Mobile-native runtime kernel and adapter-facing API for the Napaxi agent SDK. A Rust-based agent engine with skill evolution, memory management, and tool execution capabilities for mobile platforms (Android/iOS).
Napaxi exposes 18 tools with complete JSON schemas and basic descriptions, but falls short of production quality. Tool names lack consistent verb-noun structure (e.g., 'execute_linux_program_json' vs 'send_to_group_member'). Descriptions are present but generic, most under 100 chars, lacking WHEN/WHY context or recovery guidance. Parameter descriptions exist but are minimal (5-20 chars). No output schemas documented. Error handling is absent, no guidance on retryability, user-fixable vs fatal errors, or recovery paths. High-risk tools (execute_linux_program_json, tool_broker_call_tool_json_handle) lack confirmation patterns or dry-run support. Tool composition is poor: 'answer_human_request' and 'answer_human_request_typed' duplicate functionality; 'platform_tool_descriptors_json' and 'platform_tool_descriptors' are redundant variants. No pagination, no result limits, no chaining IDs in responses.
Inject a host response for a pending human-in-the-loop request; returns whether a matching pending request was found.
Typed counterpart: Err(ToolError::NotFound) when the request id is not pending (either never registered or already answered), Ok(()) on successful injection.
Typed counterpart: returns the descriptors directly without a JSON round-trip.
Built-in browser tool descriptors as a JSON array.
Close an Android Linux PTY session, returning JSON status.
Drain queued Android Linux PTY events for a session, returning JSON.
Duplicate tools with identical functionality: 'answer_human_request' and 'answer_human_request_typed'; 'platform_tool_descriptors_json' and 'platform_tool_descriptors'; 'browser_tool_descriptors_json' and 'browser_tool_descriptors'. LLMs waste reasoning cycles choosing between them.
High-risk tools (execute_linux_program_json, tool_broker_call_tool_json_handle) lack confirmation patterns, dry-run support, or error recovery guidance. Agents can execute arbitrary commands without safeguards.
Descriptions are generic and lack WHEN/WHY context. Most under 100 chars. No guidance on when to use 'send_to_group_member' vs 'tool_broker_call_tool_json_handle', or recovery paths on failure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Execute a native Linux program request encoded as JSON, returning JSON.
Whether name is a built-in browser tool.
Whether name is a registered platform tool.
List members of the current agent group.
Open an Android Linux PTY session from a JSON request, returning JSON.
Typed counterpart: returns the descriptors directly without a JSON round-trip. Adapter code that does not need to forward the JSON to Dart should prefer this.
Platform (host-provided) tool descriptors as a JSON array.
Resize an Android Linux PTY session, returning JSON status.
Send a task to a specific member of the current agent group and return that member's response.
Broker call: invoke a tool on an agent engine (JSON in, JSON out).
Broker call: list the tools an agent engine exposes (JSON in, JSON out).
Write UTF-8 input to an Android Linux PTY session, returning JSON status.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., session_id from open_linux_pty_session_json).
Inconsistent naming: '_json' suffix on some tools (execute_linux_program_json, open_linux_pty_session_json) but not others (send_to_group_member). Unclear whether '_json' variants are preferred or legacy.