feat(skills): add memory-tidy, the judgment half of the memory plane

CONTRIBUTING.md

@@ -56,6 +56,27 @@ Beyond the table, see [`docs/PHILOSOPHY.md`](docs/PHILOSOPHY.md) "What we explic
 4. Use conventional commits: `feat(scope): ...`, `fix(scope): ...`, `docs: ...`. One logical change per commit. Keep spec/task IDs OUT of the subject line (no `TASK-3`, no trailing `(SPEC-002 ...)` tags); that context belongs in the body or PR description.
 5. Update `docs/CHANGELOG.md` under an `[Unreleased]` section if your PR is non-trivial (the root `CHANGELOG.md` is a thin pointer stub, SPEC-185). The maintainer moves it to a versioned section at release time.
 
+## Introducing a component (skill, command, agent)
+
+A new component never lands as a bare file drop; it registers everywhere the kit
+already tracks its kind, in the same PR:
+
+1. **Name**: kebab-case, and the plain-words rule below applies to the name
+   itself (`memory-tidy`, `skill-review`, `get-api-docs`, a non-engineer PM
+   should parse it cold). Config keys derived from it use snake_case
+   (`memory_tidy`).
+2. **Description discipline (skills)**: frontmatter `description` starts with
+   "Use when ..." and lists ONLY triggering conditions, never a summary of the
+   workflow (an agent will follow a summarized description instead of reading
+   the body).
+3. **Register it**: add the README roster row for its kind (Skill / Command /
+   Agent table). If it has a consumer-side toggle or an unattended cadence, add
+   a `[consumer]`-tagged key to `kit.toml` `[features]`, the harness config is
+   the single control surface; skills and commands themselves always install
+   (no `[modules]` gate).
+4. **Evidence (skills)**: the PR body carries writing-skills RED/GREEN
+   evidence (baseline failure without the skill, compliance with it).
+5. **Changelog**: step 5 above applies, a new component is always non-trivial.
 
 ## Plain words rule (2026-07-16)
 

README.md

@@ -290,7 +290,7 @@ Which hooks BLOCK vs warn vs neither is a declared contract: `docs/architecture.
 </details>
 
 <details>
-<summary><b>Agents</b> (25, dispatched by commands) and <b>Skills</b> (3, Claude-triggered)</summary>
+<summary><b>Agents</b> (25, dispatched by commands) and <b>Skills</b> (4, Claude-triggered)</summary>
 
 | Agent | Dispatched by | What it does |
 |-------|--------------|-------------|
@@ -323,6 +323,7 @@ Which hooks BLOCK vs warn vs neither is a declared contract: `docs/architecture.
 | Skill | What it does |
 |-------|-------------|
 | get-api-docs | Fetches curated API docs via Context Hub before coding |
+| memory-tidy | Audits a repo's `.claude/memory` store: evidence-gated verdicts, PR-gated merges/deletions, index rebuild (judgment half of `stats memory-sweep`) |
 | skill-review | Reviews + promotes skill drafts staged by skill-curator |
 | stats | Queries/renders the ledger read plane (relocated from `lib/stats/skill/` per ADR-0034 so it actually installs) |
 
@@ -372,7 +373,7 @@ dwarves-kit/
   lib/board/parse-board.sh            The one structured BACKLOG.md parser other tools reuse (SPEC-146): `rows <file>` (id/status/full-line) and `queue-rows <file> <repo-name> <repo-root>` (allow-listed `#queue{repo=...,pointer=...}` token extraction -- charset gate, repo self-consistency, `../` traversal hardening, existence check; every failure is a skip with a stderr reason, never a hard error)
   lib/board/board-mirror.sh           The git<->Hermes kanban bridge engine (SPEC-147), backing `board.sh mirror`/`status`: extracts opted-in BACKLOG.md rows (reusing `lib/board/parse-board.sh`) + active mega-goal roadmaps into normalized rows, diffs them (bash + `jq`/awk keyed comparison, no DuckDB) against an incremental NDJSON snapshot, and loads via `hermes kanban` CLI verbs only. Reachable native states are `{triage, ready, blocked, done}` only -- `todo`/`running` were probed live and found to have no durable CLI-only path (auto-promote back to `ready` within seconds); `_target_native`/`_create_flags_for`/`_followup_for` document the full state-mapping + Hermes-CLI-reality findings. `apply-plan` decodes each op's argv over NUL-delimited jq output (never a templated shell string), so a multi-line card body is preserved as one opaque argv element end to end -- the fix for a real bug this build's own live dev-home E2E caught (a newline-delimited decode silently split a multi-line body into extra positional args, rejected by the real CLI, masked by a naive stub). `HERMES_BIN` overrides the binary for tests (a stub logs argv; no real Hermes calls in the automated suite).
   lib/board/board-writeback.sh        The git<->Hermes kanban bridge WRITEBACK leg (SPEC-149), backing `board.sh writeback`: sources `lib/board/board-mirror.sh` (not re-forked) for extract/hash/native-state machinery. `diff` reads each opted-in repo's live board (`hermes kanban --board <b> list --json`, one batched call per board) + the SPEC-147 mirror snapshot, and builds a validated changeset of rows whose Hermes status moved -- validating opted-in repo (defense in depth on top of the mirror's own filter), a legal `backlog.sh` reverse-mapped target status, and the row_hash CONFLICT RULE (a Hermes-side edit applies only if the row's current git-side hash still equals the snapshot's recorded value; git wins, always). A missing or corrupt snapshot REFUSES ALL edits (explicit error, nonzero exit) rather than degrading to "no conflicts, apply everything". `apply` builds the `chore/board-sync` branch in an ISOLATED `git worktree` off the CURRENT HEAD (never the caller's own checkout, never a stale ref -- this is what keeps a concurrent append-only writer's row safe), edits ONLY the Status column of matched rows (reusing `lib/board/backlog.sh`'s own `set`), commits with `actor=hermes` in the body, pushes, and opens a HELD PR via `gh pr create` (argv-only, never a templated shell string; never auto-merged). Snapshot refresh updates ONLY `hermes_status` (to stop re-diffing the same not-yet-merged move); `row_hash` passes through UNCHANGED (the git SoT hasn't actually changed until the PR merges) -- `mirror`'s own idempotence self-heals the row once it does. `GH_BIN` (mirrors `HERMES_BIN`) overrides the `gh` binary for tests; no real Hermes write call or `gh` API call happens in the automated suite.
-  skills/                       Claude-triggered skills (get-api-docs, skill-review, stats)
+  skills/                       Claude-triggered skills (one folder per skill; roster in the Skill table above)
   rules/                        Path-scoped coding-standard templates
   examples/hello-spec/          Demo: small CLAUDE.md + SPEC.md walkthrough
   tests/test-hooks.sh           Hook behavior assertions

docs/CHANGELOG.md

@@ -5,6 +5,17 @@ All notable changes to dwarves-kit are documented here.
 ## [Unreleased]
 
 ### Added
+- **`skills/memory-tidy`, the judgment half of the memory plane.** Pairs with the
+  read-only `stats memory-sweep` scanner (SPEC-136): evidence-required per-note
+  verdicts (KEEP / MERGE / STALE / UNSURE) via agent fan-out, danger check for
+  notes whose advice contradicts current policy, index rebuild with parity
+  checks, all confined to a worktree branch and shipped as a PR (the operator's
+  approval gate). Consumer cadence toggle `[features].memory_tidy` in `kit.toml`
+  (`[consumer]`, default off); the scheduled runner lives consumer-side
+  (ops-toolkit `tools/memory-tidy/`). CONTRIBUTING gains an "Introducing a
+  component" convention (naming, description discipline, registration,
+  evidence) so new skills/commands/agents never land as unregistered file
+  drops. First production run: tieubao/ops-toolkit#972.
 - **`sync` one-way insert-only push to a foreign team board** (`notion-taskboard`
   app, SPEC-003, implements ops-toolkit ID-138). Pushes a repo's board rows out
   to a team-OWNED Notion board (create page, never update, never read for merge,

kit.toml

@@ -83,6 +83,12 @@ learning_ledger  = true         # [consumer] the orchestration is an external sk
                                 #        (ops-toolkit/dotfiles, SPEC-126); the kit only
                                 #        collects (weekend-batch.sh) + routes. Toggle kept
                                 #        here so the harness config is the single surface.
+memory_tidy      = false        # [consumer] the unattended biweekly memory-audit cadence.
+                                #        The judgment skill itself ships with the kit
+                                #        (skills/memory-tidy, pairing with `stats memory-sweep`)
+                                #        and is always available; this flag gates only the
+                                #        scheduled consumer-side runner (e.g. ops-toolkit
+                                #        tools/memory-tidy LaunchAgent). Default off (opt-in).
 
 # =============== TEAM COLLABORATION (designed, not built , all off) ===============
 # Source: ops-toolkit/_meta/megagoals/team-mode/ (ROADMAP + DESIGN-01..07),

skills/memory-tidy/SKILL.md

@@ -0,0 +1,45 @@
+---
+name: memory-tidy
+description: Use when auditing, consolidating, or cleaning a repo's .claude/memory store, "dọn memory", "memory tidy", the biweekly memory audit, duplicate or stale memory notes, MEMORY.md index drift (notes missing from the index, index entries with no file), or when a scheduled run asks for the memory-tidy audit of a repository. NOT for the built-in machine-local auto-memory under ~/.claude/projects (not git-tracked, no PR gate) and NOT for editing a single note (just edit it).
+disable-model-invocation: false
+---
+
+# Memory tidy
+
+## Overview
+
+Audit a repo's git-tracked `.claude/memory/` store and ship the cleanup as a PR. Core principle: **every verdict cites checkable evidence, and deletions reach main only through a PR merge**, the PR is the operator's approval gate, so the audit itself never needs to ask permission.
+
+This is the judgment half of the kit's memory plane. The mechanical half is `stats memory-sweep` (lib/stats, SPEC-136): a read-only scanner that never writes. This skill consumes its output and is the only path that edits a store, always behind a PR.
+
+## Process
+
+1. **Branch in a worktree first** (native worktree tool). All edits and deletions ride this branch. Touching the store on the current branch is the failure this skill exists to prevent.
+
+2. **Mechanical pre-pass.** Run `uv run stats memory-sweep` from the kit's `lib/stats/` and filter its JSON for this repo's store: dead path refs per note, stale index entries. Always also diff both directions: the note files on disk vs the entries in `MEMORY.md` (files not indexed are invisible to sessions; entries with no file are ghosts).
+
+3. **Fan-out judgment.** Split the notes into 2-4 subsystem clusters (roughly one cluster per 40 notes; a tiny store is audited directly) and dispatch parallel read-only agents. Each agent reads every note in its cluster fully and returns, per note: a 1-line gist plus one verdict:
+
+   | Verdict | Required evidence |
+   |---|---|
+   | KEEP | distinct fact, referents spot-checked alive |
+   | MERGE into `<sibling>` | quote the overlapping claim from both notes |
+   | STALE | concrete reason: referenced path tested and gone, superseded by a NAMED doc/ADR/newer note, the event concluded (PR merged, tool graduated), or the rule now codified verbatim in an always-loaded file (CLAUDE.md) |
+   | UNSURE | what only the operator can answer |
+
+   A verdict with no checkable evidence is not actionable: re-check it or downgrade to UNSURE. UNSURE notes are never deleted; list them in the report.
+
+4. **Danger check.** A note whose recommended fix contradicts current policy is worse than stale, someone may follow it. Fold its still-true diagnostics into the policy-carrying note, then delete it, and say so in the PR body.
+
+5. **Apply.** Merges preserve incident detail and dates. Fix path-drift and dangling `[[wikilinks]]` in surviving notes. Before deleting any note, grep the store for references to its name and rewrite them.
+
+6. **Rebuild the index.** Every remaining note gets one short line in `MEMORY.md` derived from its frontmatter `description`. Verify: entry count equals note count, and a grep for every deleted note's name returns nothing.
+
+7. **Ship.** Follow the repo's own session-close conventions (lab log, backlog rows, proof/gate ledger, whatever that repo's CLAUDE.md requires), commit, push, open a PR whose body lists each removal with its reason. If nothing needed changing, create no branch and report CLEAN.
+
+## Red flags
+
+- Editing or deleting notes without the worktree branch: stop, branch, start over.
+- "Obviously stale" with no tested referent.
+- Keeping a doubtful note because deleting feels safer: the honest form is UNSURE + operator list.
+- Hand-editing index lines instead of deriving them from frontmatter.