FastMCP 3.2 bridge for Vision-Language-Action: Wall-OSS-0.5, WALL-WM, DMuon co-training, and fleet simulation loops
VLA-MCP defines 9 tools with complete input schemas and descriptions. Naming follows verb_noun convention (vla_*) but lacks action verbs (get_, list_, create_). Descriptions are domain-specific but average ~150 chars, within baseline (194 chars p50). All parameters have types and descriptions. However, output schemas are not documented in the source code provided, and error handling guidance is absent. Tool composition is reasonable but some tools combine multiple operations (e.g., vla_weights with list/download/status). No evidence of pagination, rate limiting, or recovery guidance.
VLA_DATASET - Dataset ingestion, episode management, export, validation, and telemetry segmentation.
VLA_FLEET - Fleet peer discovery, health probes, and MCP bridge status.
VLA_PIPELINE_LIVENESS - Stale VLA loops and unreachable fleet peers. Surfaces when the last E2E pipeline run is older than stale_hours and probes worldlabs/robotics/avatar peers for reachability. Intended for supervisors (meta-mcp, fleet-agent, aiwatcher) and daily checks.
VLA_STATUS - Snapshot of Wall-OSS, WALL-WM, DMuon, HF cache, and dataset.
VLA_TRAINING - Launch co-training loops (DMuon, Wall-OSS, WALL-WM) with confirmation gate.
VLA_WALL - Portmanteau for Wall-OSS-0.5 VLA (gradient-bridged MoT + flow matching).
Output schemas not documented. No evidence of return type definitions for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
Tool names lack action verbs. All tools use 'vla_*' prefix but no leading verb (get_, list_, create_, etc.). Reduces LLM's ability to infer intent from name alone.
vla_status description is only 50 chars ('Snapshot of Wall-OSS, WALL-WM, DMuon, HF cache, and dataset.'). Below baseline minimum of 34 chars but lacks context on when/why to call it.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | <=2025-11-25 | v2 |
VLA_WEIGHTS - Hugging Face checkpoint download for Wall-OSS and WALL-WM.
VLA_WORLD_MODEL - WALL-WM world action model (event joints, Wan prior).
VLA_XVLA - X-VLA 0.9B flow-matching VLA with PEFT for edge agents (Raspbot/Boomy).
No error handling guidance. Tools lack recovery hints (e.g., 'If peer unreachable, try list_peers first'). Errors will not guide LLM on retry strategy or alternatives.
vla_training combines multiple operations (launch_co_train, status, cancel) in one tool. Should split into separate tools (launch_training, get_training_status, cancel_training) for single responsibility.