MCP addon that lets jcode use arena.ai Agent Mode as sub-agents.
arena-addon exposes 5 tools with basic schemas and descriptions, but lacks depth in parameter documentation and output schema clarity. Tool names follow verb_noun convention (arena_spawn, arena_status, arena_wait, arena_list, arena_download_workspace), which is good. However, descriptions are minimal (34-194 chars, within baseline but sparse), and critical parameter descriptions are missing or trivial. For example, arena_status and arena_wait both accept 'session_id' with description 'Session identifier', no guidance on format, origin, or lifecycle. arena_spawn's 'upload_files' lacks constraints (file size, count, types). Output schemas are not documented anywhere in the code, LLMs cannot predict what fields to expect from responses. Error handling is present in the Coordinator class (e.g., 'no such session' errors) but not surfaced in tool descriptions, so agents cannot plan recovery. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (arena_spawn and arena_download_workspace are WRITE operations). This is a functional but under-documented server typical of early-stage MCP implementations.
Fetch the Arena agent's workspace files into a local directory (when supported by the driver).
List previously spawned Arena agent sessions and their statuses.
Spawn an arena.ai Agent Mode run as a sub-agent of jcode. Returns a session_id you can poll with arena_status or block on with arena_wait. The Arena agent autonomously plans and completes the task in its own workspace.
Get current status/output/result of a previously spawned Arena agent.
Block until the spawned Arena agent finishes (or timeout). Returns final result.
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting). Coordinator.status(), wait(), list(), download() return dicts but no schema is published in tool_specs().
Parameter descriptions are trivial or missing. 'session_id' described as 'Session identifier' (no format, origin, or lifecycle guidance). 'upload_files' has no constraints (size, count, file types). 'timeout_secs' lacks range (min/max). 'dest_dir' has no validation rules (must exist? writable?).
No tool annotations despite clear risk levels. arena_spawn (WRITE) and arena_download_workspace (WRITE) should have destructiveHint=true. arena_status and arena_list should have readOnlyHint=true. Agents cannot distinguish safe from risky operations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Error handling not surfaced in tool descriptions. Coordinator._monitor() catches exceptions and stores 'failed' status, but tool descriptions do not explain error conditions, recovery steps, or when to retry. Agents cannot plan recovery.
arena_list returns all sessions without pagination. If hundreds of sessions exist, response bloats context window. No limit, offset, or cursor parameters.