fix: improve code execution mode tool handling

[rabi]

Summary

• Surface extension server instructions in system prompt by removing conditional code_execution_mode block from system.md
• Use prefixed tool names (code_execution__*) consistently in MOIM, server instructions, and tool descriptions to prevent model confusion
• Add explicit "NEVER use JSON.parse()" warnings since tool results are already parsed JavaScript objects
• Set RepetitionInspector limit to 5 to prevent infinite loops when model repeatedly calls same tools

Type of Change

• Feature
• Bug fix
• Refactor / Code quality
• Performance improvement
• Documentation
• Tests
• Security fix
• Build / Release
• Other (specify below)

AI Assistance

• This PR was created or reviewed with AI assistance

Testing

Tested locally with Gemini and hybrid models like Nemotron.

[rabi]

Hey @alexhancock, PTAL. This makes code_execution work better with smaller hybrid models like Nemotron.

[Copilot]

Pull request overview

This PR improves code execution mode tool handling to better support models that omit tool prefixes.

Changes:

• Added Code Execution Mode section to system prompt explaining that execute_code must be used to call other tools
• Added fallback in dispatch to auto-prefix code_execution tools when models omit the prefix
• Normalized extension names consistently in add_client and is_extension_enabled methods
• Clarified in tool documentation that calls return parsed objects (never use JSON.parse)

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
crates/goose/src/prompts/system.md	Added Code Execution Mode section with examples and instructions for models
crates/goose/src/agents/extension_manager.rs	Normalized extension names in add_client/is_extension_enabled; refactored dispatch fallback logic
crates/goose/src/agents/code_execution_extension.rs	Added reminders about using execute_code; clarified that tool calls return parsed objects
crates/goose-cli/src/session/output.rs	Updated rendering to handle both prefixed and non-prefixed code_execution tool names

[alexhancock]

I worry that this prompt info is duplicative with:

Server instructions: https://github.com/block/goose/pull/6497/changes#diff-40aeaabae1c639df10ef13ae9edf99b906c8bcbf761b0270597d74284b5fb80aR436

Tool description: https://github.com/block/goose/pull/6497/changes#diff-40aeaabae1c639df10ef13ae9edf99b906c8bcbf761b0270597d74284b5fb80aR735

It looks like right now maybe those never make it into the system prompt because of the {% if not code_execution_mode %} block?

But I am confused because I thought in my original change I did have a way to surface this info to the model.

What do you think about refactoring this such that when code mode is active the server instructions + tool descriptions end up included in this, and we use that as the way of surfacing this info to the model vs hardcoding a new copy here.

[rabi]

Thanks Alex. Yeah, I think the tool descriptions are sent (as part of the tool spec), but the server instructions don't surface to the system prompt when code_execution_mode is true.

The naming inconsistency i.e tools are prefixed (code_execution__execute_code) in tool list but server instructions and MOIM using unprefixed names (execute_code) confuses the model and it ends up many times making calls without prefix and going in a loop. Also, the model sees the tool signatures and wants to call them immediately, forgetting the code_execution wrapper is required.

I can change to surface the server instructions and test. However, I've seen #6765 and it surely looks promising and may work better with smaller models. if we decide to go with that, I can test that with Nemotron and abandon this PR.

[rabi]

I've updated the PR and with my local testing it seems to work well with Nemotron-3-nano.

[rabi]

I've proposed a separate improvement PR after testing with pctx and smaller models.