fix: prevent infinite retry loop from orphaned tool_use blocks

[ryankass-cb]

Summary

Fixes an unrecoverable session crash caused by orphaned tool_use blocks triggering an infinite retry loop against the Anthropic API.

Error: tool_use ids were found without tool_result blocks immediately after: toolu_vrtx_...

Root Cause

Three bugs combine to create the infinite loop:

1. Retry loop skips orphan cleanup (processor.ts): The continue statement in the retry path jumps back to while(true), bypassing the orphan cleanup code (lines 393-409) that marks pending/running tools as errors. Orphaned tool_use blocks persist in message history.

2. Stale messages on retry (processor.ts): streamInput.messages is built once before the retry loop and never refreshed from the database. Even if orphan cleanup ran, retries send the same broken messages.

3. invalid_request_error incorrectly classified as retryable (retry.ts): The catch-all return JSON.stringify(json) makes ALL JSON error bodies retryable, including invalid_request_error — a structural issue that fails identically on every attempt.

Fixes

File	Change
processor.ts	Move orphan cleanup before the retry continue; rebuild messages from DB on retry via new rebuildMessages callback
llm.ts	Add optional rebuildMessages callback to StreamInput interface
prompt.ts	Wire up rebuildMessages to re-read messages from DB and convert via toModelMessages()
retry.ts	Mark invalid_request_error as non-retryable; replace catch-all return JSON.stringify(json) with return undefined
transform.ts	Add defensive repairOrphanedToolUse() validation as last-resort before API calls — injects synthetic error tool_result for any orphaned tool_use blocks

Defense in Depth

The fix operates at three layers:

1. Prevention: Orphan cleanup runs before retry, messages rebuilt from clean DB state
2. Circuit breaker: invalid_request_error stops the retry loop immediately
3. Last resort: repairOrphanedToolUse() in the transform layer catches any orphans that slip through other defenses

Reproduction

1. Start a session with tools enabled
2. Trigger a tool call, then interrupt/abort mid-execution (timeout, network error, etc.)
3. The session enters an infinite retry loop showing the invalid_request_error repeatedly
4. The session becomes unrecoverable — no way to continue without starting fresh

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search, I found a potentially related PR:

PR #8497: "fix: handle dangling tool_use blocks for LiteLLM proxy compatibility"

• fix: handle dangling tool_use blocks for LiteLLM proxy compatibility #8497
• Why related: Both PRs address issues with orphaned/dangling tool_use blocks and their impact on message validation and API retries. PR fix: handle dangling tool_use blocks for LiteLLM proxy compatibility #8497 handles a similar problem of tool_use blocks without proper tool_result pairs.

PR #14456: "fix(core): repair interleaved text/tool-call parts in assistant messages"

• fix(core): repair interleaved text/tool-call parts in assistant messages #14456
• Why related: This PR also deals with structural issues in message formatting related to tool calls, which could be tangentially related to the orphaned tool_use block problem.

However, PR #15152 (the current PR) appears to be the primary/newest fix addressing the specific infinite retry loop issue with orphaned tool_use blocks, while the older PRs address related but distinct aspects of tool_use handling.

[github-actions]

This pull request has been automatically closed because it was not updated to meet our contributing guidelines within the 2-hour window.

Feel free to open a new pull request that follows our guidelines.