Flow Next Memory Migrate

/flow-next:memory-migrate — agent-native legacy migration

Pre-fn-30 flow-next stored memory as three flat markdown files: .flow/memory/pitfalls.md, conventions.md, decisions.md. Each was a sequence of ----delimited segments with ad-hoc headings and no schema. fn-30 introduced the categorized schema (track / category / module / tags / status frontmatter, one entry per file). Existing flat files persisted but became invisible to memory list, memory search, and flow-next-audit because there's no frontmatter to scope or stale-flag.

This skill IS the migration. The host agent (Claude Code / Codex / Droid) reads each legacy entry, applies the mechanical default (track, category) from the source filename, overrides only when the entry's content warrants, and writes a categorized entry via flowctl memory add. Optional autofix mode accepts every mechanical default and marks ambiguous entries as needs-review in the report.

There is no Python classifier subprocess, no codex/copilot dispatch, no fast-model probability scoring. The host agent is already an LLM with full repo context and does the work directly. flowctl provides only thin parsing + persistence plumbing (memory list-legacy --json, existing memory add).

Read workflow.md for the full phase-by-phase execution. Read phases.md for the (track, category) decision tree with mechanical baseline + override examples.

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 1 (Classify) needs user choice on ambiguous entries in interactive mode; Phase 4 (Cleanup) needs consent before renaming originals. (sync-codex.sh rewrites this to a plain-text numbered prompt in the Codex mirror.)

Mode Detection

Parse $ARGUMENTS for the literal token mode:autofix. If present, strip it from the arguments — the remainder is the scope hint (a legacy filename like pitfalls.md to narrow the run, or empty to migrate all).

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 blocking-question tool.
• Process every legacy entry in scope. No scope-narrowing question. If no scope hint was provided, migrate all three legacy files.
• Mechanical default wins on borderline. Override only when the entry body unambiguously points at a different (track, category) (e.g. an entry titled "race condition in worker pool" inside pitfalls.md clearly warrants bug/runtime-errors over the mechanical bug/build-errors).
• Ambiguous → mechanical default + log as needs-review. Genuine "could be A or B" cases take the mechanical default and surface in the report so the user can re-classify later.
• Default-decline Phase 4 cleanup. Originals stay in place. Surface the rename suggestion as a recommendation in the report.
• 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 one question at a time 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 (always the mechanical default unless the body warrants otherwise) and a one-sentence rationale.
• Do not ask the user to make decisions before the entry has been read — Phase 1 reads first, asks second.
• Group obvious mechanical-default migrations together for batched confirmation. Present overrides and ambiguous cases one at a time.

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

Subagent dispatch (mostly N/A)

This skill runs almost entirely on the main thread. Phase 1's "one entry per tool call" rule means classification iterates serially in the orchestrator — there is no investigation step independent enough to dispatch in parallel. Cross-platform tool naming (Task on Claude Code, spawn_agent on Codex, platform-equivalent on Droid) is documented here only for the rare case where the agent needs to spawn a focused investigation subagent (e.g. resolving an ambiguous override by reading a referenced file): keep such dispatches read-only (Read / Grep / Glob), do not let subagents call flowctl memory add directly, and merge results back on the main thread before Phase 2.

Forbidden

• Migrating files outside MEMORY_LEGACY_FILES (pitfalls.md, conventions.md, decisions.md at .flow/memory/ root). Any other .md at the memory root is user data — leave it alone.
• Migrating entries inside categorized directories (.flow/memory/{bug,knowledge}/<category>/*.md). Those are already migrated; re-running on them is a bug.
• Auto-deleting legacy flat files. Phase 4 renames originals to .flow/memory/_migrated/<filename>.bak for traceability — never rm. User can git rm later if they want.
• Inventing flowctl subcommands beyond the shipped memory list-legacy. Phase 2 writes via existing flowctl memory add. Mechanical map is documented in phases.md so the agent doesn't need to call a flowctl helper for it.
• Batch-classifying multiple entries in a single tool call. Phase 1 enforces one entry per tool call. Agents under context pressure batch-classify in-prompt and silently skip entries (practice-scout flagged this real failure mode).
• Setting context: fork — blocking-question tools must stay reachable.
• Re-running on already-migrated files. Phase 0 checks .flow/memory/_migrated/<filename>.bak and skips with an "already migrated" log line.

Workflow

Execute the phases in workflow.md in order:

0. Detect & enumerate — run flowctl memory list-legacy --json, check _migrated/ for prior runs, apply scope hint, decide interaction path.
1. Classify (one entry per tool call) — for each entry: read title + body + filename context, default to mechanical (track, category), override only with body-driven evidence. Interactive: ask on ambiguity. Autofix: take mechanical default + log needs-review.
2. Write categorized entries — invoke flowctl memory add --track <t> --category <c> --title "..." --body-file <tmpfile> per classified entry. Slug uniqueness handled by existing helper.
3. Verify + Report — re-read newly created entries, print summary (legacy files processed, entries migrated, overrides, needs-review).
4. Optional cleanup — interactive: ask whether to rename originals to .flow/memory/_migrated/<filename>.bak. Autofix: default-decline + surface as recommendation. On first cleanup, write .flow/memory/_migrated/.gitignore containing * (self-ignoring directory pattern). Legacy originals are renamed, never deleted — a run that removed a legacy file has broken this.

Output rules

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

The report's exact shape lives in one place — workflow.md §3.2 (summary skeleton) and §3.3 (the autofix Applied / Recommended split). Render it from there; this file does not restate it.

What the report must carry either way: files processed, entries migrated, overrides, needs-review, plus per-entry detail (id, source filename, mechanical default, final classification, override rationale) and, for each needs-review entry, why the agent could not decide.

Done when: the full report reached stdout as markdown. A run that summarized internally and emitted a one-liner has broken this.