diff --git a/quest/m0/follow-up-duplicates.md b/quest/m0/follow-up-duplicates.md new file mode 100644 index 00000000..4b220b1e --- /dev/null +++ b/quest/m0/follow-up-duplicates.md @@ -0,0 +1,14 @@ +# [S] Agents stop adding follow-ups an open PR already owns + +## Goal + +Before an agent adds a follow-up quest, it finds any quest that already owns the work, on the base branch or in an open PR's quest files, and edits or links that one instead, so a run's merge picks stop asking to drop the same duplicate again. + +## Plan + +Planned 2026-11-15 with the maintainer in kixel-ai/rind's `/quest-run` of 2026-11-14 and 2026-11-15: + +- That run's agents added the rind-fips fork-test hang as a new follow-up five times and the dart-ios hang three times, in separate PRs, each while another open PR or main already had the quest. Each cost a pick round and a manual `quest rm` on the PR. +- `assets/AGENTS.md` says to "search the tree and git history before making a new quest", which covers main but not open PRs. Extend it to open PRs' quest files: read the `quest/**` files each open PR adds or changes. A `gh pr list --search` on the quest's subject is only a supplement, since GitHub's PR search does not read changed files and misses a PR whose text never names the subject. +- Have `quest check`, or the PR flow, warn when a PR adds a quest whose path or title another open PR also adds. A warning, not an error: two PRs may legitimately touch one quest. +- ✅ m0: duplicates cost a pick round each time in every repository using quest (rejected: m1). diff --git a/quest/m0/land-fixer-crash.md b/quest/m0/land-fixer-crash.md new file mode 100644 index 00000000..bb7aab10 --- /dev/null +++ b/quest/m0/land-fixer-crash.md @@ -0,0 +1,15 @@ +# [M] The land fixer stops crashing on kixel-ai/rind + +## Goal + +On kixel-ai/rind, a PR the merge queue drops for `failed_checks` gets a land fixer run that ends in a fix, a flaky verdict, or a handback naming what it needs, never a bare `fixer crashed (exit 1)`. A fixer that does fail hands back with the cause it printed, so the next failure names itself. + +## Plan + +Planned 2026-11-15 with the maintainer in rind's `/quest-run` of 2026-11-14 and 2026-11-15: + +- Every fixer handback that run saw on rind said `Codex land fixer: fixer crashed (exit 1); run /quest-iterate.` (`src/cli/land-fixer.ts`), on kixel-ai/rind#742, #748 (twice), #758 (twice), #786 and #789. The maintainer's session ran `/quest-iterate` by hand each time. +- On kixel-ai/rind#789 the rind watcher logged `fixing 789` and `handback 789` 66 seconds apart, and the maintainer's host journal shows no fixer container starting in that minute, while containers for other runs started and finished cleanly. So the run exits 1 before Codex starts: in Quest's launch of the fixer, or in the host's restricted launcher it calls. +- Here: make the handback comment carry the failed launch's stderr tail and exit reason instead of `crashed (exit 1)` alone, reproduce the failure, and fix it with a regression test where the cause is in Quest's code. +- Where the cause is in the host's fixer service (its socket, container, or credentials), that fix is kixel-ai/dotfiles' work, per this repository's `AGENTS.md`: add it there as a follow-up, record it under this quest's `## Waiting` as `kixel-ai/dotfiles:` with the PR or quest that fixes it, since `## Required` links only quests in this tree, and finish this quest once rind's next dropped PR gets a real fixer run. +- ✅ m0: the fixer is how [the queue watcher](/quest/m0/queue-watcher/README.md) absorbs the quest-file conflicts GitHub cannot resolve, so a crashing fixer leaves those for a human (rejected: m1).