Shinuk Yi

Classify engineering changes by approval tier and identify which actions must stop for explicit human approval. Use when a change may touch destructive, privileged, production-affecting, or policy-sensitive surfaces and the system must distinguish what can proceed automatically from what cannot.

Execute complex product-development changes with strong scope control, explicit verification, test updates, and documentation sync. Use when implementing multi-file fixes or features where shallow solutions, unrelated edits, and missing regression coverage are major risks.

Perform a focused security review on auth, permission, data exposure, secret handling, auditability, and trust-boundary changes. Use when a code change touches security-sensitive surfaces and a separate security lane should evaluate exploitability rather than general code quality alone.

Review complex changes for boundary erosion, coupling growth, migration sequencing mistakes, and unjustified abstractions. Use when code can pass tests yet still leave the system harder to evolve, operate, or safely migrate.

Build narrow, high-signal context packs for complex engineering tasks by selecting only the relevant code, tests, recent changes, docs, and risks. Use when a task spans multiple files or subsystems and the execution agent would otherwise be harmed by too much or too little context.

Extend benchmark traces with cost, latency, ownership, approval, and intervention signals so quality decisions can be compared against operational cost. Use when CTO- or org-level routing policy needs more than pass/fail evidence.

Keep design docs, ADRs, runbooks, and operator guidance aligned with complex code changes. Use when a task changes behavior, operations, migrations, integrations, or risk posture and documentation drift would create future defects.

Apply stricter decision rules for hotfixes and incidents, including freeze handling, required communications, approval compression, and rollback obligations. Use when production impact or degraded service requires a temporary mode switch without losing auditability.

Build a narrow context pack that includes owner team, service criticality, runbooks, recent incidents, and approval hints. Use when a change touches services where review, release, and escalation quality depend on ownership-aware context rather than code context alone.

Route complex changes to the correct owner teams, reviewers, and approval paths based on changed files and service criticality. Use when the system must decide who is accountable for a change and which stricter gates or escalations should apply.

Select and run deterministic quality gates for complex engineering tasks, including tests, lint, typecheck, security checks, docs sync, and task-specific gates such as migration, retry, or performance checks. Use when a task must be judged by repeatable verification rather than narrative confidence.

Generate a release-ready handoff package including release notes, operator checklist deltas, and rollback notes. Use when a code change is moving toward release and operators or approvers need a concise but actionable artifact package.

Decide whether a complex engineering change is truly ready to release by combining gate results, review findings, approval state, rollback readiness, and operator guidance. Use when passing tests is not enough and a deployment decision requires explicit blocker handling.

Use primary sources to de-risk complex implementation choices involving external SDKs, platform rules, protocols, or policies. Use when a coding decision cannot be made safely from local repository context alone and unsupported assumptions would create costly rework.

Perform an independent review pass focused on regressions, missing requirements, test gaps, security risks, and operational hazards. Use after complex implementations where a separate reviewer must report concrete findings instead of re-implementing the change.

Turn ambiguous or broad product-development requests into strict complex-case task briefs with explicit goal, done conditions, non-goals, risks, approvals, and verification scope. Use when a coding task is multi-file, regression-prone, policy-constrained, or likely to sprawl without a tight execution contract.

Record structured traces, evaluation metadata, and failure taxonomy for complex-case agent runs. Use when comparing candidate agents, skills, or model-effort combinations and you need reproducible evidence instead of impressions.