Harden Claude stream interruption handling

[juliusmarminge]

Split out from #1891.

Summary

• preserve interrupted Claude turns separately from runtime stream failures
• ignore ede_diagnostic-only results and mark interrupted tool results as declined
• add regression coverage for interrupted Claude stream edge cases

Testing

• bun fmt
• bun lint
• bun typecheck
• bun run test

---

Note

Medium Risk
Changes core Claude runtime turn/stream error classification, which can affect how sessions and tool calls are marked completed/failed/interrupted. Risk is moderate due to new branching around interrupts/diagnostics and updated tool-result status mapping.

Overview
Hardens Claude stream/turn error handling so user-initiated interrupts and SDK aborts are treated as interrupted instead of generic stream failures, using tagged ClaudeStream*Error types plus an interruptRequested flag.

Filters SDK-internal [ede_diagnostic] entries so they don’t surface as runtime errors (including is_error=false error results), and marks tool results that represent rejected tool use after interruptTurn as declined rather than failed.

Adds regression tests covering ede_diagnostic-only results, interrupted tool-result flows, and stream failures/abort messages after interruption; also tightens a stopSession assertion and slightly standardizes Claude provider probe error messages.

^{Reviewed by Cursor Bugbot for commit 625da23. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Harden Claude stream interruption handling to distinguish interrupted, diagnostic, and failed states

• Introduces ClaudeStreamInterruptedError and ClaudeStreamFailedError typed errors in ClaudeAdapter.ts to classify stream failures early, enabling differentiated downstream handling.
• Adds interruptRequested flag to ClaudeSessionContext so that subsequent messages, tool results, and stream exits can be handled appropriately after interruptTurn is called.
• Filters SDK diagnostic messages (containing [ede_diagnostic]) from user-facing error paths; these now complete turns as interrupted without emitting a runtime error.
• Tool results arriving after an interrupt with rejection text are marked declined instead of failed.
• Behavioral Change: stream exits that were previously surfaced as runtime errors may now resolve as interrupted when caused by a user-requested interrupt or SDK diagnostics.

^{Macroscope summarized 625da23.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 890ca798-7c07-455d-9eac-5001938198c8

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch t3code/claude-stream-interrupts

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[macroscopeapp]

Approvability

Verdict: Approved

This PR hardens error handling in the Claude adapter to correctly distinguish between user-initiated interruptions and actual failures, avoiding false error reports and marking interrupted tools appropriately. Changes are self-contained with comprehensive tests, and the author is the file's original maintainer.

^{You can customize Macroscope's approvability policy. Learn more.}