The agentrouter.org upstream WAF returns 400 content-blocked
intermittently when:
1. messages[].content contains a blocked keyword (Lorem ipsum, the
phrase 'language model' alone, 'virtual assistant', etc.); or
2. Requests from the same IP/key arrive in a burst, after which the
WAF's per-IP suspicion bucket starts blocking content that would
normally pass. The bucket relaxes after ~5-10s of idle.
Apply three mitigations:
1. Burst guard (open-sse/services/wafRateLimit.ts)
Per-bucket (provider+url) gate that enforces a 500ms minimum gap
between outbound requests to agentrouter. Configurable via
configureWafRateLimit(). Tested in tests/unit/wafRateLimit.test.ts.
2. Reactive retry (BaseExecutor.WAF_RETRY_CONFIG in base.ts)
New WAF_RETRY_CONFIG with maxAttempts=2, delayMs=1500,
backoffMultiplier=2. When the upstream returns 400 with a body that
matches /content[_-]blocked/i, retry the same URL with exponential
backoff (1.5s, 3.0s) before falling through to the 429/401/fallback
chain. Tested in tests/unit/base-executor-waf-retry.test.ts.
3. Documentation (docs/security/AGENTROUTER_WAF.md)
Blocklist of always-blocked and almost-always-blocked patterns,
behavior under load, guidance for prompts/tool output, and pointers
to the relevant code paths in OmniRoute.
These are belt-and-suspenders: the burst guard prevents the WAF from
activating on normal traffic, and the reactive retry recovers when it
does anyway. Together they should eliminate the intermittent
400 content-blocked that Claude Code sees when running through
agentrouter via OmniRoute.
Refs #9275 follow-up. Test: 'WAF retry config shape' and 'WAF retry
differs from generic' guard the WAF_RETRY_CONFIG contract so future
refactors don't accidentally collapse the two retry paths.
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* fix(routing): bare model ids route to codex first; validate synced candidates
Two bare-model-routing bugs surfaced in the field when an OmniRoute
deployment had a codex subscription whose cookie quota was exhausted
(retry-after 429047s / ~5 days) AND an active kiro connection whose
upstream sync briefly advertised 'claude-opus-5' before kiro vendored
it into the static registry.
1. Bare 'gpt-5.6-sol' (and friends) routed to the codex provider even
when the user had explicitly configured 'agentrouter' as their
provider (via model_provider in codex CLI). With codex in cooldown,
every bare request 429'd. Fix: extend CODEX_NATIVE_UNPREFIXED_MODELS
to include the full gpt-5.6-sol tier set + gpt-5.5 + the related
codex-native ids. The Codex CLI default is now actually honored;
users can still prefix 'agentrouter/gpt-5.6-sol' to opt into a
specific provider.
2. Bare 'claude-opus-5' silently routed to 'kiro' when kiro's synced
/v1/models catalog had that id (likely from a transient upstream
quirk). kiro's static registry never cataloged claude-opus-5, so
the upstream call 404'd. Fix: validate activeSyncedProviders against
MODEL_TO_PROVIDERS before merging them into the candidate list.
Auto-discovery still wins when the model id has no static entry
(brand-new models from upstream keep working).
Bonus: when handleNoCredentials returns a 404 'No active credentials for
provider: X' error, surface the top-3 candidate aliases (e.g.
'anthropic/claude-opus-5, claude/claude-opus-5, agentrouter/claude-opus-5')
so the operator can pick a working prefix instead of staring at a wall.
Tests (all pass, 25 regression tests preserved):
- tests/unit/fix-bare-model-precedence.test.ts (7 tests)
- tests/unit/fix-synced-model-validation.test.ts (3 tests)
- tests/unit/fix-error-message-candidates.test.ts (3 tests)
- tests/unit/fix-bare-routing-fallback.test.ts (7 tests)
* fix(tests): replace lorem ipsum with neutral text to avoid agentrouter WAF
The agentrouter.org WAF blocks requests containing 'lorem ipsum' in
messages[].content. When Claude Code reads test files via the Read tool,
the content appears in tool_result blocks which can trigger the filter.
Replace 'lorem ipsum dolor sit amet' with 'example content for testing
purposes' in compression harness test to avoid false positives.
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
fix(agentrouter): infer protocol from client endpoint
- /v1/responses resolves AgentRouter as openai-responses
- /v1/chat/completions resolves AgentRouter as openai
- /v1/messages resolves AgentRouter as claude
- Per-request protocol and credential cloning (no SQLite mutation)
- Codex 0.146.0 and Claude Code 2.1.220 identity alignment
- response.completed.usage.total_tokens normalization for strict Codex clients
Closes#9224
G1 (v3.8.51): section (2) of check-known-symbols no longer regex-scans
strategy === "..." literals from combo source. The handled set now comes
from a runtime-imported dispatch registry (open-sse/services/combo/
strategyDispatch.ts) that imports the real ordering functions and enumerates
which strategies they implement. This keeps the canonical-not-handled gate
correct under the upcoming R0.3 registry dispatch, which removes the
strategy === branches the regex relied on.
- Adds HANDLED_COMBO_STRATEGIES registry (all 20 canonical strategies) + binds
the real dispatch leaves (applyStrategyOrdering, resolveAutoStrategyOrder,
tryFusionDispatch, tryPipelineDispatch, resolveComboTargetPipeline).
- main() imports the registry instead of reading/sourcing combo files.
- extractHandledStrategies + diffComboStrategies stay exported (pure, tested).
- New TDD test proves the runtime enumeration covers canonical exactly.
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
- memoryTools: apiKeyId now optional; falls back to caller principal
(HTTP auth headers / OMNIROUTE_API_KEY env) so agents can store memory
without knowing their key id
- memorySkillsInjection: server-side memory_* builtins only injected for
non-stream requests (stream clients execute tools client-side via MCP)
- memoryBuiltins: memory_save/update/search/delete builtin tools with
per-provider schemas + interception dispatch
- retrieval: fix toFts5MatchQuery import (ReferenceError on FTS5 path)
- tests: MCP auto-owner fallback, cross-principal isolation, stream guard
injectSkills() lists the in-memory skillRegistry, which is empty after a
cold start until something calls loadFromDatabase(). The interception path
already warms the cache (#2815); the injection path did not, so skills
were silently skipped (no_enabled_skills) for the first requests after
restart. Warm the cache for the chat owner before injection.
Registers Cheaper Inference (api.cheaperinference.com) as an OSS-sponsor gateway provider.
- Canonical provider `cheaperinference` (alias `cinf`) + routing registry with 39 measured text models
- Dedicated executor: forces `store:false` on the native /v1/responses endpoint (the shared strip in
chatCore.ts deletes `store` for every provider != openai, so without this every Responses request
400'd) and resolves chat-vs-responses URL from the per-model targetFormat
- 3 image models (grok-imagine, nano-banana-pro, nano-banana-2), prefix-only: the two nano-banana ids
already belong to adobe-firefly, which keeps the bare-id routing
- Resale pricing measured from GET /v1/models (30% off list); sponsor rail Kimi 1st / Cheaper
Inference 2nd via an explicit rank map; supporter badge in 43 locales; README row
No quota card: the gateway exposes no balance API (/v1/wallet and /v1/balance both 404).
Validated live end-to-end through OmniRoute: chat, native Responses, streaming and image generation
all 200 with real content; the Firefly collision guard verified at runtime.
- responseHeaders: promote NVIDIA NIM quota/usage headers (x-nvcf-*,
x-quota-*, x-ratelimit-*) to priority 2 so they survive the 768-byte
upstream header forwarding budget (were previously dropped as priority 3)
- new GET /api/v1/quotas: lists every provider connection with saturation
(0..1), remaining percent, and data source using existing saturation
signals — lets operators check key budgets via API instead of live monitoring
- chatCore: maxAttempts 1→2 for all providers (previously only model-scope and codex got retries)
- combo.ts: add x-omniroute-attempt response header showing how many attempts were made
- memory/settings: enable memory injection by default (maxTokens 1000) — auto-extraction already wired in chatCore
Claude→Gemini direct translators dropped thoughtSignature, so Gemini 3
tool follow-ups returned 400. Store on the response path, re-attach (or
context-fallback) on the request path, and thread signatureNamespace.
Closes#8979
Stop aliasing canonical `poe` to PoeWebExecutor so API-key requests hit
api.poe.com Chat Completions / Responses / Claude-only Messages instead of
the web GraphQL path that returned HTTP 405.
* feat(sse): deprecate the gemini-cli upstream provider with a real migration path
Stored `gemini-cli` connections were being kept alive for nothing. Measured before
touching anything:
routable? absent from PROVIDERS, from REGISTRY, from OAUTH_PROVIDERS, and no
executor references it → the connection can NEVER serve a request
refreshing? yes, and successfully — it redeemed against PROVIDERS.gemini's client
(681255809395-oo8ft2o…), the same public Gemini CLI / Code Assist OAuth
client
So the scheduler made periodic upstream calls to Google to keep a credential fresh
that had nowhere to go. That is the waste this removes.
This is a deprecation, not a deletion, and the difference is deliberate. The path was
not dead code: #8232 added it after a user report (the UI advertises automatic OAuth
rotation and these rows never rotated), and #8275 narrowed it to exactly the legacy
refresh. Simply dropping it from `supportsTokenRefresh` would have produced a SILENT
skip — `Skipping … (refresh unsupported)` — leaving the row at "active" forever, doing
nothing. Worse than before.
Instead:
DEPRECATED_PROVIDERS + isDeprecatedProvider/getDeprecationNotice in tokenRefresh
one place naming the provider and where to migrate. A test asserts the migration
target is itself routable, so the notice can never point somewhere useless.
_getAccessTokenInternal returns the ESTABLISHED unrecoverable envelope
{ error: "unrecoverable_refresh_error", code: "provider_deprecated", migrateTo }
Reusing `error` means isUnrecoverableRefreshError and the manual-refresh route
already stop retrying — no new contract for callers to learn. The distinct `code`
is what makes it legible. A bare `null` would read as transient and retry forever.
tokenHealthCheck marks the connection terminal with the reason
Placed after the existing terminal-status guard, which makes it idempotent for
free: once "expired", later sweeps skip the row, so it writes once instead of
rewriting the same reason every cycle.
the manual-refresh route stops lying
It said "Refresh token expired. Please re-authenticate this account." — false
here: the token is fine, the provider is gone. Re-authenticating would loop
against something that no longer exists. It now reports the deprecation and the
migration target.
`gemini` uses the same OAuth client, so re-adding the account there is a working path,
not advice to start over.
Deliberately NOT touched:
Category A — the gemini-cli CLIENT identity (#7034): clientIdentityProfiles.ts,
clientApi.ts, googApiKeyAuth.ts. Same string, opposite direction — requests
ARRIVING from the Gemini CLI, where OmniRoute is the server. Deleting these is the
failure this change must never cause, so a test now asserts the profile survives.
Audited: `git diff --name-only` touches none of those files.
errorClassifier.ts's isCloudCodeProvider list still names gemini-cli. It is a
defensive 403→PROJECT_ROUTE_ERROR list shared with cloudcode/cloud-code; the entry
is unreachable for a non-routable provider, and editing a shared classification
path for a dead string is risk without upside.
Tests — 42 across the six files that mention the identifier, all green:
gemini-cli-legacy-refresh.test.ts 5 (3 assertions REWRITTEN, see below)
gemini-cli-deprecation.test.ts 5 (new)
client-identity-profiles.test.ts 9 (category A, untouched)
service-token-refresh.test.ts 14
errorclassifier-antigravity-403.test.ts 4
gemini-cli-ansi-sanitization.test.ts 5 (category C, untouched)
The three rewritten assertions in the legacy file are alignment, not weakening, and the
gate is right to ask: each is now STRONGER. "refresh succeeds against Google's token
endpoint" became "zero upstream calls happen at all"; "a 400 surfaces invalid_grant"
became "the envelope is unchanged but the code says provider_deprecated" plus a control
asserting `gemini` still reports invalid_grant, proving the real path was not blunted.
The file's header keeps the whole #8232 → #8275 → deprecation arc, because each step is
why the next made sense. Count unchanged; no test deleted, so no allowlist entry needed.
* docs(changelog): fragment for #8980
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* chore(sse): drop the iflow entry from the token-refresh TTL map
The `iflow` provider was removed from the product, but its 24-hour refresh-lead
entry outlived it in REFRESH_LEAD_MS. Surfaced during v3.8.49 homologation on the
production VPS, where startup logs carry:
[CREDENTIALS] Warning: unknown provider "iflow" in credentials file, skipping.
Measured across src/, open-sse/, tests/ and docs/ — the identifier had exactly two
occurrences repo-wide: the map entry and one test assertion. Nothing dispatches on
it, so `getRefreshLeadMs("iflow")` now falls through to TOKEN_EXPIRY_BUFFER_MS like
any other unknown provider.
The test assertion was not deleted, it was MOVED: from "returns explicit lead time
for known providers" to "falls back to TOKEN_EXPIRY_BUFFER_MS for unknown
providers". That is alignment to the new behavior and strictly more coverage than
before — a silent reintroduction of the entry now turns the fallback case red
instead of passing unnoticed. Flagged explicitly because the test-masking gate
rightly treats a removed assertion as suspicious.
Also removes the now-redundant "Non-rotating providers" section header: every
remaining entry under it is Google-backed and the following comment already says
"permanent (non-rotating)".
node --import tsx/esm --test tests/unit/service-token-refresh.test.ts
# 14 pass, 0 fail
* docs(changelog): fragment for #8966
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
Claude models on Vertex AI were being sent through the generic OpenAI-
compatible partner endpoint, which 404s/errors for Claude on at least
some projects. Route them through Vertex's native Anthropic Messages
API (publishers/anthropic/.../rawPredict) instead, stripping the
body-level model field rawPredict rejects and injecting the required
anthropic_version field.
rawPredict only ever returns a complete JSON body, never real SSE
framing, so streaming requests now get a genuine Anthropic-format SSE
stream synthesized from that JSON (message_start/content_block_*/
message_delta/message_stop), which the existing claude-to-openai
response translator already knows how to parse.
Also fixes two response-format resolution bugs that silently dropped
a custom model's DB-stored targetFormat override whenever the model
id also existed in the static provider registry (as claude-sonnet-4-6
and claude-opus-4-7 do under vertex): resolveModelOrError had its own
ad-hoc resolution that never consulted the override, and even once
fixed, executeChatWithBreaker discarded the correctly-resolved format
before handleChatCore's own resolution ran a second time.
Cut from the v3.8.49 freeze tip (parallel-cycle model). Bumps package.json x3,
openapi.yaml, lockfile, adds the living [3.8.50] CHANGELOG section with the three
canonical headings (New Features / Bug Fixes / Maintenance) so aggregate-changelog.mjs
cannot mis-target an older published section, and syncs the 42 i18n mirrors.
* fix(adobe-firefly): default gpt-image detailLevel to maximal (5)
GPT Image 2 quality is dominated by generationSettings.detailLevel (1-5).
The SPA often defaults to 3 (medium); missing/auto quality previously mapped
to 3 as well. Default now to 5 (high/max) so API clients and Media without
an explicit quality still get maximal detail. Explicit low/medium still honored.
* chore(quality): rebaseline adobeFireflyClient + changelog fragment
adobeFireflyClient.ts 2317->2322 (+5) — this PR's own growth at the existing
payload-build site. Covered by tests/unit/adobe-firefly.test.ts.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* feat(sse): add alternateFormats registry field and resolver
* feat(sse): honor per-connection targetFormat in getTargetFormat
Registry-driven format lookup now resolves an alternate protocol
declared for the provider when the connection's providerSpecificData
carries a matching targetFormat, falling back to the entry's default
format otherwise.
* feat(sse): resolve base URL from selected alternate format
resolveBaseUrl now falls back to the connection's selected alternate
protocol (providerSpecificData.targetFormat) before the provider's
default base URL, while a manual providerSpecificData.baseUrl override
still wins over both.
* feat(sse): apply alternate format auth header and extra headers
DefaultExecutor's registry authHeader lookup and BaseExecutor's shared
header preamble now both honor a connection's selected alternate
protocol: the alternate's authHeader wins over the registry default,
and its extra headers (e.g. Anthropic-Version) are merged in.
* refactor(sse): extract resolveAlternate helper into BaseExecutor
Centralizes the getRegistryEntry() + resolveAlternateFormat() pair
that resolveBaseUrl, buildHeadersPreamble, and DefaultExecutor's
authHeader lookup each duplicated, so a future call-site can't diverge
from the shared precedence. Also translates the PT-BR comments added
in the previous three commits to match the surrounding English. Pure
refactor — no behavior change.
* feat(sse): declare Anthropic-compatible variant for xiaomi-mimo
The provider publishes the same catalog over /anthropic/v1/messages on the same
host. Selecting it also required bypassing the per-provider URL normalizers in
DefaultExecutor.buildUrl(): normalizeXiaomiMimoChatUrl() appends /chat/completions
unconditionally, which mangled the alternate's already-complete endpoint into
.../anthropic/v1/messages/chat/completions.
* feat(sse): add xiaomi-mimo-token-plan provider with monthly quota
Token Plan is a separate product: tp- keys authenticate only on the regional
token-plan-sgp host and return 401 on api.xiaomimimo.com, where the existing
xiaomi-mimo provider points. Same pattern as qwen-cloud-token-plan.
Registers the monthly token allowance (no balance API upstream) and declares
the Anthropic-compatible variant on the token-plan host.
* feat(dashboard): add API protocol selector to connection modal
Providers that declare alternateFormats in the registry now expose an opt-in
protocol dropdown on the connection modal. The choice persists to
providerSpecificData.targetFormat as an explicit null when set back to the
default, since the PUT route merges { ...existing, ...incoming } and an omitted
key would keep the previous override.
* fix(i18n,quality): vi parity for the protocol selector + own-growth rebaselines
The three new provider keys landed only in en/pt-BR, so the vi parity test failed
(tests/unit/i18n-vi-completeness.test.ts asserts key parity AND no __MISSING__
markers — running i18n:sync-ui would have satisfied the first and broken the
second). Added translated values instead. Scoped to vi: it and pt-BR are the only
locales with a parity test.
Rebaselines are this PR's own growth: EditConnectionModal.tsx 1283->1316 (the
selector field) and open-sse/executors/base.ts 1540->1562 (alternate-format
resolution at the existing buildUrl/headers chokepoint).
Adds the changelog fragment.
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* fix(token-refresh): discover projectId during token refresh
The token refresh path (tokenRefresh.ts) did not discover projectId
for antigravity/agy accounts. Dashboard and health check refresh use
this path, not the executor path.
Add ensureAntigravityProjectAssigned call after refreshGoogleToken
for antigravity/agy providers when projectId is empty.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* chore(quality): rebaseline the token-refresh test file + changelog fragment
tests/unit/token-refresh-service.test.ts 1311->1378 (+67) — the four cases
covering projectId discovery on the tokenRefresh.ts path. Growth is the tests
this PR adds, nothing else.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Signed-off-by: Minxi Hou <houminxi@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(executors): route current Claude generations through Vertex partner endpoint
PARTNER_MODELS pinned three Claude 3.x prefixes (claude-3-5-sonnet,
claude-3-opus, claude-3-haiku). Every newer Claude generation on Vertex
(claude-sonnet-4-6, claude-haiku-4-5, etc.) fell through to the
Google-publisher branch instead, producing an invalid
publishers/google/models/claude-... path.
Replace the pinned prefixes with a single generic "claude-" prefix:
any Claude model on Vertex is always an Anthropic partner model, never
a Google one, so this can't go stale again the way pinned version
strings did.
Fixes#1985
* docs: add changelog fragment for #8852
* fix(antigravity): discover projectId during token refresh
The initial OAuth exchange can fail to populate projectId via
loadCodeAssist (network timeout, account not yet onboarded). The
runtime transformRequest path already recovers via
ensureAntigravityProjectAssigned, but refreshCredentials did not --
after a token refresh the per-token memoization cache is invalidated
(new access token = new cache key), so every subsequent request
triggers a fresh loadCodeAssist round-trip that may fail again.
Add a best-effort ensureAntigravityProjectAssigned call in
refreshCredentials when projectId is empty, mirroring the pattern in
transformRequest. Persist the discovered id so it survives the next
refresh or restart.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* chore(quality): rebaseline antigravity.ts and test file-size
antigravity.ts grew from 1493 to 1528 lines (+35) with projectId
discovery in refreshCredentials. executor-antigravity.test.ts is a new
test file at 1098 lines (above cap 1000) with 4 new test cases.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* fix(antigravity): match ExecutorLog arity in the refresh discovery log
`ExecutorLog.info` (open-sse/executors/base.ts) is `(tag, message) => void` —
two parameters. The discovery log passed a third metadata object, which failed
typecheck:core with TS2554 on antigravity.ts:777. Bind the message to a local
and pass two arguments, matching the sibling warn on the catch branch. Kept to
two lines so the file stays at its frozen size; ran Prettier, which also wrapped
the pre-existing over-width `const msg` line below.
Also adds the changelog fragment for the fix.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Signed-off-by: Minxi Hou <houminxi@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(ghe-copilot): route OpenAI-native models via Responses API
- Add targetFormat: 'openai-responses' to gpt-5.4-mini, gpt-5.3-codex, gpt-5-mini,
mai-code-1-flash, and oswe-vscode-prime in ghe-copilot registry.
- Register gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna with openai-responses targetFormat.
- Update GheCopilotExecutor.buildUrl() to route openai-responses and codex models to
<gheUrl>/responses while keeping Claude and Gemini on /chat/completions.
- Add unit tests verifying targetFormat parity and buildUrl endpoint routing.
* docs(changelog): add fragment for #8835
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Alex <sefias_methue@hotmail.de>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(providers): correct Codex GPT-5.6 context window (#7702)
* docs(changelog): add fragment for #8838
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* test(providers): align the remaining GPT-5.6 context-window assertions
Three suites assert the pinned GPT_5_6_CODEX_CAPABILITIES contract through the
VS Code and provider-models routes, and still expected 372000. They only surface
in a full run, so the focused loop on this PR stayed green while `npm run test:unit`
failed with five `272000 !== 372000`.
The two conservative-merge cases in provider-models-route-codex keep testing what
they tested: live 999999 still exceeds the pinned value (pinned wins) and live
100000 is still below it (live wins). Only the pinned number and the comments
naming it move.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(reasoning): sanitize streamed K3 think tags
* refactor(stream): move think-tag helpers into thinkTagParser leaf module
open-sse/utils/stream.ts is frozen at 2887 lines with zero headroom, and the
kimi-coding-apikey think-tag handling added here pushed it to 2951. The parsing
half of that work has no SSE dependency, so it moves to thinkTagParser.ts (an
unfrozen leaf): the open-tag lookahead predicate and the end-of-stream flush
delta assembly. stream.ts keeps only the SSE envelope - enqueue, payload
collection, logging.
Getting back under the frozen ceiling needed more than just the PR's own
added lines, since even the fully self-gating helpers still cost a handful of
call-site lines stream.ts has zero room for. Along the way this also
deduplicates a synthetic chat-completion-chunk literal that was copy-pasted
three times in stream.ts (textual tool-call flush, think-tag flush, terminal
finish_reason synthesis) into one buildSyntheticChatChunk() in
streamHelpers.ts - pure DRY, no behavior change.
Behavior is unchanged: same tests, same counts.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: rinseaid <rinseaid@rinseaid.net>
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
