Repository navigation
feat(fleet): shift executor mode — launch + shepherd OpenCode sessions for opencode-lane tasks (OpenCode-half step 2) #1655
Description
Part of the OpenCode-half redesign (design: files/opencode-half-redesign-2026-10-28.md in the keep-the-four-repos-healthy goal workspace; tracking: #1425). Drew's order, 2026-10-28: at least half of fleet tasks execute via OpenCode sessions.
Scope
Executor mode for queue shifts: when a claimed task's recorded lane is opencode, the shift launches and shepherds an OpenCode session instead of working the task itself.
- Depends on: feat(fleet): attach-first opencode client helper — forward lifecycle, service discovery, stdin-safe run wrapper #1357 (attach-first opencode client helper — the launch/poll surface) and the claim-time lane assignment (companion issue: ledger +
laneverb in repo-queue.py). - The shift keeps every current duty: claim, revalidation, hard-stop labels, label sync, laptop worktree under
~/git/fleet/, branch convention, heartbeats, checkpoints, verification against fresh GitHub state, and the terminal verb. The session never touches the queue helper. - Launch: one session per task on the laptop server via the helper, directory = the task worktree, model per the provider registry POOL MAP ladder under the ≤60-min measurement gate (workhorse default DeepSeek Flash high; Meta only via the
metaprovider; nevermuse/*;claude-codepause stays). Prompt = assignment packet (issue, acceptance criteria, branch, repo AGENTS.md rules, PR/CI discipline, completion contract). - Shepherding: poll the session; answer its questions from packet/issue context; stall → one "Continue." nudge; error-stop → one retry, one failover per the ladder, then fallback.
- Fallbacks execute directly and are recorded with a reason (
server_unreachable,helper_error,session_create_failed,quota,saturated,laptop_load,session_failed). Concurrency cap for queue-spawned sessions: 4 fleet-wide during the pilot. - Instruction + job-body text changes land through Drew's review of the quoted drafts (standing hold on job-body edits) — this issue implements them once approved; it does not pre-empt that review.
Acceptance criteria
- On the code queue, an opencode-assigned task produces a session whose work the shift verifies and completes with the standard structured result plus the execution block (
lane,session_id,model). - CI-pending tasks park with the session id recorded and reattach (or continue in the same worktree) on resume.
- Every fallback in a one-week pilot window carries a recorded reason; zero unrecorded fallbacks found by spot audit.
- Zero evaporated claims attributable to session death (the shift's heartbeat/terminal discipline is unchanged).
- Pilot report after one week on the code queue: executed share, miss breakdown, session failure classes.
Activity
developing-today commentedon Oct 28, 2026
CODE-669
dezren39 commentedon Oct 28, 2026
Intake review
## Acceptance criteriasection, so no body edit was made. One body item needs Drew, not a worker: the instruction + job-body text changes land through Drew's review of the quoted drafts (standing hold on job-body edits) — this issue implements them once approved and does not pre-empt that review.priority/high(core mechanism of a Drew-ordered redesign),effort/l(executor + shepherding + one-week pilot),source/human,status/needs-triage,area/infra,meta/enriched.dezren39 commentedon Oct 31, 2026
Status — owner order 2026-10-28, stalled, re-surfaced 2026-10-31 (supersession check vs #2132).
The Oct-28 left-on-the-ground audit owed this step a supersession check against the session-conductor issue #2132 (filed 2026-10-29) before re-dispatch. Verdict, 2026-10-31:
NOT superseded by #2132. #2132 (and its implementation follow-on #2153) merge the OpenChamber dispatcher + babysitter jobs into one conductor. They do not touch queue-shift behaviour at all — this step is about the queue lanes' own executor mode (a claimed
opencode-lane task is executed by launching + shepherding an OpenCode session from the shift), which no part of #2132/#2153 covers.Queue state, verified by direct read 2026-10-31: this step already has a live triage task — task 2623 (repo-health-code-pilot, queued, priority 70, attempts 0/3). Its dependencies are in place (step 1, #1654, is closed; the attach-first helper #1357 is closed). No duplicate packet was enqueued; the stall is queue position, not a missing row. Leaving this note so the next lane treats the step as live, in-scope work under Drew's 2026-10-28 order (at least half of fleet tasks execute via OpenCode sessions).
dezren39 commentedon Nov 2, 2026
Claimed as queue task 2623 (triage). Verifying executor-mode state against current main, then routing or closing with evidence.
dezren39 commentedon Nov 2, 2026
Triage verdict: SUPERSEDED by #3026 (closed 2026-11-02, stand-down executed)
This issue proposes shift-side executor mode (the shift itself launches + shepherds an OpenCode session for
opencode-lane tasks). That direction was retired today by #3026 "Move opencode executions to laptop-side scheduled tasks — retire the Muse boot handoff":bind_repo_task_lane_session, migration 0021). A shift that holds an opencode-lane claim yields (handed_off), never launches.origin/main:pm-stack/opencode-claimant/(prompt.md, poll.py with testedis_candidate, install.mjs). Verified on this laptop 2026-11-02 ~23:10 UTC: claimant mode file readslive,fires.jsonlactively firing (last fire 23:03Z).Remaining work is already tracked elsewhere: supervision #1656 (open, blocked), measurement #1657 (open, blocked). No new follow-up packet from this triage. Closing as superseded.