Rsi

RSI — session close

Thin Tier 1 close for the goal-gated learning loop. Loads this session's close digest, verifies one principal claim, captures one structural lesson, acks.

Do not run mid-goal unless the user explicitly invoked /rsi close. SessionEnd already skipped Tier 1 when goal is still active.

When to invoke

• User says /rsi close or asks for session retrospective with evidence
• SessionStart nudge: "Prior session … has an RSI close digest"
• Goal was achieved this session and digest exists

Workflow

1. Load digest

uv run --project ~/Projects/agent-infra python3 \
  ~/Projects/agent-infra/scripts/reflect_session_close.py \
  --latest-digest "$(cat .claude/current-session-id 2>/dev/null)"

Selects this session's last reflect.close-digest.v1 row — never a close-ack.v1 (acks share the file and dominate the tail; blind tail -1 returns an ack). An empty session arg falls back to the latest un-acked digest — the SessionStart-nudge case, where the digest belongs to a prior session.

If exit 1 (no digest): drain the queue, then retry:

uv run --project ~/Projects/agent-infra python3 \
  ~/Projects/agent-infra/scripts/reflect_session_close.py --drain

The drain is fail-loud: it prints N intents read, M digests written, K skipped (…) and exits nonzero on a silent zero. Still no digest after a clean drain → this session had no Tier-1-eligible close; stop here.

2. Verify one load-bearing claim

First read the digest's real_issue_kinds + verify_hint — they name WHY this close fired and point the verify at the right claim:

• unsupported_completion → fabrication risk: re-run the claimed-successful command, confirm the outcome.
• user_rescued_failure → verify the fix actually landed (test exit / gate / hash), not the recovery narration.
• goal_achieved / operator_flag → verify the achievement or what the operator flagged.

Pick ONE claim from the session episode (not the /goal Haiku evaluator — that is a proxy):

• Pipeline: just sample-state, receipt hash, _STATUS.json mtime
• Code: test exit code, just validate leaf gate
• Artifact: file hash, gate_truth output

Record the command + output snippet. If verification fails, report failure — do not attach evidence.

3. Close: structural lesson + ack

The close centers on the Step-2 verify plus ONE durable capture — write the structural lesson where the next session finds it: a project memory file (+ MEMORY.md pointer), OR at most one [obs] line in the project's improvement log / MAINTAIN.md. One destination, not several.

RSI-hindsight CLOSE (observe 2026-07-10): if the human said "why didn't you find…" / RSI-meta this session, do not chat-apologize. Cite the file artifact:

just -f ~/Projects/agent-infra/justfile rsi-hindsight-enqueue
# then Read artifacts/rsi-hindsight/LATEST.md — quote the matching queue row id

Convert the top matching row into a maintain draft or detector proposal (steward / maintain-candidates.json), then leave the row queued for /observe maintain. Do not expand REDISCOVERY regexes while the over_caution ablation window is open.

arc-agi MEMORY discipline (AP10): never paste session blocks into MEMORY.md. Write the fact to project_*.md Revisions (or research/ / decisions/), then add/update one MEMORY.md pointer line. Soft cap 8KB — see .claude/rules/memory-write-discipline.md. Run just memory-lint if you touched memory.

If digest has operator_dx_reflex (non-empty) or operator_dx in real_issue_kinds: fill the four fields for the top intervention (max one action):

{
  "operator_added_value": "what the operator contributed",
  "miss_class": "discovery|tool_reuse|stopping|dx|queue|eval|other",
  "would_have_prevented": "existing tool|new hook proposal|skill update|backlog row|noop",
  "action_taken": "path written OR explicit noop reason"
}

Rules (steward 2026-07-09-rsi-dx-reflex.md):

• Obvious no-downside local fix → do it, put path in action_taken.
• Global/hook-level → write one steward proposal, cite it in action_taken.
• No strictly-better change → would_have_prevented: noop + reason in action_taken.
• Do NOT dump RSI prose into every answer — only at close.

fm.py attach-evidence is optional — fm-evidence.jsonl is auto-fed by the capture path (~1.7K rows); attach manually ONLY when the verify surfaced novel evidence no automated row carries:

cd ~/Projects/agent-infra && uv run python3 scripts/fm.py attach-evidence <FM_ID> \
  --session <SESSION_ID> --quote "<verified fact + command output>"

Then ack the digest (stops SessionStart nudge). arc-agi + operator_dx: append Mode-3 grade (only when you have a derivability judgment — skip if NON-DERIVABLE taste call):

uv run --project ~/Projects/agent-infra python3 \
  ~/Projects/agent-infra/scripts/reflect_session_close.py --ack <SESSION_ID> \
  --hindsight '{"item":"operator:<slug>","grade":"DERIVABLE","evidence":"<what operator added>","gap":"<would_have_prevented>"}'

Grades: DERIVABLE | NON-DERIVABLE (taste/telos — logged, no fault). No auto-append on ack without --hindsight — avoids empty stub cruft.

uv run --project ~/Projects/agent-infra python3 \
  ~/Projects/agent-infra/scripts/reflect_session_close.py --ack <SESSION_ID>

4. Steward proposal (optional, max one)

If a concrete, reversible guard would prevent recurrence:

# Write to ~/.claude/steward-proposals/YYYY-MM-DD-<slug>.md
# /observe maintain reads these on next tick

One proposal max. Prefer attach-evidence over new infrastructure.

Backpressure gate: skip the proposal entirely when just steward-reconcile reports > 40 open — the queue is frozen until drained (just questions-drain --dispatch). The cap is measured, not a style choice (2026-09-02 tabula rasa: 105 open proposals, 118 stale questions, zero human dispositions in 22 days).

Vetoes

• Never hook Stop for this workflow
• No LLM on SessionEnd (capture is already done)
• Do not trust goal-achieved without independent verification
• Max one [obs] OR one steward proposal per close

Related

• ADR: agent-infra/decisions/2026-06-15-rsi-session-close-gate.md
• Capture: agent-infra/scripts/reflect_capture.py
• Digest: agent-infra/scripts/reflect_session_close.py
• Deep retro: /observe retro (manual, heavier)