Intent-firing system with MCP server exposing tools to promote thoughts into intents, list/snooze/complete intents, and force-fire for testing. Runs on 127.0.0.1:9085 by default.
Kairos has well-structured tool definitions with clear naming (verb_noun pattern) and complete input schemas. All 5 tools are explicitly registered with gomcp.Tool and include parameter descriptions. However, descriptions are terse (avg ~60 chars), lack context on WHEN to use each tool, and omit guidance on error recovery. Output schemas are not documented, handlers return text summaries without specifying field structure. Error handling returns actionable messages but lacks categorization (retryable vs fatal). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles (3 WRITE, 2 READ_ONLY). Parameter validation is present but descriptions don't state constraints (e.g., brain_cost enum values, RFC3339 format for 'until'). Overall: solid foundation, but descriptions and output documentation need expansion to meet production LLM-optimization standards.
Mark an intent as done.
Bypass rules and fire an intent immediately. For testing only.
List intents, optionally filtered by status.
Promote a Wooj-brain thought into a Kairos intent. Provide the thought_id and fill in the required fields for the intent.
Snooze an intent until a given time.
Output schemas not documented. Handlers return text summaries (e.g., handleListIntents returns JSON-formatted text) but LLMs cannot parse structure without explicit schema documentation. Agents cannot reliably extract intent IDs or chain to downstream tools.
Tool descriptions lack WHEN-to-use context and error recovery guidance. E.g., 'Snooze an intent until a given time' does not explain when snoozing is preferred over completing, or what happens if the timestamp is in the past. Descriptions average 55 chars; production baseline is 194 chars.
Parameter descriptions omit format/constraint details. 'brain_cost' accepts 'low, med, or high' but description does not state this as an enum. 'until' requires RFC3339 format but constraint is buried in the description text, not formalized. LLMs cannot reliably validate without explicit constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). kairos_list_intents and kairos_promote_thought have clear risk profiles (READ vs WRITE) but are not annotated. Agents cannot infer safety without explicit hints.
Error handling returns actionable messages but lacks categorization. errResult() returns text without indicating if the error is retryable, user-fixable, or fatal. Agents cannot determine next steps (retry, ask user, abort).