fix: Refine Bash PTC File Guidance

[danny-avila]

Summary

I refined bash PTC file/output guidance and added a small runtime reminder for /tmp scratch files.

• Clarified that handoff artifacts must use /mnt/data and that /tmp is same-call scratch, not later-call storage.
• Added an output-side /tmp reminder when submitted code mentions /tmp, so the model gets feedback right after the risky pattern appears.
• Updated bash PTC guidance and examples to save reusable parallel outputs under /mnt/data and parse saved tool stdout with jq -r 'fromjson? // . | ...' for object or stringified-JSON results.
• Preserved the existing compact generated-file summary behavior while keeping host artifacts unchanged.

Change Type

• Bug fix (non-breaking change which fixes an issue)

Testing

• npx jest src/tools/__tests__/ProgrammaticToolCalling.test.ts src/tools/__tests__/BashExecutor.test.ts --runInBand
• npx eslint src/tools/CodeExecutor.ts src/tools/BashExecutor.ts src/tools/ProgrammaticToolCalling.ts src/tools/BashProgrammaticToolCalling.ts src/tools/__tests__/ProgrammaticToolCalling.test.ts src/tools/__tests__/BashExecutor.test.ts
• npx tsc --noEmit
• Live smoke with claude-sonnet-4-6 against http://localhost:3112/v1: Claude saved reusable files under /mnt/data, reused them in a follow-up code-tool call, and used fromjson? // . when parsing saved tool stdout. It still inspected raw output once when it guessed unknown field names, which is schema discovery rather than the encoding issue.

Test Configuration:

• Code API: http://localhost:3112/v1
• Model: claude-sonnet-4-6

Checklist

• My code adheres to this project's style guidelines
• I have performed a self-review of my own code
• My changes do not introduce new warnings
• I have written tests demonstrating that my changes are effective or that my feature works
• Local unit tests pass with my changes

[danny-avila]

@codex review

[chatgpt-codex-connector]

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

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".

[danny-avila]

Live smoke against claude-sonnet-4-6 and http://localhost:3112/v1 after opening the PR:

• Claude saved reusable bash PTC outputs under /mnt/data on the first try, not /tmp.
• Follow-up run_tools_with_bash successfully read those saved /mnt/data files from the session.
• Claude used the new fromjson? // . parsing pattern for stringified JSON output.
• Exercised a deliberate /tmp/smoke.txt call and confirmed the tool output appended: Note: /tmp files are same-call scratch only and were not persisted; use /mnt/data for files needed later.
• One residual behavior: when the mock payload shape was not obvious, Claude still did one raw-output inspection before using the correct .result.data[0] path. That looks like schema discovery, not a persistence or stringified-JSON handling issue.