Conversation
…e anthropic provider experimental_output (JSON-schema enforcement) + tools silently drops tool_use blocks on the native Anthropic Messages API surface (provider 'anthropic'/'bedrock', incl. via a proxy/base-URL override): the model returns finishReason=tool-calls but zero tool calls are parsed, so the agent loop ends at step 1 and only the preamble text is returned — surfacing as fabricated tool-success replies and structured-output retry storms. The static gate previously excluded only Gemini; its comment assumed Anthropic supports both. Extend isToolsSchemaExclusionInForce() to also exclude the native anthropic surface via isNativeAnthropicProvider(), mirroring the Gemini fallback. Vertex+Claude (provider 'vertex', claude- model) is intentionally NOT matched — different transport, no conflict — so the primary production config keeps strict structured output. Verified end-to-end against an anthropic-via-proxy deployment: tool calls now execute (clean 2-step tool loop, real tool return values) instead of being dropped.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
📝 WalkthroughWalkthroughStructured output policy now covers native Anthropic tool usage and temperature-deprecation detection. GenerationHandler retries once without temperature on matching failures. Anthropic provider conversion now handles file parts and shared image/PDF block helpers, with new continuous tests for both policy and multimodal conversion. ChangesStructured output, Anthropic conversion, and retry
Sequence Diagram(s)sequenceDiagram
participant GenerationHandler
participant structuredOutputPolicy
participant callGenerateText
GenerationHandler->>structuredOutputPolicy: isTemperatureDeprecatedError(error)
structuredOutputPolicy-->>GenerationHandler: true
GenerationHandler->>callGenerateText: retry with temperature undefined
callGenerateText-->>GenerationHandler: retry result
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Suggested reviewers
Possibly related PRs
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
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. 🔧 ESLint
ESLint install failed: private package registry requires authentication. Disable ESLint in CodeRabbit settings or use public packages. 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. Comment |
✅ Single Commit Policy - COMPLIANTStatus: Policy requirements met • 1 commit • Valid format • Ready for merge 📊 View validation details📝 Commit Details
✅ Validation Results
🤖 Automated validation by NeuroLink Single Commit Enforcement |
🤖 AI Review & Build Compliance ✅Status: AI analysis complete • Build rules validated • Ready for review 📊 View detailed analysis results🛡️ Analysis Complete
📋 Ready for Merge When
🤖 AI analysis complete - check individual code comments for specific feedback |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
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 `@src/lib/core/modules/structuredOutputPolicy.ts`:
- Around line 58-61: The `isToolsSchemaExclusionInForce` test coverage needs to
be extended for the new native provider branches. Update
`test/continuous-test-suite-json.ts` by adding assertions alongside the existing
Gemini/tools test to verify that `isToolsSchemaExclusionInForce("anthropic",
"claude-3-sonnet", true, 5)` and `isToolsSchemaExclusionInForce("bedrock",
"anthropic.claude-3-sonnet", true, 5)` both return `true`. Use the existing test
block around `isToolsSchemaExclusionInForce` as the insertion point and keep the
assertions consistent with the current `assertEqual` style.
🪄 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
Run ID: d16c2c4f-7d3c-4806-9dae-d41b056d1de5
📒 Files selected for processing (1)
src/lib/core/modules/structuredOutputPolicy.ts
| (isGeminiProvider(providerName, modelName) || | ||
| isNativeAnthropicProvider(providerName)) && | ||
| shouldUseTools && | ||
| toolCount > 0 |
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟠 Major
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
# Description: Verify that structured-output exclusion tests cover the new native Anthropic surfaces.
# Expect: assertions for anthropic+tools and bedrock+tools returning true, plus vertex+claude+tools returning false.
rg -n -C 3 'isToolsSchemaExclusionInForce\("(anthropic|bedrock|vertex)"' test/continuous-test-suite-json.tsRepository: juspay/neurolink
Length of output: 655
🏁 Script executed:
#!/bin/bash
# Check for any existing tests involving "anthropic" or "bedrock" in the test file
rg -n "(anthropic|bedrock)" test/continuous-test-suite-json.ts --type tsRepository: juspay/neurolink
Length of output: 324
Add assertions for native Anthropic and Bedrock to isToolsSchemaExclusionInForce tests.
The current test suite in test/continuous-test-suite-json.ts lacks coverage for the new provider-specific logic. You must add assertions verifying that:
isToolsSchemaExclusionInForce("anthropic", "claude-3-sonnet", true, 5)returnstrue.isToolsSchemaExclusionInForce("bedrock", "anthropic.claude-3-sonnet", true, 5)returnstrue.
<details>
<summary>Current test coverage</summary>
```typescript
// Existing test (Line 58-61)
await test("exclusion fires only for Gemini + tools present", () => {
// Verifies Vertex+Claude = false, Vertex+Gemini = true
assertEqual(isToolsSchemaExclusionInForce("vertex", "claude-sonnet-4-6", true, 5), false);
assertEqual(isToolsSchemaExclusionInForce("vertex", "gemini-2.5-pro", true, 5), true);
});Add similar blocks for the native providers to prevent regression of the new exclusion logic.
🤖 Prompt for 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.
In `@src/lib/core/modules/structuredOutputPolicy.ts` around lines 58 - 61, The
`isToolsSchemaExclusionInForce` test coverage needs to be extended for the new
native provider branches. Update `test/continuous-test-suite-json.ts` by adding
assertions alongside the existing Gemini/tools test to verify that
`isToolsSchemaExclusionInForce("anthropic", "claude-3-sonnet", true, 5)` and
`isToolsSchemaExclusionInForce("bedrock", "anthropic.claude-3-sonnet", true, 5)`
both return `true`. Use the existing test block around
`isToolsSchemaExclusionInForce` as the insertion point and keep the assertions
consistent with the current `assertEqual` style.
Tara-ag
left a comment
Tara-ag
left a comment
There was a problem hiding this comment.
✅ Approved
This PR correctly fixes a critical bug where experimental_output combined with tools silently drops tool calls on the native Anthropic Messages API surface.
Review Summary
Files Changed: 1 (src/lib/core/modules/structuredOutputPolicy.ts)
Additions: 28 lines
Deletions: 7 lines
Key Observations
-
Correct Fix: The change extends the tools↔schema exclusion to native Anthropic providers (
anthropic/bedrock) while intentionally preserving Vertex+Claude behavior. This aligns with CLAUDE.md Critical Rule 3's pattern. -
Backward Compatible: The change is purely additive - it adds new providers to the exclusion list without affecting existing behavior for other providers. Downstream
validateOrRecoverStructured()ensures no regression in response format. -
Well-Documented: Comments clearly explain the distinction between:
- Gemini (has conflict)
- Native Anthropic/Bedrock (has conflict)
- Vertex+Claude (no conflict - different transport)
-
Clean Implementation: The new
isNativeAnthropicProvider()function follows the same simple predicate pattern asisGeminiProvider().
Suggestion (Non-blocking)
Consider adding test cases to test/continuous-test-suite-json.ts to verify:
isToolsSchemaExclusionInForce("anthropic", "claude-sonnet-4-6", true, 5)returnstrueisToolsSchemaExclusionInForce("bedrock", "anthropic.claude-sonnet-4-6", true, 5)returnstrueisToolsSchemaExclusionInForce("anthropic", "claude-sonnet-4-6", false, 0)returnsfalse(no tools)
This ensures the new behavior is covered by the continuous test suite.
Overall: Clean, focused fix that addresses the root cause without introducing risks. Approved for merge.
Verification update — why disabling (not a different mode) is the right callConfirmed the conflict is with Anthropic's native structured outputs ( End-to-end on an anthropic-via-proxy deployment with this patch, a single turn:
So disabling AI-SDK enforcement does not lose structured output: the consumer's downstream text→schema recovery reconstructs and validates it, exactly as already happens on Gemini. Tools + structured result both work. |
The newest Anthropic models (e.g. claude-opus-4-8 with tools + advanced beta features) reject temperature with HTTP 400 ('temperature is deprecated for this model.') in favour of reasoning-effort controls. NeuroLink sent it unconditionally, so the whole turn failed ('All providers failed').
Add isTemperatureDeprecatedError() and, in GenerationHandler's catch, retry the generation once with temperature omitted (mirrors the existing isToolsSchemaConflictError structured-output fallback). Add a deterministic test suite covering the native-anthropic structured-output gate + the temperature detector.
🤖 AI Review & Build Compliance ✅Status: AI analysis complete • Build rules validated • Ready for review 📊 View detailed analysis results🛡️ Analysis Complete
📋 Ready for Merge When
🤖 AI analysis complete - check individual code comments for specific feedback |
There was a problem hiding this comment.
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 `@src/lib/core/modules/GenerationHandler.ts`:
- Around line 577-585: The temperature-retry return path in GenerationHandler is
missing the same span telemetry recorded in the primary and structured-output
fallback paths. Before returning the result in the retry branch, use the same
usage/cost attributes as the other paths by setting gen_ai.usage.input_tokens,
gen_ai.usage.output_tokens, and neurolink.cost on the span alongside the
existing finish_reason and retry.count updates. Refer to the result handling
inside GenerationHandler so the retry path mirrors the telemetry behavior of the
other branches.
🪄 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
Run ID: 47168301-076e-4736-b32d-85ba498a9fd5
📒 Files selected for processing (3)
src/lib/core/modules/GenerationHandler.tssrc/lib/core/modules/structuredOutputPolicy.tstest/continuous-test-suite-anthropic-tools-policy.ts
| span.setAttribute("retry.count", 1); | ||
| if (result.finishReason) { | ||
| span.setAttribute( | ||
| "gen_ai.response.finish_reason", | ||
| result.finishReason, | ||
| ); | ||
| } | ||
| span.setStatus({ code: SpanStatusCode.OK }); | ||
| return result; |
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win
Temperature-retry path drops usage/cost span telemetry.
The primary (Lines 404-422) and structured-output fallback (Lines 502-523) paths both set gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, and neurolink.cost on the span. This retry path sets only finish_reason, so traces recovered via the temperature retry report a cost of 0 and no token usage, breaking the "what did this trace cost?" query for that path.
📊 Proposed fix to record usage/cost before returning
span.addEvent("retry.recovered", {
"retry.attempts": 2,
"retry.strategy": "temperature_omitted",
});
span.setAttribute("retry.count", 1);
+ if (result.usage) {
+ span.setAttribute(
+ "gen_ai.usage.input_tokens",
+ result.usage.inputTokens || 0,
+ );
+ span.setAttribute(
+ "gen_ai.usage.output_tokens",
+ result.usage.outputTokens || 0,
+ );
+ const retryCost = calculateCost(
+ this.providerName,
+ this.modelName,
+ {
+ input: result.usage.inputTokens || 0,
+ output: result.usage.outputTokens || 0,
+ total:
+ (result.usage.inputTokens || 0) +
+ (result.usage.outputTokens || 0),
+ },
+ );
+ span.setAttribute("neurolink.cost", retryCost ?? 0);
+ }
if (result.finishReason) {📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| span.setAttribute("retry.count", 1); | |
| if (result.finishReason) { | |
| span.setAttribute( | |
| "gen_ai.response.finish_reason", | |
| result.finishReason, | |
| ); | |
| } | |
| span.setStatus({ code: SpanStatusCode.OK }); | |
| return result; | |
| span.setAttribute("retry.count", 1); | |
| if (result.usage) { | |
| span.setAttribute( | |
| "gen_ai.usage.input_tokens", | |
| result.usage.inputTokens || 0, | |
| ); | |
| span.setAttribute( | |
| "gen_ai.usage.output_tokens", | |
| result.usage.outputTokens || 0, | |
| ); | |
| const retryCost = calculateCost( | |
| this.providerName, | |
| this.modelName, | |
| { | |
| input: result.usage.inputTokens || 0, | |
| output: result.usage.outputTokens || 0, | |
| total: | |
| (result.usage.inputTokens || 0) + | |
| (result.usage.outputTokens || 0), | |
| }, | |
| ); | |
| span.setAttribute("neurolink.cost", retryCost ?? 0); | |
| } | |
| if (result.finishReason) { | |
| span.setAttribute( | |
| "gen_ai.response.finish_reason", | |
| result.finishReason, | |
| ); | |
| } | |
| span.setStatus({ code: SpanStatusCode.OK }); | |
| return result; |
🤖 Prompt for 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.
In `@src/lib/core/modules/GenerationHandler.ts` around lines 577 - 585, The
temperature-retry return path in GenerationHandler is missing the same span
telemetry recorded in the primary and structured-output fallback paths. Before
returning the result in the retry branch, use the same usage/cost attributes as
the other paths by setting gen_ai.usage.input_tokens,
gen_ai.usage.output_tokens, and neurolink.cost on the span alongside the
existing finish_reason and retry.count updates. Refer to the result handling
inside GenerationHandler so the retry path mirrors the telemetry behavior of the
other branches.
Tara-ag
left a comment
Tara-ag
left a comment
There was a problem hiding this comment.
Review Summary
Files reviewed: 3
New issues raised: 1 (MINOR)
Existing comments acknowledged: 1 (CodeRabbit's test coverage comment - addressed by the new test file)
Assessment
This PR correctly fixes a critical production issue where the native Anthropic Messages API surface (including via proxy) silently drops tool calls when experimental_output (structured output) is combined with tools. The fix properly distinguishes between:
- Native Anthropic (
anthropic/bedrock): Disables structured output when tools are active ✅ - Vertex+Claude (
vertexwithclaude-*models): Preserves structured output with tools ✅
Changes Analysis
| File | Assessment |
|---|---|
structuredOutputPolicy.ts |
Clean implementation of isNativeAnthropicProvider() with clear documentation explaining the transport difference. Correctly extends isToolsSchemaExclusionInForce() without breaking existing Gemini logic. |
GenerationHandler.ts |
Temperature deprecation retry follows established fallback patterns. Proper telemetry attribution. No infinite loop risk (single retry only). |
continuous-test-suite-anthropic-tools-policy.ts |
Comprehensive test coverage for all new logic paths. Well-structured using the canonical harness. |
Issue Raised
- MINOR: New test file not registered in
package.jsonscripts. Consider adding"test:anthropic-tools-policy": "npx tsx test/continuous-test-suite-anthropic-tools-policy.ts"and including it intest:unitsince it's a pure logic test.
Compliance Check
- ✅ CLAUDE.md Rule 3 (Gemini tools + JSON schema mutual exclusion) - correctly extended to native Anthropic
- ✅ Backward compatibility preserved - Vertex+Claude unaffected
- ✅ Error handling follows existing patterns
- ✅ Tests included for new behavior
Recommendation: Approve after addressing the minor package.json registration.
There was a problem hiding this comment.
💡 MINOR: Test file not registered in package.json scripts
This new test suite is comprehensive and well-structured, but it's not registered in package.json scripts. Per CONTRIBUTING.md:
"Or create a new suite
test/continuous-test-suite-<name>.tsand add a matchingtest:<name>script inpackage.json"
Suggested addition to package.json scripts:
"test:anthropic-tools-policy": "npx tsx test/continuous-test-suite-anthropic-tools-policy.ts",Since this is a pure logic test (no API calls), also consider adding it to test:unit:
"test:unit": "pnpm run test:envguard && pnpm run test:bugfixes && ... && pnpm run test:anthropic-tools-policy",Without registration, these tests won't run in CI and could regress unnoticed.
…mperature proactively AI-SDK v6 encodes images AND PDFs as `type:"file"` parts in the LanguageModel prompt that the provider's doGenerate receives. messagesToAnthropic only handled text/image/image_url, so on the tool-using generate path the image was silently dropped and the model answered "no image detected". Vertex/Gemini are immune (they read input.images directly in their own builders). - Add anthropicImageBlocks.ts: pure file/image -> Anthropic block converters. Image media type is taken from the AI-SDK mediaType, then sniffed from magic bytes, then defaulted (a hardcoded image/png corrupts JPEG/GIF/WebP -> HTTP 400). PDF file parts become document blocks. - messagesToAnthropic: handle the `type:"file"` content part. - modelDeprecatesTemperature(): omit `temperature` proactively for opus-4-8+ (which reject it) instead of relying on the fail-then-retry safety net — the guaranteed-to-fail first request was a wasted round-trip on every turn. - Deterministic suites for both (no API).
🤖 AI Review & Build Compliance ✅Status: AI analysis complete • Build rules validated • Ready for review 📊 View detailed analysis results🛡️ Analysis Complete
📋 Ready for Merge When
🤖 AI analysis complete - check individual code comments for specific feedback |
Tara-ag
left a comment
Tara-ag
left a comment
There was a problem hiding this comment.
Review Summary
Files reviewed: 6 changed files
New issues raised: 2 (both MINOR)
Blocking issues: 0
Assessment
This PR effectively addresses a critical production issue where experimental_output combined with tools silently drops tool calls on the native Anthropic Messages API surface. The fix is well-architected:
- Correctly extends the structured output exclusion to native Anthropic providers (
anthropic/bedrock) while intentionally preserving Vertex+Claude behavior - Adds temperature deprecation handling for newer Anthropic models (claude-opus-4-8+) with proactive omission and reactive retry
- Fixes multimodal image handling for AI-SDK v6's
type:"file"encoding that was dropping images on the tool-using generate path - Includes comprehensive test coverage for both the policy logic and multimodal conversion
New Comments Raised
- 💡 MINOR: Type assertion safety suggestion in
anthropic.tsfor file part handling - 💡 MINOR: Test file registration in
package.jsonscripts (CONTRIBUTING.md compliance)
Already Addressed by Existing Comments
- Test coverage extension for native Anthropic/Bedrock assertions (CodeRabbit)
- Temperature-retry telemetry parity (CodeRabbit)
- Package.json script registration for multimodal test (Tara-ag)
Architectural Verification
✅ Static import from structuredOutputPolicy.ts to anthropic.ts is acceptable - this is a utility module, not a provider (CLAUDE.md Critical Rule 1 applies to provider-to-provider imports only)
✅ No breaking changes to public SDK API
✅ No security vulnerabilities identified
✅ Proper error handling and telemetry throughout
Recommendation: Approve with minor suggestions. The core fix is solid and addresses a real production pain point.
| const cc = cacheControlOf(p); | ||
| blocks.push(cc ? { ...img, cache_control: cc } : img); | ||
| } | ||
| } else if (p?.type === "file") { |
There was a problem hiding this comment.
💡 MINOR: Type assertion safety for file part handling
The type assertion p as { mediaType?: string; data?: unknown } on line 449 assumes the part object has this shape, but p originates from message parsing and may not fully match at runtime. Consider adding a runtime validation check:
} else if (p?.type === "file") {
// Runtime validation before type assertion
const filePart = p as { mediaType?: unknown; data?: unknown };
if (typeof filePart.mediaType === "string" || filePart.mediaType === undefined) {
const block = fileToAnthropicBlock({
mediaType: filePart.mediaType,
data: filePart.data,
});
// ... rest of handling
}
}This is a minor defensive coding suggestion - the current implementation is functional given the controlled input sources.
There was a problem hiding this comment.
💡 MINOR: Test file not registered in package.json scripts
Per CONTRIBUTING.md, new test suites should be registered in package.json scripts. Consider adding:
"test:anthropic-tools-policy": "npx tsx test/continuous-test-suite-anthropic-tools-policy.ts",
"test:anthropic-multimodal": "npx tsx test/continuous-test-suite-anthropic-multimodal.ts",And optionally include them in test:unit since they're pure logic tests (no API calls):
"test:unit": "pnpm run test:envguard && pnpm run test:bugfixes && ... && pnpm run test:anthropic-tools-policy && pnpm run test:anthropic-multimodal",Without registration, these tests won't run in CI and could regress unnoticed.
There was a problem hiding this comment.
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 `@src/lib/providers/anthropicImageBlocks.ts`:
- Around line 232-234: The `anthropicImageBlocks` flow is passing every
`image/*` media type into `toAnthropicImageBlock()`, which causes unsupported
hints like `image/svg+xml` or `image/bmp` to be mislabeled as PNG instead of
being omitted. Update the `mediaType.startsWith("image/")` branch to only call
`toAnthropicImageBlock()` for supported image types and return `undefined` for
unsupported ones, so the function’s behavior matches its contract. Use
`toAnthropicImageBlock` and the `mediaType.startsWith("image/")` check as the
main places to adjust.
🪄 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
Run ID: a96d9efd-459f-42bc-9872-c52fd44110d1
📒 Files selected for processing (5)
src/lib/core/modules/structuredOutputPolicy.tssrc/lib/providers/anthropic.tssrc/lib/providers/anthropicImageBlocks.tstest/continuous-test-suite-anthropic-multimodal.tstest/continuous-test-suite-anthropic-tools-policy.ts
| if (mediaType.startsWith("image/")) { | ||
| return toAnthropicImageBlock(data, mediaType); | ||
| } |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Skip unsupported image/* hints instead of relabeling them as PNG.
Lines 232-234 currently send every image/* file part into toAnthropicImageBlock(). For unsupported hints like image/svg+xml or image/bmp, that helper falls through to the "image/png" default and emits a bad Anthropic block instead of returning undefined, so the request fails even though this function promises unsupported types are omitted.
Suggested fix
if (mediaType.startsWith("image/")) {
+ if (!normalizeImageMediaType(mediaType)) {
+ return undefined;
+ }
return toAnthropicImageBlock(data, mediaType);
}📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| if (mediaType.startsWith("image/")) { | |
| return toAnthropicImageBlock(data, mediaType); | |
| } | |
| if (mediaType.startsWith("image/")) { | |
| if (!normalizeImageMediaType(mediaType)) { | |
| return undefined; | |
| } | |
| return toAnthropicImageBlock(data, mediaType); | |
| } |
🤖 Prompt for 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.
In `@src/lib/providers/anthropicImageBlocks.ts` around lines 232 - 234, The
`anthropicImageBlocks` flow is passing every `image/*` media type into
`toAnthropicImageBlock()`, which causes unsupported hints like `image/svg+xml`
or `image/bmp` to be mislabeled as PNG instead of being omitted. Update the
`mediaType.startsWith("image/")` branch to only call `toAnthropicImageBlock()`
for supported image types and return `undefined` for unsupported ones, so the
function’s behavior matches its contract. Use `toAnthropicImageBlock` and the
`mediaType.startsWith("image/")` check as the main places to adjust.
|
Superseded by #1117, which consolidates these fixes (structured-output, temperature, vision, exceljs, mammoth) plus the OAuth-proxy relocation + tracing fixes into a single commit per the single-commit policy. All review comments from this PR are addressed in #1117:
Closing in favour of #1117. |
Problem
experimental_output(AI-SDK JSON-schema enforcement) combined with tools silently drops tool calls on the native Anthropic Messages API surface (provideranthropic/bedrock, including via a proxy /ANTHROPIC_BASE_URLoverride).Symptoms observed on an anthropic-via-proxy deployment (Tara):
finishReason=tool-callsbut zero tool calls are parsed (has_tools=false), so the agent loop ends at step 1.Root cause
isToolsSchemaExclusionInForce()disables structured-output-with-tools only for Gemini; its comment assumed "Anthropic Claude … supports tools and structured output simultaneously." That holds for Vertex+Claude (different transport) but not for the native Anthropic surface.Fix
Extend the static gate to also exclude the native anthropic surface (
anthropic/bedrock) when tools are active, via a newisNativeAnthropicProvider()— mirroring the existing Gemini fallback. Vertex+Claude is intentionally left untouched (providervertex,claude-model), preserving strict structured output for the primary production config.Verification
End-to-end on an anthropic-via-proxy deployment with the equivalent change applied:
Downstream
validateOrRecoverStructured()recovers the structured shape from text mode, so there is no regression to the response format.Summary by CodeRabbit