| name | code-time-discipline |
| description | Decide whether a software task is code time or not code time, then execute with evidence-backed restraint. Use when asked to optimize, refactor, clean up, harden, review, fix, or improve code; when a request could tempt speculative edits; when deciding whether to patch now, measure first, report findings only, or stop because the change would weaken correctness, contracts, security, or maintainability. |
Code Time Discipline
Use this skill as a gate before editing code. The goal is not to avoid work; it is to edit only when the code path, cost model, and correctness boundary justify it.
Decision Gate
Classify the task before changing files:
- Code time: source inspection shows a real bug, scaling risk, repeated cost, missing guardrail, unsafe lifecycle, inefficient I/O, or clear contract drift, and a focused correctness-preserving patch is available.
- Measure first: the suspected issue is non-trivial, performance-sensitive, concurrency-sensitive, or contract-sensitive, and source inspection alone is not enough. Use existing tests, benchmarks, profiles, logs, query plans, or small reproducible checks before editing.
- Report only: the finding is plausible but speculative, the path is cold, the savings are tiny, the code would become harder to reason about, or the fix needs product/contract input.
- Not code time: the request is answered by explanation, docs, command output, configuration guidance, or review findings; or the edit would weaken validation, authorization, security, transactionality, ordering, public API contracts, or maintainability.
If the classification is not code time, say so directly and provide the next useful action.
Required Inspection
Before editing, read the relevant source and its immediate boundaries:
- Call sites and public exports.
- Tests, fixtures, benchmarks, profiles, logs, or query-plan tooling if present.
- Data schemas, route handlers, workers, jobs, UI render paths, lifecycle ownership, and external I/O boundaries touched by the change.
- Project docs and local agent instructions that define behavior.
Do not infer hotness from file names alone. Decide whether the path is hot, warm, cold, boundary, or pure logic.
Evidence Standard
Every acted-on finding needs:
- Location: file, function, route, query, worker, component, or code path.
- Trigger: when it runs.
- Cost: CPU, memory, allocation, I/O, network, database, render, lock contention, scheduling, startup, or build.
- Growth: constant, linear,
O(n log n), quadratic, unbounded, input-dependent, concurrency-dependent, or dataset-dependent.
- Evidence: proven, strongly likely, plausible, or speculative.
- Recommendation: concrete action that preserves correctness and maintainability.
- Priority:
P0 actual bottleneck or unsafe scaling, P1 likely production issue, P2 cleanup only if already touching the code.
Act on P0 and justified P1 findings. Avoid P2 edits unless they fall out naturally from the same patch.
Editing Rules
- Prefer removing unnecessary work over adding caching, memoization, batching, parallelism, queues, workers, or dependencies.
- Prefer better data access, explicit limits, and bounded concurrency over hiding bad shapes behind caches.
- Preserve validation, authorization, tenant/user/role boundaries, transaction semantics, ordering, idempotency, public API behavior, and readable control flow.
- Do not add compatibility branches, fallback logic, or workaround paths unless the external boundary requires them.
- Keep changes scoped to one coherent responsibility. Do not mix style-only cleanup with functional work.
- Work with existing user changes in the tree. Do not revert unrelated edits.
Verification
After code time, run the smallest meaningful gate set that protects the touched behavior:
- Unit or integration tests for the package/app touched.
- Typecheck and lint for typed code.
- Format check when formatting is enforced.
- Benchmark, profile, query plan, or load check when the improvement is non-trivial and tooling exists.
- Build when public exports, bundling, app startup, or runtime packaging could be affected.
If a gate cannot run, state the exact blocker. Do not claim a change is verified without a passing gate or a clear reason why verification was impossible.
Final Response
Lead with what changed and why. Include the performance finding format for acted-on work. Mention verification. Call out unrelated dirty worktree files when relevant.
