| name | change-plan-with-graph |
| description | Generate a graph-backed implementation plan before modifying repository code. |
Use this skill for significant, unfamiliar, high-risk or cross-module implementation work, refactors, renames, removals or migrations. Also use it when the user explicitly asks for graph-backed planning. Do not trigger it for simple single-file edits, formatting, typos, straightforward test updates or direct command-output requests.
Goal: Produce a graph-backed implementation plan when the change is large or risky enough that caller, callee, test or public-surface impact matters.
Shared tool policy:
- Follow
../../reference/codegraph-tool-policy.md for common CodeGraph tool selection, fallback and unavailable handling.
Task-specific tools:
- Prefer
codegraph_context, codegraph_search, codegraph_node, codegraph_callers, codegraph_callees, codegraph_impact and codegraph_explore for unfamiliar subsystems.
Workflow:
- Summarise the user's goal in a single sentence to anchor the task.
- Call
codegraph_status to ensure the graph index is available and fresh. If not, ask the user to initialise or sync it.
- Build a task context using
codegraph_context (e.g. "implement rate limiting for authenticated routes"). Limit nodes to a manageable number.
- If the target area is unfamiliar, use
codegraph_search to find concrete anchors and then one bounded codegraph_explore call for a subsystem survey.
- Identify relevant symbols and modules using
codegraph_search. Use codegraph_node to confirm details. Resolve ambiguities.
- Run
impact-analysis on the key symbols to understand who calls them, who they call and what tests are affected.
- Use
test-target-finder to determine which tests will need to run.
- Draft a short implementation plan that includes:
- High-level steps (analysis, code changes, tests, validation).
- Files or modules to edit.
- Supporting tasks (e.g. update documentation).
- Estimated risks and unknowns.
- Proceed after the plan for low-risk technical changes. Ask the user before editing only when the plan affects product behavior, cost, time, security, privacy, data, permissions or other externally visible outcomes.
Rules:
- For changes that trigger this skill, produce a graph-backed plan before editing. Do not block trivial edits that do not meet the trigger criteria.
- User confirmation is required only for the higher-risk cases above.
- Prefer CodeGraph findings; if using CLI fallback, clearly state that results may be incomplete.
- Encourage the user to ask clarifying questions or adjust the plan before execution.