swallowkit 2.0.0-alpha.6 → 2.0.0-alpha.8

package/src/core/operations/provision-operations.ts

@@ -20,7 +20,7 @@ import { deletePlanState, listProvisionRuns, loadPlanState, loadProvisionRun, lo
 import { hashFileIfExists } from "./file-session";
 import {
   DEFAULT_PROVISION_WAIT_SECONDS, DEPLOYMENT_QUERY, DeploymentObservation, DeploymentReference,
-  MAX_PROVISION_WAIT_SECONDS, queryAzureDeployment, queryDeploymentOperationErrors, waitForAzureDeployment,
+  MAX_PROVISION_WAIT_SECONDS, queryAzureDeployment, queryDeploymentOperationErrors, UNBOUNDED_WAIT_TRANSIENT_RETRIES, waitForAzureDeployment,
 } from "./provision-deployment";
 import {
   MIGRATION_MANIFEST_PATH,
@@ -625,6 +625,8 @@ export interface ApplyProvisionOptions {
   planId: string;
   approve?: boolean;
   waitSeconds?: number;
+  /** Human CLI: keep polling until Azure reaches a terminal state instead of using the bounded wait budget. */
+  waitUntilComplete?: boolean;
 }
 
 export interface ApplyProvisionResult {
@@ -848,7 +850,10 @@ export async function applyProvisionOperation(options: ApplyProvisionOptions): P
     );
   }
 
-  const result = await executeProvisionPlan(plan, { approve: options.approve, waitSeconds: options.waitSeconds, retryHint: MACHINE_RETRY_HINT, projectRoot });
+  const result = await executeProvisionPlan(plan, {
+    approve: options.approve, waitSeconds: options.waitSeconds, waitUntilComplete: options.waitUntilComplete,
+    retryHint: MACHINE_RETRY_HINT, projectRoot,
+  });
   if (result.status === "complete") {
     try { deletePlanState(plan.planId, projectRoot); } catch {
       result.warnings.push("Azure and migration bookkeeping completed, but the old preview could not be removed. The completed execution checkpoint prevents replay.");
@@ -863,11 +868,17 @@ export async function applyProvisionOperation(options: ApplyProvisionOptions): P
  */
 export async function executeProvisionPlan(
   plan: ProvisionPlanData,
-  options: { approve?: boolean; retryHint?: string; projectRoot?: string; waitSeconds?: number; signal?: AbortSignal; pollIntervalMs?: number } = {}
+  options: {
+    approve?: boolean; retryHint?: string; projectRoot?: string; waitSeconds?: number;
+    /** Human CLI only: no overall deadline; each az call stays bounded and the checkpoint stays resumable. */
+    waitUntilComplete?: boolean;
+    signal?: AbortSignal; pollIntervalMs?: number;
+  } = {}
 ): Promise<ApplyProvisionResult> {
   const projectRoot = options.projectRoot ?? process.cwd();
   const retryHint = options.retryHint ?? MACHINE_RETRY_HINT;
-  const waitSeconds = validateWaitSeconds(options.waitSeconds);
+  const waitUntilComplete = options.waitUntilComplete === true && options.waitSeconds === undefined;
+  const waitSeconds = waitUntilComplete ? 0 : validateWaitSeconds(options.waitSeconds);
   if (!options.approve) {
     throw new MachineCommandError(
       "approval-required",
@@ -937,7 +948,7 @@ export async function executeProvisionPlan(
   options.signal?.addEventListener("abort", externalAbort, { once: true });
   if (options.signal?.aborted) controller.abort();
   const signal = controller.signal;
-  const deadline = Date.now() + waitSeconds * 1_000;
+  const deadline = waitUntilComplete ? Number.POSITIVE_INFINITY : Date.now() + waitSeconds * 1_000;
   const persist = () => {
     if (!run) return;
     run.warnings = run.warnings.slice(-20);
@@ -958,6 +969,7 @@ export async function executeProvisionPlan(
       ...(signal.aborted ? {} : { signal }),
       ...(signal.aborted ? { queryTimeoutMs: 5_000 } : {}),
       pollIntervalMs: options.pollIntervalMs,
+      ...(waitUntilComplete ? { transientRetries: UNBOUNDED_WAIT_TRANSIENT_RETRIES } : {}),
       secretValues,
       onObservation: (value) => {
         run!.observation = value;

package/src/core/project/agent-templates/AGENTS.md

@@ -224,6 +224,10 @@ required.
 
 Never invent or implicitly grant human approval.
 
+Humans normally invoke `swallowkit provision` directly. Its invocation requests
+execution without an additional confirmation. Agents must still use the
+machine/MCP approval flow and must not switch to normal CLI to bypass it.
+
 ## Model Changes
 
 For an existing model change, the normal pattern is:

package/src/core/project/agent-templates/CLAUDE.md

@@ -24,6 +24,8 @@ Always-on rules:
   after apply).
 - Use plan/apply for deterministic generation.
 - Respect all `requires-human` states and provisioning approval gates.
+- Normal `swallowkit provision` is for humans. Agents must not use it to bypass
+  the machine/MCP approval flow.
 - Provision timeouts/Ctrl+C do not prove Azure failed. Inspect the original
   execution checkpoint with `machine inspect provision --plan <runId>`; resume
   `machine apply provision --plan <runId> --approve` only after explicit human

package/src/core/project/agent-templates/copilot-instructions.md

@@ -23,6 +23,8 @@ Key rules:
   after apply).
 - Use plan/apply for deterministic generation.
 - Respect `requires-human` states and other human approval gates.
+- Normal `swallowkit provision` is a human-facing deployment command. Agents must
+  use the machine/MCP approval flow, not normal CLI as an approval bypass.
 - Provision timeouts/Ctrl+C do not prove Azure failed. Read `status`, `outcome`,
   and `nextActions`; inspect the original run with `machine inspect provision
   --plan <runId>`, then resume `machine apply provision --plan <runId> --approve`

package/src/core/project/capabilities.ts

@@ -233,12 +233,12 @@ export function inspectCapabilities(): CapabilitiesContract {
       "(regular CLI) configure-mcp [--version auto|latest|<semver>] — re-select the MCP launcher version in .mcp.json",
     ],
     provisioning: {
-      execution: "Submit Bicep once with az deployment group create --no-wait, then bounded state polling (default 300 seconds, configurable 0-3600). Azure CLI invocations are bounded separately. Warning-only stderr with exit 0 is not a failure; output capture limits do not terminate az.",
+      execution: "Submit Bicep once with az deployment group create --no-wait, then bounded state polling (default 300 seconds, configurable 0-3600) for machine/MCP; the human `swallowkit provision` command waits until Azure reaches a terminal state. Azure CLI invocations are bounded separately. Warning-only stderr with exit 0 is not a failure; output capture limits do not terminate az.",
       completion: "status complete/outcome succeeded requires Azure Succeeded plus verified migration tags and local ledger completion. Running/Accepted -> in-progress; unreadable/not-found state -> blocked/outcome unknown; bookkeeping interruption after Azure Succeeded -> blocked/outcome partial; Failed/Canceled -> failed with Azure causes.",
       tracking: "Approved runs persist before Azure writes under .swallowkit/state/provision/. runId equals the original planId and preserves the subscription, RG, deployment ID/correlation, fingerprints and migration state. inspect provision is read-only; apply provision --plan resumes without re-submitting attempted deployments. Ctrl+C/termination may leave Azure running; never cancel/delete without explicit consent.",
       recovery: "For an older untracked baseline, plan provision --resume-deployment <name> compares the existing ARM template and non-secure parameters to the local Bicep/parameters. It observes that deployment and only tags/records it after Succeeded. Mismatches, unreadable evidence, secure parameters or non-baseline migrations block; do not use reconcile just to repair bookkeeping.",
-      diagnostics: "Results include bounded/redacted stdout/stderr, omitted-output flags, exit code, signal, failure kind, elapsed time and deployment identity. Deployment output values are not returned. Never print --debug output, keys, tokens or connection strings.",
-      approval: "Initial apply and every resume require explicit human approval. A passing local command or an ok envelope is not proof of completed Azure provisioning; inspect status, outcome and nextActions.",
+      diagnostics: "Machine/MCP results include bounded/redacted stdout/stderr, omitted-output flags, exit code, signal, failure kind, elapsed time and deployment identity, but no deployment outputs. Normal CLI shows named resource information and CI/CD setup guidance after success. Never print --debug output, keys, tokens or connection strings.",
+      approval: "Machine/MCP initial apply and every resume require explicit human approval. A human invoking normal provision or provision --resume explicitly requests execution without another confirmation prompt. Agents must not switch to normal CLI to bypass approval. A passing local command or an ok envelope is not proof of completed Azure provisioning; inspect status, outcome and nextActions.",
     },
     verifyChecks: {
       defaultChecks: ["structure", "drift", "typecheck", "backend-build"],

package/src/core/project/workflows.ts

@@ -211,6 +211,8 @@ Goal: bootstrap Azure once, then apply only pending versioned infrastructure mig
   \`status: "requires-human"\` and \`requiresApproval: true\` by design.
 - Never pass \`--approve\` to \`apply provision\` unless a human explicitly
   instructed you to provision in this session.
+- This plan/apply workflow is for agents. Humans normally run \`swallowkit provision\`
+  directly; do not switch to that command to bypass the agent approval gate.
 - Inspect infra locally first — no \`az\` calls are needed for inspection.
 - Never edit an applied migration. Create a corrective migration instead.
 - Never use \`--reconcile\` for an ordinary model or resource addition.