[stanleytejakusuma]

Summary

Add an opt-out for the shared combo cooldown wait without changing the global policy or existing defaults.

A two-target priority fallback chain can currently dispatch A, then B, wait out a short cooldown, and dispatch A again even when maxSetRetries is zero. Operators who want the error after both targets fail cannot disable that shared wait for just one combo.

Changes

• Add optional boolean config.comboCooldownWaitEnabled; only literal false disables the shared post-dispatch cooldown and pre-dispatch circuit-open/model-lock waits.
• Pass the resolved combo config to both wait eligibility and default target-timeout calculation, removing the wait-budget timeout-floor extension for opted-out combos. Explicit targetTimeoutMs still takes precedence.
• Preserve the existing global behavior when the key is absent or true. The key cannot enable waiting when the global policy is disabled.
• Document the setting and the unchanged, separate native-Codex turn-pin wait.

Example configuration:

{
  "config": {
    "comboCooldownWaitEnabled": false
  }
}

Validation

Original implementation evidence, recorded before the maintainer base merge:

• TDD: the focused test command failed on the pinned unmodified base with 6 failures (52/58 passing), then passed after the fix (58/58).
• Focused command: node --import tsx/esm --test --test-reporter=tap --test-concurrency=1 tests/unit/combo-config.test.ts tests/unit/combo-cooldown-wait-opt-out.test.ts.
• Additional targeted rerun: 82/82 across the two focused files plus combo-quota-share-cooldown-wait, combo-loop-safety-timer-cleanup, and combo-set-retry-durable-exhaustion.
• The complete implementation verification recorded 222 distinct targeted/adjacent tests passing, core typecheck, configured ESLint, and Prettier.
• Behavioral coverage asserts A/B-only dispatch with no cooldown sleep, the unchanged A/B/A positive control, pre-dispatch model-lock/circuit-open paths, strict boolean validation, config cascade, timeout precedence, and SQLite create/update persistence.

Current CI and inherited timing sensitivity

⚠️ base-red inherited: #18791 (the current open release-branch tracker; #17219 is closed).

The failed CI run checked out synthetic merge cc5be8554ffdae247cc925265758e7bf1d552655, whose exact base parent is 38b0fbb1ed4a58f7c23d7a91cbc8a5b3ae471689 and topic parent is 36cd9cfe3a389b4c58de119df10ed97c0c8c9f0c.

• Unit Tests (6/8) had one failure: tests/unit/codex-usage-status-throttle-15172.test.ts:81, gap 213ms between fetch 0 and 1 is below one interval. Gate / CI failed downstream. The other unit shards and Vitest passed; downstream coverage/quality work was skipped, not passed.
• The original timing test, throttle, usage dispatcher/Codex fetcher, connection-proxy resolver, DB settings, Auto-Ping module and dependency lockfile are unchanged between that exact base and merge.
• Matched probes on Node 24.22.2: both refs pass the original two-test file with an immediate synthetic proxy lookup; adding a 120ms first-connection lookup delay through the existing test seam makes BOTH fail at the same assertion with the identical gap 130ms message. No production/test source or assertion was changed for the probes. The historical CI latency source remains uninstrumented: this demonstrates pre-existing timing sensitivity, not a definitive diagnosis of that historical run.
• Focused recheck of this PR on the exact Actions merge: 78/78 passing across the five combo configuration/cooldown/retry test files on supported Node 26.10.0. This is focused local evidence, not a full CI claim.
• Exact-base evidence recorded on the base-red tracker. No unrelated throttle repair, tolerance relaxation or CI-policy change is included here. A new CI cycle is being requested by this status update; a green rerun does not resolve the underlying timing sensitivity.

Scope

The cooldown opt-out implementation, regression tests, resilience documentation, mutation registration and changelog fragment. The maintainer's base merge and changelog contribution are retained. No provider-specific defaults, migrations, dependency changes, live routing changes or deployment. The separate native-Codex turn-pin wait remains unchanged.

[stanleytejakusuma]

Published base sync 0d5e20023665d28274a0e6f9b38139114f6d3fd6 by normal non-force push. The prior contributor/maintainer commits are preserved; the feature delta is still the same8files (+291/-6), with no feature changes to CI, CHANGELOG.md, the quota throttle or timing assertions.

The candidate CHANGELOG is byte-identical to base24ba770e4a, and the pinned-base integrity guard passes. Independent immutable-candidate verification:82/82 focused combo tests and242/242 cooldown-expansion tests, plus changed-file lint/format, on supported local Node26.10.0. These are local checks, not a CI pass.

Fresh checks on the published head:

• CI: https://github.com/cheaper-inference/OmniRoute/actions/runs/41605732377
• Quality Gates: https://github.com/cheaper-inference/OmniRoute/actions/runs/41605732336

The earlier unchanged Codex timing sensitivity, moving-base changelog comparison and Coverage runner ENOSPC are documented under base-red#18791. Coverage capacity is not claimed fixed by this sync; it remains to be verified in the fresh run.

[stanleytejakusuma]

CI verified green on the published head 0d5e20023665d28274a0e6f9b38139114f6d3fd6:31successful checks,10workflow-defined skips, no failures or pending checks. GitHub reports CLEAN/MERGEABLE.

• CI: https://github.com/cheaper-inference/OmniRoute/actions/runs/41605732377 — all8unit shards, Vitest, Build, Coverage, Quality Ratchet and Gate / CI pass.
• Quality Gates: https://github.com/cheaper-inference/OmniRoute/actions/runs/41605732336 — changelog integrity, Fast Quality and Gate / Quality pass.
• API Route Typecheck and semgrep also pass on the same head.

The only new commit is the normal base sync; the8-file feature delta remains unchanged. Earlier timing sensitivity and hosted-runner ENOSPC evidence stay recorded under#18791; this successful run is not claimed as a repair of those underlying conditions. No timing assertion, CI policy or coverage threshold was weakened. No merge or deployment.