Factory Foreman Loop Iteration

[HARD] Before the first action of a session running this skill, read BOTH role-gated rule files in full: .claude/rules/moai/workflow/factory-dispatch.md and .claude/rules/moai/workflow/cross-session-messaging.md. A Codex-only project deploys the same files under .moai/policies/ (for example .moai/policies/workflow/factory-dispatch.md) — read whichever layout this project carries. The always-loaded surface carries only their stubs; the dispatch protocol, the card classes, and the queue boundaries this loop consumes live in those two files.

One unattended pass of the factory foreman: watch the backlog queue, dispatch the next operator-picked card to an isolated worker, collect completion evidence, report. The queue surface is moai todo (the moai gtd spelling serves the same queue); the dispatch protocol and card classes live in the factory dispatch rule (.claude/rules/moai/workflow/factory-dispatch.md). foreman — an auxiliary role of the leader: the unattended watcher that dispatches the already-picked card to an isolated worker when no leader session is holding the queue.

Running unattended

AskUserQuestion is removed from the tool pool while this skill is active — that is the mechanical guarantee that the loop cannot stop and ask. Anything that would have been a question becomes a line in the iteration report, and anything that genuinely needs the operator's decision becomes a blocked card under Boundaries below.

Deployment: start a session in the project, run bare /loop, then background the session — loop tasks carry over to the background session and keep running without a terminal. Esc cancels the pending wakeup of a waiting loop. A recurring loop expires seven days after creation; restart it when the queue still needs a foreman. Background monitors do not survive a session resume — the first iteration after a resume re-arms the queue watch.

The session's permission settings must already allow what this loop uses (queue reads, git inspection, the worker spawn). A permission prompt that surfaces while unattended stalls the iteration until someone attaches; pre-approving that surface in project settings is the operator's setup step, not something this loop can do for itself.

Boundaries (hard)

1. The operator admits and picks work. Only backlog items whose state is already picked are dispatchable. Never run moai gtd add; never run moai gtd next <n> — that mutation is the operator's pick. Never invent, reword, or reorder cards. An empty queue is a legitimate state: say so and idle. A batch authorization (/moai:todo --auto — the operator's typed invocation-as-approval) is the card-pick gate's autonomous form (.claude/rules/moai/workflow/auto-semantics.md §9): within it, serial consumption on its own judgment, outside the keep-set (bar a [보류 card, which is ranked last, not excluded), is authorized; queue ADMISSION stays the operator's. A lane takes its card through moai factory next --card <id>, never through this loop.
2. No approval gate is answered on the operator's behalf. When a card's next step needs a human decision that is not already recorded as made (plan-to-run kickoff approval, a review severity call, a scope choice), do not proceed. Leave the card picked, name it blocked-for-operator in the report together with the decision it waits on, and move on.
3. One write-capable worker at a time. While a worker is in flight the iteration only reads. This is the foreman's own serialization, stricter than the doctrine it sits under — one writer per tree, owned by .claude/rules/moai/core/agent-common-protocol.md § Background Agent Execution. The foreman keeps one worker in flight so a failed iteration has exactly one author to read.
4. Every worker runs in its own worktree (isolation: "worktree" on the spawn; relative paths in the prompt — the worker's CWD is its worktree root). Nothing writes to the shared checkout.
5. No integration actions. No push, no pull request, no merge, no branch deletion, no worktree disposal. The card's branch is unpushed and its worktree is the work's only instance; both stay until the operator integrates them. The report names the branch and the worktree path.
6. Verification is lane-local. The worker runs only the checks its own change can affect; the full suite belongs to CI. Never spawn background CPU load — the queue watch below is the only long-running process this loop arms.
7. Completion is read, never trusted. A card advances only on evidence this iteration actually read.

The iteration

1. Queue watch. If no backlog monitor is live (first iteration, or after a resume), arm one Monitor on the queue file, re-arming it at each expiry:

   • command:

     # The queue directory, resolved the way factory.StateDirForRoot does for a
     # standard git-repository project: <moai-home>/db/<project-key>/todo,
     # keyed by the repository's canonical (primary-checkout) root.
     mh=${MOAI_HOME:-$HOME/.moai}
     root=$(git rev-parse --show-toplevel 2>/dev/null) || root=$PWD
     top=$(git worktree list --porcelain 2>/dev/null | sed -n 's/^worktree //p' | head -n 1)
     [ -n "$top" ] && root=$top
     root=$(cd "$root" && pwd -P)
     key=$(basename "$root"); key=$(printf '%s' "$key" | tr -c 'A-Za-z0-9._-' '-')
     sum=$(printf '%s' "$root" | sha256sum 2>/dev/null | cut -c1-8)
     [ -n "$sum" ] || sum=$(printf '%s' "$root" | shasum -a 256 | cut -c1-8)
     d=$mh/db/$key-$sum/todo
     last=init
     while true; do
       cur=$(cksum "$d"/backlog.db "$d"/backlog.db-wal 2>/dev/null)
       [ -n "$cur" ] || cur=missing
       if [ "$cur" != "$last" ]; then
         [ "$last" != init ] && echo "backlog changed"
         last=$cur
       fi
       sleep 5
     done
     

   • timeout_ms: 1800000

   • description: backlog queue watch

   The persistent option no longer exists — every Monitor now carries a deadline, timeout_ms is a required input, and the watch above is an unbounded loop that never ends on its own. 1800000 is the longest deadline the tool documents accepting; a larger value is capped rather than honoured. At expiry the loop is killed and one notice arrives with the event count, so an iteration that still needs the watch re-arms it. A non-interactive (-p) session is reported to cap the deadline lower than an interactive one; that lower cap is not observable from the tool input schema, so treat a shorter-than-requested expiry there as expected rather than as a fault.

   The watch resolves the queue directory the way factory.StateDirForRoot does for a standard git-repository project — a project-keyed directory under the moai home (MOAI_HOME when that is set to an absolute path, otherwise ~/.moai), keyed by the primary checkout's root — so a linked worktree watches the primary checkout's queue, not a directory local to its own tree. The queue is the database, and a backlog.json beside it is an export or a legacy leftover — never the queue — so a watch pointed at the JSON on a migrated project polls a file that never changes and reports nothing, forever. The write-ahead log is watched alongside the database because a committed write can sit in backlog.db-wal with the database image byte-identical until a checkpoint folds it back; watching the database alone misses exactly those mutations. The checksum poll emits one line per change and costs two tiny reads every five seconds. Do not tighten the interval, and do not arm a second watcher. Each emitted line, like each scheduled wakeup, is a prompt to run this same idempotent iteration — an iteration that finds nothing to do ends quickly.

2. Read the queue. moai gtd list --json (lock-free). A missing queue file is an empty queue, never an error. Records carry id, text, spec_id, and state (queued | picked | dropped).

3. Collect before dispatch. If a worker dispatched by an earlier iteration has returned, read its evidence file first (step 6). If a worker is still running, end the iteration with a one-line status.

4. Choose the dispatchable card. The oldest picked item with no live worker and no recorded blocker. queued items are not yours to pick outside a batch authorization.

5. Dispatch one worker with the Agent tool, isolation: "worktree". The dispatch is a fixed-field address block — a pointer, not a copy, ten lines at most:

   card: <id>
   spec: <SPEC-ID>            # only when the card carries one
   cmd: <the card's work in one line; the phase command when a SPEC is attached>
   evidence: .moai/reports/<card-id>/evidence.md
   

   Route the spawn to the agent the card's work matches under the standard delegation rules; where no specialist matches, a general-purpose spawn with a domain whitelist. The worker prompt carries the block plus these standing orders: rename the worktree branch to WT-<slug> (a descriptive slug, never the card id) first; implement the card; verify lane-locally; write the evidence file with decisions, verbatim output tails of the checks run, explicit gaps, and residual risk; commit by explicit pathspec; never push.

6. Collect on evidence. When the worker returns, read the evidence file it names. Advance the card — moai gtd done <t-id> — only when the evidence shows the work complete: verbatim passing output present, gaps named. A missing, unreadable, or stale evidence file is a gap: the card stays picked, the report says why, and the card is not re-dispatched this iteration. Absence of a failure signal is not a pass.

7. Report. Close with two to six lines: queue summary, what was dispatched or collected, which evidence was read, blocked cards and the decisions they wait on. This report is what the operator reads on reattach — name what you read, not what you were told.

Factory seam (reserved, not implemented)

The single-worker dispatch above is the only mode. Fanning a card out to numbered factory lanes — the multi-lane launcher surface — is separate work; when the foreman grows that routing, it lands here as a second dispatch mode chosen per card. Until then this loop spawns one subagent per card, reads no factory state, and launches no lanes.

Failure handling

Stop the loop (ScheduleWakeup with stop: true), with a one-line reason, when the loop cannot do its job: the queue file is repeatedly unreadable, the queue watch cannot be armed, or this skill's own surface is broken. A transient worker failure is not a loop failure — record it on the card and let the next iteration decide whether to re-dispatch.