Greptile Summary

This PR fixes three interrelated bugs in the MCP approval-resuming flow: org ID partition mismatches caused resume to look in the wrong database partition; the tool loop continued after parking (sending a continuation to the model), producing a stray trailing assistant message that poisoned the resume transcript; and replayed mcp_approval_request history items were being forwarded to the upstream model, which couldn't consume them.

• Org resolution: Extracts resolve_request_org() as a shared helper and calls it on both the park path (PipelinePrincipal::from_auth) and the resume route, ensuring anonymous/no-auth requests resolve identically in both directions and land in the same DB partition.
• Loop stop signal: Adds stop_loop: bool to ToolCallResult; the MCP approval gate sets it true when parking, and the runner breaks before issuing a continuation request. All other tools explicitly set stop_loop: false.
• History cleanup: rewrite_mcp_history now filters out mcp_approval_request items before reconstructing history for the upstream provider, since resume has already folded the approved call into a function_call/function_call_output pair.

Confidence Score: 4/5

The changes are targeted correctness fixes for the MCP approval flow with comprehensive test coverage; no regressions expected in unrelated paths.

All three bugs are fixed at their root: the org-partition mismatch is resolved by sharing a single resolver; the trailing assistant-message problem is fixed by the stop_loop signal; and stale mcp_approval_request history items are now dropped before upstream forwarding. Each fix is covered by a dedicated unit or integration test.

No files require special attention; the core logic changes are in responses_pipeline.rs, chat.rs, runner.rs, and executor.rs, all of which read clearly and are well-tested.

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Client
    participant Route as chat.rs
    participant Runner as ToolLoopRunner
    participant Executor as McpExecutor
    participant DB as mcp_pending_approvals

    Note over Client,DB: Park turn
    Client->>Route: POST /v1/responses
    Route->>Runner: wrap_streaming
    Runner->>Executor: execute(mcp_call)
    Executor->>DB: "insert(org_id=resolve_request_org())"
    Executor-->>Runner: "ToolCallResult stop_loop=true"
    Runner-->>Client: mcp_approval_request + response.completed

    Note over Client,DB: Resume turn
    Client->>Route: POST /v1/responses + mcp_approval_response
    Route->>Route: resolve_request_org() same helper
    Route->>DB: take_by_id_and_org(approval_id, org_id)
    DB-->>Route: parked call row
    Route->>Route: resume_mcp_approvals rewrites input
    Route->>Runner: wrap_streaming rewritten input
    Runner-->>Client: real tool result

_{Reviews (1): Last reviewed commit: "Fix MCP approval resuming" | Re-trigger Greptile}