Add interception example environments

[xeophon]

Overview

This adds four small runnable environments that show the intended @vf.intercept interface directly on a task.

interception-v1 shows deterministic and ordinary LLM judge guards that rewrite matching messages, plus a termination guard that ends the rollout without a replacement message. The judge can use the canonical trace.

grayscale-interception-v1 shows typed user-message rewriting. It converts every inline image in the user message to grayscale with Pillow, preserves the other content parts, and returns a rewritten UserMessage before the request reaches the model.

bash-interception-v1 shows that an AssistantMessage interceptor can replace a Bash tool call before the Bash harness executes it. It records whether the call was intercepted and whether its sentinel side effect occurred.

web-search-interception-v1 enables Codex's provider-native web search. Its AssistantMessage interceptor detects a completed web_search_call, inspects the search-backed response, and terminates when the response contains a blocked word. Because intercepted responses are buffered, the blocked response does not reach the Codex harness. The provider has already executed the search at that point.

What this PR adds

• Deterministic and judge-backed content guards.
• A termination guard with a final reward.
• A Bash example that blocks a local tool call before execution.
• A Pillow example that rewrites user images before the model request.
• A Codex example that blocks a native web-search response before the harness receives it.
• A small Codex harness option for enabling native web search.
• Example-only dependencies, with no root dependency change.
• Editable example installation in the existing V1 live job.

Stack

This is PR 4 of 4. The stack should merge in order:

1. Serialize model exchanges within a rollout #2164 — Serialize model exchanges and buffer intercepted streams. Establishes one active exchange per rollout and buffers streams only when response interception is active.
2. Add typed model interception #2165 — Add typed @vf.intercept and graph-owned rewrites. Adds typed request and response interception and stores canonical rewrites in the trace graph.
3. Make interception termination final #2166 — Add Terminate and make its reward final. Adds non-retryable early termination and prevents later scoring from changing the terminal reward.
4. Add interception example environments #2178 — Add interception examples. Shows deterministic, judge-backed, termination, image-rewrite, Bash-tool, and Codex native-search interception.

Note

Add interception example environments for bash, grayscale, web search, and judge-based blocking

Adds four new example environments under environments/ to demonstrate the @vf.intercept() API across different interception patterns:

• interception_v1: Shows three intercept handlers — deterministic string blocking, early termination on a marker, and async judge-based blocking with configurable OpenAI/PRIME credentials.
• bash_interception_v1: Intercepts a specific bash tool call before it executes a sentinel command and rewards successful blocking.
• grayscale_interception_v1: Rewrites inline base64 image data URLs to grayscale using Pillow before they reach the model.
• web_search_interception_v1: Terminates rollouts with a positive reward when a Codex web-search response contains a blocked word.
• Adds a search flag to CodexHarnessConfig that passes web_search="live" to the Codex process when enabled.
• All four packages are installed as editable dependencies in CI via test.yml.

^{Macroscope summarized ff31809.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d41b94349e

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".

[macroscopeapp]

Approvability

Verdict: Approved ff31809

Adds 4 self-contained example environments demonstrating interception patterns, with minimal core changes: publicly exporting an existing function and adding an optional disabled-by-default search config flag to CodexHarness. Low-risk additive changes that don't affect existing runtime behavior.

^{You can customize Macroscope's approvability policy. Learn more.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7aae47f960

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: 311342505e

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: b12342b8be

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]

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.

[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.

[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.

[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.

[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.

[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.

[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]

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.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 1dde940. Configure here.}

[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.

[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.

[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.

[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.

[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.