fix(gastown): PR-strategy MR beads stuck in_progress when github_token is missing — agents re-dispatched for completed work

[jrf0110]

Bug

When a town uses the PR merge strategy but townConfig.git_auth.github_token is not set (user configured a PAT in github_cli_pat instead, or uses the GitHub App integration token), the poll_pr side effect silently fails on every tick. MR beads with PR URLs stay in_progress permanently. This causes:

1. 4 MR beads stuck in_progress simultaneously (violating the "at most 1 per rig" invariant)
2. 604 invariant violations per hour (one per reconciler tick)
3. Reconciler polling 3-4 PRs every 5 seconds indefinitely, with ~1.5s wall clock spikes from the failed HTTP calls
4. Polecats being re-dispatched for already-completed work — source beads go back to open via orphan recovery (because MR beads are stuck), and the reconciler assigns new agents to them

Diagnostic Evidence

Analytics Engine data for town 8a6f9375-b806-4ee0-ad6e-1697ea2dbfff shows:

• 604 reconciler ticks in 1 hour, every one with invariant violations
• Most common action pattern: {"poll_pr": 3} or {"poll_pr": 4} — same PRs polled repeatedly
• 4 agent.dispatch_failed events — polecats dispatched to beads that are already done
• 0 bead.* lifecycle events — no beads are transitioning, everything is frozen
• Wall clock spikes to 1500-1976ms per tick (healthy is <100ms)

Root Cause

checkPRStatus() in Town.do.ts:3493-3583 reads townConfig.git_auth.github_token at line 3501. If this is null or empty, it logs a warning and returns null:

const token = townConfig.git_auth.github_token;
if (!token) {
  console.warn(`checkPRStatus: no github_token configured, cannot poll ${prUrl}`);
  return null;
}

Back in the poll_pr handler at actions.ts:566:

const status = await ctx.checkPRStatus(action.pr_url);
if (status && status !== 'open') {
  ctx.insertEvent('pr_status_changed', { ... });
}

null fails the truthiness check — no event is emitted. The MR bead stays in_progress. On the next tick, reconcileReviewQueue Rule 1 emits another poll_pr. The cycle repeats forever.

The user may have configured git credentials via github_cli_pat (the PAT text input in settings) rather than github_token (the OAuth token from #1350). Or they're using the GitHub App integration token which lives in a different field. checkPRStatus only looks at github_token.

Fix

Fix 1: Fall back to github_cli_pat when github_token is missing

const token = townConfig.git_auth.github_token ?? townConfig.github_cli_pat;
if (!token) {
  console.warn(`checkPRStatus: no GitHub token configured, cannot poll ${prUrl}`);
  return null;
}

The github_cli_pat is a valid GitHub PAT that can call the GitHub API. It's the same token the container uses for gh CLI operations.

Fix 2: Also try the platform integration token

If neither github_token nor github_cli_pat is set, try resolving the token from the rig's platform_integration_id (the GitHub App installation token). This requires looking up the integration and minting an installation token — more complex but covers the case where the user connected via GitHub App and never set a PAT.

Fix 3: Stop polling when token is missing

If no token is available, the poll_pr action should NOT silently retry every 5 seconds. Instead:

• Emit a pr_poll_failed event with reason no_token
• After N consecutive failures (e.g., 10), transition the MR bead to failed with a clear error message: "Cannot poll PR status — no GitHub token configured. Please add a token in town settings."
• Or escalate to the mayor: create an escalation bead that tells the user to configure their git credentials

Fix 4: Rate-limit PR polling

Even when the token is valid, polling 3-4 PRs every 5 seconds is excessive. PR status doesn't change that frequently. Add a per-MR-bead polling interval (e.g., poll each PR at most once per minute) tracked via last_poll_at on the review metadata.

Files

• src/dos/Town.do.ts — checkPRStatus() (line 3493-3583)
• src/dos/town/actions.ts — poll_pr handler (line 547-576)
• src/dos/town/reconciler.ts — reconcileReviewQueue Rule 1 (emits poll_pr)

Acceptance Criteria

• checkPRStatus falls back to github_cli_pat when github_token is missing
• PR polling rate-limited to once per minute per MR bead
• After N consecutive poll failures, MR bead fails with clear error message
• No more 600+ invariant violations per hour from stuck MR beads
• Reconciler wall clock returns to <100ms after fix