Repository navigation
Rationalize the controller's manual work into the poll, scheduler and supervisor #649
Description
From the operator (cgwalters), 16 Oct: "we need to rationalize what you are doing with the poller".
Below is everything the interactive controller did by hand on 15–16 Oct, and which component should own it. The goal is that the controller session becomes unnecessary. Its remaining job is conversation with the operator, which turns into issues and plans.
| Manual action (examples) | Owner | Status |
|---|---|---|
| Queue approved green PRs | poll | done |
| Re-queue PRs removed by a cascade; fix real queue failures (#640, #641, #777, #829) | poll (agent-board#92) | bug: it re-queued #777 three times on the same failure instead of fixing. See the escalation issue. |
| Rebase/fix rounds; then a targeted fix with the failing log excerpt once rounds run out (#640, #698, #740, #829) | poll | gap: escalation issue below |
| Mark complete drafts ready (#826–#829, #647, #648) | agentic-job apply | done (tracker#648 ready_when_complete) |
| Continue checkpoint drafts; retitle a partial PR once continued (#776) | poll (agent-board#90/#98) | partial: retitle is open, folded into the escalation issue |
Close issues whose PR merged without Closes (#2, #715, #717, #812, …) |
apply (agentic-job#760), scheduler (#645) | done |
| Prevent duplicate dispatches (#630, #22) | apply Closes (#760), dispatch overlap (agent-board#88/#89) |
done |
Re-dispatch runs lost to transients (binary 403, egress-cargo, secret scan) |
scheduler (#645) | done for scheduled items |
| Notice an idle scheduler and feed it (07:19, 08:19) | scheduler idle report (#636) plus supervisor | gap: supervisor agent |
| Read operator#1 and act on answers (D1–D4, D14) | supervisor | gap: supervisor agent |
| Post safe outputs gh-aw refused (#719: 54 links) | agentic-job in-session validator (#854) | in progress |
| Version bump, release, then pin bumps in actions, tracker and agent-board (v0.5.0, 0.5.1, 0.5.2) | release train | gap: release-train issue below |
Root-cause investigations (GLM window, edit tools, missing ready_when_complete) |
supervisor agent (research dispatches), operator | gap: supervisor |
| Turning operator chat into issues and plans | controller, or the operator directly via operator#1 | stays |
Gap issues (filed alongside, linked here as they open):
- Supervisor agent and executor: tracker#626 steps 2–4 on the D14 typed-action allowlist already in
.github/supervisor/actions.cjs. Also find out whysupervisor.yml's scheduled runs are all skipped. - Poll escalation: after fix rounds run out, one diagnostic round with the failing job's log excerpt; the bootc CI: extend package artifact retention beyond 1 day #92 repeat-requeue bug; retitling continued partial PRs.
- Release train: cut agentic-job releases automatically when changes accumulate, then open the pin-bump PRs.
The operator decides whether this becomes an epic.
Generated-by: https://github.com/cgwalters/#llms
Activity
supervisor: Run the health pass, and let an agent ask a validated exe…
cgwalters-bot commentedon Oct 16, 2026
Operator, 16 Oct, verbatim: "a key goal is to move your observation and steering hosted".
SUPERVISOR=applyis now set on tracker (11:24 UTC). From here, every check the controller makes on its 15-minute tick should be a supervisor health check, and every correction it makes should be a supervisor typed action. The table above is the list. The controller's tick shrinks as each row moves.Generated-by: https://github.com/cgwalters/#llms
cgwalters-bot commentedon Oct 22, 2026
✅ dispatch ended well. Result: #649 (comment)
cgwalters-bot commentedon Oct 22, 2026
Triage (21 Oct): done as a tracking issue. Recommend closing it as completed. I can only comment from this run, so closing it is left to someone else. Every gap row in the table has its own issue or a merged PR now, so nothing is left that only #649 tracks. Nothing here needs code.
What I found for each gap:
supervisor.ymlruns were all skipped. That is answered: since e34aae5, every poll pass starts the supervisor itself (see theworkflow_run/workflow_dispatchtriggers in.github/workflows/supervisor.yml). Recent supervisor runs showworkflow_dispatchsucceeding about every 8 minutes. Only theissue_commenttriggers are skipped, and that is expected. The rest of the work has its own issues: supervisor: Configure the requests and log issues so the agent runs #657 (requests/log issues), supervisor: Observe what the controller watches by hand, and stop warning on missing data #659 (observe what the controller watches), supervisor: Act on routine corrections instead of asking (close, plan, re-dispatch, enqueue), with a dispatch cap #666 (routine corrections with a dispatch cap).diagnosticPoolhere). Both are open. 5b3c128 (poll owns the merge queue; a repeated merge-group failure starts a fix round) covers part of the bootc CI: extend package artifact retention beyond 1 day #92 symptom.Unclear: the issue leaves it to the operator whether this becomes an
epic. If@cgwalterswants an umbrella kept, the child issues above could become its sub-issues instead of closing it.Generated-by: https://github.com/cgwalters/#llms
cgwalters-bot commentedon Oct 23, 2026
Done as a tracking issue: every row now has its own issue or merged PR (the last was #626), so nothing remains that only this issue tracks.
Generated-by: https://github.com/cgwalters/#llms