typescript-ui-mockup-parity-worker-stewardship

Answer

Root cause. In hermes-canopy Tick 122 the worker commits code but never runs gitreins task complete, so tasks stall at committed. The naive foreman fix — invoking gitreins task complete — is wrong because that CLI path re-runs the 9-minute judge eval. The stewarded pattern is: run the judge once in the background with a timeout, resolve the CLI-vs-disk hash discrepancy by trusting the newest snapshot, write the completion record directly into tasks.yaml, and sync board-v2 with an explicit-id event INSERT + metadata.

Fix 1 — steward completion path (tested, /tmp/canopy/steward.py):

def trust_newest(cli_hash, disk_dir, cli_printed_at=None):
    """Trust the NEWEST snapshot: disk dir beats stale CLI hash iff disk was
    touched after the CLI printed its snapshot."""
    disk_hash, disk_mtime = None, 0.0
    if disk_dir.exists():
        newest = max(disk_dir.rglob("*"), key=lambda p: p.stat().st_mtime)
        disk_mtime, disk_hash = newest.stat().st_mtime, newest.name
    if cli_printed_at is None:
        cli_printed_at = time.time()
    if disk_hash and (disk_mtime >= cli_printed_at or not cli_hash):
        return {"hash": disk_hash, "source": "disk"}
    return {"hash": cli_hash, "source": "cli"}

def run_judge(task_id):
    proc = subprocess.Popen(f"timeout 900 gitreins judge {task_id}",
                            shell=True, stdout=subprocess.PIPE, text=True)
    out = "".join(proc.stdout.readlines())          # ~9 min in prod
    return {"exit": proc.returncode,
            "cli_hash": re.search(r"hash[:=]\s*([0-9a-f]{8,40})", out).group(1),
            "verdict": "PASS" if "PASS" in out else "FAIL",
            "acs": out.count("AC")}

def write_completion(task_id, rec):                 # idempotent direct write
    data = yaml.safe_load(TASKS_YAML.read_text()); tasks = data["tasks"]
    for t in tasks:
        if t.get("id") == task_id:
            if t.get("status") == "complete" and t["judge_hash"] == rec["judge_hash"]:
                return False                        # no-op re-run
            t.update(rec); break                    # supersede in place
    else:
        tasks.append({"id": task_id, **rec})
    TASKS_YAML.write_text(yaml.safe_dump(data, sort_keys=False)); return True

Fix 2 — board-v2 sync (UPDATE + explicit-id INSERT + metadata + abs-path COPY):

BEGIN IMMEDIATE;  -- serializes MAX(id) allocation across concurrent stewards
UPDATE tasks SET status='complete', judge_hash=?, verdict=?, acs=?, judged_at=datetime('now')
 WHERE id = ?;
INSERT INTO events(id, task_id, event_type, ticks, ts, metadata)
VALUES ((SELECT COALESCE(MAX(id),0)+1 FROM events), ?, 'task_complete',
        (SELECT COALESCE(MAX(ticks),0) FROM events WHERE task_id=?), datetime('now'),
        json_object('ticks_total', (SELECT MAX(ticks) FROM events WHERE task_id=?),
                    'cooldown', 120));            -- 120 pinned in fleet.toml, NOT stale 60
COPY board_tasks TO '/var/canopy/warehouse/tasks_UI-05.parquet';  -- absolute path only
COMMIT;

Fix 3 — UI-06 dispatch prompt (verified facts, no guessing): implement composer bar per mockup-1; handleSendMessage is a console.log stub → replace with apiPost('/api/nodes', { parent_id, content_format, node_type, content }, ...) — API contract is snake_case (parent_id/content_format/node_type); no upload endpoint exists, so send content as JSON body (base64 only if binary); use the existing apiPost helper rather than adding a fetch wrapper.