fix: preserve gemini thought signatures by ibetitsmike · Pull Request #25933 · coder/coder

ibetitsmike · 2026-06-01T18:01:43Z

AI Bridge reserializes OpenAI chat-completions requests before sending them upstream. For Gemini OpenAI-compatible routes, that OpenAI typed-parameter round trip drops tool_calls[].extra_content.google.thought_signature, so Google rejects tool-result continuations with Function call is missing a thought_signature.

This PR:

patches the AI Bridge upstream serialization boundary for Gemini OpenAI-compatible chat completions
shares the Gemini thought-signature patching helpers with chatd's OpenAI-compatible transport patch to keep behavior consistent
treats direct Google OpenAI-compatible upstream endpoints as Gemini-scoped even when the request model is an alias
adds the Google fallback thought signature to every assistant tool call in the active turn, including parallel tool calls
covers the regression that extra_content is dropped before the upstream body is patched

Mux updated this PR description on behalf of Mike.

ibetitsmike · 2026-06-01T20:15:48Z

/coder-agents-review

coder-agents-review · 2026-06-01T20:15:53Z

Chat: Review posted | View chat
Requested: 2026-06-01 21:42 UTC by @ibetitsmike
Spend: $75.39 / $100.00

Review history

R1 (2026-06-01): 16 reviewers, 4 Nit, 1 P2, 7 P3, 1 P4, COMMENT. Review
R2 (2026-06-01): 15 reviewers, 7 Nit, 1 P2, 8 P3, 1 P4, APPROVE. Review
R3 (2026-06-01), 7 Nit, 1 P2, 8 P3, 1 P4, COMMENT. Review
R4 (2026-06-01): 6 reviewers, 7 Nit, 1 P2, 8 P3, 1 P4, APPROVE. Review

deep-review v0.6.1 | Round 4 | fc01aee..6eae190

Last posted: Round 4, 17 findings (1 P2, 8 P3, 1 P4, 7 Nit), APPROVE. Review

Finding inventory

Findings

#	Sev	Status	Location	Summary	Round	Reviewer	Posted
CRF-1	P2	Author fixed (`c33e4a0`)	google_openai_compat.go:100	Five functions + constant duplicated from chatd/chatprovider with divergent parallel tool call behavior	R1	Robin P2, Mafuuu P2, Mafu-san P2, Zoro P2	Yes
CRF-2	P3	Author fixed (`c33e4a0`)	google_openai_compat.go:76	Current-turn-only patching is correct per Google docs but the assumption is undocumented	R1	Pariston P2, Kite P2, Meruem P2, Razor Note (confirms correctness)	Yes
CRF-3	P3	Author fixed (`c33e4a0`)	google_openai_compat.go:11	Constant has no provenance comment; chatd copy has doc link to Google thought-signatures page	R1	Leorio P3, Mafu-san P3, Zoro P3, Razor P3	Yes
CRF-4	P3	Author fixed (`c33e4a0`)	google_openai_compat.go:1	134 lines of vendor workaround with zero file-level or function-level comments	R1	Leorio P2	Yes
CRF-5	P3	Author fixed (`c33e4a0`)	blocking.go:289	Blocking path passes full typed params then overrides with WithRequestBody; streaming path passes empty params	R1	Hisoka P3, Ryosuke P3	Yes
CRF-6	P3	Author fixed (`c33e4a0`)	google_openai_compat.go:62	Host check pinned to generativelanguage.googleapis.com; Vertex AI endpoints not covered	R1	Meruem P3, Razor P2	Yes
CRF-7	P3	Author fixed (`c33e4a0`)	google_openai_compat_internal_test.go:53	No test exercises isAssistantRole("model"), the Gemini-specific role value	R1	Bisky P3	Yes
CRF-8	P3	Author fixed (`c33e4a0`)	google_openai_compat_internal_test.go:12	No multi-turn conversation test to document current-turn-only scoping	R1	Bisky P3	Yes
CRF-9	P4	Author fixed (`c33e4a0`)	google_openai_compat.go:21	chatCompletionRequestOptions (blocking path entry point) has no direct test	R1	Bisky P4, Chopper P3	Yes
CRF-10	Nit	Author fixed (`c33e4a0`)	google_openai_compat_internal_test.go:70	Guard test uses positional assertions without named subtests	R1	Chopper Nit, Kite P3	Yes
CRF-11	Nit	Author fixed (`c33e4a0`)	google_openai_compat.go:39	Silent error fallback returns unpatched body without logging	R1	Chopper P3, Meruem Nit	Yes
CRF-12	Nit	Author fixed (`c33e4a0`)	streaming.go:186	Comment now incomplete; does not mention thought signature patching added by chatCompletionRequestBody	R1	Leorio Nit, Gon P2	Yes
CRF-13	Nit	Author fixed (`c33e4a0`)	google_openai_compat.go:113	isAssistantRole checks for "model" without explaining why	R1	Leorio Nit	Yes
CRF-14	Note	Dropped by orchestrator (informational, no code change needed)	google_openai_compat.go:11	Google documents that skip_thought_signature_validator degrades reasoning quality	R1	Razor Note, Hisoka Note	No
CRF-15	P3	Author fixed (`6eae190`)	thought_signature_test.go:12	ShouldPatchOpenAICompatRequest has no unit test at home package level; behavioral difference from ShouldPatchGoogleUpstreamRequest untested	R2	Netero P3	Yes
CRF-16	P3	Author fixed (PR title update)	PR title	PR title says "tool signatures" but code and docs consistently say "thought signatures"	R2	Leorio P3	Yes (body)
CRF-17	Nit	Author fixed (`6eae190`)	thought_signature.go:34	Five helpers exported without external callers; test file needs exports but could use internal test package	R2	Meruem Nit, Ryosuke Nit, Zoro P3, Knov Nit	Yes
CRF-18	Nit	Author fixed (`6eae190`)	thought_signature.go:76	Doc comments on AddThoughtSignaturesToLatestTurn, IsAssistantRole, EnsureThoughtSignature have restating first sentences	R2	Gon P2	Yes
CRF-19	Nit	Author fixed (`6eae190`)	google_openai_compat.go:34	append([]option.RequestOption{}, opts...) should be slices.Clone(opts)	R2	Ging-Go Nit	Yes
CRF-20	Nit	Dropped by orchestrator (common Go test pattern, cross-package helpers have trade-offs)	thought_signature_test.go:111	Test helper thoughtSignature reimplemented in 3-4 places across packages	R2	Robin Nit, Zoro Nit	No
CRF-21	Nit	Dropped by orchestrator (intentional, tested, confirmed correct by Hisoka and Pariston)	thought_signature.go:30	ShouldPatchGoogleUpstreamRequest requires model prefix for direct endpoints while sibling does not	R4	Mafuuu P3	No

Round log

Round 1

Panel. Netero: 0 findings. Panel (16 reviewers): 1 P2, 7 P3, 1 P4, 4 Nit, 1 Note (dropped). Reviewed against fc01aee..6cfa027.

Round 2

All 13 R1 findings addressed in c33e4a0. Netero: 1 P3 (CRF-15). Panel (15 reviewers): 2 P3, 3 Nit, 1 Nit (dropped). 7 reviewers reported no findings. Reviewed against fc01aee..c33e4a0.

Round 3

BLOCKED. CRF-15/17/18/19 addressed in 6eae190. CRF-16 (PR title terminology) silent. No review.

Round 4

CRF-16 addressed via PR title update. All findings resolved. Netero: 0 findings. Panel (6 reviewers): 1 Nit (dropped, intentional per Hisoka+Pariston). 5 reported no findings. Reviewed against fc01aee..6eae190.

About deep-review

CRF = Coder Review Finding (P0-P4, Nit, Note)

Reviewer	Focus
Bisky	tests
Chopper	ops/errors
Churn-guard	change verification
Ging	language modernization
Gon	naming
Hisoka	edge cases
Killua	perf
Kite	change integrity
Knov	contracts
Knuckle	SQL
Kurapika	security
Law	decomposition
Leorio	docs
Luffy	product
Mafu-san	process
Mafuuu	contracts
Melody	dispatch/pairing
Meruem	structural
Nami	frontend
Netero	mechanical checks
Pariston	premise testing
Pen-botter	product gaps
Razor	verification
Robin	duplication
Ryosuke	Go arch
Takumi	concurrency
Zoro	shape

🤖 Managed by Coder Agents.

coder-agents-review

Well-targeted fix for the Gemini thought-signature round-trip loss. The root cause diagnosis is solid: the test at line 42 proves the SDK drops extra_content, and the patch injects Google's documented bypass value at the serialization boundary. The approach is proportional, the guard logic is clean (host + model double-check), and ensureGoogleOpenAICompatThoughtSignature is properly idempotent.

Severity breakdown: 1 P2, 7 P3, 1 P4, 4 Nit.

The P2 is code duplication. Five functions and one constant are copied from coderd/x/chatd/chatprovider/openai_compat_patches.go (4 days old, same module), and the two implementations diverge on parallel tool calls: this PR patches all tool calls per assistant message, while the chatd version patches only the first. If patching all is correct (as the PR description states), the chatd version has a sibling bug for parallel tool calls.

Multiple reviewers (8+) flagged the current-turn-only scoping in addGoogleOpenAICompatThoughtSignatures as potentially incorrect for multi-turn conversations. Razor resolved this with evidence from Google's thought-signatures documentation: "Only current turn is required; we don't validate on previous turns." The scoping is correct, but the assumption is undocumented in the code.

"The agent used an existing verified value (trust signal: not fabricated), but dropped the trail that makes it verifiable by others." (Mafu-san)

🤖 This review was automatically generated with Coder Agents.

ibetitsmike · 2026-06-01T20:57:00Z

/coder-agents-review

coder-agents-review

All 13 Round 1 findings addressed thoroughly. The extraction to internal/googleopenai/ resolves the duplication (CRF-1), fixes the chatd parallel tool call bug, and centralizes the thought-signature logic with proper documentation. The blocking path now correctly passes empty params when the body is overridden (CRF-5). Every documentation gap from R1 is filled.

Round 2 severity: 2 P3, 3 Nit. 7 of 15 reviewers reported no new findings.

P3 [CRF-16] The PR title says "tool signatures" but all code, tests, and documentation consistently say "thought signatures." Future git log --grep=thought.signature searches will miss this PR. Consider updating to "fix: preserve gemini thought signatures." (Leorio P3)

"Five helpers exported without external callers. The test file forces the exports, but EnsureThoughtSignature without AddThoughtSignaturesToLatestTurn is a loaded gun without the safety." (Knov)

🤖 This review was automatically generated with Coder Agents.

ibetitsmike · 2026-06-01T21:38:14Z

/coder-agents-review

coder-agents-review

CRF-15, CRF-17, CRF-18, and CRF-19 addressed in 6eae190. Thank you.

Further review is blocked on one unaddressed finding:

CRF-16 (P3): The PR title says "tool signatures" but all code, tests, and documentation consistently say "thought signatures." This was raised in the Round 2 review body. Please either update the PR title to "fix: preserve gemini thought signatures" or respond with why the current title is intentional.

🤖 This review was automatically generated with Coder Agents.

ibetitsmike · 2026-06-01T21:42:31Z

/coder-agents-review

coder-agents-review

All 21 findings resolved or dropped across 4 rounds. Netero clean. Panel clean (5 of 6 no findings; 1 Nit dropped as intentional per independent confirmation from 2 prior reviewers).

This PR correctly diagnoses the Gemini thought-signature round-trip loss through the OpenAI SDK, patches it at the narrowest possible point in the serialization boundary, extracts shared logic into internal/googleopenai/ to keep both callers (aibridge and chatd) in sync, and fixes the pre-existing chatd bug for parallel tool calls. Test density is 58.8% with targeted behavioral tests that double as living documentation.

🤖 This review was automatically generated with Coder Agents.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1969d97d8c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

ibetitsmike · 2026-06-04T11:04:20Z

@codex review

chatgpt-codex-connector · 2026-06-04T11:08:26Z

Codex Review: Didn't find any major issues. Swish!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

ssncferreira · 2026-06-05T15:56:37Z

+	return patched, nil
+}
+
+func (i *interceptionBase) chatCompletionRequestOptions(opts []option.RequestOption) ([]option.RequestOption, bool, error) {


nit: would be nice to have a short summary explaining the returned values. Same on the above 🙂

ssncferreira · 2026-06-05T16:33:08Z

+	if !googleopenai.ShouldPatchGoogleUpstreamRequest(i.cfg.BaseURL, i.req.Model) {
+		return body, nil
+	}
+	patched, _, err := googleopenai.PatchThoughtSignatures(body)


IIUC, for every request, we are checking if the request is going to Google Gemini. Instead of doing this every time at the interception layer, we could do it at the provider layer once. A Google provider (in this case, a OpenAI provider with base url "generativelanguage.googleapis.com") will always be sending request to "generativelanguage.googleapis.com" , and vice-versa a non-google provider wll never send request to "generativelanguage.googleapis.com" .

ssncferreira · 2026-06-05T16:48:09Z

+// history in body. It returns changed=false when no patch is needed.
+func PatchThoughtSignatures(body []byte) ([]byte, bool, error) {
+	var payload map[string]any
+	if err := json.Unmarshal(body, &payload); err != nil {


We are marshalling ChatCompletionNewParams to byte[], then here unmarshalling to map, and then again marshalling to map 😅 unfortunately, give the sdk struct, I don't think there is a better alternative

Port parseProviderBaseURL logic from main into internal/googleopenai. Handles bare hostnames without schemes and empty strings.

ssncferreira

LGTM with some nits, but nothing blocking

ssncferreira · 2026-06-09T11:08:47Z

+	if isDirectGeminiOpenAIEndpoint(baseURL) {
+		return true
+	}
+	return isCoderAIBridgeEndpoint(baseURL) && isGeminiModelID(modelID)


I think this else branch is unnecessary, because whenever requests go through AI Gateway, it already performs the patch. IIUC, the flow would be:

Coder agent makes a request to Gemini via AIBridge

isCoderAIBridgeEndpoint(baseURL) && isGeminiModelID(modelID) returns true

AIBridge removes the though signature due to the SDK

AIBridge does chatCompletionRequestOptions which patches the request body to include the thought signature again.

github-actions Bot assigned ibetitsmike Jun 1, 2026

ibetitsmike force-pushed the mike/no-workspace-terminal-tool-errors branch from 4879cd2 to e555857 Compare June 1, 2026 19:03

ibetitsmike changed the title ~~fix(coderd/x/chatd): stop no-workspace tool continuations~~ fix(coderd/x/chatd/chatprovider): patch gemini bridge tool calls Jun 1, 2026

ibetitsmike force-pushed the mike/no-workspace-terminal-tool-errors branch from e555857 to 6cfa027 Compare June 1, 2026 19:39

ibetitsmike changed the title ~~fix(coderd/x/chatd/chatprovider): patch gemini bridge tool calls~~ fix(aibridge/intercept/chatcompletions): preserve gemini tool signatures Jun 1, 2026

coder-agents-review Bot reviewed Jun 1, 2026

View reviewed changes

ibetitsmike force-pushed the mike/no-workspace-terminal-tool-errors branch from 6cfa027 to c33e4a0 Compare June 1, 2026 20:56

ibetitsmike changed the title ~~fix(aibridge/intercept/chatcompletions): preserve gemini tool signatures~~ fix: preserve gemini tool signatures Jun 1, 2026

coder-agents-review Bot approved these changes Jun 1, 2026

View reviewed changes

Comment thread internal/googleopenai/thought_signature_test.go

Comment thread internal/googleopenai/thought_signature.go Outdated

Comment thread internal/googleopenai/thought_signature.go Outdated

Comment thread aibridge/intercept/chatcompletions/google_openai_compat.go Outdated

coder-agents-review Bot reviewed Jun 1, 2026

View reviewed changes

ibetitsmike changed the title ~~fix: preserve gemini tool signatures~~ fix: preserve gemini thought signatures Jun 1, 2026

coder-agents-review Bot approved these changes Jun 1, 2026

View reviewed changes

ibetitsmike force-pushed the mike/no-workspace-terminal-tool-errors branch from 6eae190 to 440e251 Compare June 2, 2026 08:10

dannykopping requested a review from ssncferreira June 2, 2026 08:41

ibetitsmike marked this pull request as ready for review June 4, 2026 10:07

ibetitsmike force-pushed the mike/no-workspace-terminal-tool-errors branch from 9cc161b to 1969d97 Compare June 4, 2026 10:07

chatgpt-codex-connector Bot reviewed Jun 4, 2026

View reviewed changes

Comment thread internal/googleopenai/thought_signature.go Outdated

ssncferreira reviewed Jun 5, 2026

View reviewed changes

ibetitsmike and others added 4 commits June 8, 2026 11:23

fix: preserve gemini tool signatures

5d5962f

fix: tighten gemini signature helpers

2235b0a

fix(internal/googleopenai): patch google upstream aliases

478d0ff

fix: use robust URL parsing for Google endpoint detection

b06967c

Port parseProviderBaseURL logic from main into internal/googleopenai. Handles bare hostnames without schemes and empty strings.

ssncferreira force-pushed the mike/no-workspace-terminal-tool-errors branch from d4809b3 to b06967c Compare June 8, 2026 11:30

ssncferreira self-assigned this Jun 8, 2026

chore: address review comments

10cd6d8

ssncferreira approved these changes Jun 9, 2026

View reviewed changes

ssncferreira merged commit c349ea6 into main Jun 9, 2026
28 checks passed

ssncferreira deleted the mike/no-workspace-terminal-tool-errors branch June 9, 2026 11:11

github-actions Bot locked and limited conversation to collaborators Jun 9, 2026

ssncferreira added the backport label Jun 9, 2026

Conversation

ibetitsmike commented Jun 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ibetitsmike commented Jun 1, 2026

Uh oh!

coder-agents-review Bot commented Jun 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Finding inventory

Findings

Round log

Round 1

Round 2

Round 3

Round 4

Uh oh!

coder-agents-review Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ibetitsmike commented Jun 1, 2026

Uh oh!

coder-agents-review Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ibetitsmike commented Jun 1, 2026

Uh oh!

coder-agents-review Bot left a comment

Choose a reason for hiding this comment

Uh oh!

ibetitsmike commented Jun 1, 2026

Uh oh!

coder-agents-review Bot left a comment

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

ibetitsmike commented Jun 4, 2026

Uh oh!

chatgpt-codex-connector Bot commented Jun 4, 2026

Uh oh!

Uh oh!

ssncferreira Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

ssncferreira Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ssncferreira Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

ssncferreira left a comment

Choose a reason for hiding this comment

Uh oh!

ssncferreira Jun 9, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

ibetitsmike commented Jun 1, 2026 •

edited

Loading

coder-agents-review Bot commented Jun 1, 2026 •

edited

Loading