opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone

Summary

pm-stack/opencode-claimant/poll.py reads GET /api/session once and treats the reply as the complete session list. That endpoint truncates at 50 entries (ordered by update), so the claimant both undercounts the concurrency cap — it can launch past CAP = 12 — and reports live bound sessions as missing, which the launch rules say must never happen.

Observed 2026-11-03T04:42Z on the laptop server (http://127.0.0.1:13420):

GET /api/session → exactly 50 sessions, 50 roots.
Tasks 6687 (ses_e5e60c36744ffeUY2rSsCrjMoj), 6639 (ses_e5e613367a0ffe3AfMkKP2LOjB), 6722 (ses_e5e60bd3d74ffewwDg1Qd0n2Qa), 4876 (ses_e5e61a3c885ffeJ9P4DjRpBFVy) — all bound in progress.session_id, all absent from that list — all four return 200 from GET /api/session/<id>.

Impact

Cap overcount of room. summarize_cap() (poll.py:112-128) intersects the 50-row page with /api/session/active. Any active root outside the page is invisible, so active_root_sessions is understated and cap_room overstated. A fire could then launch into a saturated fleet, breaking the one hard rule that is purely about not overloading the host.
False "session is gone". Supervision (claimant instructions step 1) says to check a bound session "still exists / is progressing via GET /api/session". On the truncated page an alive session reads as gone. The relaunch rule is "never relaunch over a live session"; a list-based check turns that into a duplicate-PR class defect waiting to happen.

Proposed fix

Count actives from /api/session/active directly (its id set is authoritative and small) rather than filtering a page of /api/session; if root-ness is still needed, fetch each active id with GET /api/session/<id>.
Supervision existence must be a per-id GET /api/session/<id>, never membership in the list.
If a page is genuinely needed, paginate or pass a large limit and fail closed (cap_room: null) when the returned row count equals the page size — the same fail-closed rule the claimant already applies to an unreadable list.

Evidence trail

Fire log lines for 2026-11-03 in ~/.local/state/opencode-claimant/fires.jsonl carry the full supervision table for that fire.

Acceptance criteria

Cap accounting counts active roots from /api/session/active directly (per-id GET for root-ness if needed); summarize_cap() no longer intersects a truncated /api/session page.
Supervision existence checks for a bound session use GET /api/session/, never membership in the list endpoint.
If any list read remains, it paginates or fails closed (cap_room: null) when the returned count equals the page size.
A regression test or fire-log evidence shows correct cap accounting with >50 sessions on the server.

Intake review

Evidence in the issue: exactly 50 sessions returned while four bound sessions (tasks 6687, 6639, 6722, 4876) were absent from the list yet returned 200 per-id. Sibling defect: #4758 (same claimant, same morning). Still live at intake: no merged PR references this issue (verified this run). Risks: this is the fleet's host-overload guard — undercounting actives can launch into a saturated laptop. Suggested approach: patch poll.py per the issue's proposed fix (active-set counting + per-id existence), which is small and localized.

— fleet monitor intake (repo-fleet-monitor, 2026-11-03)

Fresh evidence from the 2026-11-03T13:03Z claimant fire (live mode, cap closed the launch gate, nothing launched):

GET /api/session → exactly 50 rows again. GET /api/session?limit=100 → 100, ?limit=200 → 200, so the default page is the cap, not the session count.
Six bound ids are claimed + progress.lane=opencode + progress.session_id (tasks 1930, 4858, 4876, 5697, 5893, 6418) and all six were absent from the 50-row page while each returned 200 from GET /api/session/<id>, time.idle set and outcome=succeeded. A list-membership existence check would have called six live-but-terminal sessions "gone" and, once their leases expired, made them eligible for a relaunch.
Cap count: poll.py's page-based read reported 18 active roots; the per-id count from /api/session/active (fetch each id, keep non-child roots) reported 20. Undercount confirmed in the direction the issue predicts. Both are ≥ CAP = 12, so this fire's decision (launch nothing) was unaffected — the risk is real but did not bite this time.

Supervision this fire already used per-id GET /api/session/<id> for existence, per the proposed fix; poll.py still needs the summarize_cap() change (active ids from /api/session/active, root-ness per-id) and the fail-closed page-size check.

Full fire line: ~/.local/state/opencode-claimant/fires.jsonl at 2026-11-03T13:03:29Z.

Adopted as Phase 1 of design #4912 (judge a session by its own record — the opencode-claimant's session contract), filed by the fleet design architect, pass 44 (2026-11-03). Design doc: files/designs/code-claimant-session-contract.md in the keep-the-four-repos-healthy goal workspace.

Phase structure for triage: do NOT re-file, re-scope, or split this issue — it is the phase. Its acceptance stands as written and is decision 1 of the design (cap from the /api/session/active id set with per-id root checks; supervision per-id; any remaining list read fails closed at page-size equality — the same treatment summarize_cap already gives an unreadable active set). Phases 2 (#4870) and 3 (#4758) are independent; Phase 4 (#4916, closing audit) is gated on all three.

agent-note:v1 — design-architect adoption note, pass 44.

🤖 Agent note · handed off

Agent: queue worker (repo-worker-code), lane code-y
Session / ID: task 7566
Found via: queue claim
Task: task 7566 (triage, issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757 — opencode-claimant session-cap read truncates at 50) assigned to the opencode execution lane; handing off to the laptop claimant. Issue re-verified OPEN this shift; claim labels were stale (queue/queued) and were synced to queue/claimed + status/in-progress + work/assigned this shift. No hard-stop labels.
Expectation: claimant launches the bound opencode session and drives task 7566 to a terminal state.
Links: issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757

🤖 Agent note · handed off

Agent: queue worker (repo-worker-code), lane code-c
Session / ID: task 7566
Found via: queue claim · repo-health-code-pilot
Task: task 7566 (triage, issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757) claimed by lane code-c and assigned to the opencode execution lane; handing off to the laptop opencode claimant, which launches the bound session.
Expectation: done = task 7566 acceptance met for issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757.
Links: issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757

🤖 Agent note · handed off

Agent: queue worker (repo-worker-code), lane code-s
Session / ID: task 7566
Found via: queue claim
Task: task 7566 (triage for issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757) was assigned to the opencode execution lane; handing off to the opencode claimant, which launches the bound session. Revalidated fresh: issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757 is OPEN with no hard-stop labels (labels already claimed-synced). Noted for the claimant/fleet: issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757 is itself about the claimant — its cap and supervision reads of GET /api/session truncate at 50 sessions, which is a plausible mechanism for the unlaunched-handoff cycles observed across lanes tonight (handed-off tasks returning after lease expiry). task 7566 is at attempts 22/24, so if this lease expires unlaunched it will cap out and need a cap raise or monitor repair.
Expectation: effort s; done = the claimant's bound session triages issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757 at current main with evidence and routes it, driving task 7566 to a terminal state.
Links: task 7566 attempts 22/24

Agent note — handed off to opencode claimant (task 7566) — ATTEMPTS 23/24, ONE CYCLE FROM CAP

Task: task 7566 (repo-health-code-pilot, lane code-s, execution lane opencode, 2026-11-05). Triage packet for issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757: the opencode-claimant's cap/supervision reads of GET /api/session truncate at 50 sessions.
Fresh revalidation (2026-11-05T06:58Z): issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757 OPEN, no hard-stop labels; claim labels already in sync (queue/claimed, status/in-progress, work/assigned).
Handoff: execution lane assigned opencode; the laptop claimant launches the bound session. This is attempt 23 of 24: if this lease also expires unlaunched (the pattern all last shift — every opencode handoff expired unlaunched, plausibly because of this very issue's truncation bug), task 7566 hits its attempt cap and becomes unclaimable until a cap raise or monitor repair. Flagging for the fleet monitor / Drew: cap raise for task 7566, or direct-lane reassignment so a worker lane can triage issue opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757 itself.

🤖 Agent note · handed off

Agent: fleet monitor
Session / ID: task 7566 (repo-fleet-monitor run 2026-11-05T15:32:18Z)
Found via: monitor reconciliation (Duty 1 claim/label divergence repair)
Task: Claim evaporated: task 7566 is queued and unclaimed in repo-health-code-pilot, but this issue still carried claimed labels. Labels reset to the truth (queue/queued + status/needs-triage). Checkpoint stage at evaporation: handed_off.
Expectation: The next lane claim picks task 7566 up from the queued state and posts a fresh pickup note.

Navigation Menu

Repository navigation

opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757

Summary

Impact

Proposed fix

Evidence trail

Acceptance criteria

Intake review

25 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

opencode-claimant: cap and supervision read GET /api/session, which truncates at 50 — active roots undercounted, live bound sessions reported gone #4757

Description

Summary

Impact

Proposed fix

Evidence trail

Acceptance criteria

Activity

dezren39 commented on Nov 3, 2026

Intake review

dezren39 commented on Nov 3, 2026

dezren39 commented on Nov 3, 2026

25 remaining items

dezren39 commented on Nov 5, 2026

dezren39 commented on Nov 5, 2026

dezren39 commented on Nov 5, 2026

dezren39 commented on Nov 5, 2026

dezren39 commented on Nov 5, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions