[xeophon]

Overview

Use the provider SDKs for inference transport and stream assembly, with dialects projecting native requests and responses into Verifiers traces. Keep evaluation and training clients separate and retain the interception server's rollout, cancellation, and replay ownership.

Details

• Share sampling overrides and Chat stream accumulation, while preserving each provider's authentication, terminal events, usage extensions, error statuses, and reasoning state.
• Preserve native content and provider metadata through task-hook edits using per-message setters and per-part provenance.
• Keep network capability mediation separate from wire-format translation, with conservative handling of provider domain filters.

Note

Rework dialects around native dict responses, setters, and SDK transport

Simplifies the Chat, Responses, and Anthropic dialects by removing per-dialect SSE parsers, Pydantic response models, and local capability mediation. Responses are now passed around as native dictionaries, streamed completions are assembled through shared SDK readers (read_chat_completion, AsyncMessageStream), and request edits go through setter callbacks returned by parse_request.

• The eval client now sends requests through OpenAI and Anthropic SDK clients instead of a hand-rolled HTTPX relay; provider auth comes from the selected SDK, and error status mapping covers Anthropic, httpx, and connection-reset failures in errors.py.
• Adds NativeContentPart so unrecognized provider content blocks are preserved instead of dropped, and exports it from verifiers.v1 in types.py.
• Network-policy request mediation moves from dialect methods to a reusable RequestFilter base in policy.py.
• Risk: parse_request now returns a (Request, setters) tuple instead of just a Request, so all callers (e.g. train.py, taskset.py, server.py) must unpack it; StreamParser and the removed dialect filter classes are no longer importable from verifiers.v1.dialects.

^{Macroscope summarized 7022952.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-22T22:43:54.431930Z	7022952	New commits

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 97fbff7ea6

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".

[cursor]

Stale Bugbot comment from a previous run.

[macroscopeapp]

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — This PR replaces the inference relay with SDK-backed OpenAI/Anthropic transports and rewires streaming, request rewriting, authentication filtering, and error handling. It also introduces a substantial network-policy mediation module affecting provider tools and remote content, creating broad production and security-sensitive impact.

^{You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Macroscope has since reviewed this pull request. An earlier review was skipped by a cost limit; a review has now completed, so that notice no longer applies.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 60c20f9074

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".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 99b2ebee17

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".

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[cursor]

Stale Bugbot comment from a previous run.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c777d364fa

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".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2704d26f8b

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".

[cursor]

Stale Bugbot comment from a previous run.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 242fac6055

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".

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 242fac6. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

verifiers/verifiers/v1/dialects/chat.py

Lines 194 to 196 in 7022952

	native["content"] = patch_content(
	message.content, _content_to_wire(message.content)
	)

Pass the original content when patching scalar rewrites

When a request interceptor replaces a one-part Chat content list with a scalar string, this call omits the original native content passed as patch_content's third argument. Consequently the scalar branch that preserves a single text block's native metadata never runs, so fields such as prompt-cache controls are dropped and the upstream request changes from a content-part list to plain text. The Responses setters at responses.py:104-110 have the same issue; pass each setter's original native content into patch_content.

---

verifiers/verifiers/v1/utils/artifacts.py

Line 177 in 7022952

declared = {str(PurePosixPath(workdir) / a.source): a for a in artifacts}

Resolve saved sidecar artifacts with their own workdirs

For a Harbor Compose artifact whose service is a sidecar and whose source is relative, collection resolves the source against that sidecar's inspected WorkingDir, but this lookup resolves every declaration against the main agent workdir supplied by write_episode. If the workdirs differ, the collected absolute source misses declared, so save_artifacts falls back to a generic artifact and silently loses the configured destination and service in the saved tree and manifest. Preserve the resolved declaration per collected source or resolve each service's artifacts with that service's workdir.

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".