Durable coordination runtime for heterogeneous AI agent sessions. Provides provider-neutral task leasing, checkpoints, crash recovery, and multi-agent coordination.
RelayMesh defines 11 tools with complete JSON schemas and descriptions. All tools have clear verb-noun naming (relay_*) and detailed parameter documentation. Schemas use proper type definitions with constraints (UUID format, integer minimums, string length). However, output schemas are not documented, LLMs cannot predict return structure. Error handling guidance is absent; tools do not explain recovery paths or categorize failures. Some descriptions are verbose (>200 chars) and could be more concise for LLM parsing. Tool composition is strong: each tool has a single responsibility and output IDs chain to downstream tools (taskId, leaseId, fencingToken). Security-sensitive parameters (fencingToken) are properly typed but lack explicit permission documentation.
Mark one durable inbox message as processed.
Persist provider-neutral recovery state after meaningful progress and before risky or external work.
Lease the highest-priority ready task matching this session's role and capabilities. Returns the latest recovery checkpoint when one exists.
Commit a task result using the active lease and fencing token. RelayMesh rejects late writes from crashed or superseded sessions.
Report a blocker or failure. Retryable work returns to the queue while attempts remain; terminal failure may fail the mission.
Checkpoint progress, release the active lease, preserve the attempt budget, retarget the task, and send a durable handoff message as one atomic operation.
Output schemas not documented. LLMs cannot predict return structure (fields, types, pagination). Downstream tool calls risk field-mapping errors.
No error handling guidance. Tools do not explain recovery paths, categorize failures as retryable/fatal, or suggest next steps on failure.
Descriptions exceed 200 chars for several tools (relay_sync, relay_handoff, relay_send_message). Verbose descriptions waste tokens and bury key details. Baseline: 194 chars average, p90=392.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2025-06-18+ | v2 |
Read replayable messages addressed to this session, role, or mission.
Publish immutable base64 content for other model sessions and receive its SHA-256 identity.
Send a typed, replayable message to one session, a role, or every agent in this mission.
Renews the current agent session and returns its durable mission identity and unread-message count. Use relay_sync at startup.
Always call first and after reconnecting. Atomically renews this session, returns peers and inbox messages, and resumes or claims exactly one compatible task as a provider-neutral context envelope.
No permission/scope declarations. Tools do not state required permissions (e.g., 'write:task', 'read:inbox'). Limits audit trail clarity and least-privilege configuration.
Destructive operations (relay_fail_task, relay_complete_task) lack confirmation or dry-run support. Agents cannot preview consequences before committing irreversible state changes.