ch-build-map

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# UE Map Subagent Workflow

## Core contract

The confirmed parent is the design authority, contract owner, and final visual gate. Use `gpt-5.6-luna` for light or moderate bounded work and `gpt-6.1-sol` for complex design, integration, map writing, or adjudication. Workers receive bounded deliverables. Exactly one final-map builder writes the target `.umap`; no other agent edits it concurrently.

## Profile selection and syntax

Invoke `$ch-build-map [low|med|high] <request>`. `medium` aliases `med`; omitting the level selects `med`. A level is recognized only as the first token after the skill name. Model IDs are not positional arguments.

- `low`: light, bounded changes with straightforward geometry, layout, and checklist-based QA.
- `med`: moderate work with complex asset preparation, layout, integration, or structural review; simple surveys and bounded technical modules use Luna.
- `high`: complex work with coupled geometry, layout, technical systems, and structural decisions; decision-making workers use Sol 6.1.

Use exactly one selected profile's table. Assign only roles actually needed. If a deliverable exceeds the selected profile's scope, report the mismatch and request a profile change before dispatching it; do not silently mix tables. Every profile retains all applicable acceptance gates.

### Parent model gate

- `low` requires `gpt-5.6-luna / medium`.
- `med` and `high` require `gpt-6.1-sol / xhigh`.
- Verify the exact runtime parent model and reasoning before planning, reading project assets, or dispatching workers. If the required pair cannot be verified or is unavailable, return `BLOCKED_MODEL` without starting the workflow.
- Do not switch the runtime model yourself. Ask the user to switch to the required pair when it does not match.

## Role table — profile `low`

| Role | Model | Reasoning |
|---|---|---|
| Parent / art director / integrator | `gpt-5.6-luna` | `medium` |
| Geometry explorer | `gpt-5.6-luna` | `medium` |
| Asset inventory and license scout | `gpt-5.6-luna` | `medium` |
| Blender hero-asset worker | `gpt-6.1-sol` | `high` |
| Layout worker | `gpt-5.6-luna` | `medium` |
| Technical worker | `gpt-5.6-luna` | `medium` |
| Final-map builder | `gpt-6.1-sol` | `high` |
| Evidence worker | `gpt-5.6-luna` | `medium` |
| Structural QA | `gpt-5.6-luna` | `medium` |
| Final visual QA | `gpt-5.6-luna` | `medium` |

## Role table — profile `med`

| Role | Model | Reasoning |
|---|---|---|
| Parent / art director / integrator | `gpt-6.1-sol` | `xhigh` |
| Geometry explorer | `gpt-5.6-luna` | `medium` |
| Asset inventory and license scout | `gpt-5.6-luna` | `medium` |
| Blender hero-asset worker | `gpt-6.1-sol` | `high` |
| Layout worker | `gpt-6.1-sol` | `high` |
| Technical worker | `gpt-5.6-luna` | `medium` |
| Final-map builder | `gpt-6.1-sol` | `high` |
| Evidence worker | `gpt-5.6-luna` | `medium` |
| Structural QA | `gpt-6.1-sol` | `high` |
| Final visual QA | `gpt-6.1-sol` | `xhigh` |

## Role table — profile `high`

| Role | Model | Reasoning |
|---|---|---|
| Parent / art director / integrator | `gpt-6.1-sol` | `xhigh` |
| Geometry explorer | `gpt-6.1-sol` | `high` |
| Asset inventory and license scout | `gpt-5.6-luna` | `medium` |
| Blender hero-asset worker | `gpt-6.1-sol` | `high` |
| Layout worker | `gpt-6.1-sol` | `high` |
| Technical worker | `gpt-6.1-sol` | `high` |
| Final-map builder | `gpt-6.1-sol` | `high` |
| Evidence worker | `gpt-5.6-luna` | `medium` |
| Structural QA | `gpt-6.1-sol` | `high` |
| Final visual QA | `gpt-6.1-sol` | `xhigh` |

## Shared role ownership

- Parent owns the brief, room contract, integration decisions, and acceptance; it remains read-only against the target map while the builder runs.
- Geometry explorer surveys the map read-only; inventory/license scout produces a read-only inventory and report.
- Hero-asset worker creates isolated source/export/import-ready assets. This complex authoring role uses Sol 6.1 even in `low`.
- Layout worker produces a transform proposal, footprint manifest, or sandbox level. Technical worker produces isolated cable, bracket, service, collision, or material modules.
- Final-map builder is the sole writer of the named target `.umap`; this integration role uses Sol 6.1 in every profile.
- Evidence worker captures screenshots, hashes, counts, logs, and reports only. Structural QA audits collision, overlap, clearance, support, reload, and PIE read-only.
- Final visual QA opens the images directly and decides PASS/FAIL.

Only the user may explicitly override one role. An override must name the role, exact model, and exact reasoning. Do not infer a global override from prose and do not restore the former “one model for every worker” behavior.

If the runtime rejects any required model/reasoning pair, stop before dispatching that role and report the exact rejection. Do not lower reasoning or substitute another model automatically. This skill uses only `gpt-5.6-luna` and `gpt-6.1-sol`.

## Reasoning promotion ladder

Use `medium` for light, repeatable, bounded work assigned to `gpt-5.6-luna`. Use `high` or `xhigh` for complex geometry, assets, collision, integration, or QA assigned to `gpt-6.1-sol`. More actor count, more files, or a tighter deadline never alone justifies higher reasoning.

Promote one worker from `high` to `xhigh` only when the parent records at least one observable trigger:

- two or more hard constraints conflict (for example geometry, clearance, sightline, mounting, or lighting);
- the change spans two coupled systems (for example geometry plus collision, or material plus final visibility) and a `high` pass did not resolve it;
- its same deliverable failed review after one targeted correction; or
- the available evidence cannot distinguish a visual/design defect from a spatial/integration defect.

Apply that promotion only to the bounded `gpt-6.1-sol` worker and deliverable that needs it. Do not escalate Luna inventory/evidence work merely because it has many files.

`max` is not a routine final gate and is never assigned to a worker. The confirmed `gpt-6.1-sol` parent may use it temporarily for one adjudication only when both conditions hold: (1) two `xhigh` attempts or reviews remain in material conflict or fail to resolve a non-reversible decision; and (2) the decision changes the room contract, protected architecture, or a costly-to-rework integration choice. Record the conflict, the options, and the decision. Otherwise remain at parent `xhigh`.

## Required sequence

1. **Survey without mutation.** Measure the real level, coordinate frame, floor, walls, glass, doors, ceiling, collision, existing assets, dirty files, protected actors, and player capsule. Illustrative coordinates never override measured geometry.
2. **Freeze the room contract.** The parent writes the manifest and acceptance rubric before any target-map mutation.
3. **Prepare assets in isolation.** Inventory, hero assets, layout proposals, and technical modules may run in parallel only when their files and ownership do not overlap.
4. **Approve assets before placement.** The confirmed parent inspects actual asset evidence. Rejected assets return to their owner; the map builder does not imitate missing assets with primitives.
5. **Run one builder.** The named `gpt-6.1-sol` builder writes only the target map and explicitly scoped integration files from the approved manifest.
6. **Collect independent evidence.** The evidence worker and structural reviewer use the selected profile's exact assignments. Neither edits the target map during QA.
7. **Parent visual gate.** The confirmed parent opens every required eye-level and top-down image. A failed visual gate returns targeted corrections to the same builder or asset owner.
8. **Reload and report.** Reopen the saved level, rerun affected checks, and distinguish completed work from unresolved or unverified work.

## Manifest required before placement

Each managed actor or functional cluster includes:

- stable ID and owning role;
- asset path, source, license, and redistribution status;
- location, named-axis rotation, scale, footprint, and bottom/top bounds;
- supporting surface or mounting actor;
- access side and required clearances;
- power, data, sample, or fluid endpoints;
- collision policy;

From chauipp/dotagents, MIT.

More from chauipp/dotagents