Capability status
| Capability | Status |
|---|---|
| Multi-turn Session | Implemented |
| Anthropic Messages model/tool loop | Implemented, protocol tests included |
| OpenAI Responses adapter | Planned |
| Typed events | Implemented |
| Tool timeout | Implemented |
| Context compaction | Implemented, approximate token counting |
| Bash and file tools | Implemented |
| WebFetch | Implemented, basic content extraction |
| WebSearch | Placeholder |
| Asynchronous sub-agent | Experimental |
| Tool approval | Incomplete; rejected by the one-shot CLI |
| MCP | Not implemented |
| Skills | Not implemented |
| Durable memory | Not implemented |
| Prompt-cache accounting | Not implemented |
| OTel observability | Deferred until lifecycle semantics stabilize |
| Harness abstraction | Deliberately deferred until two integrations exist |
The priority is runtime correctness and clear semantics. Capability count is not used as a substitute for compatibility testing.