[jfro]

Closes #39

Adds a Mark Unavailable / Mark Available toggle for agents, on the dashboard's agent list and in the agent page header, for maintenance like updating VM images.

What an unavailable agent does

• Gets no new runners. Its free capacity counts as 0, so on-demand jobs (webhook and polling) pass it over, and fixed-capacity pools leave it out (both the target and the runner count).
• Idle fixed-capacity runners on it are drained. On each reconcile the runner is unregistered from the forge first, and only then is DestroyRunner sent for its VM. GitHub refuses to unregister a runner that's running a job (the busy flag, or a 422 on the delete), so a job can't land on a runner while it's being destroyed.
• Busy runners finish their job, then exit like any ephemeral runner and aren't replaced. Runners that aren't registered yet get 10 minutes to finish booting: one that registers in time is drained as idle, and one that never registers is destroyed. Drained runners aren't recorded as agent failures. Runners made for a specific job (webhook and polling modes) are left alone.
• Stays unavailable across restarts. The flag is a new registered_agents.unavailable_at column. The registry loads it at startup and applies it inside register(), under the agents lock, so an agent restarting mid-maintenance never gets work in between.
• Marking it available again wakes the fleet manager, so queued jobs and pool top-ups go out right away instead of on the next tick.

UI / API

• Status column and agent page show an Unavailable badge. The agent page also says since when.

• A pending job whose only matching agents are unavailable shows Agents unavailable, not "Agents busy".

• /api/v1/agents returns unavailable and unavailable_since for each agent. /api/v1/status returns unavailable_agents.

• Toggling is logged with the agent, the admin user, and its active runner count. The "No agent available" warning now includes unavailable= for each agent.

• GitHub runner lookups now filter by name. They used to read only the first page of runners (30), which also broke remove_runner for scopes with more runners than that.

Not in this PR

• No kuiper-forge agent CLI subcommand for this yet. It's dashboard and DB only.

Testing

• New unit tests: the registry (no selection, capacity, or pools while unavailable, and the flag surviving reconnects) and the auth store (round-trip; status syncs don't clear it).
• cargo test -p kuiper-forge passes. Clippy is clean on sqlite and postgres.
• Manual run: dry-run coordinator in webhook mode with two test agents.
  • Marked both unavailable and sent a queued workflow_job webhook. The job showed "Agents unavailable", and the log listed unavailable=true for each agent.
  • Marked one available. The job was dispatched to it right away.
  • Restarted the coordinator. The flag held.
• Manual drain run: fixed-capacity dry-run with two fake agents whose VMs stay up, one pool runner each. Marked A1 unavailable.
  • Its runner was kept while it was younger than the boot grace (the mock forge reports it unregistered).
  • Once the runner was backdated past the grace, it was drained: DestroyRunner went to A1 and the record was removed.
  • It wasn't replaced, no failure was recorded, and A2's runner was untouched.
• The forge-side Removed/Busy outcomes are covered by a unit test of the drain decision. The GitHub calls themselves aren't exercised here.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-19T18:32:10.706116Z	8878a38	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8878a38c49

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".