Repository navigation
opencode-route-proxy: known-100% seating gate — exclude legs whose binding quota window is known exhausted until reset (#4370c) #5295
Description
opencode-route-proxy: known-100% seating gate — exclude legs whose binding quota window is known exhausted until reset (#4370 acceptance (c))
Implementation slice scoped by triage task 8950 (lane W, 2026-11-04) extending #3325 leg-shedding to the #4370 death classes. Findings and full evidence are on #3325 (task 8950 comment) and #4370.
Problem (verified at main, blob 70f01618, bin/opencode-route-proxy.mjs, 1095 lines)
- Seating has no known-100% gate.
demotedPairs(lines ~330–345) learns only from the attempt ledger (≥3 quota failures in 4 h), andapplyDemotion(lines ~662–670) reorders demoted pairs to the back — it never excludes them. - A leg whose binding window is already known to be at 100% with a future reset (cline-pass monthly, fleet session deaths rotated legs: provider.transport timeouts on the gpt-6-luna leg and cline-pass leg wall/rate-limit deaths are not tracked by #3325 #4370: 11 sessions killed 2026-11-01) is a normal first-choice candidate until it has failed three times.
Scope
- Seating-time gate in
bin/opencode-route-proxy.mjs: a (provider, model) pair whose binding quota window is known exhausted (ledger quota detail carrying a future reset, or an equivalent known-100% record the proxy already stores) is EXCLUDED from candidates until that reset — the credential-less exclusion pattern (opencode-route-proxy: opendesign leaves sit in pools but the provider's credential is inactive — walks burn attempts and surface a misleading auth error #3228). - Fail-safe: if exclusion would empty the route, keep current demoted order rather than hard-failing; record the gate decision in existing diagnostics.
- Tests with the pinned-now clock (
OPENCODE_ROUTE_PROXY_NOW_MS): known-100% pair excluded while a live pair exists; pair returns after reset; all-exhausted route still walks in demoted order.
Out of scope: external quota-service reads at request time, demotion threshold changes, transport timeouts (companion slice), in-stream shedding (task 8932).
Acceptance criteria
- Sessions are not seated on a known-100% leg while any non-exhausted candidate exists (fleet session deaths rotated legs: provider.transport timeouts on the gpt-6-luna leg and cline-pass leg wall/rate-limit deaths are not tracked by #3325 #4370 acceptance (c)).
- Gate is per (provider, model) and lifts automatically at the recorded reset.
- Verified by proxy tests plus a route-attempt ledger extract in the PR showing a known-exhausted pair skipped.
Activity
dezren39 commentedon Nov 4, 2026
🤖 Agent note · picked up → PR opened (task 9239)
queue/9239-5295-known100-gate, commit8f0bef24e)Implemented:
knownExhaustedPairs+applyKnownExhaustedGateinbin/opencode-route-proxy.mjs— a pair whose latest quota-bearing ledger failure carries a future reset (Retry-After-derived or absolute) and no later success is excluded from candidates after demotion; fail-safe keeps the demoted order when exclusion would empty the route; decisions logged to the proxy's stderr diagnostics; clock vianowMs()soOPENCODE_ROUTE_PROXY_NOW_MSpins it.Verified on the laptop:
node --test bin/opencode-route-proxy.test.mjs— 31/31 pass, including 4 new pinned-clock tests (exclusion with ledger extract, return after reset, all-exhausted bypass, success-clears-gate). This issue closes via the PR'sFixes #5295on merge.fix(route-proxy): Codex round on #5295 — credential-aware gate fallba…
fix(route-proxy): Codex round 2 on #5295 — order gate observations by…
Merge pull request #5296 from developing-today/queue/9239-5295-known1…
dezren39 commentedon Nov 4, 2026
🤖 Agent note · done (task 9239)
9c8f91e48477c7908f55bf50ce5fff3c9ee87414; this issue closed by itsFixes #5295.Delivered: the known-100% seating gate in
bin/opencode-route-proxy.mjs(#4370 acceptance (c)): pairs whose latest quota-bearing ledger failure carries a future reset and no later success are excluded at seating time; fail-safe keeps the demoted order when no credentialed candidate survives; decisions on the proxy's stderr diagnostics; clock vianowMs().Verification:
node --test bin/opencode-route-proxy.test.mjs— 34/34 on the laptop, including 7 new pinned-clock gate tests. Codex review: three rounds — credential-aware fallback (P2), receipt-anchored Retry-After (P2), completion-time ordering (P2) — all fixed with regression tests and replied on-thread; final round completed with no findings.Claim labels cleared with this note.