/reflect — session reflection

Produce a single-file interactive HTML dashboard analyzing a Claude Code session for places the agent took a wrong turn, paired with concrete repo additions that would prevent the same wrong turn next time.

When to use

• User invokes /reflect (no args) → analyze the current session from the in-context conversation. Do not re-read the session JSONL — work from the agent's own memory of what happened.
• User invokes /reflect <session-id> or /reflect <path-to-jsonl> → analyze a different session. Run scripts/analyze_session.py <path> directly (this skill runs in a forked context — context: fork — so the compressed transcript can enter your context safely). The script outputs a compact markdown transcript: system reminders stripped, tool calls/results collapsed to one-liners, compaction blocks expanded with embedded user quotes. Then you analyze that transcript to identify wrong-turn moments — the script does NOT classify incidents, it only compresses.

Session JSONLs live under: ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl

What to detect (incidents)

Open-ended — but the common shapes are:

• correction — user pushed back on an approach ("no", "don't", "stop", "actually")
• retry — agent ran the same tool 2+ times with variations before it worked
• waste — many Reads/Greps/Globs before finding the right file
• reversal — agent edited then unwound (Edit → revert, Write → delete)
• dead-end — tool failed because of the environment (missing binary, wrong path, OS mismatch)
• self-correction — agent caught its own mistake mid-stream

Severity:

• high — explicit user correction, repeated correction, or substantial wasted turns
• med — single retry/reversal, modest waste
• low — minor dead-end, easily recovered

Confidence (0–100): how sure you are this is a real wrong turn worth surfacing, vs. signal noise. Be honest — low-confidence items are not bugs, they let the user filter out speculation.

• 90–100 — explicit user correction or unambiguous failure
• 70–89 — strong pattern (repeated retries, clear reversal) with minor interpretation
• 50–69 — plausible wrong turn, could also be normal exploration
• <50 — speculative; surface only if the pattern is interesting

Render as data-conf="<n>" on each .incident plus a small badge. The dashboard slider hides incidents below the chosen threshold.

What to recommend

Open-ended. Anything that would help a future agent reach the goal faster. Examples (not a fixed taxonomy — pick what fits the incident):

• rule → CLAUDE.md block (project or user)
• doc → docs/ARCHITECTURE.md, README pointer, architecture map
• script → scripts/<name>.sh wrapping a known-good command
• hook → .claude/settings.json PreToolUse / PostToolUse for hard blocks (e.g. block edits to a generated file)
• memory → user/feedback/project/reference entry under ~/.claude/.../memory/
• skill → a new sub-skill in ~/.claude/skills/
• agent → an agent definition in .claude/agents/
• allowlist → permissions in .claude/settings.json
• env → environment variable, MCP server, etc.

Each recommendation should map to one or more incidents (the dashboard renders these as addresses #N links).

How to render

Synthesize a fresh single-file HTML dashboard each run, using reference/example.html as inspiration (not a template — override anything that doesn't fit the session). The reference file establishes:

• Two-pane layout: incidents left, recommendations right
• CSS tokens (--bg, --panel, --fg, --muted, --accent, etc.) — keep this scheme
• Type pairing: Newsreader italic for the wordmark, JetBrains Mono for everything technical, system sans for prose
• Click incident card to expand; click addresses #N link to scroll-and-highlight the matching incident
• Filter chips toggle severity and category

Output directory — do not write inside the skill folder. Resolve a temp dir from the environment, in this preference order:

1. $TMPDIR (Unix/macOS)
2. $TMP or $TEMP (Windows / Git Bash)
3. /tmp as fallback

Then create a reflect/ subdir inside it (mkdir -p) and write the report as <that-dir>/reflect/<slug>.html. Slug rules — kebab-case, derived from session goal:

• 2–5 words, lowercase, hyphen-separated, ASCII only
• describe the task, not the session id (e.g. auth-middleware-rewrite)
• if the goal is unclear, fall back to <YYYY-MM-DD>-<topic>.html

Open it in the browser when done: start "" <path> (Git Bash on Windows).

Files in this skill

• reference/example.html — canonical inspiration HTML. Read before generating.
• scripts/analyze_session.py — JSONL compressor. Strips system reminders, preserves real user messages, collapses tool calls/results to one-liners, expands compaction blocks. Output is markdown to stdout. Use only for explicit sessions — never on the current in-context session.

Workflow

Default (no args) — reflect on current session:

1. From conversation memory, list wrong-turn moments (incidents) with severity + category + turn marker + a 1-line excerpt.
2. For each incident (or cluster), draft a recommendation: title, target (kind + path), snippet, list of incident IDs it addresses.
3. Read reference/example.html for the current aesthetic.
4. Write the dashboard to the temp dir resolved above (<temp>/reflect/<slug>.html) — fresh HTML, same look-and-feel as the reference, populated with the real incidents/recs.
5. Open it with start "" <path>.

Explicit session — /reflect <id-or-path>:

1. Resolve to a JSONL path (if just an id, look under ~/.claude/projects/<encoded-cwd>/<id>.jsonl).
2. Run scripts/analyze_session.py <path>. Skill is forked (context: fork), so the compressed transcript (tens of KB even for multi-MB JSONLs) is safe in context. No subagent.
3. Identify wrong-turn incidents from the transcript yourself — the script does not classify. Focus on user/assistant exchange (what the user wanted vs. what the agent did), not raw tool patterns.
4. Fallback to JSONL when transcript is lossy. Transcript truncates tool args/results and long messages. When detail matters (exact rejected input, full error body, Edit diff, full user message), Read the JSONL directly with offset/limit scoped to the event by turn/tool. JSONL is source of truth; transcript is the index.
5. Draft recommendations and render (default steps 2–5).

Notes

• Incidents are claims about what happened. Be honest — include the agent's own mistakes, not just user corrections.
• Recommendations should be specific and actionable (real paths, real snippet content). A vague "improve docs" rec is not useful.