You are improving a subagent's prompts/config for a software engineering (SWE) automation system, based on recent run trajectories. The subagent is used by an AI main agent to address code issues.

CONTEXT
- You will receive trajectory summaries below, starting with the main agent's trajectory, followed by any subagent trajectories in call order.
- Each summary shows what the agent did, what was observed, and how far it progressed.

GOAL
Analyze the subagent's performance and suggest improvements to make it:
1. More discoverable by the main agent (when appropriate)
2. More reliable in its behavior
3. More useful in its output

ANALYSIS FRAMEWORK
Consider these questions:
- Did the main agent discover and use the subagent when it should have?
- Did the subagent behave as expected and return useful information?
- Were there missed opportunities or inefficient behaviors?

IMPROVEMENT TYPES
Focus on one or more of these areas:
1. **docstring**: Make the subagent easier for the main agent to discover and choose appropriately. CRITICAL: Make sure to include "[subagent]" at the beginning of the docstring.
2. **context_description**: Improve the description of the 'context' argument (the only argument) to be clearer and more helpful
3. **instance_template**: Better incorporation of context, clearer framing for each problem instance

Note that the docstring and context_description are visible only to the main agent, while the instance_template is visible only to the subagent.
Thus, if the subagent was not called, you should not edit instance_template. Similarly, if the only issue is the subagent's trajectory, not how it was called, do not edit docstring or context_description.

PRINCIPLES
- Make surgical, targeted improvements rather than broad rewrites
- Preserve existing style and capabilities. Only edit components that need improvement.
- Focus on clarity, discoverability, and reliability
- Ensure generality. Avoid repo- or issue-specific assumptions.
- CRITICAL: DO NOT WRITE ANYTHING SPECIFIC TO THE PARTICULAR CODEBASE, PROJECT, OR DOMAIN.


OUTPUT FORMAT
First, explain your reasoning about what issues you noticed with the provided trajectory and what improvements you're making. Then, output the YAML in a code block. 
**IMPORTANT**: Only suggest edits when you identify a clear, specific problem. If the entire subagent is working well, use an empty updates dictionary.

Sample outputs:

**When improvements are needed:**
Explain your reasoning here.
```yaml
updates:
  docstring: "<improved docstring if needed>"
  context_description: "<improved context argument description if needed>"
  instance_template: "<improved instance template if needed>"
```

RULES
- Only include keys that you intend to change.
- Start with `updates:` as the top-level key. If there are no updates to make, the value for `updates` should be an empty dictionary.
- You may update any combination of the three fields (docstring, context_description, instance_template), but only include a field if it needs improvement.
- No explanations or extra content in the YAML
- Keep each field concise but complete

HEURISTICS
- **Discovery issues**: Strengthen docstring with clear use cases and when to invoke
- **Insufficient context passed to subagent**: Improve context_description with clearer argument explanation
- **Subagent behavior and output (Incorrect subagent trajectory or return information)**: Improve instance_template with better instructions and output specifications