning_effort "none" is left without a summary.
Adds regression tests for the effort-only default, the none case, and
keeps the existing explicit-reasoning-object behavior unchanged.
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
* fix(responses): escape literal control chars in tool call JSON; emit status=failed on upstream error #6785
Two bugfixes in the Responses API translator:
1. escapeJsonStringValues() sanitizes tool call arguments containing
literal 0x0A/0x0D/0x09 bytes (emitted by Gemma4 models) into valid
JSON \n/\r/\t escapes, preventing SSE framing corruption. Only
escapes inside JSON string contexts — already-escaped sequences
and structural JSON pass through unchanged.
2. sendCompleted() checks state.upstreamError and emits status="failed"
with error.code + error.message instead of silently hardcoding
status="completed" + error=null, so mid-stream errors (e.g. Gemini
503 after partial content) are properly surfaced to the client.
3. stream.ts: calls translateResponse(null,...) before controller.error()
so the translator can emit close events (reasoning item done,
response.completed) before the stream is terminated.
* test(boundary): fix ESLint no-explicit-any warnings and quality gates
Green the PR against release/v3.8.47 quality gates without weakening tests:
- Replace @typescript-eslint/no-explicit-any in the new boundary/gemma4
tests with proper interfaces (ResponseBody, ToolDef, ToolArgs, SseEvent
item accessors) — fixes the "No new ESLint warnings" gate.
- Split tests/unit/translator-resp-openai-responses.test.ts (1079 LOC) by
extracting the round-trip suite into a sibling file so both stay under
the 800-line test cap — fixes check:file-size.
- Rename the 5 live boundary tests to *.live.test.ts, gate them behind
RUN_BOUNDARY_LIVE=1, add a test:boundary:live npm script and register the
glob in check-test-discovery COLLECTORS — fixes check:test-discovery
(they hit a live remote and must never run unopted in CI).
Co-authored-by: Markus Hartung <mail@hartmark.se>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* feat(codex): echo requested effort-suffixed model id in Responses payloads (#3697)
Codex CLI compatibility shim: the Responses API response.created/
response.in_progress/response.completed payloads now carry a `model`
field (previously absent), and for Codex-CLI-originated requests it
echoes the client-requested effort-suffixed model id (e.g.
gpt-5.5-xhigh) instead of the bare upstream id (gpt-5.5), so the Codex
CLI status line/model button shows the active reasoning effort.
- openai-responses.ts translator threads the upstream model into the
Responses event objects (additive, omitted when unknown).
- New isCodexOriginatedHeaders() (codexIdentity.ts) reuses PR #3481's
originator/User-Agent detection, header-based so it still fires when
a combo routes codex/gpt-5.5-xhigh to a non-codex upstream.
- chatCore's existing opt-in #1311 echoModel pipeline now also fires
automatically for Codex clients on the Responses API, regardless of
the echoRequestedModelName setting.
- responseModelEcho.ts now also rewrites the nested response.model
field the Responses API uses (previously only top-level model).
- /v1/models keeps returning models: [] for Codex (unchanged, #3481).
Regression guard: tests/unit/codex-effort-model-echo-3697.test.ts.
Closes#3697
* chore(merge): re-sync with release/v3.8.47 (restore CHANGELOG, keep own bullet)
* chore(merge): re-sync with release/v3.8.47; move changelog bullet to changelog.d fragment (merge-storm proof)
* fix(translator): defer content_block_start until GLM streams the tool name (port from 9router#2077)
GLM 5.2 (and similar OpenAI-compatible upstreams) stream a tool call's id and
function.name across separate SSE delta chunks. The openai-to-claude streaming
translator emitted content_block_start immediately on the id-only chunk with an empty
name; the Claude SSE protocol cannot patch a block after emission, so the later
name-only chunk was dropped and Claude Code rejected the tool_use with an empty tool
name / "No such tool available:". Defer content_block_start until the name arrives
(start on args if they arrive first), and emit a start for any orphaned id-only tool
call at finish so content_block_stop is never orphaned.
Reported-by: itiwant (https://github.com/decolua/9router/issues/2077)
* chore(6730): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
* fix(translator): strip empty cloud_base_branch from Cursor Subagent tool call (port from 9router#2446)
The Responses->Chat tool-arg cleanup (stripEmptyOptionalToolArgs) only stripped
empty-string/empty-array optional args for Claude Code's Read tool. Cursor's local
Subagent tool call therefore passed through with the cloud-only field
cloud_base_branch: "", which Cursor rejects ("cloud_base_branch may only be specified
when environment equals cloud") before starting the subagent. Extend the cleanup to an
allowlist of Read + Subagent; arbitrary tools stay untouched.
Reported-by: like3213934360-lab (https://github.com/decolua/9router/issues/2446)
* chore(6729): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
* fix(antigravity): surface aborted Gemini tool calls off end_turn
Gemini/Antigravity aborts a turn with finishReason MALFORMED_FUNCTION_CALL
(or a sibling like UNEXPECTED_TOOL_CALL) instead of completing cleanly. Both
Claude-facing translators collapsed these to a clean end_turn, hiding the
aborted tool call as a successful completion:
- the OpenAI hub path (openai-to-claude.ts convertFinishReason default), and
- the DIRECT Gemini->Claude path (gemini-to-claude.ts), which is the one
Claude Code actually hits through an antigravity/Gemini-routed model.
Add isAbortFinishReason() to finishReason.ts and map these reasons to
tool_use on both paths; genuinely unknown reasons still fall back to end_turn.
Co-authored-by: anhdiepmmk <n08ni.dieppn@gmail.com>
Inspired-by: https://github.com/decolua/9router/pull/2462
* chore(6713): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: anhdiepmmk <n08ni.dieppn@gmail.com>
* fix(volcengine): clamp Kimi max_tokens to Ark endpoint cap
VolcEngine Ark's Kimi coding-plan endpoint (ark.cn-beijing.volces.com)
enforces max_tokens <= 32768 server-side and returns 400 "integer above
maximum value, expected a value <= 32768" for anything over that ceiling.
OmniRoute's StripRule only supported dropping params outright, with no
numeric clamp mechanism, so a client sending a larger max_tokens (common
default, e.g. 65536) 400s outright against volcengine's kimi-k2-5-260127.
The 32768 cap is independently confirmed against two live-endpoint bug
reports hitting this exact Ark endpoint for both kimi-k2.5 and
kimi-k2.7-code (NousResearch/hermes-agent#51773, MoonshotAI/kimi-cli#1124),
not just upstream's own value — same cap upstream 9router#2460 uses.
StripRule gains two optional fields: `clampToModelMaxOutput` (clamp to the
model's own catalog maxOutputTokens ceiling, when set) and `maxOutputCap`
(a fixed endpoint-imposed ceiling); when both apply, the lower wins. The
new rule is scoped to the literal id `kimi-k2-5-260127` (OmniRoute's real
volcengine Kimi model, not upstream's `Kimi-K2.7-Code`), not a broad
/kimi/i regex, so it can never clamp an unrelated future Kimi listing
whose Ark cap may differ. glm-4-7-251222 (the other volcengine model) is
unaffected.
Inspired-by: https://github.com/decolua/9router/pull/2460
Co-authored-by: whale9820 <whale9820@users.noreply.github.com>
* chore(6712): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: whale9820 <whale9820@users.noreply.github.com>
open-sse/translator/request/claude-to-gemini.ts already guards against
sending thinkingConfig for gemma-4-* models (Gemma doesn't support it —
Vertex returns 400: "Thinking budget is not supported for this model"),
but the OpenAI-shape path (openai-to-gemini.ts) lacked the same guard, so
OpenAI-shape clients hitting a vertex gemma-4-* model still got a 400.
Mirrors the existing claude-to-gemini.ts guard: wrap the reasoning_effort
and Claude-shape thinking.budget_tokens branches with a model.startsWith
("gemma-4") check. Branch 3 (default includeThoughts for modern Gemini
models) already excludes non-"gemini" model ids and needed no change.
Inspired-by: https://github.com/decolua/9router/pull/2480
Co-authored-by: chy1211 <31048289+chy1211@users.noreply.github.com>
* fix(sse): unwrap bare {function:{…}} tools in openai→claude translation
Some OpenAI-shape clients send a tool as a bare `{ function: {...} }`
object, omitting the spec-required `type: "function"` parent wrapper.
The tools-mapping in openai-to-claude.ts (~line 366) only unwrapped
`tool.function` when `tool.type === "function"` was ALSO true, so a
bare-function tool fell through to `toolData = tool` (the wrapper
itself, with no `.name`), producing an empty `originalName` and
silently dropping the tool from the translated request — worse than
a 400, since the caller has no signal the tool never made it
upstream. Unwrap `tool.function` whenever present, independent of
the parent `type` field. Regression guard:
tests/unit/openai-to-claude-bare-tool.test.ts.
Co-authored-by: Samir Abis <me@samirabis.com>
Inspired-by: https://github.com/decolua/9router/pull/2473
* chore(6704): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: Samir Abis <me@samirabis.com>
