Find Simplifications

Finding qwen-code Simplifications

AGENTS.md § Simplicity First — "Minimum code that solves the problem. Nothing speculative." This skill hunts what that principle already rejected but that shipped anyway. It owns correct code nobody needs, repo-wide, with no diff to anchor on.

Every candidate names the surface, names every consumer, and says what breaks when it goes. A candidate whose consumers cannot all be named is dropped — not downgraded, dropped.

Read this file, then the one document for your phase. Do not survey or land from this file alone; if the phase document cannot be read, stop and say so.

Issue First, PR Second

A run's deliverable is a comment on the tracking issue, not a PR.

Survey → file candidates on the tracking issue → maintainer says yes to one
       → land that one → one PR, one candidate.

This is not ceremony. AGENTS.md § Core Infrastructure ends "When in doubt, escalate. Better to wrongly escalate than to wrongly approve," and an unrequested batch of deletions is the shape that gets closed. It also makes the ledger free: the issue that carries the proposals is the same object the next run reads to avoid re-proposing them.

A maintainer asking for a PR in this session counts as assent — record it on the ledger before opening the PR.

Boundaries

Territory

The split is not taste. packages/core/package.json exports "./src/*" and "./dist/*", and packages/core/src/index.ts carries ~179 export * from lines, so every file under packages/core/src is reachable from outside this repo. The release workflow npm-publishes @qwen-code/audio-capture and the eight @qwen-code/channel-* packages with --access public, so a symbol re-exported by their package entry is reachable the same way. No grep inside this repo can prove such a symbol has no consumer. The same is true of two surfaces whose consumers are not imports at all: packages/core/vendor/** and packages/web-shell ship inside the published @qwen-code/qwen-code tarball — packages/core/package.json lists vendor in files, and scripts/copy_bundle_assets.js copies both vendor/ and web-shell/dist into the bundle, where qwen serve hands the latter to browsers. Their consumers are registry-, tarball-, or browser-side, so an in-repo grep for them returns zero hits and the consumer proof passes vacuously — getBuiltinRipgrep() even assembles the vendor path from segments, so no literal path exists to grep for. Import reachability is the floor, not the definition: consumption also happens through runtime reads, loaders, manifests, and tool configs that never appear as imports. The consumer proof must name the mechanism that consumes the surface; a grep blind to that mechanism proves nothing, and the rows below mark every such path this repo ships or loads. A path named by more than one row takes the most restrictive outcome: Never-a-target beats Report-only, which beats Landable.

A settings key with zero read sites is still not cleanup. The key is likely documented under docs/users/ and completed from the generated schema; removing it withdraws a documented, user-settable option and turns an accepted setting into a silently ignored one — warned nowhere, since the unknown-key check compares top-level keys only and its output is a session-gated debug-log append, never the terminal. That is a deprecation decision. File it; never land it.

Report-only does not mean worthless — a named, proven, un-landable finding is exactly what a maintainer needs to make the call. It means the run stops at the issue comment.

Decision Table

AGENTS.md § Core Infrastructure Is Maintainer-Only governs packages/core/src/**, packages/*/src/ under auth, providers, models, config, tools, services, and any cross-package change: 500+ production logic lines of refactor there is a hard block for non-maintainer PRs (excluding *.test.ts(x), *.spec.ts(x), __tests__/**, *.schema.{ts,json}, *.generated.ts, **/generated/**), and anything smaller "must be 100% confident … name every downstream consumer; if it cannot, escalate." Breadth alone is not size: a sweep touching many files with a line or two each is judged on confidence, not file count.

Recurring-Run Design

Rotation

Survey one slice per run, picked by the calendar, plus at most one other if the first comes up empty. Rotation is what stops the third run from re-searching the same hot directories.

git fetch origin || exit 1
# Fetch only updates the ref, but every grep below reads the working tree —
# survey fresh code from a throwaway worktree at origin/main. Never switch
# the user's checkout: the survey is read-only and must leave the checkout
# exactly as it found it.
SURVEY_PARENT="${TMPDIR:-/tmp}/find-simplifications-survey"
SURVEY="$SURVEY_PARENT/main"
# Clear any leftover from a failed or interrupted earlier run first. The
# fixed path (not mktemp) lets every call re-derive the worktree with no
# shared state, so at most one leftover can ever exist. There is no EXIT
# trap: the consuming harness spawns a fresh shell per command, so a trap
# fires when THIS call ends — before any survey command runs — and its
# variables do not survive to the call that would need to clear it.
# Delete the directory, never `git worktree remove --force` on the fixed
# path: remove resolves symlinks, so a planted link to another registered
# worktree of this repo would force-delete that foreign tree, uncommitted
# work included. rm -rf unlinks a symlink without following it; prune
# then clears the stale registration. Fail closed on removal errors: a
# leftover that cannot be deleted is owned by someone else, and the
# checkout must never go into a parent this user does not control.
rm -rf "$SURVEY_PARENT" || exit 1
# Close the create side too: `git worktree add` accepts a planted EMPTY
# directory (exit 0, the planter keeps ownership and mode) and a symlink to
# an empty directory (exit 0, the checkout written through the link), so
# recreate the parent as a fresh 0700 dir owned by the running user — no
# other local user can then plant the worktree path. mkdir, not install -d:
# install -d exits 0 adopting an existing directory and following a planted
# symlink, so no exit-status check could tell creation from adoption; mkdir
# fails when anything is still at the path, aborting the block instead.
mkdir -m 0700 "$SURVEY_PARENT" || exit 1
git worktree prune
git worktree add --detach "$SURVEY" origin/main || exit 1
# Slice from the month, not the ISO week: a monthly run advances ISO weeks
# by ~4, so week % 4 repeated the same slice for 3–6 monthly runs at a
# time. POSIX slice computation: the `10#` radix prefix is bash/ksh-only
# and is a hard syntax error under /bin/sh (dash), hence the leading-zero
# strip.
M=$(date -u +%m); SLICE=$(( ${M#0} % 4 ))
echo "SURVEY=$SURVEY SLICE=$SLICE"

Each fenced block in this skill is ONE command: the consuming harness spawns a fresh shell per command, so environment variables, cwd, and traps set here do not survive to later calls. Every later call that needs the worktree re-derives the same fixed path (${TMPDIR:-/tmp}/find-simplifications-survey/main) and runs its commands from there (cd "$SURVEY" && …, or by setting the call's working directory).

Skip a slice the ledger shows was surveyed in the last three runs; the fallback is the lowest-numbered slice NOT surveyed in those runs — a skip with no fallback surveys nothing, and three empty runs in a row trip the stop-loss below without any sweep having happened. Survey a report-only territory only when a human asks for it by name.

Check churn per candidate rather than trusting a static list — it moves:

git log --since='3 months ago' --name-only --pretty=format: -- <dir> \
  | grep -v '^$' | wc -l

A candidate sitting in a directory the team touches constantly loses the merge race on any slow cadence, and its test files are exactly where conflicts land. File those; do not try to land them.

The ledger

State lives in ONE long-lived GitHub issue, [find-simplifications] candidate ledger, and nowhere else. Do not commit a state file: a ledger in the repo turns "found nothing" into a diff, conflicts across bot branches, and is deleted by the autofix branch sweep.

Read it before surveying, append to it after. If the search returns no issue, create it first through /create-issue, with exactly the title [find-simplifications] candidate ledger — a first run has nowhere else to post, and an improvised title splits the state this section keeps in one issue:

gh issue list --state all --search '"find-simplifications" ledger in:title'
gh issue comment <number> --body-file <comment>.md

• Each candidate gets an id derived from the surface, never from prose: slug(<primary symbol, file, or directory>) — enum-selector, locale-orphans-auth-subcommand. A prose-derived id reappears forever.
• One line per candidate: id — territory — status — date. Statuses: filed, landed, declined, dropped-recency, dropped-consumers.
• An id on the ledger with status declined is a permanent tombstone. Never re-propose it, however good the new evidence looks. A closed-unmerged simplify/* PR tombstones its id only when it closed on the finding's merits — the same test the stop-loss rule below uses. Read the closing comment and the id's ledger line; when neither records a merits decline, treat the close as operational. A PR closed for operational reasons (a stale base, conflicts, an infrastructure retry, a superseding re-file) is not a tombstone: re-survey the id and, if it still survives the proof protocol, re-file it. Search with --state all — the default open filter never returns the closed-unmerged PRs this rule targets — and quote the marker: GitHub tokenizes on hyphens, so gh pr list --state all --search 'enum-selector in:body' returns unrelated PRs, while --search '"find-simplifications:id=enum-selector" in:body' does not.
• Append only. Never rewrite, reorder, prune, or summarize it. A maintainer removing a line is the only supported retraction.
• Record what a run rejected and why, not only what it filed. That is the whole anti-churn mechanism: without it the next run re-derives and re-rejects the same hundred symbols.
• If the ledger cannot be read, stop and say so in the run's output — no candidates this run. The ledger is the only copy of the permanent tombstones and the recent-slice rotation this section keeps, so surveying through a read failure can re-propose a permanently declined id or re-sweep a slice the last run already covered. Fail closed; there is no equally authoritative snapshot to fall back on. In headless mode (§ Output) the read is a file, not gh: the caller supplies the ledger text as <workdir>/ledger.md before the run, and a missing or unreadable file is exactly this read failure.

Never treat a previous run's finding as evidence. Evidence is code: a call site, a file:line, a git log result.

Stopping

Finding nothing is a successful run, and in a repo this skill has already swept it is the expected outcome most of the time. Do not lower the evidence bar to produce output. On an empty run: open nothing, le

…(truncated)