How

Build a working mental model from the current code. This is a read-only investigation, including when another skill delegates it.

Trace the behavior

1. Resolve the requested entry point, scenario, and revision. Inspect applicable instructions and working-tree state. Keep committed code and local changes distinguishable.
2. Read relevant CONTEXT.md, ADRs, and entry points. Trace a representative input through the real callers, state transitions, boundaries, and observable result. Follow actual dispatch, configuration, and error paths rather than describing filenames in isolation.
3. Identify who owns the state, who can mutate it, and where work starts and ends. Include ordering, persistence, retries, and cleanup only when they affect the question.
4. Check material claims against the source. Cite file locations and label unexecuted behavior as source-derived. A passing existing test is evidence only for the path it exercises.

For a narrow question, investigate directly. For independent, substantial parts of a subsystem, use bounded read-only subagents when permitted and useful. Give each the same scope, revision, and no-write restriction, and a distinct question. The parent verifies citations and reconciles disagreements.

Do not edit files, start services, install tools, fetch Git refs, or run code that can mutate state to answer a read-only question. If runtime observation would settle an uncertainty, describe that check as pending. Read existing observations when available.

Explain the result

Lead with the behavior the reader needs to understand, then explain the key concepts, one execution path, ownership, and consequential failure paths. Link to a small set of relevant locations. Use a diagram only when it reduces the explanation.

Distinguish facts, inferences, and unknowns. Current code alone does not establish why its authors chose that shape; use why when rationale matters. Mention supported defects without starting a fix.

When used inside an already authorized implementation task, return the traced model and uncertainties to the caller. This investigation adds no write authority.

Source

Adapted from Lauren Tan's pstack how. Local changes: direct investigation for small questions, strict read-only scope, runtime-neutral delegation, and parent verification of evidence. See LICENSE.