repo-b-mcp-comfy-bridge

name: repo-b-mcp-comfy-bridge description: Canonical MCP adapter and Comfy bridge operations for . Use when enabling MCP, validating shim.comfy resources/tools, operating workflow/pipeline submissions, or running fail-closed Comfy diagnostics with optional AMUSE and CapCut contract checks.

REPO_B Shim MCP Comfy Bridge

Use this skill as the canonical MCP + Comfy lane in <PRIVATE_REPO_B>.

A legacy local Comfy wrapper is retained for backward-compatible triggers and should route here for new work.

Workflow

1. Enforce loopback-only MCP and fail-closed defaults.
2. Validate MCP runtime state via /api/mcp/status.
3. Apply /api/mcp/config changes only when configuration drift is confirmed.
4. Validate required Comfy resources (shim.comfy.status, shim.comfy.queue, shim.comfy.history).
5. Validate Comfy tools (shim.comfy.prompt.submit, shim.comfy.workflow.submit, shim.comfy.pipeline.run).
6. Validate optional AMUSE status/capabilities and profile-level CapCut export metadata contracts.
7. Fail closed on stale status or schema contract violations.

Scope Boundary

Use this skill for MCP adapter state/config plus shim.comfy.* resource health.

Do not use this skill for:

1. OpenAI-thin-waist chat/vision/jobs route debugging.
2. Agent Bridge task/write-mode safety lanes.
3. Hybrid Windows-host/WSL network topology checks.

Required Environment (PowerShell)

$env:SHIM_ENABLE_MCP = "1"
$env:SHIM_MCP_HOST = "127.0.0.1"
$env:SHIM_MCP_PORT = "9550"
$env:SHIM_MCP_ALLOW_LAN = "0"
$env:MX3_COMFYUI_ENABLED = "1"
$env:MX3_COMFYUI_BASE_URL = "http://127.0.0.1:8188"
$env:MX3_COMFYUI_TIMEOUT_S = "10"
$env:MX3_COMFYUI_DEFAULT_WORKFLOW_PROFILE = "small_video"
$env:MX3_AMUSE_ENABLED = "1"
$env:MX3_AMUSE_BASE_URL = "http://127.0.0.1:3001"
$env:MX3_AMUSE_TIMEOUT_S = "15"
$env:REPO_B_LOCAL_COMFY_ORCH_FAIL_CLOSED = "1"

Runtime/API Checks

curl http://127.0.0.1:9000/api/mcp/status
curl http://127.0.0.1:9000/api/comfy/workflows/templates
curl http://127.0.0.1:9000/api/comfy/pipelines/profiles
curl http://127.0.0.1:9000/api/amuse/status
curl http://127.0.0.1:9000/api/amuse/capabilities

Apply config:

curl -X POST http://127.0.0.1:9000/api/mcp/config \
  -H "content-type: application/json" \
  -d '{"enabled":true,"persist":true,"apply_now":true}'

Resource Contract Checks

Validate these resource contracts before accepting diagnostics:

1. shim.comfy.status:
   • enabled=true
   • reachable=true
   • empty last_error
   • fresh checked_at
2. shim.comfy.queue:
   • running_prompt_ids and pending_prompt_ids lists
   • running_count and pending_count must match list lengths
3. shim.comfy.history:
   • entries list
   • count must match entries length

Workflow/Pipeline Contract Checks

1. shim.comfy.workflow.submit accepts one of:
   • workflow object
   • workflow_id
   • workflow_path
   • workflow_profile
2. shim.comfy.pipeline.run supports profile defaults and optional AMUSE stage:
   • profile values include small_video_capcut and quality_video_capcut
   • wait_for_state must normalize to NONE|QUEUED|RUNNING|DONE
   • when capcut_preset=true, response includes capcut_export.editor=capcut
   • when amuse_enhance=true, response includes additive amuse payload

Pipeline run example:

curl -X POST http://127.0.0.1:9000/api/comfy/pipelines/run \
  -H "content-type: application/json" \
  -d '{"profile":"small_video_capcut","wait_for_state":"RUNNING","capcut_preset":true}'

Advanced Diagnostics (Optional Legacy Tooling)

When deterministic artifact output is needed, use the legacy local orchestrator drop-ins (tools/local_comfy_orchestrator.py, tools/local_comfy_validate.py):

python tools/local_comfy_orchestrator.py \
  --task comfy-health-check \
  --json-out .codex/local_comfy/comfy-health-check.json \
  --repo-root .

Comfy Bridge Control Vars

• MX3_COMFYUI_ENABLED
• MX3_COMFYUI_BASE_URL
• MX3_COMFYUI_API_KEY
• MX3_COMFYUI_TIMEOUT_S
• MX3_COMFYUI_DEFAULT_WORKFLOW_PROFILE
• MX3_AMUSE_ENABLED
• MX3_AMUSE_BASE_URL
• MX3_AMUSE_TIMEOUT_S

References

• references/mcp-comfy-checklist.md

Loopback

If this lane is unresolved, blocked, or ambiguous:

1. Capture current evidence and failure context.
2. Route back through $skill-hub for chain recalculation.
3. Resume only after the updated chain returns a deterministic next step.