Sunday Review

/sunday-review — weekly meta-orchestrator

{SKILL_DIR} = this skill's own folder (locally: the directory this SKILL.md lives in; a served brain substitutes the real absolute path before you read this). Shared starter files live at the repo root two levels up: {SKILL_DIR}/../... If a path does not resolve, name the missing file and stop — never guess another location.

You are running the Sunday meta-review. This skill doesn't reinvent the existing weekly skills — it orchestrates them in the right sequence, surfaces the cross-cutting signal, and produces ONE clean note instead of N independent reports.

Order of operations

Run each step in order. After each step, capture the headline finding (1-2 sentences) into the running synthesis. Do NOT dump full reports inline — link to them.

Step 1 — Pattern recognition (/weekly)

Invoke the existing /weekly insights skill. It produces the panel-driven journal pattern recognition for the past 7 days. Capture: which floor was dominant, what pattern repeated, what was avoided.

Step 2 — Instinct Engine (/patterns)

Invoke /patterns. It scans recent sessions, journals, and decisions for hardening patterns and turns them into concrete captures. Capture: any new patterns ready to codify into a CLAUDE.md rule, concept note, or skill improvement.

Step 3 — Vault hygiene scan

Run:

python3 "{SKILL_DIR}/../../scripts/vault-hygiene.py" --quiet

It writes a fresh report to ⚙️ Meta/Vault Hygiene.md. Capture: how many broken wikilinks, empty notes, stale notes, duplicate concepts.

Step 3b — Backup restore drill (NO SKIP)

Run:

bash "{SKILL_DIR}/../../scripts/vault-backup.sh" verify --vault "$VAULT_ROOT"

verify decrypts the newest snapshot into a temp dir, extracts it, and counts the files. It is the only backup check that opens the archive. Three outcomes, three different reports:

• Exit 0 — capture the file count restored and the snapshot's date. A count far below the vault's real file count is a FAIL, not a pass.
• Exit non-zero saying vault not configured or no snapshot found — there is no backup at all. That is the week's one thing to do; give the setup command: bash "{SKILL_DIR}/../../scripts/vault-backup.sh" setup --vault "$VAULT_ROOT".
• Any other non-zero — an archive exists and does not restore. FAIL.

⛔ vault-backup.sh status does not substitute for this, and neither does the backup check in /diagnose. Both assert that an archive file exists and read its mtime — nothing about its contents. One real vault, moved inside a cloud-sync folder the scheduled job had no OS permission to read, wrote 32 consecutive nights of 10 KB archives against an 18 MB vault (tar failed, its stderr went to /dev/null). Every one of those nights was reported as "OK, 0.7 days old". verify caught it in 4 seconds.

Cheap corroborating signal, no decryption needed: list the snapshot directory with sizes. Two consecutive snapshots identical to the byte means the job is writing the same failure over and over.

Report a failure as a FAIL, never as a WARN, and carry it into Step 7 under "Vault state". Bug class STAMP-GREEN-WHILE-ARTIFACT-GONE: a check that probes the proxy (the file, the timestamp) instead of the leaf (the restorable content) stays green for exactly as long as the artifact is broken. Background: docs/BACKUP.md.

Step 4 — CLAUDE.md drift

Run:

python3 "{SKILL_DIR}/../../scripts/check-claude-md-drift.py" --quiet

It writes a report to ⚙️ Meta/CLAUDE-md drift.md. Capture: any dormant people, archived projects, broken links, or old codifications that need review.

Step 4b — Multi-edit semantic drift (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/drift-detection.py"

It writes a report to ⚙️ Meta/Drift Audit.md. Lists vault files edited 5+ times in the last 30 days (configurable via --days, --min-edits, --top, --include). Codified-rule files at the top are highest-leverage drift candidates. Capture: which rule file shows highest churn, and whether any of the diffs softened a guard or shifted a number without a Decisions/ entry. Skip silently if the script is missing (vault hasn't installed it yet). Inspired by Microsoft DELEGATE-52 (arxiv.org/abs/2604.15597) finding that frontier LLMs corrupt ~25% of professional content over 20 edits.

Step 4c — Cross-document rule conflicts (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/check-rule-conflicts.py" --scan-all

It writes a report to ⚙️ Meta/Rule Conflicts.md. Engram-inspired (github.com/Gentleman-Programming/engram) keyword-anchor detector — catches always X vs never X contradictions across the rules corpus. Add --semantic if ANTHROPIC_API_KEY is set for vocabulary-different contradiction detection via claude-haiku. Capture: any candidate conflicts at confidence ≥0.5 that need reconciliation. Pair signal with drift detection: drift = single-document shift over time; conflicts = cross-document clash at write time. Skip silently if the script is missing.

Step 4c.5 — Storage tier audit (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/storage-tier-audit.py"

Walks every vault subdirectory, reads its .tier declaration (GIT / DB / OS), flags directories missing the declaration or holding GIT-tier files >1MB. Output uses Compiled-Truth + Timeline format at ⚙️ Meta/Storage Tier Audit.md. Pattern source: garrytan/gbrain (cherry-picked). Capture: how many directories missing tier declarations, any oversized GIT-tier files. Add .tier files to the highest-traffic missing directories before next week. Skip silently if the script is missing.

Step 4c.6 — Zero-LLM typed-relationship refresh (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/wire_typed_relationships.py" \
  --output "$VAULT_ROOT/⚙️ Meta/typed-edges.jsonl"

Re-extracts typed edges (frontmatter + wikilinks) across the full vault using regex + entity-type rules. Zero LLM calls. Pattern source: garrytan/gbrain (cherry-picked). The output JSONL feeds /graphify Part A.5 so the next graphify run skips the structural extraction LLM cost. Capture: total edges by type (works_at / journaled_about / attended / etc.), runtime in ms, any extraction failures logged to stderr. Skip silently if the script is missing.

Step 4c.7 — Public repo standards audit (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/audit-public-repo-standards.py"

Audits every public non-archived non-fork repo under your GitHub org for baseline standards: LICENSE, README, real CI workflow (Dependabot auto-merge alone does not count), last-push freshness (>60d = stale). Classifies repo type (code-python, code-node, code-go, code-rust, skill, docs, meta) so CI requirements only apply to actual code repos. Output uses Compiled-Truth + Timeline format at ⚙️ Meta/Public Repo Standards Audit.md. Capture: any repos with gaps, especially missing LICENSE (MIT add via gh api -X PUT contents/LICENSE) or no CI workflow (extend the harden script's Layer 6 to handle the missing case). Skip silently if the script is missing.

Step 4c.8 — Hallucination tracking (if the scripts exist in this vault)

Run the aggregator first (fast, no LLM calls):

python3 "$VAULT_ROOT/⚙️ Meta/scripts/hallucination-watch.py" --quiet

Then run the sample audit (slower, requires Claude auth — ANTHROPIC_API_KEY in env OR a claude CLI subprocess can reach Max OAuth):

python3 "$VAULT_ROOT/⚙️ Meta/scripts/hallucination-sample-audit.py" --turns 6 --quiet

The aggregator tallies pre-ship catches from the hookify-blocks log, post-ship corrections from the Critical Failure Inventory, and fabrication-class feedback memories. Output: ⚙️ Meta/Hallucination Watch.md. The sample audit random-samples recent assistant turns, extracts verifiable factual claims, deterministic vault-greps the keywords, and Claude-classifies each claim as supported / contradicted / unverifiable. Output: ⚙️ Meta/Hallucination Sample Audit.md (overwrites) + appended history to ⚙️ Meta/Hallucination Sample Audit History.jsonl. Together they form a per-vault verification harness in the family DELEGATE-52 (arxiv 2604.15597, Apr 2026) identifies as the only consistent mitigation for multi-turn LLM corruption. Capture: this week's fabrication hook fires, any new Critical Failure Inventory rows, this run's verified_fraction, and the 4-week trend on verified_fraction (if 4+ history rows exist). Skip silently if either script is missing.

Step 4c.9 — Agent Memory orphan-debt trend (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/orphan-memo-trend.py"

Reports how many Agent Memory memos are NOT referenced in the MEMORY.md index, split into total (mostly expected cold mass) and non-audit (feedback_/project_/performance_/user_ — the actionable hot tier the discoverability close-gate flags), with the delta since last run persisted to ⚙️ Meta/logs/orphan-memo-trend.json. Policy: MEMORY.md is a curated, capped index (~200 lines / 24.4 KB Read-tool cliff), NOT a complete catalog — do NOT drive the count to zero; "orphan" is the correct resting state for cold / superseded memos. Only a RISING non-audit count is a regression (new hot memos written without an index entry — index them with the atomic memo / memo-finalize helper). Capture: this week's non-audit orphan count + delta. Skip silently if the script is missing.

Step 4d — Passive captures triage (if the script exists in this vault)

Run:

SINCE=$(date -v-7d +%Y-%m-%d 2>/dev/null || date -d "7 days ago" +%Y-%m-%d)
python3 "$VAULT_ROOT/⚙️ Meta/scripts/passive-capture.py" --scan-since "$SINCE"
python3 "$VAULT_ROOT/⚙️ Meta/scripts/passive-capture.py" --triage

Engram-inspired (mem_capture_passive) — scans the past week's session transcripts for utterances pattern-matched as rules, decisions, or lessons that were NOT explicitly filed via /journal or /decision. Writes triageable stubs to ⚙️ Meta/Passive Captures/{date}-{slug}.md. Idempotent via state file. Capture: how many pending captures, broken down by type (rule/decision/lesson). Adopt the load-bearing ones into CLAUDE.md or rules/ files; reject and archive the rest. Skip silently if the script is missing.

Step 4f — Closed-loop week report (if the script exists in this vault)

Run:

python3 "{SKILL_DIR}/../../scripts/closed-loop-week-report.py" \
  --vault-root "$VAULT_ROOT" --days 7

Reports the previous week's automatic activity in the episodic→procedural memory loop: how many promotion candidates landed in Meta/Promotion-Candidates/ awaiting human ratification, how many got auto-promoted into Meta/Workflows/ / Meta/Exceptions/ / Meta/Facts/, how many rules were demoted to status: superseded, and the current resolver conflict count. Inline the script's output into the synthesis under "Closed-loop activity." If candidates are pending, the user's "One thing to do this week" should be reviewing them. The closed-loop runs hourly (promote) and weekly (demote/conflicts), but the human ratification gate only fires once a week, here. Skip silently if the script is missing.

Step 4g — Security cadence check (if the script exists in this vault)

Run:

python3 "$VAULT_ROOT/⚙️ Meta/scripts/security-cadence-report.py"

Reads telemetry from every autonomous security layer (gh-harden daily fleet, workflow-injection scanner, endpoint-inventory daily NDJSON, secret-detection log, permission-denied log, Dependabot backlog per repo, hookify auto-commit volume) and writes ⚙️ Meta/Security Cadence Report.md. Designed as the weekly meta-audit of automation coverage — the question Layer-1 cron substrate doesn't answer: "is the automation still catching what it should?"

After the script lands the report, inline a 5-minute Claude pass answering the 6 drift questions at the report tail:

1. Did any cron-layer fail silently this week? (Baselines missing days, gh-harden log gap > 24h, scanner returning identical lines for multiple days)
2. Did any hookify rule fire zero times in 30d? (Rule earning its slot, or drifted off the actual risk surface — candidate for deletion)
3. Did secret-detection or permission-denied counts spike? (Spike = new risk surface OR false-positive class needing carve-out; investigate)
4. Any production-tier repo with >5 open Dependabot PRs? (Tier-protected prod repos cannot accumulate unreviewed vuln PRs)
5. Has the threat landscape shifted since last check? (Read 2-3 CISA + OWASP + vendor advisories — does any layer need a new pattern?)
6. Any new public repo / endpoint / agentic surface shipped this week not classified into the 4-layer detection model? (CI-publish workflow-injection / post-publish Dependabot / outbound SSRF / endpoint-installed-state inventory)

The Claude pass surfaces signals INTO the /sunday-review synthesis under a "Security cadence" subsection. If any question's answer is non-trivial → escalate to a full code-security review as event-driven, NOT calendar-driven. The three-layer cadence model:

• Layer 1 — daily cron (gh-harden + workflow-injection scan + endpoint inventory + Dependabot + hookify family). Auto, no skill invocation.
• Layer 2 — weekly meta-audit (this Step 4g via security-cadence-report.py).
• Layer 3 — event-driven full code-security review (new endpoint, new agentic surface, new CVE class, drift escalation from Layer 2). Not calendar. Not shipped here: /security-snapshot is passive and external-only, /secret-warn is edit-time guardrails, and both disclaim deep review. Bring your own reviewer.

Source: panel synthesis (Schneier + Charity Majors + Patrick Collison + DHH-dissent, 2026-05-27) on cadence-vs-event-driven security. Bug class prevented: ARTIFACT-WITH-OVER-STRICT-VERIFICATION applied to time — over-scheduled verification trains a bypass habit; real fires get ignored. Skip silently if the script is missing.

Step 5 — Decision retrospective

Run:

python3 "{SKILL_DIR}/../../scripts/decision-retrospective.py" --apply-prompt

This surfaces decisions older than 90 days with empty Outcome and appends review-ready prompts to ⚙️ Meta/Decision Retrospective.md. Capture: how many stale decisions need their Outcome filled in.

Step 6 — Skill usage curatorial pass (if telemetry enabled)

Run:

python3 "{SKILL_DIR}/../../scripts/skill-usage-report.py"
python3 "{SKILL_DIR}/../../scripts/curate-skills-surface.py" --top 5 --days 7

Capture: the 3 most-used skills this week + any skills that haven't been used in 30+ days (dormant — candidates for pruning or re-promoting).

Step 7 — Synthesize into one note

Write a single markdown note to 📓 Journals/Reviews/Sunday Review {YYYY-MM-DD}.md with this structure:

---
creationDate: {ISO timestamp}
type: review
category: sunday-meta
week_of: {YYYY-MM-DD}
---

# Sunday Review — {YYYY-MM-DD}

## Highlights this week
[1-2 sentences from /weekly + /patterns: what was the dominant emotional pattern, what shipped, what stalled]

## What's hardening (from /patterns)
[any new patterns that appeared 3+ times → ready to codify]

## Vault state
- Hygiene: [link to Meta/Vault Hygiene.md] · {broken wikilinks count}, {empty notes count}, {stale notes count}
- Backup restore drill: {PASS — N files restored from {snapshot date} | FAIL — what broke | NOT CONFIGURED}
- CLAUDE.md drift: [link to Meta/CLAUDE-md drift.md] · {N signals flagged}
- Stale decisions: [link to Meta/Decision Retrospective.md] · {N candidates for outcome backfill}

## Skill usage
- Top this week: {skill 1}, {skill 2}, {skill 3}
- Dormant: {names of skills not used in 30 days}

## One thing to do this week
[Pick the highest-leverage action from the captures above. Just one.]

Step 8 — Surface to the user

Reply with a 2-3 sentence summary. Link to the Sunday Review note. Name the one thing to do this week. End there. Do NOT inline the whole review.

Why this exists

Five Sunday-relevant skills exist already (/weekly, /patterns, vault-hygiene, claude-md-drift, decision-retrospective). Running them sequentially by hand is friction. Running them via this orchestrator produces one synthesized output that respects the user's attention budget.

The Matuschak panel critique was that more skills don't deepen thinking unless they compound. This is the compounding layer: it forces the existing tools to interlock once a week.

When NOT to use this

• During the week (these are weekly-cadence checks; daily use is overkill).
• When the user explicitly only wants /weekly (the journal pattern read), not the full meta-review.
• When the user says they want to skip the system and just journal — respect that.

Configuration

In CLAUDE.md frontmatter, you can disable specific steps:

sundayReview:
  skipHygiene: false
  skipDrift: false
  skipRetro: false
  skipTelemetry: true   # if telemetry not opted-in

If a step's underlying script is missing, skip silently and note it in the synthesis under "Pending steps."