Simplify Action docs and use version tags

.github/workflows/ci.yml

@@ -27,7 +27,7 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 15
     steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/checkout@v7
         with:
           persist-credentials: false
       - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7.1.0
@@ -42,7 +42,7 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/checkout@v7
         with:
           persist-credentials: false
       - name: Run native mock evaluation through the packaged Action

.github/workflows/copilot-e2e.yml

@@ -28,7 +28,7 @@ jobs:
       group: evals-${{ github.workflow_ref }}
       cancel-in-progress: false
     steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/checkout@v7
         with:
           persist-credentials: false
       - uses: ./

AGENTS.md

@@ -2,7 +2,7 @@
 
 - Keep this a thin GitHub adapter over the acquired Vally CLI. Translate
   supported commands and consume native JSONL; never duplicate selection,
-  configuration parsing, grading or execution. No Runtime/Core bundle is required.
+  configuration parsing, grading or execution.
 - Use Node 24, the pinned pnpm version, and one lockfile. Install with scripts
   disabled. Run focused tests during development and `pnpm check` before review.
 - Rebuild and commit `dist` after source changes. Never manually edit bundles.
@@ -12,6 +12,7 @@
   completion or use a report-only override for native failures.
 - Treat prompts, source, raw evidence and credentials as private. Default
   artifacts are allowlisted structured projections only.
-- Use argument-array process calls and SHA-pinned third-party Actions.
+- Use argument-array process calls. Use `actions/checkout@v7`; keep other
+  third-party Actions SHA-pinned.
 - Qualify native behavior with the built-in mock executor and deterministic
   graders. Do not substitute model-backed runs for model-free acceptance.

CONTRIBUTING.md

@@ -21,8 +21,9 @@ source-to-bundle comparison, and a clean consumer-directory packaging control.
 The clean consumer-directory packaging control deliberately lacks artifact
 credentials and verifies invalid-input failure. Linux CI separately exercises
 the packaged Action with real Vally acquisition, native mock-executor pass/fail
-and lint, plus hosted artifact upload. No model credentials belong in CI or
-test fixtures.
+and lint, plus hosted artifact upload. Do not add model credentials to the
+model-free workflow or unit-test fixtures. The separate Copilot E2E workflow
+uses its automatic job token for the approved live example.
 
 PR check names map directly to these commands:
 
@@ -33,14 +34,14 @@ PR check names map directly to these commands:
 | Unit tests | `pnpm test:coverage`: input, policy, reporting, mocked GitHub, tool reuse, orchestration and real model-free native CLI tests with coverage thresholds. |
 | Bundle verification | `pnpm check:package`: committed bundle freshness and a clean consumer-directory packaged entry-point test. |
 | Action smoke test | Executes `uses: ./` on Ubuntu with actual acquisition, native mock-executor pass/fail, lint, installation reuse and artifact upload. No models run. |
-| Copilot E2E | For relevant same-repository PRs, lints the example and runs one real Copilot skill eval. This consumes organization Copilot AI credits. |
+| evals (Copilot E2E workflow) | For relevant same-repository PRs, directly invokes the default Action on the example. This consumes organization Copilot AI credits. |
 
 Comment networking tests use synthetic GitHub responses, including serialized
 workflow guards, pagination, bot ownership, late runs/attempts, permission
 failures and readback. Installer tests use synthetic packages and exercise
 concurrent install/reuse and tampered files/receipts. `tests/native.test.ts`
 separately qualifies the real pinned Vally CLI with the built-in mock executor
-and deterministic graders. No test claims live GitHub bot-token qualification.
+and deterministic graders. Unit tests do not establish live GitHub token acceptance.
 The live `Copilot E2E` workflow directly exercises the default Action with one
 checkout and one `uses: ./` step. It deliberately uses combined Copilot and
 PR-write permissions in one job, just like the simplest consumer setup.
@@ -54,11 +55,12 @@ without installing this repository. `pnpm check:dist` compares build bytes with
 the checked-in bundle. Never hand-edit generated files. Update the source,
 rebuild, and review both source and distribution changes.
 
-Use strict TypeScript, argument-array process invocation, exact dependencies,
-and SHA-pinned third-party Actions. Do not log raw evaluator/process output or
+Use strict TypeScript, argument-array process invocation, and exact dependencies.
+Use `actions/checkout@v7`; other third-party Actions remain SHA-pinned.
+Do not log raw evaluator/process output or
 copy native runtime behavior into the Action. New native capabilities require
 verified CLI arguments/output contracts and model-free integration controls.
-Only trusted own-repository code may eventually receive execution credentials;
+Only trusted same-repository code may receive execution credentials;
 environment filtering is not an OS sandbox.
 
 The repository's `Copilot E2E` workflow is an explicitly approved live control
@@ -72,6 +74,19 @@ capability through repository write access and branch governance.
 
 ## Release preparation
 
+The first prerelease is [`v0.1.0`](https://github.com/github/evals-action/releases/tag/v0.1.0).
+Keep published version tags unchanged. For future releases, verify the packaged
+Action at the target commit and update consumer examples to the intended tag.
+Action release versions are independent of the selected Vally version.
+
+The moving `v0` tag points to the latest approved 0.x release. After publishing
+a new 0.x minor or patch release and verifying its checks, advance only `v0`
+to that release's exact commit. Verify the tag target afterward. Do not move
+version-specific tags, publish a separate GitHub release for the moving tag,
+or use a blanket force-push of all tags. Release publication and advancing the
+alias are separate steps; no automatic tag-moving workflow is configured.
+Keep `v0` on the 0.x line when introducing a future `v1` release.
+
 Before proposing a release, verify supported CLI versions and their native
 evidence contracts, run the Linux model-free controls, and qualify the optional
 reporting client with the supported hosted token. Do not advertise a plan