postmortem

name	postmortem
description	Session-end review of recent chiron activity. Scores across 5 axes (design thinking, code quality, idioms, testing, engineering maturity) and names one concrete thing to practice next time. Read-only in v0.3.0 — does not persist scores.
user-invocable	true
argument-hint	[optional session summary]
allowed-tools	Read, Grep, Glob, LS
compatibility	Run /teach-chiron first to generate .chiron-context.md

/postmortem — session-end review and scoring

Step 0 — Load project context

Check if .chiron-context.md exists in the project root. If it exists: read it. DO NOT scan the codebase or read additional files. If not: tell the user: "No project context found. Run /teach-chiron first." Then stop.

Optional user-supplied session summary (or blank):

$ARGUMENTS

CRITICAL — user instructions always win

If the current project's CLAUDE.md or AGENTS.md says to skip session reviews or use different scoring, follow those instructions instead. This command is opt-in.

---

Current level

Apply the voice level from .chiron-context.md. If missing or unrecognized, use default.

---

What this command does

Reviews the current Claude Code conversation's recent chiron activity and presents a session-end report. Based on chiron's "After each session" pattern:

1. Session summary — 1–2 sentences recapping what was worked on
2. Scores — /10 across 5 axes
3. One thing to practice — concrete, bounded next step

Read-only in v0.3.0. Does NOT write to ~/.chiron/profile.json. Persistence of postmortem scores over time is deferred — if proven useful in practice, add in v0.3.x.

Graceful degradation

If you re-read the recent conversation and find NO recent chiron activity (no /chiron, no /challenge, no /hint, no pervasive-mode chiron behavior signals), respond:

No recent chiron session to review. Run /postmortem after a /chiron or /challenge session, or after working on a coding task with pervasive chiron mode active.

Stop there. Do not invent a session to review. If the user is new to chiron or just started the conversation, /postmortem correctly reports nothing to review.

If $ARGUMENTS contains a user-supplied summary (e.g., /postmortem fan-out worker pool implementation), use that as context but still look at the recent conversation for the actual scoring evidence. The arguments are a hint about what the user cared about — not a substitute for evidence.

The 5 scoring axes

Axis	What you're scoring
Design thinking	Did the user propose a design before coding? Consider trade-offs? Identify the right abstractions?
Code quality	Correctness, readability, handling of edge cases, appropriate error handling.
Idioms	Idiomatic use of the language's patterns (e.g., Go errgroup, defer mu.Unlock, early-return, table-driven tests). Language-specific.
Testing	Was test design discussed or written? Were edge cases considered? Table-driven tests?
Engineering maturity	How did the user respond to feedback? Did they acknowledge mistakes, push back appropriately, or dig in defensively?

Each axis is scored /10. Be honest and specific. Never cruel.

Self-verification loop

After generating initial scores, verify each one before delivering (silent — the user sees only verified scores):

1. For each axis score, ask yourself one specific verification question:

   • Design: "What specific design decision did the user make (or fail to make) that justifies this score?"
   • Code: "What specific line or pattern in the code justifies points lost?"
   • Idioms: "Can I name the specific idiom that was used correctly or missed?"
   • Testing: "What specific test was written or should have been?"
   • Maturity: "What specific feedback exchange justifies this score?"

2. If you cannot answer the verification question with a specific, concrete example from the conversation: revise the score. A score without evidence is wrong by definition.

3. If two axes have the same score AND the same justification, one of them is wrong — they're measuring different things. Two axes with the same score for different reasons is fine.

This loop improves accuracy without adding output length.

Profile-informed trends (optional)

Read ~/.chiron/profile.json if it exists. This is a read-only operation — /postmortem never writes to the profile. The file is owned by /challenge's Step 8.

When profile data is available, use it to ground scoring in longitudinal evidence rather than only the current conversation:

1. Filter recent history. Last 30 days of entries, grouped by tag.

2. Identify trend signals:

• Recurring weakness: Tags with 2+ drill_attempted or drill_gaveup in the window
• Emerging mastery: Tags with 2+ drill_solved in the window and no recent failures
• Language/domain focus: Which languages and domains the user has practiced most in the window

3. Integrate into scoring. Use trend signals to refine the 5 axis scores where relevant:

• Idioms axis: If the current session touched a tag that's a recurring weakness in profile history, the idioms score reflects both the current evidence AND the historical pattern. Example: "Idioms 6/10 — the current session used errgroup.WithContext correctly, but profile history shows 3 missed attempts on this pattern over the last 2 weeks, so the grasp is still developing."
• Engineering maturity axis: Emerging mastery (profile shows recent improvement on a tag) counts positively: "Maturity 8/10 — accepted feedback on ctx shadowing immediately, and profile shows this is the first session where that pattern was solved cleanly."

4. Surface a trend callout in the session summary (at most one per session).

If a profile-backed trend is genuinely relevant to the current session, mention it in the Session line:

Session: Implemented fan-out worker pool with errgroup. Profile note: this is your third session touching go:errgroup-with-context and your first clean solve.

Rules for trend callouts:

• One trend mention per session, maximum. More is noise.
• Only if genuinely relevant. Don't force a trend callout when the current session doesn't intersect with profile data.
• No moralizing. State the trend as fact, not judgment. "Third session" yes. "Finally solved it" no.
• Never shame history. Past failures are data, not character flaws.

Schema-aware reading (v0.16.0+):

• schema_version === 1 or schema_version === 2 → supported. Read entries[] and compute trend data normally. In v1 profiles, ignore any top-level install_id field (it's unused legacy). Entry shape is identical across v1 and v2.
• schema_version is an integer > 2 → unknown future version. The file was produced by a newer chiron than this one; the entries may have a shape this skill doesn't recognize. Skip trend analysis entirely and surface ONE terse line in the Session summary: "Profile note: ~/.chiron/profile.json is schema_version <N>, newer than this chiron understands (max 2) — scoring on current conversation only." Fall through to conversation-only scoring.
• schema_version missing but entries is a valid array → assume v1. Same behavior as v1 above.
• schema_version is a non-integer, negative, or otherwise invalid type → treat as corrupt (see below).

Error handling (silent fallback — never crash):

• Profile file missing → no trend data, score only on current conversation, no trend callouts
• File unreadable / corrupt JSON → same behavior as missing. Do NOT rename or repair the file here — /postmortem is strictly read-only; repair is /challenge Step 8's responsibility.
• File is very large (>500 entries) → read last 200 entries only for trend analysis
• No relevant trend signals (all tags are one-off or no overlap with current session) → score only on current conversation, no trend callouts

Read-only invariant: Never write to ~/.chiron/profile.json from /postmortem. This preserves the v0.3.0 read-only contract — /postmortem never persists scores, renames the file, or modifies any chiron state file. The file is owned exclusively by /challenge's Step 8, which is the only writer and the only component allowed to perform migrations or backups.

Response format — keep it terse

Compact 3-section format. ~10 lines total instead of decorative ## headers:

Session: <1–2 sentence summary, not 2–3. Be specific about what was worked on.>

Scores:
- Design 7/10: <one-line justification>
- Code 8/10: <one-line justification>
- Idioms 6/10: <one-line justification>
- Testing 5/10: <one-line justification>
- Maturity 7/10: <one-line justification>

Practice: <concrete, bounded next step — specific exercise, not vague advice>

Example (shape reference, not content to copy):

Session: Implemented fan-out worker pool with errgroup, iterated through L0–L3 before landing on bounded concurrency.

Scores:
- Design 7/10: considered cancel-on-error; missed backpressure discussion
- Code 8/10: correct errgroup usage, clean goroutine lifecycle
- Idioms 6/10: shadowed ctx inside goroutine, unbuffered channel with known capacity
- Testing 4/10: no tests written, no edge cases discussed
- Maturity 8/10: accepted feedback on ctx shadowing immediately, pushed back on channel sizing with valid reasoning

Practice: Write a table-driven test for the fan-out function covering cancel-on-error and zero-input edge cases — `/challenge src/worker/pool.go`

Style rules:

• Session: one line, not a ## Session summary header block
• Score list is flat bullets with axis name + score + colon + justification; no **bold labels**
• Practice: one line, not a ## One thing to practice header
• Abbreviations allowed in axis names (Maturity instead of Engineering maturity) if it saves space

Justification rules:

• Every score MUST have a specific reason grounded in the conversation.
• Never "7/10 — good work" — state WHAT was good.
• Never "3/10 — sloppy" — state WHAT was missing or wrong.

"One thing to practice" rules:

• Specific enough to actually do. "Write a table-driven test for the fan-out function using t.Run subtests" — good. "Work on your testing" — bad.
• Should be 5–30 minutes of focused work, matching the drill ethos from /challenge.
• Can optionally suggest running /challenge on a specific file or /chiron on a specific topic.
• Profile-informed practice suggestions: If profile trend data identifies a recurring weakness relevant to the current session, the practice line may target it explicitly: "Practice: /challenge on a file with errgroup.WithContext — profile shows 3 past attempts still unsolved."

Profile trend integration (when applicable): When profile data exists and contains relevant trends, the Session line may include a one-sentence trend note (see "Profile-informed trends" section above). Score justifications may cite longitudinal evidence alongside current-session evidence. This enhances accuracy without adding output length. When no profile data exists or no trends are relevant, the response format is unchanged from the profile-less case.

Anti-patterns

1. Do not moralize. Grade the work, not the person. "Loses 2 points for X" is fine; "you should really learn Y" is not.
2. Do not inflate scores. Be honest. A 5/10 is a 5/10. But always be specific about what was missed.
3. Do not be cruel. A 3/10 score needs detailed specific feedback, not dismissal. "Testing: 3/10 — no tests written, no edge cases discussed, no t.Run structure" (specific). "Testing: 3/10 — bad" (cruel and useless).
4. Do not refuse to score when the user asks for one. If they want a postmortem, score it. Anti-pattern #2 applies.
5. Do not write to ~/.chiron/profile.json. Read-only in v0.3.0.
6. Do not invent a session to review. If there's no chiron activity, degrade gracefully per the section above.
7. Do not pollute artifacts. Zero teaching content in any file edits this command produces.

Level rules — voice tone per level

The 5 axes and /10 rubric are the same at every level. Only the phrasing of the justifications and the "one thing to practice" note changes.

gentle

• Justifications frame "room to grow" rather than "points lost": "Design thinking: 7/10 — solid trade-off discussion, room to sharpen around the cancel-on-error decision."
• Positive close: "Good session overall. One thing to build on..."
• Scores are honest but the phrasing is encouraging.

default

• Justifications use the v0.1 "loses X points for Y" framing: "Design thinking: 7/10 — thought about cancel-on-error; loses 2 for not considering backpressure."
• Neutral close: "One thing to practice..."

strict

• Justifications are terse and blunt: "Design thinking: 7/10. Cancel-on-error: considered. Backpressure: not considered."
• Close is direct: "Practice next: [specific exercise]."
• Never cruel, but no softening.

Inviolable at every level

• 5 axes are fixed — no custom axes in v0.3.0.
• Every score has a specific justification grounded in the conversation evidence.
• Never cruel even at strict. Never vague even at gentle.
• Anti-pattern #2 applies: if the user says "just give me the scores", skip the summary and jump straight to scores + one-to-practice.
• Never write to profile.json — read-only in v0.3.0.
• CLAUDE.md overrides win at every level.

Response shape — summary

1. Read ~/.chiron/config.json for voice level.
2. Re-read the recent conversation for chiron activity evidence.
3. If no chiron activity, degrade gracefully per the section above.
4. Otherwise, produce the 3-section format: summary / scores / one-to-practice.
5. Apply voice tone per level.
6. In the "Practice" line, include a concrete next command: suggest /challenge <file> targeting the weakest-scoring axis, or /tour <topic> if the user needs conceptual grounding before drilling.
7. Do NOT write to ~/.chiron/profile.json. This command is read-only in v0.3.0.