handleComboChat/handleRoundRobinCombo tracked lastStatus (first-write-wins),
lastError (last-write-wins), and earliestRetryAfter (global MIN across all
targets) independently, so the final unavailableResponse() could surface a
status/message pair from two different failing targets and decorate a
config-class error (e.g. Antigravity's 422 missing_project_id, which carries
no retryAfter of its own) with an unrelated target's long reset window.
- lastStatus now overwrites on every failure (last-write-wins), matching
lastError, so status and message always come from the same target.
- the "(reset after ...)" decoration is only applied when the surfaced
status is itself rate-limit-class (429/503) — see the new
open-sse/services/combo/unavailableRetryGate.ts leaf module (both
combo.ts and chat.ts are already over their file-size baseline, so the
gate logic lives in a new module and combo.ts only wires it in).
Co-authored-by: ikelvingo <im.kelvinwong@gmail.com>
`normalizeExecutorResult()` has always accepted `Response | { response, url, headers,
transformedBody }` — the bare arm is what the web/scraping executors return from their
error and passthrough paths, and `chatcore-upstream-timeouts.test.ts` already covers
that both shapes are handled. But `BaseExecutor.execute` has no explicit return type,
so TypeScript inferred it from the method's single `return` — the object shape alone.
Every override returning a bare `Response` was therefore reported as incompatible:
* 14 × TS2739 in `duckduckgo-web.ts`, whose `execute()` additionally pinned its own
signature to just the object shape while returning `errorResponse()` /
`processResponse()` (both `Response`) from 14 valid paths
* TS2416 in `felo-web.ts` and `gitlab.ts`, which declare `Promise<Response>`
Fix the declaration rather than the call sites: export `ExecutorExecuteResult` from
`base.ts` — the same union `normalizeExecutorResult()` accepts — and annotate
`BaseExecutor.execute` with it. `duckduckgo-web.ts` then drops its over-narrow
annotation, matching BaseExecutor and the ~38 other executors that let the return type
be inferred.
Two subclasses read `.response` straight off `super.execute()` and now narrow first:
* `github.ts` — the existing `!result.response` guard already meant "bare Response,
nothing to materialize"; it is now expressed as `result instanceof Response`, which
is the same branch for every input (bare / object / nullish)
* `pollinations.ts` — reads the status through both arms for its pool bookkeeping
Wrapping DuckDuckGo's 14 returns would have been the wrong fix: the values are already
correct, and `normalizeExecutorResult()` produces exactly `{ response, url: "",
headers: {}, transformedBody: null }` for them.
Validation: full tsc error-set diff against the base config — 335 -> 319, **zero new
errors** (line-number-agnostic diff is empty; the two `duckduckgo-web.ts` TS2345s that
appear to move are the same two pre-existing errors renumbered by added comments, and
are left for a later slice). `typecheck:core` clean, `check:type-coverage` 92.17% ->
94.17%, and 49 of the 50 existing test files importing a touched executor pass —
`plan3-p0.test.ts` fails identically with and without this change (it reads the
developer's real ~/.omniroute DB rather than a test-scoped DATA_DIR).
The new test pins the runtime behavior of the narrowing so a later simplification
cannot quietly drop the bare-Response arm.
`gemini-business.ts` built its upstream fetch options with `combineAbortSignals(...)`,
which is defined nowhere in the repository. The module imports `mergeAbortSignals`
from `./base.ts` on line 31 and never used it — a rename that was only half applied.
Because the call sits inside the fetch options object literal, the ReferenceError was
thrown while *constructing* the arguments, before `fetch()` ran, and the surrounding
try/catch turned it into `makeErrorResult(502, "Gemini Business network error: ...")`.
So every Gemini Business request failed with what reads like an upstream outage. The
provider is registered and reachable (`open-sse/executors/index.ts`), so this affects
the whole provider, not an edge case.
`mergeAbortSignals(primary, secondary)` requires two real signals while
`ExecuteInput.signal` is `AbortSignal | null | undefined`, so the call is guarded and
falls back to the timeout alone — the same shape huggingchat, grok-web, claude-web,
and ninerouter already use.
Why it went unnoticed: this file is only type-checked by `open-sse/tsconfig.json`,
whose runs abort at `TS5101` (the deprecated `baseUrl`) before any file is checked,
and `typecheck:core` covers a curated 26-file allowlist that excludes every executor.
Removing that config error is #8473; this bug is what the first full run surfaced.
TDD: the two new tests fail on the parent commit — `execute()` never reaches the
stubbed `fetch` — and pass with the fix. They also cover the null-signal path, since
that is where an unguarded `mergeAbortSignals` would throw next.
First slice of the TypeScript 7 migration split requested on #7697: resolve the
type diagnostics under `open-sse/tsconfig.json` in the lowest-risk modules, with
no toolchain change. 12 diagnostics across 8 files, all outside the hot path —
`chatCore.ts` and `stream.ts` are deliberately left for a later, standalone slice.
Fixes, by cause:
* `Transformer.cancel` (progressTracker, sseHeartbeat, and stream.ts's existing
handler) — the WHATWG Streams standard defines `transformer.cancel(reason)` and
Node implements it (verified on v24: cancelling the readable side invokes it),
but `lib.dom.d.ts` still omits it from `Transformer`, so every such handler was
TS2353. These handlers clear the heartbeat/progress intervals when an SSE client
disconnects, so deleting them to satisfy the checker would leak a timer per
abandoned stream. The interface is patched in `open-sse/types.d.ts` instead.
* `earlyStreamKeepalive` — `SettledHandler` was discriminated by `ok: true | false`.
This workspace compiles with `strictNullChecks: false`, where a boolean-literal
discriminant narrows the positive branch but not the negative one, so reading
`.error` off the rejected arm did not type-check (the two `.response` reads
elsewhere in the file did, which is why only one site errored). Retagged with a
string discriminant, which narrows both branches under the same settings.
* `toolCallShim` / `openai-responses` — assigning back to a property declared
`unknown` resets the `typeof` narrowing, so the following comparison no longer
saw a number/array. Both now read through a local. The `Read` limit clamp is
behavior-identical: its two branches are mutually exclusive at READ_MAX_LIMIT 2000.
* `sanitizeToolResultId` — takes `unknown` but forwards to a `string` parameter; a
non-string id previously reached `.replace()` and threw. Coerced instead.
* `openaiHelper` — `opts = {}` inferred `{}`; typed as `FilterToOpenAIFormatOptions`.
* `cursorAgentProtobuf` — `Buffer.alloc(0)` infers `Buffer<ArrayBuffer>` under
@types/node 26 while the decoded field is `Buffer<ArrayBufferLike>`; the locals
now use bare `Buffer`, matching `requestMetadata` a few lines above.
Validation: 335 -> 321 diagnostics with zero new errors (full tsc error-set diff
against the base config). typecheck:core clean, lint clean, check:type-coverage
92.17% -> 94.17%. All 114 existing test files that import a touched module pass;
`plan3-p0.test.ts` fails identically with and without this change (it reads the
developer's real ~/.omniroute DB instead of a test-scoped DATA_DIR).
The new test covers the three behavioral surfaces rather than the refactors the
existing keepalive/heartbeat suites already hold: that `transformer.cancel()`
really fires and can clear an interval, the id coercion, and the limit-clamp bounds.
TypeScript 6.x raises TS5101 on `open-sse/tsconfig.json`: `baseUrl` is
deprecated and stops functioning in TypeScript 7.0. It was paired with
`ignoreDeprecations: "5.0"`, which no longer silences it under TS 6 (the
compiler now demands "6.0").
Remove `baseUrl: ".."` and rewrite the `paths` mappings relative to the
tsconfig's own directory, which is how TypeScript resolves them with no
baseUrl set:
"@/*" ./src/* -> ../src/*
"@omniroute/open-sse" ./open-sse -> ../open-sse
"@omniroute/open-sse/*" ./open-sse/* -> ../open-sse/*
`ignoreDeprecations` goes with it — baseUrl was the only deprecated option
it was suppressing.
Verified by diffing the full tsc error set against the previous config (the
base run used `ignoreDeprecations: "6.0"` so compilation proceeds past the
config error, which otherwise aborts type-checking and masks everything):
zero new errors, 28 fewer. All 28 were in `electron/*.js`, which
`baseUrl: ".."` had been dragging into the open-sse program via
root-relative resolution. Scoping the program back to open-sse also moves
`check:type-coverage` from 92.17% to 94.01%; the ratchet direction is up so
the gate passes, and the baseline is deliberately left alone because the
gain is a measurement-scope change rather than new typing work.
The guard test asserts no tsconfig reintroduces `baseUrl` or
`ignoreDeprecations`, and that every `paths` target still resolves to a real
directory — the second half is the part that matters, since dropping baseUrl
silently changes what those mappings point at.
The openai-responses translator tracked tool calls in its own funcCallIds/
funcNames/funcArgsBuf bookkeeping without ever writing to the shared
state.toolCalls Map that stream.ts's completion-log summary builder reads.
Every openai->openai-responses translated stream with a tool call was
persisted with finish_reason "stop" and no tool_calls, even though the
actual SSE events sent to the client were correct.
* fix(providers): prevent health sweep from expiring devin-cli local credentials
* fix(auth): drop devin-cli from supportsTokenRefresh explicit set (#8407)
Root cause: listing "devin-cli" as refresh-capable made tokenHealthCheck
force-expire local CLI connections that never have a refresh token.
Remove it from the explicit set (keep windsurf) and drop the health-check
provider hardcode — the existing supportsTokenRefresh=false guard is enough.
* fix(api): stop the 2000-token safety buffer from inflating usage.prompt_tokens in the client response (#8331)
* fix(sse): scope #8331's usage-buffer fix around Claude-Code-compatible providers
The #8331 fix correctly stopped folding the 2000-token context-window safety
margin into client-visible prompt_tokens/input_tokens/total_tokens for normal
API metering clients. But it also silently changed the response shape for
Claude-Code-compatible providers, whose own context accounting reads the
buffered number straight out of usage — regressing
tests/unit/cc-compatible-provider.test.ts (expected 2007, got 7).
Fold the computed context_budget_* fields back into the visible usage fields
for that one path only (applyClientUsageBuffer's new
preserveContextBudgetInVisibleUsage option, gated on the existing
isClaudeCodeCompatible flag in chatCore.ts). Every other caller keeps the
real, unbuffered #8331 numbers.
Live incident: streamed assistant text was losing the spaces BETWEEN words
(e.g. "Bilden är en riktig JPEG nu" -> "Bildenärenriktig JPEG nu") on the
Responses-API and Claude streaming paths.
stripInternalReasoningPlaceholder() (#8081/#8162) is called on every
individual delta.content chunk, and unconditionally called .trim() even when
its sentinel ("(prior reasoning summary unavailable)") was never present in
that chunk. Tokenizers commonly emit sub-word tokens with a leading space as
part of the token (e.g. " en", " riktig") -- each such chunk got its only
whitespace character (the inter-word space) silently trimmed away before
being appended to the accumulated message, while the words themselves stayed
intact. Punctuation-only chunks were largely unaffected, matching what was
observed live.
Only trims when the sentinel is actually present -- preserves the original
#8081 intent (collapse a placeholder-only chunk to "") without touching the
overwhelming majority of chunks that never contain it.
Co-authored-by: Markus Hartung <markus.hartream@gmail.com>
* fix(providers): reconcile Kimi K3 vision when attachment contradicts modalities
Synced models.dev rows for kimi-coding*/k3 can ship attachment=false while
modalities_input still lists image/video. Prefer the modality signal (and
normalize at sync + resolve) so supportsVision, attachment, and exposed
modalities agree.
Closes#8250
* fix(providers): keep Kimi K3 static fallback text-only (#8250)
The Kimi K3 vision reconciliation added supportsVision=true directly
to the kimi-coding registry entry for id "k3". That entry is the
static/stable fallback catalog used when discovered capabilities are
unavailable, and it must stay text-only per #4071 — the vision fix is
already applied correctly on the discovered path via
MODEL_SPECS["kimi-k3"] (aliases: ["k3"]) and
modelCapabilities.ts::resolveVisionCapability.
Restores the invariants guarded by
tests/unit/kimi-k2.7-code-registration.test.ts ("Kimi Code k3 fallback
leaves discovered capabilities unset") and
tests/unit/catalog-updates-v3829-kimi-qwen.test.ts ("kmca stable
fallback only carries documented static capabilities").
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(translator): cap thinking budget on explicit budget_tokens path
* fix(translator): stop dropping thinkingConfig on cap-0 reasoning_effort path
The thinking-budget-cap guard added in this branch skipped thinkingConfig
entirely whenever a model's thinkingBudgetCap was 0 (e.g. gemini-3-flash),
including on the reasoning_effort/budgetMap path. That regressed the
pre-#6943 native-defaults contract (thinkingBudget 0 / includeThoughts
false must still be present) and crashed callers that read
`.thinkingConfig.thinkingBudget` unconditionally
(translator-openai-to-gemini-defaults.test.ts).
Also restore includeThoughts:true on the Claude-format explicit
thinking.budget_tokens path (openai-to-gemini.ts's Claude-format field and
claude-to-gemini.ts's native thinking field): budget_tokens:0 there is the
client's dynamic-thinking sentinel (#6813), not an off-switch, and must
stay true even after the new capping — the cap must only clamp positive
explicit values, never flip the zero sentinel's semantics.
Updates two tests this branch added that encoded the incorrect
"omit thinkingConfig / includeThoughts:false for the 0 sentinel" behavior,
to match the pre-existing, still-required contracts above.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(providers): revert scope-creep flip of Gemini 3.5/3.6 Flash supportsThinking
The thinking-budget-cap fix accidentally expanded 5 shorthand modelSpecs
entries (gemini-3.5-flash, gemini-3.5-flash-low, gemini-3.6-flash-high/
medium/low) into explicit objects setting supportsThinking:true and
thinkingBudgetCap:24576. That flip was unrelated to the two proven test
regressions (translator-openai-to-gemini-defaults.test.ts and
claude-to-gemini-budget-tokens-zero-6813.test.ts, which only exercise
gemini-3-flash-preview, gemini-3.1-pro and gemini-2.5-pro) and reopens a
deliberately closed path from #8013: Antigravity still rejects
client-supplied thinking params for these Gemini 3.5/3.6 Flash tier ids,
so supportsThinking must stay false (inherited from
GEMINI_35_FLASH_MODEL_SPEC).
Reverted all 5 entries back to the release shorthand
`{ ...GEMINI_35_FLASH_MODEL_SPEC }`. gemini-3-flash, gemini-3.1-pro and
gemini-2.5-pro (the models the regression tests actually exercise) were
already correctly specced in the release baseline and are untouched.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
#8010 registered the codex-responses engine in the compression catalog
(engineCatalog.ts, stackPriority 12 between rtk's 10 and headroom's 15)
but never added it to adaptiveCompression/ladder.ts's AGGRESSIVENESS and
REDUCTION_FACTOR maps. Those maps' own header documents that they must
cover every real catalog/registry engine, not just the 7 in
DEFAULT_LADDER, so an operator adding codex-responses via ladderOverride
silently fell back to aggressivenessOf() === 0 (same as "off") and
expectedReductionFactor() === 0.9 (the generic default), breaking
floor-mode escalation ranking for any ladder that includes it.
Add "codex-responses" to both maps between rtk and ionizer, matching
its stackPriority (12) sitting between rtk's (10) and ionizer's (13):
- AGGRESSIVENESS: 22 (between rtk's 20 and ionizer's 25)
- REDUCTION_FACTOR: 0.84 (between rtk's 0.85 and ionizer's 0.83),
reflecting its "lossless-first, bounded diagnostic" guidance in
engineCatalog.ts
Validation: tests/unit/ladder-engine-maps-6533.test.ts red -> green
(2 of 3 tests were failing on the missing engine; all 3 pass after the
fix). Sanity-checked neighbors compression-exclusions.test.ts and
compression/adaptive-resolve-plan.test.ts still pass.
Refs #8010
Context: #8183 introduced AUTO_COMBO_NOAUTH_ALLOWLIST (opencode, felo-web) to
gate no-auth providers out of every auto/* candidate pool, motivated by public
HTTP egress reliability on the reference VPS (.15) — several no-auth backends
(duckduckgo-web, theoldllm, chipotle, aihorde) were flaky there. Its own tests
(noauth-autocombo-allowlist.test.ts, virtual-auto-combo.test.ts) never
exercised the auto/<family> path, so the gate silently applied there too.
auto/<family> combos (#6453, e.g. auto/glm, auto/zai) are a different axis:
an identity selector ("route to whatever genuinely serves GLM"), not a
reliability-curated pool. auggie (local CLI subprocess, zero HTTP egress —
the reliability concern #8183 targets doesn't even apply to it) advertises a
literal glm-5.2 model and had an explicit design-test seat in auto/glm since
#7032, but the #8183 allowlist silently excluded it from that pool.
Operator decision (2026-07-24): the no-auth allowlist gate keeps applying to
category/tier and flat-variant auto/* pools (auto/best-free, auto/coding:fast,
...), but auto/<family> pools bypass it — any no-auth backend that genuinely
serves the family is admitted.
Fix: thread a `bypassAllowlist` flag through isChatAutoComboNoAuthProvider()
and getNoAuthCandidates(), set to `Boolean(spec?.family)` at the single call
site in createVirtualAutoCombo(). Family narrowing (buildFamilyCandidateFilter)
still runs afterward, so a bypassed no-auth candidate only survives if its
model actually belongs to the requested family. Category/tier and flat-variant
pools (spec.family unset) keep the gate fully intact.
Validation:
- tests/unit/autoCombo/provider-family-combos.test.ts:136 was red (expected
["auggie","glm","zai"], got ["glm","zai"]) — now green (11/11 passing).
- tests/unit/noauth-autocombo-allowlist.test.ts (3/3) and
tests/unit/virtual-auto-combo.test.ts (10/10, including "restricts the
no-auth pool to the allowlist") stay green — the #8183 gate is untouched for
spec-less/category/tier pools.
- Full tests/unit/autoCombo/ vitest sweep: 5 files, 36/36 passing.
- npm run typecheck:core clean.
- 4 unrelated failures pre-exist on origin/release/v3.8.49 (verified via `git
show HEAD:<path>` swap, no stash) in
tests/unit/auto-combo-credentialed-model-pool.test.ts (antigravity/gemini-3.5
credentialed-pool logic, untouched by this change).
Refs #8183, Refs #6453, Refs #7032
Contract vs cap: #6863 requires a model lockout to honor a VERIFIED
upstream quota reset exactly (e.g. Antigravity "Resets in 92h27m28s",
shipped in v3.8.47). #7940 requires SYNTHETIC exact-cooldown estimates
(the quota_exhausted until-midnight heuristic) to respect the
operator's maxCooldownMs so they cannot balloon unbounded. Both are
legitimate, non-conflicting contracts — they apply to different kinds
of values.
Root cause: #7980 (fixing #7940) changed recordModelLockoutFailure()
in open-sse/services/accountFallback.ts to unconditionally clamp
every exactCooldownMs against maxCooldownMs, with no way to
distinguish a verified upstream reset from a synthetic estimate. A
real ~92h reset got clamped to the operator's ~30min cap, and the
router went on hammering 429 against quota that was known not to
recover for days — regressing #6863's contract by omission, not by
new policy (the "honor it exactly" docstrings on
selectLockoutCooldownMs() and its call sites were left untouched and
now describe dead code).
Fix: add an opt-in `exactCooldownVerified` flag to
recordModelLockoutFailure()'s options. When true, exactCooldownMs
bypasses the maxCooldownMs clamp entirely; when false/omitted
(the default), behavior is byte-identical to before this change.
Set the flag only at the 4 call sites that already carry upstream
provenance for the value they pass — usedUpstreamRetryHint /
quotaResetHintMs from checkFallbackError():
- open-sse/services/combo.ts (2 sites): exactCooldownVerified
mirrors lockoutHintMs > 0, which is only ever nonzero when it
traces back to a genuine upstream signal.
- src/sse/services/auth.ts (2 sites): exactCooldownVerified mirrors
the same usedUpstreamRetryHint / quotaResetHintMs check already
used to derive exactCooldownMs at each site.
The quota_exhausted → until-midnight synthetic default and plain
exponential backoff are untouched and stay capped, per #7940. The
two other recordModelLockoutFailure call sites (combo.ts quality
failure, auth.ts local-404/grok-web-403) never carry a verified hint
and were left unmodified.
Validation (TDD): tests/unit/combo-lockout-quota-reset-6863.test.ts
red→green with its assertions unchanged (was clamping ~332,848,000ms
to ~1,799,995ms; now honors the parsed reset). Added a boundary pair
to tests/unit/model-lockout-exact-cooldown-cap.test.ts proving the
same magnitude resolves differently by provenance: synthetic stays
capped, verified passes through whole. Full existing suite in that
file plus combo-model-lockout-honors-reset-1308.test.ts stay green
unmodified. Swept 45 lockout/cooldown-adjacent test files (502/505
passing); the 3 failures reproduce byte-identical on a pristine
origin/release/v3.8.49 checkout (PROVIDER_BREAKER_FAILURE_STATUSES
ReferenceError in untouched chat.ts, and a documented timing-sensitive
serial test) — confirmed pre-existing, out of this fix's scope.
npm run typecheck:core and npm run lint are clean.
Refs #6863
Refs #7940
Refs #7980
Regression: #8162 (port of #8081) added an unconditional `.trim()` to
stripInternalReasoningPlaceholder(), applied to every streaming
delta.content chunk across 3 call-sites (openai-to-claude.ts,
openai-responses.ts, responsesTransformer.ts). Leading/trailing
whitespace at a chunk boundary is a real word boundary between
streaming fragments; trimming it glues adjacent chunks together on
the client ("Hello, " + "world." + " Bye." -> "Hello,world.Bye.").
Fix: early-return via .includes() before the replaceAll+trim, so the
function is a true no-op when the sentinel is absent from the chunk.
Behavior when the sentinel IS present is unchanged.
Validation:
- tests/unit/streaming-reasoning-dedup-5786.test.ts: the "(A-guard)"
test was RED on the base branch ('Hello,world.Bye.' vs
'Hello, world. Bye.'); GREEN after the fix (4/4 passing).
- tests/unit/translator-resp-openai-to-claude.test.ts: added a new
multi-chunk boundary-whitespace regression test, proven RED against
the pre-fix code (12/13), GREEN after (13/13).
- No regressions in responses-transformer.test.ts (17/17),
responses-transformer-dense-output.test.ts (3/3), or the other
suites exercising the shared placeholder utility (160/160 total
across all consumers).
Refs #8162
Refs #8081
Claude→OpenAI translation was emitting a literal </think> into
delta.content for ordinary Chat Completions clients. Reasoning already
ships as reasoning_content, so default to suppress and keep
x-omniroute-thinking-marker: on as the #4633 opt-in.
* fix(cpa): isolate credential pool failures
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(cpa): forward transport through the chatCore key-health wrapper
The local recordKeyHealthStatus wrapper in handleChatCore only declared
(status, creds), so the transport argument added for CPA credential-pool
isolation was silently dropped at the call site (TS2554 "Expected 2
arguments, but got 3" once chatCore.ts is typechecked with tsc directly —
this file is not in tsconfig.typecheck-core.json's file list, so `npm run
typecheck:core` did not surface it). The CPA isolation guard in
keyHealth.ts never received `transport`, so it never fired.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Ravi Tharuma <RaviTharuma@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Node/undici fetch is rejected by Notion's edge with HTTP 200
temporarily-unavailable and empty assistant text (messages appear in the
thread, UI shows 502 No response from Notion AI). The same cookie and body
succeed via curl/Schannel and a browser Chrome JA3 handshake.
Route inference through tls-client-node (chrome_146), matching Claude/
Perplexity web providers. Also detect nested patch-start error objects so
operators see temporarily-unavailable instead of a misleading empty-body
502, and treat that subtype as retryable.
Verified live: notion-web/fable-5, hyperagent/fable, and
promptql/vertex-claude-fable-5 all return PONG through the packaged
backend; unit tests 83/83.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* feat(chaos+ponytail): parallel chaos-mode dispatch + ponytail output style (rebased on v3.8.49)
- Chaos mode: new auto/chaos variant fans the prompt out to the top-N
stable models in parallel and returns a single merged SSE stream.
- Progressive streaming: each panel model's answer is enqueued as it
lands (omni-chaos-part event), instead of awaiting the whole panel.
- withTimeout now aborts the underlying request (modelAbortSignal) on
timeout so the connection is released, not leaked.
- concatSseText parses both OpenAI and Anthropic SSE wire formats.
- autoPrefix/modePacks add the chaos-mode weight pack; virtualFactory
materializes auto/chaos with fusion strategy + chaos config flag.
- Ponytail (lazy-senior-dev mode) integrated into the existing
OUTPUT_STYLE_CATALOG registry (id 'ponytail') so it rides the production
output-style injector, instead of a bespoke duplicate module. Dev-only
scripts and the duplicate ponytail/ module are removed.
- Tests: chaosEngine/chaosVirtualCombo cover panel dispatch, progressive
broadcast, timeout abort, and Anthropic parsing; autoCombo pack count
updated to 6.
Rebased onto release/v3.8.49 (no provider-registry or validation changes —
those are split out per review).
* optimize(chaos+ponytail): i18n ponytail, dedupl chaos dispatch, provider diversity
- Ponytail: add vi/ja/pt-BR/id i18n with lite/full/ultra levels
- chaosEngine: extract dispatchOnePanelModel (shared), add onResult for
progressive SSE streaming, fix withTimeout anti-pattern
- virtualFactory: deduplicate chaos panel by provider, add tuning overrides
- dispatchChaosFromCombo: accept ChaosTuning, enforce minPanel
- Add/port 8 node-runner tests for ponytail i18n + catalog integrity
- Add muse-spark-web.ts to KNOWN_MISSING_ERROR_HELPER (pre-existing)
* fix(8264): use HandleSingleModel type in chaosEngine dispatch (base-drift)
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
sanitizeStreamingChunk() only stripped zero-width characters (U+200B,
U+200C, U+200D, U+FEFF) from OpenAI-format choices[].delta.content.
Anthropic-native content_block_delta events with text_delta or
thinking_delta payloads bypassed that path entirely, leaking U+200D
to clients on the Messages API streaming route.
Add a content_block_delta branch that strips zero-width characters
from delta.text and delta.thinking before returning the event,
matching the existing OpenAI path behavior from #5857.
Co-authored-by: Austin Liu <austinliu@Austins-MacBook-Air-3.local>
#8232 correctly targeted OAuth auto-refresh for legacy stored Gemini CLI connections, but exceeded that compatibility goal by recreating a complete routable and UI-visible provider with an Antigravity model catalog.
Preserve legacy refresh by mapping gemini-cli rows to the existing Gemini OAuth credentials. Remove the public provider registry, OAuth preset, model routing snapshot, and canonical-provider tests while keeping Gemini API and Antigravity unchanged.
#2101 still hard-stopped model-not-supported failures when upstream
wrapped them as invalid_request_error / Bad Request. Keep models in the
combo and try the next target (#8251, residual of #5249).
- export MODEL_ACCESS_DENIED_PATTERNS + broaden does-not-support shapes
- add isModelScoped400() and exempt it from the body-specific stop guard
- regression tests for wrapper forms; body-specific stop still intact
Co-authored-by: Ravi Tharuma <RaviTharuma@users.noreply.github.com>
* feat(providers): add Typhoon and Inception Mercury API-key providers
Two OpenAI-compatible API-key providers, each verified against a live
endpoint smoke test with a negative control before registration: an
unknown route answers 404 while /v1/chat/completions answers 401, which
rules out gateways that reply identically to every path.
- typhoon (SCB 10X, Thailand): first Thai-first provider in the catalog.
/v1/models answers 200 unauthenticated and serves exactly one chat
model, typhoon-v2.5-30b-a3b-instruct (128K ctx). The docs also list
typhoon-v2.1-12b-instruct, but the live endpoint no longer serves it,
so it is deliberately not registered. The typhoon-ocr* and typhoon-asr*
entries are OCR and speech models, not chat, and are omitted.
- inception (Inception Labs): first diffusion LLM (dLLM) in the catalog.
mercury-2 has a 128K context, 50K max output, and supports tools,
json_mode and structured outputs. The mercury-coder models advertised
on the vendor blog are no longer served by the live endpoint and are
therefore not registered.
Both expose a working /v1/models catalog, so they are added to
NAMED_OPENAI_STYLE_PROVIDERS for discovery and key validation.
Free-tier metadata is claimed only where it is documented and durable:
Typhoon issues a free API key rate-limited to 5 req/s and 200 req/m, and
Inception grants 10M tokens on signup with no card, so both are
registered with hasFree: true.
Provider count moves from 280 to 282; README/AGENTS/CLAUDE counters were
stale at 278 and are resynced against docs/reference/PROVIDER_REFERENCE.md.
* fix(8170): union frontier-labs Inception+Writer, regen ref+golden
* fix(8170): close inception object + regen ref/golden
---------
Co-authored-by: Álvaro Ángel Molina <alvaretto@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* feat(providers): add Sarvam AI, Writer Palmyra and PLaMo API-key providers
Three OpenAI-compatible API-key providers, each verified against a live
endpoint smoke test before registration:
- sarvam (India): /v1/models answers 200 unauthenticated and lists
sarvam-105b (128K ctx) and sarvam-30b (64K ctx). The older sarvam-m is
discontinued upstream and is deliberately not registered.
- writer (Palmyra): api.writer.com exposes the OpenAI alias
/v1/chat/completions alongside its native /v1/chat — confirmed with a
negative control, since an unknown route answers 404 'endpoint not
available via API gateway' while /v1/chat/completions answers 401.
Registers palmyra-x5 (1M ctx) and palmyra-x4 (128K ctx); the
medical/financial/creative/vision variants are deprecated upstream and
are omitted.
- plamo (Preferred Networks, Japan): only plamo-3.0-prime (262K ctx) is
registered. plamo-3.0-prime-beta is discontinued on 2026-07-31 and
plamo-2.2-prime on 2026-09-30, so neither is worth wiring up.
Free-tier metadata is claimed only where it is documented and durable:
Sarvam ships a permanent signup credit, while PLaMo's 10M-token grant is
a campaign that expires on 2026-07-31 and Writer documents no free tier,
so both are registered with hasFree: false.
* regen golden+ref
---------
Co-authored-by: Álvaro Ángel Molina <alvaretto@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
