🤖 fix: handle lost OpenAI previousResponseIds on frontend retry

[ammar-agent]

Problem

OpenAI sometimes invalidates stored previousResponseId values that were working moments ago. When this happens mid-stream, the frontend automatically retries, but the retry would include the same invalid ID and fail again, creating an unrecoverable error loop.

Solution

Track invalidated response IDs in memory when OpenAI rejects them with previous_response_not_found. When buildProviderOptions constructs the next request (during frontend retry), filter out any IDs that have been marked as lost.

Also fixed ripgrep stdin detection: bash tool now redirects stdin from /dev/null so CLI tools don't treat it as a pipe.

Changes

• StreamManager: Added lostResponseIds Set to track invalidated OpenAI response IDs
• StreamManager.recordLostResponseIdIfApplicable(): Detect previous_response_not_found errors during streaming and extract the rejected ID from the error message
• StreamManager.isResponseIdLost(): Public API for checking if an ID is invalid
• buildProviderOptions(): Filter out lost IDs when extracting previousResponseId from message history
• bash.ts: Prepend exec </dev/null to all commands to close stdin (fixes ripgrep treating stdin as a pipe)

Testing

• Unit tests for lost ID tracking and extraction logic
• Existing bash tool tests verify stdin behavior
• Manually verified fix resolves the unrecoverable error

Generated with mux