Skip to content

fix: filter non-Anthropic content blocks before sending to Vertex API #9618

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
mrubens merged 4 commits into from
Nov 26, 2025
Merged

fix: filter non-Anthropic content blocks before sending to Vertex API #9618

mrubens merged 4 commits into from
Nov 26, 2025

Conversation

@hannesrudolph
Copy link
Collaborator

@hannesrudolph hannesrudolph commented Nov 26, 2025 •
edited
Loading

Summary

Fixes the context condensation error when using reasoning with Anthropic models on Vertex:

Input tag 'reasoning' found using 'type' does not match any of the expected tags

Root Cause

  1. Roo Code stores reasoning content internally with type: 'reasoning' (provider-agnostic format)
  2. Gemini stores thought signatures with type: 'thoughtSignature'
  3. When messages are replayed to Anthropic/Vertex APIs during context condensation or model switching, these non-standard types cause 400 errors

Solution

Added filterNonAnthropicBlocks() to both anthropic-vertex.ts and anthropic.ts handlers that uses an allowlist approach to only keep valid Anthropic content types.

Valid Anthropic content types (allowlist): text, image, tool_use, tool_result, thinking, redacted_thinking, document

This automatically filters out:

  • type: 'reasoning' - Internal Roo Code format
  • type: 'thoughtSignature' - Gemini's encrypted reasoning signature
  • Any other unknown/future block types

Implementation Details

This follows established patterns in the codebase:

  • Gemini handler (line 98-104): Filters type: 'reasoning' messages
  • gemini-format.ts (line 131-136): Skips unsupported blocks with console.warn
  • Claude Code handler: Uses filterMessagesForClaudeCode() for image blocks
  • Cerebras handler: Strips thinking tokens from assistant messages

Testing

  • Added tests for filtering internal reasoning blocks
  • Added tests for filtering empty messages after removing all reasoning blocks
  • Added test for filtering unknown future block types
  • All 48 tests pass

Related

Fixes #9583
Related to #9584

@dosubot dosubot bot added size:L This PR changes 100-499 lines, ignoring generated files. bug Something isn't working labels Nov 26, 2025
@roomote
Copy link
Contributor

roomote bot commented Nov 26, 2025 •
edited
Loading

Oroocle Clock   Follow along on Roo Cloud

Review complete. One issue was flagged related to Anthropic prompt caching model coverage.

  • Ensure claude-opus-4-5-20251101 is included in the prompt-caching switch and beta-header logic in src/api/providers/anthropic.ts so it uses the same caching behavior as the other Opus models.
Previous reviews

Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.

@hannesrudolph hannesrudolph added the Issue/PR - Triage New issue. Needs quick review to confirm validity and assign labels. label Nov 26, 2025
roomote[bot]
roomote bot reviewed Nov 26, 2025
View reviewed changes
hannesrudolph
hannesrudolph commented Nov 26, 2025
View reviewed changes
Copy link
Collaborator Author

@hannesrudolph hannesrudolph left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Review complete! The fix correctly addresses issue #9583 by filtering out non-Anthropic content blocks before sending to the API.

What was reviewed:

  • ✅ Filters reasoning and thoughtSignature blocks that cause API 400 errors
  • ✅ Proper handling of empty messages after filtering
  • ✅ Tests cover both mixed content and edge cases
  • ✅ Follows established patterns from gemini-format.ts

Additional improvements made during review:

  • Extracted filterNonAnthropicBlocks to shared utility src/api/transform/anthropic-filter.ts
  • Added 9 unit tests for the shared utility
  • Eliminates code duplication between anthropic.ts and anthropic-vertex.ts

All 48 tests pass. Ready to merge.

@hannesrudolph
fix: filter non-Anthropic content blocks before sending to Vertex API
308f966 
Fixes context condensation error when using reasoning with Anthropic models on Vertex:
'Input tag reasoning found using type does not match any of the expected tags'

The issue occurs because:
1. Roo Code stores reasoning content internally with type: 'reasoning'
2. Gemini stores thought signatures with type: 'thoughtSignature'
3. When messages are replayed to Anthropic/Vertex APIs, these non-standard types cause 400 errors

Solution:
- Added filterNonAnthropicBlocks() to anthropic-vertex.ts and anthropic.ts
- Filters out 'reasoning' (internal Roo format) and 'thoughtSignature' (Gemini format)
- Removes empty messages after filtering

This follows the same pattern used by:
- Gemini handler (filters 'reasoning' messages)
- gemini-format.ts (skips unsupported block types)
- Claude Code handler (filterMessagesForClaudeCode for images)

Fixes #9583
Related to #9584
@hannesrudolph
fix: restore claude-opus-4-5-20251101 prompt caching and remove dead …
ccdf0df 
…code

- Add claude-opus-4-5-20251101 back to switch statements for prompt caching support
- Remove unused lastThinkingSignature property and getThinkingSignature() method
@hannesrudolph
refactor: extract filterNonAnthropicBlocks to shared utility
f25c4c5 
- Created src/api/transform/anthropic-filter.ts with shared filtering logic
- Updated anthropic.ts and anthropic-vertex.ts to import from shared utility
- Added tests for the shared utility (9 new tests)
- Eliminates code duplication between the two handlers
@hannesrudolph hannesrudolph force-pushed the fix/9583-filter-reasoning-blocks-vertex branch from db4448f to f25c4c5 Compare November 26, 2025 20:40
@hannesrudolph hannesrudolph moved this from Triage to PR [Needs Prelim Review] in Roo Code Roadmap Nov 26, 2025
@hannesrudolph hannesrudolph added PR - Needs Preliminary Review and removed Issue/PR - Triage New issue. Needs quick review to confirm validity and assign labels. labels Nov 26, 2025
@dosubot dosubot bot added the lgtm This PR has been approved by a maintainer label Nov 26, 2025
@hannesrudolph hannesrudolph moved this from PR [Needs Prelim Review] to PR [Needs Review] in Roo Code Roadmap Nov 26, 2025
mrubens
mrubens reviewed Nov 26, 2025
View reviewed changes
@hannesrudolph
refactor: use allowlist instead of denylist for Anthropic block types
67129d5 
Address review feedback from mrubens to use an allowlist approach instead
of a denylist. This is more robust as any new/unknown block types will
automatically be filtered out.

VALID_ANTHROPIC_BLOCK_TYPES now explicitly lists accepted types:
- text, image, tool_use, tool_result, thinking, redacted_thinking, document
@mrubens mrubens merged commit fb9c57e into main Nov 26, 2025
10 checks passed
@mrubens mrubens deleted the fix/9583-filter-reasoning-blocks-vertex branch November 26, 2025 22:47
@github-project-automation github-project-automation bot moved this from PR [Needs Review] to Done in Roo Code Roadmap Nov 26, 2025
@github-project-automation github-project-automation bot moved this from New to Done in Roo Code Roadmap Nov 26, 2025
@cardil
Copy link

cardil commented Nov 27, 2025

🎉 Confirmed.

After upgrading to v3.34.6 I was able to resume the previously failing thread, and it properly condenses.

Screenshot From 2025-11-27 01-08-30

mini2s added a commit to zgsm-ai/costrict that referenced this pull request Nov 27, 2025
@mini2s @daniel-lxs
@cte @roomote @hannesrudolph @roomote @mrubens @github-actions @ellipsis-dev @brunobergher @jr @AlexKer @sebastiand-cerebras @ggoranov-smar @Smartsheet-JB-Brown
Roo to main (#712