Chained on top of #342. Quick fixes from the review of that PR.

What this changes

When the agent inserts a code block via addAndExecuteCodeBlock, the runtime tracks the outcome of that kernel execution and surfaces it through onBlockDone. Previously:

• A kernel.execute resolution with success: false (e.g. SyntaxError) was reported via onBlockDone with success: true — the success flag was dropped between blockOutputs and the post-loop.
• A kernel.execute rejection was never reported via onBlockDone at all — the catch path returned the error string to the agent but never pushed to blockOutputs, so the post-loop skipped it. The block was already inserted into notebook.blocks, leaving consumers with an inserted-but-never-completed block.
• When executeAgentBlock itself threw partway through, the post-loop never ran, so even successfully-completed added blocks weren't reported.

The fix

• Track success: boolean on each blockOutputs entry.
• Push an entry on kernel.execute rejection (with a synthesized error output via the existing createErrorOutput helper) so the failed block still gets onBlockDone.
• Wrap the executeAgentBlock call in try/finally so added blocks are drained even when the agent throws.
• Pass bo.success through to onBlockDone instead of hardcoded true.
• Drop the redundant collectedOutputs.set in the post-loop (the per-block set inside addAndExecuteCodeBlock already covers all paths now).
• Type blockOutputs.outputs as IOutput[] to drop the two as IOutput[] casts.
• Tighten the add_code_block tool description (it claimed "stdout, stderr, and execution results" but actually returns a single formatted string).

Tests

• Augmented surfaces addAndExecuteCodeBlock failures when kernel execution rejects to assert the failed code block now gets onBlockDone with success: false and a synthesized error output.
• Augmented returns failure when kernel execution fails to assert success: false propagates for the resolved-failure path too.

All 204 tests in runtime-core pass; typecheck and biome clean.

Out of scope (deliberately)

From the review:

• onLog regression (docs: init readmes with spellcheck and formatting needed for them #1) — context.onLog was never set by any caller, so the lost log lines were already no-ops.
• serializeNotebookContext export ambiguity (chore: setup spell-check and root pnpm config #2) — JSDoc nudge worth doing but a separate concern.
• Direct test coverage for serializeNotebookContextFromBlocks (Enable Renovate for automated dependency pinning and updates #7) — covered transitively; not blocking.
• Verbose helper name (Decide on linting configuration: ESLint, Biome, or blend #8) — public-API rename, worth its own discussion.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes

  • Enhanced error handling and reporting for code blocks executed by agents. Callbacks are now reliably emitted even when execution fails, and each block's success/failure status is accurately tracked and reported.

• Tests

  • Added tests to verify proper error reporting and callback emission when agent-executed code blocks encounter execution failures.