CI-esta: faster GitHub Actions with evidence

A measured Windows optimization reduced pytest-selective's full validation gate from 249s to 185s and its dependency critical path from 242s to 177s, with every original case and coverage retained. Raw reports and a fresh unchanged control accompany the claim. The same investigation reduced Saleor's full conformance test duration from 223.8s to 205.1s with warm caches on both sides, retaining all 226 original cases and outcomes.

For slow test execution after routing is correct, see critical-path optimizations: isolated PostgreSQL worker databases and native Git on Windows, with full conformance and coverage retained.

Let proven work rest. Measure the rest.

For Bun TypeScript/React suites, use the Bun affected-test adapter. It keeps the repository's test inventory and contracts, uses Bun's native graph, and reports through the same CI-esta work-selection API as other runners.

Find the 279-second job hiding in a 295-second CI run. That is what this real report exposed. Pair measurement with pytest-selective: its matched README experiment used 92.3% less runner time, 627s → 48s, through input routing. The measured results and attribution are below; no universal or receipt-reuse saving is claimed.

CI-esta is an open-source toolkit for GitHub Actions optimization and CI observability. It measures where a workflow spends time and reuses successful validation only when matching execution evidence exists. Your application can use any language.

Repeated CI work is easy to skip incorrectly: the previous workflow was green but the relevant lane never ran, the PR base moved, a newer rerun failed, or the change list was incomplete. CI-esta makes those decisions explicit and falls back to doing the work when proof is uncertain.

Have your agent improve CI

Paste this into your coding agent in the repository you want to improve:

Read https://raw.githubusercontent.com/gigaverse-app/ci-esta/main/skills/ci-esta/SKILL.md
and use it to measure and improve this project's GitHub Actions CI. Preserve
required checks and show measured savings. Use pytest-selective for Python tests.

Install the skills or add AGENTS.md instructions. Any application language on GitHub Actions; no new MCP server or hosted service.

What is included

All actions live in this repository. No hosted CI-esta service is required. Evidence and gate actions use JavaScript; reporting uses an isolated Python runtime. A pytest application is not required.

Start by measuring one workflow

# .github/workflows/observe-ci.yml
name: CI observability
on:
  workflow_run:
    workflows: [CI] # Your validation workflow's name
    types: [completed]
permissions:
  contents: read
  actions: read
  checks: read
  pull-requests: read
jobs:
  report:
    uses: gigaverse-app/ci-esta/.github/workflows/report.yml@v0.1
    with:
      run-id: ${{ github.event.workflow_run.id }}
      run-attempt: ${{ github.event.workflow_run.run_attempt }}
      head-sha: ${{ github.event.workflow_run.head_sha }}
      required-gates-json: '[{"kind":"actions_job","job_name":"Validation gate"}]'
      analyze-critical-path: true

Replace the workflow and required job names with yours. Commit this observer to your default branch; it reports future completed runs without becoming a validation prerequisite. The pinned reusable workflow writes a summary and JSON/Markdown/Prometheus artifacts with read-only permissions. See the observability guide for costs and local CLI usage, then adopt evidence reuse and strict gates. Our validation workflow and observer are complete examples.

Who runs CI-esta

The skipped-lane savings come from pytest-selective routing; CI-esta's part is the gate that keeps a skip honest and the reports that show what to cut next. CI-esta also renders pytest-selective's selection explanations, importing its models instead of keeping a copy. A reused lane is not a conformance certificate: Due and pytest-obligation keep their executed pytest-obligation evidence checks. See how the projects fit together.

Running CI-esta? Say so in your README:

[![CI powered by CI-esta](https://img.shields.io/badge/CI_powered_by-CI--esta-EA580C?logo=data%3Aimage%2Fsvg%2Bxml%3Bbase64%2CPHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHZpZXdCb3g9IjAgMCA2NCA2NCI%2BPHJlY3Qgd2lkdGg9IjY0IiBoZWlnaHQ9IjY0IiByeD0iMTUiIGZpbGw9IiNFQTU4MEMiLz48cGF0aCBkPSJNMTUgMzRsMTEgMTEgMjAtMjIiIGZpbGw9Im5vbmUiIHN0cm9rZT0iI2ZmZiIgc3Ryb2tlLXdpZHRoPSI3IiBzdHJva2UtbGluZWNhcD0icm91bmQiIHN0cm9rZS1saW5lam9pbj0icm91bmQiLz48cGF0aCBkPSJNNDYgN2gxMGwtMTAgMTBoMTAiIGZpbGw9Im5vbmUiIHN0cm9rZT0iI2ZmZiIgc3Ryb2tlLXdpZHRoPSIzLjUiIHN0cm9rZS1saW5lY2FwPSJyb3VuZCIgc3Ryb2tlLWxpbmVqb2luPSJyb3VuZCIvPjwvc3ZnPg%3D%3D)](https://github.com/gigaverse-app/ci-esta)

What measured runs show

Explicit input routing with pytest-selective reduced work in these successful GitHub Actions pairs:

These are one sequential pair per scenario on the same PR head and merge checkout, including planner overhead. Raw timestamps, run links and limitations are published with the selector. Runner-seconds sum executed-job durations; they are not billed minutes. Due's separate widget benchmark is excluded here. These savings come from selection and consumer input declarations, not CI-esta receipt reuse.

The README-only paired experiment also reduced pytest-selective's elapsed time from 211s to 26s, retaining routing regressions, docs and package validation. The same report records a Due docs-only preservation check at 26s elapsed / 91 runner-seconds; it has no paired control and is not presented as a measured speedup.

CI-esta makes the next optimization visible. In a real pytest-selective report, required jobs completed in 295s. The slowest Windows job took 279s; all jobs totaled 731 runner-seconds. Its separate completed-run reporter used 19s of runner time and was not a validation prerequisite. How this was measured includes the source JSON, run identities and limits.

Start with reporting, then narrow inputs and preserve full checkpoints. For Due and pytest-obligation conformance lanes, review every runtime/configuration input before enabling receipt reuse. A previous green workflow alone is insufficient evidence.

CI-esta now uses the selector too. Its same-checkout README probe used 49 versus 105 runner-seconds (53.3% less) and reached the gate in 45s versus 88s, with quality and package checks preserved. The new evidence report also computes real dependency critical paths and slack for all three consumer adoption runs.

What a reused lane means

declared inputs + actual tool bytes
             |
       matching receipt?
       /             \
fresh passing      missing, stale,
lane found         failed or uncertain
    |                    |
reuse evidence       execute lane
       \             /
        strict aggregate gate

Receipts are recorded only after required work passes. Explicit reruns execute again. Missing or contradictory evidence must not silently green the merge gate. Input declarations remain your responsibility: a changed external service cannot be inferred from an unchanged Git tree.

Documentation

Install the Python tools

pip install "ci-esta>=0.1.0,<0.2.0"
ci-esta --help

Actions and reusable workflows use the compatible @v0.1 release tag. Python lockfiles retain exact resolved versions and artifact hashes.

Status and limits

Extracted from Gigaverse's backend-common CI tools. APIs and schemas may evolve before 1.0. On the current Git revision, base reporting installs neither pytest nor pytest-selective (the published 0.1.0 release predates this split). Install ci-esta[selection] for Python selection reports or ci-esta[pytest] for pytest execution tools. Python 3.11+ is required for the CLI; Actions do not require your application to use Python.

Cost figures are estimates using configured rates, not billing records. Admission delay includes more than runner queue time. Evidence reuse is bounded by declared inputs, proof freshness and GitHub API visibility. No universal speedup is claimed. The backend-common compatibility migration is tracked separately from publication; see adoption.

Apache-2.0. See source provenance.

CI toolkit catalog

CI-esta consolidates reusable CI execution, measurement and evidence tooling. Selection and test-isolation guards live in pytest-selective; durable-work conformance tests live in pytest-obligation; the Django/PostgreSQL/Celery application runtime lives in Due.

See the action catalog, migration inventory and preservation test plan. Some actions retain the originating project's defaults (Python 3.14, coverage schedule, security labels); review their documented inputs. The optional dependency-token receipt adapter recognizes backend-common pins. Without that adapter, evidence uses exact dependency pins.

The wheel includes action support resources. ci_esta.resources.source_root() locates them for compatibility imports; normal Actions users can use compatible GitHub uses: tags such as @v0.1. Install ci-esta[pytest] when loading ci_esta.testing.pytest_completion as a pytest plugin.

Runner sizing, distributed shard pools and opt-in resource/pytest hang diagnostics: runner recipes. Standard import rules use import-linter; filesystem-specific rules remain in the shared boundary checker.