retrospective — from recurring friction to structural enforcement

The baseline's enforcement funnel has three rungs: a landmine (a remembered failure pattern surfaced by advisory memory hooks), an advisory hook (a PreToolUse hook that surfaces the landmine inline before the risky call), and a hard gate (a blocking hook or consent gate). Friction that recurs should climb this funnel instead of being re-litigated every session. This skill is the climb's entry point.

This skill is a generator, not a workflow phase. It enters no Track Guard ordering and blocks nothing. It writes ONLY memory entries (via the canonical entry shape) and an inline report. It never edits hooks, never touches seed.md/CLAUDE.md, never creates enforcement — those are amendment-path changes the user drives.

Inputs

Gather in main context (Article II — no subagent judgment):

• The cycle window: since the last release tag, the epic's first commit, or a user-named range.
• The landmines and decisions categories under .claude/memory/ — what is already recorded. Read them shape-agnostically: a sharded store keeps one fact per file under <category>/, a flat one keeps ## key blocks in <category>.md.
• Friction evidence inside the window: guard-trip messages in .claude/state/logs/ (hook log lines), docs/rca/*.md postmortems, integrate-failure yields in .claude/state/harness/*.log, and user corrections you can cite verbatim from the session or _resume.md.

Method

1. Collect candidate frictions

Sweep the inputs for events that happened more than once in the window, or once with high cost (a broken commit, a failed release, a consent near-miss). One-off annoyances are not retro material.

2. Classify each recurring friction

• Already a landmine? Re-verify the entry (Article IX.2); refresh last-touched and strengthen the entry body with the new occurrence.
• New pattern? Draft a landmines.md entry in the canonical shape (stable key, occurrence citations, verified-at, verbatim quote when the source is a user correction).
• Not memory-shaped (a spec gap, a missing test)? Route it to backlog.md as an entry instead — the retro records it; a future workflow fixes it.

3. Propose graduations

For each landmine with ≥ 2 recorded occurrences ACROSS cycles, assess the next rung:

Current rung	Graduation	What it takes
landmine entry	advisory hook surfaces it inline (e.g. process_lifecycle_guard pattern)	seed.md §4.1 amendment + hook change — user-driven
advisory hook	hard block / consent gate	same amendment path, stronger justification

A proposal names: the landmine key, occurrence count, the proposed rung, the hook/gate shape, and the cost of NOT graduating. It is advisory — put it in the report, not in any enforcement file.

4. Report

# Retrospective — <window>

## Recurring frictions (N)
- <key> — <occurrences, cost, where recorded>

## Memory writes
- <landmines.md/backlog.md key> — new | strengthened | re-verified

## Graduation candidates (advisory)
- <key> — landmine → advisory hook: <one-line shape> — evidence: <counts>

## Not retro material (dropped)
- <item> — <why>

Constraints

• Memory writes only — landmines.md / backlog.md / decisions.md in the canonical entry shape, respecting size-cap: 500 and verbatim-provenance rules (Article IX). No other file writes.
• No enforcement changes. Proposing is this skill; amending is the user's seed.md §4.1 path.
• Evidence or it didn't happen. Every friction cites its occurrences (log line, RCA path, commit, or verbatim correction). No vibes-based landmines.
• Read-only outside memory. No state files, no workflow.json, no consent paths.