refactor: enhance retry mechanism with flexible context management

[xtco3o]

Refactors the retry mechanism with flexible context management and enhanced safety.

Key Changes:

• Flexible Context Control: Added reset_context support.
• Rich Failure Feedback: Captures stdout, stderr, and exit_status.
• Advanced Context Protection: Automatically truncates large outputs based on both line count (200) AND character count (20,000). This prevents LLM context overflow even from massive single-line outputs (like Base64 or minified JSON).
• Privacy & UI Integrity: Marked retry feedback as agent_only() to keep UI clean.
• Improved Type Safety: Introduced SuccessCheckResult and ExecutionDetails.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 961b9855cc

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[xtco3o]

All requested refactors have been completed:

• Reduced Redundancy: Extracted append_output_details helper to unify output processing.
• Improved Maintainability: Added MAX_RETRY_OUTPUT_LINES constant (200) for consistent log truncation.
• Context Protection: Implemented automatic truncation of stdout/stderr to prevent token overflow during retries.
• Verified: Unit tests for truncation logic have been added and passed.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e579037a3d

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[xtco3o]

Resolved Codex Review:
Non-zero exit status for on_failure commands is now correctly treated as a terminal error. This prevents the retry logic from proceeding with a failed recovery state.

• Updated execute_on_failure_command to return Err on non-zero exit.
• Updated related unit tests to reflect this change.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 091b5e5bc8

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[DOsinga]

Review

Thanks for the contribution — the reset_context feature is a useful idea. A few things need addressing before this can move forward:

Needs an issue first

This adds a new feature (the reset_context field and the feedback-loop retry path) without a prior issue or discussion. Per our contributing guidelines, please open an issue describing the use case and proposed design so we can discuss it before the code: https://github.com/block/goose/blob/main/CONTRIBUTING.md#discussions-issues-and-prs

on_failure error handling is broken in the non-reset path

Codex flagged this correctly. execute_on_failure_command returns Err on non-zero exit, and handle_retry_logic propagates it with ?:

let on_failure_details = if let Some(on_failure_cmd) = &retry_config.on_failure {
    Some(execute_on_failure_command(on_failure_cmd, retry_config).await?)
} else {
    None
};

This means when reset_context: false and the on_failure command fails, the retry aborts entirely. The code below that builds a message with "with failure" status is dead — it can never be reached. You need to decide: should a failing on_failure command abort the retry, or should its output be fed back to the agent? The current code tries to do both and achieves neither.

Remove test_retry_result_enum

This test just asserts that enum variants are not equal to each other and that Clone/Debug work. It can never catch a real bug — please remove it.

Integration tests needed

The new reset_context: false path has no integration test that exercises handle_retry_logic end-to-end. crates/goose/tests/agent.rs already has a retry_tests module — please add tests there (following the mock-provider pattern from max_turns_tests) that cover:

• reset_context: true resets conversation history (existing behavior, but worth pinning)
• reset_context: false appends failure feedback as an agent-only user message with stdout/stderr
• reset_context: false with a failing on_failure command (once you fix the error handling)
• truncation actually kicks in when output exceeds the limits

Minor

• Comments like // Execute on_failure if configured and // Add Success Check failure info just restate what the next line does — remove them.
• truncate_output mixes len() (bytes) with the "chars" label in the truncation message — pick one unit and be consistent.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e809fe9239

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

goose/crates/goose/src/agents/agent.rs

Line 1734 in b71ec20

messages_to_add = Conversation::default();

Preserve failed assistant turn when reset_context is false

In the retry path, handle_retry_logic runs before messages_to_add is merged into conversation, and when a retry is requested this branch unconditionally clears messages_to_add. With reset_context: false, that drops the just-produced assistant response from persisted history, so the next attempt only sees synthetic failure feedback and loses the model’s immediate prior turn, which defeats the intended context-preserving behavior.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5556cc6fab

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f67e6e97bd

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0e26f05bbc

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 577d4a652b

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Keep append-only persistence on non-retry turns

This branch now persists with replace_conversation even when retry logic returns false because it was skipped or all checks passed, so ordinary no-tool turns rewrite the entire transcript. Since replace_conversation deletes and reinserts all session messages, long chats now pay O(n) database writes per turn (O(n²) across a session), which can materially slow interactions and increase lock contention compared with the previous append-only add_message path.

Useful? React with 👍 / 👎.