24 tools with explicit schemas and descriptions, but significant quality gaps. All tools have descriptions (10-200 chars, mostly adequate), and input schemas are visible with type declarations. However, parameter descriptions are sparse or generic, output schemas are undocumented, and error handling guidance is absent. Many tools lack actionable context for LLM selection. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Composition is reasonable (single responsibility per tool), but chaining support is weak, responses don't consistently include IDs needed for downstream calls. Security concerns: owner_ack and owner_code patterns suggest gating, but no explicit permission checks or audit trail documentation visible.
Promote proposed (or provided) content to canonical NEXT_STEP.md — owner-gated. Default gate (nonce): call once WITHOUT owner_code; a one-time code is delivered to the owner (never to this client). Ask the owner for it and call again with owner_code=<code>. The code is bound to the exact content and is single-use. Legacy gate (ack_legacy): owner_ack=true + reason.
Grant a time-bounded session lease tied to mission ID and exact contract version.
Bind a device-local directory to a logical project_id mount table entry. Requires owner_ack=true + reason.
Deterministically classify planned or observed activity against the mission contract.
Snapshot ACTIVE_CONTEXT and optional project artifacts under HAI_HOME/history/checkpoints.
Complete with verified per-criterion evidence, or abandon (owner-gated). Abandon: first call delivers a one-time code to the owner; call again with owner_code. Legacy gate: owner_ack=true.
Output schemas undocumented. All 24 tools return JSON via _json() helper, but no schema documentation visible for what fields clients should expect. LLMs cannot plan downstream calls or extract data reliably.
Parameter descriptions sparse or generic. Many parameters (e.g., 'done_criteria', 'constraints', 'evidence', 'changes') lack actionable descriptions of expected structure, format, or constraints. LLMs cannot construct valid payloads without trial-and-error.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Distill intake items into structured decisions and next steps.
Check for drift in activity against mission contract constraints.
Return the exact canonical mission contract for a valid session lease (not a summary).
Read the canonical NEXT_STEP.md for a project (or report none).
Check HAI-MCP server health and optional project path usability. No model calls.
Intake a raw idea or request into the HAI system for processing.
Start a new mission with problem statement, artifact, done criteria, and optional constraints.
Open a bounded mission with a versioned canonical contract. One active mission globally.
Park a thought/meta item in HAI inbox without changing ACTIVE lanes.
Park a mission-linked idea with full context. Grants no execution right; contract unchanged.
Verify evidence against done criteria for mission completion.
Write NEXT_STEP.proposed.md (not canonical). Promote via hai_accept_next_step.
Read-only summary of HAI Projek-Managment run-contract artifacts.
Apply a visible field-level contract diff; revokes all leases. Owner-gated. First call (no owner_code) returns the diff and delivers a one-time code to the owner; call again with owner_code=<code from the owner>. Legacy gate: owner_ack=true.
Return the smallest recovery next action from latest or named checkpoint (read-only advice).
Set focus and register an ACTIVE lane (max 2). Fails if a third ACTIVE is requested.
Return ACTIVE lanes, focus, inbox count, and optional project next-step flags.
Stop the current day's work and optionally close the active mission.
No tool annotations. None of the 24 tools declare readOnlyHint, destructiveHint, or idempotentHint. Agents cannot distinguish safe reads from irreversible writes (e.g., hai_close_mission marked IRREVERSIBLE in metadata but not in schema).
Error handling absent. No visible error recovery guidance, categorization (retryable vs fatal), or actionable error messages. Tools like hai_accept_next_step and hai_recontract have complex gating logic but no documented error paths.
Chaining support weak. Response schemas not documented, so LLMs cannot reliably extract IDs for downstream calls. E.g., hai_open_mission likely returns mission_id, but this is not documented, forcing agents to guess or make extra discovery calls.