afterglow

afterglow — Retrospective Pattern Surfacer

aftercare and afterglow are two entry points to the same goal — capturing reusable learning from the project — and differ only in discovery:

• aftercare: in-session capture, using the rich current conversation context.
• afterglow: retrospective scan across past sessions, surfacing patterns that repeated across runs.

afterglow finds candidates and presents them with evidence. It then asks whether to hand the selected ones to aftercare, which owns the judgment (skill / memory-file / drop) and the drafting. afterglow itself never drafts and never writes files.

When to use

• The user explicitly invokes /afterglow. Never auto-trigger.
• For in-session capture of a just-fixed mistake, use aftercare instead.

Orchestration

1. Compute paths

1. Get the current working directory: pwd.
2. Encode it as a project path: replace every / with -. Example: /Users/x/y → -Users-x-y.
3. Resolve the Claude config base: ${CLAUDE_CONFIG_DIR:-$HOME/.claude}.
4. SESSIONS_DIR = ${BASE}/projects/${ENCODED}.
5. CACHE_DIR = ${BASE}/afterglow/${ENCODED}. Create it if absent: mkdir -p.

If SESSIONS_DIR does not exist, print "No prior sessions found for this project." and exit.

2. List recent substantive sessions

Call the bundled helper script in a single Bash invocation (path is relative to this skill's root):

bash scripts/list-substantive-sessions.sh "${SESSIONS_DIR}"

The script sorts by mtime descending, drops sessions with user_message_count < 2 OR size < 4096 bytes, and caps the output at the most recent 50 by default. Override with --limit N (or --limit 0 for unlimited) only when the user explicitly asks for a wider scan.

Do NOT inline this as a shell loop (cd ~/.claude/projects/<encoded> && for f in *.jsonl ... or any for/while over the sessions directory with grep/wc substitutions). Auto-mode permission classifier blocks those patterns. Always go through the helper.

If the output is empty, print "No substantive sessions found for this project." and exit.

3. Ensure compressed transcripts are cached

For each session path, with session_id = <filename without .jsonl>:

1. Target: ${CACHE_DIR}/<session_id>.txt.
2. If the target exists AND its mtime is newer than the source .jsonl, mark as cached — skip compression.
3. Otherwise run bash scripts/compress-session.sh <jsonl_path> ${CACHE_DIR}/<session_id>.txt. The script writes [User]: ... / [Assistant]: ... / [Tool: NAME] lines (typical reduction 40–50×).

Print a one-line summary: "Compressed N new (M cached)."

4. Dispatch the analyst

Dispatch a single subagent (one Agent tool use):

• Prompt: body of analyst-prompt.md.
• Input: a list of (session_id, transcript_path) pairs covering the substantive set. The subagent will Read each .txt itself — the main context never reads transcript content.
• Expected output: YAML matching the analyst schema (sessions_scanned, candidates, also_seen).

Parse the YAML. On parse failure, re-dispatch once with a structure-correction note. If it still fails, print the raw output and ask the user how to proceed.

5. Present candidates

• candidates: [] → "No repeated patterns found across N sessions." Print also_seen topics as a one-liner if any (for transparency). Exit.

• Otherwise, for each candidate, render the kind field and its suggested bucket so the user (and downstream aftercare) sees the analyst's tag:

  N. [×<count>] <topic>
     Kind: <kind>  →  suggested bucket: <skill | CLAUDE.md>
     Evidence:
       • "<verbatim quote>"  (session <id[:8]>)
       • "<verbatim quote>"  (session <id[:8]>)
     Pattern: <detail>
  

  Bucket-hint mapping:

  • repeated_task_kickoff → skill
  • multi_step_procedure → skill
  • convention_correction → CLAUDE.md

  The hint is a suggestion, not a decision. aftercare's 4-criteria judge still owns the final routing (skill / memory-file / drop).

• After the list, if also_seen is non-empty, append one line: "Also seen once each: <topic>, <topic>, ..." (truncate to ~5).

6. Offer hand-off to aftercare

Ask the user:

"캡처할 패턴이 있나요? aftercare를 트리거하면 메모리/스킬로 만들 수 있습니다. 선택: (숫자 / all / no)"

Branches:

• no or cancel → "OK. /aftercare 직접 호출하셔도 됩니다." Exit cleanly. No files written.
• 숫자 / all → Echo the selection back in plain text so it becomes part of the conversation context: "선택된 패턴: 1, 3 — aftercare로 넘깁니다." Then invoke Skill(aftercare) so its stage-1 judge picks up the presented candidates from the conversation. afterglow's job ends here; the rest is aftercare's flow.
• Skill(aftercare) unavailable or fails → Fall back to printing "/aftercare 를 다음 메시지로 호출해주세요." while keeping the candidate list on screen. Same destination either way.

Errors / Edge Cases

Situation	Handling
SESSIONS_DIR missing	Step 1: "No prior sessions found for this project." Exit.
Substantive set empty	Step 2: "No substantive sessions found for this project." Exit.
Single compression failure	Log a one-line warning, continue with the remaining sessions.
Analyst YAML parse failure	Re-dispatch once with a correction note. If still bad: show raw output, ask user.
Zero candidates	Step 5: clean exit, with also_seen summary if any.
User rejects hand-off	Step 6: clean exit, no writes.
Skill(aftercare) chaining unsupported	Step 6: fall back to "type /aftercare next".

Constraints

• User-explicit invocation only. Never auto-trigger.
• No drafting in afterglow. It does not write skills, memory entries, or any files under .claude/skills/, AGENTS.md, CLAUDE.md, or ~/.claude/MEMORY.md. That work belongs to aftercare.
• No bucket decision in afterglow. aftercare's 4-criteria judge owns skill / memory-file / drop.
• Main context never reads transcripts. Only metadata (mtime/size/grep counts via the helper script) and the analyst's structured YAML output enter the main context.
• Compressed .txt cache is append-only; do not delete on failure. Stale .yaml files from previous afterglow versions in the cache directory are harmless — leave them.