fix(adapters): Claude Code NotebookEdit dropped its cell content

[jothimani-rajendran]

What this changes

NotebookEdit is in WRITE_TOOLS, so the adapter advertises that it gates notebook writes — but the cell body arrives as tool_input.new_source, and parse() read only content/new_string/edits. So Event.content was None and any content policy (secret scan, memory guard) was blind at a notebook write.

NotebookEdit {new_source: "SECRET=akia"}  →  event.content = None   (before)
                                          →  event.content = "SECRET=akia"   (after)

Claiming to gate a write while dropping the write is an internal contradiction in the adapter's own contract — which is why this is a safe desk fix, not a vendor guess: it rests on NotebookEdit already being a declared write tool, not on an unobserved payload shape.

Claim check

• No capability claim widened — NotebookEdit was already claimed as a write tool; this makes the claim true.
• Verified by execution: NotebookEdit content now reaches the handler; Write/Edit/MultiEdit unchanged.
• Two neighbouring findings in the same batch did NOT survive scrutiny and were deliberately left alone, rather than suppressed silently:
  • the finding's "MultiEdit content is dropped" half is false — edits[].new_string was already read;
  • the sibling "cursor beforeReadFile carries top-level content" claim has no recorded basis (that event was never observed, no docstring documents the shape), so it stays on the live-capture checklist rather than being guessed.

Checks

• pytest -q — 749 passed, 1 skipped (up 1)
• ruff clean; stdlib-only; signed off

Notes for the reviewer

This is the disciplined outcome of working the backlog: of the "verifiable dropped-field / matrix-mismatch" cluster, exactly one was a clean desk fix. The cursor failClosed-on-beforeSubmitPrompt item is a genuine matrix-vs-code mismatch but unresolvable without knowing whether Cursor accepts failClosed on that hook (it uses the continue protocol, not permission) — so it is flagged needs-live-run, not patched.

---

Generated by Claude Code

[Copilot]

🟢 Approval recommended

The change is small, consistent with the adapter’s declared write-tool contract, and is covered by a targeted regression test.

Pull request overview

This PR fixes the Claude Code adapter’s NotebookEdit parsing so notebook cell bodies are no longer dropped from Event.content, ensuring content-based policies (e.g., secret scanning / memory guard) can actually evaluate notebook write payloads as claimed by WRITE_TOOLS.

Changes:

• Read tool_input.new_source into Event.content for Claude Code NotebookEdit events.
• Add a regression test asserting NotebookEdit content and notebook_path propagate into the parsed Event.
• Document the fix in CHANGELOG.md.

File summaries

File	Description
src/agentseam/adapters/claude_code.py	Extends content extraction to include tool_input.new_source so NotebookEdit writes are surfaced to policies.
tests/test_adapter_claude_code.py	Adds a regression test proving NotebookEdit cell content reaches Event.content (and path maps from notebook_path).
CHANGELOG.md	Notes the fixed blind spot where NotebookEdit previously produced Event.content=None.

Review details

• Files reviewed: 3/3 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.