fix: report stop_reason tool_use for tool calls on Anthropic-compatible egress

[eyeveil]

Problem

When Bifrost's Anthropic-compatible ingress (/anthropic/v1/messages) routes to a non-Anthropic backend (e.g. Azure OpenAI), a tool-calling turn emits a correct tool_use content block but the terminal message_delta.delta.stop_reason is "end_turn" instead of "tool_use". Anthropic SDK consumers that dispatch tools on stop_reason == "tool_use" (the documented signal) fail to detect the tool call.

Root cause

The Anthropic Responses streaming egress converter defaulted response.completed to stop_reason: "end_turn" unless BifrostResponsesResponse.StopReason was explicitly set. OpenAI/Azure Responses-style completed payloads can carry function_call output items with no Bifrost stop reason — the converter had enough information to know the turn requested a tool, but didn't infer "tool_use".

Fix

Centralized Anthropic stop-reason inference for Responses egress (inferAnthropicStopReasonFromBifrostResponse), used by both the non-streaming conversion and the streaming response.completed → message_delta conversion:

• explicit stop reason mapping preserved ("tool_calls" → "tool_use", etc. via the existing mapping table);
• nil stop reason + function_call output items ⇒ "tool_use";
• nil stop reason + converted tool_use content blocks ⇒ "tool_use" (existing non-streaming inference, kept);
• otherwise "end_turn".

Testing

• Extended the non-streaming stop-reason table test (tool_calls + function-call output ⇒ tool_use).
• Added TestToAnthropicResponsesStreamResponse_CompletedWithFunctionCallInfersToolUseStopReason covering the reported streaming case: response.completed, nil StopReason, function-call output ⇒ message_delta with stop_reason: "tool_use".
• go build ./..., go vet ./providers/anthropic/, go test -count=1 ./providers/anthropic/ (core, cgo enabled) — all pass.

Fixes #4065

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 5126b101-228a-41e3-95ee-a9d15ea937ab

📥 Commits

Reviewing files that changed from the base of the PR and between f45804f and dfd0878.

📒 Files selected for processing (2)

• core/providers/anthropic/compaction_test.go
• core/providers/anthropic/responses.go

🚧 Files skipped from review as they are similar to previous changes (1)

• core/providers/anthropic/responses.go

---

📝 Walkthrough

Summary by CodeRabbit

• Bug Fixes
  • Improved stop-reason handling for Anthropic responses and streams so tool usage is now detected more consistently.
  • Responses with tool-call data now correctly return tool_calls instead of defaulting to end_turn.
  • Streamed completed responses now emit the expected message_delta followed by message_stop with the correct stop reason.
  • Updated behavior for cases with reasoning-only function calls to avoid incorrectly inferring tool use.

Walkthrough

This PR changes Anthropic stop-reason derivation for streaming and non-streaming response conversion. It adds shared inference logic for function_call and tool_use handling, updates both conversion paths to use it, and expands tests around completed and streaming responses.

Changes

Stop reason inference fix

Layer / File(s)	Summary
Stop-reason inference helper implementation core/providers/anthropic/responses.go	Adds helper functions that detect function_call output and tool_use content blocks, then infer the Anthropic stop reason from Bifrost response data with end_turn as the fallback.
Wiring inference into response conversion paths core/providers/anthropic/responses.go	Updates the streaming message_delta path and ToAnthropicResponsesResponse to use the shared stop-reason inference helper instead of the previous hard-coded logic.
Test coverage for function_call stop reason inference core/providers/anthropic/compaction_test.go	Refactors the existing stop-reason test cases for function_call inputs and adds a streaming test that asserts tool_calls is emitted for completed responses with tool-call data.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• maximhq/bifrost#3640: Both PRs adjust Anthropic stop-reason handling for tool-call responses in conversion paths.
• maximhq/bifrost#3685: Both PRs update stop-reason mapping logic in core/providers/anthropic/responses.go.

Suggested reviewers: danpiths, TejasGhatte, akshaydeo

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly states the main fix: reporting tool-use stop_reason for Anthropic-compatible egress.
Description check	✅ Passed	The description covers the problem, root cause, fix, testing, and related issue, though several template sections are omitted.
Linked Issues check	✅ Passed	The changes map backend tool_calls/function_call outputs to Anthropic tool_use stop_reason and add regression tests for the reported case.
Out of Scope Changes check	✅ Passed	The modified files stay within Anthropic response logic and tests, with no unrelated out-of-scope changes evident.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

Warning

There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure.

🔧 golangci-lint (2.12.2)

level=error msg="[linters_context] typechecking error: pattern ./...: directory prefix . does not contain main module or its selected dependencies"

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Confidence Score: 5/5

Safe to merge — the change is a targeted, pure-function extraction that fixes a well-scoped streaming egress bug with no side effects on other provider paths.

The helper is a pure transformation with no HTTP calls or shared state; the inference logic is correct and all branches are exercised by existing or new tests. The only behavior difference outside the bug fix is that message_delta.stop_sequence now serializes as null instead of "", which is more spec-compliant. No regressions in the non-streaming path are introduced by the ordering of the new output check.

No files require special attention.

Important Files Changed

Filename	Overview
core/providers/anthropic/responses.go	Introduces inferAnthropicStopReasonFromBifrostResponse helper that correctly infers tool_use from function-call output items; wires it into both streaming and non-streaming Responses egress paths; removes the spurious stop_sequence: "" from message_delta.
core/providers/anthropic/compaction_test.go	Adds streaming test for function-call inferred tool_use stop reason and expands the non-streaming stop-reason table with explicit-mapping and reasoning-exclusion cases.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["response.completed / ToAnthropicResponsesResponse"] --> B["inferAnthropicStopReasonFromBifrostResponse(bifrostResp, contentBlocks)"]
    B --> C{bifrostResp.StopReason != nil?}
    C -- Yes --> D["ConvertBifrostFinishReasonToAnthropic(stopReason)\ne.g. 'tool_calls' → tool_use"]
    C -- No --> E{"bifrostResponsesOutputHasFunctionCall(Output)?"}
    E -- "Yes (has ResponsesToolMessage, no ResponsesReasoning)" --> F["tool_use"]
    E -- No --> G{"anthropicContentBlocksHaveToolUse(contentBlocks)?"}
    G -- Yes --> H["tool_use"]
    G -- No --> I["end_turn"]

    D --> Z["Emit message_delta.stop_reason"]
    F --> Z
    H --> Z
    I --> Z

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["response.completed / ToAnthropicResponsesResponse"] --> B["inferAnthropicStopReasonFromBifrostResponse(bifrostResp, contentBlocks)"]
    B --> C{bifrostResp.StopReason != nil?}
    C -- Yes --> D["ConvertBifrostFinishReasonToAnthropic(stopReason)\ne.g. 'tool_calls' → tool_use"]
    C -- No --> E{"bifrostResponsesOutputHasFunctionCall(Output)?"}
    E -- "Yes (has ResponsesToolMessage, no ResponsesReasoning)" --> F["tool_use"]
    E -- No --> G{"anthropicContentBlocksHaveToolUse(contentBlocks)?"}
    G -- Yes --> H["tool_use"]
    G -- No --> I["end_turn"]

    D --> Z["Emit message_delta.stop_reason"]
    F --> Z
    H --> Z
    I --> Z

Loading

_{Reviews (2): Last reviewed commit: "[fix]: Anthropic provider - do not infer..." | Re-trigger Greptile}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@core/providers/anthropic/responses.go`:
- Around line 3541-3545: The helper bifrostResponsesOutputHasFunctionCall is
incorrectly treating all ResponsesMessageTypeFunctionCall items as tool calls,
even when they are reasoning-only. Update this function to ignore function_call
messages that have ResponsesReasoning set, matching the nearby streaming
conversion logic, so only real tool-use items cause a tool-use stop_reason.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 1b228e3f-9200-43d8-97cd-73c422e0b47e

📥 Commits

Reviewing files that changed from the base of the PR and between 6484317 and f45804f.

📒 Files selected for processing (2)

• core/providers/anthropic/compaction_test.go
• core/providers/anthropic/responses.go

[eyeveil]

@coderabbitai review — the reasoning-only function_call skip in bifrostResponsesOutputHasFunctionCall already landed in dfd0878 (with a regression test); the CHANGES_REQUESTED review predates that commit.

[coderabbitai]

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail
gh pr view 4900 --repo maximhq/bifrost --json commits --jq '.commits[] | .oid + " " + .messageHeadline' | head -50
echo "---"
gh api repos/maximhq/bifrost/commits/dfd0878d6 --jq '.sha, .commit.message' 2>/dev/null || echo "commit not found directly"

Length of output: 1067

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail
gh api repos/maximhq/bifrost/commits/dfd0878d68aa8be3b79d6e1641d5ac50172b848d --jq '.files[].filename'
echo "---diff---"
gh api repos/maximhq/bifrost/commits/dfd0878d68aa8be3b79d6e1641d5ac50172b848d --jq '.files[].patch' 2>/dev/null | head -150

Length of output: 2069

---

@nnNyx, confirmed — commit dfd0878 adds the guard to bifrostResponsesOutputHasFunctionCall (now requiring ResponsesReasoning == nil && ResponsesToolMessage != nil) plus the regression test "reasoning-only function_call output does not infer tool_use" in core/providers/anthropic/compaction_test.go. This directly addresses the concern from my earlier CHANGES_REQUESTED review. Triggering a fresh review now to re-evaluate against the latest commits.

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.