/audit-cron-arc

name: audit-cron-arc
description: Meta-analyze the effectiveness of a /loop arc — per-iteration metrics, LOC delta trend, saturation detection, and a keep/lengthen/delete recommendation.

/audit-cron-arc [--cron-id=<id>]

Analyzes a completed or running `/loop` arc to determine whether it's still producing value or has hit saturation. Emits a markdown report with per-iteration table, ASCII trend chart, saturation verdict, and an actionable recommendation.

Usage

/audit-cron-arc
/audit-cron-arc --cron-id=loop-1718643000

Arguments

| Flag | Default | Description | |------|---------|-------------| | `--cron-id=<id>` | most recent CronCreate'd job | Loop job identifier to audit |

Logic

Step 0 — Arc-type classifier (run FIRST; gates all downstream thresholds)

Scan files modified across all iterations via `git diff --name-only <start-sha>..<end-sha>`.

**Classify each modified file:**

| Bucket | File patterns | |--------|--------------| | DOCTRINAL | `rules/*.md`, `_router.md`, `prompt-cache.md`, `commands/*.md`, `skills/**/*.md` | | BUG-FIXING | `mcp-servers/*/src/index.ts`, `bin/validate*.mjs`, `rules/*-discipline.md` derived from incidents (check git log for "Principle #", "fix:", "patch:", "guard") | | META | `retrospectives/*.md`, `rules/audit-baselines/*.json`, `settings.json`, hook files |

**Assign arc type from file counts across ALL iterations:**

• ≥60% DOCTRINAL → `arc_type = DOCTRINAL`
• ≥60% BUG-FIXING → `arc_type = BUG-FIXING`
• Otherwise → `arc_type = MIXED`

Emit arc type at report top. Example:

Arc type: BUG-FIXING (14 of 19 modified files match bug-fix patterns across iters 14–18)

---

1. **Resolve cron job** — if `--cron-id` omitted, read `~/.claude/loops/active.json`; pick highest `created_at`. 2. **Load task log** — call `TaskList` filtered by `cron_id` tag; sort by `iteration` asc. 3. **Compute per-iteration metrics:**

• `task_count` — tasks spawned this iteration
• `files_created` — `git diff --stat` between iteration start+end SHA (new files only)
• `loc_delta` — net lines added/removed (`git diff --numstat`)
• `tools_used` — unique tool call types from the task transcript
• `duration_s` — wall-clock seconds between task start and last child completion
• `new_findings` — (BUG-FIXING only) distinct violation classes reported; parse transcript for `N violations`, `N findings`, `N call sites`, `N errors`
• `findings_patched` — (BUG-FIXING only) violations resolved this iteration

4. **Detect saturation curve — per arc type:**

**DOCTRINAL (LOC-based):**

• `loc_per_iter[i] < loc_per_iter[i-1] * 0.80` (>20% drop) → "converging"
• Three consecutive converging iterations → `saturated`
• Any growing or flat iteration → `still-runway`

**BUG-FIXING (seam-fork; LOC is NOT the primary signal):**

• `net_findings[i] = new_findings[i] - findings_patched[i]`
• `net_findings[i] > 0` → seam forking → `still-runway`
• `new_findings[i] == 0` AND `findings_patched[i] > 0` → `converging`
• `new_findings[i] == 0` AND `findings_patched[i] == 0` for ≥2 consecutive iters AND backlog empty → `saturated`
• LOC-drop threshold does NOT trigger convergence; surgical edits have volatile LOC curves.
• task_count plateau at 3 for 5+ iters is NOT a convergence signal.

**MIXED:**

• `saturated` only when BOTH DOCTRINAL (LOC-based) AND BUG-FIXING (zero-finds) criteria simultaneously indicate saturation. Either alone = `still-runway`.

5. **Render report** (see Output format). 6. **Recommend — per arc type:**

**DOCTRINAL:**

• `still-runway` → `keep-cron`
• `converging` → `lengthen-interval` (2× current interval)
• `saturated` → `delete-cron`

**BUG-FIXING:**

• `still-runway` → `keep-cron` (seam forking; keep running validators)
• `converging` → `keep-cron` (backlog draining; do not lengthen)
• `saturated` → `rotate-validator` first; if rotating produces zero new findings → `delete-cron`

**MIXED:**

• Apply the more conservative recommendation (never delete if either domain shows runway).

Output format

# Cron Arc Audit — <cron-id>

**Audited:** <timestamp>
**Arc type:** <DOCTRINAL | BUG-FIXING | MIXED>  ← always emit; gates thresholds below
**Arc-type rationale:** <N of M modified files matched <pattern>; <brief reason>>
**Iterations:** <N>
**Total LOC delta:** <+X / -Y>
**Total duration:** <HH:MM:SS>

## Per-Iteration Table

<!-- For DOCTRINAL arcs: -->
| Iter | Tasks | Files+ | LOC Δ | Tools | Duration | Trend |
|------|-------|--------|-------|-------|----------|-------|
|  1   |   4   |   3    | +210  |   7   |   2m14s  |  —    |
|  2   |   6   |   5    | +380  |   9   |   3m01s  |  ↑    |
|  3   |   5   |   2    | +190  |   6   |   2m30s  |  ↓    |
|  4   |   3   |   1    |  +45  |   4   |   1m55s  |  ↓↓   |

<!-- For BUG-FIXING arcs: add New Findings + Patched columns; LOC Trend column is informational only -->
| Iter | Tasks | Files+ | LOC Δ | New Findings | Patched | Net | Seam |
|------|-------|--------|-------|--------------|---------|-----|------|
| 14   |   3   |   4    | +820  |      4       |    0    |  +4 | fork |
| 15   |   3   |   5    | +480  |   1134       |    0    | +1134| fork |
| 16   |   3   |   3    | +170  |    581       |  1134   | -553| close|
| 17   |   3   |   5    | +747  |      0       |  747    | -747| drain|
| 18   |   3   |   6    | +310  |      0       |    0    |   0 | dry? |

## LOC-per-Iteration Trend (ASCII)

LOC 400 │ ██ 350 │ ██ 300 │ ██ ██ 250 │ ██ ██ 200 │ ██ ██ ██ 150 │ ██ ██ ██ 100 │ ██ ██ ██ 50 │ ██ ██ ██ ██ 0 └────────────────── I1 I2 I3 I4

## Verdict

**<still-runway | converging | saturated>**

<1-2 sentence rationale.>

## Recommendation

**<keep-cron | lengthen-interval | delete-cron>**

<What to do and why. If lengthen: suggest new interval. If delete: surface any remaining seam to tackle manually.>

Execution script

CRON_ID=""

for arg