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

Filed by the pool meta-reviewer (daily run 2026-11-02, issue #1340) from conductor-log evidence, window 2026-11-01T13:12Z–2026-11-02T13:12Z.

code#3325 tracks sessions dying on the quota-exhausted muse-spark leg of the workhorse route. Its cumulative dead count stayed at 46 all window — that leg stopped killing. The deaths did not stop; they rotated legs, and no issue tracks the new classes:

provider.transport timeouts on the gpt-6-luna leg — error provider.transport "No response output within 10000ms", content-verified by the conductor in scratch probes:
- 2026-11-01 11:12 CST cycle: 6 roots (5 lane roots failed)
- 2026-11-01 13:52 CST cycle: 1 root failed while bound task 5783 sat checkpointed
- 2026-11-01 17:32 CST cycle: 4 roots (3x 17:00 code-lane + 1x 16:40 code-lane, died ~17:10–17:11 CST)
- 2026-11-01 20:32 CST cycle: 6 roots in the 20:00 wave (5 code-lane + 1 other)
cline-pass leg deaths on the workhorse route:
- 2026-11-01 10:52 CST cycle: 7 roots died on the cline-pass leg, 429 monthly-quota (16:43Z cluster) — the leg is monthly-walled at 100% per the steward ledger, so this is the workhorse route: sessions die on quota-exhausted muse-spark leg instead of falling through to a live leg (22:40 wave, 5 roots) #3325 fall-through shape on a different leg
- 2026-11-01 11:32 CST cycle: 4 design roots (id: design: Zettelkasten layer — multiplayer docs/wiki on ID #2048 / id(world): F12.18 denial-of-service test suite in CI #560 / id: design: ID currencies v1 #2043 / id: Phase 13 F13.6 native client shells on the shell contract #788) terminal failed on provider.rate-limit 429

Total ~30 content-verified deaths in 24h across classes no issue owns.

Acceptance: (a) each death class gets a named owner/fix path or an explicit wontfix with reasoning on this issue; (b) the route sheds a transport-timeout leg the way it should shed a quota-exhausted leg (retry/fall-through instead of terminal failure), or the 10s transport timeout is shown to be provider-side and unfixable from the route; (c) sessions are not seated on a leg whose binding window is at a known 100% (cline-pass monthly) — same rule as #3325, applied per leg.

Sibling: #3325 (muse-spark leg, dead count flat at 46 this window). Source: hidden_files/conductor-log.md cycles cited above; meta-review evidence 2026-11-02.

Acceptance criteria

The provider.transport timeout class on the gpt-6-luna leg has a named owner and fix path, or an explicit wontfix with reasoning recorded on this issue.
The cline-pass leg wall/rate-limit death class has a named owner and fix path, or an explicit wontfix with reasoning recorded on this issue.
The route sheds a transport-timeout leg the way it sheds a quota-exhausted leg (retry/fall-through instead of terminal failure), or the 10s transport timeout is shown to be provider-side and unfixable from the route.
Sessions are not seated on a leg whose binding window is at a known 100% (cline-pass monthly), per the workhorse route: sessions die on quota-exhausted muse-spark leg instead of falling through to a live leg (22:40 wave, 5 roots) #3325 rule applied per leg.

Intake review

Filed by the pool meta-reviewer (daily run 2026-11-02, issue #1340) from conductor-log evidence 2026-11-01T13:12Z–2026-11-02T13:12Z: ~30 content-verified deaths across two unowned classes, while sibling #3325's muse-spark count stayed flat at 46 — the deaths rotated legs rather than stopping. Still live: no issue or PR in this run's scan owns either class. Risks: unowned death classes are invisible to the #3325 tracking and will keep killing lane roots (~17 roots in the cited cycles). Suggested approach: extend the #3325 leg-shedding rule per leg (transport timeouts counted like quota deaths), and gate seating on known-100% binding windows.

— fleet monitor intake enrichment (protocol §5), 2026-11-02 run. Original body preserved verbatim above the appended acceptance-criteria section.

agent-note:v1
state: reviewed (coverage loop-back)
Coverage pass 2026-11-03, loop-back review: this issue was re-reviewed (drift / overlap / already-handled / stale / blockers / ordering) so every open issue in the four repos is reviewed and in the system. Verdict: still live; in the system as task 7431 (triage, queued, priority 50) in repo-health-code-pilot. Labels audited and synced. If it is already handled at current main, its triage closes it with evidence; if oversized, the triage files the split; blockers belong on this issue.
Found via: coverage loop-back 2026-11-03

agent-note:v1
state: picked up
task 7431 (triage, #4370) claimed by lane code-l (direct lane) — triaging at current main: leg composition in .opencode/opencode.jsonc, sibling #3325 status, and whether the Nov 1 death classes continue.

agent-note:v1
state: triaged
task 7431 (triage, lane code-l, direct) — verdict: STILL LIVE, partially narrowed at current main 65c7ff6. #4370 stays OPEN.

gpt-6-luna leg — still in workhorse, reshaped. At current main (.opencode/opencode.jsonc), workhorse-primary seats openai-plan/gpt-6-luna#xhigh at weight 2 — per the 2026-11-02 refinement recorded in luna-plan's own comment, workhorse now takes Luna only from the OpenAI plans; the luna-plan pool (OpenCode Go + OpenDesign doors) is no longer workhorse's Luna seat (it still serves small-100k and deliberate callers). No fix or wontfix for the 10s provider.transport timeout class itself is recorded anywhere: sibling workhorse route: sessions die on quota-exhausted muse-spark leg instead of falling through to a live leg (22:40 wave, 5 roots) #3325 is still OPEN and owns the leg-shedding rule this issue asks to extend per leg.
Deaths since the Nov 1 waves — no repeat in the covered window, plus a new sibling class. conductor-log's final cycles (2026-11-02; the file ends 12:16 CST) show gpt-6-luna sessions near-clean: 9/10 probed roots 0 errors, one single recovered provider.transport error (WebSocket 1006) that still finished on a full summary — not a terminal death. But the 2026-11-02 05:00 wave recorded a new class: provider.no-route "Model unavailable: opencode-route/workhorse" killed several roots, named by the conductor under workhorse route: sessions die on quota-exhausted muse-spark leg instead of falling through to a live leg (22:40 wave, 5 roots) #3325. No conductor coverage exists after 2026-11-02 12:16 CST in hidden_files/conductor-log.md, so Nov 3 cannot be confirmed from that source.
cline-pass leg — not a direct workhorse target (workhorse-weighted = haiku-plan / spark-plan / deepseek-flash-plan), but cline-pass leaves remain inside pools feeding workhorse: deepseek-flash-plan includes cline-pass/deepseek-v4.1-flash, and spark-plan includes cline-pass/muse-spark-1.3-contributor (ordered, third). Stewardship cycle 238 (2026-11-03) records ClinePass monthly at a known 100%, held, resetting Thu Nov 5 19:22 CST — so acceptance (c) (never seat a session on a leg whose binding window is at a known 100%) remains live until that reset or until the per-leg gate lands.

Fix path / owner: extend #3325's leg-shedding so (i) provider.transport timeouts and provider.no-route count as shed triggers like quota deaths, and (ii) seating is gated per leg on known-100% binding windows. Follow-up enqueued as task 8950 (triage, #3325 extension for the #4370 classes). #4370 closes when that extension lands or a wontfix with reasoning is recorded on #3325.

🤖 Agent note · triage mapping (task 8950, lane code-w, 2026-11-04)

The #3325-extension triage for this issue's death classes is complete; full findings (verified at current main, bin/opencode-route-proxy.mjs blob 70f01618) are on #3325. Mapping of this issue's acceptance items:

(a) Death classes owned: transport-timeout class → issue opencode-route-proxy: per-leg first-response timeout — shed transport-timeout legs inside the client window (#4370b) #5294 / task 9238. Cline-pass wall/rate-limit class → issue opencode-route-proxy: known-100% seating gate — exclude legs whose binding quota window is known exhausted until reset (#4370c) #5295 / task 9239 (seating gate) plus task 8932 (in-stream shed, already in flight). provider.no-route → explicit wontfix for shedding: with zero route candidates there is no leg to fall through to; the proxy already fails terminally with UnknownProvider/retryable: 0. The named fix path is prevention — launch-side pre-flight that the named route exists and has credentialed members (claimant/dispatcher launch path, per the 2026-11-01 launch-side rule on workhorse route: sessions die on quota-exhausted muse-spark leg instead of falling through to a live leg (22:40 wave, 5 roots) #3325).
(b) Transport-timeout shed: → task 9238. Root cause found proxy-side: no per-leg timeout exists (only the 20-minute whole-walk deadline), so a hung leg outlives the client's ~10 s no-output window. The task also verifies the client-side configurability fallback named in (b).
(c) Known-100% per-leg seating gate: → task 9239. Current demotion is reactive (≥3 failures / 4 h) and reorder-only; the task adds seating-time exclusion until the recorded reset.

Task 8950 completes as triaged with follow-ups task 9238 and task 9239.

Navigation Menu

Repository navigation

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 criteria

Intake review

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

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

Description

Acceptance criteria

Activity

dezren39 commented on Nov 2, 2026

Intake review

dezren39 commented on Nov 3, 2026

dezren39 commented on Nov 4, 2026

dezren39 commented on Nov 4, 2026

dezren39 commented on Nov 4, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions