ai-safe2-framework

Game Developer Quickstart

What the developer receives

The starter adds policy and evidence files. It does not replace the engine, build system, Claude configuration, Codex configuration, or repository tests.

Installation

  1. Copy repository-template/ to the game repository.
  2. Copy profiles/game.json to .ai-safe2/profiles/game.json.
  3. Replace the owner placeholders in .ai-safe2/workflow.json.
  4. Enter the actual build, test, lint, and packaging commands.
  5. Validate the package and generate the repository-specific explanation:

    python scripts/validate_workflow.py --root .
    python scripts/explain_workflow.py \
      --workflow .ai-safe2/workflow.json \
      --profile .ai-safe2/profiles/game.json \
      --output AI-SAFE2-DEVELOPER-WORKFLOW.md
    
  6. Read AI-SAFE2-DEVELOPER-WORKFLOW.md and confirm the routes, trust boundaries, owner, review options, and limitations match the game.
  7. Commit the setup through a normal pull request.

The starter GitHub workflow verifies the pinned CLI, creates the deterministic route, runs CLI decision routing, initializes the revision-bound human decision record to hold, repeats the repository explanation, and uploads those artifacts. A green validation proves that these required records were produced and that the configuration is internally consistent. It does not prove that an optional AI reviewer ran or that the game is secure or ready to release.

What changes after installation

Before After
The same review habit may be used for every change Documentation, gameplay code, and critical trust-boundary changes can take different routes
Claude, Codex, and scanners can produce disconnected comments Results can be retained as attributed, revision-bound evidence
A missing reviewer may be easy to overlook Failed, unavailable, and stale evidence do not satisfy a required capability
“Checks passed” can sound like release approval The workflow stops at ready_for_human_decision; a named human decides

Start with the route and deterministic checks. Add planning, semantic review, security review, or runtime governance only where the repository’s risks and goals justify them. See the Adoption Guide for the module menu and tradeoffs.

Game-specific review lenses

Claude and Codex together

Suggested first pull request

Use a small, reversible change with tests. The pull request should demonstrate:

  1. classification;
  2. deterministic evidence;
  3. one optional semantic review;
  4. a completion receipt;
  5. a named human decision.

Do not start with payment, authentication, multiplayer authority, build signing, or destructive save migration.