fix(responses): preserve DeepSeek reasoning across tool turns by SantiagoDePolonia · Pull Request #623 · ENTERPILOT/GoModel

SantiagoDePolonia · 2026-07-31T13:45:58Z

Summary

preserve raw reasoning content when converting Chat Completions output into Responses API events and items
replay reasoning content only where providers require it for assistant tool-call messages, including DeepSeek
normalize developer messages, usage fields, output indexes, and late reasoning deltas for OpenAI-compatible behavior
keep the streaming conversion hot path within the existing allocation and byte thresholds

Validation

go test ./...
repository pre-commit suite: formatting, race tests, tidy check, performance guard, fix-check, and lint

Summary by CodeRabbit

New Features
- Added support for displaying streamed reasoning content in Responses API outputs.
- Reasoning now appears in the correct order before assistant messages and tool calls.
- Reasoning is preserved when assistants invoke tools, improving continuity across tool-call workflows.
- Developer messages are normalized for broader compatibility.
Bug Fixes
- Improved handling of unsupported and malformed input and usage data.
- Prevented late reasoning updates from disrupting response ordering.
- Standardized token usage details in converted responses.

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

coderabbitai · 2026-07-31T13:46:21Z

📝 Walkthrough

Walkthrough

The PR adds reasoning replay for tool-call inputs, reasoning output items for non-streaming responses, and reasoning delta handling for streaming Responses conversion. It also normalizes usage fields and adds coverage for ordering, suppression, and malformed usage.

Changes

Reasoning content support

Layer / File(s)	Summary
Input reasoning replay and message normalization `internal/providers/responses_input.go`, `internal/providers/responses_adapter_test.go`, `internal/providers/deepseek/deepseek_test.go`	Reasoning items are extracted and attached to assistant tool-call messages as `reasoning_content`. Standalone reasoning is omitted. Developer roles become `system`. Tests cover message merging and DeepSeek request translation.
Reasoning output items `internal/providers/responses_output_state.go`, `internal/providers/responses_output.go`, `internal/providers/responses_adapter_test.go`	Reasoning output items receive IDs, reasoning text, summary data, delta events, and completion events. Non-streaming output places reasoning before assistant content.
Streaming conversion and usage mapping `internal/providers/responses_converter.go`	Streaming conversion reserves reasoning output indexes, emits deltas, completes reasoning before assistant or tool-call output, suppresses late reasoning, and maps valid usage fields to Responses fields.
Streaming regression coverage `internal/providers/responses_converter_test.go`, `tests/contract/testdata/golden/groq/responses_stream.golden.json`	Tests cover reasoning order, late reasoning, tool-call transitions, usage details, normalized field names, malformed usage omission, and the updated golden response.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant ModelStream
  participant ResponsesStreamConverter
  participant ResponsesOutputEventState
  ModelStream->>ResponsesStreamConverter: send reasoning_content delta
  ResponsesStreamConverter->>ResponsesOutputEventState: reserve and append reasoning
  ResponsesOutputEventState-->>ResponsesStreamConverter: emit reasoning delta event
  ModelStream->>ResponsesStreamConverter: send assistant or tool-call delta
  ResponsesStreamConverter->>ResponsesOutputEventState: complete reasoning
  ResponsesStreamConverter->>ResponsesOutputEventState: start later output item

Possibly related PRs

ENTERPILOT/GoModel#285: Both changes cover reasoning-content replay in DeepSeek Responses translation.
ENTERPILOT/GoModel#581: Both changes modify OpenAI-compatible Responses streaming conversion.

Suggested reviewers: copilot

Poem

A rabbit sends reasoning through the stream,
Then places tool calls in order.
Index zero holds the thought,
Usage fields take their proper names.
Tests guard every transition.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly identifies the main change: preserving DeepSeek reasoning across tool-call turns.
Description check	✅ Passed	The description explains the changes, rationale, affected behavior, and validation steps, although it uses Summary instead of the template's Description heading.

✨ Finishing Touches

📝 Generate docstrings

Create stacked PR
Commit on current branch

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch fix/codex2

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

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

codecov-commenter · 2026-07-31T13:51:13Z

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 80.80357% with 43 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/providers/responses_input.go	60.97%	24 Missing and 8 partials ⚠️
internal/providers/responses_converter.go	85.48%	5 Missing and 4 partials ⚠️
internal/providers/responses_output.go	90.90%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

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 `@internal/providers/responses_converter.go`:
- Around line 488-514: Update chatUsageToResponsesUsage to remove the
bytes.Contains checks and represent the three required token counts as pointer
fields in the local JSON-unmarshal struct. Unmarshal once, return false when any
pointer is nil or unmarshalling fails, and dereference the validated pointers
when constructing responsesStreamUsage; preserve the existing omission behavior
for malformed usage.

🪄 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: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 87c6d21e-82dc-4ba0-9dc0-7b783c8a6a3b

📥 Commits

Reviewing files that changed from the base of the PR and between 61493f4 and 0ee3a91.

📒 Files selected for processing (7)

internal/providers/deepseek/deepseek_test.go
internal/providers/responses_adapter_test.go
internal/providers/responses_converter.go
internal/providers/responses_converter_test.go
internal/providers/responses_input.go
internal/providers/responses_output.go
internal/providers/responses_output_state.go

greptile-apps · 2026-07-31T13:57:07Z

Confidence Score: 3/5

T-Rex Logs

What T-Rex did

T-Rex produced a finding-proof for a posted P1 finding and linked it to the corresponding review comment.
T-Rex produced a second finding-proof for another posted P1 finding and linked it to its review comment.
T-Rex performed a general contract validation for an index-collision scenario, capturing the before and after data and listing the observed events (reasoning, function_call, and assistant message outputs).

_{Ran code and verified through T-Rex}

Comments Outside Diff (2)

internal/providers/responses_converter.go, line 146-155 (link)

Started tool call and later assistant message share Responses output index

When a stream emits reasoning, starts a tool call, and then emits assistant text, reserveAssistantOutput keeps the assistant at the post-reasoning index while only shifting tool calls that have not started. The already-started function call therefore remains at output index 1 and the later assistant message is also emitted at output index 1. Responses clients receive two distinct output items with the same output_index, breaking ordered output-item processing. Allocate a distinct assistant index after started tool calls, or otherwise prevent this event ordering from producing colliding items.

Context Used: CLAUDE.md (source)
Artifacts

Evidence from the check
- Captured the exact Go test source used to feed the converter the claimed event order, providing the reproducible input.
Command output from the check
- Executed the focused converter test; it records reasoning at index 0, function call at index 1, and assistant message at index 1 before failing, confirming the collision.
_{Ran code and verified through T-Rex}
General comment

Started tool call and later assistant message share Responses output index
- Bug
  - For an upstream stream that emits reasoning, starts tool call 0, then emits assistant content, the converter produces response.output_item.added for the function call and assistant message with output_index: 1. Responses output items therefore collide at the same output index.
- Cause
  - reserveAssistantOutput at internal/providers/responses_converter.go:146-155 increments indexes only for tool-call states that have not started. Once handleToolCallDeltas starts the tool call at index 1, subsequent assistant text reserves its fixed post-reasoning index 1 without moving the started tool call or allocating a distinct assistant index.
- Fix
  - Allocate the assistant output index according to preceding/started tool-call output items, or reject/otherwise serialize assistant text after a started tool call. Add the executed reasoning → tool-call → text regression case to the permanent converter suite.
_{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "fix(responses): preserve reasoning acros..." | Re-trigger Greptile}

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

SantiagoDePolonia · 2026-07-31T14:08:18Z

Addressed the review findings in 7abcc27:

reasoning → started function call → assistant text now assigns output indexes 0, 1, and 2 respectively; emitted items are never renumbered
usage-field presence is validated from the decoded top-level object without pointer allocations
added regression coverage for both cases
updated the Groq Responses-stream contract golden to expect Responses usage field names

Validation passed: go test ./..., contract replay tests, race tests, lint, and the hot-path performance guard (87 allocations/op, 10,296 bytes/op).

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 `@internal/providers/responses_converter_test.go`:
- Around line 286-315: Convert the new assistant-after-started-tool case at
internal/providers/responses_converter_test.go:286-315 into an entry in the
relevant table-driven converter test, preserving its expected output indexes.
Also convert the nested required usage-field omission case at
internal/providers/responses_converter_test.go:597-620 into an entry in the
appropriate table-driven test, retaining its expected behavior.

🪄 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: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 3fb82e59-1a06-4c00-891f-01e7d933cb0d

📥 Commits

Reviewing files that changed from the base of the PR and between 0ee3a91 and 7abcc27.

📒 Files selected for processing (3)

internal/providers/responses_converter.go
internal/providers/responses_converter_test.go
tests/contract/testdata/golden/groq/responses_stream.golden.json

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

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 `@internal/providers/responses_converter_test.go`:
- Around line 269-285: Update the test’s event tracking around the
response.output_item.added handling to preserve insertion order in an addedTypes
slice instead of relying only on addedIndexes. Add the expected added-item
sequence to each table-driven case, then assert the observed sequence; for
“assistant after started tool call,” require reasoning, function_call, then
message while retaining the existing index and reasoning assertions.

🪄 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: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 0fcf1267-4ceb-488e-9cd1-7c4207de3dc2

📥 Commits

Reviewing files that changed from the base of the PR and between 7abcc27 and ac8ccf5.

📒 Files selected for processing (1)

internal/providers/responses_converter_test.go

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

fix(responses): preserve reasoning across Codex tool turns

0ee3a91

Copilot AI review requested due to automatic review settings July 31, 2026 13:45

Copilot AI reviewed Jul 31, 2026

coderabbitai Bot reviewed Jul 31, 2026

View reviewed changes

Comment thread internal/providers/responses_converter.go

fix(responses): keep streamed output indexes unique

7abcc27

Copilot AI review requested due to automatic review settings July 31, 2026 14:07

Copilot AI reviewed Jul 31, 2026

coderabbitai Bot reviewed Jul 31, 2026

View reviewed changes

Comment thread internal/providers/responses_converter_test.go Outdated

test(responses): table-drive converter regressions

ac8ccf5

Copilot AI review requested due to automatic review settings July 31, 2026 14:16

Copilot AI reviewed Jul 31, 2026

coderabbitai Bot reviewed Jul 31, 2026

View reviewed changes

Comment thread internal/providers/responses_converter_test.go

test(responses): assert streamed item order

1e83ed9

Copilot AI review requested due to automatic review settings July 31, 2026 14:21

Copilot AI reviewed Jul 31, 2026

SantiagoDePolonia changed the title ~~fix(responses): preserve reasoning across Codex tool turns~~ fix(responses): preserve DeepSeek reasoning across tool turns Jul 31, 2026

SantiagoDePolonia merged commit c0b5d40 into main Jul 31, 2026
19 checks passed

SantiagoDePolonia deleted the fix/codex2 branch July 31, 2026 15:13

coderabbitai Bot mentioned this pull request Jul 31, 2026

fix(responses): tolerate unsupported Codex tools #625

Merged

Uh oh!

Uh oh!

Conversation

SantiagoDePolonia commented Jul 31, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Validation

Summary by CodeRabbit

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Sequence Diagram(s)

Uh oh!

codecov-commenter commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

greptile-apps Bot commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 3/5

T-Rex Logs

Comments Outside Diff (2)

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

SantiagoDePolonia commented Jul 31, 2026

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

SantiagoDePolonia commented Jul 31, 2026 •

edited by coderabbitai Bot

Loading

coderabbitai Bot commented Jul 31, 2026 •

edited

Loading

codecov-commenter commented Jul 31, 2026 •

edited

Loading

greptile-apps Bot commented Jul 31, 2026 •

edited

Loading