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`) — landed by Task 2 of this spec.

**Read [workflow.md](workflow.md) for the full phase-by-phase execution. Read [phases.md](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`:

```bash
FLOWCTL="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/flowctl"
[ -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).

```bash
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
```

| Mode | When | Behavior |
|------|------|----------|
| **Interactive** (default) | User is at the terminal | Ask via blocking-question tool when an entry's content suggests overriding the mechanical default; confirm Phase 4 cleanup; show triage summary before writes |
| **Autofix** (`mode:autofix` in arguments) | Ralph or batch usage | No user questions. Apply mechanical defaults for every entry. Override only when the agent has high-confidence evidence from the entry body. Mark genuinely ambiguous entries as `needs-review` in the report. Default-decline Phase 4 cleanup. Print full report |

### 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_LEG