Self-hosted TypeScript agent runtime with durable approvals and verifiable run records. Own it, approve it, audit it.
Open Multi-Agent defines 10 tools across memory, release, and artifact domains. Tool naming follows verb_noun convention (engram_commit, engram_query, tdam_search_*). Descriptions are present and contextual (avg ~120 chars), explaining WHEN to use each tool. Input schemas are well-structured with enums, type constraints, and descriptions. However, output schemas are not documented, LLMs cannot predict what fields to expect from responses. Error handling guidance is absent; tools do not indicate recovery paths or retryability. No tool annotations (readOnlyHint, destructiveHint) despite clear risk classifications (WRITE vs READ_ONLY). Parameter descriptions are adequate but lack format/range constraints for numeric fields (e.g., ttl_days, limit have no min/max bounds). Composition is sound, tools are single-responsibility and chainable (e.g., engram_query → engram_resolve). Security: no credentials exposed as parameters. Overall: solid foundation with gaps in output documentation and error guidance.
Commit a verified fact to Engram shared team memory. Use this to record discoveries, decisions, or corrections that other agents should see.
List conflicts between facts in Engram shared memory. Conflicts are auto-resolved by Claude (with ANTHROPIC_API_KEY) or heuristic — this tool is for auditing resolutions, not triggering them.
Query Engram shared memory for facts about a topic. Call this before starting any task to see what the team already knows.
Override an auto-resolution for a conflict between facts. Use this when the automatic resolution was incorrect and you need to pick a different winner or merge.
Return immutable release evidence collected from the latest core tag through the current HEAD.
Read the repository release contract from .github/RELEASING.md.
Output schemas not documented. LLMs cannot predict response structure, forcing them to guess field names and types. This breaks tool chaining and increases hallucination risk.
No error handling guidance. Tools do not indicate what errors are retryable, user-fixable, or fatal. Agents cannot recover from failures intelligently.
Numeric parameters lack bounds. ttl_days, limit, confidence have no min/max constraints. LLMs may pass invalid values (negative days, limits > 1000).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2025-06-18+ | v2 |
Read one deterministic, risk-ranked, size-bounded bundle of base-to-HEAD diffs. The bot selects the paths; model-supplied repository paths are not accepted. Unselected paths remain listed in the full evidence metadata.
Render an image preview for a named artifact.
Search the raw conversation history captured in long-term memory. Use this to quote or verify exactly what was said in a previous session.
Search the team's long-term memory for distilled facts from previous sessions. Use this when you need background the current conversation does not contain — prior findings, decisions, or user preferences.
Tool annotations missing. WRITE tools (engram_commit, engram_resolve) lack destructiveHint. READ_ONLY tools lack readOnlyHint. Agents cannot infer safety properties.
render_preview and release tools have minimal descriptions (<70 chars). Insufficient context for LLM selection. render_preview description does not explain when to call it or what 'label' means.