The Docker image bakes NODE_OPTIONS=--max-old-space-size=256, and the standalone
launcher (scripts/dev/run-standalone.mjs, the Docker CMD) spawned 'node
server.js' without overriding it — so the server inherited the 256 MB cap and
OOMed randomly under load or with a large SQLite DB. `omniroute serve` already
honored OMNIROUTE_MEMORY_MB but the Docker path did not.
Add a shared resolveMaxOldSpaceMb() helper (OMNIROUTE_MEMORY_MB, default 512,
clamped [64,16384]) and have the launcher append --max-old-space-size to the
child NODE_OPTIONS (a trailing flag wins, overriding the baked 256 without
clobbering other flags). Update the .env.example doc to reflect the 512 default.
(A) For a Codex-only account, a bare gpt-5.5 Responses request was rerouted to
codex with the model hardcoded to gpt-5.5-medium (chatHelpers.ts). The Codex
executor reads a model-name suffix as an explicit modelEffort that, per #2331,
overrides the client's reasoning.effort — so a genuine reasoning.effort=xhigh
was silently demoted to medium. Keep the bare gpt-5.5 id (the connection
fallback still supplies the default effort); the executor precedence is
untouched, so #2331 stays intact.
(B) gpt-5.5-xhigh/-high/-low misrouted to the openai provider (only bare gpt-5.5
was codex-preferred), so codex-only users got 'No credentials for provider:
openai'. Add the suffixed variants to CODEX_PREFERRED_UNPREFIXED_MODELS so they
infer codex before the /^gpt-/ → openai fallback.
Closes#2877
handleChatCore declared `const settings = cachedSettings ?? await
getCachedSettings()` twice in the same function scope — the consolidated
"fetch once, reuse" const near the top, plus a duplicate added alongside the
per-key stream-default-mode feature. esbuild/tsx rejected the same-scope
redeclaration ("The symbol 'settings' has already been declared"), which made
every unit test that imports chatCore fail to transform and broke the
production build.
Remove the duplicate; reuse the earlier consolidated `settings`. Add an
import-guard regression test.
GET /api/quota/pools/[id]/usage returns { usage: snapshot }, but usePoolUsage
stored the whole wrapper, leaving usage.dimensions undefined. StackedAllocationBar
then dereferenced usage.dimensions[dimensionIndex] (undefined[0]) and crashed the
entire quota-share page for any pool that has allocations. Unwrap data.usage in the
hook + defensively guard usage.dimensions?.[i] / dim.perKey ?? [] in the bar.
- duckduckgo-web.ts: Add poolConfig (2-5 sessions, 1s cooldown)
- providerRegistry.ts: Add poolConfig field to RegistryEntry type, configure for llm7
- default.ts: Read poolConfig from registry in constructor
DuckDuckGo Web now uses pool for VQD token rotation.
LLM7.io now uses pool with 2s cooldown for its 1 req/s rate limit.
Move pool integration from Pollinations-specific code into BaseExecutor
so any executor can opt in via a protected poolConfig property.
Changes:
- base.ts: Add poolConfig, getPool(), buildPoolHeaders() to BaseExecutor
- pollinations.ts: Remove static pool, use base class pattern
- session-pool-modular.test.ts: 36 tests for provider-agnostic pool behavior
Closes#2953
Codex can emit a function_call with an empty/missing call_id; the matching
function_call_output then becomes a role:'tool' message with no preceding
tool_call, which the upstream rejects: "Messages with role 'tool' must be a
response to a preceding message with 'tool_calls'". The orphan filter let it
slip through because its guard (`&& rec.tool_call_id`) short-circuited on the
empty id.
Two changes in openaiResponsesToOpenAIRequest:
- Skip function_call items with an empty call_id (mirrors the empty-name skip),
so no dangling assistant tool_call with an unmatched id is emitted.
- Drop ANY role:'tool' message whose tool_call_id (including empty/missing) has
no matching tool_call, instead of only those with a truthy id.
Closes#2893
opencode-zen serves the public, signup-free OpenCode Zen endpoint
(https://opencode.ai/zen/v1). With no API-key connection configured,
getProviderCredentials returned null, so the Playground/combos surfaced
"No credentials for provider: opencode-zen" when selecting an OpenCode free
model.
Fall back to synthetic no-auth credentials for opencode-zen when no usable
connection exists (a configured active key is still selected first; a
rate-limited/terminal key returns its own signal before this point). Other
api-key providers are unaffected.
Closes#2962
Fireworks Fire Pass (fpk_*) keys return 403 "Fire Pass API keys are not
authorized for this route." on /models while still serving chat. Two paths
wrongly penalized the key:
- validateOpenAILikeProvider returned "Invalid API key" for any 403 on the
models endpoint without trying the chat probe.
- checkFallbackError classified the 403 as AUTH_ERROR (retryable cooldown), and
even a generic fall-through hit the transient-cooldown default — marking the
connection unavailable.
Fix: validateOpenAILikeProvider now inspects the 403 body and falls through to
the chat probe for "not authorized for this route" responses (401 and generic
403 still fail fast). checkFallbackError short-circuits such route-restriction
403s to { shouldFallback: false, cooldownMs: 0 } so the connection is not cooled
down. Per CLAUDE.md, a generic api-key 403 should be recoverable unless terminal.
Closes#2929
The no-auth OpenCode provider has id "opencode" and alias "oc". The combo
builder built qualifiedModel from the provider id ("opencode/big-pickle"), but
parseModel("opencode/...") resolves to the opencode-zen api-key tier via a
manual ALIAS_TO_PROVIDER_ID override — not the no-auth "opencode" provider.
"oc/<model>" resolves correctly.
Rewrite qualifiedModel to the provider alias for no-auth providers (only when
the alias differs from the id), keeping providerId for getModelIsHidden. Aligned
the route test that previously asserted the buggy "opencode/" prefix.
Closes#2901
Adversarial review of the peer-IP fix surfaced: (1) CRITICAL — cliTokenAuth.ts
still derived loopback from new URL(request.url).hostname (the same spoofable
Host class), letting a remote caller with a stolen CLI token reach management
APIs via Host: 127.0.0.1; now it trusts the middleware-stamped locality verdict
(AUTHZ_HEADER_PEER_LOCALITY, a client-stripped trusted header). (2) HIGH —
isLoopbackHost mangled bare IPv6 (::1, ::ffff:127.0.0.1) via split(":")[0],
a fail-closed DoS on IPv6 deploys. (3) HIGH — the Docker entrypoint ran bare
server.js (no peer stamp); run-standalone.mjs now prefers server-ws.mjs.
Codex Desktop injects an image_generation hosted tool into every Responses API
request (even text-only ones). The tool-type validator threw
unsupportedFeature() (400) for it, breaking every Codex Desktop request.
Mirror the tool_search handling: add IMAGE_GENERATION_TOOL_TYPES, allow it past
the validator guard, and drop it from the tools array before forwarding to Chat
Completions.
Closes#2950
The github provider has format:"openai" (baseUrl .../chat/completions) plus a
separate responsesBaseUrl (.../responses); a model only uses the Responses API
when it sets targetFormat:"openai-responses". GitHub Copilot's Responses API
does not serve Claude/Gemini models, so claude-opus-4.7, claude-opus-4-5-20251101,
gemini-3.1-pro-preview and gemini-3-flash-preview failed with [400]. The working
claude-opus-4.6 carries no targetFormat and uses chat/completions.
Drop targetFormat:"openai-responses" from those four Claude/Gemini entries so
they use the provider default. Native OpenAI gpt-* models (and oswe-vscode-prime)
keep the Responses API.
Closes#2911
The middleware runtime exposes no socket, so a prior fix derived LOCAL_ONLY
locality from the Host header — letting a remote caller send Host: 127.0.0.1
and reach spawn-capable routes (RCE class). The custom Node servers now stamp
the real socket.remoteAddress into a token-signed internal header; the policy
trusts only a stamp whose token matches this process's secret, and fails closed
otherwise. Preserves the owner-authorized loopback + private-LAN access without
trusting any client-controlled header.
big-pickle's OpenCode/Zen upstream runs DeepSeek thinking mode, but the model
id reveals no DeepSeek signal, so requiresReasoningReplay (called with
allowLegacyFallback:false) never triggered. Follow-up/tool-use turns failed
with [400] 'The reasoning_content in the thinking mode must be passed back to
the API'. Note: requiresReasoningReplay does not consume supportsReasoning, so
the registry flag alone would not have fixed it.
Add RegistryModel.interleavedField (mirrors models.dev interleaved_field),
declare interleavedField:'reasoning_content' (+ supportsReasoning:true) on
big-pickle in both opencode and opencode-zen, and surface the registry value
in getResolvedModelCapabilities so requiresReasoningReplay returns true.
Closes#2900
077_api_key_stream_default_mode.sql and 077_quota_pools.sql both claimed
prefix 077, so getMigrationFiles() threw a version-collision error and
getDbInstance() failed at every startup (app would not boot; all DB-touching
unit tests were red on release/v3.8.8).
Renumber the dependency-free, idempotent quota_pools migration 077 -> 085
(no other migration references quota_pools/quota_allocations), keep the
non-idempotent api_key_stream_default_mode ALTER at 077, add a retroactive
isSchemaAlreadyApplied guard (case 085) for DBs that already applied it under
077, and add a regression test enforcing unique migration prefixes.
1. check-permissions.sh: swap NODE_OPTIONS flag order so runtime
OMNIROUTE_MEMORY_MB override wins (last flag takes precedence)
2. comboMetrics.ts: evictOldestMetric(targetMap) now accepts the
target map as parameter; cleanup timer iterates BOTH metrics
AND shadowMetrics; null lastUsedAt falls back to Date.now()
instead of epoch 0 (prevents premature eviction of intent-only
entries)
3. providerRegistry.ts: remove Proxy wrapper — adds CPU/complexity
overhead with zero memory savings since _REGISTRY_EAGER is
already fully allocated and generator functions iterate all
entries at startup. Simple re-export instead.
4. usage.ts: use per-cache TTL constants in cleanup timer
(ANTIGRAVITY_MODELS_CACHE_TTL_MS=1min, others=5min) instead of
single 10min SUB_CACHE_TTL_MS for all caches.
5. Add 18 unit tests for comboMetrics memory management covering:
basic CRUD, shadow metrics, intent tracking, eviction at capacity,
strategy tracking, and reset operations.
- PoolCreateSchema: add optional connectionIds[] + .refine() that enforces primary membership
- PoolWizard: replace single-select dropdown with checkbox multi-select; first checked = primary (badge); step-2 adds helper note for additional connections; step-3 preview grouped by provider with +N more; POST body sends both connectionId and connectionIds
- PoolCard: optional providers[] prop renders a row of ProviderIcon (up to 3 + badge) instead of a single icon when pool has multiple connections
- i18n: 4 new keys added to both en.json and pt-BR.json (wizardConnectionsLabel, wizardPrimaryBadge, wizardAdditionalConnectionsNote, wizardPreviewMoreModels) — parity maintained (23 wizard keys each)
- Tests: quota-pool-wizard-multi.test.ts (21 tests) covering schema accept/reject, structural wizard assertions, and i18n parity
- quotaKey.ts (resolveQuotaKeyScope): iterate pool.connectionIds (fall back to
[connectionId] for un-backfilled rows); each connection contributes its own
connId + provider to the scope. poolSlugs logic unchanged (one slug per pool).
- quotaCombos.ts (syncQuotaCombos): replace single-connection resolvePoolProvider
with resolvePoolForSync that returns all connectionIds; iterate each connId to
build the desiredNames union across all providers; upsert combos pinned to the
CORRECT per-connection connId; prune against the full union so only truly stale
combos (no longer produced by any current connection) are deleted.
- enforce.ts (enforceQuotaShare + recordConsumption): both pool-matching loops
changed from equality (p.connectionId === input.connectionId) to membership
(p.connectionIds.includes) with fallback for un-backfilled rows. Fail-open
(B16) and pool-level dimension key semantics are preserved unchanged.
- quotaPools.ts: no logic change needed — connectionIds already flows through
getPool (D1); syncQuotaCombosGuarded passes poolId and syncQuotaCombos
resolves the full QuotaPool internally.
- tests/unit/quota-multiprovider.test.ts: 6 new tests covering D2 (scope,
enforce primary/secondary membership, combos 2-provider create + prune).
All 22 tests pass (14 new + 8 existing enforce + pool-connections suites).
Full SOLO remote-agent integration against solo.trae.ai's reverse-engineered
API (upgrades the previous import_token stub).
- open-sse/executors/trae.ts: streaming executor for the solo_agent_remote API
(POST /chat_sessions + GET /events SSE -> OpenAI chat.completions), accumulating
plan_item thoughts and mapping token_usage. Headless Cloud-IDE-JWT refresh via
the ExchangeToken endpoint.
- Session mode via model id: `trae/work` runs the fast work-mode auto agent;
`trae/auto` and named models (gpt-5.4, kimi-k2.5, gemini-3.1-pro, ...) run in
code mode.
- Provider registry entry (models + 272k context) and executor wiring.
- Two credential paths: browser /authorize loopback callback (captures the JWT +
long-lived refresh token) and manual Cloud-IDE-JWT paste via
POST /api/oauth/trae/import (Zod-validated).
- TraeAuthModal dashboard UI wired into the provider detail page.
- mapTokens/providerSpecificData carry the SOLO common_params identity fields.
- Tests: executor (stream/non-stream/error/refresh/work-mode + callback parser)
and updated oauth-trae provider tests.
Adds migration 086 to create the `quota_pool_connections` join table with a backfill
that seeds every existing pool's single connection_id as its first member. Updates
QuotaPool type with `connectionIds: string[]`, wires createPool/updatePool/deletePool
to maintain the join table transactionally, and keeps `connection_id` as the primary
back-compat column synced to `connectionIds[0]`.
