how

---
name: how
description: 'Use when the user asks "how does X work" and wants an architecture walkthrough of the code path or subsystem, including its gotchas. Don''t use for tasks that require source or remote-system changes.'
---

# How

## Contract

| Field | Bound contract |
|---|---|
| Trigger | User asks to explain how a code path or subsystem works |
| Authority | Read-only. No file, VCS, credential, paid, published, deployed, or remote mutation. |
| Side effect | Chat output only; read-only investigation |
| Done | Grounded architecture walkthrough and gotchas returned |

## Inputs

- Target: the specific code path, function, module, class, or subsystem the user names. Required.
- User question: the full natural-language phrasing of the how request. Required.

## Procedure

1. Identify the named target and confirm the scope implied by the user's phrasing. If the target is ambiguous, ask one clarifying question before proceeding. Done when: the target is identified and scope confirmed, or a clarifying question is asked.
2. Locate the entry point of the target in the codebase. Use structural read operations (declaration-only summaries, line-range reads) to identify the primary file and its dependencies. Do not open every file blindly. Done when: the entry point file and its direct dependencies are identified.
3. Trace the control flow from the entry point: identify which functions, classes, or modules are called, in what order, and under what conditions. Map data that flows through the path. Done when: the control flow and data flow from the entry point are mapped.
4. Identify architectural seams: where the code delegates to an external library, makes a system call, crosses a process or network boundary, or performs I/O. Done when: architectural seams are identified with their type recorded.
5. For each major step, note the behavior that a caller or newcomer would find surprising: implicit assumptions, edge conditions handled silently, global state relied upon, error paths that behave differently from happy paths, and performance characteristics that are not obvious from signatures alone. Done when: gotchas are noted for each major step.
6. Turn the findings into a grounded walkthrough of the architecture as it exists, not as intended or documented. Cite concrete files and line ranges for key decisions. Done when: the walkthrough is grounded with concrete file and line-range citations.
7. Return a structured chat report containing the walkthrough and gotchas. Done when: the structured report is returned.

## Failure and recovery
- Target not found: the named path, symbol, or module does not exist in the codebase. Return `no-evidence` and state exactly which target was not found. Do not guess or infer the target's location.
- Ambiguous target: the name matches multiple distinct locations. Ask the user to specify which one they mean. Do not proceed on multiple targets simultaneously.
- Insufficient evidence: the target exists but its implementation is behind an opaque boundary (binary blob, generated code without source, external service). State that the evidence is unavailable and end with `inconclusive`.

## Output
A grounded architecture walkthrough describing:

- The entry point and control flow
- Data that moves through the path
- Architectural seams and their consequences
- Gotchas: surprising behaviors, silent edge cases, and non-obvious assumptions

Output ends with `done` or one of the named failure states: `no-evidence`, `inconclusive`.

• SKILL.md 3 KB
• agents/openai.yaml 1 KB