[Harrison (iZiTTMarvin)]

Summary

Addresses #204.

Editing a human message had two problems in src/components/thread/messages/human.tsx:

• Duplicate message in the optimistic UI. optimisticValues started from getMessagesMetadata(message).firstSeenState.values. In @langchain/langgraph-sdk 1.12.4, firstSeenState is the oldest history state that already contains the edited message (dist/react/stream.lgp.js:271, findLast(history, state => messages.includes(messageId))), so appending the new message rendered the old and new text together until the run streamed back.
• Unpaired tool calls after a stopped run. If a run was cancelled mid tool call, the normal composer pairs the dangling call with ensureToolCallsHaveResponses (index.tsx), but the edit path did not. Forking from the parent checkpoint then sent [..., ai(tool_calls), human], which providers like Anthropic reject (tool_use ids were found without tool_result blocks).

Changes (single file):

• Build optimistic values from the parent checkpoint's state (looked up in thread.history by checkpoint_id), i.e. exactly what the server forks from. This does not depend on message ids.
• Pair dangling tool calls with ensureToolCallsHaveResponses, same as a normal submit.
• Give the edited message a uuidv4() id, like the main composer.
• If the parent checkpoint is unavailable, show a toast and keep the editor open instead of submitting.

The server-side fork itself (submit(input, { checkpoint: parent })) was already correct, so updateState + submit(null) is not needed.

Repro (deterministic local LangGraph.js graph, no model keys)

1. Send "weather?" (one tool call), then "hello"; edit "hello". Before: optimistic messages [..., "hello", "edited"]. After: [..., "edited"]. Final server state is unchanged and tool call/result ids stay paired.
2. Send a message that triggers a slow tool, cancel mid tool call, send "hello", edit it. Before: run fails with tool_use ids were found without tool_result blocks immediately after: call_1 (provider validation mimicked in the graph). After: the run completes.

Verification

• pnpm format:check and pnpm build: pass.
• eslint .: 0 errors and the same 18 existing warnings as main (on Windows, run with NODE_OPTIONS set manually because the lint script uses POSIX env syntax).
• Manually checked against langgraphjs dev with the graph above.

Limitations

• The provider error was reproduced with a validator that mimics Anthropic, not a real provider call.
• Not tested against the Python server.
• The original report doesn't mention cancelled runs, so there may be other triggers for the tool_use error.
• As before, editing replaces the content with text, so non-text blocks of the original message are not kept.

[vercel]

Harrison (@iZiTTMarvin) is attempting to deploy a commit to the LangChain Team on Vercel.

A member of the Team first needs to authorize it.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.