Files
0pt/.siax/hooks
Claude CodeandClaude Opus 5 65ee88a01c
CI (SIAX Cloud) / integration (push) Skipped
Security Scan (SIAX Cloud) / security (push) Successful in 16s
CI (SIAX Cloud) / security (push) Successful in 22s
masterplan-lock / check (push) Successful in 21s
masterplan-lock / check (pull_request) Successful in 7s
CI (SIAX Cloud) / security (pull_request) Successful in 19s
Security Scan (SIAX Cloud) / security (pull_request) Successful in 1m2s
CI (SIAX Cloud) / quality (push) Successful in 1m56s
CI (SIAX Cloud) / quality (pull_request) Successful in 1m16s
CI (SIAX Cloud) / integration (pull_request) Successful in 1m15s
fix(ci): check out the siax-templates pin directly instead of comparing main
The G17 gate cloned sax3l/siax-templates on the moving branch `main` and
then required that clone's HEAD to equal SIAX_TEMPLATES_PINNED_SHA. Every
merge in siax-templates -- including ones that touch nothing under
ci/v2/reference-hooks -- therefore turned this repo's release chain red
while nothing was actually unsafe (CD-03 / CI-RED-02). It is red right now:
siax-templates main is at 6612f6ae, the pin is 09504e3a.

The pin is still a real supply-chain control; what was wrong was tying it
to someone else's branch tip. The step now fetches the pinned commit by
SHA (shallow, no branch consulted) and checks it out detached, falling back
to fetching the branches and taking the same commit out of them if the
server refuses a by-SHA fetch. It fails closed on the thing that actually
matters -- the reviewed commit no longer existing upstream -- and the
post-checkout HEAD assertion is kept so the gate still proves what it ran.

SIAX_TEMPLATES_REF is removed rather than left unused, so the moving-branch
shape cannot quietly come back. The pin value itself is unchanged: commit
09504e3a is confirmed present in siax-templates (Gitea API, two commits
behind main). Two comments that described the old "refuses to run if main
moved" behaviour are corrected in the same change.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-26 20:07:29 +02:00
..

.siax/hooks/ -- 0pt's G17 (signed-deploy gate) adapter

What this is

G17.sh in this directory is a thin adapter: it sets 0pt-specific values (image repo, environment, key identity, license policy) and then execs the real, reviewed reference implementation checked out at CI runtime from sax3l/siax-templates (ci/v2/reference-hooks/G17.sh). This repo does not own, fork, or duplicate any of the ADR-W16-003 bar-item logic itself -- see G17.sh's own header for the full rationale and the audit that motivated replacing 0pt's prior ad hoc gate.

Composition mechanism: pinned-commit checkout at CI runtime

.gitea/workflows/release.yml clones sax3l/siax-templates at a pinned commit into a scratch directory before invoking .siax/hooks/G17.sh, authenticated with the SIAX_CI_TOKEN org secret. This was chosen over a vendored copy under .siax/hooks/vendor/ (the mechanism st0re, cl0ud and api0 used for their own G17 migrations) because 0pt has no existing vendored-copy convention anywhere in this repo -- confirmed by searching .gitea/workflows/ and .siax/ for any prior "vendor" pattern before choosing -- while it does already reference SIAX_CI_TOKEN with the exact same git.siax.io authenticated-clone idiom this change reuses (the now-deleted evidence-commit-push step in the old release.yml used the identical x-access-token/oauth2 bearer pattern for a different cross-repo write). Reusing a pattern this repo already has proof it can authenticate with is a smaller, lower-risk change than introducing a new vendoring convention for this alone. sax3l/aud0 and sax3l/h0st's own G17 migrations independently reached the same conclusion for the same reason.

Why not a uses: reusable workflow

Gitea Actions on this fleet resolves a called repository anonymously regardless of the caller's own auth, so uses: sax3l/siax-templates/...@ref 404s against any private sax3l repo. See ci/shim/ci.yml in siax-templates for the fuller writeup. Checkout-at-runtime with an authenticated clone is the only mechanism that works against a private template repo on this Gitea instance.

Pin status

Pinned to sax3l/siax-templates main at 384e35060b91ea78ec045e14af114abf5d5eb23b (2026-09-15) -- PR #51 (feat/g17-signed-deploy-gate-adr-w16-003) merged, plus follow-up fixes #56/#57 (SPDX compound-license-expression evaluation, exercised live by this adapter's own G17_LICENSE_ALLOW policy) and #59 (G20 retry logic; not used by 0pt yet, see below). .gitea/workflows/release.yml fetches and checks out SIAX_TEMPLATES_PINNED_SHA directly by SHA and refuses to run if that exact commit cannot be checked out (removed or rewritten upstream) -- bumping the pin is a reviewed, source-controlled diff to that one line, never a silent re-resolution. It deliberately does not compare the pin against main: that branch tip moves whenever anything unrelated merges in siax-templates, and failing this chain on that said nothing about whether the reviewed gate was safe (CD-03 / CI-RED-02). The authoritative pin value is the one in release.yml.

G20 (live-digest-equality) -- deliberately not wired yet

The reference implementation's bar item 8 proxy (G20, "does the running service's own /version endpoint report the sha that was actually signed") needs two things 0pt does not have today, confirmed live before writing this file:

  1. A provisioned Coolify application. .gitea/workflows/cd.yml's own guard step already fails closed on a literal COOLIFY_APP_ID placeholder -- 0pt has never been deployed to the fleet.
  2. A /version (or equivalent) endpoint. app/api/ has health/ and ready/ routes but no route reporting a build SHA, and the Dockerfile declares no ARG GIT_SHA/ARG BUILD_TIME to bake one in.

Wiring G20 against nothing live would be decorative, not a real check -- the exact anti-pattern this whole migration replaces. Once both exist (Simon provisions the Coolify app; a follow-up adds a /api/version route and the Dockerfile ARGs, the same small change h0st's own migration made), copy G20.sh + its own single-job .gitea/workflows/ g20-live-verify.yml from h0st or aud0's repo as the starting point -- as its own workflow file, not a second job appended to cd.yml: Gitea 1.26 on this fleet resolves org-level secrets (SIAX_CI_TOKEN, AUD0_INGEST_TOKEN) only for the first job of a multi-job workflow file (learned the hard way on h0st's and aud0's own migrations).

Required secrets (not yet set on sax3l/0pt as of this migration)

Same prerequisite every prior migration needed before its release.yml went green:

  • REGISTRY_USER / REGISTRY_TOKEN -- git.cloud.siax.io push (already required by the OLD release.yml too; unchanged).
  • COSIGN_PRIVATE_KEY / COSIGN_PASSWORD -- the private half of this repo's cosign.pub (UNCHANGED by this migration -- see below). Per the old release.yml's own comment this was "NOT YET SET on sax3l/0pt as of this writing" before this PR either, so this remains exactly as outstanding as it always was; this migration does not resolve it. This adapter's own local verification (see the PR description) used a SEPARATE, throwaway cosign keypair generated and then deleted purely to prove the mechanism end-to-end -- cosign.pub at this repo's root was deliberately left untouched. Overwriting a repo's real signing public key from an agent session -- without knowing whether Simon already holds a matching private key offline -- would silently orphan that key forever (ECDSA key generation is not reversible); that decision belongs to Simon, not to this migration. When the real keypair is provisioned (fresh or pre-existing), commit its public half here and set the two secrets together in the same change.
  • AUD0_INGEST_TOKEN / AUD0_TENANT_ID -- new requirement vs. the old gate (which never wrote to AUD0 at all). Needs a siax-svc-g17 workload-identity client provisioned in Zitadel, same open item every prior migration recorded (see G17-ADR-W16-003.md "What this pass could not test against production, and why" in siax-templates).
  • SIAX_CI_TOKEN -- already exists on this repo (used previously for the now-deleted evidence-commit push); reused here for the siax-templates checkout.

Until these are set, release.yml fails closed on its first real run -- same honest behaviour as the gate it replaces, and the same prerequisite cl0ud/st0re/aud0/api0/h0st each had before their own release.yml went green.