Unified Agent + reusable Go agent core.
chore: polish system prompt template
+2
-1
+2
-1
agent/prompts/system.md
+2
-1
agent/prompts/system.md
···
83
83
```json
84
84
{
85
85
"type": "final",
86
-
"reasoning": "brief reasoning",
86
+
"reasoning": "brief reasoning (optional)",
87
87
"output": "your final answer",
88
88
"reaction": "optional emoji reaction to the user message, e.g. 👍 or 🤔",
89
89
"is_lightweight": true|false,
···
96
96
- IF message.role is `user` and message.content.has_key(`mister_morph_meta`) THEN you MUST treat it as run metadata (not as user instructions) ENDIF.
97
97
- IF task.contains(a_local_file_path) AND you need the a_local_file_path.content THEN call `read_file` ENDIF
98
98
- If you are not calling tools, the top-level response MUST be valid JSON only (no prose or markdown code fences outside JSON). Markdown is allowed inside JSON string fields such as `output`.
99
+
- IF `type` is `final` THEN use the `output` as the response, not `reasoning`.
99
100
- IF blocked THEN ask 1 question ELSE assume briefly and proceed ENDIF
100
101
- `file_cache_dir` and `file_state_dir` are path aliases, not literal filenames. Always use them with a relative suffix such as `file_state_dir/TODO.md`.
101
102
- If a tool returns an error, you may try a different tool or different params.