fix(ai-chat): emit output-denied state for rejected tool approvals

[whoiskatrin]

Summary

Fixes #955.

When CF_AGENT_TOOL_APPROVAL is sent with approved: false, AIChatAgent now transitions tool parts to output-denied (instead of approval-responded).

This aligns with AI SDK expectations so convertToModelMessages can emit a tool_result for denied approvals (required by Anthropic tool-use sequencing).

What changed

• packages/ai-chat/src/index.ts
  • _applyToolApproval() now sets:
    • approved: true -> approval-responded
    • approved: false -> output-denied
  • Keeps merging existing approval object to preserve approval.id.
• packages/ai-chat/src/tests/client-tool-duplicate-message.test.ts
  • Updated existing rejection assertions to expect output-denied.
  • Added regression test: rejection from approval-requested state also transitions to output-denied and preserves approval.id.
• Added changeset for @cloudflare/ai-chat patch release.

Reviewer guide (quick pass)

1. Start with _applyToolApproval in packages/ai-chat/src/index.ts.
2. Check rejection test updates around:
   • CF_AGENT_TOOL_APPROVAL handles rejection (approved: false)
   • CF_AGENT_TOOL_APPROVAL with approved: false and autoContinue does NOT continue
3. Check the new test:
   • CF_AGENT_TOOL_APPROVAL rejection from approval-requested sets output-denied

Validation

• Ran targeted workers tests for rejection flows:
  • npm run test:workers --workspace @cloudflare/ai-chat -- src/tests/client-tool-duplicate-message.test.ts -t "rejection|approved: false"

[changeset-bot]

🦋 Changeset detected

Latest commit: 21c702c

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

This PR includes changesets to release 1 package

Name	Type
@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

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/cloudflare/agents@956

npm i https://pkg.pr.new/cloudflare/agents/@cloudflare/ai-chat@956

npm i https://pkg.pr.new/cloudflare/agents/@cloudflare/codemode@956

npm i https://pkg.pr.new/cloudflare/agents/hono-agents@956

commit: 21c702c

[whoiskatrin]

/bonk review this PR

[ask-bonk]

All 173 tests pass, including all 22 test files. Here's my review:

---

PR Review: fix(ai-chat): emit output-denied state for rejected tool approvals

Verdict: Approve -- this is a clean, well-scoped bug fix with good test coverage.

Summary

The core change is a one-line fix at packages/ai-chat/src/index.ts:1818: rejected tool approvals (approved: false) now transition to output-denied instead of approval-responded. This aligns with the AI SDK's convertToModelMessages expectations, which needs output-denied to emit a proper tool_result for providers like Anthropic that require tool-use sequencing.

Code review

packages/ai-chat/src/index.ts (+5/-2)

The fix itself is minimal and correct:

state: approved ? "approval-responded" : "output-denied",

• The ternary is clear and matches the existing pattern used by message-builder.ts:254 which already handles tool-output-denied -> output-denied transitions for streaming chunks.
• The approval object merge is preserved for both paths, so approval.id survives (needed by convertToModelMessages for the approvalId field).
• The matchStates array (["input-available", "approval-requested"]) is unchanged and correct for both approval and rejection.
• The auto-continue guard at line 621 (if (applied && approved && autoContinue)) already prevents continuation on rejection, so the output-denied state won't trigger an unintended LLM call.
• The JSDoc update accurately describes the new behavior.

packages/ai-chat/src/tests/client-tool-duplicate-message.test.ts (+64/-2)

• Two existing assertions updated from "approval-responded" to "output-denied" for the rejection cases (lines 678, 1131). Correct.
• New test "CF_AGENT_TOOL_APPROVAL rejection from approval-requested sets output-denied" (line 749) covers the case where a tool is already in approval-requested state and gets rejected. It verifies both state: "output-denied" and that approval.id is preserved through the merge. Good regression coverage.
• Test structure follows the existing patterns in the file.

.changeset/ten-llamas-exist.md (+9/-0)

Changeset is well-written and scoped to @cloudflare/ai-chat patch.

examples/playground/testing.md (+1/-1)

Escapes * in X-Agent-\* for markdown rendering. Unrelated but harmless housekeeping.

Minor observations (non-blocking)

1. The _resolveMessageForToolMerge method (line 1043) checks for output-available, approval-responded, and approval-requested states but does not check for output-denied. If a message with an output-denied tool part gets re-persisted through a different code path, it won't get the toolCallId-based dedup. This seems fine in practice since denied tools are terminal state and won't be re-sent, but worth noting.

2. All 173 ai-chat workers tests pass locally.

Verdict

Clean fix, correct logic, good tests. Ship it.

github run

[threepointone]

very good pr, I added an extra paranoid e2e test for my sake, and merged from main. approved, will land when ci passes