Session Curation

Session analysis & memory curation

This is the session_curation dimension of the machine-equality matrix (docs/reports/machine-equality-matrix.html, GROUP heading "Session analysis & memory curation — daily freshness"). Unlike the other matrix dimensions — which grade a machine against its peers (DIVERGES / NO-MAJORITY) or against a baseline (BELOW-BASELINE) — this cell grades a machine against real wall-clock now: it answers "did this box analyze its session activity and curate its memory recently enough?"

Each box, every 6h, does two things and stamps the time it finished:

1. Analyze all AI-provider session logs present on THIS box — Claude (~/.claude/projects), Codex (~/.codex/sessions), Gemini (~/.gemini/tmp), Hermes (~/.hermes/sessions) — into per-provider activity counts. Bounded, mtime-first, never reads gigabytes (Hermes' ~27 GB tree is capped at MAX_FILES_PER_PROVIDER, never fully walked).
2. Curate the workspace memory delta — basenames of .claude/memory/*.md changed since the last run — into a portable digest, and publish a fingerprint to a dedicated git ref so sibling machines can pull it.

Counts / enums / timestamps only — never prompt text, tokens, absolute paths, or client identifiers (matches the collect-equality allowlist and repo legal rules).

Daily-freshness semantics

freshness_verdict(report, now=None) in scripts/readiness/build-equality-matrix.py (line 198) grades the cell by the age of last_curated_at:

Only CURATED-FRESH is in OK_VERDICTS. The verdict fails closed: a missing, unparseable, or future-dated timestamp (clock skew) is graded MISSING-EVIDENCE, never "fresh". The 6h cron cadence keeps a healthy box comfortably under the 12h green threshold (two refreshes per window).

Why the matrix MUST rebuild daily — the dead-man's-switch

The freshness cell only ages if something re-renders it. The collector (collect-equality.sh) only references the state file — it never re-runs the engine. So if the session-curation cron silently dies, the last curation run left a green cell in the last rendered matrix, and nothing would ever recolour it: a dead cron would look healthy forever.

The equality-matrix-refresh cron (control-plane, daily 50 5 * * *) closes that hole. It rebuilds the matrix independent of the curation cron, re-grading every cell against a fresh now. With it, a dead curation cron's cell visibly ages green → orange (>12h) → red (>24h) on its own. Per schedule-tasks.yaml: "so a DEAD curation cron still ages its freshness cell past green … Without this daily rebuild the weekly equality-report (#2801) would freeze the last-green render and the dead-man's-switch would never fire."

Pieces (all repo-root relative)

Cron (config/scheduled-tasks/schedule-tasks.yaml)

• session-curation — 47 */6 * * *, every box. Runs the wrapper; refreshes this box's equality column each cycle.
• equality-matrix-refresh — 50 5 * * *, control-plane only. The daily dead-man's-switch rebuild described above.

Cross-machine transfer

The engine publishes this box's fingerprint to the dedicated git ref session-curation-state via scripts/monitoring/equivalence_state.py (the same CAS-backed equivalence-state store the harness uses). Sibling machines pull every box's fingerprint and render a merged fleet view (session-curation-fleet.md) with --collect. The transfer carries the same leak-safe payload as the state file — counts, active-provider enum, last-curated stamp, memory-delta count — nothing else.

Operator commands

# Inspect this box's state WITHOUT writing or publishing (dry run)
uv run --script scripts/curation/curate_session_memory.py --stdout

# Curate this box now + publish the fingerprint (what the cron runs)
uv run --script scripts/curation/curate_session_memory.py
# …or the full wrapper (curate → collect → rebuild matrix, fail-loud):
bash scripts/curation/curate-session-memory.sh

# Curate locally but skip the git-ref publish
uv run --script scripts/curation/curate_session_memory.py --no-publish

# Fleet view — merge every box's fingerprint from the session-curation-state ref
uv run --script scripts/curation/curate_session_memory.py --collect

# Re-render the matrix to re-grade freshness against a fresh `now`
bash scripts/readiness/collect-equality.sh
uv run --no-project --with pyyaml python scripts/readiness/build-equality-matrix.py

(uv unavailable? the wrapper falls back to python3; the engine's only dependency is pyyaml via the PEP-723 header.)

Remediating a non-green cell

remediate() (build-equality-matrix.py line 446) maps the verdicts:

A cell that stays orange/red after a successful manual curation means the render isn't refreshing — check the equality-matrix-refresh cron, not the engine.

Related

• ecosystem-equivalence-reconcile — drives THIS box back to full matrix equivalence; its verdict→action playbook is the parent of this cell's remediation.
• repo-ecosystem-hygiene — read-only detector for the repo-hygiene surface the reconcile skill complements.