1. Eliminate work that should not run
Start with duplicate workflow triggers, clearly superseded runs, unnecessary retries, or expensive workflows triggered by changes they cannot meaningfully validate. Avoiding a job entirely is usually more valuable than making the same unnecessary job marginally cheaper.
Every proposed removal still needs an evidence-backed reason. Unknown impact should fall back to the existing verification behavior.
2. Reuse equivalent proof conservatively
When an authoritative successful verification already covers the same repository, revision-relevant inputs, environment, policy, and risk class, equivalent-proof reuse can avoid duplicate work. Equivalence must be deterministic and repository/tenant scoped; an LLM assertion is not sufficient proof.
3. Narrow necessary verification only when justified
Affected-scope optimization can reduce runtime, but it carries more false-negative risk than obvious duplicate elimination. Shared schemas, migrations, security, deployment, generated interfaces, or ambiguous dependency relationships should fall back to broader verification.
4. Reduce runtime and setup overhead
For jobs that remain necessary, profile repeated dependency setup, deterministic build work, cache behavior, matrix structure, and long setup-dominated phases. Prefer measurements tied to real job history rather than generic optimization advice.
5. Optimize the price of residual work
Provider and runner placement matters after the residual necessary workload is known. Queue latency, isolation, restartability, risk, and fixed-capacity economics must remain part of the placement decision instead of selecting the cheapest nominal rate in isolation.