| name | perf-plan |
| description | Investigate and drive major performance work using repository-local planning, LSP-assisted code navigation, and benchmark evidence. Use when the user asks for performance optimization planning, wants dramatic latency, throughput, memory, CPU, I/O, or cost improvements, or invokes `/perf-plan`-style work with an LSP, focus, or scope. |
Perf Plan
Use this skill to run a performance investigation from scoping through implementation without losing context between iterations.
Resolve Inputs
- Determine effective
LSP, FOCUS, and SCOPE from the request.
- If one is omitted, infer a conservative default from the repository and current conversation, then state the resolved values once.
- Follow the repository's planning convention when it exists. Otherwise, use
.agents/plans/<descriptive-name>.md as the living plan file.
Workflow
- Rebuild context from repository instructions, any existing plan file, and the current performance-sensitive code paths.
- Use the selected LSP MCP server for navigation, symbol references, and static analysis throughout the investigation.
- Summarize the current performance context for the effective scope.
- Identify high-leverage strategies and rank them by expected impact, risk, and implementation cost.
- Preserve external contracts unless the user explicitly approves breaking changes.
- Apply the
execplan workflow for the implementation plan and keep the living plan current as work proceeds.
- Continue through investigation, implementation, and validation unless blocked by missing information or an unsafe decision.
- Benchmark before and after each material optimization using the language-appropriate toolchain.
- If a change regresses performance or correctness, revise the approach instead of rationalizing the regression.
Planning and Memory
- Save key findings, trade-offs, chosen strategy, and next actions in the living plan file after significant decisions.
- Treat the plan file as the resume point for future runs.
- Keep plan updates concrete enough that a fresh contributor can continue without chat history.
Benchmark Rules
- For Go with
gopls, capture baseline and current runs with identical benchmark parameters and compare them with benchstat.
- For Rust with
rust-analyzer, use cargo bench and criterion with like-for-like comparisons.
- When the repository uses a different harness, follow that harness but keep baseline/current comparability strict.
- Do not claim an improvement without reproducible evidence.
Response Shape
- State resolved
LSP, FOCUS, and SCOPE.
- Summarize the current bottlenecks or likely bottlenecks.
- Present concrete improvement strategies tied to the requested focus.
- Provide the implementation sequence and the plan-file location updated.
- Note benchmark evidence, blockers, and remaining risks.
