hbzheng0942/greybox-host-bridge

Read {{SKILL_ROOT}}/SKILL.md for the production workflow and its human decisions. Use python "{{SKILL_ROOT}}/scripts/greybox.py" as the common command prefix. No Node, MCP or model switch is required. This wrapper is optional: WorkBuddy can also read the complete Skill and execute its CLI directly. Use the command prefix returned by setup; it records the actual Python and verified package identity.

For long-running art cases, pin the actual entry in the existing PLAN using task-context pin and continue with task-context run; read {{SKILL_ROOT}}/references/task-context.md. Recheck after a handoff or model/tool change. Report unavailable model/schema facts as unknown. task-context inspect distinguishes installed copies and each original ledger's legacy/executor recovery route; never recreate paid jobs to make them fit the newer executor.

1. Read the current task by intent, observed state and requested endpoint. Use intent route --file REQUEST --project PROJECT for structured request routing. Discussion stays discussion; greybox requests default to a greybox endpoint; continue uses original task records. Run host setup --host workbuddy --scope greybox --project PROJECT for greybox work, or host setup --host workbuddy --scope art --project PROJECT --capability-report REPORT --session-id SESSION for the art/image branch. Setup resolves the actual native project root and returns the command prefix. Bind current native ImageGen invocation evidence when available; no API profile is then needed. Tool discovery alone is insufficient. Image services and unrelated tools are not greybox prerequisites. For Blender scene work, use host setup --host workbuddy --scope spatial --engine blender --operation create|edit|review --project SOURCE_OR_WORKSPACE when CLI preparation is needed; available native bpy/CLI/MCP operations can then follow the spatial route. Setup is readiness only; application discovery does not inspect a scene. A pure spatial endpoint needs no game project or supplier.

2. For Godot/Unity greybox creation/revision, read {{SKILL_ROOT}}/references/greybox-authoring.md and use greybox init/build/play/validate/status/handoff; stop at the requested endpoint. For the art branch, if generation is missing, read {{SKILL_ROOT}}/references/image-services.md and configure the user's chosen image service with image configure. Bind its non-secret name using host configure --host workbuddy --image-profile NAME.

3. For new S7 modeling references, read {{SKILL_ROOT}}/references/modeling-references.md, record each asset and its views once, and enter boards workbuddy start --file REGISTRY --project PROJECT --workspace WORKSPACE. This compiles the page jobs and registers/installs their scoped S7 guard. Run its exact harmless native Bash probe, then boards workbuddy bind with the current native tool schema. Consume the returned jobs; boards workbuddy claim writes the exact call-ready arguments and reserves the existing spending ceiling. A planned single follows this same path and needs its computed or sourced reason. Record the original result, then crop/review through boards production; a bad slot does not erase good siblings, and corrected ROIs create crop attempts without regenerating. A new chat renews the free probe and preserves original pending calls. For a chosen image profile, consume the exact prepared page requests through the existing configured service ledger. Do not lower quality requirements or submit an uncertain call again.

4. When the active branch produces images, open the actual returned local image through the current host's pixel-reading tool. Present that file using the host's supported image UI. CLI success does not establish either capability. Bind exact current tools and evidence in the session capability report. Greybox review uses its native captured views and does not require image generation.

5. Continue the commands for the current route and requested endpoint. Deliver and stop after greybox handoff when stop_at is greybox; enter art only when requested. Preserve the original project state and approval history. Delegation is optional; image/model multi-task concurrency uses the common executor described in {{SKILL_ROOT}}/references/execution-contract.md. Native tools remain agent-driven unless an actual independent bridge is verified; API workers can continue while the model connection stalls.

Before H3, compare the approved concept with the current whole-scene proxy via scene-review --help. At H4, new production contracts require current feature/view comparison evidence. Read {{SKILL_ROOT}}/references/scene-comparison.md; use observed gaps to repair environment, major forms, light and contact before expansion. Valid evidence coverage never supplies the user's visual acceptance.

Commands return JSON; failed commands return a nonzero code. Use each command's --help for exact current options. Keep credentials in local configuration or the configured environment, never in chat, connector files or portable project records. A service credential is required only for the chosen service; connector initialization does not log into a provider or create a paid task.

Only when the current operation needs Lux, run host setup --host workbuddy --scope runtime --runtime lux. It prepares/reuses the pinned private runtime and can recover from a failed venv/ensurepip probe using an existing compatible interpreter. It does not query suppliers or submit generation. Preserve its returned Python path across commands. An explicit --python is never silently replaced.