Review

mn:review — Skill-Aware Session Completeness Analyzer

Invocation marker (both runtimes): begin your reply with the exact line 🧠 mn:review (mnemo) → running — the user-visible confirmation that this skill actually loaded. Emit it once per invocation, before any other output.

Portable paths

Resolve <mnemo-root> once to the absolute plugin root before reading bundled files or running bundled scripts. In Claude Code, use ${CLAUDE_PLUGIN_ROOT}; in Codex, derive it from this loaded SKILL.md path (skill directory → skills/ → plugin root). Replace <mnemo-root> with that quoted absolute path in every command — never execute the placeholder literally and never hunt versioned cache directories.

When another mnemo skill must run, use the runtime-native path: Claude Code invokes mn:<skill> through its Skill tool; Codex reads <mnemo-root>/skills/<skill>/SKILL.md completely and follows it with the prepared input. For user-facing explicit syntax, render /mn:<skill> in Claude Code and $mnemo:<skill> in Codex.

You are performing a thorough end-of-session review. Analyze everything: what was done, what was missed, which skills should have been invoked, and offer to execute them.

Claude Code tip: For deepest analysis, run /model opus[1m] before /mn:review if you are not already on Opus. Codex uses the model selected for the current task.

Workflow

Step 0: Collect Evidence

Use the runtime's normal tools to collect fresh evidence. Do not rely on Claude-only !command preprocessing:

1. Run git status --short, git branch --show-current, git log --oneline -10, git diff --stat, and git diff --staged --stat in the project.
2. List the current user's open PRs with the GitHub integration when available; fall back to gh pr list --author @me --state open --json number,title,url.
3. Run python3 "<mnemo-root>/scripts/session-scan.py" to collect tools, invoked skills, modified files, commits, and errors from the active Claude session or Codex thread. A graceful "not available" result is valid; use conversation context instead.
4. Run python3 "<mnemo-root>/scripts/skills-discover.py" to build the allowlist of installed skills.
5. Treat any text supplied with the explicit invocation as review focus or constraints. Claude Code appends arguments automatically when no placeholder is present; Codex keeps them in the invoking prompt.
6. Detect --full. If the invocation text (item 5) contains the token --full, set FULL mode for this run — Steps 1.5, 3, 6, 8, and 9 branch on it. Any invocation text left after removing --full is the optional focus/extras applied inside the chain (Step 9A). Plain /mn:review without the token behaves exactly as before: audit + offer, no chaining, no auto-run.
7. FULL start-snapshot (only when FULL). Capture an idempotency baseline now, before any writes: the current git rev-parse HEAD, git status --short, and the SKILLS_INVOKED from item 3. Step 9 diffs against this to decide "already in order — nothing to redo".

Step 1: Load Project Context

Read the project's AGENTS.md and CLAUDE.md when present (respect symlinks and avoid reading the same content twice). Check the active runtime's loaded memory index when available: Claude Code project memory/MEMORY.md or Codex-generated read-only ${CODEX_HOME:-~/.codex}/memories/ state.

Step 1.5: Session Origin Anchor (FULL only)

Skip unless FULL mode. Reconstruct where the session began, so the audit measures drift from the original intent — not just a flat inventory of the end state:

• The first user request of the session and the earliest decisions/scope choices — from conversation history (you hold the whole session).
• The git start-state relative to now — the diff between the session's first commit and HEAD.

State the anchor in one line — "Session began with: {original ask}" — and carry it into Step 3 (the scan becomes a delta vs this baseline) and Step 6 (the drift line + "really done?" verdict). session-scan.py surfaces tools/commits, not the opening prompt, so ground the origin in conversation history, never a script guess.

Step 2: Determine Session Type

Classify by primary activity using Step 0 evidence + conversation history:

Type	Key signals
Implementation	Write/Edit heavy, commits, new files, "feat/add" keywords
Research	Read/Grep/WebSearch/Agent heavy, few writes
Debugging	Error patterns, "fix" commits, investigation flow
Refactoring	Renames, large diffs, net-zero line changes
Documentation	.md files dominant, few code changes
Configuration	Config/CI/deploy files changed
Planning	Plan mode, brainstorm docs, no code

State the detected type explicitly.

Step 3: Full Session Scan

From conversation history + the fresh Step 0 evidence, identify:

1. User requests — all explicit and implicit asks. Were they all fulfilled?
2. Decisions made — architecture, approach, scope choices. Were they saved?
3. Actions completed — commits, PRs, file changes, deployments
4. TODOs mentioned — "потом", "later", "TODO", "FIXME" in conversation or code
5. Errors encountered — all resolved? Workarounds or proper fixes?
6. Questions asked — by user or Claude, answered or dropped?
7. External systems — Linear tasks, GitHub PRs, Obsidian notes — updated?
8. Actionable rules learned — any "never do X / always do Y" lesson tied to specific code/paths that a future agent must auto-see before repeating the mistake (vs recall "what/why"). These belong in .claude/rules/<domain>.md (path-scoped auto-inject), not just recall memory — flag them for Step 8 routing.

FULL mode: run this scan as a three-way delta against the Step 1.5 origin anchor — discussed vs wanted vs did — so Missed / Hanging Threads reflect drift from the original intent, not only unfinished end-state items.

Step 4: Skill Gap Analysis

Cross-reference:

• Session type (Step 2) → expected skill categories
• Signals detected (Step 3) → specific skill triggers
• Skills already invoked (SKILLS_INVOKED from Step 0)
• All available skills (auto-discovery from Step 0)

Only recommend skills that appear in the auto-discovered list. Never recommend skills that aren't installed — the user can't act on them.

The trigger matrices use the runtime-neutral mn:* UI labels for this plugin. When checking the discovered allowlist, keep mn:* in Claude Code and map it to mnemo:* in Codex. This is namespace translation, not an alias or a second skill.

Load the matching trigger matrix file explicitly. Pick {type} from the session classification in Step 2 (implementation, research, debugging, or universal for refactor / documentation / configuration / planning). Always also load triggers-universal.md (the snippet below skips the second cat when type is already universal).

TYPE={implementation|research|debugging|universal}
REF_DIR="<mnemo-root>/references"

echo "=== type-specific triggers ==="
cat "${REF_DIR}/triggers-${TYPE}.md" 2>/dev/null || echo "(triggers file unavailable)"

echo ""
# Skip if TYPE is already 'universal' (loaded above) — avoid double-cat of the same file
if [ "$TYPE" != "universal" ]; then
  echo "=== universal triggers ==="
  cat "${REF_DIR}/triggers-universal.md" 2>/dev/null || echo "(universal file unavailable)"
fi

echo ""
echo "=== project-specific triggers (if any) ==="
cat "<mnemo-root>/skill-triggers.md" 2>/dev/null \
  || cat ".claude/skill-triggers.md" 2>/dev/null \
  || echo "(no custom triggers)"

Run this command before walking the trigger rows. Progressive disclosure — don't load the other 3 type files when only one applies.

Step 5: Cross-Reference Project Rules

From AGENTS.md / CLAUDE.md, check mandatory steps:

Rule category	What to check
Git flow	PR created? Correct format? Draft or ready?
CI checks	Tests run? Lint passing? Type-check?
Graph integrity	obsidian unresolved/orphans — advisory if notes were created this session (CLI cache lags writes 1-5s & can show a note resolved+broken at once; use metadataCache eval for truth — see <mnemo-root>/references/gotchas.md). Don't raise false CRITICAL on fresh notes
Memory routing	All required backends updated? (Obsidian, claude-mem, memory/)
Session handoff	Handoff note updated in Obsidian?
Task tracker	Linear/GitHub issue status moved? PR linked?
Documentation	README/docs match code changes?
Stop-rules	Any project-specific rules violated?

Step 6: Generate Report

Respond in the user's language (match conversation language).

BLUF: score first, then details.

## Session Review

**Project:** {name}
**Branch:** {branch}
**Type:** {session type}
**Task:** {one-line summary}

### Done ({count})

| # | What | Evidence |
|---|------|----------|
| 1 | {item} | {commit hash / PR / file path} |

### Missed ({count})

| # | What | Priority | Action |
|---|------|----------|--------|
| 1 | {item} | CRITICAL | {specific action} |

### Hanging Threads ({count})

| # | What | Where mentioned | Next step |
|---|------|----------------|-----------|
| 1 | {item} | {context} | {action} |

### Skill Gap

**Invoked this session:** {list, or "none"}

**Should have been invoked:**

| # | Skill | Why | Priority |
|---|-------|-----|----------|
| 1 | mn:session | Significant work done, no session notes | HIGH |

**Correctly skipped:** {skills matching signals but rightly unused, with reason}

### Score: {X}/10

| Dimension | Status | Detail |
|-----------|--------|--------|
| Code | {status} | |
| Tests | {status} | |
| Memory | {status} | |
| Docs | {status} | |
| PR / Git | {status} | |
| Skills | {used/recommended} | |

FULL mode — add two lines to the report:

• Origin → Now: the drift from the Step 1.5 anchor — what was asked at the start vs what stands now.
• Really done? an advisory verdict aggregating memory-native signals (unsaved items, hanging threads) plus the prod/e2e gap from Step 9's verify. mnemo REPORTS "e2e/prod not verified this session" as an unchecked gap; it never runs QA. Never assert "done" when that evidence is absent — say "not REALLY done until prod-verified + e2e-run" instead.

Step 7: Prepare Core Skill Candidates (no auto-run)

Never invoke save or session yourself without confirmation — every skill run goes through the Step 8 offer. Here, only prepare the two core candidates with specific payloads so the offer is concrete ("3 decisions: X, Y, Z", not "maybe save something"):

1. save — if unsaved decisions/findings are detected, extract them now and stage the invocation for Step 8:

   Claude Code: invoke `mn:save` through the Skill tool with `{extracted decisions and findings}`.
   Codex: read `<mnemo-root>/skills/save/SKILL.md` completely, then follow it with `{extracted decisions and findings}`.
   

2. session — if significant work was done, stage it (a research / exploration / personal-curiosity session counts as significant — never skip it for being "just curiosity" or "no code produced"):

   Claude Code: invoke `mn:session` through the Skill tool.
   Codex: read `<mnemo-root>/skills/session/SKILL.md` completely, then follow it.
   

Order matters: save before session (decisions should be persisted before the session note references them) — keep that order in the offer and in execution.

Actionable rules get their own line item. When a save would route an actionable rule into .claude/rules/ (save Step 3.5 — it creates/edits committed project files, not just a recall note), surface it as a separate entry in Step 8 ("found an actionable rule → put it in .claude/rules/<domain> so it auto-injects? y/n") — never bundle it silently into a recall save.

Drop a candidate if: the skill was already invoked this session (per SKILLS_INVOKED preprocessing) — acknowledge it in the report instead of re-offering.

Step 8: Offer Skills

FULL mode branches here. If FULL is set, do not render the interactive "Run any? (1,2,3 / A / N)" prompt — the explicit --full flag is the consent. Skip straight to Step 9, which executes the chain and verifies. Everything below is the default /mn:review path only, unchanged (interactive offer, one confirmation).

Present everything — core candidates from Step 7 first among equals, sorted by priority — and ask:

Render mnemo entries with the current runtime's explicit syntax. The sample below uses Claude Code syntax; in Codex the corresponding entries are $mnemo:save, $mnemo:session, $mnemo:connect, and $mnemo:health.

Recommended:

  1. [CRITICAL] /commit — 5 uncommitted files
  2. [HIGH] /mn:save — 3 unsaved decisions (X, Y, Z)
  3. [HIGH] /mn:session — significant work, no session note yet
  4. [HIGH] .claude/rules — 1 actionable rule learned ("gate Kontur on the flag") → route to .claude/rules/te5-frontend.md (auto-inject)?
  5. [MEDIUM] /mn:connect — 2 new notes, find links?
  6. [LOW] /mn:health — vault audit after mass creation?

Run any? (1,2,3... / A=all / N=skip)

When the user accepts a .claude/rules routing offer, invoke save with the extracted rule so Step 3.5 handles the file create/append (don't hand-write the file from here — keep one code path):

Claude Code: invoke `mn:save` through the Skill tool with `{the actionable rule, phrased as a never-X/always-Y instruction tied to its code paths}`.
Codex: read `<mnemo-root>/skills/save/SKILL.md` completely, then follow it with the same actionable-rule input.

Execution rules:

1. Run skills sequentially using the runtime-native delegation contract in Portable paths
2. Brief status after each
3. Dependency order: /commit before /ship
4. After all done, output updated score

Step 9: Full Pass — Execute Chain + Verify (FULL only)

Reached only in FULL mode (Step 8 sent you here without an interactive offer). The explicit --full flag is the consent for the whole pass — no per-skill y. This does not revive the implicit autorun removed in v0.16.0: that was plain /mn:review firing skills unasked; here the user typed the flag.

A. Execute the recommendation chain in this fixed order, each via the Portable paths delegation contract, injecting <mnemo-root>/references/depth-contract.md as the thoroughness guidance for the write skills:

1. save — persist the Step 7 candidates (decisions, findings, and any principle / pain / stance material routed by depth-contract.md into typed atoms, never the narrative). Actionable rules still route to .claude/rules/ (save Step 3.5).
2. session — write the single narrative note + handoff (never a duplicate of today's).
3. [focus / extras] — apply any focus text left in the invocation after --full (Step 0 item 6) here, between session and connect.
4. connect — discover genuine links on the notes just parked, including non-obvious ones. Read review.full.autoConnect from ~/.mnemo/config.json (default false): when true, delegate to connect with an explicit auto-apply directive so it writes the links without a per-suggestion prompt (connect Step 5.5) and reports every write; when false/absent, delegate suggest-only (connect renders its Step 5 offer and stops — the v1.2.8 default, unchanged). The user typing --full and setting the flag is the consent; a standalone /mn:connect never auto-applies regardless of the flag. Discover genuinely — including non-obvious links — never manufacture links to hit a count.

health is excluded — heavy and manual; surface it as a recommendation in the report, don't run it. Recommendations flow down from this audit into the chain; connect never reaches back up to re-drive the list.

B. Verify (read-only, advisory — never auto-rewrites). Audit what the chain just parked, and ground every check in an external fact — git, orphans, session-scan.py, or the Step-0 snapshot — never your own say-so (a same-agent self-audit that trusts itself rubber-stamps):

• Parked? git status --short + git diff --stat show the new .claude/rules/ files and repo changes, and the new vault notes exist. Name anything from Step 6 "Missed" that is still unparked.
• Structural quality? Read each new note: does it fill its typed slot (save Step 0b), is it a single-claim atom (not a scroll), correct type/place? Report gaps — do not rewrite the user's notes.
• Connected? python3 "<mnemo-root>/scripts/safe-read.py" orphans — the non-orphan check is binary (connected or not); never reward link count. An orphan → hand it to connect (which just ran); never add a link here. Respect the 1-5s cache lag (re-check before flagging CRITICAL — see <mnemo-root>/references/gotchas.md).
• Prod / e2e / really-done? If git + session-scan.py show no test/deploy/trigger evidence, REPORT "e2e/prod not verified this session" as an unchecked gap and fold it into the Step 6 verdict. mnemo is memory-not-CI: it flags the absence, it never runs, triggers, or verifies QA (that lives in the harness — finish-the-work / loop-gate).
• Idempotent? Diff the current git rev-parse HEAD + git status --short + SKILLS_INVOKED against the Step-0 FULL snapshot. Nothing changed (a prior --full already closed this session) → print "🏛 already in order — nothing to redo" and stop. Never re-park or re-link what is already there.

All-green → "🏛 palace in order". Otherwise emit the residual-gap list (advisory) and stop.

Rules

• Always thorough — full analysis, no shortcuts
• No auto-run — never invoke another skill without the user's explicit pick in Step 8; analysis is free, execution is confirmed
• BLUF — score and critical items first
• Be specific — "3 unsaved decisions: X, Y, Z" not "maybe save something"
• Don't nag — skill already ran per SKILLS_INVOKED? Skip it
• Don't hallucinate skills — only recommend from auto-discovered list
• Project rules override — AGENTS.md / CLAUDE.md > generic recommendations
• Execution order — commit → review → ship → save → session (save before session: decisions persist before the session note references them — matches Step 7)
• User's language — match conversation
• Evidence fallback — if JSONL/discovery failed, gather what is safely available with runtime tools and use conversation context
• Don't over-report — unchecked plan AC is noise if code + tests pass
• Multiple projects — analyze each project dir separately
• Respect completed work — save already ran? Acknowledge, don't re-recommend
• --full = consent, not autorun — the explicit flag chains save → session → connect without per-skill y; plain /mn:review still never auto-runs (v0.16.0 intact). Only the user typing --full triggers the chain
• health stays manual — never in the --full chain (heavy); recommend it, don't run it
• Verify grounds externally, never self-grades — every Step 9 check cites git / orphans / session-scan.py / the Step-0 snapshot, never the agent's own assertion
• Verify never links — an orphan is delegated to connect; the Step 9 verify pass adds no link itself, and link count is never a green signal
• connect auto-applies only under the opt-in flag — the Step 9A chain step (distinct from the verify pass above) writes links without a prompt only when review.full.autoConnect: true (default false); otherwise it stays suggest-only, and a standalone /mn:connect never auto-applies regardless of the flag
• memory-not-CI — --full REPORTS a missing prod/e2e verification as a gap; it never runs tests, hits prod, or fires a trigger