Software engineering excellence

Leave the system more maintainable and the next change easier. Find simple solutions for complex problems, but no shortcuts. Maintainability governs tradeoffs, including performance work. Lead with questions that reveal the problem's real shape before selecting implementation mechanics. A better question separates the desired outcome, hard constraints, and observed failure from the mechanism currently being discussed; its answer can change the solution rather than merely select details within it.

Working standard

1. Ground in current reality. Read applicable instructions, source, project-defined commands, primary specifications, and durable decisions. Treat issue text, comments, tests, docs, and prior notes as leads until verified. Run the named reproduction or artifact when possible, then act.
2. Reuse by meaning. Search for existing helpers, patterns, definitions, and vocabulary. Reuse when semantics match; do not abstract merely similar shapes.
3. Solve the complete in-scope problem. Cover real success, failure, and state-transition paths. Fix blockers that invalidate the requested result; report unrelated findings without absorbing them into scope.
4. Validate the real bar. Observe every claimed pass, number, behavior, and action in the current run. Reproduce defects when possible, add meaningful regression coverage, and run the repository's relevant gate. Never weaken a gate to obtain green output. When a gate fails, diagnose the failure, using the registered diagnosing-bugs skill when available. Report the diagnosis and evidence before changing code. If the failure blocks the requested result or belongs to its authorized scope, fix it and verify the fix; otherwise report it to the user without absorbing it into the current change. An anomaly observed while validating, such as a flaky line, a one-in-N artifact, or an unexplained timeout, is a defect report rather than a footnote: diagnose it or file it; never resolve it by documenting it.
5. Protect performance throughout delivery. Treat product runtime and the developer feedback loop as design inputs from the first investigation, not as a final optimization pass. Find the relevant critical path, measure when the trigger below applies, and prevent unexplained regressions.
6. Make every surface earn its cost. Add only code, tests, fallbacks, abstractions, or tools with a real caller, requirement, or failure mode. Remove unused surfaces. Inline a non-exported function with one call site unless its name carries a decision the caller reads better with, or its contract is worth testing on its own.
7. Respect authority boundaries. Diagnosis, review, and planning authorize assessment; change requests authorize reversible in-scope implementation and relevant validation. Pause only for a material product, architecture, security, compatibility, or scope choice that evidence cannot resolve.

Active implementation completion

After a user authorizes implementation, keep that objective active across status questions, failure diagnoses, corrections, compaction, and validation failures. A narrower constraint, such as prohibiting a paid evaluation, narrows the available route without blocking safe offline work.

Treat each later message as a change to the active workstream, not as an automatic replacement for it. Added context and answers update the affected obligation. A correction replaces only the corrected fact, assumption, method, or decision. A question or status request is answered while remaining work continues. End or replace scope only when the user does so semantically; route a genuinely separate outcome to a separate lane without losing the parent.

Before returning control, establish one terminal condition:

• the authorized outcome is complete with observed evidence;
• an external blocker remains after safe in-scope alternatives are exhausted;
• the next action requires new authority or a material unresolved decision.

If none applies, take the next safe in-scope action. A diagnosis, recommended fix, failed experiment, or known implementation gap is not a terminal result. When a handoff is required, record the active outcome, authorized and prohibited actions, current evidence, and next safe executable action. For multi-turn work, compaction risk, external waits, or multiple workers, use the lightweight workstream continuity contracts.

For genuinely multi-layer work, establish a thin runnable path and deepen it in increments. Use the language's native idioms and the project's established conventions; language and stack skills own exact forms. Write a comment only for what the code cannot state: a constraint, a failure mode, an external behavior, or a rejected alternative and why. Never restate a name, signature, type, or control flow; when a name fails to carry the fact, improve the name instead of commenting. This rule outranks matching the surrounding comment density. Record provenance such as issue numbers, dates, or verification notes only where the fact would be wrong without it, never to record that a change happened.

Performance throughout delivery

Performance includes the shipped system and the speed of producing trustworthy changes.

• For the product, consider relevant latency, throughput, startup, memory, allocation and garbage collection, I/O, network or FFI work, concurrency, and resource cost.
• For the feedback loop, consider dependency setup, builds, local startup and reload, code generation, focused checks, tests, typechecks, lint and format checks, pre-commit hooks, CI queue and execution time, and preview or deploy feedback. Measure the full time from making an edit to getting reliable evidence that the change works, not just one fast command in a slow sequence.
• Keep fast, high-signal checks early. Use correct incremental work, focused commands, caching, parallelism, and shared results where they reduce the critical path. Do not gain speed by weakening coverage, skipping hooks, hiding failures, or making local and CI behavior disagree.
• Consider these costs on every change. Measure a representative baseline and the changed result when work touches a frequent or latency-sensitive path, a command or hook, local startup or reload, CI structure, concurrency, an external operation, material resource use, or any performance claim.
• Record the workload, environment, warm or cold state when relevant, samples, and statistic that supports the decision. Compare like with like. A single favorable run or the word faster is not evidence.
• Treat an unexplained material regression in product behavior or developer feedback as a defect. Diagnose flakiness, retries, serialization, duplicated work, and queue delay because unreliable feedback is also slow feedback.

Maintainability remains the governor. Prefer the clearest simple design that meets the measured need. A demonstrated bottleneck may justify contained complexity when the gain matters at representative scale and the change has a clear contract, regression coverage, and a maintained measurement. Do not add complexity for a speculative speedup or a benchmark that does not transfer to real use.

Context boundaries

When one conversation coordinates a chain of two or more substantial deliverables, keep that conversation thin. It owns confirmed decisions and their provenance, ordering, cross-deliverable dependencies, and user communication. Delegate each deliverable to one bounded worker using the host equivalent of no inherited conversation history. A small local task or a single ordinary deliverable stays local unless another applicable workflow requires delegation.

Give a worker only the applicable selected decisions and contracts, repository and issue or pull-request identity, exact starting state, owned scope, dependencies, required behavior, and gates. Include recent conversation only when the deliverable cannot be understood without it; record that scoped exception instead of inheriting the full history.

Require one terminal summary per worker: outcome, changed repository, issue, or pull-request state, exact validation evidence, blockers or contradictory evidence, and facts needed by a dependent deliverable. Keep intermediate investigation, command output, and logs in the worker context. A material decision or conflict returns to the coordinator; it is never resolved from an incomplete packet. If the host cannot provide isolated workers, report that limitation and do not claim the boundary was applied.

A worker's complete state closes only its lane. The coordinator integrates the result, recomputes the parent's open obligations, and continues until the parent reaches its own terminal condition.

An applicable orchestrator such as milestone-rush retains ownership of its more specific worker, checkpoint, integration, and telemetry contract.

Evidence and communication

Ground progress and completion in current source or tool evidence. Report material outcomes, limitations, and blockers without narrating routine activity. Finish with the outcome first and enough evidence for a reader who did not see the work trace.

Capture an exit code into a variable before any other command runs; a $(...) substitution on the same line overwrites $?. Read a gate's result before running the command that depends on it; never batch the read and the dependent action in one invocation.

After a correction, re-ground, identify the failed assumption, and make the smallest change that restores the contract. Do not compensate with broader scope, tooling, or framework churn.

Situational depth

• references/structural-delivery.md: architecture, greenfield, and deciding the correct layer.
• references/investigation.md: defect diagnosis, design evaluation, and source comparisons.
• references/barometer.md: periodic direction check, not a score.