/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:

1. Poisoning — early hallucinated content gets quoted by later turns, compounding the error.
2. Distraction — irrelevant earlier context dilutes the model's attention to current task.
3. Confusion — contradicted facts pile up; the model doesn't know which to trust.
4. 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