/flow-next:audit — agent-native memory staleness review

Memory entries decay. A .flow/memory/bug/runtime-errors/ entry logged six months ago might reference a renamed file, a deleted function, or a codepath that no longer exists. Without periodic review, the store accumulates zombie entries and memory-scout surfaces outdated advice.

This skill IS the audit. The host agent (Claude Code / Codex / Droid) walks .flow/memory/, reads each entry, uses Read/Grep/Glob/git to verify references against the current codebase, applies engineering judgment, and decides per entry whether to Keep / Update / Consolidate / Replace / Delete / Harden. Optional autofix mode applies unambiguous actions and marks ambiguous as stale.

Harden is the graduation path: a lesson that keeps getting re-learned and states a mechanically checkable rule should stop riding the context window and become a gate. The audit proposes an artifact in a surface the repo already has — a lint rule, a CI step, or a rule in the substantive CLAUDE.md / AGENTS.md — verifies the gate actually fires, and only then demotes the entry to a pointer at it (file stays on disk, provenance intact). Propose-and-confirm by design: gate surfaces are shared repo infrastructure, so Harden never applies unattended.

Decision entries (.flow/memory/knowledge/decisions/) and glossary terms (GLOSSARY.md files at the repo root and on the ancestor chain) are walked alongside the rest of memory. Decisions get a calibrated judging question — "does the constraint that motivated this choice still hold?" — and Replace becomes a two-step supersession (write successor, mark old decision_status: superseded, never git rm). Glossary terms are scanned for code usage; zero-hit terms get a <!-- stale: ... --> HTML comment via Edit tool (no flowctl glossary mark-stale exists), _Avoid_ aliases appearing in code surface as alias-creep findings.

There is no Python audit-engine, no codex/copilot subprocess dispatch, no deterministic scorer. The host agent is already an LLM and does the work directly. flowctl provides only thin persistence plumbing (memory mark-stale, memory mark-fresh, memory mark-hardened, memory search --status). All judgment — is this recurring, is it mechanizable, which gate surface, what should the rule say — stays in this skill; there is no flowctl gate subcommand and never will be.

Read workflow.md for the full phase-by-phase execution. Read phases.md for the 6-outcomes lookup with memory-schema-specific calibration.

Preamble

CRITICAL: flowctl is BUNDLED — NOT installed globally. which flowctl will fail (expected). Define once; subsequent blocks (here and in workflow.md) use $FLOWCTL:

FLOWCTL="${CODEX_HOME:-$HOME/.codex}/scripts/flowctl"
[ -x "$FLOWCTL" ] || FLOWCTL="<plugin-root>/scripts/flowctl"   # <plugin-root> = the directory two levels above this skill's SKILL.md file (the harness gave you that file's absolute path when the skill loaded); substitute it literally
[ -x "$FLOWCTL" ] || FLOWCTL=".flow/bin/flowctl"

Inline skill (no context: fork) — plain-text numbered prompt must stay reachable across phases. Subagents can't call plain-text numbered prompts (Claude Code issues #12890, #34592). Phase 3 (Ask) and Phase 6 (Discoverability check) both require user choice in interactive mode.

Mode Detection

Parse $ARGUMENTS for the literal token mode:autofix. If present, strip it from the arguments — the remainder is the scope hint.

RAW_ARGS="$ARGUMENTS"
MODE="interactive"
if [[ "$RAW_ARGS" == *"mode:autofix"* ]]; then
  MODE="autofix"
  # Strip token, collapse whitespace, trim.
  SCOPE_HINT=$(printf "%s" "$RAW_ARGS" | sed 's/mode:autofix//' | tr -s ' ' | sed 's/^ //;s/ $//')
else
  SCOPE_HINT="$RAW_ARGS"
fi

Autofix mode rules

• No user questions. Never call the plain-text numbered prompt.
• Process all entries in scope. No scope-narrowing question. If no scope hint was provided, process every categorized entry.
• Attempt all safe actions. Keep (no-op), Update (write tool), Consolidate (merge + git rm subsumed), auto-Delete (only when code AND problem domain both gone), Replace (only with sufficient evidence to write a trustworthy successor).
• Never apply Harden. Classify and report Harden candidates (and un-graduation proposals) under Recommended with full detail — gate type, draft artifact, evidence, the --gate-ref that would be recorded — but write no artifact and demote no entry.
• Mark ambiguous as stale. When classification is genuinely ambiguous (Update vs Replace vs Consolidate vs Delete) or Replace evidence is insufficient, run flowctl memory mark-stale <id> --reason "..." instead of guessing. Stale-marking writes are atomic and round-trip safe.
• Conservative confidence. Borderline cases get marked stale; never deleted on autofix.
• Always print the full report. The report is the sole deliverable — there is no user to ask follow-ups.

Interaction Principles (interactive mode only)

In autofix mode, skip user questions entirely and apply the rules above.

In interactive mode, follow these principles:

Ask the user via plain text. Render the options below as a numbered list 1. … N., followed by a final option N+1. Other — type your own answer. Print the question, then the numbered list, then stop and wait for the user's next message before continuing. Parse the reply as: a bare number 1–N+1 → that option; the literal text of an option label → that option; free text after Other → custom answer.

• Ask one question at a time via plain-text numbered prompt. Never silently skip the question.
• Prefer multiple choice when natural options exist.
• Lead with the recommended option and a one-sentence rationale.
• Do not ask the user to make decisions before evidence is gathered — Phase 1 investigates first, Phase 3 asks.
• Group obvious Keeps and obvious Updates together for batched confirmation. Present Consolidate / Replace / Delete one at a time.

The goal is automated maintenance with human oversight on judgment calls — not a question for every finding.

Forbidden

• Auditing legacy flat files (.flow/memory/pitfalls.md, conventions.md, decisions.md at the memory root). Skip with a warning that recommends /flow-next:memory-migrate first. Report includes the skipped count.
• Auditing under _audit/, _review/, or any other _* directory under .flow/memory/.
• Deleting silently. Delete is reserved for unambiguous cases (code gone AND problem domain gone). Default to Replace or Consolidate when there's still value to preserve.
• git rm on superseded decision entries. Decision history stays on disk. Replace for knowledge/decisions/ entries means write a new entry and mark the old decision_status: superseded with superseded_by: <new-id> — never delete the old file.
• Deleting glossary terms. When a term has zero code hits, mark stale via Edit-tool HTML comment. Removing the term entry is the operator's call, surfaced in the report.
• Auto-applying Harden. In mode:autofix (and therefore any pilot / Ralph invocation) Harden never applies: no gate artifact is written, no entry is demoted, no un-graduation is executed. Candidates surface under Recommended only. Graduation edits files outside .flow/memory/ — lint config, CI, CLAUDE.md — and silent edits to shared repo infrastructure from an autonomous sweep are unacceptable. Audit proposes; a human accepts.
• Demoting a lesson to a gate that was never verified to fire. memory mark-hardened runs only after the gate is confirmed live (resolved lint config / a job that actually runs / the substantive instruction file). Verification failure leaves the entry active and reports a failed graduation. A gate that does not fire is worse than no gate.
• git rm on Harden. Ever, on any track. The entry file stays on disk as a pointer at the gate — that is what keeps "why does this rule exist?" answerable.
• Scaffolding infrastructure to host a gate. Never create a linter setup, a CI pipeline, or a config file that does not already exist. The gate lands in a surface the repo already has, degrades to the substantive instruction file, or the entry stays Keep.
• Inventing flowctl subcommands beyond what ships (memory mark-stale, memory mark-fresh, memory mark-hardened, memory search --status). There is no flowctl gate subcommand — the gate artifact is skill-authored prose/config written via Edit/Write. fn-38 task 2 ships only glossary {add,list,read,remove} — there is no flowctl glossary mark-stale; use Edit tool. Use Write tool + git for moves and deletes.
• Mass-renaming code from a glossary alias-creep finding. The audit reports file:line locations and stops there; code rename is the operator's call.
• Auto-committing without user awareness in interactive mode. Phase 5 detects git context and asks. Autofix uses sensible defaults.
• Setting context: fork — plain-text numbered prompt must stay reachable.
• Running parallel replacement subagents. Investigation subagents can run in parallel for 3+ independent entries; replacement subagents run sequentially to protect orchestrator context.

Workflow

Execute the phases in workflow.md in order:

0. Discover & Triage — walk .flow/memory/{bug,knowledge}/<category>/, group by module / category, count, choose interaction path (focused / batch / broad), skip legacy + _* directories with a counted warning. knowledge/decisions/ entries are picked up automatically by the same glob. 0.5 Glossary scan — enumerate GLOSSARY.md files via flowctl glossary list --json; per term, grep tracked code for the term and each _Avoid_ alias (case-insensitive whole-word, normalized whitespace); zero hits + zero alias hits → mark stale via Edit tool (HTML comment after the term heading); alias hits → surface as alias-creep finding for Phase 3 (interactive) or report (autofix); skip husk files (count: 0) with a single advisory. 0.75 Change-detection pre-filter — pre-scan recurrence artifacts (## Update heading count, entry-file commit count, related_to size) BEFORE the auto-Keep decision, so a recurrence-qualified entry reaches Phase 1 even when its module is unchanged; hardened entries get a cheap gate-liveness check instead of re-investigation.
1. Investigate — per entry: read frontmatter + body, verify referenced files / symbols / modules against current code via Read / Grep / Glob, check git log in the area, form Keep / Update / Consolidate / Replace / Delete / Harden recommendation with 2-4 evidence bullets and confidence. For 3+ independent entries, dispatch parallel investigation subagents (read-only). Decision entries use the calibrated judging question — "does the constraint still hold?" — see phases.md §Decision-entry calibration. 1.75 Cross-doc analysis — compare entries sharing module / category for overlap (problem, solution, root cause, files), supersession (newer canonical entry covers older narrower precursor), contradictions.
2. Classify — apply phases.md decision criteria and the outcome-precedence rule (correctness > Consolidate > Harden). For Replace, verify evidence is sufficient to write a trustworthy successor; mark stale otherwise. For decision entries, Replace = supersede (write new entry; mark old decision_status: superseded, superseded_by: <new-id>; never git rm the old). Harden requires a recurrence signal AND mechanizability, and passes the duplication guard first.
3. Ask — interactive only; autofix skips. Group obvious Keeps + Updates → confirm batch. Present Consolidate / Replace / non-auto-Delete individually. Present each Harden candidate individually with gate type, draft artifact, evidence bullets, and accept / different-gate-type / decline. Surface glossary alias-creep findings per alias. Lead with recommendation. One question at a time.
4. Execute — Keep: no edit. Update: agent edits frontmatter / body via Write tool, preserving unknown fields. Consolidate: merge unique content into canonical, git rm subsumed. Replace: write new entry, git rm old (decisions: write new + edit old's frontmatter to mark superseded, never git rm). Delete: git rm (only when code AND problem domain both gone). Harden: write the artifact, verify the gate fires, then flowctl memory mark-hardened <id> --gate-ref "<path>#<rule-id> -- <note>" — never git rm; verification failure leaves the entry active. Glossary stale: Edit comment after term heading. Ambiguous in autofix: flowctl memory mark-stale.
5. Report + Commit — print Kept / Updated / Consolidated / Replaced / Deleted / Hardened / Marked-stale / Skipped counts plus per-entry detail and a Glossary section (Kept / Marked stale / Alias-creep / Husks). Detect git context (current branch, dirty tree). Interactive: ask commit options. Autofix: branch-and-PR on main, commit on feature branch, stage only audit-modified files.
6. Discoverability check — verify the substantive CLAUDE.md / AGENTS.md (the one not just @-including the other) mentions .flow/memory/ with schema basics (track / category / module / tags / status) and when to consult. Add a minimal line if missing — interactive asks consent, autofix surfaces as recommendation.

Output rules

The full report is the deliverable — print it as markdown to stdout. Do not summarize internally and emit a one-liner.

Report structure (see workflow.md §5 for full schema):

Memory Audit Summary
====================
Scanned: N entries
Skipped legacy: M (run `/flow-next:memory-migrate` first to make these auditable)

Kept: X
Updated: Y
Consolidated: C
Replaced: Z
Deleted: W
Hardened: H  (failed graduations: HF; un-graduated: HU)
Marked stale: S

Glossary
--------
Files scanned: F (H husks)
Terms scanned: T
Kept: K_g
Marked stale: S_g
Alias-creep flagged: A_g

Then per-entry detail (id, classification, evidence, action taken). For Consolidate: which entry was canonical, what unique content was merged, what was deleted. For Replace: what the old entry recommended vs what current code does, path to successor (decision Replace also notes the old entry now carries decision_status: superseded). For Marked stale: why ambiguous. For Harden: gate type, artifact path, --gate-ref, and how the gate was verified live (a failed graduation names the reason and states the entry was left active). For glossary terms: only stale + alias-creep cases get per-term lines (Keep is silent); husks get a one-line advisory each.

Autofix mode splits actions into Applied (writes succeeded) and Recommended (writes failed — e.g. permission denied — plus every Harden candidate, which autofix never attempts). The structure is the same; only the bucket differs.