ftown-bridge 0.18.2 → 0.19.1

package/skills/factory/factory-template/bin/dispatch.py

@@ -0,0 +1,472 @@
+#!/usr/bin/env python3
+"""No-LLM factory dispatcher — one tick per invocation, then exit.
+
+Run by an ftown interval loop from a deployed project's repo root:
+
+    uv run --with fticket,pyyaml python factory/bin/dispatch.py [--dry-run]
+
+Responsibilities in a single tick: run the fticket scheduler, dead-letter stragglers,
+claim queued tickets up to capacity and spawn one worker session per claim (rejecting
+first-stage tickets with a missing/empty request.md and tickets with 3+ zero-progress
+claim expiries — both dead-lettered instead of respawned), track worker_id -> session_id
+in .ffactory/workers.json and reap the recorded session on every event that retires a
+worker's claim (claim expiry, dead-letter, or a normal terminal outcome — advance,
+complete, reject, release; backstops the worker's own self-close), and forward
+dead-letter/orphan events to the operator session. All coordination is server-side and
+atomic (claims fence), so two dispatchers may race safely with no lockfile.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+import traceback
+import uuid
+from dataclasses import dataclass
+from pathlib import Path
+
+import yaml
+
+from fticket import FTS
+from fticket.errors import FTSError
+from fticket.types import EventType, HistoryKind, Status
+
+# The ftown spawn CLI. Overridable (a test points it at a stub script via the env var).
+FTOWN_CLI = str(Path(os.environ.get("FTOWN_SESSIONS_BIN", "~/.ftown/ftown-sessions")).expanduser())
+
+TERMINAL_STATUSES = (Status.DONE, Status.DEAD_LETTER)
+FORWARD_EVENTS = (EventType.TICKET_DEAD_LETTER, EventType.TICKET_ORPHANED)
+# Reap (ftown remove) the recorded session on every lifecycle event that retires a
+# worker's claim: crash/hang past claim TTL, ticket killed under it, or a normal
+# terminal outcome. Self-close is LLM compliance, not a guarantee — the dispatcher backs
+# it up by removing the session itself; an already-self-closed session errors harmlessly
+# (non-fatal, stderr-logged in ftown_remove).
+REAP_EVENTS = (
+    EventType.TICKET_CLAIM_EXPIRED,
+    EventType.TICKET_DEAD_LETTER,
+    EventType.TICKET_ADVANCED,
+    EventType.TICKET_COMPLETED,
+    EventType.TICKET_REJECTED,
+    EventType.TICKET_RELEASED,
+)
+CURSOR_FILE = Path(".ffactory") / "dispatch.cursor"
+WORKERS_FILE = Path(".ffactory") / "workers.json"
+DEAD_LETTER_ACTOR = "dispatcher"
+ZERO_PROGRESS_LIMIT = 3
+
+
+@dataclass(frozen=True, slots=True)
+class StageCfg:
+    name: str
+    harness: str
+    model: str | None
+    max_workers: int
+    next: str
+    bounce: str
+    claim_ttl_ms: int | None
+
+
+@dataclass(frozen=True, slots=True)
+class Limits:
+    max_sessions: int
+    bounce_limit: int
+    claim_ttl_ms: int
+    max_ticket_age_h: int
+
+
+@dataclass(frozen=True, slots=True)
+class FactoryConfig:
+    project: str
+    operator_session: str
+    limits: Limits
+    stages: tuple[StageCfg, ...]
+
+
+def load_config(yaml_path: Path) -> FactoryConfig:
+    raw = yaml.safe_load(yaml_path.read_text())
+    limits_raw = raw.get("limits", {})
+    limits = Limits(
+        max_sessions=int(limits_raw.get("max_sessions", 1)),
+        bounce_limit=int(limits_raw.get("bounce_limit", 3)),
+        claim_ttl_ms=int(limits_raw.get("claim_ttl_ms", 1_800_000)),
+        max_ticket_age_h=int(limits_raw.get("max_ticket_age_h", 48)),
+    )
+    stages = tuple(
+        StageCfg(
+            name=str(s["name"]),
+            harness=str(s["harness"]),
+            model=(str(s["model"]) if s.get("model") else None),
+            max_workers=int(s.get("max_workers", 1)),
+            next=str(s.get("next", "-")),
+            bounce=str(s.get("bounce", "-")),
+            claim_ttl_ms=(int(s["claim_ttl_ms"]) if s.get("claim_ttl_ms") else None),
+        )
+        for s in raw.get("stages", [])
+    )
+    return FactoryConfig(
+        project=str(raw.get("project", "factory")),
+        operator_session=str(raw.get("operator_session", "-")),
+        limits=limits,
+        stages=stages,
+    )
+
+
+def compose_briefing(
+    *,
+    ticket_id: int,
+    stage: StageCfg,
+    db_path: Path,
+    ticket_dir: Path,
+    repo_root: Path,
+    epoch: int,
+    worker_id: str,
+    operator_session: str,
+) -> str:
+    lines = [
+        f"TICKET_ID={ticket_id}",
+        f"STAGE={stage.name}",
+        f"NEXT_STAGE={stage.next or '-'}",
+        f"BOUNCE_STAGE={stage.bounce or '-'}",
+        f"FTS_DB={db_path}",
+        f"TICKET_DIR={ticket_dir}",
+        f"REPO_ROOT={repo_root}",
+        f"EPOCH={epoch}",
+        f"WORKER_ID={worker_id}",
+        "",
+        "Task: Read and follow $REPO_ROOT/factory/skills/_protocol.md, "
+        f"then execute your stage per $REPO_ROOT/factory/skills/{stage.name}.md.",
+    ]
+    if operator_session == "-":
+        lines.append(
+            "You have no parent session: skip only the `tell --parent` mail command and "
+            "end your final message with the one-line result instead. The self-close "
+            'command (`~/.ftown/ftown-sessions remove "$FTOWN_SESSION_ID"`) is still '
+            "MANDATORY and must be the very last command you run."
+        )
+    return "\n".join(lines)
+
+
+def spawn_worker(
+    *, cfg: FactoryConfig, stage: StageCfg, ticket_id: int, briefing: str, repo_root: Path
+) -> str | None:
+    """Fire an ftown worker session. Returns the new session id ("" if the CLI output
+    could not be parsed) on success, None on failure; never retries in a tick."""
+    cmd: list[str] = [FTOWN_CLI, "create", "--shell", stage.harness]
+    if stage.model:
+        cmd += ["--model", stage.model]
+    cmd += ["--workdir", str(repo_root), "--name", f"{cfg.project}-t{ticket_id}-{stage.name}"]
+    if cfg.operator_session != "-":
+        cmd += ["--parent-id", cfg.operator_session]
+    cmd += ["--prompt", briefing]
+    try:
+        proc = subprocess.run(cmd, capture_output=True, text=True, check=True)
+    except (OSError, subprocess.CalledProcessError) as exc:
+        detail = getattr(exc, "stderr", "") or str(exc)
+        print(f"spawn failed for ticket {ticket_id} stage {stage.name}: {detail}", file=sys.stderr)
+        return None
+    # `create` prints JSON by default: {"session": {"id": "...", ...}}.
+    try:
+        session_id = str(json.loads(proc.stdout)["session"]["id"])
+    except (ValueError, KeyError, TypeError):
+        print(f"spawn ok but no session id parsed for ticket {ticket_id}", file=sys.stderr)
+        return ""
+    return session_id
+
+
+def load_workers(repo_root: Path) -> dict[str, dict]:
+    """worker_id -> {session_id, ticket_id, epoch}. Missing/corrupt file -> empty map."""
+    try:
+        data = json.loads((repo_root / WORKERS_FILE).read_text())
+    except (OSError, ValueError):
+        return {}
+    return {str(k): v for k, v in data.items() if isinstance(v, dict)} if isinstance(data, dict) else {}
+
+
+def save_workers(repo_root: Path, workers: dict[str, dict]) -> None:
+    path = repo_root / WORKERS_FILE
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(workers, indent=2, sort_keys=True) + "\n")
+
+
+def ftown_remove(session_id: str) -> None:
+    try:
+        subprocess.run([FTOWN_CLI, "remove", session_id], capture_output=True, text=True, check=True)
+    except (OSError, subprocess.CalledProcessError) as exc:
+        detail = getattr(exc, "stderr", "") or str(exc)
+        print(f"remove failed for session {session_id}: {detail}", file=sys.stderr)
+
+
+def ftown_tell(session_id: str, message: str) -> None:
+    cmd = [FTOWN_CLI, "tell", session_id, "--type", "task", message]
+    try:
+        subprocess.run(cmd, capture_output=True, text=True, check=True)
+    except (OSError, subprocess.CalledProcessError) as exc:
+        detail = getattr(exc, "stderr", "") or str(exc)
+        print(f"tell failed for session {session_id}: {detail}", file=sys.stderr)
+
+
+def run_scheduler(fts: FTS) -> tuple[int, int]:
+    """Drain the scheduler up to 5 rounds. Returns (requeued, promoted)."""
+    requeued = 0
+    promoted = 0
+    for _ in range(5):
+        report = fts.tick()
+        requeued += report.claims_expired
+        promoted += report.tickets_unblocked
+        if not report.more_pending:
+            break
+    return requeued, promoted
+
+
+def sweep_stragglers(fts: FTS, cfg: FactoryConfig) -> int:
+    """Dead-letter every non-terminal ticket older than max_ticket_age_h. Runs in dry-run too."""
+    now_ms = int(time.time() * 1000)
+    max_age_ms = cfg.limits.max_ticket_age_h * 3_600_000
+    deadlettered = 0
+    for node in fts.dag().nodes:
+        if node.status in TERMINAL_STATUSES:
+            continue
+        try:
+            ticket = fts.get_ticket(node.id)
+        except FTSError as exc:
+            print(f"straggler read failed for ticket {node.id}: {exc}", file=sys.stderr)
+            continue
+        if now_ms - ticket.created_at_ms <= max_age_ms:
+            continue
+        try:
+            fts.dead_letter(ticket.id, actor=DEAD_LETTER_ACTOR, reason="max_ticket_age exceeded")
+            deadlettered += 1
+        except FTSError as exc:
+            print(f"dead_letter failed for ticket {ticket.id}: {exc}", file=sys.stderr)
+    return deadlettered
+
+
+def zero_progress_expiries(fts: FTS, ticket_id: int) -> int:
+    """Count claim expiries where the worker never started (claimed -> queued).
+
+    fticket's history does not record renew_count; an expiry annotated with
+    from_status=claimed means the worker never even called start() — the closest
+    observable to "renew_count == 0". One fts.get_history call per candidate.
+    """
+    return sum(
+        1
+        for h in fts.get_history(ticket_id)
+        if h.kind is HistoryKind.ANNOTATION
+        and h.note == "claim_expired"
+        and h.from_status is Status.CLAIMED
+    )
+
+
+def capacity_spawn(
+    fts: FTS,
+    cfg: FactoryConfig,
+    workers: dict[str, dict],
+    *,
+    repo_root: Path,
+    db_path: Path,
+    dry_run: bool,
+) -> tuple[list[int], int, int, int]:
+    """Claim + spawn up to capacity, stages in yaml order.
+
+    Returns (spawned_ids, rejected_input, expiry_capped, active_total). Guards run just
+    after the claim (fts.claim picks the ticket, so preflight can't run earlier); a guard
+    hit dead-letters the freshly claimed ticket instead of spawning.
+    """
+    depths = {d.stage: d for d in fts.board()}
+    first_stage = cfg.stages[0].name if cfg.stages else None
+
+    def active_of(name: str) -> int:
+        d = depths.get(name)
+        return (d.claimed + d.in_progress + d.blocked) if d is not None else 0
+
+    active_total = sum(active_of(s.name) for s in cfg.stages)
+    budget = cfg.limits.max_sessions - active_total
+    spawned: list[int] = []
+    rejected_input = 0
+    expiry_capped = 0
+
+    def guard_dead_letter(ticket_id: int, reason: str) -> None:
+        try:
+            fts.dead_letter(ticket_id, actor=DEAD_LETTER_ACTOR, reason=reason)
+        except FTSError as exc:
+            print(f"dead_letter failed for ticket {ticket_id}: {exc}", file=sys.stderr)
+
+    for stage in cfg.stages:
+        free = min(stage.max_workers - active_of(stage.name), budget)
+        if free <= 0:
+            continue
+        ttl = stage.claim_ttl_ms or cfg.limits.claim_ttl_ms
+
+        if dry_run:
+            depth = depths.get(stage.name)
+            queued = depth.queued if depth is not None else 0
+            would = min(free, queued)
+            if would > 0:
+                print(f"[dry-run] would claim up to {would} ticket(s) in stage {stage.name}")
+                budget -= would
+            continue
+
+        for _ in range(free):
+            worker_id = f"{cfg.project}-{stage.name}-{uuid.uuid4().hex[:8]}"
+            try:
+                result = fts.claim(stage.name, worker_id, ttl_ms=ttl)
+            except FTSError as exc:
+                print(f"claim failed for stage {stage.name}: {exc}", file=sys.stderr)
+                break
+            if result is None:
+                break
+            ticket = result.ticket
+            ticket_dir = (repo_root / ticket.folder_path).resolve()
+            if zero_progress_expiries(fts, ticket.id) >= ZERO_PROGRESS_LIMIT:
+                guard_dead_letter(
+                    ticket.id, "3+ claims expired with no progress — inspect worker logs"
+                )
+                expiry_capped += 1
+                continue
+            if stage.name == first_stage:
+                request_md = ticket_dir / "request.md"
+                try:
+                    seeded = request_md.is_file() and bool(request_md.read_text().strip())
+                except OSError:
+                    seeded = False
+                if not seeded:
+                    guard_dead_letter(
+                        ticket.id, "missing request.md — seed the ticket folder and revive"
+                    )
+                    rejected_input += 1
+                    continue
+            briefing = compose_briefing(
+                ticket_id=ticket.id,
+                stage=stage,
+                db_path=db_path,
+                ticket_dir=ticket_dir,
+                repo_root=repo_root,
+                epoch=result.epoch,
+                worker_id=worker_id,
+                operator_session=cfg.operator_session,
+            )
+            session_id = spawn_worker(
+                cfg=cfg, stage=stage, ticket_id=ticket.id, briefing=briefing, repo_root=repo_root
+            )
+            if session_id is not None:
+                spawned.append(ticket.id)
+                if session_id:
+                    workers[worker_id] = {
+                        "session_id": session_id,
+                        "ticket_id": ticket.id,
+                        "epoch": result.epoch,
+                    }
+            budget -= 1
+            if budget <= 0:
+                break
+        if budget <= 0:
+            break
+    return spawned, rejected_input, expiry_capped, active_total
+
+
+def reap_workers(workers: dict[str, dict], ev) -> None:
+    """Apply one event to the worker map: reap (ftown remove) or prune the mapping.
+
+    Expiry events carry the exact worker_id; the other lifecycle events match by
+    ticket_id, guarded by claim epoch (ev.cursor > entry epoch) so a stale event can
+    never evict a session spawned later in the same or a following tick.
+    """
+    if ev.type is EventType.TICKET_CLAIM_EXPIRED:
+        wid = str(ev.payload.get("worker_id", ""))
+        entry = workers.pop(wid, None)
+        if entry and entry.get("session_id"):
+            ftown_remove(str(entry["session_id"]))
+        return
+    if ev.type in REAP_EVENTS:
+        stale = [
+            wid
+            for wid, e in workers.items()
+            if e.get("ticket_id") == ev.ticket_id and ev.cursor > int(e.get("epoch", 0))
+        ]
+        for wid in stale:
+            entry = workers.pop(wid)
+            if entry.get("session_id"):
+                ftown_remove(str(entry["session_id"]))
+
+
+def bridge_events(
+    fts: FTS, cfg: FactoryConfig, workers: dict[str, dict], *, repo_root: Path, dry_run: bool
+) -> int:
+    """Forward dead-letter/orphan events to the operator, reap/prune worker sessions on
+    lifecycle events; advance the cursor past all events."""
+    cursor_path = repo_root / CURSOR_FILE
+    try:
+        cursor = int(cursor_path.read_text().strip())
+    except (OSError, ValueError):
+        cursor = 0
+    events = fts.poll_events(after_cursor=cursor)
+    for ev in events:
+        if not dry_run:
+            reap_workers(workers, ev)
+        if ev.type not in FORWARD_EVENTS:
+            continue
+        msg = f"[{cfg.project}] {ev.type} ticket={ev.ticket_id} (cursor {ev.cursor})"
+        if dry_run or cfg.operator_session == "-":
+            print(f"[event] {msg}")
+        else:
+            ftown_tell(cfg.operator_session, msg)
+    if events and not dry_run:
+        cursor_path.write_text(str(events[-1].cursor))
+    return len(events)
+
+
+def run_tick(repo_root: Path, *, dry_run: bool) -> int:
+    yaml_path = repo_root / "factory" / "factory.yaml"
+    if not yaml_path.is_file():
+        print(f"error: {yaml_path} not found (run from the project repo root)", file=sys.stderr)
+        return 2
+    db_path = (repo_root / ".ffactory" / "factory.db").resolve()
+    if not db_path.is_file():
+        print(f"error: {db_path} not found (run `factory init` first)", file=sys.stderr)
+        return 2
+
+    cfg = load_config(yaml_path)
+    workers = load_workers(repo_root)
+    fts = FTS(db_path)
+    try:
+        requeued, promoted = run_scheduler(fts)
+        deadlettered = sweep_stragglers(fts, cfg)
+        spawned, rejected_input, expiry_capped, active_total = capacity_spawn(
+            fts, cfg, workers, repo_root=repo_root, db_path=db_path, dry_run=dry_run
+        )
+        events = bridge_events(fts, cfg, workers, repo_root=repo_root, dry_run=dry_run)
+    finally:
+        fts.close()
+    if not dry_run:
+        save_workers(repo_root, workers)
+
+    spawned_str = ",".join(str(i) for i in spawned) if spawned else "0"
+    print(
+        f"tick: requeued={requeued} promoted={promoted} spawned={spawned_str} "
+        f"rejected_input={rejected_input} expiry_capped={expiry_capped} "
+        f"deadlettered={deadlettered} events={events} "
+        f"active={active_total}/{cfg.limits.max_sessions}"
+    )
+    return 0
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="No-LLM factory dispatcher (one tick).")
+    parser.add_argument("--dry-run", action="store_true", help="Never claim or spawn; print plan.")
+    args = parser.parse_args()
+    repo_root = Path.cwd().resolve()
+    try:
+        return run_tick(repo_root, dry_run=args.dry_run)
+    except SystemExit:
+        raise
+    except Exception:
+        traceback.print_exc()
+        return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/skills/factory/factory-template/factory.yaml

@@ -0,0 +1,36 @@
+# Factory definition — checked into the project repo. Runtime state lives in
+# .ffactory/ (gitignored). Copied from ffactory/factory-template by `factory init`.
+
+project: my-project            # used to namespace ftown loops and worker names
+operator_session: "-"          # ftown session id for digests/worker mail; "-" = none
+
+limits:
+  max_sessions: 6              # global concurrent worker cap for this factory
+  bounce_limit: 3              # rejections before auto dead-letter (fts config)
+  claim_ttl_ms: 1800000        # default claim TTL (30m); stages may override
+  max_ticket_age_h: 48         # dispatcher dead-letters tickets older than this
+  dispatch_every: 30s          # dispatcher loop interval (skip-overlap)
+
+# Harness per stage: claude | codex | cursor | zai | kimi | deepseek | fireworks
+# next: where advance() sends the ticket on success (- = terminal)
+# bounce: where reject() sends it (- = rejection not allowed)
+stages:
+  - { name: groom,     harness: claude, model: sonnet, max_workers: 1, next: design,    bounce: "-" }
+  - { name: design,    harness: claude, model: opus,   max_workers: 1, next: implement, bounce: groom }
+  - { name: implement, harness: codex,                 max_workers: 3, next: review,    bounce: design,    claim_ttl_ms: 2700000 }
+  - { name: review,    harness: claude, model: opus,   max_workers: 2, next: qa,        bounce: implement }
+  - { name: qa,        harness: claude, model: sonnet, max_workers: 1, next: pr,        bounce: implement }
+  - { name: pr,        harness: claude, model: sonnet, max_workers: 1, next: "-",         bounce: implement, allow_push: true }
+
+resources:                     # shared surfaces workers must lease before touching
+  - { name: staging, mode_policy: exclusive_only }
+
+triage:
+  every: 10m                   # interval sweep; preflight skips (no session) when idle
+  harness: claude
+  model: sonnet
+
+digest:
+  cron: "0 9 * * *"            # daily shift report mailed to operator_session
+  harness: claude
+  model: sonnet

package/skills/factory/factory-template/skills/_protocol.md

@@ -0,0 +1,85 @@
+# Worker Protocol (read this before your stage skill — it is binding)
+
+You are an autonomous worker session inside a software factory. You were spawned to do
+ONE stage of ONE ticket, then exit. You do not choose what to work on; the dispatcher
+already claimed the ticket for you.
+
+## Your briefing variables
+
+Your spawn prompt defines these values. Refer to them exactly as named:
+
+- `TICKET_ID` — the ticket you own.
+- `STAGE` — the stage you are executing (matches your skill file).
+- `NEXT_STAGE` — where the ticket goes if you succeed (`-` means terminal).
+- `BOUNCE_STAGE` — where the ticket goes if you reject it (`-` means rejection not allowed at this stage).
+- `FTS_DB` — path to the factory database. Every `fts` command needs `--db "$FTS_DB"`.
+- `TICKET_DIR` — the ticket's artifact folder. NEVER move or rename it.
+- `REPO_ROOT` — the project repository root.
+- `EPOCH` — your claim fence token. Pass it to every `--epoch` flag shown below.
+- `WORKER_ID` — your identity, exactly as given here (the dispatcher derives it before
+  your session exists; use it verbatim, never substitute `$FTOWN_SESSION_ID`).
+  ($FTOWN_SESSION_ID is your session id — used only for the final self-close.)
+
+## The fts commands you are allowed to use
+
+Copy these shapes exactly; do not invent flags:
+
+```bash
+fts show   --db "$FTS_DB" "$TICKET_ID" --json                 # read ticket + history
+fts start  --db "$FTS_DB" --ticket "$TICKET_ID" --worker "$WORKER_ID" --epoch "$EPOCH"
+fts renew  --db "$FTS_DB" --ticket "$TICKET_ID" --worker "$WORKER_ID" --epoch "$EPOCH"
+fts acquire --db "$FTS_DB" --resource <name> --ticket "$TICKET_ID" --worker "$WORKER_ID" --mode exclusive
+fts release --db "$FTS_DB" --ticket "$TICKET_ID" --worker "$WORKER_ID" --resource <name> --epoch "$EPOCH"
+fts complete --db "$FTS_DB" --ticket "$TICKET_ID" --worker "$WORKER_ID" --epoch "$EPOCH" --note "<one line>"
+fts advance  --db "$FTS_DB" --ticket "$TICKET_ID" --worker "$WORKER_ID" --to-stage "$NEXT_STAGE" --note "<one line>"
+fts reject   --db "$FTS_DB" --ticket "$TICKET_ID" --worker "$WORKER_ID" --epoch "$EPOCH" \
+             --reason "<structured feedback, see your skill>" --to-stage "$BOUNCE_STAGE"
+```
+
+You may NOT use any other fts subcommand — unless your stage skill explicitly grants one
+(e.g. groom may `create`/`add-dep` to split work). Never open or modify the database file
+directly.
+
+## Lifecycle (always the same)
+
+1. `fts start` FIRST, before any work. If it fails, STOP and exit — you do not own this ticket.
+2. Read context: `fts show --json`, then read `TICKET_DIR` artifacts from earlier stages.
+3. Do the work your stage skill prescribes.
+4. **Heartbeat:** run `fts renew` after EVERY substantial step (any test run, file written,
+   long command). If renew ever fails (`ClaimExpired`, `ClaimFenced`, `NotClaimOwner`):
+   STOP IMMEDIATELY. Do not write anything more, do not "finish up". Another worker owns
+   the ticket now. Mail your parent that you were fenced, then self-close (step 7) and exit.
+5. Run your skill's GATE checklist. Every box must be checked with evidence, not assumed.
+6. Outcome — exactly one of:
+   - PASS: `fts complete` then (if `NEXT_STAGE` != `-`) `fts advance`.
+   - FAIL with actionable feedback and `BOUNCE_STAGE` != `-`: `fts reject` with the
+     structured reason your skill defines.
+   - STUCK (missing input, broken environment, contradictory requirements): do NOT
+     complete, do NOT reject. Mail your parent the exact blocker and exit; your claim
+     will expire and the ticket re-queues. Then self-close (step 7) — never leave your
+     session idling.
+7. Mail a result summary, then self-close. This applies to EVERY outcome — PASS, REJECT,
+   and STUCK alike — and the self-close command must be the very last command you run;
+   nothing after it executes:
+   ```bash
+   ~/.ftown/ftown-sessions tell --parent --type result \
+     "ticket $TICKET_ID $STAGE: <PASS|REJECTED|STUCK> — <one sentence>"
+   ~/.ftown/ftown-sessions remove "$FTOWN_SESSION_ID"   # self-close; your session ends here
+   ```
+
+## Hard rules
+
+- Write ONLY inside `TICKET_DIR` (and, for the implement stage, your assigned git
+  worktree). Never write to other tickets' folders, never touch `.ffactory/factory.db*`.
+- Artifacts accrete: never delete or rewrite an earlier stage's artifact. If you disagree
+  with it, that is a `reject` with reasons, not an edit.
+- Never move `TICKET_DIR` (paths are stable for the ticket's lifetime).
+- Never `git push`, create PRs, or call external services unless your stage skill
+  explicitly instructs it.
+- Never add credentials to code or artifacts.
+- If you touch a shared external surface (staging env, test database), `acquire` a lease
+  first and `release` when done. Acquire multiple resources in alphabetical order.
+- One ticket, one stage, one session. When your outcome is recorded: exit. Do not pick up
+  other work, do not "improve" unrelated code.
+- If the same error defeats you twice, your assumption is wrong: re-read the artifacts
+  and this protocol instead of retrying a third time; if still stuck, use the STUCK path.