Local voice output for coding agents, backed by a local TTS bridge
Single tool 'speak' has a clear action verb and reasonable description (127 chars, within 10-1024 range). Input schema is properly defined with Zod, including text (required) and kind (optional enum). However, the tool lacks output schema documentation, LLMs cannot predict what fields to expect from the response. Error handling returns plain text messages but does not categorize errors as retryable/user-fixable/fatal or provide recovery guidance. The 'kind' parameter is marked optional with vague metadata-only semantics, reducing clarity. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present despite being a read-only operation.
Speak a concise question, blocker, alert, or completion summary through local TTS. Do not send code, logs, paths, or long output.
Output schema not documented. Response structure is inferred from code (text content + isError flag) but not declared in tool definition. LLMs cannot plan downstream actions or extract structured data.
Error handling lacks recovery guidance. Errors return raw message strings without categorization (retryable vs fatal) or actionable next steps. E.g., 'Connection refused' tells the agent nothing about whether to retry or escalate.
'kind' parameter semantics unclear. Marked optional with description 'Metadata only in the MVP', does not explain what values mean or when to use each. LLMs cannot determine when to pass 'question' vs 'alert' vs 'custom'.
No tool annotations. Tool is read-only (does not modify state) but lacks readOnlyHint annotation. Agents cannot infer safety properties without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | 2026-07-28+ | v2 |
Text parameter description warns against code/logs/paths but does not specify length limits or format constraints. LLMs may still pass multi-paragraph text despite the guidance.