setup-codex now generates Codex profiles for compatible text models from the live /v1/models catalog when the model id doesn't match a hand-tuned pattern, skipping media/embedding models. Integrated into release/v3.8.43.
Re-applies two dropped gemini-code-assist hardening fixes (defaultClaudeToolType non-object passthrough; mergeConsecutiveSameRoleContents shallow-copy) with regression tests. Follow-up to #5661/#5662. Integrated into release/v3.8.43.
BLOCO E3 of the god-files campaign. db/settings.ts (1154 LOC) mixed five concerns; the
three cleanly-separable ones plus the shared toRecord/JsonRecord helper were lifted out
verbatim into a new src/lib/db/settings/ subdirectory, leaving the Settings-core + Proxy
config concerns in the host (host now 646 LOC). The host re-exports every moved public
symbol so the module's public API (consumed by ~93 test files + localDb) is unchanged.
- settings/shared.ts — toRecord + JsonRecord (9 LOC)
- settings/pricing.ts — pricing layers/sources/per-model + update/reset (254 LOC)
- settings/lkgp.ts — Last-Known-Good-Provider get/set/clear (49 LOC)
- settings/cacheMetrics.ts — cache metrics + trend (235 LOC)
Settings-core + the Proxy-config concern stay in the host: proxy is the most tangled
(245-line resolveProxyForConnection, resolution cache, imports from ./proxies) and
getSettings is the most central function — leaving them is the correct coupled-core stop.
Pricing/LKGP/Cache have NO dependency on Settings/Proxy helpers (verified); the
dependency DAG is acyclic (check:cycles).
Adds tests/unit/db-settings-split.test.ts: characterization of the shared toRecord helper
+ a guard asserting the host preserves its full public export surface.
Validated: typecheck:core, check:cycles (no cycles), 149 existing+new db/settings consumer
tests green (db-settings-crud/extended, 8 pricing suites, cache-metrics, 2 proxy-resolution
suites + 29 new), ESLint, Prettier, check:file-size (host 646 < frozen 1155).
BLOCO E3 of the god-files campaign. db/models.ts (1250 LOC) mixed six concerns; the
three cleanly-separable ones plus the shared key_value helpers were lifted out verbatim
into a new src/lib/db/models/ subdirectory, leaving the tightly-coupled custom/synced/
flags trio in the host (host now 936 LOC). The host re-exports every moved public symbol
so the module's public API (consumed by ~29 test files + localDb) is unchanged.
- models/shared.ts — asRecord / toNonEmptyString / getKeyValue + JsonRecord (19 LOC)
- models/compat.ts — model-compat overrides + sanitizeUpstreamHeadersMap (249 LOC)
- models/aliases.ts — model-alias CRUD + cascade delete (61 LOC)
- models/mitmAlias.ts — MITM alias get/set (32 LOC)
The custom/synced/flags trio stays in the host because it is genuinely coupled
(flags->getCustomModelRow, flags->readCompatList, custom->removeModelCompatOverride,
synced->getModelIsDeleted, setModelIsHidden->updateCustomModel) — splitting it cleanly
is a follow-up. Dependency DAG is acyclic (verified by check:cycles).
Adds tests/unit/db-models-split.test.ts: characterization of the pure extracted helpers
+ a guard asserting the host preserves its full public export surface.
Validated: typecheck:core, check:cycles (no cycles), 77 existing db/models consumer
tests (db-models-crud/extended/aliases-cascade + 7 more) green, 7 new tests, ESLint,
Prettier, check:file-size (host 936 < frozen 1259; no new violations).
Adds an opt-in, default-off LLM-tier compression engine ('llm') that condenses non-system message prose via a pluggable chat-completion backend, mirroring the llmlingua contract. Safe by construction: no-op default backend (pass-through out of the box), not in the default stacked pipeline, enabled defaults false, fenced code blocks + system messages never sent to the model, fail-open everywhere, minTokens floor. Real production backend is a VPS-validated follow-up (Hard Rule #18). Regression guard: tests/unit/compression/llm-compressor-engine.test.ts (8). gaps v3.8.42 — T05/C3.
BLOCO E2 of the god-files campaign. The module-level pure/standalone helpers in
src/app/api/v1/models/catalog.ts (1611 LOC) were lifted out verbatim into five
cohesive leaf modules so the catalog host shrinks toward the 800-LOC file-size cap
without any behavior change (host now 1345 LOC; the heavy getUnifiedModelsResponse
orchestrator is untouched — its in-function closures stay put):
- catalogHelpers.ts — pure numeric/array/shape helpers + shared catalog types
- catalogOpenrouter.ts — OpenRouter id/modality/free-model/display-name helpers
- catalogVision.ts — vision-capability field derivation (+ isVisionModelId re-export)
- catalogProviderMaps.ts — alias<->providerId resolution maps (buildAliasMaps)
- catalogRequest.ts — /v1/models API-key auth gating + Codex CLI client detection
The host re-exports getCustomVisionCapabilityFields and isVisionModelId so the public
API consumed by other tests (llm-selector-custom-vision-models, vision-detection-
consistency) is unchanged; all 9 catalog/vision suites stay green.
Adds tests/unit/catalog-helpers-extraction.test.ts: characterization tests for every
extracted helper + a guard asserting the host preserves its public exports.
Validated: typecheck:core, 50 catalog characterization tests, 12 new leaf tests,
integration-wiring, check:cycles, check:file-size (no new violations), ESLint, Prettier.
Integrated into release/v3.8.43. Validated bun==node output for all 3 gates (provider-consistency, compression-budget, known-symbols). Thanks @KooshaPari!
Integrated into release/v3.8.43. Includes the legacy-boolean back-compat derivation so existing preserveSystemPrompt=false installs keep whenNoCache behavior.
Closes#5483. OpenAI-compatible providers receiving a Responses-shaped request carrying MCP / tool_search tools now route to the upstream /responses endpoint instead of downgrading to /chat/completions, preserving Codex deferred tool discovery. Detection helpers extracted to open-sse/executors/forceResponsesUpstream.ts. Thanks to @KooshaPari.
Behavior-preserving decomposition: src/shared/constants/sidebarVisibility.ts
1197 -> 291 LOC by extracting two leaves under sidebarVisibility/:
- types.ts (160): HIDEABLE_SIDEBAR_ITEM_IDS + all sidebar types (self-contained).
- sections.ts (762): section building-block consts + SIDEBAR_SECTIONS (imports
types only — cycle-safe). COMPRESSION_CONTEXT_GROUP + SIDEBAR_SECTIONS stay
exported; host re-exports both + 'export *' of types, so every consumer import
path is unchanged.
Byte-identical data verified via JSON.stringify of HIDEABLE_SIDEBAR_ITEM_IDS /
SIDEBAR_ICON_ACCENTS / COMPRESSION_CONTEXT_GROUP / SIDEBAR_SECTIONS / SIDEBAR_PRESETS
+ getSectionItems output (identical before/after). typecheck:core, check:cycles
(no cycles), check:file-size (3 files <800), and the 3 sidebar suites (20/20) pass.
No logic changed.
Note: file-size frozen baseline for sidebarVisibility.ts (1198) can ratchet to 291
to lock the shrink (left for the release ratchet / operator).
Root cause (NOT a timeout): the nightly-mutation run fails on cold-cache nights
because the blocking mutation-ratchet job measures modules below baseline, while
warm-cache nights pass — the verdict tracked GitHub Actions cache state, not code
quality. Proven via a local Stryker probe on headers.ts: covering unit tests
(no-memory-header, strip-reasoning) had drifted OUT of stryker.conf.json
tap.testFiles, so their mutants went covered-but-unkilled = Survived on a cold
full run (COVERED score 61.73 vs 94.29 baseline); adding them restores the kills.
- Add scripts/check/check-mutation-test-coverage.mjs: guards that every UNIT test
importing a Stryker-mutated module is listed in tap.testFiles. Advisory by
default, --strict in CI (wired in quality.yml fast-gates). Prevents recurrence.
- Add the 38 drifted covering unit tests to stryker.conf.json tap.testFiles
(138 -> 176). Monotonically safe: more covering tests only raise/hold the score.
- Add MUTATION_RATCHET_EPS (1.0pt) anti-flake tolerance to check-mutation-ratchet
so sub-point tap-runner jitter no longer false-fails the gate. Lowers no baseline.
- Tests: check-mutation-test-coverage (3) + eps cases in check-mutation-ratchet.
Residual: a clean post-merge nightly confirms scores return to/above baseline;
any marginal residual gets a baseline re-seed (operator).
Integrated into release/v3.8.43. HuggingChat web session-routing fix (root parent-message fetch + cookie propagation + encrypted-credential guard) + 24-model catalog refresh. Maintainer adjustments (co-authored): reverted the freeModelCatalog.data.ts whole-file reformat down to the surgical 24-record huggingchat change (preserving the auto-generated compact format), and added a 502 regression test for the null parent-message-id path (Rule #18).
Integrated into release/v3.8.43. Windows path-normalization fix for the route-guard membership gate + regression test (Rule #18). Co-authored test added by maintainer.
## Problem
When Antigravity returns HTTP 429 with `rate_limit_exceeded` error code,
the model lockout system never records the failure, so the model is not
cooled down despite being rate-limited.
### Root Cause
Antigravity's 429 error text is: `"Resource has been exhausted (e.g. check
quota)."`
The QUOTA_PATTERNS in `classify429.ts` contained overly broad regexes:
- `/resource.*exhaust/i` — matches "Resource has been exhausted"
- `/check.*quota/i` — matches "check quota"
This caused `classifyErrorText()` to return `QUOTA_EXHAUSTED` (wrong),
which set `providerExhausted = true` in the combo target exhaustion logic.
With `providerExhausted`, the retry path was skipped entirely, and while
the "done retrying" path should still record lockout, the misclassification
cascaded into incorrect provider-level exhaustion state.
Additionally, `targetExhaustion.ts` used the raw error text string instead
of the structured error code (`rate_limit_exceeded`) that was already
parsed from the response body.
## Fix
1. **classify429.ts** — Removed overly broad `/resource.*exhaust/i` and
`/check.*quota/i` from QUOTA_PATTERNS. Antigravity's rate-limit wording
is not a true quota exhaustion signal.
2. **targetExhaustion.ts** — Added optional `structuredError` to
`ApplyComboTargetExhaustionOptions`. When available, the structured
error code (e.g. `rate_limit_exceeded`) takes precedence over raw error
text for exhaustion classification.
3. **combo.ts** — Passes `structuredError` to both `applyComboTargetExhaustion`
call sites (dispatch path + retry-or-rotate path).
## Effect
`structuredError.code = "rate_limit_exceeded"` → classified as rate-limit
(not quota) → `providerExhausted = false` → retry proceeds →
`recordModelLockoutFailure` called → model enters lockout with proper
cooldown (120s base, exponential backoff).
## Tests
Added 2 new tests for `structuredError.code` precedence in exhaustion
classification. All 28 related tests pass.
