KeiSeiKit-1.0/_primitives/templates/deep-sleep-trigger-prompt.md

Deep-sleep NREM consolidation (KeiSeiKit v0.13.0)

Clone: {REPO_URL} Cadence: every {DEEP_SLEEP_CRON_DAYS} day(s), counted from first install Fork-output mode: WITH_FORK={WITH_FORK} (1 = plan + fork; 0 = plan only)

Cycle order with Phase C

Phase A (incubation, v0.12.0) → Phase B (REM consolidation, v0.12.0) → Phase C (deep-sleep NREM, THIS document, v0.13.0)

Phase C runs AFTER Phase B, and ONLY when today is a multiple of DEEP_SLEEP_CRON_DAYS from the install date (file sync-repo/reports/install-anchor.txt). If the file is missing on a first run, Phase C silently no-ops and writes the anchor for next time.

If Phase A selected a marathon: true task, Phase B is skipped per v0.12.0 rules AND Phase C is skipped too — the marathon owns the night.

Phase C — Task

1. Scan. Run kei-conflict-scan --path sync-repo/ --format=json --exit-on-hit and capture the JSON. Categories: rules, hooks, blocks, orphans, cp.

2. Plan. Pipe the JSON into kei-refactor-engine --input - --plan-out sync-repo/sleep-deep/YYYY-MM-DD-plan.md. The plan markdown always lists:

   • Auto-apply candidates (safe; engine-proposed)
   • "Requires human decision" items (zero-conflict guarantee: NEVER included in the generated patch)

3. Optional fork (only if WITH_FORK=1):

   a. kei-refactor-engine --input - --apply-to-branch deep-sleep/YYYY-MM-DD --patch-out sync-repo/sleep-deep/YYYY-MM-DD-autoresolve.md (re-run on same JSON; the auto-resolve markdown lists auto-apply items only — NOT a unified diff, see note below).

   b. Review the auto-resolve markdown and apply each change manually in a new local branch: git checkout -b deep-sleep/YYYY-MM-DD Open <autoresolve>.md, edit the listed files accordingly, then git add <files> && git commit.

   c. Gate: kei-graph-check --path sync-repo/. If broken refs after your edits → delete branch, append "graph check failed — fork aborted, plan kept" note to the plan file.

   d. If clean → push the fork branch for morning review.

NOTE (v0.14.1 retraction): earlier docs claimed the engine emits a git apply-ready patch. It does not — see patch.rs header for the reason (engine cannot synthesise file-content hunks without reading source files, which risks RULE 0.4 hallucination). The companion file is a markdown summary reviewed and applied by hand.

4. Commit + push. The plan markdown is always committed to main with message NREM: deep-sleep YYYY-MM-DD. If a fork branch was produced, it is pushed as a separate ref for the user's morning review. The user merges (or rejects) the fork manually.

5. Model-router calibration (v0.14.0+). Run kei-model-router calibrate once Phase C reaches step 5. The CLI reads the local ledger at ~/.claude/agents/ledger.sqlite (or $KEI_LEDGER_DB), performs leave-one-out grid search over kernel weights, and prints baseline-vs-best MSE plus the calibrated (α_role, α_caps, α_scope, α_body) tuple. Append the JSON line to ~/.claude/memory/router-calibration.jsonl:

   {"ts":"YYYY-MM-DDThh:mm:ssZ","rows":N,"baseline_mse":B,
    "best_mse":M,"weights":{...}}
   

   If rows < 5, write {"ts":"...","rows":N,"skipped":true} and continue — too few outcomes to recalibrate. The router falls back to the last persisted weights or compile-time defaults.

6. Affect-matrix aggregation (v0.14.1+). Aggregate the per-session affect files at ~/.claude/memory/affect/*.jsonl (written live by affect-live-scan.sh UserPromptSubmit hook) into a Phase C summary. Steps:

   a. Concatenate all affect files dated within the last DEEP_SLEEP_CRON_DAYS interval. b. Group by category and count. For any category whose count exceeds its threshold (repeat-signal ≥ 3, frustration-tone ≥ 5, conservative-framing ≥ 4, data-contamination ≥ 2) AND whose top-3 quotes share a semantic theme (manual review by the deep-sleep agent), emit a proposed_rule block in the plan markdown listing: - the recurring pattern (1 sentence) - 2-3 example user quotes (with timestamps) - suggested severity tier (remind / warn / enforce / block) - suggested hook event (UserPromptSubmit / PreToolUse / Stop) c. Append a one-line audit row to sync-repo/reports/affect-trends.jsonl:

      {"date":"YYYY-MM-DD","window_days":N,
       "counts":{"repeat-signal":R,"frustration-tone":F,
                 "conservative-framing":C,"data-contamination":D},
       "proposed_rules":N_proposed}
   

   This is the LEARNING LOOP — Phase C surfaces affect patterns to the user via the morning plan, the user clicks /escalate-recurrence to codify them. Without step 6, the affect matrix is a passive log; with it, the matrix feeds back into rule formation per RULE 0.10.

Zero-conflict guarantee

Any conflict the refactor-engine marks requires_human_decision is EXCLUDED from the generated patch and listed plainly in the plan under the matching section. The user sees every such item explicitly. No silent auto-apply of ambiguous changes.

Invariants

• Plan is ALWAYS written, even if the engine finds nothing (body reads "no conflicts this cycle").
• Fork branch never auto-merges to main.
• Phase C never touches traces/*.jsonl (append-only, inherited).
• Store backend is whatever kei-store status reports — the Phase C pipeline is store-agnostic.

Failure handling

• kei-conflict-scan fails → record the error in the plan body and skip fork.
• kei-refactor-engine fails → same; keep any partial plan markdown.
• Manual edits in step 3b produce merge conflicts → delete fork branch; append the conflict summary to the plan.
• Push fails → retry once; on second failure leave local commit and exit 1. Local state is recoverable on next run.