autodev-parallel

name	autodev-parallel
description	AutoDeveloper orchestrator that runs multiple experiments in parallel using worktrees. Triggers: "parallel experiments", "autodev parallel", "simultaneous experiments", "worktree experiments" Anti-triggers: "sequential experiments", "one at a time"
user-invocable	true
disable-model-invocation	false
allowed-tools	Read, Write, Edit, Bash, Grep, Glob, Agent

AutoDeveloper Parallel -- Parallel Experiment Orchestrator

Runs multiple experiments simultaneously via git worktrees and selects the best result.

Core Concept

main (or current branch)
 |
 +-- worktree A -- Agent 1: Experiment idea 1
 +-- worktree B -- Agent 2: Experiment idea 2
 +-- worktree C -- Agent 3: Experiment idea 3
 |
 +-- Orchestrator (this skill)
      - Generate & assign ideas
      - Collect & compare results
      - Select best branch
      - Repeat for next round

Phase 0: Configuration Collection

Confirm with user:

goal: "What to achieve"
scope: ["Modifiable file patterns"]
metric: "Evaluation command"
parallel: 3                       # Concurrent execution count (default 3)
rounds: 5                         # Number of rounds (default 5, total experiments = parallel * rounds)

Phase 1: Baseline

# Measure baseline from current state
mkdir -p .autodev
{metric_command} > .autodev/baseline.log 2>&1
bash ~/.claude/hooks/autodev-judge.sh
BASELINE_SCORE=$(cat .autodev-score)

# Initialize results file
echo -e "round\tagent\tcommit\tscore\tstatus\tdescription" > .autodev/results.tsv

Phase 2: Round Loop

for round in 1..rounds:

  1. BRAINSTORM
     - Generate {parallel} ideas based on scope file analysis + goal
     - Reference previous round results (build on kept changes)
     - Ideas must be independent of each other (avoid conflicts)

  2. LAUNCH (parallel)
     - Invoke {parallel} Agents simultaneously
     - Each Agent is isolated with isolation: "worktree"
     - Prompt passed to each Agent:

     """
     Perform a single AutoDeveloper experiment.

     Goal: {goal}
     Scope: {scope}
     Metric: {metric}
     Idea: {specific_idea}

     Procedure:
     1. Read files within scope and apply the idea
     2. git commit -m "[autodev] {idea_summary}"
     3. Run {metric} -> .autodev/run.log
     4. Run bash ~/.claude/hooks/autodev-judge.sh
     5. Return score as final message:
        AUTODEV_RESULT: score={N}, commit={hash}, description="{desc}"

     On build failure, attempt recovery once. On 2nd failure, give up:
        AUTODEV_RESULT: score=-999, commit=none, description="{desc} (crash)"
     """

  3. COLLECT
     - Wait for each Agent to complete
     - Parse score from results
     - Record in results.tsv

  4. SELECT
     - Identify the worktree/branch of the highest-scoring Agent
     - If highest score > current best_score:
       - Cherry-pick changes from that branch to current branch
       - Update best_score
     - Remaining worktrees are cleaned up (Agent tool handles automatically)

  5. REPORT (per round)
     Round {round}/{rounds} complete:
     - Agent 1: score={n1} ({status1}) -- {desc1}
     - Agent 2: score={n2} ({status2}) -- {desc2}
     - Current best: score={best}

  6. CONTINUE
     Proceed to next round (apply new ideas on top of previous best)

Phase 3: Final Report

# AutoDev Parallel Experiment Report

## Summary
- Total rounds: {rounds}
- Total experiments: {rounds * parallel}
- Keep: {K} / Discard: {D} / Crash: {C}
- Baseline -> Final: {baseline} -> {best} ({improvement}%)

## Per-Round Results
| Round | Best Experiment | Score | Description |
|-------|-----------------|-------|-------------|
| 1 | Agent 2 | 85 | Added cache layer |
| 2 | Agent 1 | 120 | Query batching |
| ... | ... | ... | ... |

## Cumulative Kept Changes
1. [commit1] Added cache layer
2. [commit2] Query batching

## Next Steps
- Additional optimization areas: ...
- Ready to merge into main: `git merge autodev/{tag}`

Parallel Count Guidelines

Situation	Recommended parallel
Independent files/modules	5 (maximum)
Changes within the same file	1-2 (conflict risk)
Includes performance benchmarks	2-3 (resource sharing)
Tests only for evaluation	3-5

Safety Measures

1. Worktree isolation: Uses Agent tool's isolation: "worktree". No mutual interference
2. Main protection: Uses cherry-pick only. No force push
3. Existing test protection: Unconditionally discard if broken
4. Resource limits: Do not exceed the parallel count
5. Cross-round synchronization: Use previous round's best as the next round's base

Leveraging Existing Agents (Optional)

TTH team members can be used as experiment agents:

Team Member	Suitable Experiment Types
pichai (architect)	Structural changes, module separation
jensen (backend)	API optimization, DB query improvement
zuckerberg (frontend)	Component optimization, bundle reduction
bezos (QA)	Code deletion, removing unnecessary abstractions

Include the team role file path in the prompt to provide expertise:

Read the role file for the team member and perform the experiment from that perspective:
~/.claude/team-roles/{role}.md