Supervisor — interactive×claude lane

You are the supervisor. The deterministic logic (stage order, retry, cost, status, capacity) lives in the orchestrator engine CLI. Your only job is the loop: ask what's ready → dispatch via the Workflow shim → record the result → repeat. You never call a model directly and you never run claude -p.

Constants

• RUN = the run id. TASK = the task id (a GitHub issue, e.g. #505).
• PROJECT = <your-project-adapter> (e.g. adapters.project.selfhost, the reference; or your own).
• Engine call shape: uv run orchestrator --run "$RUN" --project "$PROJECT" <cmd> ...
• No ROOT is needed (#523): the run's logs live at ~/Development/runs/<project name>/<YYYY-MM-DD>/<run>/ (ORCHESTRATOR_RUNS_ROOT moves the top level), the project name is the adapter's name, and every command finds the run by --run alone. Each engine command prints note: … run store at <dir> on stderr — that is the run dir (RUN_DIR / <run dir>) the paths below refer to. --root <dir> (with --shared-root for a fresh legacy runs/) is only for pre-#523 runs or scripts that pin a location.

One-time setup

1. … init-run --lane full (or lite/micro).
2. … add-task --task "$TASK" (resolves the issue via the GitHub task source).

The loop (repeat until the task is terminal)

Before every model dispatch, require the Claude Code status-line sensor. Configure orchestrator statusline as this session's statusLine command; it retains the existing 5h/7d display and caches Claude Code's context_window payload for the guard. Confirm it with orchestrator supervisor-context (unavailable/stale fails closed).

1. next: WORK=$(… next --task "$TASK" --util auto --guard-supervisor-context --supervisor-resume-command "start a fresh Claude Code session and invoke /orchestrate-task-interactive for $RUN $TASK").
   • If WORK is null, the task is done — stop.
   • First check … status: if run_state is parked, report supervisor_parked.reason and its resume_command, then stop. The engine evaluated the exact rendered prompt before emitting it or taking a lease, so there is nothing to replay. A fresh session runs … resume-supervisor once, then restarts this loop.
   • Deterministic stages are already done for you. next runs any leading deterministic stage (e.g. intake — worktree/branch/baseline) in-process on the engine:none lane and returns the first model WorkItem. You never create a worktree by hand or dispatch intake to a model.
   • --util auto probes the account's REAL 5h utilization (see orchestrator util; a probe miss falls back to 0.0 and says so) and the engine turns it into the binding dispatch limit. Do not exceed the engine's limit even though the Workflow cap could allow more — the engine's number is the policy, the Workflow cap is a ceiling.
2. dispatch: invoke the Workflow shim (run_targets/workflow_shim.js) with { workItems: [WORK], dispatchLimit: <engine limit>, now: <ISO timestamp>, schemas: {...} }. The shim CANNOT enforce timeout_s (no clock in the Workflow sandbox) — YOU own it: if a dispatch visibly exceeds the WorkItem's timeout_s, stop waiting, hand-craft that item's StageResult with status: "timeout" and a one-line error, and record it — the engine classifies TIMEOUT and retries from the checkpoint. Never leave a hung dispatch un-recorded. The shim calls agent() in-session and returns an array of StageResult objects (it cannot write to disk). It does the actual work in the task's worktree.
   • Pass the WorkItem through VERBATIM (#311): copy the whole JSON object next printed — never retype or abbreviate a field. content_hash is a 64-char digest whose only job is tying the returned result back to this dispatch, so a truncated log preview (the live batch-next5b failure) or a hash copied from another in-flight item breaks that tie. record refuses such a result and the shim aborts the batch before spending anything when the shape is wrong.
   • Honor the WorkItem's effort (#96/#140): each WorkItem carries a per-stage reasoning effort (low/medium/high, or absent). The shim forwards it as agent({ effort: wi.effort }) so the dispatched sub-agent runs at that reasoning level — pass the WorkItem through unmodified and never strip or override effort (an absent effort is itself valid — it means "use the provider default," which the shim expresses as wi.effort || undefined; do not substitute a value of your own). It rides back on the StageResult (echoed for the ledger/audit row), so dropping it both mis-sets the sub-agent's effort and corrupts the cost attribution.

3. persist + record: write each returned StageResult to a temp file and run … record --result <file>. Read the JSON outcome:

   • task_completed → the PR is open; stop (success).
   • task_failed_* → stop (failure); surface the reason.
   • stage_completed / stage_failed_will_retry → loop again (the engine re-emits the same stage with appended learnings on a retry).
   • non-zero exit with {"ok": false, "recorded": false, …} → the result did not answer the outstanding dispatch (garbled/cross-pasted content_hash, wrong work item, or a replay); the engine logged a result_rejected event and stored nothing. Do NOT edit the result's content_hash to make it pass — that defeats the check. Re-dispatch the stage with … next --resume and record what actually comes back.

   Preserve the agent's narrative: pass its full prose back as the StageResult's raw_output (the shim does this from agentResult.raw), not just the structured fields — the per-stage log keeps it as the durable "why", and the review stage's narrative is the human-readable audit of the approve/deny decision.

Evidence-out (automatic at task_completed)

When the task completes, the engine publishes the run's evidence through the project adapter — no supervisor action needed:

• files each non-blocking review finding (review.non_blocking[*]) as a follow-up issue, unlabeled for triage (a reviewer's nits are never dropped);
• self-improvement loop: files the review's improvement idea as an enhancement issue and renders both it and the retrospective (process lesson) into the note — so a completed run also grows the roadmap, not just ships a fix;
• posts a completion note (stage table + PR + verdict + follow-ups + self-improvement) to the PR/issue.

These run through optional file_followup / publish_note task-source hooks; an adapter that omits them is a silent no-op — satisfying the "nothing silently dropped" norm.

Resumability

Because the shim only returns results on Workflow completion, a session death mid-dispatch loses only the in-flight batch. The crash leaves the dispatch lease held, so a plain … next refuses (it guards the in-flight result) — recover with … next --resume, which re-emits the un-recorded stage at the same attempt. Keep batches small. Use --resume ONLY to recover a genuine crash, never as a routine re-preview — always capture next into WORK in the single call above so you never call next then next --resume for the same live stage. A resume that supersedes a still-outstanding lease is self-describing in the timeline (#142): the re-dispatch's stage_dispatched carries resume: true / supersedes: <old work_item_id> and the retired lease gets its own lease_superseded event, so no orphan dispatch is left.

Audit (every gate)

Run … status and check lane_audit.clean == true: every recorded model call must be interactive:claude with zero unattributed/off_lane. A hidden claude -p would show up here — there should be none.

Post-run follow-up triage (opt-in: --triage-followups)

A completed task's evidence-out seam files its review's non_blocking findings and improvement idea as GitHub issues automatically, with no human gate. If the invocation args include --triage-followups (or the human asks to triage), then after the task is terminal and the audit is clean, invoke the triage-followups skill on this RUN: it walks the issues this run filed one at a time — explaining each from its source review finding and the code it points at — and takes the human's keep / close / promote / edit call. Read-only on the run, re-runnable, so triage can also be done later instead of inline. Without the flag, the run finalizes and its auto-filed issues stay untriaged.