Design — Harness Configuration Review

Delegate the full review to the harness-config-advisor agent. The agent
holds the merged context for agent-roster and command-set design and runs in
its own context window. Present its recommendations and wait for approval
before applying changes.

What to ask the agent

Hand the agent the four review axes below and request a single consolidated
report. Do not duplicate espalier audit . — that command already verifies
manifest integrity, broken refs, and JSON validity. The skill body covers
judgment calls the audit cannot make.

1. Agent roster

• Each agent's name, model tier, and current role
• Coverage gaps (frameworks/layers no agent owns)
• Stale agents (added for features that no longer exist)
• Model tier fit (Opus for reasoning-heavy, Sonnet for mechanical)
• Count discipline: 3–6 agents is a soft default, not a rule — flag count only alongside observed dysfunction (overlap, unused agents, coverage gaps). 7 tightly-scoped agents beats 4 bloated ones.

2. Command set

• One-line purpose per command
• Universal commands present (/context-load, /handoff, /status, /smoke)
• Commands reference real paths in this repo
• Workflows the developer hits weekly that lack a command
• Count discipline: 10–17 commands (more = developer forgets they exist)

3. Hook coverage

Standard 12 hooks (verify all active in .claude/settings.json):

• session_start.py (SessionStart) — load blueprint chain
• task_router.py (UserPromptSubmit) — route to /implement-task or --multi
• plan_guard.py (PreToolUse) — block source writes without active plan
• write_guard.py (PreToolUse) — block mutations of protected zones (a write, a delete, a move)
• config_guard.py (ConfigChange) — block unsafe project/local/user settings changes
• post_write_check.py (PostToolUse) — validate writes
• reflect_trigger.py (PostToolUse) — trigger reflect every 10th write
• stop_gate.py (Stop) — lightweight session hygiene; full pytest gate via ESPALIER_STOP_GATE=full
• subagent_stop.py (SubagentStop) — append subagent reasoning to active blueprint
• post_compact.py (PostCompact) — re-inject critical context
• subagent_start.py (SubagentStart) — inject cold-subagent orientation
• context_reinject_failure.py (PostToolUseFailure) — re-derive after a failed Edit/Write

Flag hooks referencing paths that no longer exist.

4. Documentation freshness

• ESPALIER_MEMORY.md lean and current (the harness caps it at 120 lines)
• docs/CONVENTIONS.md matches actual code patterns (run /analyze if uncertain)
• docs/SHARP_EDGES.md covers real footguns (not generic advice)
• Session log current (last 5 rows)

for f in ESPALIER_MEMORY.md docs/CONVENTIONS.md docs/SHARP_EDGES.md docs/CHEAT-SHEET.md docs/TASK_RECIPES.md; do
  [ -f "$f" ] && echo "$f: $(git log -1 --format='%ar' -- $f 2>/dev/null)" || echo "$f: MISSING"
done
wc -l ESPALIER_MEMORY.md 2>/dev/null

Expected output

HARNESS CONFIGURATION REVIEW
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Agents ({N}):   {list}
Commands ({N}): {list}
Hooks:          {list}

Recommended changes:
  ADD:    {list or NONE}
  REMOVE: {list or NONE}
  MODIFY: {list or NONE}

Per-finding format:

COMPONENT: {name}
ACTION:    KEEP | ADD | REMOVE | MODIFY
REASON:    {concrete evidence}
DETAILS:   {for ADD/MODIFY — exactly what to change}
PRIORITY:  ESSENTIAL | RECOMMENDED | NICE-TO-HAVE

After approval

Apply approved changes directly, then run python -m espalier audit . to verify
surface integrity.