openai bed: prove the static-key model-access path (ADR 0050)

A single-node VM bed: the operator sets an API key on an openai licence, the mesh seals it to the consumer, the host unseals and mounts it, and the consumer writes it as OPENAI_API_KEY (env + Codex auth.json). Asserts the written key equals the one set — the other shape ADR 0050 defines, and the ADR 0054 branch where a static-key vendor records no usage. Green on the first run. Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
2026-09-07 04:25:26 +02:00
parent 0513163119
commit a1231a7f89
2 changed files with 287 additions and 0 deletions
@@ -0,0 +1,41 @@
+# One machine that becomes a mesh, then plays out the STATIC-KEY half of model-access (novox/hq
+# ADR 0050) for OpenAI. Where the Anthropic bed proves the refreshable-grant shape (a manager holds a
+# refresh token, the mesh delivers only the access token), this proves the OTHER shape ADR 0050
+# defines and no vendor had exercised: a single operator-supplied key, sealed to the consumer holder,
+# unsealed by the host, and written where an OpenAI/Codex client reads it. There is no manager, no
+# refresh, and no usage — a static-key vendor exposes none, which ADR 0054 says is fine.
+#
+# The flow the test drives:
+#   the operator sets the API key on the licence  ->  the mesh seals it to the consumer holder  ->
+#   the host unseals it and mounts it at the consumer's secret path  ->  the consumer writes it as
+#   OPENAI_API_KEY (env file + the Codex auth.json). The test asserts the written key equals the one
+#   the operator set — through the sealed delivery path, unchanged.
+#
+#   MESH_LAB_HOST_BINARY=.../mesh-host   MESH_LAB_BUNDLE=.../examples/substrate-first-node.lock
+#   Build the consumer runtime image into the local daemon first (raise() stocks it from there):
+#     scripts/build-module-runtime.sh openai-consumer  /tmp/openai-consumer.tar
+scenario: openai-bed
+
+segments:
+  hosting:
+    kind: public
+    cidr: [192.0.2.0/24]
+
+machines:
+  anchor:
+    at: { segment: hosting, address: [192.0.2.10] }
+    inbound: allow
+    memory: 3GiB
+    cpus: 2
+
+images:
+  # The first-node substrate: store, broker, control.
+  - postgres:17-alpine
+  - cloudamqp/lavinmq:latest
+  - mesh-control:development
+  # The static-key consumer runtime, built by scripts/build-module-runtime.sh into the local daemon and
+  # stocked into the scenario's own registry, which is where the host pulls it from.
+  - mesh-runtime-openai-consumer:development
+
+place:
+  all: [host, runtime]
@@ -0,0 +1,246 @@
+/**
+ * The mesh plays out the STATIC-KEY half of model-access for OpenAI, end to end (novox/hq ADR 0050).
+ * Where the Anthropic bed proves the refreshable-grant shape, this proves the other shape ADR 0050
+ * defines — and the ADR 0054 branch where a static-key vendor exposes no usage and records nothing:
+ *
+ *   the operator sets one API key on the licence (`licence key`, read from a file, never a CLI arg) ->
+ *   the mesh seals it to the consumer holder (the generic static-key adapter) -> the host unseals it
+ *   with the node's private key and mounts it at the module's secret path -> the consumer runtime
+ *   writes it where an OpenAI/Codex client reads it (OPENAI_API_KEY env + the Codex auth.json).
+ *
+ * The property proven: the key the consumer writes equals the key the operator set, carried through the
+ * ordinary sealed-delivery path — with NO manager, NO refresh, NO access/refresh split, NO usage. The
+ * whole of OpenAI's integration in the control plane is one registry line (vendor -> static-key).
+ *
+ *   MESH_LAB_HOST_BINARY=.../mesh-host   MESH_LAB_BUNDLE=.../examples/substrate-first-node.lock
+ *   Build the consumer runtime image into the local daemon first:
+ *     scripts/build-module-runtime.sh openai-consumer  /tmp/openai-consumer.tar
+ */
+
+import { test, before, after } from "node:test";
+import assert from "node:assert/strict";
+import { existsSync, readFileSync } from "node:fs";
+import { loadScenario } from "../../src/declaration/parse.ts";
+import { raise } from "../../src/lifecycle/raise.ts";
+import { destroy, exec } from "../../src/lifecycle/operate.ts";
+import { hostBinaryPath, HOST_PATH } from "../../src/lifecycle/place.ts";
+import { labIsUsable, destroyAll } from "./harness.ts";
+
+const capability = await labIsUsable();
+const binary = hostBinaryPath();
+const bundle = process.env["MESH_LAB_BUNDLE"] ?? "";
+
+const skip = !capability.usable
+  ? `lab not usable: ${capability.why}`
+  : !binary || !existsSync(binary)
+    ? "MESH_LAB_HOST_BINARY is not set to a built mesh-host"
+    : !bundle || !existsSync(bundle)
+      ? "MESH_LAB_BUNDLE is not set to a substrate bundle (mesh-host examples/)"
+      : false;
+
+const SCENARIO = "openai-bed";
+const MACHINE = "anchor";
+
+// The API key the flow moves. The whole test is that the key the consumer writes equals this one,
+// carried through the sealed delivery path unchanged. A fake value shaped like a real key.
+const API_KEY = "sk-lab-openai-static-key-value-for-the-bed-only";
+
+let instanceId = "";
+let stocked: string[] = [];
+
+function quote(s: string): string {
+  return `'${s.replaceAll("'", `'\\''`)}'`;
+}
+
+async function on(command: string, timeoutMs?: number): Promise<{ out: string; ok: boolean }> {
+  const { stdout } = await exec(instanceId, MACHINE, [
+    "sh", "-c", `exec 2>&1\n${command}\necho "__exit=$?"`,
+  ], timeoutMs);
+  const marker = stdout.lastIndexOf("__exit=");
+  if (marker < 0) return { out: stdout, ok: false };
+  return { out: stdout.slice(0, marker), ok: stdout.slice(marker + 7).trim() === "0" };
+}
+
+async function must(command: string, timeoutMs?: number): Promise<string> {
+  const { out, ok } = await on(command, timeoutMs);
+  if (!ok) throw new Error(`${MACHINE}: ${command}\n${out}`);
+  return out;
+}
+
+async function mesh(command: string, timeoutMs?: number): Promise<string> {
+  return must(`docker exec mesh-control /mesh-control ${command}`, timeoutMs);
+}
+
+async function meshTry(command: string): Promise<{ out: string; ok: boolean }> {
+  return on(`docker exec mesh-control /mesh-control ${command}`);
+}
+
+function pinned(repository: string): string {
+  const found = stocked.find((r) => r.slice(r.indexOf("/") + 1, r.indexOf("@")) === repository);
+  assert.ok(found, `the scenario stocks no ${repository}; it serves ${stocked.join(", ")}`);
+  return found;
+}
+
+function bundleFor(images: string[]): string {
+  let text = readFileSync(bundle, "utf8");
+  for (const ref of images) {
+    const repository = ref.slice(ref.indexOf("/") + 1, ref.indexOf("@"));
+    const escaped = repository.replaceAll("/", "\\/").replaceAll(".", "\\.");
+    text = text.replaceAll(new RegExp(`[A-Za-z0-9_.:-]+\\/${escaped}@sha256:[0-9a-f]+`, "g"), ref);
+  }
+  return text;
+}
+
+function tokenFrom(said: string): string {
+  const found = said.split("\n").map((l) => l.trim()).find((l) => l.length > 100 && !l.includes(" "));
+  assert.ok(found, `no token in:\n${said}`);
+  return found;
+}
+
+/** The local image id the host pulled for a runtime, so the test can drive it deterministically. */
+async function imageId(substr: string): Promise<string> {
+  const out = (await must(`docker images --no-trunc --format '{{.ID}} {{.Repository}}' | grep ${quote(substr)} | head -1`)).trim();
+  const id = out.split(/\s+/)[0] ?? "";
+  assert.ok(id.startsWith("sha256:") || id.length > 0, `no local image matched ${substr}:\n${out}`);
+  return id;
+}
+
+async function settled(withinMs = 600_000): Promise<void> {
+  const until = Date.now() + withinMs;
+  let last = "";
+  while (Date.now() < until) {
+    const asked = await meshTry(`status --json`);
+    if (asked.ok) {
+      try {
+        const state = JSON.parse(asked.out) as {
+          wrong: { node: string; outcome: string }[];
+          waiting: { node: string }[];
+          reported: { node: string; outcome: string; current: boolean }[];
+        };
+        const bad = state.wrong.find((w) => w.node === MACHINE);
+        if (bad) throw new Error(`${MACHINE} did not apply what it was sent: ${bad.outcome}\n${asked.out}`);
+        const word = state.reported.find((r) => r.node === MACHINE);
+        if (!state.waiting.some((w) => w.node === MACHINE) && word?.outcome === "applied" && word.current) return;
+        last = asked.out;
+      } catch (err) {
+        if (err instanceof Error && err.message.includes("did not apply")) throw err;
+        last = asked.out;
+      }
+    }
+    await new Promise((r) => setTimeout(r, 5000));
+  }
+  throw new Error(`${MACHINE} never caught up within ${Math.round(withinMs / 1000)}s. Last:\n${last}`);
+}
+
+before(async () => {
+  if (skip) return;
+
+  const raised = await raise(loadScenario(`scenarios/${SCENARIO}.yml`), {
+    onProgress: (m) => console.log(`raise: ${m}`),
+  });
+  instanceId = raised.instanceId;
+  stocked = raised.images;
+
+  await must(`cat > /tmp/substrate.lock <<'MESHBUNDLE'\n${bundleFor(raised.images)}\nMESHBUNDLE`);
+  await must(`${HOST_PATH} apply /tmp/substrate.lock`, 600_000);
+  const up = await must(`docker ps --format '{{.Names}}'`);
+  for (const c of ["mesh-store", "mesh-broker", "mesh-control"]) {
+    assert.match(up, new RegExp(c), `the substrate did not raise ${c}:\n${up}`);
+  }
+
+  await mesh(`node add ${MACHINE}`);
+  const token = tokenFrom(await mesh(`token issue --node ${MACHINE}`));
+  await must(`${HOST_PATH} enrol --token ${quote(token)}`);
+  await must(`nohup ${HOST_PATH} run > /var/log/mesh-host.log 2>&1 & sleep 3`);
+}, { timeout: 1_800_000 });
+
+after(async () => {
+  if (instanceId) await destroy(instanceId);
+  await destroyAll(`${SCENARIO}-`);
+}, { timeout: 600_000 });
+
+test("a static-key model-access licence delivers the operator's API key to the consumer, unchanged", {
+  skip, timeout: 1_500_000,
+}, async () => {
+  const consumerImage = pinned("mesh-runtime-openai-consumer");
+
+  // --- the licence, a record with vendor openai (static-key) -------------------------------------
+  // No manager: a static-key licence has none (mesh-control refuses `licence manager` on it). The
+  // consumer is put on the licence BEFORE the key is set — the static-key adapter's Accept seals to the
+  // CURRENT holders, so a holder must exist first, or the key would seal to nobody.
+  await mesh(`licence add openai personal --serves '{"model":"gpt-model"}'`);
+  await mesh(`licence use personal ${MACHINE} openai-consumer`);
+
+  // The operator sets the static key. `licence key` reads it from a file (never a CLI arg) and seals it
+  // to the holder; the plaintext is discarded by the control plane. Staged 0644 so distroless
+  // mesh-control can read the file (docker cp preserves the mode).
+  await must(`printf %s ${quote(API_KEY)} > /tmp/openai-key && chmod 0644 /tmp/openai-key && docker cp /tmp/openai-key mesh-control:/openai-key`);
+  await mesh(`licence key personal --file /openai-key`);
+
+  // --- deploy the consumer -----------------------------------------------------------------------
+  // Inline manifest mirroring the committed module.json: a model-access holder whose delivered key
+  // arrives at its secret path. The scheduled apply container installs as present state (ADR 0053) and
+  // its image is pulled at apply (schedule-pull); the test drives apply directly for a deterministic
+  // flow rather than waiting on cron.
+  const consumerManifest = JSON.stringify({
+    module: "openai-consumer",
+    version: "1",
+    requires: ["model-access"],
+    binds: { "model-access": "/var/lib/openai-consumer/model.json" },
+    secrets: { "model-access": "/var/lib/openai-consumer/api-key" },
+    resources: [
+      { id: "state", type: "directory", path: "/var/lib/openai-consumer", mode: "0700" },
+      { id: "config", type: "directory", path: "/var/lib/openai-consumer/config", mode: "0700" },
+      {
+        id: "apply", type: "container", name: "mesh-openai-consumer-apply",
+        image: consumerImage, network: "host", schedule: "*/5 * * * *",
+        args: ["run", "/app/modules/openai-consumer/dist/apply/index.js"],
+        volumes: ["/var/lib/openai-consumer:/run/state"],
+        env: {
+          MESH_MODEL_ACCESS_SECRET_FILE: "/run/state/api-key",
+          MESH_MODEL_ACCESS_BIND_FILE: "/run/state/model.json",
+          MESH_OPENAI_ENV_FILE: "/run/state/config/openai.env",
+          MESH_OPENAI_CREDENTIALS_FILE: "/run/state/config/auth.json",
+        },
+      },
+    ],
+  });
+  await must(`printf %s ${quote(consumerManifest)} > /tmp/openai-consumer.json && docker cp /tmp/openai-consumer.json mesh-control:/openai-consumer.json`);
+  await mesh(`module add /openai-consumer.json`);
+  await mesh(`module issue openai-consumer --node ${MACHINE}`);
+  await mesh(`assign ${MACHINE} openai-consumer`);
+  await mesh(`push ${MACHINE}`);
+  await settled();
+
+  // --- the host unsealed and mounted the key -----------------------------------------------------
+  let delivered = false;
+  for (let i = 0; i < 20 && !delivered; i++) {
+    delivered = (await on(`test -s /var/lib/openai-consumer/api-key`)).ok;
+    if (!delivered) await new Promise((r) => setTimeout(r, 3000));
+  }
+  assert.ok(delivered, "the sealed API key was never delivered to the consumer's secret path");
+
+  // The delivered key equals the operator's — sealed by the mesh, unsealed by the host, unchanged. The
+  // control plane sealed it and never held it in the clear; here it is, in the clear, on the node the
+  // licence was granted to.
+  const mounted = (await must(`cat /var/lib/openai-consumer/api-key`)).trim();
+  assert.equal(mounted, API_KEY, "the host mounted a different key than the operator set");
+
+  // --- the consumer writes the credential --------------------------------------------------------
+  const consumerId = await imageId("openai-consumer");
+  await must(
+    `docker run --rm --network host -v /var/lib/openai-consumer:/run/state ` +
+      `-e MESH_MODEL_ACCESS_SECRET_FILE=/run/state/api-key ` +
+      `-e MESH_MODEL_ACCESS_BIND_FILE=/run/state/model.json ` +
+      `-e MESH_OPENAI_ENV_FILE=/run/state/config/openai.env ` +
+      `-e MESH_OPENAI_CREDENTIALS_FILE=/run/state/config/auth.json ` +
+      `${consumerId} run /app/modules/openai-consumer/dist/apply/index.js`,
+  );
+
+  // Written where an OpenAI/Codex client reads it: the env file and the Codex auth.json, both the key.
+  const env = await must(`cat /var/lib/openai-consumer/config/openai.env`);
+  assert.match(env, new RegExp(`OPENAI_API_KEY=${API_KEY}`), `the env file does not carry the key:\n${env}`);
+  const auth = await must(`cat /var/lib/openai-consumer/config/auth.json`);
+  const parsed = JSON.parse(auth) as { OPENAI_API_KEY?: string };
+  assert.equal(parsed.OPENAI_API_KEY, API_KEY, `the Codex auth.json does not carry the key:\n${auth}`);
+});