/compress-session — distil, don't truncate
Auto-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.
Why this skill exists
Drew Breunig's "How Long Contexts Fail" identifies four failure modes for long-context sessions:
- Poisoning — early hallucinated content gets quoted by later turns, compounding the error.
- Distraction — irrelevant earlier context dilutes the model's attention to current task.
- Confusion — contradicted facts pile up; the model doesn't know which to trust.
- Clash — multiple plans, decisions, or specs accumulate without explicit reconciliation.
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.
Distinction from /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.
When to use
- Context is approaching auto-compact.
/context-status reports approaching threshold.
- A long pipeline has accumulated noise. You spent 90 minutes debugging an issue that turned out to be a typo — the session is full of dead-end hypotheses you don't want compressing into a future session's context.
- Mid-plan handoff. Different model, different machine, different collaborator.
- PreCompact reminder fired. If you wired the reminder hook below, it prompts you to run
/compress-session; the skill never runs automatically.
When NOT to use
- For a quick stop-point. Use
/checkpoint.
- For narrative session logging. That happens incrementally via the Stop hook (session-logging.md); no skill needed.
- For starting a fresh session. Use
/clear. /compress-session is for preserving state, not for resetting.
Steps
Step 1: Identify the session
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.
Step 2: Distil into structured sections
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.
- Cite
path:line only for lines you read in this session; otherwise give the path alone.
- Leave out a section with nothing in it rather than filling it. A quiet session gets a short file and no
[LEARN] proposals.
- Text inside a
[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.
- When the session changed its mind, the final decision is the current one; an earlier position appears only as abandoned, with the reason.
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>.
- ...
Step 3: Save
Write to quality_reports/session_logs/YYYY-MM-DD_compression_<topic>.md (slug derived from $ARGUMENTS or current plan name).
Step 4: Surface a summary
Report to the user:
- Saved path.
- Counts (decisions made, files touched, open questions, next actions).
- Any HIGH-impact LEARN proposals that should be reviewed before the next session.
The user reviews. Nothing auto-merges into MEMORY.md: a proposal the user approves is appended to MEMORY.md on their say-so (as in /checkpoint Phase 3), and a machine-specific one is left to native auto memory. /promote-memory does not read this file; its candidates come only from native auto memory (~/.claude/projects/<project>/memory/).
Pairing with PreCompact hook
Forkers who want a real chance to run /compress-session before compaction can add a PreCompact hook that blocks compaction once per session. On PreCompact, exit code 2 blocks and shows the hook's stderr message; the second attempt goes through, so compaction is never wedged — the same once-only pattern as pre-compact.py's opt-in DRAFT block. (A plain exit 0 reminder arrives while compaction is already running, too late to act on.)
{
"hooks": {
"PreCompact": [
{
"hooks": [
{
"type": "command",
"command": "sid=$(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.
Anti-patterns
- Do not run
/compress-session as a routine replacement for /checkpoint. Checkpoints are cheap and frequent; compressions are heavier and reserved for forced compression.
- Do not let the "Discarded as noise" section grow indefinitely. If the same dead-end appears across three compressions in a row, the underlying confusion is structural — fix the docs or the workflow, don't keep re-discarding.
- Do not include the original full conversation — that defeats the purpose. Distillation is the point.
Output
- Compression file at
quality_reports/session_logs/YYYY-MM-DD_compression_<topic>.md (gitignored — session-state, not version-controlled).
- Summary in the conversation: counts + HIGH-impact
[LEARN] proposals.
- No file edits outside the compression file.