Patch Steering via Prompted Tool Selection

Coding agents with access to multiple patching or refactoring tools (text-based search/replace, AST-based refactoring, semantic migration) may choose suboptimal tools if not explicitly guided. This leads to:

• Unnecessary complexity: the agent picks a generic text-replace tool instead of a specialized AST-aware refactoring tool.
• Inconsistent results: without explicit instructions, tool selection varies between runs for the same task.
• Safety risks: text-based patching on refactoring tasks can break imports, miss references, or introduce syntax errors.

The edit-format choice is measurable. In aider's refactoring benchmark, prompting GPT-4 Turbo to emit unified diffs instead of the previous search/replace block format raised the score from 20% to 61%, and lazy "rest of code unchanged" edits dropped from 12 of 89 tasks to 4. The patch format the prompt asks for is part of the tool interface.

Steer the agent's patch-tool and edit-format selection through explicit natural language instructions in the prompt. This pattern is the patch-tool specialization of the catalogue's Tool Use Steering via Prompting; apply that general pattern when the tools in question are anything other than code-editing tools.

Patch-specific techniques:

1. Name the tool and the format

• Prepend: "Use the apply_patch tool to insert a new function validate_input in auth_service.py."
• When the model's own edit format is the variable, specify it: "Return a unified diff, one hunk per function."

2. Teach the tool in the prompt

• Provide a mini-manual in the context:

"Our `ASTRefactor` tool takes JSON describing node edits:
{"file": string, "pattern": string, "replacement": string}.
Use it for safe refactors rather than raw sed commands."

3. Prefer familiar, forgiving formats

• Aider's benchmarks found that simple, familiar formats (unified diffs) beat structured formats such as JSON tool calls for code edits, because escaping code is error-prone, and that flexibly applying imperfect patches mattered: strict application increased editing errors about 9x.

4. Negative constraints and fallback order

• Specify what NOT to use: "Do NOT use text-based patching for signature changes (it breaks imports)."
• Specify a fallback chain: semantic → AST → text, for when the safer tool fails.

• Tool registry: expose tool metadata (name, usage example, input schema, capabilities and limits) in the agent's initialization context, with use_when and avoid_when criteria per tool.
• Prompt templates: reuse templates with placeholders, for example "Task: {task_description}. Preferred tool: {tool_name}. Usage example: {tool_usage_snippet}."
• Fallback handling: add a directive such as "If ASTRefactor fails, fall back to apply_patch," and test the chain.

• Pros:
  • Predictable behavior: reduces variance in tool usage for the same task.
  • Higher code quality: steers the agent to semantically safe tools and formats for the task.
  • Measured effect: edit-format steering shows large benchmark deltas (aider: 20% to 61% on GPT-4 Turbo).
• Cons/Considerations:
  • Prompt length: tool documentation consumes tokens.
  • Maintenance: as patching tools change, templates and the tool registry need updates.
  • Format lock-in: a format chosen for one model may not be optimal for the next; re-run benchmarks when the model changes.

• Primary source: Unified diffs make GPT-4 Turbo 3X less lazy — aider's write-up of prompting GPT to produce unified diffs, with refactoring benchmark results (20% to 61%) and experiments over alternative edit formats. It covers edit-format steering for code changes, not multi-tool registries.
• Parent pattern: Tool Use Steering via Prompting — the general technique this file specializes.
• Yao, S. et al. (2022). "ReAct: Synergizing Reasoning and Acting in Language Models." ICLR 2023. https://arxiv.org/abs/2210.03629 — general foundation for interleaved reasoning and tool use; not patch-specific.
• An earlier revision cited a Prime Intellect talk on agentic RL as the primary source; that talk does not discuss patch-tool steering, so it was removed.