evaluate-presets（评估预设）

Evaluate Presets

Overview

Systematically test all hat collection presets using shell scripts. Direct CLI invocation—no meta-orchestration complexity.

When to Use

• Testing preset configurations after changes
• Auditing the preset library for quality
• Validating new presets work correctly
• After modifying hat routing logic

Quick Start

Evaluate a single preset:

./tools/evaluate-preset.sh tdd-red-green claude

Evaluate all presets:

./tools/evaluate-all-presets.sh claude

Arguments:

• First arg: preset name (without .yml extension)
• Second arg: backend (claude or kiro, defaults to claude)

Bash Tool Configuration

IMPORTANT: When invoking these scripts via the Bash tool, use these settings:

• Single preset evaluation: Use timeout: 600000 (10 minutes max) and run_in_background: true
• All presets evaluation: Use timeout: 600000 (10 minutes max) and run_in_background: true

Since preset evaluations can run for hours (especially the full suite), always run in background mode and use the TaskOutput tool to check progress periodically.

Example invocation pattern:

Bash tool with:
  command: "./tools/evaluate-preset.sh tdd-red-green claude"
  timeout: 600000
  run_in_background: true

After launching, use TaskOutput with block: false to check status without waiting for completion.

What the Scripts Do

evaluate-preset.sh

1. Loads test task from tools/preset-test-tasks.yml (if yq available)
2. Creates merged config with evaluation settings
3. Runs Ralph with --record-session for metrics capture
4. Captures output logs, exit codes, and timing
5. Extracts metrics: iterations, hats activated, events published

Output structure:

.eval/
├── logs/<preset>/<timestamp>/
│   ├── output.log          # Full stdout/stderr
│   ├── session.jsonl       # Recorded session
│   ├── metrics.json        # Extracted metrics
│   ├── environment.json    # Runtime environment
│   └── merged-config.yml   # Config used
└── logs/<preset>/latest -> <timestamp>

evaluate-all-presets.sh

Runs all 12 presets sequentially and generates a summary:

.eval/results/<suite-id>/
├── SUMMARY.md              # Markdown report
├── <preset>.json           # Per-preset metrics
└── latest -> <suite-id>

Presets Under Evaluation

Interpreting Results

Exit codes from evaluate-preset.sh:

• 0 — Success (LOOP_COMPLETE reached)
• 124 — Timeout (preset hung or took too long)
• Other — Failure (check output.log)

Metrics in metrics.json:

• iterations — How many event loop cycles
• hats_activated — Which hats were triggered
• events_published — Total events emitted
• completed — Whether completion promise was reached

Hat Routing Performance

Critical: Validate that hats get fresh context per Tenet #1 ("Fresh Context Is Reliability").

What Good Looks Like

Each hat should execute in its own iteration:

Iter 1: Ralph → publishes starting event → STOPS
Iter 2: Hat A → does work → publishes next event → STOPS
Iter 3: Hat B → does work → publishes next event → STOPS
Iter 4: Hat C → does work → LOOP_COMPLETE

Red Flags (Same-Iteration Hat Switching)

BAD: Multiple hat personas in one iteration:

Iter 2: Ralph does Blue Team + Red Team + Fixer work
        ^^^ All in one bloated context!

How to Check

1. Count iterations vs events in session.jsonl:

# Count iterations
grep -c "_meta.loop_start\|ITERATION" .eval/logs/<preset>/latest/output.log

# Count events published
grep -c "bus.publish" .eval/logs/<preset>/latest/session.jsonl

Expected: iterations ≈ events published (one event per iteration) Bad sign: 2-3 iterations but 5+ events (all work in single iteration)

2. Check for same-iteration hat switching in output.log:

grep -E "ITERATION|Now I need to perform|Let me put on|I'll switch to" \
    .eval/logs/<preset>/latest/output.log

Red flag: Hat-switching phrases WITHOUT an ITERATION separator between them.

3. Check event timestamps in session.jsonl:

cat .eval/logs/<preset>/latest/session.jsonl | jq -r '.ts'

Red flag: Multiple events with identical timestamps (published in same iteration).

Routing Performance Triage

Root Cause Checklist

If hat routing is broken:

1. Check workflow prompt in hatless_ralph.rs:

- Does it say "CRITICAL: STOP after publishing"? - Is the DELEGATE section clear about yielding control?

1. Check hat instructions propagation:

- Does HatInfo include instructions field? - Are instructions rendered in the ## HATS section?

1. Check events context:

- Is build_prompt(context) using the context parameter? - Does prompt include ## PENDING EVENTS section?

Autonomous Fix Workflow

After evaluation, delegate fixes to subagents:

Step 1: Triage Results

Read .eval/results/latest/SUMMARY.md and identify:

• ❌ FAIL → Create code tasks for fixes
• ⏱️ TIMEOUT → Investigate infinite loops
• ⚠️ PARTIAL → Check for edge cases

Step 2: Dispatch Task Creation

For each issue, spawn a Task agent:

"Use /code-task-generator to create a task for fixing: [issue from evaluation]
Output to: tasks/preset-fixes/"

Step 3: Dispatch Implementation

For each created task:

"Use /code-assist to implement: tasks/preset-fixes/[task-file].code-task.md
Mode: auto"

Step 4: Re-evaluate

./tools/evaluate-preset.sh <fixed-preset> claude

Prerequisites

• yq (optional): For loading test tasks from YAML. Install: brew install yq
• Cargo: Must be able to build Ralph

Related Files

• tools/evaluate-preset.sh — Single preset evaluation
• tools/evaluate-all-presets.sh — Full suite evaluation
• tools/preset-test-tasks.yml — Test task definitions
• tools/preset-evaluation-findings.md — Manual findings doc
• presets/ — The preset collection being evaluated