Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
120 changes: 120 additions & 0 deletions quest/m0/ask-less/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,120 @@
# Ask the maintainer only what their answer changes

## Goal

Agents ask the maintainer only the decisions their answer is likely to
change. Everything else proceeds on a standing rule the maintainer set and
recorded, and stays reviewable after the fact. Each repository's work loop
runs without being launched by hand, and the maintainer opens one session
when they choose, or when a push says decisions wait.

This epic's own work is its closing re-measure: rerun the baseline below
over a week of sessions on every host, and record the comparison in its
closing PR.

Non-goals: releases and deploys, which keep their own approvals; the
online reviewers; changing what reviews a PR before it is ready.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan), from the transcripts on the
maintainer's three Linux hosts and the forge records of moq-dev/moq,
moq-dev/moq.pro, moq-dev/quest, kixel-ai/rust2 and kixel-ai/rind:

- Goal? ✅ As above (rejected: keep every merge pick and stop only
repeated and rubber-stamped questions; agents merge whatever they
recommend and the maintainer only sets direction).
- Scope? ✅ All four: merge picks by change class, never asking twice,
the loop running without the maintainer, and the machine-side blockers
(rejected: none dropped).
- Which merges land on a standing pick? ✅ Every PR the coordinator would
merge, except a protected list
([standing picks](/quest/m0/ask-less/standing-picks.md)) (rejected:
plan PRs and bug fixes only; plan PRs only). The maintainer's override
rate barely varies by kind of change, and the pick does not catch
escapes: nearly every later culprit had been picked.
- Review after the fact? ✅ A list of what landed on a standing pick,
acted on by naming one ([landed list](/quest/m0/ask-less/landed-list.md))
(rejected: session summaries only; the list plus a weekly second-model
audit). Either way the [mistake audit](/quest/m0/mistake-audit.md)
counts escapes by standing and maintainer pick, and proposes narrowing a
class whose escapes rise.
- Which planning questions stop? ✅ Milestone placement only
([milestone by rule](/quest/m0/ask-less/milestone-by-rule.md)); which
follow-ups to plan is still asked (rejected: both; neither).
- The loop? ✅ Each repository's `/quest-run` runs headless on a schedule
and never asks, a push goes out when only decisions remain, and one
session asks them across every repository
([headless loop](/quest/m0/ask-less/headless-loop.md),
[decisions session](/quest/m0/ask-less/decisions-session.md))
(rejected: the same without a push; keep launching by hand, with one
cross-repository `/quest-complete`).
- Push channel? ✅ The Discord webhook kixel-ai/dotfiles' storage-alert
quest adds, so there is one place to look (rejected: a separate Discord
channel; ntfy; decide later). Quest runs a configured command; the
host owns the channel.
- How long does a recorded skip hold? ✅ Until the PR's head changes or the
maintainer lifts it
([recorded picks](/quest/m0/ask-less/recorded-picks.md)) (rejected:
until lifted; until the next day).
- Which repositories? ✅ Every repository, opting out or adding protected
paths in its `CONTRIBUTING.md` as today (rejected: opt in per
repository; all but those running production services).
- Milestone? ✅ m0 (rejected: m1; fixes in m0 and the loop in m1).

Host work, decided the same day, goes to kixel-ai/dotfiles' m0, since its
consolidate-hosts quest makes it the one repository for the maintainer's
hosts and archives kixel-ai/nixos:

- The service running each loop and posting to Discord.
- The dev-shell guard, which refused 601 agent session starts since 11-03.
✅ Keep its HEAD-on-origin check and accept the flake files at any version
merged on the default branch, not only its tip. First picked without the
HEAD check, then reversed once it turned out to reopen the fork-PR hole
kixel-ai/dotfiles@7202fdf closed (rejected: drop the HEAD check too; also
accept the agent's own flake changes; keep the guard as is).

Baseline. Typed messages and answers come from `quest messages --json` run
on each host, 2026-10-20 to 11-13; sessions whose Codex TUI turns it
rejects were skipped ([messages fix](/quest/m0/ask-less/messages-codex-tui.md)).
Blocking time comes from the session logs since 11-03; PR stages from the
forge timelines of PRs created and merged 11-03 to 11-13.

- 12,900 answers to agents' blocking questions, about 500 a day, with the
maintainer active 7 to 12 hours every day. 74% took the recommendation
unchanged, 8% picked another listed option, 12% typed something else
(9% in the first week, 15% in the last), and 5% had no recommendation.
- 2,752 answers where the agent recommended merging, 5.1% overridden; 4.2%
where the question named no API, breaking, decision, outside-author, or
release signal. Since the narrow standing pick landed on 11-07, about
127 such answers a day.
- 345 merge-pick answers and 80 skips repeated an answer already given for
the same PR; 325 PRs were asked about three or more times.
- Milestone placement was asked 902 times (85% accepted), and which
follow-ups to plan 602 times (86%).
- Typed overrides that were not just the recommendation retyped: 13% asked
what something is or why, 12% cut scope or deleted, and 8% were about API
shape or naming. Those, goal and scope, and product knowledge are where
the maintainer's answer changed outcomes.
- Typed messages: 192 hand-run `/quest-complete`, 112 `/afk`, and status
nudges such as "any new prs?", "finish up", and "continue" among the
commonest short messages.
- 2,954 blocking questions since 11-03 kept their agents waiting 389 hours
in total: median 2.4 minutes, p90 16 minutes, 8 over six hours.
- 2,354 merged PRs: median lead time 5.8 hours, p90 25.6. Of the summed
lead time, implementation and local review took 38%, waiting from ready
to the first pick 40% (median 1.0 hour, p90 12.2), and the queue 21%.
A PR readied between 23:00 and 04:00 Pacific waited a median 5 to 9
hours for its pick. 96% were first picked from the maintainer's account.
- 213 merged PRs (9%) are linked under a later PR's `## Caused by`, an
upper bound, since the mistake audit found most such links are not real
escapes.

## Related

- [Mistake audit](/quest/m0/mistake-audit.md) - tallies escapes by pick and
the answers taken as recommended each week
- [Late signals](/quest/m0/late-signals.md) - holds a head on a marked
decision comment, which recorded picks can share
- [Queue watcher](/quest/m0/queue-watcher/README.md) - re-enqueues the
merge queue's drops, 21% of merged PRs' lead time
30 changes: 30 additions & 0 deletions quest/m0/ask-less/decisions-session.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,30 @@
# [L] Settle every repository's decisions in one session

## Goal

One `/quest-complete` session covers every repository the user works in:
it first says how many decisions wait, and where, then shows the
[landed list](/quest/m0/ask-less/landed-list.md), then asks the decisions
in its usual rounds, ranked across repositories by what each unblocks. It
edits each repository's quests in a checkout of that repository, one PR per
repository, as it does for one today.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan,
[the epic](/quest/m0/ask-less/README.md)), with the
[headless loop](/quest/m0/ask-less/headless-loop.md): one session asks the
decisions every loop deferred.

Agent's choices, for the implementer to revisit:

- The repositories come from one list in the user's own Quest settings,
which the headless loop's host service reads too, rather than a scan of
checkouts.
- Reuse `quest waiting --json` per repository and merge the rankings,
rather than a cross-repository command, unless the merge needs one.

## Related

- [Recorded picks](/quest/m0/ask-less/recorded-picks.md) - keeps it from
asking what an earlier session settled
43 changes: 43 additions & 0 deletions quest/m0/ask-less/headless-loop.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
# [L] Run a repository's quest loop headless and push when decisions wait

## Goal

One command runs a repository's `/quest-run` as a headless agent that never
asks: it takes the standing picks, starts ready quests, iterates and
merges, and defers every decision the maintainer owns, as `/quest-afk`
does. When it goes idle with nothing left but the maintainer's decisions,
it runs a command the user configured once, with how many decisions wait
in which repositories. A host service runs it on a schedule, and the
maintainer answers in the
[decisions session](/quest/m0/ask-less/decisions-session.md).

Non-goal: the service and the channel, which the host owns: the
maintainer's are a kixel-ai/dotfiles quest posting to Discord.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan,
[the epic](/quest/m0/ask-less/README.md)): ✅ headless loops with a push
when decisions wait (rejected: the same without a push; keep launching by
hand). The core runs a configured command, so it names no service.

From 2026-10-20 to 11-13 the maintainer hand-ran `/quest-complete` 192
times and `/afk` 112 times, each in one repository, and PRs readied
between 23:00 and 04:00 Pacific waited a median 5 to 9 hours for a pick.

Agent's choices, for the implementer to revisit:

- Build on `quest agent start`, which already runs a headless turn on the
CLI's own login and reports `limited` on a usage limit, rather than a
second runner. A limit ends the run cleanly; the next scheduled run
resumes it. `quest agent start` admits itself, so the loop starts only
what the host and account can carry.
- Push once per transition to "only decisions wait", never per run, so a
schedule that finds nothing new says nothing.
- Keep the configured command in the user's own Quest settings, where
`quest setup` writes, never in a repository.

## Related

- [Standing picks](/quest/m0/ask-less/standing-picks.md) - what it may
merge without the maintainer
30 changes: 30 additions & 0 deletions quest/m0/ask-less/landed-list.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,30 @@
# [M] List what landed on a standing pick

## Goal

The maintainer sees, in one list across their repositories, every PR that
merged on a standing pick since they last looked, grouped by the class it
qualified under, one line each. Naming one starts its revert or its redo,
as the maintainer says, as a new PR. The
[decisions session](/quest/m0/ask-less/decisions-session.md) shows the
list first.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan,
[the epic](/quest/m0/ask-less/README.md)): ✅ a landed list the maintainer
acts on (rejected: session summaries only; the list plus a weekly
second-model audit of a sample).

Agent's choices, for the implementer to revisit:

- A CLI command reading the [recorded picks](/quest/m0/ask-less/recorded-picks.md)
of merged PRs, with `--since` and `--json`, so the skill step stays
prose-light.
- "Since they last looked" is the time the list was last shown, kept in the
user's own state, never in a repository.

## Required

- [Recorded picks](/quest/m0/ask-less/recorded-picks.md) - which PRs merged
on a standing pick, and under which class
31 changes: 31 additions & 0 deletions quest/m0/ask-less/messages-codex-tui.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
# [S] quest messages reads Codex TUI turns that carry only content kinds

## Goal

`quest messages` reads Codex 0.162 TUI sessions, whose user turns carry
`content_item_kinds`, such as `user.text` for typed text and
`agents_md.instructions` or `environments.environment_context` for
injected context, with no `content_item_metadata` provenance. It counts
`user.text` parts as typed, sets the injected kinds aside, and still fails
on a kind it does not know.

## Plan

Found 2026-11-12 while measuring [the epic](/quest/m0/ask-less/README.md)'s
baseline: the command failed on two of the maintainer's three hosts, 4
sessions on one and 21 on the other, every one `codex-tui` 0.162.1, so the
baseline was taken with those sessions removed from a copy of
`~/.codex/sessions`.

`readCodexMessages` in `src/cli/messages.ts` accepts `content_item_kinds`
only when every kind is `unknown`, a compaction's replay, and
`test/messages.test.ts` expects `["user.text", "unknown"]` to throw. Keep
that refusal for parts whose kind says nothing.

Caused by [moq-dev/quest#626](https://github.com/moq-dev/quest/pull/626),
which required provenance on every typed Codex turn.

## Related

- [Mistake audit](/quest/m0/mistake-audit.md) - reads typed messages from
every host each week
20 changes: 20 additions & 0 deletions quest/m0/ask-less/milestone-by-rule.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
# [S] Place new quests by rule instead of asking

## Goal

`/quest-plan` places each new quest in the milestone a recorded rule or its
siblings' placement gives: a milestone rule the maintainer stated, in this
session or recorded in the tree, or the milestone of the epic or quests it
joins. It records the placement and the reason in the quest's Plan, lists
it in the summary, and asks only when neither covers the quest.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan,
[the epic](/quest/m0/ask-less/README.md)): ✅ milestone placement stops
being asked; which follow-ups to plan is still asked (rejected: also plan
every recommended follow-up unasked; keep asking both).

Milestone placement was asked 902 times from 2026-10-20 to 11-13, 85% taken
as recommended. `assets/skills/plan.md` asks it of every new quest today,
unless a rule from the same session covers it.
33 changes: 33 additions & 0 deletions quest/m0/ask-less/recorded-picks.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
# [M] Record merge picks and skips on the PR

## Goal

A merge pick or skip, the maintainer's or a standing one, is recorded on
the PR itself, so no session asks it again. `quest status` prints it, and
every coordinator honors it: a pick holds until the PR lands or hands back
a decision, and a skip holds until the PR's head changes or the maintainer
lifts it. A skip whose head changed is asked once more, saying what
changed. A standing pick records the class it qualified under.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan,
[the epic](/quest/m0/ask-less/README.md)): ✅ a skip holds until the head
changes (rejected: until the maintainer lifts it; until the next day).

Today a pick lives in the asking session and its sub-agents' briefs, and a
skip "stands for the session" (`assets/skills/run.md`, `complete.md`), so
from 2026-10-20 to 11-13 345 merge-pick answers and 80 skips repeated one
already given for the same PR, and 325 PRs were asked about three or more
times.

Agent's choices, for the implementer to revisit:

- Record each as a PR comment opening with a fixed marker naming the kind,
the head it was given on, and for a standing pick its class, beside
[late signals](/quest/m0/late-signals.md)' "Maintainer decision"
marker, rather than a label, which cannot carry the head. Only comments
from accounts with write access count.
- GitHub auto-merge enabled on the current head stays the pick's record
while it lasts; the comment covers the gaps, such as a PR waiting for
`/quest-iterate`.
53 changes: 53 additions & 0 deletions quest/m0/ask-less/standing-picks.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,53 @@
# [M] Merge everything outside a protected list on a standing pick

## Goal

A coordinator that would recommend `/quest-merge` for a PR runs it without
asking, unless the PR is on the protected list:

- a public API or wire change, or a breaking (`!`) title
- a release, deploy, or backport, or anything that ships one
- a new dependency
- shipped agent rules or skills, such as Quest's own `assets/`
- an author without write access
- open decisions: a `## Waiting` section, or a plan PR holding decisions
the maintainer did not make in its interview

A plan PR that records only the maintainer's own interview answers
qualifies, so adding or expanding quests no longer needs a pick by itself.
A repository's `CONTRIBUTING.md` can opt out, as today, or add protected
paths, such as deploy or billing configuration. The summary names each
PR merged this way, with the class it qualified under.

## Plan

Decided 2026-11-12 (maintainer, /quest-plan,
[the epic](/quest/m0/ask-less/README.md)): ✅ protected list (rejected:
plan PRs and bug fixes only; plan PRs only), in every repository with
`CONTRIBUTING.md` opt-out as today (rejected: opt in per repository; all
but production services).

The standing pick of `assets/skills/run.md` (landed in
[moq-dev/quest#534](https://github.com/moq-dev/quest/pull/534)) covers only
PRs with no user-facing behavior change that add no quests. This
replaces that rule there, in `assets/skills/afk.md` and `merge.md` where
they restate it, and in this repository's `AGENTS.md`, whose merge line
quotes the old rule.

Agent's choices, for the implementer to revisit:

- Judge the diff, not the title, as today, and ask when unsure. A pick
stays standing only while each new head still qualifies.
- Record the class each merge qualified under, through
[recorded picks](/quest/m0/ask-less/recorded-picks.md) once it lands,
so the [landed list](/quest/m0/ask-less/landed-list.md) and the
[mistake audit](/quest/m0/mistake-audit.md) can tell standing picks from
the maintainer's.
- Before review, replay the rule over a sample of last week's merge-pick
questions and report how many it would have taken and how many of those
the maintainer overrode, from `quest messages --json`.

## Related

- [Mistake audit](/quest/m0/mistake-audit.md) - proposes narrowing a class
whose escapes rise
7 changes: 7 additions & 0 deletions quest/m0/mistake-audit.md
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,13 @@ Check first next run:
[late signals](/quest/m0/late-signals.md) lands.
- Whether moq-dev/moq and moq-dev/moq.pro planned the runtime follow-ups
of moq-dev/quest's 2026-11-11 run PR, and whether their escapes recurred.
- Once [standing picks](/quest/m0/ask-less/standing-picks.md) widen and
[recorded picks](/quest/m0/ask-less/recorded-picks.md) record their
class, count escapes separately for standing and maintainer picks, by
class, and propose narrowing a class whose escapes rise (maintainer,
2026-11-12, [ask less](/quest/m0/ask-less/README.md)). Compare the
answers taken as recommended with that epic's baseline, read from
`quest messages` on every host.

Baseline from the 2026-11-11 run (PRs updated 2026-11-09 to 11-11; typed
messages 2026-10-26 to 11-11, classed by one sub-agent, approximate):
Expand Down