agent-capability-routing

Files (1)

• SKILL.md

SKILL.md (first lines)

## agent-capability-routing

Goal: when the current agent (for example Domain Orchestrator) has no access to a required tool, find a subagent that has it and delegate the task.

### Subagent capability map

| Subagent               | run_in_terminal | read_file | write_file | search |
| ---------------------- | --------------- | --------- | ---------- | ------ |
| **Test Finisher**      | YES             | YES       | YES        | YES    |
| **Domain Implementer** | YES             | YES       | YES        | YES    |
| **Code Reviewer**      | NO              | YES       | NO         | YES    |
| **Explore**            | NO              | YES       | NO         | YES    |
| **Scope Mapper**       | NO              | YES       | NO         | YES    |
| **Expert Groomer**     | NO              | YES       | NO         | YES    |

> Note: tool availability can vary between sessions. If one subagent fails, try another one from the YES column.

### When to use

- Orchestrator needs to run a PS1/bash script and does not have `run_in_terminal`
- Orchestrator needs to run validation (for example `validate-rag-quality.ps1`)
- Orchestrator needs to inspect process output (build, tests, lint)
- Orchestrator needs to perform a file-writing action outside the standard flow

### Delegation procedure

1. **Identify the need**: what exactly must be executed (command, path, parameters)?
2. **Choose a subagent**: `Test Finisher` or `Domain Implementer` for terminal commands
3. **Craft the prompt**: include the exact command, working directory (for example `cd <repo-root>`), and instruction "return FULL output without truncation"
4. **Verify result**: check exit code and stdout/stderr from the subagent result
5. **Continue workflow**: use the returned output as input for the next step

### Prompt template for subagent (terminal)

```
Run the following command in terminal and return FULL output (stdout + stderr):

cd <repo-root>
& <script-path> <parameters>

Do not truncate output. Do not edit any files except what is required by the command.
```

### Diagnostics when subagent also fails

- If `Test Finisher` has no terminal access -> try `Domain Implementer`
- If both fail -> inform the user and request manual execution + pasted output
- Never estimate output manually instead of actual execution

### Checklist before delegation

- [ ] Is the command safe (non-destructive, does not modify production data)?
- [ ] Is the working directory set correctly?
- [ ] Does the prompt include the instruction "return full output"?
- [ ] Do I know what I will do with the output after receiving it?

From ArkadiuszSieracki/mastermind-agentic-sdlc-vscode-copilot, Apache-2.0.

More from ArkadiuszSieracki/mastermind-agentic-sdlc-vscode-copilot