Flow Next Audit

/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 Without periodic review, the store accumulates zombie entries and memory-scout surfaces outdated advice.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 — and Replace becomes a two-step supersession (write successor, mark old decision_status: superseded, never git rm).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 subprocess judgment or deterministic classification. memory audit-scan --json collects mechanical evidence; memory apply --plan <file> --json persists host-authored actions. 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.

Read working-rules.md first unless you already have this run; it holds for every step of this skill.

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="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}/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) — AskUserQuestion must stay reachable across phases. Subagents can't call blocking question tools (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"* || "$RAW_ARGS" == *"mode:autonomous"* || "${FLOW_AUTONOMOUS:-}" == "1" || "${AUTONOMOUS:-}" == "1" ]]; then
  MODE="autofix"
  # Strip token, collapse whitespace, trim.
  SCOPE_HINT=$(printf "%s" "$RAW_ARGS" | sed 's/mode:autofix//; s/mode:autonomous//' | tr -s ' ' | sed 's/^ //;s/ $//')
else
  SCOPE_HINT="$RAW_ARGS"
fi

Autofix mode rules

• No user questions. Never call the blocking-question tool.
• Process all entries in scope. No scope-narrowing question. If no scope hint was provided, process every categorized entry.
• Attempt all safe actions. Keep (mark-fresh stamp only), 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 via AskUserQuestion (call ToolSearch with select:AskUserQuestion first if its schema isn't loaded). Fall back to numbered options in plain text only if the tool is unreachable or errors. 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.