/compact-summary — view a session's compaction summary

A Claude Code skill that surfaces the compaction summary of the current session — the text Claude Code generates when it compresses a long conversation, whether you ran /compact manually or auto-compact fired on its own.

The catch it solves: that summary is stored as a single line in the session transcript and Claude Code does not let you re-read it in the TUI (compaction prints only a Conversation compacted line). Asking the assistant to reprint it costs ~3-6k output tokens every time — a summary measures ~10-20KB. So this skill extracts the summary to a file and returns only the file path plus a one-line meta; the body never re-enters the chat.

Quick start

/compact-summary

Output is two lines — a path and a meta line:

C:\Users\you\.claude\compact-summaries\236219d3-...-2.md
latest of 2 | 14,821 chars | ~4,234 tokens | excludes messages after this compaction

Open that file to read the summary. Nothing is printed into the conversation.

What it shows (and what it does not)

The summary is the compressed stand-in for the conversation before the compaction point. It is not the session's full current context: the live context also holds the system prompt, your CLAUDE.md/memory, and every turn after the compaction — none of which are in this file. Hence the meta's excludes messages after this compaction.

If a long session compacted more than once, each compaction is its own summary line; this skill writes the latest one and reports latest of N so you know earlier ones exist.

How it works

[invocation]  /compact-summary
   └─ Claude runs scripts/compact_summary.py and relays its stdout verbatim

[script]
   session id = $CLAUDE_CODE_SESSION_ID
   transcript = the one  ~/.claude/projects/*/<id>.jsonl  matching that id   (else: fail closed)
   summaries  = every transcript line where the PARSED field isCompactSummary is True
   none?  -> print "not compacted yet", write nothing, exit 0
   else   -> write the LAST summary to ~/.claude/compact-summaries/<id>-<N>.md
             print: path + "latest of N | chars | ~tokens | excludes post-compaction messages"

Design choices that matter:

• Discovery by session id, not mtime. The live id is in $CLAUDE_CODE_SESSION_ID; globbing ~/.claude/projects/*/<id>.jsonl by filename sidesteps having to reproduce Claude Code's cwd→project-dir escaping. An mtime "newest file" fallback is deliberately omitted — it could silently pick a different project's conversation. If the id is unset or the file isn't uniquely found, the script errors instead of guessing.
• Parsed field, never a substring. Lines are matched by json.loads(...)["isCompactSummary"] is True. A raw substring scan would be fooled by an assistant message quoting the field name and would miss a "isCompactSummary": true written with a space after the colon.
• Tolerates a live transcript. The current session's file is still being appended to, so a truncated final line is expected; the parser skips unparseable lines rather than aborting.
• Meta is script-computed. Char/token counts come from the script, so the assistant never needs to open the file to describe it — preserving the token saving end to end.

Files

compact-summary/
├── README.md    this file
├── LICENSE      MIT
├── SKILL.md     skill body Claude follows at runtime (run the script, relay stdout, never open the file)
└── scripts/
    └── compact_summary.py   locate transcript, extract latest summary, write file, print path + meta

Output lives outside the skill, at ~/.claude/compact-summaries/<session-id>-<N>.md. Files are overwritten in place per (session, compaction number), so re-running never accumulates copies. They persist until you delete them (the original transcript keeps its own copy regardless).

Install

Copy this directory to ~/.claude/skills/compact-summary/:

git clone <this-repo> ~/.claude/skills/compact-summary

The directory name compact-summary is significant — it must match the slash command (/compact-summary). No hook and no settings change is needed; the skill is invoked on demand.

Verify

python ~/.claude/skills/compact-summary/scripts/compact_summary.py --selftest

Expect selftest ok (checks parsed-field extraction, space-after-colon tolerance, list-content normalization, partial-line skipping, fail-closed transcript lookup, and utf-8 round-trip). For a live check, run /compact-summary in a session that has compacted and open the printed path.

Limits / out-of-scope

• Current session only — targets $CLAUDE_CODE_SESSION_ID. Viewing an arbitrary past session, or a summary from before a /clear, is out of scope.
• The summary, not the live context — for a real-time breakdown of what the context currently holds, use Claude Code's built-in /context.
• Latest by default — selecting an earlier compaction by index is not implemented (the meta reports how many exist); add a --index N flag if that becomes necessary.

Dependencies

• Python 3.9+ (standard library only — no third-party packages)
• A Claude Code build that sets CLAUDE_CODE_SESSION_ID and stores per-session JSONL transcripts under ~/.claude/projects/

License

MIT — see LICENSE.