fix(agent): instruct background review fork to call skill_view before patching skills (#62397)

[webtecnica]

Summary

The background self-improvement review fork is instructed to patch skills, but the prompt never tells the model to call skill_view(name) first. The guard _background_review_write_guard in tools/skill_manager_tool.py enforces a read-before-write requirement and refuses the patch because the current SKILL.md hasn't been loaded. Net effect: corrections are silently dropped.

Change

Updated _SKILL_REVIEW_PROMPT and _COMBINED_REVIEW_PROMPT in agent/background_review.py with a "Read-before-write requirement" instruction that tells the model to:

1. Always call skill_view(name) BEFORE any skill_manage write operation
2. If the tool returns _read_before_write_required: true, read first then retry

Also improved the deny message format to tell the model which tools are available, reducing trial-and-error.

Verification

76 background review tests pass.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #60331 (earliest open, 2026-07-07) — same file (agent/background_review.py), same mechanism: adds a read-before-write instruction to _SKILL_REVIEW_PROMPT / _COMBINED_REVIEW_PROMPT so the review fork calls skill_view(name) before skill_manage writes, aligning with the shipped guard #55906. Sibling #62414 is likewise a dup of #60331. Closes #62397 (the spec/discussion). related_to #62414, #62397.

[teknium1]

Thanks for tracing the background-review guard to the prompts; the current-main mismatch is real.

Problems

• The new wording says skill_view(name) before all listed writes, but that is insufficient for existing supporting files. tools/skill_manager_tool.py:1167-1172 and :1228-1232 require the exact file path to have been read; tests/tools/test_skill_manager_tool.py:1330-1348 verifies that reading SKILL.md does not authorize overwriting references/workflow.md.
• The PR changes both prompt constants without extending the existing behavior-level prompt suite in tests/run_agent/test_review_prompt_class_first.py.

Suggested changes

• Teach the retry path to use skill_view(name, file_path=...) when the rejected mutation targets an existing supporting file, while retaining skill_view(name) for SKILL.md.
• Add focused assertions for both review prompts covering that distinction and the _read_before_write_required retry instruction.

Automated hermes-sweeper review.

[teknium1]

This must distinguish SKILL.md from supporting files. The guard requires the exact target to be read, so skill_view(name) will still reject an overwrite/remove/patch of an existing references/... file; instruct skill_view(name, file_path=...) for that case.

[webtecnica]

Closing as duplicate — another PR already addressed this. Appreciate the contribution!