[kixelated]

Coding agents stayed alive waiting on GitHub checks and reviews, and each wake reread a large context. An audit of a week of Claude transcripts on three hosts traced about 65% of subagents' 5–60 minute idle gaps to CI/PR waits and polling, and 14% to tests and builds.

This PR stops agents waiting on GitHub and splits the work so background workers fit a 5-minute prompt cache while the foreground coordinator keeps a 1-hour one.

Behavior

• No GitHub waits. quest wait, the runner's wait tool, and the webhook wake transport are removed. One-shot quest status, queue submission, and landing reconciliation stay. Nothing local watches a PR after publication; a later invocation reads its state and acts. Runner tools refuse further work after a successful delivery and allow a retry after a failed one.
• Workers do one burst. Every worker tool call returns within 4 minutes. A worker commits, reports, and terminates, pushing only to a draft. Open decisions go under ## Decisions needed in its PR. The coordinator asks the user, records the answers in the PR, and starts a fresh agent instead of resuming the one that asked.
• The coordinator runs the long waits. Full checks, quest review rounds, and gated commands run in the coordinator, in the worker's checkout, with each wait under an hour. It gives each round's findings to a fresh /quest-iterate worker, which works under the coordinator's hold, and it pushes to a ready PR only after checks and review pass. A foreground session working alone does both.
• /quest-complete, /quest-spawn, and /quest-afk stay long-running coordinators. That restores the interactive prompt after /quest-review, PR-list refreshes, and automatic /quest-plan of suggested follow-ups, which an earlier commit here had removed on the assumption that coordinators would exit.

Follow-up quests

• Gate tickets: quest gate returns within 4 minutes, either admitted with a token or queued with a ticket, so waiting for admission stops outlasting a worker's cache.
• Fix ejected landings: an ejection webhook starts a fresh kix merge job that fixes the PR and lands it again, or labels it for the maintainer.
• moq and moq.pro each plan just check and just test under 5 minutes per diff, in their own PRs.

Decisions needed

Re-read on 2026-11-01 against main at a27cacd, 57 commits past this PR's base. The PR now conflicts with main in README.md and in the complete, merge, review, and spawn skills. The code removals still merge cleanly.

What main already covers

• The skills stopped waiting on GitHub. /quest-merge now only queues inline, "never ... waits on GitHub", and leaves pending checks to a later pick (moq-dev/quest#292). /quest-complete's sub-agents "never wait on GitHub CI, reviews, or merges", and /quest-spawn stops at the PRs (moq-dev/quest#300). That covers this PR's merge, complete, afk, and spawn changes in a different form.
• Fix ejected landings is superseded. The host-land epic says it "replaces the hook-started kix fixer that refactor!: stop waiting on GitHub and keep workers in a 5-minute cache #276 proposes", and the land daemon's fixer now triages failing batches instead of ejecting them (moq-dev/quest#303, moq-dev/quest#355).

What conflicts

• Removing quest wait, the runner's wait tool, and the webhook wake. The guide's Workspace section on main still says "Wait on a PR's CI or reviews with quest wait ... Bound each wait under an hour". On 2026-11-01 the maintainer also dropped their personal "never wait or poll for GitHub CI" rule. Waiting after publication is still discouraged ("report the PR and terminate"), but waiting is no longer banned outright. Main still uses all three removed pieces.
• The worker/coordinator split, where workers never run /quest-review, reverses main's current flow. On main, each started agent runs /quest-review itself, and /quest-complete runs /quest-iterate and /quest-review in sub-agents that report when done.

What still applies

• The 5-minute worker cache idea: one burst per worker, tool calls under 4 minutes, and the coordinator running the long checks and review rounds. The gate tickets quest depends on it. Neither is on main. The open P2 finding from the 2026-10-29 review is still unfixed: a finished worker's draft stays held by its starter, so the coordinator's quest ready is refused.

1. What to do with this PR?
   • Close as superseded (recommended). Main already ships the "skills don't wait on GitHub" half. The quest wait removal contradicts the current guide and the maintainer's latest stance, and the eject fixer is replaced. Comment with this analysis when closing.
   • Rework to fit: drop the quest wait, runner wait, and webhook-wake removals and the eject-fix quest, then rebuild the worker/coordinator split on main's current skills and fix the P2 handoff. That amounts to a new PR, and it rests on decision 2.
   • Merge as is: not possible. It conflicts with main, contradicts the current guide, and has an open finding.
2. Is the 5-minute worker split still wanted? (only if the PR is closed or reworked)
   • Re-plan it with /quest-plan as its own quest (recommended). First confirm the premise that sub-agents keep only a 5-minute cache. Include the gate-tickets quest and the starter-hold handoff, so the decision gets made against main's current skills.
   • Drop it, along with gate tickets.

Validation

• just check passed at 811deae.
• just test passed at abd9c65: 557 CLI/runner tests (2 skipped), 426 Worker tests, shell regression suites, and both CLI lifecycle checks. The only change after it is documentation.
• test/agent.test.ts fails inside an agent session unless CLAUDE_CONFIG_DIR is unset. That's pre-existing and already tracked in agent test environment.

Review

Round 2: quest review, Codex gpt-6.1-sol, medium effort. Reviewed head 811deaef7e4550901f0de158217e970e24bc6977 against origin/main (01d051c): no findings.

• Round 1 (head abd9c65) had one finding: a worker could push to a ready PR before the coordinator's checks and review. Fixed: workers push only to drafts, and the coordinator pushes to ready PRs after checks and review pass.
• The earlier commits (through dc44496) were reviewed by Claude Opus 5.5 and approved.

🤖 Generated with Claude Code

(Written by Claude Opus 5.5)

[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 2026-10-29T17:58:39.851393Z	811deae	New commits
🔒 Security Review	✅ Completed 2026-10-29T18:00:57.083190Z	811deae	New commits

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.

[kixelated]

Automated review by review (OpenAI)

Reviewed commit: dc44496

No actionable bugs found. The removed CLI/runner waits and webhook wake transport are consistently disconnected, while one-shot context/status and landing-hold reconciliation remain. Successful delivery closes subsequent runner tools; a failed publication leaves delivery retryable. The added tests cover those boundaries without treating publication as proof that CI or landing finished.

Direction: separating a finite coding attempt from GitHub monitoring is a coherent simplification. The skills correctly distinguish the user's continuing merge decision from temporary ownership, require a fresh hold after ejection, and do not promise that this change starts a monitor.

Verification: GitHub-only static inspection of changed runtime, webhook/protocol, CLI and regression tests, plus an independent skills/ownership-handoff review. No builds, tests, agent lifecycle runs or deployment executed.

[kixelated]

Automated review by review (OpenAI)

Re-reviewed commit 811deaef7e4550901f0de158217e970e24bc6977, focusing on the two commits since my review of dc44496cb215374a0f2a8fa0e76938456fa892ce.

The short-lived worker/foreground coordinator split is coherent, and keeping ready-PR pushes behind the coordinator's checks and review is the right boundary. One new handoff gap remains:

• P2 — Release the starter-held draft when a completed worker hands it to the coordinator (assets/skills/start.md:26). A worker with no open decisions now pushes its draft and terminates without running /quest-review, but it leaves no ## Decisions needed handback marker. src/cli/hold.ts:128–129,303–306 therefore still classifies that draft as held by its starter and refuses the coordinator's quest ready <pr> with “taking it over is the user's choice.” The review skill's bypass is explicitly for a draft started in the same session, not one started by the terminated worker. This stalls the normal successful /quest-spawn path before the coordinator can perform the promised review. Define an explicit completed-worker ownership handoff that the coordinator can consume while keeping the PR draft, and cover worker completion → coordinator acquisition → local review in a regression.

The earlier CLI/runner wait removal is unchanged, not a new finding. GitHub-only static review of the delta, skills, and hold-acquisition implementation, with an independent handoff review; I did not run tests or agent lifecycle workflows.

[kixelated]

Closing as superseded, per the maintainer's decision on 2026-11-01:

• Skills no longer wait on GitHub. Main stopped this through moq-dev/quest#292 and moq-dev/quest#300.
• The fixer is handled elsewhere. The ejected-landings fixer is replaced by the quest land fixer (moq-dev/quest#303, moq-dev/quest#355).
• The quest wait removal conflicts with the current guide. It removes quest wait and the wake path, which the guide still uses to wait on CI and reviews.

The 5-minute worker/coordinator split and gate tickets will be planned again as their own quest, starting with a check that sub-agents really keep only a 5-minute cache. The branch stays available for reference.

(Written by Claude Opus 5.5)