fix(ai): buffer streaming preamble text before code fence appears

[temrjan]

Summary

• Buffer incoming AI stream chunks in CellCreationStream until a code fence (```) appears, preventing conversational preamble from becoming a separate Python cell
• Flush buffered content as a cell in stop() if the stream ends without any fence (backward compatibility for models that return code without fences)
• Add tests for: preamble + fence, code without fence, fence from first chunk

Root cause

CellCreationStream.stream() called codeToCells(buffer) on every chunk. Before any fence arrived, codeToCells treated the entire buffer as Python code (line 166 of completion-utils.ts: if (!code.includes("```")) return [{ language: "python", code }]), creating a cell from preamble text like "I'll create a fibonacci function...".

What this does NOT fix

The "Inline AI edit" flow (problem #2 in the issue) uses a different code path through the backend without_wrapping_backticks function. That is a separate issue.

Test plan

• All 25 staged-cells tests pass
• TypeScript clean (no new errors)
• Test: preamble text buffered, cell created only when fence arrives
• Test: code without fence → cell created on stream end (backward compat)
• Test: fence in first chunk → cell created immediately (no delay)

Fixes #8880

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
marimo-docs	Ready	Preview, Comment	Apr 8, 2026 5:27am

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[temrjan]

I have read the CLA Document and I hereby sign the CLA

[manzt]

Nice. I think buffering until a fence appears makes sense, and the stop() fallback for fence-less responses is a good backward-compat touch.

[manzt]

This test covered the case where the first chunk is a partial fence, which got replaced. Would we add that scenario back?

[temrjan]

Added! The new test (should buffer partial fence and create cell when fence completes) covers the split-fence scenario: first chunk has just ``````, second chunk completes the fence — no premature cell creation.

[Copilot]

Pull request overview

This PR adjusts the frontend AI “staged cells” streaming logic to avoid turning conversational preamble into its own Python cell by delaying parsing until a code fence is observed (or until the stream ends).

Changes:

• Buffer streamed text in CellCreationStream.stream() until a triple-backtick code fence appears (unless cells were already created).
• On stream end (stop()), flush buffered content into a cell if no cells were created (maintains behavior for models that emit code without fences).
• Add/adjust staged-cells streaming tests for preamble buffering, no-fence flushing, fences in first chunk, and partial-fence completion.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
frontend/src/core/ai/staged-cells.ts	Buffers streamed chunks until a code fence appears; flushes buffered content on stop when no cells were created.
frontend/src/core/ai/tests/staged-cells.test.ts	Adds test coverage for buffering behavior and stop-time flushing across multiple stream chunk patterns.

[Copilot]

stop() says it clears all state, but it only resets buffer. createdCells (and hasMarimoImport) remain populated, which can leak state if any additional text-delta chunks arrive after text-end/finish, and also makes the comment inaccurate. Consider resetting createdCells/hasMarimoImport here (and/or nulling the stream ref after stop) so a completed stream cannot update previous cells.

Suggested change

	this.buffer = "";
	this.buffer = "";
	this.createdCells = [];
	this.hasMarimoImport = false;

[temrjan]

Good catch — applied in 2d784fc. CellCreationStream is single-use (new instance per text-start), so this is purely defensive, but it makes the comment accurate.