App Subagent Orchestrator (local, read-only)

This folder contains prompts + scripts to coordinate App bug-hunt subagents.

Files

• prompts/orchestrator-prompt.md — orchestrator-level instructions for Codex
• prompts/subagent-prompt.md — one-off finding prompt (grounded, one issue)
• prompts/fixer-prompt.md — simple-fix handoff prompt
• schema/bug-report.schema.json — expected output shape
• scripts/run_bug_hunt.py — orchestrates subagents, dedupes findings, optionally dispatches simple fixes

Quick start

cd /path/to/repo
python3 .omx/subagent_orchestrator/scripts/run_bug_hunt.py \
  --target 100 \
  --max-agents 5 \
  --prompt .omx/subagent_orchestrator/prompts/subagent-prompt.md

Results appear under:

• .omx/subagent_orchestrator/runs/<run-id>/raw/
• .omx/subagent_orchestrator/runs/<run-id>/merged/unique_findings.json
• .omx/subagent_orchestrator/runs/<run-id>/final/unique_findings.md
• .omx/subagent_orchestrator/runs/<run-id>/final/needs_decision.md
• .omx/subagent_orchestrator/runs/<run-id>/final/simple_findings.md
• .omx/subagent_orchestrator/runs/<run-id>/final/run_log.md

Duplicate safety

Each finding is deduped by signature:

{file}:{line}:{category}:{normalized_issue}

If two agents return the same anchor, only the first is kept.

Fix pass behavior

When target is reached and --no-fix is not set, subagents are spawned for all simple findings (requires_decision == false). Their outputs are written to:

• final/fixes/*.json
• final/simple_fixes.jsonl

Items requiring direction are kept in:

• final/needs_decision.md

Stop conditions

• stop at 100 unique findings by default,
• or earlier if no new findings appear for three consecutive rounds,
• or when tasks are exhausted.

Notes

• Everything is designed for read-only discovery mode; scripts do not modify source.
• No tests are run by the orchestrator.