fix: ChainPlanner._parse null-guard + TaskRecorder.input default_factory

[Ricardo-M-L]

Summary

Fixes #261. Two small defensive-coding fixes in utu/agents/:

1. ChainPlanner._parse (utu/agents/orchestrator/chain.py) — guard the <plan> regex match so a malformed LLM output surfaces a ValueError with a clear message instead of AttributeError: 'NoneType' object has no attribute 'group'. The preceding <analysis> match on line 82 already uses the same defensive style (if match else "").

2. TaskRecorder.input (utu/agents/common.py) — use default_factory=list so the runtime default matches the declared type str | list[TResponseInputItem]. Other list-typed fields on the same dataclass (trajectories, raw_run_results) already use default_factory=list; input was the outlier with default_factory=dict.

Changes

 # utu/agents/orchestrator/chain.py
 match = re.search(r"<plan>\s*\[(.*?)\]\s*</plan>", text, re.DOTALL)
+if match is None:
+    raise ValueError("Failed to parse plan: <plan>[...]</plan> block not found in LLM output")
 plan_content = match.group(1).strip()

 # utu/agents/common.py
-input: str | list[TResponseInputItem] = field(default_factory=dict)
+input: str | list[TResponseInputItem] = field(default_factory=list)

Net: +3/-1 lines across two files, split into two commits so each fix is easy to bisect or cherry-pick.

Test plan

• pre-commit run (ruff check + ruff format) passes locally
• No new tests added — each fix is a one-line behavioural tightening whose effect is observable by reading the surrounding code; happy to add targeted unit tests if you'd prefer.

🤖 Generated with Claude Code