omar/panopainter
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Document multi-agent director workflow

Browse Source
This commit is contained in:
omigamedev
2026-06-12 18:42:00 +02:00
parent c9c3df2733
commit c98130a51f
3 changed files with 199 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

190
docs/modernization/director-workflow.md Normal file
View File

@@ -0,0 +1,190 @@
# Modernization Director Workflow
Status: live
Last updated: 2026-06-12
Use this workflow when the user explicitly asks for subagents, delegation, a
director, captains, or parallel agent work. Do not spawn subagents just because
a task is complex. The default is still one agent executing one task from
`docs/modernization/tasks.md`.
## Goals
- Save main-thread tokens by keeping long exploration and simple edits out of
the director context.
- Use stronger models only where they change the outcome.
- Keep each implementation slice measurable, validated, committed, and pushed.
- Avoid merge conflicts by giving every agent a disjoint task and file scope.
## Roles
### Director
The director is the main agent in the user thread. The director owns:
- choosing the task group from `docs/modernization/tasks.md`
- deciding whether delegation is worth the coordination cost
- spawning one `gpt-5.4` captain per independent task group
- integrating returned changes
- running final validation
- updating docs/debt/tasks
- committing and pushing the verified slice
- resetting conversation context after the slice when practical
The director should not delegate the immediate blocking step. If the next local
action is required before anything else can proceed, do that locally.
### Team Captain
Use `gpt-5.4` for each captain. A captain owns one coherent task group, for
example:
- renderer/export boundary
- legacy adapter retirement
- platform/package parity
- UI lifetime
- dependency cleanup
- test/hardening work
The captain turns one task-group objective into smaller disjoint subtasks. If
nested subagents are available, the captain may spawn workers. If nested
subagents are not available in the current surface, the captain returns a worker
plan and the director performs the second-level spawns.
The captain must not edit broad shared files unless assigned. The captain must
not rewrite the task tracker or roadmap except for a requested status note.
### Workers And Explorers
Workers perform bounded edits in assigned files. Explorers answer specific
questions and should usually not edit files.
Every worker and explorer must be told:
- this repository may have other agents working in parallel
- do not revert or overwrite unrelated changes
- stay inside the assigned scope
- report changed files, validation run, and blockers
## Model Selection
| Work Type | Model | Reasoning Effort | Use |
| --- | --- | --- | --- |
| Captain for a task group | `gpt-5.4` | `medium` default, `high` for renderer/platform architecture | Split task, supervise workers, summarize integration. |
| Risky implementation with cross-file C++ behavior | `gpt-5.4` | `medium` or `high` | Code changes touching contracts, build graph, or live adapters. |
| Bounded implementation in known files | `gpt-5.4-mini` | `medium` | Localized tests, simple adapters, docs/debt updates, small refactors. |
| Fast lookup or inventory | `gpt-5.3-codex-spark` | `low` or `medium` | `rg` inventory, file ownership map, simple grep-based answers. |
| Mechanical docs cleanup | `gpt-5.3-codex-spark` | `low` | Formatting, table updates, command list normalization. |
| Ambiguous strategy or failed captain escalation | `gpt-5.5` | inherited or `high` | Only when cheaper models are likely to make architectural mistakes. |
Prefer the inherited model unless the user requested this director workflow or
there is a clear task-specific reason to override. In this workflow, the user
has requested `gpt-5.4` captains, so use that override for captains.
## Token Discipline
- Use `fork_context=false` by default. Pass the task id, relevant files, debt
ids, validation commands, and only the necessary excerpts.
- Use `fork_context=true` only when prior conversation details are essential and
not already in repo docs.
- Do not paste large logs into prompts. Point agents at log paths and ask for
the smallest relevant excerpt.
- Do not ask captains or workers to "read the roadmap" generally. Name the
exact task row and debt ids.
- Keep subagent prompts under about 250 words for explorers and 500 words for
workers/captains unless the task truly needs more.
- Ask for compact final reports: changed files, result, validation, blockers,
next recommendation.
- Close completed agents after their results are integrated or rejected.
- Prefer one captain and two or three workers over many agents. Coordination
cost can exceed token savings.
## Delegation Flow
1. Director picks one `Ready` task or one small group of adjacent tasks from
`docs/modernization/tasks.md`.
2. Director identifies the immediate local blocking step and keeps it local.
3. Director spawns one `gpt-5.4` captain for sidecar decomposition or an
independent task group.
4. Captain returns either:
- a worker plan with disjoint write scopes, or
- completed bounded edits if the captain can finish the task directly.
5. Director or captain spawns worker/explorer agents only for disjoint subtasks.
6. Workers edit only their assigned files and run focused validation when cheap.
7. Director reviews and integrates changes.
8. Director runs the task's listed validation command.
9. Director updates `tasks.md`, `debt.md`, and `roadmap.md` if the task moved.
10. Director commits and pushes the verified slice.
## Director Prompt Template For A Captain
```text
You are the gpt-5.4 team captain for the <group> task group in PanoPainter.
Task source: docs/modernization/tasks.md task <TASK-ID>.
Debt ids: <DEBT-LIST>.
Assigned scope: <FILES/DIRS>.
Validation: <COMMANDS>.
Goal: <ONE PARAGRAPH>.
First, split this into disjoint worker tasks suitable for gpt-5.4-mini or
gpt-5.3-codex-spark. Use gpt-5.4 only for risky C++ behavior changes. Keep
tasks small enough to validate. Do not edit outside the assigned scope. Other
agents may be working in parallel; do not revert unrelated changes.
Return:
- recommended worker tasks with model choice and file scope
- any task you can safely complete yourself
- risks/blockers
- minimal context each worker needs
```
## Captain Prompt Template For A Worker
```text
You are a worker on PanoPainter. Other agents may be editing nearby files; do
not revert unrelated changes.
Task: <WORKER-TASK>.
Model fit: <WHY THIS MODEL IS ENOUGH>.
Write scope: <FILES/DIRS ONLY>.
Read scope: <FILES/DIRS>.
Validation: <COMMANDS>.
Make the smallest behavior-preserving change that satisfies the done checks.
Update tests/docs only if listed. If the task is larger than expected, stop and
report the split instead of broadening scope.
Final report:
- files changed
- behavior changed
- validation run and result
- blockers or follow-up
```
## Explorer Prompt Template
```text
Answer one codebase question for PanoPainter.
Question: <QUESTION>.
Search scope: <FILES/DIRS>.
Use compiler-aware navigation if this depends on C++ symbols.
Do not edit files.
Return only:
- answer
- supporting file references
- confidence and caveats
```
## Final Integration Checklist
- No worker changed files outside its assigned scope without calling it out.
- No generated logs or build output were committed.
- Focused validation for the task passed or the failure is documented.
- `docs/modernization/debt.md` changed when debt was narrowed or closed.
- `docs/modernization/tasks.md` score changed only for `Done` tasks.
- The commit contains one coherent slice.
- The branch was pushed.