add training datasets: memory, persona, agent tools, serpent data

data/serpent_git.jsonl

@@ -0,0 +1 @@
+{"messages": [{"role": "system", "content": "You are a Python coding agent. Execute implementation tasks precisely and completely.\n\n## Constraints\n\n- Type hints on ALL new function signatures (params + return)\n- `pathlib.Path` over `os.path` for filesystem operations\n- `uv` for dependency management if `pyproject.toml` present — never raw `pip install`\n- Pydantic v2 for data models (no v1 `BaseModel` patterns: `.dict()` → `.model_dump()`)\n- `pytest` + `pytest-asyncio` for testing — no `unittest`, no `nose`\n- `ruff` for lint + format — respect existing `ruff.toml` / `pyproject.toml [tool.ruff]`\n\n## Workflow\n\n1. Read `pyproject.toml` (or `setup.py`, `requirements.txt`) to understand deps and tooling\n2. Check for `ruff.toml`, `.flake8`, `mypy.ini` — respect existing lint config\n3. Match existing style — run `git log -5 --oneline` for commit convention\n4. Implement with focused, minimal diffs\n5. Run `ruff check .` — fix violations\n6. Run `ruff format .` — apply formatting\n7. Run `pytest` (or project-specific test command) — fix failures\n8. If `mypy` is configured: run `mypy --strict <changed_paths>` — fix type errors\n\n## Verification\n\nAfter every change cycle:\n```\nruff check .\nruff format --check .\npytest                    # (or project-specific test command)\n```\n\nIf mypy is in the project's dev dependencies, also:\n```\nmypy --strict <changed_files>\n```\n\nAll must pass before reporting success.\n\n## Framework Detection\n\nCheck project structure:\n- `app/` + `manage.py` → Django\n- `app.py` or `main.py` with `Flask` import → Flask\n- `main.py` with `FastAPI` → FastAPI (use `uvicorn` to test-run)\n- `noxfile.py` → use `nox` for test orchestration\n- `tox.ini` → use `tox` for test orchestration\n\nFollow the framework's conventions for route definitions, middleware, configuration.\n\n## Commit Style\n\n- Stage only changed files: `git add <paths>`\n- Match repo's format (check `git log -5 --oneline`)\n- Common format: `<type>(<scope>): <subject>`\n\n## Memory Trail\n\nAfter completing work, store results:\n- What you built, file paths, key decisions → subject: `project.<name>.build-result`\n- Anything unresolved or worked around → subject: `project.<name>.unresolved`\n- Things you noticed out of scope → subject: `project.<name>.out-of-scope`\n- Bugs found and how you fixed them → subject: `project.<name>.errors`"}, {"role": "user", "content": "fix: parse tool_call arguments from JSON string to dict for Qwen3.5 template"}, {"role": "assistant", "content": null, "tool_calls": [{"type": "function", "function": {"name": "edit", "arguments": {"filePath": "train_qwen35_27b.py", "oldString": "        ex[\"messages\"], tokenize=False, add_generation_prompt=False", "newString": "def fix_tool_calls(messages):\n    \"\"\"Parse tool_call arguments from JSON strings to dicts for Qwen3.5 template.\"\"\"\n    import json as _json\n    fixed = []\n    for msg in messages:\n        msg = dict(msg)\n        if msg.get(\"tool_calls\"):\n            new_tcs = []\n            for tc in msg[\"tool_calls\"]:\n                tc = dict(tc)\n                if \"function\" in tc:\n                    fn = dict(tc[\"function\"])\n                    if isinstance(fn.get(\"arguments\"), str):\n                        try:\n                            fn[\"arguments\"] = _json.loads(fn[\"arguments\"])\n                        except (ValueError, TypeError):\n                            fn[\"arguments\"] = {\"raw\": fn[\"arguments\"]}\n                    tc[\"function\"] = fn\n                new_tcs.append(tc)\n            msg[\"tool_calls\"] = new_tcs\n        fixed.append(msg)\n    return fixed\n\n    messages = fix_tool_calls(ex[\"messages\"])\n        messages, tokenize=False, add_generation_prompt=False"}}}]}, {"role": "tool", "content": "Edit applied successfully."}, {"role": "assistant", "content": "Applied changes to train_qwen35_27b.py."}], "metadata": {"sha": "8137d278db4add8f3a1118ebda5ee4186d029e38"}}