fix(claude-sessions): unwrap NBI preamble when joined with user prompt

[pjdoland]

Summary

Closes #329. When a Claude Code session has only a single user turn, the chat-history picker showed no preview / summary for that session — only the id and timestamp. The cause was a skip filter that treated the joined preamble + prompt string as if it were only the preamble.

Solution

claude.py joins consecutive user-role chat-history entries with \n before handing them to the Claude Agent SDK. extension.py emits the NBI context preamble ("Additional context: Current directory open in Jupyter is: ...") as its own user-role entry, so a single-turn session records ONE combined user message whose first line is the preamble and whose remainder is the user's actual prompt.

The previous skip filter at notebook_intelligence/claude_sessions.py checked stripped.startswith(_SKIPPABLE_PREFIXES) against the full text, matched the preamble, and dropped the whole message. The session's only user message was thereby skipped, leaving the picker preview blank.

The load-bearing change is a new _strip_nbi_context_preamble helper that drops the preamble line (everything up to and including the first newline) when present. It plugs into the two consumers:

• _is_skippable_text unwraps before applying the rest of the skip rules so the message is judged by the real-prompt content. The unwrap recurses at most once because NBI_CONTEXT_PREFIX is single-line: the unwrapped tail can never itself lead with the preamble, and the recursive call falls through to the legacy prefix / control-command checks (or bottoms out on the empty-stripped guard).
• _extract_preview unwraps before whitespace-collapsing so the preview reads as the user's prompt rather than the boilerplate.

The helper deliberately doesn't trim or normalize the unwrapped tail beyond stripping the preamble line: both callers normalize whitespace themselves (_is_skippable_text via .strip(), _extract_preview via " ".join(text.split())), and trailing whitespace in the result belongs to the user's prompt either way.

Testing

Five new tests in tests/test_claude_sessions.py pin the contract:

• test_unwraps_joined_preamble_and_prompt_in_one_message — the exact bug from the issue body (preamble + \n + prompt as a single string content).
• test_unwraps_joined_preamble_in_structured_content — same shape, Anthropic structured-content form with the joined string inside one text block.
• test_joined_preamble_with_skippable_prompt_still_skips — preamble joined with a control slash command (/exit); the combined message must remain skippable so the picker keeps scanning for a real prompt.
• test_joined_preamble_alone_without_newline_still_skips — preamble recorded with no trailing newline; helper returns empty string and the recursive call bottoms out on the empty-stripped guard.
• test_joined_preamble_with_empty_post_newline_still_skips — preamble followed by whitespace-only tail; same end state as preamble-alone.

The first two are verified to fail against the pre-fix code; the latter three document downstream behavior the picker has always honored.

All four gates green: pytest tests/ --ignore=tests/test_claude_client.py (1055 passed, 66 in test_claude_sessions.py), jlpm tsc --noEmit, jlpm lint:check, jlpm jest (254 passed).

Risks / follow-ups

• A user prompt that happens to start with the literal phrase "Additional context: Current directory open in Jupyter is:" would have its first line stripped. The pre-existing _SKIPPABLE_PREFIXES tuple already had the same trade-off (called out in the existing comment at lines 47-49); the helper inherits it.
• Multi-block structured content where one block is the preamble and another block is the prompt isn't observed today (the SDK joins at the string level, not at the block level). If that shape ever lands, _is_skippable_user_message would need to mirror _extract_preview's block-join behavior. Filed mentally as a follow-up.
• The NBI_CONTEXT_PREFIX constant remains in _SKIPPABLE_PREFIXES as a backstop. The new early branch consumes the preamble first, but the tuple entry still handles edge cases where the preamble somehow reaches the fallthrough.

Issue linkage

Closes #329.

[mbektas]

thanks!