windmill/HANDOFF.md

Handoff: AI-chat thinking-blocks fix (PR #9841)

Branch: glm/fix-copilot-thinking-blocks · base main · file touched: frontend/src/lib/components/copilot/chat/anthropic.ts

The bug

Global AI chat (/sessions, dev-gated) 400s when an assistant turn uses the native web_search tool while thinking is on AND ends in a tool call:

400 invalid_request_error
messages.N.content.M: `thinking` or `redacted_thinking` blocks in the latest assistant
message cannot be modified. These blocks must remain as they were in the original response.

Root cause: the turn was reconstructed keeping only thinking blocks, reordering them to the front and dropping the server_tool_use / web_search_tool_result blocks. With interleaved thinking (a 2nd thinking block after the web-search result) the 2nd block's signature no longer validates.

What's already done (1 commit)

frontend/src/lib/components/copilot/chat/anthropic.ts:

• parseAnthropicCompletion: store the full turn as _anthropicContent = finalMessage.content (replaces the thinking-only _anthropicThinkingBlocks extraction).
• convertOpenAIToAnthropicMessages: if _anthropicContent is present, emit it verbatim; skip the redundant standalone text message for that turn; the old _anthropicThinkingBlocks path is kept as a fallback for sessions persisted before this change.
• npm run check:fast passes. Do NOT commit frontend/package-lock.json churn (revert if the install/check rewrites it).

Verify (still TODO)

1. Run this worktree's frontend against the EE backend: from frontend/, REMOTE=http://localhost:8000 npm run dev; open it in a browser.
2. Unlock global chat: DevTools console -> localStorage.setItem('wm_dev_global_ai','1') -> reload.
3. Workspace demo-orange -> AI Sessions -> new session (Claude model, effort high).
4. Prompt: "Search the web for the current NIST SP 800-63B minimum password length. After you get the number, in a follow-up step list the resources in my workspace. Think step by step between each tool call."
5. Expected: Thinking -> web search -> Thinking -> list resources, and the auto-continuation no longer 400s (on main it does). Capture the body to /api/w/demo-orange/ai/proxy/v1/messages: the replayed assistant content should keep thinking, server_tool_use, web_search_tool_result in original order (not ["thinking","thinking","tool_use"]).

Only NEW sessions are fixed — sessions already persisted in the old shape will still fail on continue; start a fresh session to test.

Remaining tasks

• Confirm the repro is fixed in-app; attach before/after screenshots to PR #9841.
• Add a unit test for convertOpenAIToAnthropicMessages (near AIChatManager.test.ts / global/core.test.ts): a turn with _anthropicContent = [thinking, server_tool_use, web_search_tool_result, thinking, tool_use] round-trips verbatim, and the preceding standalone text message is skipped.
• Sanity-check a plain text turn (no tools) and an old-style session (only _anthropicThinkingBlocks) still convert via the fallback.
• npm run check (full) before marking PR ready.
• Remove this HANDOFF.md before the PR is marked ready for review.

Deploying to orange-dev preview

Frontend-only fix — cherry-pick the commit onto the preview's frontend branch and rebuild the frontend bundle (no backend/DB change). Already-broken sessions stay broken; verify in a new chat.