An AI-powered Android application with conversation management, extension management, and memory tools for assistants
Miffan provides 6 tools with reasonable descriptions (avg ~180 chars) and clear naming conventions. All tools follow verb_noun patterns (recent_chats, conversation_search, extensions_catalog, extensions_preview_changes, extensions_apply_changes, memory_tool). However, input schemas are visible but lack depth: parameter descriptions are present but terse, and output schemas are not documented. The memory_tool has good guidance text but mixes operational instructions with parameter docs. Error handling is implicit rather than explicit. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in metadata.
Full-text search across the user's past conversations to recall specific information they mentioned before. Use focused keywords. Run multiple searches with different keywords if needed. Each result includes the conversation title, a snippet with matched keywords wrapped in [brackets], and the date.
Apply a valid server-issued preview by its opaque previewId. The canonical preview is revalidated against current settings before one atomic settings update. This operation always requires explicit user approval. Never invent or alter a previewId.
Return a credential-redacted catalog of assistants and extension resources that can be managed. Prompt bodies and MCP connection fields are omitted. Inspect this catalog before proposing changes so you use stable resource identifiers.
Validate and normalize proposed extension configuration changes without writing anything. Always call extensions_catalog first. Show the returned summaries to the user before asking to apply a valid preview.
The memory tool stores long-term information across conversations. Use `action` to control the operation: `create` (add), `edit` (update), `delete` (remove). - No relevant record: `create` + `content` - Existing relevant record: `edit` + `id` + `content` - Outdated/irrelevant record: `delete` + `id` Memories will automatically appear in the <memories> tag in later conversations. Do not store sensitive information (e.g., ethnicity, religion, sexual orientation, political views, sex life, criminal records). You may store: preferred name, preferences, plans, work-related notes, chat style preferences, first chat time, etc. Do not show memory content directly in the conversation unless the user explicitly asks.
Output schemas not documented. Tools return results but LLMs cannot plan downstream calls without knowing field names and types. E.g., recent_chats returns 'conversation titles and dates' but exact JSON structure is opaque.
extensions_preview_changes accepts 'changes' array but parameter description lacks detail on operation fields, structure, and valid values. LLM must infer schema from context.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk classifications in metadata. extensions_apply_changes is WRITE but lacks destructiveHint annotation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
List the user's recent conversations with you to understand their preferences and ongoing topics. Returns conversation titles and the date of last activity, ordered by pinned first then most recently updated. Use this when you need quick context about what the user has been discussing lately. Only titles and dates are returned; use `conversation_search` to look up the actual content.
memory_tool description mixes operational guidance ('Do not store sensitive information...') with parameter docs. Should separate user-facing guidance from schema documentation.
Error handling not documented. Tools provide no guidance on retryability, user-fixable errors, or recovery steps. E.g., conversation_search with no results should suggest alternative queries.