bug: discoverSessionBeads re-admits agents from suspended rigs into desired state

[manikmaxster]

Before you continue

• I searched existing issues and did not find a duplicate.
• I read the relevant docs and contributor guidance.

Gas City version

v13.4

Environment

ubuntu 24.04

Reproduction

1. Start a city with packs/gastown included in a rig's config
2. Confirm the rig's agents are running
3. Run gc rig suspend
4. Wait for the next reconciler tick
5. Observe that the agents are still alive

Expected behavior

After the reconciler tick, the suspended rig's agents should be drained and stopped.

Actual behavior

The agents remain running — they are not drained or stopped.

Logs, screenshots, or traces

Additional context

No response

[manikmaxster]

Even if I manually terminate the sessions they are recreated.

[rileywhite]

Investigation: unable to reproduce on current main

We investigated this with the intent to fix, but could not reproduce the behavior on current main (commit 5054185).

Test performed

1. Confirmed two rig-scoped agents were running with live Claude processes
2. Ran gc rig suspend <rig-name>
3. After ~60s (two patrol cycles at patrol_interval = "30s"), both sessions were drained and stopped — tmux sessions killed
4. Waited an additional patrol cycle — sessions stayed dead, not recreated (the specific concern raised in the follow-up comment)
5. Ran gc rig resume <rig-name> — agents restarted normally

Why it appears to be fixed

build_desired_state.go now has suspendedRigPaths checks at every point where agents are admitted to the desired state:

• Fixed agents (line 189)
• Pool agents (line 205)
• Assigned work bead collection (line 225)
• Pool demand realization (line 244)
• Named sessions (line 291)
• discoverSessionBeads (line 526) — the specific function called out in the issue title
• Dependency floor realization (line 591)

The agentInSuspendedRig helper (line 797-807) provides a consistent check across all paths. This comprehensive coverage means there's no remaining entry point through which a suspended rig's agents can re-enter the desired state.

Recommendation

This appears to be resolved. Suggest closing once a maintainer confirms.

[julianknutsen]

Thanks for testing the fixes @rileywhite and the original report @manikmaxster. We've changed this area a lot and I'm going to close this for now. Don't hesitate to reopen if we continue to see errors in the path moving forward.