docs: update poller dispatcher, PR state machine, agent chaining (closes #7)

Replace flag-file + heartbeat approach with the production dispatcher
pattern: poller triages notifications and spawns isolated agents
directly via openclaw cron. Adds assignment scan for self-created
issues. Response time ~15-60s instead of ~30 min.

AUTOMATED_DEV.md

@@ -74,103 +74,108 @@ back to issues.
 
 ### PR State Machine
 
-Once a PR exists, it enters a finite state machine tracked by Gitea labels and
+Once a PR exists, it enters a finite state machine tracked by Gitea labels. Each
-issue assignments. Labels represent the current state; the assignment field
+PR has exactly one state label at a time, plus a `bot` label indicating it's the
-represents who's responsible for the next action.
+agent's turn to act.
 
 #### States (Gitea Labels)
 
 | Label          | Color  | Meaning                                       |
-| -------------- | ------ | ------------------------------------------------- |
+| -------------- | ------ | --------------------------------------------- |
-| `needs-rebase` | red    | PR has merge conflicts or is behind main          |
+| `needs-review` | yellow | Code pushed, `docker build .` passes, awaiting review |
-| `needs-checks` | orange | `make check` does not pass cleanly                |
-| `needs-review` | yellow | Code review not yet done                          |
 | `needs-rework` | purple | Code review found issues that need fixing      |
-| `merge-ready`  | green  | All checks pass, reviewed, rebased, conflict-free |
+| `merge-ready`  | green  | Reviewed clean, build passes, ready for human  |
 
-#### Transitions
+Earlier iterations included `needs-rebase` and `needs-checks` states, but we
+eliminated them. Rebasing is handled inline by workers and reviewers (they
+rebase onto the target branch as part of their normal work). And `docker build .`
+is the only check — it's run by workers before pushing and by reviewers before
+approving. There's no separate "checks" phase.
+
+#### The `bot` Label + Assignment Model
+
+The `bot` label signals that an issue or PR is the agent's turn to act. The
+assignment field tracks who is actively working on it:
+
+- **`bot` label + unassigned** = work available, poller dispatches an agent
+- **`bot` label + assigned to agent** = actively being worked
+- **No `bot` label** = not the agent's turn (either human's turn or done)
+
+The notification poller assigns the agent account to the issue at dispatch time,
+before the agent session even starts. This prevents race conditions — by the
+time a second poller scan runs, the issue is already assigned and gets skipped.
+
+When the agent finishes its step and spawns the next agent, it unassigns itself
+first (releasing the lock). The next agent's first action is to verify it's the
+only one working on the issue by checking comments for duplicate work.
+
+At chain-end (`merge-ready`): the agent assigns the human and removes the `bot`
+label. The human's PR inbox contains only PRs that are genuinely ready to merge.
+
+#### Agent Chaining — No Self-Review
+
+Each step in the pipeline is handled by a separate, isolated agent session.
+Agents spawn the next agent in the chain via `openclaw cron add --session
+isolated`. This enforces a critical rule: **the agent that wrote the code never
+reviews it.**
+
+The chain looks like this:
 
 ```
-New PR created
+Worker agent (writes/fixes code)
-  │
+  → docker build . → push → label needs-review
-  ▼
+  → unassign self → spawn reviewer agent → STOP
-[needs-rebase] ──rebase onto main──▶ [needs-checks]
+
-  ▲                                        │
+Reviewer agent (reviews code it didn't write)
-  │                                   run make check
+  → read diff + referenced issues → review
-  │ (main updated,                         │
+  → PASS: rebase if needed → docker build . → label merge-ready
-  │  conflicts)              ┌─────────────┴──────────────┐
+          → assign human → remove bot label → STOP
-  │                          │                              │
+  → FAIL: comment findings → label needs-rework
-  │                       passes                          fails
+          → unassign self → spawn worker agent → STOP
-  │                          │                              │
-  │                          ▼                              ▼
-  │                   [needs-review]                  [needs-checks]
-  │                          │                        (fix code, re-run)
-  │                     code review
-  │                          │
-  │                ┌─────────┴──────────┐
-  │                │                      │
-  │             approved             issues found
-  │                │                      │
-  │                ▼                      ▼
-  │          [merge-ready]          [needs-rework]
-  │                │                      │
-  │           assign human           fix issues
-  │                                       │
-  │                                       ▼
-  └───────────────────────────── [needs-rebase]
-                                 (restart cycle)
 ```
 
-The cycle can repeat multiple times: rebase → check → review → rework → rebase →
+The cycle repeats (worker → reviewer → worker → reviewer → ...) until the
-check → review → rework → ... until the PR is clean. Each iteration typically
+reviewer approves. Each agent is a fresh session with no memory of previous
-addresses a smaller set of issues until everything converges.
+iterations — it reads the issue comments and PR diff to understand context.
 
-#### Assignment Rules
+#### TOCTOU Protection
 
-- **PR in any state except `merge-ready`** → assigned to the agent. It's the
+Just before changing labels or assignments, agents re-read all comments and
-  agent's job to drive it forward through the state machine.
+current labels via the API. If the state changed since they started (another
-- **PR reaches `merge-ready`** → assigned to the human. This is the ONLY time a
+agent already acted), they report the conflict and stop. This prevents stale
-  PR should land in the human's queue.
+agents from overwriting fresh state.
-- **Human requests changes during review** → PR moves back to `needs-rework`,
-  reassigned to agent.
 
-This means the human's PR inbox contains only PRs that are genuinely ready to
+#### Race Detection
-merge — no half-finished work, no failing CI, no merge conflicts. Everything
+
-else is the agent's problem.
+If an agent starts and finds its work was already done (e.g., a reviewer sees a
+review was already posted, or a worker sees a PR was already created), it
+reports to the status channel and stops.
 
 #### The Loop in Practice
 
-A typical PR might go through this cycle:
+A typical PR goes through this cycle:
 
-1. Agent creates PR, labels `needs-rebase`
+1. Worker agent creates PR, runs `docker build .`, labels `needs-review`
-2. Agent rebases onto main → labels `needs-checks`
+2. Worker spawns reviewer agent
-3. Agent runs `make check` — lint fails → fixes lint, pushes → back to
+3. Reviewer reads diff — finds a missing error check → labels `needs-rework`
-   `needs-rebase` (new commit)
+4. Reviewer spawns worker agent
-4. Agent rebases → `needs-checks` → runs checks → passes → `needs-review`
+5. Worker fixes the error check, rebases, runs `docker build .`, labels
-5. Agent does code review — finds a missing error check → `needs-rework`
+   `needs-review`
-6. Agent fixes the error check, pushes → `needs-rebase`
+6. Worker spawns reviewer agent
-7. Agent rebases → `needs-checks` → passes → `needs-review`
+7. Reviewer reads diff — looks good → rebases → `docker build .` → labels
-8. Agent reviews — looks good → `merge-ready`
+   `merge-ready`, assigns human
-9. Agent assigns to human
+8. Human reviews, merges
-10. Human reviews, merges
 
-Steps 1-9 happen without human involvement. The human sees a clean, reviewed,
+Steps 1-7 happen without human involvement. Each step is a separate agent
-passing PR ready for a final look.
+session that spawns the next one.
 
-#### Automated Sweep
+#### Safety Net
 
-A periodic cron job (every 4 hours) scans all open PRs across all repos:
+The notification poller runs a periodic scan (every 2 minutes) of all watched
-
+repos for issues/PRs with the `bot` label that are unassigned. This catches
-- **No label** → classify into the correct state
+broken chains — if an agent crashes or times out without spawning the next agent,
-- **`needs-rebase`** → spawn agent to rebase
+the poller will eventually re-dispatch. A 30-minute cooldown prevents duplicate
-- **`needs-checks`** → spawn agent to run checks and fix failures
+dispatches during normal operation.
-- **`needs-review`** → spawn agent to do code review
-- **`needs-rework`** → spawn agent to fix review feedback
-- **`merge-ready`** → verify still true (main may have updated since), ensure
-  assigned to human
-
-This catches PRs that fell through the cracks — an agent session that timed out
-mid-rework, a rebase that became necessary when main moved forward, etc.
 
 #### Why Labels + Assignments
 
@@ -263,26 +268,45 @@ A practical setup:
 - **DM with agent** — Private conversation, sitreps, sensitive commands
 - **Project-specific channels** — For coordination with external collaborators
 
-### The Notification Poller
+### The Notification Poller + Dispatcher
 
 Because the agent can't see Gitea webhooks in Mattermost (bot-to-bot visibility
-issue), we built a lightweight Python script that polls the Gitea notifications
+issue), we built a Python script that both polls and dispatches. It polls the
-API every 2 seconds and wakes the agent via OpenClaw's `/hooks/wake` endpoint
+Gitea notifications API every 15 seconds, triages each notification (checking
-when new notifications arrive.
+@-mentions and assignment), marks them as read, and spawns one isolated agent
+session per actionable item via `openclaw cron add --session isolated`.
+
+The poller also runs a secondary **label scan** every 2 minutes, checking all
+watched repos for open issues/PRs with the `bot` label that are unassigned
+(meaning they need work but no agent has claimed them yet). This catches cases
+where the agent chain broke — an agent timed out or crashed without spawning the
+next one.
 
 Key design decisions:
 
-- **The poller never marks notifications as read.** That's the agent's job after
+- **The poller IS the dispatcher.** No flag files, no heartbeat dependency. The
-  processing. Prevents the poller and agent from racing.
+  poller triages notifications and spawns agents directly.
-- **Tracks notification IDs, not counts.** Only fires on genuinely new
+- **Marks notifications as read immediately.** Prevents re-dispatch on the next
-  notifications, not re-reads of existing ones.
+  poll cycle.
-- **The wake message tells the agent to route output to Gitea/Mattermost, not
+- **Assigns the agent account at dispatch time.** Before spawning the agent
-  DM.** Prevents chatty notification processing from disturbing the human.
+  session, the poller assigns the bot user to the issue via API. This prevents
-- **Zero dependencies.** Python stdlib only (`urllib`, `json`, `time`). Runs
+  race conditions — subsequent scans skip assigned issues.
-  anywhere.
+- **Dispatched issues are tracked in a persistent JSON file.** Survives poller
+  restarts. Entries auto-prune after 1 hour.
+- **30-minute re-dispatch cooldown.** The poller won't re-dispatch for the same
+  issue within 30 minutes, even if it appears unassigned again.
+- **Concurrency cap.** The poller checks how many agents are currently running
+  and defers dispatch if the cap is reached.
+- **Stale agent reaper.** Kills agent sessions that have been running longer
+  than 10 minutes (the `--timeout-seconds` flag isn't always enforced).
+- **`bot` label + `merge-ready` skip.** The label scan skips issues that are
+  already labeled `merge-ready` — those are in the human's court.
+- **Zero dependencies.** Python stdlib only. Runs anywhere.
+
+Response time: ~15-30 seconds from notification to agent starting work.
 
 Full source code is available in
-[OPENCLAW_TRICKS.md](OPENCLAW_TRICKS.md#the-gitea-notification-poller).
+[OPENCLAW_TRICKS.md](OPENCLAW_TRICKS.md#gitea-integration--notification-polling).
 
 ## CI: Gitea Actions
 
@@ -371,42 +395,34 @@ Everything gets a production URL with automatic TLS via Traefik.
 Putting it all together, the development lifecycle looks like this:
 
 ```
-1. Issue filed in Gitea (by human or agent)
+1. Human labels issue with `bot` (or agent files issue)
      ↓
-2. Agent picks up the issue (via notification poller)
+2. Poller detects `bot` label + unassigned → assigns agent → spawns worker
      ↓
-3. Agent posts "starting work on #N" to Mattermost #git
+3. Worker agent clones repo, writes code, runs `docker build .`
      ↓
-4. Agent (or sub-agent) creates branch, writes code, pushes
+4. Worker creates PR "(closes #N)", labels `needs-review`
      ↓
-5. Gitea webhook fires → #git shows the push
+5. Worker spawns reviewer agent → stops
      ↓
-6. CI runs docker build → passes or fails
+6. Reviewer agent reads diff + referenced issues → reviews
      ↓
-7. Agent creates PR "(closes #N)"
+7a. Review PASS → reviewer rebases if needed → `docker build .`
+    → labels `merge-ready` → assigns human → removes `bot`
      ↓
-8. Gitea webhook fires → #git shows the PR
+7b. Review FAIL → reviewer labels `needs-rework`
+    → spawns worker agent → back to step 3
      ↓
-9. Agent reviews code, runs make check locally, verifies
+8. Human reviews, merges
      ↓
-10. Agent assigns PR to human when all checks pass
+9. Gitea webhook fires → µPaaS deploys to production
      ↓
-11. Human reviews, requests changes or approves
+10. Site/service is live
-     ↓
-12. If changes requested → agent reworks, back to step 6
-     ↓
-13. Human merges PR
-     ↓
-14. Gitea webhook fires → µPaaS deploys to production
-     ↓
-15. Gitea webhook fires → #git shows the merge
-     ↓
-16. Site/service is live on production URL
 ```
 
-Steps 2-10 can happen without any human involvement. The human's role is reduced
+Steps 2-7 happen without any human involvement, driven by agent-to-agent
-to: review the PR, approve or request changes, merge. Everything else is
+chaining. The human's role is reduced to: label the issue, review the final PR,
-automated.
+merge. Everything else is automated.
 
 ### Observability
 

OPENCLAW_TRICKS.md

@@ -189,50 +189,86 @@ arrive instantly.
 Setup: add a webhook on each Gitea repo (or use an organization-level webhook)
 pointing to `https://your-openclaw-host/hooks/gitea`. OpenClaw handles the rest.
 
-#### Option B: Notification Poller (Local Machine Behind NAT)
+#### Option B: Notification Poller + Dispatcher (Local Machine Behind NAT)
 
 If your OpenClaw runs on a dedicated local machine behind NAT (like a home Mac
 or Linux workstation), Gitea can't reach it directly. This is our setup —
 OpenClaw runs on a Mac Studio on a home LAN.
 
-The solution: a lightweight Python script that polls the Gitea notifications API
+The solution: a Python script that both polls and dispatches. It polls the Gitea
-every few seconds. When new notifications appear, it writes a flag file that the
+notifications API every 15 seconds, triages each notification (checking
-agent checks during heartbeats.
+@-mentions and assignments), marks them as read, and spawns one isolated agent
+session per actionable item via `openclaw cron add --session isolated`.
+
+The poller also runs a secondary **label scan** every 2 minutes, checking all
+watched repos for open issues/PRs with the `bot` label that are unassigned. This
+catches cases where the agent chain broke — an agent timed out or crashed
+without spawning the next agent. It also picks up newly-labeled issues that
+didn't trigger a notification.
 
 Key design decisions:
 
-- **The poller never marks notifications as read.** The agent does that after
+- **The poller IS the dispatcher.** No flag files, no heartbeat dependency. The
-  processing. This prevents lost notifications if the agent fails to process.
+  poller triages notifications and spawns agents directly.
-- **It tracks notification IDs, not counts.** Only fires on genuinely new
+- **Marks notifications as read immediately.** Prevents re-dispatch on the next
-  notifications, not re-reads of existing ones.
+  poll cycle.
-- **Flag file instead of wake events.** We initially used OpenClaw's
+- **Assigns the bot user at dispatch time.** Before spawning the agent, the
-  `/hooks/wake` endpoint, but wake events target the main (DM) session — any
+  poller assigns the bot account to the issue via API. This prevents race
-  model response during processing leaked to DM as a notification. The flag file
+  conditions — subsequent scans skip assigned issues. The spawned agent doesn't
-  approach is processed during heartbeats, where output routing is controlled.
+  need to claim ownership; it's already claimed.
-- **Zero dependencies.** Just Python stdlib. Runs anywhere.
+- **Persistent dispatch tracking.** Dispatched issues are tracked in a JSON
+  file on disk (not just in memory), surviving poller restarts. Entries
+  auto-prune after 1 hour.
+- **30-minute re-dispatch cooldown.** Safety net for broken agent chains. Normal
+  operation uses agent-to-agent chaining (each agent spawns the next), so the
+  poller only re-dispatches if the chain breaks.
+- **Concurrency cap.** The poller checks how many agents are currently running
+  (`openclaw cron list`) and defers dispatch if the cap is reached.
+- **Stale agent reaper.** Each scan cycle, kills agent sessions running longer
+  than 10 minutes. The `--timeout-seconds` flag isn't always enforced by
+  OpenClaw, so the poller handles cleanup itself.
+- **`merge-ready` skip.** The label scan skips issues already labeled
+  `merge-ready` — those are in the human's court.
+- **Template-based prompts.** The poller reads two workspace files (a dispatch
+  header with `{{variable}}` placeholders, and a workflow rules document),
+  concatenates them, substitutes variables, and passes the result as the
+  agent's `--message`. This keeps all instructions in version-controlled
+  workspace files with a single source of truth.
+- **Zero dependencies.** Python stdlib only. Runs anywhere.
 
-Tradeoff: notifications are processed at heartbeat cadence (~30 min) instead of
+Response time: ~15–30s from notification to agent starting work.
-realtime. For code review and issue triage, this is fine.
 
 ```python
 #!/usr/bin/env python3
 """
-Gitea notification poller (flag-file approach).
+Gitea notification poller + dispatcher.
-Polls for unread notifications and writes a flag file when new ones
+
-appear. The agent checks this flag during heartbeats and processes
+Two polling loops:
-notifications via the Gitea API directly.
+1. Notification-based: detects new @-mentions and assignments, dispatches
+   agents for actionable notifications.
+2. Label-based: periodically scans for issues/PRs with the 'bot' label
+   that are unassigned (available for work). Catches broken agent chains
+   and newly-labeled issues.
+
+The poller assigns the bot user to the issue BEFORE spawning the agent,
+preventing race conditions where multiple scans dispatch for the same issue.
 
 Required env vars:
   GITEA_URL        - Gitea instance URL
   GITEA_TOKEN      - Gitea API token
 
 Optional env vars:
-  FLAG_PATH   - Path to flag file (default: workspace/memory/gitea-notify-flag)
+  POLL_DELAY       - Seconds between notification polls (default: 15)
-  POLL_DELAY  - Delay between polls in seconds (default: 5)
+  COOLDOWN         - Seconds between dispatch batches (default: 30)
+  BOT_SCAN_INTERVAL - Seconds between label scans (default: 120)
+  MAX_CONCURRENT_AGENTS - Max simultaneous agents (default: 10)
+  REAP_AGE_SECONDS - Kill agents older than this (default: 600)
+  OPENCLAW_BIN     - Path to openclaw binary
 """
 
 import json
 import os
+import subprocess
 import sys
 import time
 import urllib.request
@@ -240,62 +276,270 @@ import urllib.error
 
 GITEA_URL = os.environ.get("GITEA_URL", "").rstrip("/")
 GITEA_TOKEN = os.environ.get("GITEA_TOKEN", "")
-POLL_DELAY = int(os.environ.get("POLL_DELAY", "5"))
+POLL_DELAY = int(os.environ.get("POLL_DELAY", "15"))
-FLAG_PATH = os.environ.get(
+COOLDOWN = int(os.environ.get("COOLDOWN", "30"))
-    "FLAG_PATH",
+BOT_SCAN_INTERVAL = int(os.environ.get("BOT_SCAN_INTERVAL", "120"))
-    os.path.join(
+MAX_CONCURRENT_AGENTS = int(os.environ.get("MAX_CONCURRENT_AGENTS", "10"))
-        os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+REAP_AGE_SECONDS = int(os.environ.get("REAP_AGE_SECONDS", "600"))
-        "memory",
+REDISPATCH_COOLDOWN = 1800  # 30 min safety net for broken agent chains
-        "gitea-notify-flag",
+OPENCLAW_BIN = os.environ.get("OPENCLAW_BIN", "openclaw")
-    ),
+BOT_USER = os.environ.get("BOT_USER", "clawbot")
+
+WORKSPACE = os.path.expanduser("~/.openclaw/workspace")
+DISPATCH_HEADER = os.path.join(
+    WORKSPACE, "taskprompts", "how-to-handle-gitea-notifications.md"
+)
+WORKFLOW_DOC = os.path.join(
+    WORKSPACE, "taskprompts", "how-to-work-on-a-gitea-issue-or-pr.md"
+)
+DISPATCH_STATE_PATH = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)), ".dispatch-state.json"
 )
 
+# Repos to watch for bot-labeled issues
+WATCHED_REPOS = [
+    # "org/repo1",
+    # "org/repo2",
+]
+
+# Dispatch tracking (persisted to disk)
+dispatched_issues: dict[str, float] = {}
+
+
+def _load_dispatch_state() -> dict[str, float]:
+    try:
+        with open(DISPATCH_STATE_PATH) as f:
+            state = json.load(f)
+        now = time.time()
+        return {k: v for k, v in state.items() if now - v < 3600}
+    except (FileNotFoundError, json.JSONDecodeError):
+        return {}
+
+
+def _save_dispatch_state():
+    try:
+        with open(DISPATCH_STATE_PATH, "w") as f:
+            json.dump(dispatched_issues, f)
+    except OSError as e:
+        print(f"WARN: Could not save dispatch state: {e}", file=sys.stderr)
+
+
+def gitea_api(method, path, data=None):
+    url = f"{GITEA_URL}/api/v1{path}"
+    body = json.dumps(data).encode() if data else None
+    headers = {"Authorization": f"token {GITEA_TOKEN}"}
+    if body:
+        headers["Content-Type"] = "application/json"
+    req = urllib.request.Request(url, headers=headers, method=method, data=body)
+    try:
+        with urllib.request.urlopen(req, timeout=15) as resp:
+            raw = resp.read()
+            return json.loads(raw) if raw else None
+    except Exception as e:
+        print(f"WARN: {method} {path}: {e}", file=sys.stderr, flush=True)
+        return None
+
+
+def load_template() -> str:
+    """Load dispatch header + workflow doc, concatenated."""
+    parts = []
+    for path in [DISPATCH_HEADER, WORKFLOW_DOC]:
+        try:
+            with open(path) as f:
+                parts.append(f.read())
+        except FileNotFoundError:
+            print(f"ERROR: File not found: {path}", file=sys.stderr)
+            sys.exit(1)
+    return "\n\n---\n\n".join(parts)
+
+
+def render_template(template, repo_full, issue_number, title,
+                    subject_type, reason):
+    return (
+        template
+        .replace("{{repo_full}}", repo_full)
+        .replace("{{issue_number}}", str(issue_number))
+        .replace("{{title}}", title)
+        .replace("{{subject_type}}", subject_type)
+        .replace("{{reason}}", reason)
+        .replace("{{gitea_url}}", GITEA_URL)
+        .replace("{{gitea_token}}", GITEA_TOKEN)
+        .replace("{{openclaw_bin}}", OPENCLAW_BIN)
+        .replace("{{bot_user}}", BOT_USER)
+        # Add your own variables here (e.g. git_channel)
+    )
+
+
+def count_running_agents() -> int:
+    try:
+        result = subprocess.run(
+            [OPENCLAW_BIN, "cron", "list"],
+            capture_output=True, text=True, timeout=10,
+        )
+        return sum(1 for line in result.stdout.splitlines()
+                   if "running" in line or "idle" in line)
+    except Exception:
+        return 0
+
+
+def spawn_agent(template, repo_full, issue_number, title,
+                subject_type, reason):
+    dispatch_key = f"{repo_full}#{issue_number}"
+    last = dispatched_issues.get(dispatch_key)
+    if last and (time.time() - last) < REDISPATCH_COOLDOWN:
+        return
+
+    if count_running_agents() >= MAX_CONCURRENT_AGENTS:
+        print(f"    → Concurrency limit reached, deferring {dispatch_key}",
+              flush=True)
+        return
+
+    dispatched_issues[dispatch_key] = time.time()
+
+    # Assign bot user immediately to prevent races
+    gitea_api("PATCH", f"/repos/{repo_full}/issues/{issue_number}",
+              {"assignees": [BOT_USER]})
+
+    repo_short = repo_full.split("/")[-1]
+    job_name = f"gitea-{repo_short}-{issue_number}-{int(time.time())}"
+    msg = render_template(template, repo_full, issue_number, title,
+                          subject_type, reason)
+
+    try:
+        result = subprocess.run(
+            [OPENCLAW_BIN, "cron", "add",
+             "--name", job_name, "--at", "1s",
+             "--message", msg, "--delete-after-run",
+             "--session", "isolated", "--no-deliver",
+             "--thinking", "low", "--timeout-seconds", "300"],
+            capture_output=True, text=True, timeout=15,
+        )
+        if result.returncode == 0:
+            _save_dispatch_state()
+        else:
+            dispatched_issues.pop(dispatch_key, None)
+    except Exception as e:
+        print(f"Spawn error: {e}", file=sys.stderr, flush=True)
+        dispatched_issues.pop(dispatch_key, None)
+
+
+def is_actionable(notif):
+    """Check if a notification warrants spawning an agent."""
+    subject = notif.get("subject", {})
+    repo = notif.get("repository", {})
+    repo_full = repo.get("full_name", "")
+    url = subject.get("url", "")
+    number = url.rstrip("/").split("/")[-1] if url else ""
+    if not number or not number.isdigit():
+        return False, "no issue number", None
+
+    issue = gitea_api("GET", f"/repos/{repo_full}/issues/{number}")
+    if not issue:
+        return False, "couldn't fetch issue", number
+
+    # Check for @-mentions in the latest comment
+    comments = gitea_api(
+        "GET", f"/repos/{repo_full}/issues/{number}/comments"
+    )
+    if comments:
+        last = comments[-1]
+        if last.get("user", {}).get("login") == BOT_USER:
+            return False, "own comment is latest", number
+        if f"@{BOT_USER}" in (last.get("body") or ""):
+            return True, "@-mentioned in comment", number
+
+    # Check for @-mention in issue body
+    body = issue.get("body", "") or ""
+    if f"@{BOT_USER}" in body:
+        return True, "@-mentioned in body", number
+
+    return False, "not mentioned", number
+
+
+def scan_bot_labeled(template):
+    """Scan for issues/PRs with 'bot' label that are unassigned."""
+    for repo_full in WATCHED_REPOS:
+        for issue_type in ["issues", "pulls"]:
+            items = gitea_api(
+                "GET",
+                f"/repos/{repo_full}/issues?state=open&type={issue_type}"
+                f"&labels=bot&sort=updated&limit=10",
+            ) or []
+            for item in items:
+                number = str(item["number"])
+                dispatch_key = f"{repo_full}#{number}"
+
+                last = dispatched_issues.get(dispatch_key)
+                if last and (time.time() - last) < REDISPATCH_COOLDOWN:
+                    continue
+
+                assignees = [
+                    a.get("login", "") for a in item.get("assignees") or []
+                ]
+                if BOT_USER in assignees:
+                    continue
+
+                labels = [
+                    l.get("name", "") for l in item.get("labels") or []
+                ]
+                if "merge-ready" in labels:
+                    continue
+
+                kind = "PR" if issue_type == "pulls" else "issue"
+                spawn_agent(
+                    template, repo_full, number,
+                    item.get("title", "")[:60],
+                    "pull" if issue_type == "pulls" else "issue",
+                    "bot label, unassigned",
+                )
+
+
+def main():
+    global dispatched_issues
+    dispatched_issues = _load_dispatch_state()
 
-def check_config():
     if not GITEA_URL or not GITEA_TOKEN:
         print("ERROR: GITEA_URL and GITEA_TOKEN required", file=sys.stderr)
         sys.exit(1)
 
+    template = load_template()
+    print(f"Poller started (poll={POLL_DELAY}s, cooldown={COOLDOWN}s, "
+          f"bot_scan={BOT_SCAN_INTERVAL}s, repos={len(WATCHED_REPOS)})",
+          flush=True)
 
-def gitea_unread_ids():
+    seen_ids = set(
-    req = urllib.request.Request(
+        n["id"] for n in
-        f"{GITEA_URL}/api/v1/notifications?status-types=unread",
+        (gitea_api("GET", "/notifications?status-types=unread") or [])
-        headers={"Authorization": f"token {GITEA_TOKEN}"},
     )
-    try:
+    last_dispatch = 0
-        with urllib.request.urlopen(req, timeout=10) as resp:
+    last_bot_scan = 0
-            return {n["id"] for n in json.loads(resp.read())}
-    except Exception as e:
-        print(f"WARN: Gitea API failed: {e}", file=sys.stderr, flush=True)
-        return set()
-
-
-def write_flag(count):
-    os.makedirs(os.path.dirname(FLAG_PATH), exist_ok=True)
-    with open(FLAG_PATH, "w") as f:
-        f.write(json.dumps({
-            "ts": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
-            "count": count,
-        }))
-
-
-def main():
-    check_config()
-    print(f"Gitea poller started (delay={POLL_DELAY}s, flag={FLAG_PATH})", flush=True)
-    last_seen_ids = gitea_unread_ids()
-    print(f"Initial unread: {len(last_seen_ids)}", flush=True)
 
     while True:
         time.sleep(POLL_DELAY)
-        current_ids = gitea_unread_ids()
+        now = time.time()
-        new_ids = current_ids - last_seen_ids
+
-        if not new_ids:
+        # --- Notification polling ---
-            last_seen_ids = current_ids
+        notifs = gitea_api("GET", "/notifications?status-types=unread") or []
-            continue
+        current_ids = {n["id"] for n in notifs}
-        ts = time.strftime("%H:%M:%S")
+        new_ids = current_ids - seen_ids
-        print(f"[{ts}] {len(new_ids)} new ({len(current_ids)} total), flag written", flush=True)
+        if new_ids and now - last_dispatch >= COOLDOWN:
-        write_flag(len(new_ids))
+            for n in [n for n in notifs if n["id"] in new_ids]:
-        last_seen_ids = current_ids
+                nid = n.get("id")
+                if nid:
+                    gitea_api("PATCH", f"/notifications/threads/{nid}")
+                is_act, reason, num = is_actionable(n)
+                if is_act:
+                    repo = n["repository"]["full_name"]
+                    title = n["subject"]["title"][:60]
+                    stype = n["subject"].get("type", "").lower()
+                    spawn_agent(template, repo, num, title, stype, reason)
+            last_dispatch = now
+        seen_ids = current_ids
+
+        # --- Bot label scan (less frequent) ---
+        if now - last_bot_scan >= BOT_SCAN_INTERVAL:
+            scan_bot_labeled(template)
+            last_bot_scan = now
 
 
 if __name__ == "__main__":
@@ -640,25 +884,27 @@ From REPO_POLICIES.md and our operational experience:
 
 #### The PR Pipeline
 
-Our agent follows a strict PR lifecycle:
+Our agent follows a strict PR lifecycle using agent-to-agent chaining. Each step
+is handled by a separate, isolated agent session — the agent that writes code
+never reviews it:
 
 ```markdown
 ## PR pipeline (every PR, no exceptions)
 
-1. **Review/rework loop**: code review → rework → re-review → repeat until clean
+Worker agent → docker build . → push → label needs-review → spawn reviewer
-2. **Check/rework loop**: `make check` + `docker build .` → rework → re-check →
+Reviewer agent → review diff → PASS: docker build . → label merge-ready
-   repeat until clean
+                              → FAIL: label needs-rework → spawn worker
-3. Only after BOTH loops pass with zero issues: assign to human
+Repeat until reviewer approves.
 
-- "Passes checks" ≠ "ready for human"
+- docker build . is the ONLY authoritative check (runs make check inside)
 - Never weaken tests/linters. Fix the code.
 - Pre-existing failures are YOUR problem. Fix them as part of your PR.
 ```
 
-The agent doesn't just create a PR and hand it off — it drives the PR through
+The agent chain doesn't just create a PR and hand it off — it drives the PR
-review, rework, and verification until it's genuinely ready. A PR assigned to
+through review, rework, and verification until it's genuinely ready. A PR
-the human means: all checks pass, code reviewed, review feedback addressed,
+assigned to the human means: build passes, code reviewed by a separate agent,
-rebased against main, no conflicts. Anything less is the agent's open task.
+review feedback addressed, rebased. Anything less is still in the agent chain.
 
 #### New Repo Bootstrap
 
@@ -1510,12 +1756,12 @@ For complex coding tasks, spawn isolated sub-agents.
 
 ### Sub-Agent PR Quality Gate (MANDATORY)
 
-- `make check` must pass with ZERO failures. No exceptions.
+- `docker build .` must pass. This is identical to CI and the only
+  authoritative check. No exceptions.
 - Pre-existing failures are YOUR problem. Fix them as part of your PR.
 - NEVER modify linter config to make checks pass. Fix the code.
-- Every PR must include full `make check` output
 - Rebase before and after committing
-- Never self-review
+- Never self-review — each agent spawns a separate agent for review
 ```
 
 ---