Fix path injection, empty choices crash, and config inconsistencies

[Copilot]

Security and robustness fixes found during review of the agent loop implementation in PR #4.

Path injection in collect_artifacts

Artifact filenames from container ls output were interpolated into shell commands with only single-quote wrapping—vulnerable to traversal and injection. Now uses os.path.basename() + shlex.quote():

# Before
content = await self._container.exec(f"cat '{_ARTIFACTS_DIR}/{name}'", timeout=10)

# After
name = os.path.basename(name)
safe_path = shlex.quote(f"{_ARTIFACTS_DIR}/{name}")
content = await self._container.exec(f"cat {safe_path}", timeout=10)

Missing bounds check on response.choices[0]

• LLMClient.chat(), AgentLoop.run(), AgentLoop._force_final(), and Level3Agent all accessed choices[0] without guarding against empty arrays → IndexError on malformed API responses
• Added early-return/fallback at each call site

Unsafe full_name.split("/", 1) in handlers

• Both PullRequestHandler and IssueCommentHandler would raise ValueError on malformed webhook payloads missing the / separator
• Added validation with early return

Hardcoded exec timeout cap

• ContainerManager.exec() capped timeout at 120s, silently overriding the configurable container_timeout (default 150s)
• Changed to use self._settings.container_timeout as the cap

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.