Tool Call ID Attribution#

Claude Code and Codex session transcripts can contain repeated tool_use_id values within the same session 1. The correct pairing strategy is source-order binding: match each tool_use block to the next unconsumed tool_result with the same ID in traversal order.

Implemented Behavior: One-to-One Call/Result Pairing#

src/ingest.ts pairs tool calls and results one-to-one based on occurrence order [PR #178]:

1. During the block-insertion pass, each tool_result is appended to a queue stored in Map<tool_use_id, resultQueue[]>, preserving every result's blockId, isError, text, and timestamp.
2. In the linkage pass, each tool_use consumes the next unconsumed result from the queue for that tool_use_id.
3. A result occurrence counter (resultTaken) tracks how many results with a given ID have been consumed, ensuring each call pairs with the corresponding result in sequence.

Handling edge cases:

• Results that precede any call — If a result's blockId is lower than the call's blockId, the result is skipped and cannot answer that call.
• Results after all calls are answered — Surplus results remain unlinked; they cannot shift outcomes when the same ID is reused later.
• Result tool names — Tool names for results are resolved through the result_block_id linkage established during pairing, not from the result message itself.

The tool_call table stores call_block_id and result_block_id as foreign keys to block(id) 5, so the linked block is authoritative for all derived data.

Duplicate ID Diagnostics#

linkageIssues() in src/diagnostics.ts emits three issue codes:

Critical: these diagnostics are purely informational — they do not alter the queue-based linkage in ingest.ts. A duplicate_tool_use_id diagnostic does not mean those calls were dropped; both call blocks are stored in block and collected for the tool_call insertion pass 1. Raw blocks and diagnostic rows must be retained because pairing repeated IDs is an attribution policy, not evidence that the source IDs were unique; re-attributing historical links from corrected logic requires the original JSONL logs to remain available.

Impact#

The fix deployed in PR #178 corrects output volume, outcome, duration, and activity attribution for sessions with repeated tool-call IDs:

• Accurate output volume — output_bytes for each call now reflects its own result, not the last result with that ID.
• Correct error attribution — is_error matches the outcome of the specific call/result pair.
• Correct duration — duration_ms is computed from each call's timestamp to its own result's timestamp.
• Preserved bucket totals — Tool-call counts, generation tokens, and context-window tokens are no longer inflated by result reuse.

The ingest pipeline revision was bumped from 5 to 8, and the session economics format version from 4 to 7. Cached bucket vectors and available logs will refresh on the next sync, re-attributing historical sessions that had repeated IDs.