Loading ToolBench
MCP stdio bridge for connecting an AWR control plane to an IDE, exposing world state, actions, branches, capabilities, and adapters as MCP tools
22 tools with schemas present but significant quality gaps. All tools have descriptions (10-200 chars, baseline 194), but many lack parameter descriptions and output schemas. Naming follows verb_noun pattern (awr_world_*, awr_branch_*, awr_action_*) but is domain-specific and verbose. Parameters are typed (string, object, enum, integer) but lack inline descriptions explaining constraints, formats, or dependencies. No error handling guidance visible. Schemas are minimal JSON Schema with required/properties but no output documentation. Risk classification present (READ_ONLY, WRITE) but not reflected in tool annotations.
Review an action without invoking an adapter. Returns authorization, blockers, risk, current revision, and whether execution is allowed.
List adapter manifests, observables, mutables, side effects, and rollback boundaries.
Create an isolated branch descriptor from the current main state.
List the main branch and currently active isolated branches.
Preview changes, conflicts, uncertainties, and approval requirements for merging a branch.
Read the canonical state of an isolated branch.
Inspect visible operations and capability grants for a branch.
No parameter descriptions visible. Tools like awr_world_action accept 'objective', 'parameters', 'target_branch', 'mode', 'base_revision' but LLMs cannot infer what each controls without inline docs. Baseline: 100% of A+ tools have param descriptions.
No output schemas documented. Tools return state deltas, audit results, and event streams but LLMs cannot plan downstream calls without knowing response structure. Baseline: 100% of A+ tools have documented return types.
Tool annotations missing. Risk classification (READ_ONLY vs WRITE) is present in metadata but not exposed as readOnlyHint/destructiveHint in MCP tool definitions. Agents cannot distinguish safe from destructive operations without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Run one allowlisted local program through the shell adapter. Output is returned as observations.
Read the append-only event stream and verify its hash chain.
Move an experience through the explicit lifecycle state machine.
Persist an observed experience record for later evaluation and reuse.
List persisted experience records and their lifecycle states.
Read a repository file through the filesystem adapter with root containment checks.
Inspect Git status, HEAD, or diff through the Git adapter.
Run the workspace test command through the shell adapter and return stdout/stderr observations.
Execute an authorized semantic action and return its state delta. Stale revisions and capability violations are errors.
List semantic actions with side-effect, idempotency, capability, and parameter metadata.
Run a read-only world audit covering descriptor integrity, event-chain verification, state, capabilities, branches, experiences, and action catalog health.
Compile a bounded decision frame for an objective.
Read the formal world descriptor, views, actions, branch modes, and capabilities.
Inspect the AWR world descriptor and current main state.
Read canonical world state for main or an isolated branch.
No error handling guidance. Tools like awr_world_action accept 'base_revision' and reject stale revisions, but no error response documents what to do on conflict (retry? fetch new revision? ask user?). Baseline: A+ tools include recovery guidance.
Enum constraints not enforced in descriptions. awr_world_action 'mode' param has enum ['execute','simulate','shadow','replay'] but no description explaining when to use each. awr_experience_promote 'lifecycle' enum lacks guidance on state transitions.