fix(ai-chat): stop provider tool-call replays from regressing tool part state (#1404)

[threepointone]

Summary

Closes #1404. The OpenAI Responses API (and likely other providers) re-emits prior tool calls in continuation streams as a tool-input-start → tool-input-delta → tool-input-available → tool-output-available sequence carrying the same toolCallId and the same output the part already holds. AI SDK v6's updateToolPart mutates an existing tool part in place when the toolCallId matches, so a replayed tool-input-start was clobbering an output-available part back to input-streaming on the client and producing the worker warn:

(warn) [AIChatAgent] _applyToolResult: Tool part with toolCallId call_xxx
not in expected state (expected: input-available|approval-requested|approval-responded)

What changed

packages/agents/src/chat/message-builder.ts

• applyChunkToParts is now idempotent against an existing tool part with the same toolCallId for tool-input-start, tool-input-delta, tool-input-available, and tool-input-error. A replayed tool-input-start no longer pushes a duplicate part or regresses state; deltas/available chunks only mutate while still input-streaming; tool-input-error preserves an existing terminal state (first-write-wins).
• New exported helper isReplayChunk(parts, chunk) returns true when a tool-input-* chunk would visibly regress an AI SDK v6 client's tool part. Stream broadcasters use it to drop replay chunks before forwarding them. tool-output-available is intentionally not in the helper because its in-place update is safe when the data already matches.

packages/ai-chat/src/index.ts

• _streamSSEReply calls isReplayChunk on each chunk and skips applying / storing / broadcasting if it's a replay. Combined with the applyChunkToParts idempotency, the cloned server-side streaming message stays clean and the regression-inducing chunks never reach the client.
• _applyToolResult accepts output-available, output-error, and output-denied as starting states for idempotent re-application. A duplicate cf_agent_tool_result (cross-tab re-run, redelivered WS frame, provider replay round-trip) is now a silent no-op rather than a warn + skipped update. The cross-message tool-output-available / tool-output-error fallback gets the same first-write-wins semantics.
• _findAndUpdateToolPart separately tracks wasFound (a matching part was processed) and hasRealChange (the apply actually mutated state). Idempotent re-applies skip the SQLite write and MESSAGE_UPDATED broadcast. The split also handles legacy duplicate tool parts correctly: a real change to one duplicate is still persisted even when another duplicate is already terminal.

Why two layers

The visible regression on the client comes from the AI SDK's own updateToolPart, which we can't fix from this repo. So the server has to stop forwarding the regression-inducing chunks (layer 1: isReplayChunk filter in _streamSSEReply). And the worker warn comes from _applyToolResult not accepting an already-terminal state, which we fix by making the apply itself idempotent (layer 2). The two fixes are independent: either alone reduces the symptoms, both together kill the bug.

The pre-existing "first-write-wins" contract for terminal tool states (locked in by client-tool-duplicate-message.test.ts) is preserved — terminal-state re-applies always become silent no-ops, never overwrites.

Test plan

• packages/ai-chat/src/tests/message-builder.test.ts — new describe blocks for tool-input-* idempotency against existing tool parts (8 tests) and isReplayChunk (8 tests). Covers single-chunk regressions, the full provider replay sequence, approval-state preservation, and tool-input-error against terminal parts.
• packages/ai-chat/src/tests/tool-result-replay.test.ts — new file (5 tests). Idempotent re-apply for output-available / output-error (no warn, no broadcast); regression guard that real transitions still broadcast; the full continuation-stream replay scenario verifying chunks are not forwarded; legacy duplicate-tool-part edge case.
• packages/ai-chat/src/tests/worker.ts — adds a replayPriorToolCall body knob to TestChatAgent.onChatMessage simulating the OpenAI replay chunk pattern, and a testApplyToolResult helper.
• npm run check — sherif + export checks + oxfmt + oxlint + typecheck all clean.
• packages/ai-chat — 476/476 tests pass.
• packages/agents — 1372/1372 tests pass (8 skipped, unchanged).
• packages/think — 270/270 tests pass.
• Two changesets included (agents patch + @cloudflare/ai-chat patch).

Made with Cursor

---

[changeset-bot]

🦋 Changeset detected

Latest commit: 3e6a10a

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 2 packages

Name	Type
agents	Patch
@cloudflare/ai-chat	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 4 additional findings.