AGENT SKILL

repo-health-review

Systematic repo health review — survey scripts, detect duplicates, check naming, find gaps, prune dead weight, and produce a structured report. Run before major refactors or when asked to audit the repo.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 4
Forks
1

01 前言 Front Matter

name
repo-health-review
pinned
false
version
1.1.0
category
devops
description
Systematic repo health review — survey scripts, detect duplicates, check naming, find gaps, prune dead weight, and produce a structured report. Run before major refactors or when asked to audit the repo.
related_skills
[ "fleet-preflight", "documentation-auditing", "change-checklist" ]

02 资源文件 6

03 SKILL.md 正文

Repo Health Review — Systematic Audit

When to Use

  • User asks "review the repo", "audit for gaps", "look for consolidation opportunities"
  • Before a major refactor
  • When onboarding a new team member (to document current state)

Multi-Axis Audit (preferred for comprehensive reviews)

For thorough reviews, run 3+ parallel subagents via delegate_task — one per axis. Each subagent scans a different dimension simultaneously, keeping your context uncluttered while they work:

# Standard 3-axis dispatch
delegate_task(tasks=[{
    "goal": "Audit docs for stale paths, broken cross-refs, missing sections",
    "context": "Check docs/ for stale src/, deploy/, runtime/ references..."
}, {
    "goal": "Audit install/update/doctor pipeline",
    "context": "Check fleet-update.sh register() validity, doctor --quick..."
}, {
    "goal": "Audit Linux vs macOS platform support and agent types",
    "context": "Check platform-specific code, SOUL.md profiles..."
}])

After subagents return, consolidate findings, fix easy items immediately, then produce a structured report of what was fixed vs deferred.

The 7-Pass Review (use for smaller/script-focused audits)

Pass 1: Structure Survey

ls ops/scripts/*/           # All script directories
ls docs/                    # All docs
ls skills/                  # All skill categories

Pass 2: Duplicate Detection

Check for files with same name in multiple directories:

find ops/scripts/ -name '*.py' -o -name '*.sh' | sed 's|.*/||' | sort | uniq -d

Verify identical content via MD5 when found:

md5sum <path1> <path2>

Pass 3: Naming Check

Identify scripts breaking hyphen convention:

find ops/scripts/ -name '*_*' -not -path '*/__pycache__/*' -not -name '__*'

Pass 4: Registration Check (Bi-Directional)

Check BOTH directions:

  • Scripts in ops/scripts/ that are NOT registered in fleet-update.sh
  • register() entries in fleet-update.sh that point to NON-EXISTENT files
# Direction A: Which scripts are NOT registered?
grep -oP 'register\s+"[^"]+' ops/scripts/fleet-update.sh | sed 's/register "//'

# Direction B: Which registered files are MISSING on disk?
grep -oP 'register\s+"[^"]+' ops/scripts/fleet-update.sh | sed 's/register "//' | while read f; do [ -f "$f" ] || echo "MISSING: $f"; done

Stale register() entries silently fail during deploy — the file simply doesn't get copied. Over time this causes the deployed state to diverge from expectations. Fix by removing the stale line from fleet-update.sh.

Also check for duplicate registrations (same file registered twice) and path drift (file was renamed but register wasn't updated, e.g. offline_code_index_cron.sh → offline-code-index-cron.sh).

Pass 4b: Stale Deploy Check

Check for deployed files that are no longer registered in fleet-update.sh. When a register() line is removed, the deployed copy stays on disk as an orphan. The doctor detects these:

python3 ~/steadfaste-fleet/ops/scripts/manage/fleet-doctor.py 2>&1 | grep "Stale deploy"

Fix with:

bash ~/steadfaste-fleet/ops/scripts/fleet-update.sh --clean-stale

Or run fleet-update.sh (auto-cleans stales after deploy). For non-scripts deploy targets (dashboard/, bus/, systemd units, launchd plists), manual cleanup may be needed.

Pass 4c: Doctor Comprehensive Check

After any structural change, run the full doctor and check ALL sections:

python3 ~/steadfaste-fleet/ops/scripts/manage/fleet-doctor.py 2>&1 | grep -E "❌|⚠️|ℹ️" | head -20

New in this session (2026-07-23):

  • check_stale_deploys() — detects orphaned deploy files and missing sources
  • Orphan cron detection — flags crons in scheduler not in expected list
  • Network safety — scans for services on 0.0.0.0 (Ollama, Bus, Langfuse)
  • Systemd linger — verifies Linux user services survive reboot
  • Skills disk presence — verifies every listed skill exists on disk

The doctor itself was modularized into fleet_doctor/ package (8 modules). Monolith at fleet-doctor.py is now a 34-line shim. To add a new check: create a new function in checks.py and add it to the all_checks list in cli.py.

Pass 5: Doc Cross-Reference

Check which docs are NOT referenced by any other doc, AGENTS.md, or README.md:

Search every .md for each doc filename. Unreferenced docs may be lost or stale.

Pass 6: Stale File Check

  • __pycache__ directories
  • .pytest_cache directories
  • Root-level test files (should be in tests/)
  • Zero-byte files
  • runtime symlinks

Pass 7: Gap Analysis

  • PRDs with no named implementation
  • Overlapping scripts doing similar things
  • Features documented but not wired

Pass 8a: Platform & Agent-Type Consistency

Check that platform-aware code handles both Linux and macOS:

grep -rn "systemctl\|launchctl\|brew\|apt-get\|platform\|uname" ops/scripts/ --include="*.py" --include="*.sh"

Identify scripts that assume one platform without an OS guard. Key patterns:

  • systemctl used without falling back to launchctl on macOS
  • apt-get used on macOS (should be brew)
  • Hardcoded paths like /home/ (Linux-only) or /Users/ (macOS-only)

Check SOUL.md profiles for platform consistency:

  • Does each agent profile declare a platforms: field in frontmatter?
  • Do orchestrator profiles specify Linux?
  • Do dev-agent profiles specify macOS/Linux?

Check agent-type coverage in docs:

  • docs/agent-architecture.md — lists all 4 roles (orchestrator, backup-orch, server-agent, dev-agent)?
  • docs/fleet-reference.md — maps real agent names (Moses, Esther, etc.) to roles?
  • docs/cron-schedules.md — split by orch-*, agent-*, local-* prefixes?
  • AGENTS.md — mentions agent type differences?

Pass 8b: Cross-Script Deploy/Install/Doctor Consistency Audit

Purpose: Verify that the three core lifecycle scripts — install.sh, fleet-update.sh, and fleet_doctor/ — agree on agent-type branching. If they diverge, agents get false-positive doctor failures, wrong cron expectations, or orphan deploy artifacts.

Run when: User asks to audit agent-type branching, fix doctor false positives on non-orch agents, or verify deploy/install/doctor pipeline consistency. Also run before adding a new agent type.

Methodology — check 5 dimensions:

8b-1: Capability Model vs Role Model

Identify whether the system uses a capability model (FLEET_PROFILE=core|laptop|server — how much hardware) or a role model (--agent-type orchestrator|server|dev — what services to run):

# Which model does install.sh use?
grep -n 'FLEET_PROFILE' ops/install/install.sh | head -5
grep -n 'agent.type\|agent_type\|--agent-type' ops/install/install.sh | head -5

# Which model does the doctor use?
grep -n 'is_orch\|IS_ORCH\|hostname\|agent.type' ops/scripts/manage/fleet_doctor/config.py

Gap to detect: A capability model (FLEET_PROFILE) cannot distinguish between an orchestrator (needs bus server + dashboard + nginx) and a server-agent (needs bus client + crons, NOT dashboard/nginx). When both map to server, orchestrator-only services get deployed to server-agents and produce doctor false positives.

8b-2: register() Guard Coverage in fleet-update.sh

Check whether orchestrator-only service files (dashboard, fleet-bus server, nginx configs) are conditionally registered or deployed to every agent:

# Are dashboard/bus/nginx files registered unconditionally?
grep -n 'register.*dashboard\|register.*bus\|register.*nginx' ops/scripts/fleet-update.sh

# Is deploy_nginx_configs() guarded?
grep -B3 -A3 'deploy_nginx_configs' ops/scripts/fleet-update.sh

# Is deploy_system_scripts() guarded?
grep -B3 -A3 'deploy_system_scripts' ops/scripts/fleet-update.sh

Check: Is there an IS_ORCHESTRATOR, IS_SERVER, or FLEET_PROFILE guard on these? If they run unconditionally, server-agents and dev-agents get files they don't need. The only guard is FLEET_SKIP_NGINX env var — optional, not default.

8b-3: Doctor EXTERNAL_SERVICES Hardcoding

The doctor's config.py defines EXTERNAL_SERVICES — endpoints it checks for health. If this list is hardcoded with Dashboard, Langfuse, and Agent Bus, it will FAIL on every non-orch agent:

grep -A10 'EXTERNAL_SERVICES' ops/scripts/manage/fleet_doctor/config.py

Check: Is there agent-type awareness? Does the doctor skip Dashboard/Langfuse checks on server-agents that don't host them? Also check check_nginx() — does it run unconditionally?

8b-4: Cron Separation and Doctor Cron Expectation

Three cron layers should be cleanly separated. Verify cross-referencing:

# Universal crons (all agents)
grep -n 'agent-' ops/scripts/install-crons.sh | head -5

# Orchestrator-only crons
grep -n 'orch-' ops/scripts/install/install-orch-crons.sh | head -5

# Dev-agent crons — local-* prefix (created manually, no installer exists)
grep -n 'local-' ops/scripts/install-crons.sh | head -5

Check path 1: Does the doctor's parse_expected_crons() correctly exclude orch crons from the universal expected list?

grep -A15 'parse_expected_crons' ops/scripts/manage/fleet_doctor/config.py

Check path 2: Does check_crons() correctly identify orchestrator hosts and only FAIL on missing orch crons for those hosts?

grep -A15 'orch_crons_list' ops/scripts/manage/fleet_doctor/checks.py

Check path 3: Does check_crons() check the IS_ORCHESTRATOR env var (like fleet-update.sh does), or only hostname? Hostname-only is brittle — won't match fresh installs or custom hostnames.

Gap to detect: If parse_expected_crons() doesn't match install-orch-crons.sh, or if check_crons() hostname detection differs from install-orch-crons.sh, the doctor will produce inconsistent results.

8b-5: Dev-Agent Minimal Path

Check whether dev agents (e.g. titus) have a valid automated setup path:

# Is there an install-local-crons.sh script?
find ops/ -name '*local*' 2>/dev/null

# Does install.sh --agent-type dev exist?
grep 'dev\|titus' ops/install/install.sh

# Does fleet-update.sh skip dashboard/bus/nginx for dev agents?
grep 'IS_DEV\|agent.dev\|dev.minimal\|skip.*dev' ops/scripts/fleet-update.sh

Gap to detect: Dev agents with no automated installer must create local-* crons manually. If the local-* prefix is documented in comments but no scaffolding exists (no installer, no doctor awareness, no update guard), the dev-agent path is incomplete.

8b-6: Verify Hostname Detection Uniformity

The same hostname/orch-detection logic should appear in:

  1. fleet-update.sh (lines ~1412-1438) — IS_ORCHESTRATOR env var + hostname fallback
  2. fleet_doctor/checks.py (lines ~407-421) — hostname only
  3. install-orch-crons.sh (lines ~29-39) — IS_ORCHESTRATOR env var + hostname fallback

Gap: If the doctor doesn't read IS_ORCHESTRATOR from .env, it will fail to detect orchestrator hosts not named moses/esther. Scripts 1 and 3 agree; script 2 may be out of sync.

8b-7: Produce Cross-Script Consistency Table (see also: references/agent-type-system.md)

After checking all 5 dimensions, produce a table showing which files branch on what:

| Dimension | install.sh | fleet-update.sh | fleet_doctor/ |
|-----------|-----------|-----------------|----------------|
| Service install guard | FLEET_PROFILE=server | none (unconditional deploy) | none (hardcoded EXTERNAL_SERVICES) |
| Nginx deploy guard | FLEET_PROFILE=server | FLEET_SKIP_NGINX (optional) | none (always checks) |
| Orch cron guard | N/A | IS_ORCHESTRATOR + hostname | hostname only |
| Dev-agent path | none | none | none |

This table makes gaps visible at a glance. Each row that differs by column is a consistency failure.

User Preference Signals

When the user corrects your approach or style, capture the lesson:

Signal Response
"No. Refactor. Do it right" Don't take shortcuts — thin wrappers are tech debt. Do the full refactor, not a compromise.
"Finish. Don't wait for future" Don't defer actionable items. Fix them in the same session — every issue found during audit must be fixed or explicitly acknowledged before reporting done.
"Thoroughly test" Every change must be exercised with real tool output before reporting done.
"Make agent X aware of Y" If the doctor or pipeline should check for Y, add a check — don't just document it.
User repeats a correction Add a structural guardrail (skill update, pre-commit hook, doctor check, auto-cleanup) that makes the mistake impossible.

Encode these in the relevant skill's SKILL.md body — not just in memory. Skills capture "how to do this class of task for this user"; memory captures "who the user is and what the current situation is."

Deliverable

Produce a structured report with:

  1. Summary — counts: scripts, docs, skills, issues found, issues fixed
  2. Fixed Items — what was changed (moved, renamed, pruned, archived)
  3. Documented Items — what was noted for future action
  4. Priority Recommendations — ordered by effort vs impact

Output Path

Persist the report as a durable markdown document in the repo, committed so the whole fleet benefits:

docs/reviews/repo-health-review-YYYY-MM-DD.md

Treat that as relative to the repo root (e.g. steadfaste-fleet/docs/reviews/). Prior reports live alongside it in docs/reviews/ — keep new reports there, never at the docs root.

Anti-Patterns

  • ❌ Only searching disk, not git — file may exist in repo but not deployed
  • ❌ Skipping MD5 verification — two files with same name may differ
  • ❌ Not checking cross-references before renames — breaks other agents
  • ❌ Leaving stales without documenting the decision

04 同仓库更多 Skills