Engineering Orchestrator

Use this as the process coordinator for non-trivial work.

1. Classify the task as small, standard, or complex by risk and blast radius.
2. Load process skills before domain skills, and keep the active set small.
3. Establish observable acceptance criteria and the verification needed to prove them. Use deterministic ocskill inspect / ocskill evidence helpers when available to ground stack, test and working-tree facts cheaply.
4. Inspect enough repository context to plan real files, interfaces, producers, and consumers.
5. For complex or interruption-prone work, maintain a compact fact/assumption/decision/rejected-hypothesis ledger rather than relying on conversational memory.
6. For long-horizon work explicitly using the persistent workflow, create a machine-checkable plan, validate it with ues-plan-checker, and execute dependency-safe tasks through fresh ues-executor contexts instead of carrying the whole implementation in one conversation.
7. Implement in dependency order with verification after meaningful increments.
8. If a check fails, switch to diagnosis rather than stacking speculative fixes.
9. For substantial or high-risk behavior changes, run an independent critic pass that attempts to falsify assumptions and find counterexamples.
10. Repair only evidence-backed blocking findings, re-run affected verification, and bound critic/repair cycles.
11. Before completion, run the evidence gate, inspect the final diff, and disclose unresolved risks accurately.

Read routing.md when skill selection is ambiguous.
Read verification-matrix.md before verifying medium/high-risk changes.
Read retry-policy.md when implementation or verification fails repeatedly.
Read delegation.md before using subagents or parallel analysis.
Read evaluator-loop.md for the critic/repair/re-verify loop on substantial changes.
Read long-horizon.md when work spans many files/tasks, must survive compaction, or will be executed through fresh task agents.
Read model-escalation.md when user-configured model tiers are available or repeated task failures justify escalation.