| name | agent-orchestration |
| description | Orchestrate one Crash WOC unit or one function with Copilot using the repo-local helper scripts, prompt entry points, and a one-writer-at-a-time workflow. |
Agent Orchestration
Purpose
Use this skill when the task is larger than a one-off lookup but should still stay tightly scoped to one file or one function.
Entry Points
python tools/ai_orchestrate.py --unit <unit-or-path>
python tools/ai_orchestrate.py --function <symbol-or-address>
.github/prompts/orchestrate-unit.prompt.md
.github/prompts/match-function.prompt.md
.github/prompts/reverse-engineer.prompt.md
Operating Rules
- Use read-only helper scripts and read-only exploration freely.
- Keep exactly one active code-writing flow at a time.
- Keep one active function selected at all times for decomp edits.
- Prefer
ai_lookup_symbol.py, ai_lookup_unit.py, ai_context.py, ai_match_plan.py, and ai_ghidra.py over ad hoc repo-wide searching.
- Never use interactive
objdiff-cli diff sessions in automated loops.
Suggested Unit Workflow
- Run
python tools/ai_orchestrate.py --unit <unit-or-path> --no-launch --print-prompt.
- Review the chosen next function and the generated helper commands.
- Launch Copilot with the same command without
--no-launch when ready.
- Re-run
python tools/ai_match_plan.py <unit-or-path> after each measurable change.
- If a function stalls for 4 build/measure cycles, note the blocker and move on.
Suggested Function Workflow
- Run
python tools/ai_orchestrate.py --function <symbol-or-address> --no-launch --print-prompt.
- Refresh lookup and context with
ai_lookup_symbol.py and ai_context.py.
- Use
ai_ghidra.py if the original-side context is still unclear.
- Edit only the target function, then build only the owning object.
References