Distill the current conversation into a structured note (decisions made, open questions, file pointers with line numbers, next 1–3 actions) and save to `quality_reports/session_logs/` before auto-compression. Differs from `/checkpoint` (explicit stop-point snapshot) and from auto-compaction (which truncates rather than distills). Use when context is approaching auto-compact threshold, when a long pipeline has accumulated many decisions, or when the user says "compress", "distil this session", "before we hit auto-compact", "structured handoff before context resets".
/compress-session — distil, don't truncateAuto-compaction is lossy: it keeps recent turns and drops earlier ones, with no preservation of what was decided mid-session. /compress-session is the distil-not-truncate alternative — produce a structured note that the next session can resume from in under a minute.
Drew Breunig's "How Long Contexts Fail" identifies four failure modes for long-context sessions:
The template's 200-line MEMORY.md cap defends against distraction. The plan-on-disk convention defends against clash. Nothing currently defends against poisoning or directly against confusion — that's what this skill is for.
/checkpoint/checkpoint | /compress-session | |
|---|---|---|
| When | Explicit stop-point (end of working session, before model switch, before collaborator handoff) | Forced — context is about to auto-compact, or the conversation has accumulated enough noise that distillation pays for itself |
| What's preserved | Active plan, decisions, file pointers, next 1–3 actions | Same, plus an explicit "discarded as noise" line so the next session knows what was intentionally not kept |
| Output location | quality_reports/checkpoints/YYYY-MM-DD_<slug>.md | quality_reports/session_logs/YYYY-MM-DD_compression_<slug>.md |
| Triggering | User-invoked at a natural pause | User-invoked when context fatigue shows, OR prompted by the optional PreCompact reminder hook below (not wired by default) |
| Memory updates | Optional auto-proposal of [LEARN] entries | Proposes 0–3 [LEARN] entries — distillation is when generalizable lessons surface, and a quiet session proposes none |
Both skills are companions to the narrative session-log workflow at quality_reports/session_logs/. None replaces the others.
/context-status reports approaching threshold./compress-session; the skill never runs automatically./checkpoint./clear. /compress-session is for preserving state, not for resetting.The current conversation is the source. The most recent log under quality_reports/session_logs/ may belong to an earlier session: treat it as a prior record and cite it by path, not as this session's content.
Optionally use $ARGUMENTS as a topic slug for the output filename.
Write only what this session established. The next session is handed this file automatically (session-handoff.py), so anything invented here arrives there as fact.
path:line only for lines you read in this session; otherwise give the path alone.[LEARN] proposals.[Session handoff: …] or [Context Restored After Compaction] block is the previous record. Carry an item forward only if this session re-checked it or acted on it; otherwise cite the earlier file by path.Produce a note with these sections:
# Session Compression — <YYYY-MM-DD> <topic-slug>
**Source:** <session-log file or "current conversation">
**Token budget at compression:** <approximate %>
**Why compress now:** <approaching auto-compact | mid-plan handoff | accumulated noise | user-requested>
## Active state
- **Plan:** [link to active plan in `quality_reports/plans/` or "no active plan"]
- **Branch:** <git branch>
- **Last commit:** <SHA + subject>
- **Working tree:** <clean | N modified files>
## Decisions made (this session)
1. [Decision]. **Why:** [one sentence]. **Where recorded:** [file:line, plan, or memory].
2. ...
## Files touched
- [path:line] — [what changed, one sentence]
- ...
## Open questions
- [Question]. **Blocker?** [Yes/No]. **Where to resume:** [pointer].
- ...
## Next actions (1–3 only)
1. [Specific next step with file pointer]
2. ...
## Discarded as noise
Things explored during this session that should NOT carry forward — failed hypotheses, abandoned approaches, debugging dead-ends. Listing them explicitly defends against [Breunig's "poisoning" failure mode](https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html) (hallucinated / wrong content from early turns getting quoted later).
- [Hypothesis / approach / dead-end] — [why it didn't work]
- ...
## Proposed `[LEARN]` entries
Lessons worth promoting to MEMORY.md (machine-specific ones stay in native auto memory). User reviews; nothing auto-merged.
- `[LEARN:<category>]` <lesson>. **Evidence:** <pointer>.
- ...Write to quality_reports/session_logs/YYYY-MM-DD_compression_<topic>.md (slug derived from $ARGUMENTS or current plan name).
Report to the user:
The user reviews. Nothing auto-merges into MEMORY.md: a proposal the user approves is appended to MEMORY.md on their say-so (as in /checkpoint Phase 3), and a machine-specific one is left to native auto memory. /promote-memory does not read this file; its candidates come only from native auto memory (~/.claude/projects/<project>/memory/).
Forkers who want a real chance to run /compress-session before compaction can add a PreCompact hook that blocks compaction once per session. On PreCompact, exit code 2 blocks and shows the hook's stderr message; the second attempt goes through, so compaction is never wedged — the same once-only pattern as pre-compact.py's opt-in DRAFT block. (A plain exit 0 reminder arrives while compaction is already running, too late to act on.)
{
"hooks": {
"PreCompact": [
{
"hooks": [
{
"type": "command",
"command": "sid=$(python3 -c \"import sys,json; print(json.load(sys.stdin).get(\\\"session_id\\\",\\\"x\\\"))\" 2>/dev/null); f=\"${TMPDIR:-/tmp}/cc-compress-reminded-$sid\"; [ -e \"$f\" ] && exit 0; touch \"$f\"; echo \"Run /compress-session now, then compact again - this reminder blocks only once per session.\" >&2; exit 2",
"timeout": 5
}
]
}
]
}
}The hook only pauses and reminds; the user runs /compress-session, then compacts again. We deliberately don't make the hook auto-invoke the skill — that bypasses the user's review step.
/compress-session as a routine replacement for /checkpoint. Checkpoints are cheap and frequent; compressions are heavier and reserved for forced compression.quality_reports/session_logs/YYYY-MM-DD_compression_<topic>.md (gitignored — session-state, not version-controlled).[LEARN] proposals.pedrohcgs/claude-code-my-workflow
fe70774
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.