Skip to content

v0.7.0

Choose a tag to compare

@github-actions github-actions released this 21 Oct 18:00
· 11 commits to main since this release
f1cd317

Minor Changes

  • 1dd107e: Every non-secret setting now lives in .stamp/policy.yml, so you review a settings change like any policy change.

    • The bundled policy denies only edits to stamp's own policy. Auth, secrets, migrations, infrastructure, billing, public API and toolchain changes go to the reviewer as scrutiny flags with a category-specific instruction. stamp init --strict writes the policy that sends them to a human. The bundled guidance has the reviewer verify a risky change itself and escalate only a risk it cannot resolve.

    • New policy keys: model, backends, pricing, provider.{region,base_url,anthropic_base_url}, guard, max_output_tokens, cli.{claude_model,codex_model}, label, notify.on, telemetry.{endpoint,service}, advise_on_deny and branding. A non-blank environment variable overrides each one.

    • stamp config prints each effective setting and its source, and shows secrets as set or unset.

    • stamp init --model --region writes the policy, so your first run needs only the secret. Init pins the workflows to the version you ran, takes --environment <name> for the review and digest jobs, adds the digest only with --digest, and prints suggested CODEOWNERS entries.

    • stamp init --upgrade rewrites existing workflows from the current templates and keeps each one's environment, OIDC and region.

    • --policy-dir <dir> tries a policy on a real PR before you merge it. Stamp refuses it with --post.

    • advise_on_deny: true appends the model's findings to a gate refusal. The verdict stays REFUSED.

    • branding sets the review header name, an https avatar and the verdict and gate glyphs. Stamp escapes every value.

    • Write access through a team or a private org membership counts as COLLABORATOR, for PR authors and /stamp commenters.

    • Stamp reads the verdict out of a model's narration, and ends in ERROR when the reply holds conflicting verdicts. A model that skips the forced submit_verdict call gets one more turn.

    • The templates grant issues: read for PR reactions, and stamp drops blank provider variables before any provider SDK loads.

    • The run log names the variable each model key comes from, and says when STAMP_API_KEY wins over the provider's own. A failed review's ERROR names it too.

    • stamp --help and stamp <command> --help print the usage.