fix(forecast): ignore stale time-bounded runtime overlay reasons (#507) - #508

Summary

Fixes #507. forecast --live reported both accounts as unavailable (runtime skip: rate-limited) while Codex kept executing prompts successfully.

Root cause: the runtime overlay in runtime-observability.json persists a per-account skip reason on pool exhaustion and is only cleared by an explicit runtime reset, never on a subsequent successful request. Once the underlying window expired, the stale reason kept lib/forecast.ts marking the account unavailable. The proxy was unaffected because it rebuilds skip reasons fresh per request (getAccountRuntimeSkipReason -> isRateLimitedForFamily), which is why Codex still worked. doctor's forecast-runtime-alignment warning surfaced the same stale state through the shared evaluateForecastAccounts call.

Fix

In the overlay-application block of evaluateForecastAccount, validate time-bounded overlay reasons against the time-aware disk state before applying them:

• rate-limited is ignored when getRateLimitResetTimeForFamily(account, now, "codex") returns null (no active reset on disk, including model-scoped keys like codex:5h).
• cooling-down:... is ignored when coolingDownUntil is absent or <= now.

Each reason validates only against its own backing disk state, so the displayed reason string is never substituted with a misleading one. Non-time-bounded reasons (circuit-open, token-exhausted, policy-blocked) have no disk expiry and remain unconditional. The guard reuses the rateLimitResetAt value and the single now timestamp already computed in the function, so no extra disk read or clock skew is introduced.

Tests

• Split the parametrized runtime-skip test into non-time-bounded reasons (still applied unconditionally).
• Added: stale rate-limited overlay with expired disk entry -> account ready, reason dropped.
• Added: active rate-limited overlay (codex and codex:5h future reset) -> account unavailable.
• Added: stale cooling-down overlay (elapsed coolingDownUntil) -> reason dropped.
• Added: active cooling-down overlay -> account unavailable.

Verification

• npx tsc --noEmit: clean.
• npx vitest run test/forecast.test.ts: 31 passed.
• npx vitest run (full suite): 4365 passed, 6 skipped, 0 failed.

Notes / follow-up

• token-exhausted shares the same overlay-staleness class but has no disk-backed reset time to validate against, so it is intentionally left unconditional. Fully clearing it would require clearing the skip reason on a successful proxy request (CodeRabbit design-choice option 2); out of scope for this bug. Worth a tracking issue.
• Possible duplicate of [bug] Accounts look healthy but codex throws 503 #479 (noted by CodeRabbit issue enrichment); this PR addresses the [bug] forecast --live marks working accounts as unavailable #507 report directly.

note: greptile review for oc-chatgpt-multi-auth. cite files like lib/foo.ts:123. confirm regression tests + windows concurrency/token redaction coverage.

Greptile Summary

fixes stale time-bounded runtime overlay reasons persisting in runtime-observability.json long after the underlying rate-limit window or cooldown expires, causing forecast --live to incorrectly report working accounts as unavailable.

• lib/forecast.ts: adds isStaleOverlayReason computed from the already-available rateLimitResetAt and coolingDownActive values; ignores rate-limited overlay when getRateLimitResetTimeForFamily returns null, and cooling-down:… overlay when coolingDownUntil is absent or elapsed. non-time-bounded reasons (circuit-open, token-exhausted, policy-blocked) are applied unconditionally.
• test/forecast.test.ts: nine new vitest cases cover expired and absent disk entries, both overlay-source precedence paths (accountSkipReasons vs lastPoolExhaustionSkipReasons), active boundaries, and model-scoped rate-limit keys (codex:5h); the it.each is narrowed to the three genuinely unconditional reasons.

Confidence Score: 5/5

the staleness guard is tightly scoped, reuses already-computed values, and all nine new tests pass; no functional regressions introduced.

the fix reads no extra disk state and introduces no new async paths; the ternary correctly short-circuits for null overlayReason and the policyBlockedIndexes branch is unaffected. test coverage spans both staleness sources, both time-bounded reason types, active/expired/absent disk state, and model-scoped keys.

no files require special attention.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[evaluateForecastAccount] --> B[resolve overlayReason\naccountSkipReasons ?? lastPoolExhaustionSkipReasons]
    B --> C{overlayReason type?}
    C -->|rate-limited| D[rateLimitResetAt === null?]
    C -->|cooling-down:...| E[coolingDownUntil absent or <= now?]
    C -->|other non-time-bounded| F[isStaleOverlayReason = false]
    D -->|yes - stale| G[drop reason, account stays ready]
    D -->|no - active| H[apply: availability=unavailable]
    E -->|yes - stale| G
    E -->|no - active| H
    F --> H
    H --> I[reasons.push: runtime skip: ...]

_{Reviews (3): Last reviewed commit: "test(forecast): cover accountSkipReasons..." | Re-trigger Greptile}