Fix Claude Code and Codex trace import parsing

[x0ba]

Summary

• Treat Claude/Anthropic thinking and related content block types as known blocks, and extract thinking text when present
• Parse Codex rollout event_msg user_message and agent_message entries so the visible conversation imports completely
• Correct Codex function_call_output handling before generic function call matching and preserve call IDs
• Add importer regression tests for Claude Code thinking/tool blocks and Codex rollout event/function output shapes

Sources checked

• Anthropic extended thinking docs describe thinking and redacted_thinking content blocks: https://platform.claude.com/docs/en/build-with-claude/extended-thinking
• OpenAI Codex rollout files are JSONL RolloutLine entries with session_meta, event_msg, and response_item records: Session disappear. Codex trips over bad JSON history. openai/codex#24425
• Claude Code JSONL traces include text, tool_use, tool_result, and thinking blocks: Claude Code JSONL mining silently drops all tool output (49% content loss) MemPalace/mempalace#590

Verification

• vp check --fix (passes with existing shaderElement warning)
• vp test

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
peek	Ready	Preview, Comment	Jun 8, 2026 9:16pm

[greptile-apps]

Greptile Summary

This PR fixes two importer parsers — Claude Code and Codex — to correctly handle content types that were previously unknown or misclassified. It also adds a regression test suite covering the new paths.

• Claude Code: adds a extractClaudeContent wrapper that extracts thinking block text before delegating to the shared utility; extends the allowlist and tool-event loop to cover thinking, redacted_thinking, server_tool_use, and web_search_tool_result; sharpens inputSummary/outputSummary assignment and adds toolUseId to metadata.
• Codex: introduces a pre-scan (hasEventMessages) to prefer event_msg user/agent messages over response_item/message duplicates; adds an agent_message branch; moves function_call_output matching ahead of function_call to fix mis-classification; preserves call_id on all tool events; adds rawInputRedacted to function-call events.

Confidence Score: 5/5

Safe to merge — all changed parsing paths are exercised by the new regression tests and produce the expected outputs.

The changes are narrowly scoped to two importers: adding thinking-block extraction in Claude Code and fixing event-message deduplication plus function_call_output ordering in Codex. The new test file covers each modified path (thinking blocks, server_tool_use, web_search_tool_result, event_msg user/agent messages, function_call_output, and the legacy response_item fallback). No existing behaviour is removed; the only control-flow change (function_call_output checked before function_call) is a clear correctness fix, not a regression risk.

No files require special attention.

Important Files Changed

Filename	Overview
src/lib/importers/claude-code.ts	Adds thinking-block extraction and server_tool_use / web_search_tool_result tool-event support; logic is correct and well-covered by the new test.
src/lib/importers/codex.ts	Adds event_msg deduplication logic and fixes function_call_output ordering; all changed paths are exercised by the new tests.
src/lib/importers/importers.test.ts	New regression tests cover Claude Code thinking/tool blocks and Codex rollout event/function-output shapes; assertions are precise and match expected parsing outputs.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[JSONL row] --> B{item.type?}

    B -->|event_msg| C{payload.type?}
    C -->|user_message| D[codexUserMessageContent\n→ push user message]
    C -->|agent_message| E[extractTextContent\n→ push assistant message]

    B -->|response_item| F{payload.type?}
    F -->|message| G{hasEventMessages\n&& isConversation?}
    G -->|No → include| H[push message\nwith role]
    G -->|Yes → skip| I[deduplicated]

    B -->|assistant/user| J[extractClaudeContent\non content array]
    J --> K{block.type?}
    K -->|thinking| L[return block.thinking]
    K -->|text / tool_use\nredacted_thinking| M[extractTextContent]
    K -->|server_tool_use\nweb_search_tool_result\ntool_result| N[push ToolEvent\ninputSummary or outputSummary\n+ toolUseId]

    A --> O{candidateType regex?}
    O -->|function.*output\ntool.*output| P[ToolEvent\noutputSummary + call_id]
    O -->|function_call\ntool_call\nshell\ncommand| Q[ToolEvent\ninputSummary + call_id\n+ rawInputRedacted]

Loading

_{Reviews (3): Last reviewed commit: "address greptile review feedback (greplo..." | Re-trigger Greptile}