A service that reads this repository's Gitea issues, decides what to do about each thread as the Program Manager persona, and runs one worker agent per assignment that authors a change, runs the merge gate and opens a pull request. It reacts to issue creation and to comments; the assignment and all correspondence travel through the issue thread and the PR. It authenticates with an Anthropic subscription token. It never pushes to main and never merges. Three choices carry the design, with alternatives recorded in docs/plans/: - it POLLS rather than receiving webhooks (Gitea's webhook SSRF guard, lost deliveries, and no inbound reach needed at all); - it keeps NO state: the trigger is "the newest thing in this thread was not written by me", so the issue thread is the state — no PVC, no backup obligation; - the model holds NO tools: its reply is one JSON object this code validates, so nothing in the path executes text. The boundary is platform/kyverno/admission/pm-agent-job-shape.yaml (Enforce, failurePolicy: Fail), not the RBAC: Kubernetes places no `use` check on serviceAccountName, so "may create Jobs" is "may run arbitrary code as the namespace's most privileged identity". Six rules admit only the committed shape, and the worker identity is bound to nothing with no token mounted, so there is nothing to escalate to. 28/28 offline Kyverno assertions. Nothing has run: the image has never been built or pushed, and both the image digest and the two credentials are deliberate placeholders. See the handoff entry's "What I did NOT verify". Refs CLAUDE.md §1/§3/§6, R-SEC-02, R-GITOPS-02. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
In-cluster admission enforcement (scaffolding — Phase 2 / T-2.7)
Installs Kyverno and runs the same L0 policies as an in-cluster admission
controller, so a non-conforming resource that somehow bypassed CI still cannot be
admitted (defense in depth — docs/07, ADR-0011). Not applied in Phase 0 (no
cluster exists yet); this is scaffolding wired up in task T-2.7.
Enforcement COMPLEMENTS, never REPLACES, behavior tests
A policy admitting a manifest is not proof the runtime behaves (
docs/14principle 1, R-TEST-01). Kyverno confirms a resource is shaped correctly; it does not prove default-deny actually blocks a packet, that the VPN kill-switch fails closed, or that a PVC survives a rebuild. Those are proven only by the L2/L3 Chainsaw behavior and fault-injection tests (conformance/, ADR-0011), which remain mandatory. Admission is a gate, not a guarantee of behavior.
Authored once, shared
kustomization.yaml references the policy files in ../policies/ verbatim — the
identical objects the CI kyverno test gate validates — plus the Kyverno install.
No policy is forked between CI and admission, with one deliberate exception:
The one deliberate CI vs admission divergence — Secrets
../policies/disallow-plaintext-secrets.yaml (CI) requires a top-level sops block
on the Secret manifest. Flux decrypts SOPS before admission, so that block is
gone by the time the API server sees the Secret — the CI policy cannot run in-cluster.
The admission equivalent is secret-provenance.yaml, which enforces provenance
(Secret came from Flux / is an exempt controller Secret) instead of ciphertext. This
is escalated in the T-0.5 report because docs/07's literal wording ("reject a
plaintext Secret at admission") is not achievable post-decrypt.
Operational note (from ADR-0011)
Admission enforcement adds a failure mode: a bad policy can block legitimate
reconciliation. Mitigations: policies ship with their own kyverno test suite (this
directory's ../tests), roll out via GitOps, and the webhook failurePolicy: Fail
is chosen so the system fails closed (coordination-protocol.md #8) — validate for
bootstrap deadlocks in dev/test before prod.