@oneuptime/common 14.0.8 → 14.0.10

package/Tests/Server/Utils/AI/RemediationKubectlCopyParity.test.ts

@@ -355,6 +355,8 @@ describe("the tools the model reads use the shared summaries", () => {
         canRunAiCommands: true,
       },
       accessMethod: "in_cluster",
+      aiAgent: null,
+      automaticInvestigation: { incidents: false, alerts: false },
       kubectlAllowlist: ["kubectl set image deployment/web * -n web"],
       isInvestigationEnabled: true,
       isInvestigationReady: true,

package/Tests/Server/Utils/AI/RemediationPlanRunner.test.ts

@@ -18,7 +18,9 @@ import Incident from "../../../../Models/DatabaseModels/Incident";
 import Runbook from "../../../../Models/DatabaseModels/Runbook";
 import AIRunStatus from "../../../../Types/AI/AIRunStatus";
 import AutoRemediationSuggestionStatus from "../../../../Types/AutoRemediation/AutoRemediationSuggestionStatus";
+import AutoRemediationSuggestionType from "../../../../Types/AutoRemediation/AutoRemediationSuggestionType";
 import ObjectID from "../../../../Types/ObjectID";
+import AiResourceType from "../../../../Types/ResourceAiAgent/AiResourceType";
 import { afterEach, beforeEach, describe, expect, it } from "@jest/globals";
 
 /*
@@ -744,4 +746,119 @@ describe("RemediationPlanRunner.settleStrandedPlanningSuggestions", () => {
 
     expect(feed).not.toHaveBeenCalled();
   });
+
+  /*
+   * A resource round (a Docker host's, a database server's, ...) is named
+   * by its resource on the feed, never as "Auto Remediation Rule ...";
+   * rule and cluster rounds read exactly as before.
+   */
+  describe("how the feed note names the stranded round", () => {
+    const RESOURCE_ID: ObjectID = new ObjectID(
+      "dddddddd-dddd-4ddd-8ddd-dddddddddddd",
+    );
+
+    function resourceRound(
+      overrides: Partial<Record<string, unknown>> = {},
+    ): Partial<Record<string, unknown>> {
+      return {
+        autoRemediationRuleId: undefined,
+        ruleNameSnapshot: 'AI remediation for Docker host "web-1"',
+        suggestionType: AutoRemediationSuggestionType.CommandPlan,
+        resourceType: AiResourceType.DockerHost,
+        resourceId: RESOURCE_ID,
+        ...overrides,
+      };
+    }
+
+    async function settleAndReadFeed(): Promise<string> {
+      const feed: jest.SpyInstance = jest
+        .spyOn(IncidentFeedService, "createIncidentFeedItem")
+        .mockResolvedValue(undefined as never);
+      jest
+        .spyOn(AutoRemediationSuggestionService, "attemptStatusTransition")
+        .mockResolvedValue(1 as never);
+
+      await RemediationPlanRunner.settleStrandedPlanningSuggestions();
+
+      expect(feed).toHaveBeenCalledTimes(1);
+
+      return String(
+        (feed.mock.calls[0]![0] as { feedInfoInMarkdown: string })
+          .feedInfoInMarkdown,
+      );
+    }
+
+    it("reads the resource columns of the Planning rounds", async () => {
+      mockPlanningList();
+      mockRun({ status: AIRunStatus.Running });
+
+      await RemediationPlanRunner.settleStrandedPlanningSuggestions();
+
+      const select: Record<string, unknown> = (
+        (AutoRemediationSuggestionService.findBy as unknown as jest.Mock).mock
+          .calls[0]![0] as { select: Record<string, unknown> }
+      ).select;
+      expect(select["resourceType"]).toBe(true);
+      expect(select["resourceId"]).toBe(true);
+    });
+
+    it("a resource round is named by its resource", async () => {
+      mockPlanningList(resourceRound());
+      mockRun({ status: AIRunStatus.Error });
+
+      expect(await settleAndReadFeed()).toBe(
+        '⚡ **AI remediation for Docker host "web-1": AI planning did not complete** (the planning run failed) — no runbook was proposed.',
+      );
+    });
+
+    it("a resource round that had already run commands is named by its resource too", async () => {
+      mockPlanningList(
+        resourceRound({
+          commandPlan: {
+            commands: [
+              {
+                sequence: 1,
+                stepType: "ResourceCommand",
+                command: "docker restart web",
+                runnerId: RUN_ID.toString(),
+                resourceType: AiResourceType.DockerHost,
+                resourceId: RESOURCE_ID.toString(),
+                resourceNameSnapshot: "web-1",
+                policyVerdict: "AutoApproved",
+                rationale: "restart",
+                execution: { status: "Succeeded" },
+              },
+            ],
+          },
+        }),
+      );
+      mockRun({ status: AIRunStatus.Stale });
+
+      expect(await settleAndReadFeed()).toMatch(
+        /^⚡ \*\*AI remediation for Docker host "web-1": the AI command run was interrupted after executing 1 command\(s\)\.\*\*/,
+      );
+    });
+
+    it("negative control: a rule round keeps its rule label", async () => {
+      mockPlanningList();
+      mockRun({ status: AIRunStatus.Error });
+
+      expect(await settleAndReadFeed()).toBe(
+        '⚡ **Auto Remediation Rule "Restart API pods": AI planning did not complete** (the planning run failed) — no runbook was proposed.',
+      );
+    });
+
+    it("negative control: a cluster round's note is unchanged", async () => {
+      mockPlanningList({
+        autoRemediationRuleId: undefined,
+        kubernetesClusterId: RULE_ID,
+        ruleNameSnapshot: 'AI remediation for cluster "prod-us"',
+      });
+      mockRun({ status: AIRunStatus.Error });
+
+      expect(await settleAndReadFeed()).toBe(
+        '⚡ **Auto Remediation Rule "AI remediation for cluster "prod-us"": AI planning did not complete** (the planning run failed) — no runbook was proposed.',
+      );
+    });
+  });
 });

package/Tests/Server/Utils/AI/RemediationResourceCopyParity.test.ts

@@ -0,0 +1,463 @@
+import {
+  buildResourceFullAutoPersona,
+  buildResourceQuestion,
+  buildResourceSuggestPersona,
+  getResourceUndoExamples,
+} from "../../../../Server/Utils/AI/Remediation/RemediationExecutionRunner";
+import {
+  RESOURCE_ALLOWLIST_SUMMARY,
+  RESOURCE_ALWAYS_ASKS_SUMMARY,
+  RESOURCE_AUTOMATIC_MODE_SUMMARY,
+  RESOURCE_BYPASS_MODE_SUMMARY,
+  RESOURCE_EVERY_MODE_LIMITS_SUMMARY,
+  RESOURCE_NEVER_RUNS_SUMMARY,
+  RESOURCE_RISKIER_CHANGES_SUMMARY,
+  RESOURCE_SAFE_CHANGES_SUMMARY,
+  RESOURCE_UNATTENDED_ROUND_BECOMES_PROPOSAL_SUMMARY,
+} from "../../../../Types/AutoRemediation/AiRemediationCommandPlan";
+import AiResourceType, {
+  AI_RESOURCE_TYPE_INFO,
+  ALL_AI_RESOURCE_TYPES,
+} from "../../../../Types/ResourceAiAgent/AiResourceType";
+import {
+  RESOURCE_AI_ALLOW_WRITES_ENV,
+  RESOURCE_AI_WRITE_TARGETS_ENV,
+  ResourceAiAccessStatus,
+  ResourceAiRemediationMode,
+  ResourceCommandTier,
+} from "../../../../Types/ResourceAiAgent/ResourceAiAccess";
+import { AiRemediationCommandPolicyVerdict } from "../../../../Types/AutoRemediation/AiRemediationCommandPolicyVerdict";
+import ResourceCommandPolicy from "../../../../Utils/AiRemediation/Resource/ResourceCommandPolicy";
+import {
+  ResourceAutoExecutionVerdict,
+  tokenizeResourceCommand,
+} from "../../../../Utils/AiRemediation/Resource/ResourceCommandPolicyCore";
+import { RUN_INFRASTRUCTURE_COMMAND_TOOL_NAME } from "../../../../Server/Utils/AI/ResourceAccess/ResourceAccessToolNames";
+import { describe, expect, it } from "@jest/globals";
+
+/*
+ * Contract under test — every LLM-facing description of the resource
+ * command tiers and the unattended resource modes says what the resource
+ * command policy does NOW, the way RemediationKubectlCopyParity pins the
+ * kubectl copy:
+ *
+ * - the shared summaries (AiRemediationCommandPlan's RESOURCE_*_SUMMARY)
+ *   restate the canonical ResourceCommandTier / ResourceAiRemediationMode
+ *   semantics: SafeWrite is ONE named object; stop, kill, scale to zero,
+ *   power off and terminate are riskier; a node drain, a VM migration, host
+ *   maintenance and killing a host process always need a human; destructive
+ *   commands never run; the allowlist is word-wise;
+ * - each example they name, and every undo a persona recommends, is tiered
+ *   that way by ResourceCommandPolicy itself — so the copy cannot drift
+ *   from the policy unnoticed;
+ * - the resource personas and the question use those summaries, name the
+ *   resource and its programs, and send reads to run_infrastructure_command.
+ */
+
+// Each phrase of the safe summary, as the model would write it, per resource type.
+const SAFE_EXAMPLES: Array<[AiResourceType, string]> = [
+  [AiResourceType.DockerHost, "docker restart web"],
+  [AiResourceType.PodmanHost, "docker start web"],
+  [AiResourceType.DockerSwarmCluster, "docker service update --force api"],
+  [
+    AiResourceType.ProxmoxCluster,
+    "pvesh create /nodes/pve1/qemu/100/status/start",
+  ],
+  [AiResourceType.VMwareVCenter, "govc vm.power -on /DC1/vm/web-01"],
+  [AiResourceType.CephCluster, "ceph osd in 3"],
+  [AiResourceType.DatabaseServer, "db cancel-query 4242"],
+  [AiResourceType.Host, "systemctl restart nginx"],
+];
+
+const RISKIER_EXAMPLES: Array<[AiResourceType, string]> = [
+  [AiResourceType.DockerHost, "docker stop web"],
+  [AiResourceType.DockerHost, "docker kill -s TERM web"],
+  [AiResourceType.DockerHost, "docker pause web"],
+  [AiResourceType.DockerHost, "docker update --memory 1g web"],
+  [AiResourceType.DockerSwarmCluster, "docker service scale web=0"],
+  [
+    AiResourceType.ProxmoxCluster,
+    "pvesh create /nodes/pve1/qemu/100/status/shutdown",
+  ],
+  [AiResourceType.VMwareVCenter, "govc vm.power -off web-01"],
+  [AiResourceType.VMwareVCenter, "govc vm.power -reset web-01"],
+  [AiResourceType.CephCluster, "ceph osd out 3"],
+  [AiResourceType.Host, "systemctl stop nginx"],
+  [AiResourceType.DatabaseServer, "db terminate-session 4242"],
+  // Several objects at once.
+  [AiResourceType.DockerHost, "docker restart web api"],
+];
+
+const ALWAYS_ASKS_EXAMPLES: Array<[AiResourceType, string]> = [
+  [
+    AiResourceType.DockerSwarmCluster,
+    "docker node update --availability drain n1",
+  ],
+  [
+    AiResourceType.ProxmoxCluster,
+    "pvesh create /nodes/pve1/qemu/100/migrate --target pve2",
+  ],
+  [AiResourceType.VMwareVCenter, "govc host.maintenance.enter esx-01"],
+  [AiResourceType.Host, "kill -TERM 1234"],
+];
+
+const NEVER_RUNS_EXAMPLES: Array<[AiResourceType, string]> = [
+  [AiResourceType.DockerHost, "docker exec web sh"],
+  [AiResourceType.DockerHost, "docker run alpine"],
+  [AiResourceType.DockerHost, "docker rm web"],
+  [AiResourceType.DockerHost, "docker system prune -f"],
+  [AiResourceType.Host, "sudo systemctl restart nginx"],
+  [AiResourceType.Host, "systemctl restart nginx | tee /tmp/x"],
+  [AiResourceType.VMwareVCenter, "govc vm.destroy web-01"],
+  [AiResourceType.CephCluster, "ceph auth get client.admin"],
+];
+
+function readyResource(
+  resourceType: AiResourceType,
+  overrides: Partial<ResourceAiAccessStatus> = {},
+): ResourceAiAccessStatus {
+  return {
+    resourceType,
+    resourceId: "33333333-3333-4333-8333-333333333333",
+    resourceName: "web-1",
+    isAiInvestigationEnabled: true,
+    aiRemediationMode: ResourceAiRemediationMode.Automatic,
+    aiCommandAllowlist: [],
+    agent: {
+      agentId: "44444444-4444-4444-8444-444444444444",
+      connectionStatus: "connected",
+      isOnline: true,
+      posture: {
+        resourceType,
+        resourceIdentifier: "web-1",
+        allowWrites: true,
+        writeTargets: [],
+        protectedTargets: [],
+        reachable: true,
+      },
+    },
+    gaps: [],
+    isInvestigationReady: true,
+    isRemediationReady: true,
+    ...overrides,
+  };
+}
+
+describe("the shared resource tier summaries match what ResourceCommandPolicy does", () => {
+  it.each(SAFE_EXAMPLES)(
+    "a safe change is SafeWrite on a %s: %s",
+    (resourceType: AiResourceType, command: string) => {
+      expect(
+        ResourceCommandPolicy.evaluateCommand({ resourceType, command }).tier,
+      ).toBe(ResourceCommandTier.SafeWrite);
+    },
+  );
+
+  it.each(RISKIER_EXAMPLES)(
+    "a riskier change is RiskyWrite on a %s: %s",
+    (resourceType: AiResourceType, command: string) => {
+      expect(
+        ResourceCommandPolicy.evaluateCommand({ resourceType, command }).tier,
+      ).toBe(ResourceCommandTier.RiskyWrite);
+    },
+  );
+
+  it.each(ALWAYS_ASKS_EXAMPLES)(
+    "a change that always asks a human does so on a %s even bypassed and allowlisted: %s",
+    (resourceType: AiResourceType, command: string) => {
+      const verdict: ResourceAutoExecutionVerdict =
+        ResourceCommandPolicy.evaluateForAutoExecution({
+          resourceType,
+          command,
+          allowlistPatterns: [command],
+          bypassApproval: true,
+        });
+
+      expect(verdict.verdict).toBe(
+        AiRemediationCommandPolicyVerdict.RequiresApproval,
+      );
+      expect(verdict.requiresHuman).toBe(true);
+    },
+  );
+
+  it.each(NEVER_RUNS_EXAMPLES)(
+    "a destructive command never runs on a %s: %s",
+    (resourceType: AiResourceType, command: string) => {
+      expect(
+        ResourceCommandPolicy.evaluateCommand({ resourceType, command }).tier,
+      ).toBe(ResourceCommandTier.Denied);
+    },
+  );
+
+  it("the summaries name what each tier covers, in the canonical words", () => {
+    expect(RESOURCE_SAFE_CHANGES_SUMMARY).toContain("exactly ONE named object");
+    expect(RESOURCE_SAFE_CHANGES_SUMMARY).toContain(
+      "never several objects at once",
+    );
+    for (const riskier of [
+      "stop",
+      "kill",
+      "scale a service to zero",
+      "power off",
+      "terminate a session",
+      "several objects at once",
+    ]) {
+      expect(RESOURCE_RISKIER_CHANGES_SUMMARY).toContain(riskier);
+    }
+    for (const asks of [
+      "draining a Swarm node",
+      "migrating a VM",
+      "maintenance",
+      "killing a process on a host",
+      "in every mode — Bypass approval and the allowlist included",
+    ]) {
+      expect(RESOURCE_ALWAYS_ASKS_SUMMARY).toContain(asks);
+    }
+    for (const never of ["exec", "run", "rm", "prune", "credentials", "sudo"]) {
+      expect(RESOURCE_NEVER_RUNS_SUMMARY).toContain(never);
+    }
+    expect(RESOURCE_ALLOWLIST_SUMMARY).toContain(
+      "a * stands for exactly one whole word",
+    );
+  });
+
+  it("the allowlist summary describes how the matcher really reads a pattern", () => {
+    // Exactly one whole word, never part of one; same length.
+    expect(
+      ResourceCommandPolicy.matchesAllowlist({
+        resourceType: AiResourceType.DockerHost,
+        command: "docker stop web",
+        patterns: ["docker stop *"],
+      }),
+    ).toBe(true);
+    expect(
+      ResourceCommandPolicy.matchesAllowlist({
+        resourceType: AiResourceType.DockerHost,
+        command: "docker stop -t 5 web",
+        patterns: ["docker stop *"],
+      }),
+    ).toBe(false);
+    expect(
+      ResourceCommandPolicy.matchesAllowlist({
+        resourceType: AiResourceType.DockerHost,
+        command: "docker stop web-1",
+        patterns: ["docker stop web-*"],
+      }),
+    ).toBe(false);
+  });
+
+  it("the every-mode clause names the agent's own write switches and ends with the proposal rule", () => {
+    expect(RESOURCE_EVERY_MODE_LIMITS_SUMMARY).toContain(
+      `${RESOURCE_AI_ALLOW_WRITES_ENV}=true`,
+    );
+    expect(RESOURCE_EVERY_MODE_LIMITS_SUMMARY).toContain(
+      RESOURCE_AI_WRITE_TARGETS_ENV,
+    );
+    expect(
+      RESOURCE_EVERY_MODE_LIMITS_SUMMARY.endsWith(
+        `and ${RESOURCE_UNATTENDED_ROUND_BECOMES_PROPOSAL_SUMMARY}.`,
+      ),
+    ).toBe(true);
+    expect(RESOURCE_UNATTENDED_ROUND_BECOMES_PROPOSAL_SUMMARY).toBe(
+      "an unattended run becomes a proposal when the hourly per-resource circuit breaker trips or another unattended round already holds the resource",
+    );
+  });
+
+  it("the Automatic and Bypass summaries state the canonical ResourceAiRemediationMode semantics", () => {
+    expect(RESOURCE_AUTOMATIC_MODE_SUMMARY).toContain(
+      "when the round could only find riskier fixes it ends by proposing exactly those for one-click approval",
+    );
+    expect(RESOURCE_AUTOMATIC_MODE_SUMMARY).toContain(
+      "a riskier fix is proposed only if verification shows the safe ones did not recover the signal (the follow-up round, which asks)",
+    );
+    expect(RESOURCE_AUTOMATIC_MODE_SUMMARY).toContain(
+      "Shapes on the resource's command allowlist run on their own.",
+    );
+    expect(RESOURCE_BYPASS_MODE_SUMMARY).toContain("AI does not ask.");
+    expect(RESOURCE_BYPASS_MODE_SUMMARY).toContain(
+      "Every change the policy allows — safe AND riskier — runs on its own, follow-up rounds included, except for what always needs a human.",
+    );
+    expect(RESOURCE_BYPASS_MODE_SUMMARY).not.toContain("below");
+  });
+});
+
+describe("the undo every persona recommends is a safe change the policy accepts", () => {
+  /*
+   * A rollback runs unattended: on an Automatic resource the planner refuses
+   * a riskier one. Every concrete undo the personas name must therefore be
+   * a SafeWrite on ONE named object — filled in with sample names.
+   */
+  const SAMPLE_UNDOS: Array<[AiResourceType, string]> = [
+    [AiResourceType.DockerHost, "docker start web"],
+    [AiResourceType.DockerHost, "docker unpause web"],
+    [AiResourceType.PodmanHost, "docker start web"],
+    [AiResourceType.DockerSwarmCluster, "docker service rollback web"],
+    [AiResourceType.DockerSwarmCluster, "docker service scale web=3"],
+    [
+      AiResourceType.ProxmoxCluster,
+      "pvesh create /nodes/pve1/qemu/100/status/start",
+    ],
+    [AiResourceType.VMwareVCenter, "govc vm.power -on /DC1/vm/web-01"],
+    [AiResourceType.CephCluster, "ceph osd in 3"],
+    [AiResourceType.CephCluster, "ceph osd unset noout"],
+    [AiResourceType.Host, "systemctl start nginx"],
+  ];
+
+  it.each(SAMPLE_UNDOS)(
+    "%s: %s is SafeWrite",
+    (resourceType: AiResourceType, command: string) => {
+      expect(
+        ResourceCommandPolicy.evaluateCommand({ resourceType, command }).tier,
+      ).toBe(ResourceCommandTier.SafeWrite);
+    },
+  );
+
+  it.each(ALL_AI_RESOURCE_TYPES)(
+    "the %s undo examples start with the type's program (or say there is none)",
+    (resourceType: AiResourceType) => {
+      const examples: string = getResourceUndoExamples(resourceType);
+      const programs: ReadonlyArray<string> =
+        AI_RESOURCE_TYPE_INFO[resourceType].programs;
+
+      if (resourceType === AiResourceType.DatabaseServer) {
+        expect(examples).toContain("omit it");
+        return;
+      }
+
+      const firstWord: string = examples.split(" ")[0] || "";
+      expect(programs).toContain(firstWord);
+    },
+  );
+
+  it("an unknown resource type gets advice, never a command", () => {
+    expect(getResourceUndoExamples("Nope" as AiResourceType)).toContain(
+      "omit the rollbackCommand",
+    );
+  });
+});
+
+describe("the resource personas use the shared summaries", () => {
+  it.each(ALL_AI_RESOURCE_TYPES)(
+    "the Automatic persona for a %s",
+    (resourceType: AiResourceType) => {
+      const resource: ResourceAiAccessStatus = readyResource(resourceType);
+      const persona: string = buildResourceFullAutoPersona({
+        bypassApproval: false,
+        resource,
+      });
+
+      expect(persona).toContain(RESOURCE_AUTOMATIC_MODE_SUMMARY);
+      expect(persona).toContain(RESOURCE_SAFE_CHANGES_SUMMARY);
+      expect(persona).toContain(RESOURCE_RISKIER_CHANGES_SUMMARY);
+      expect(persona).toContain(RESOURCE_ALWAYS_ASKS_SUMMARY);
+      expect(persona).toContain(RESOURCE_ALLOWLIST_SUMMARY);
+      expect(persona).toContain(
+        ResourceCommandPolicy.getWriteCommandGuide(resourceType),
+      );
+      expect(persona).toContain(RUN_INFRASTRUCTURE_COMMAND_TOOL_NAME);
+      expect(persona).toContain("stepType ResourceCommand");
+      expect(persona).toContain('"web-1"');
+      expect(persona).toContain(
+        AI_RESOURCE_TYPE_INFO[resourceType].agentDisplayName,
+      );
+      for (const program of AI_RESOURCE_TYPE_INFO[resourceType].programs) {
+        expect(persona).toContain(program);
+      }
+      expect(persona).toContain(getResourceUndoExamples(resourceType));
+      expect(persona).toContain(
+        "A rollback must itself be a safe change on ONE named object",
+      );
+      // A resource's agent is never given a credential.
+      expect(persona).not.toContain("credentialId");
+      expect(persona).not.toContain("kubectl");
+    },
+  );
+
+  it("the Bypass persona asks nothing but what always needs a human", () => {
+    const persona: string = buildResourceFullAutoPersona({
+      bypassApproval: true,
+      resource: readyResource(AiResourceType.DockerHost, {
+        aiRemediationMode: ResourceAiRemediationMode.BypassApproval,
+      }),
+    });
+
+    expect(persona).toContain(RESOURCE_BYPASS_MODE_SUMMARY);
+    expect(persona).toContain("EXCEPT that");
+    expect(persona).toContain("proposes it to a human for one-click approval");
+    expect(persona).toContain(
+      RESOURCE_UNATTENDED_ROUND_BECOMES_PROPOSAL_SUMMARY,
+    );
+    expect(persona).toContain("chose to bypass approvals entirely");
+    expect(persona).not.toContain(RESOURCE_AUTOMATIC_MODE_SUMMARY);
+  });
+
+  it.each(ALL_AI_RESOURCE_TYPES)(
+    "the planning persona for a %s proposes ResourceCommand steps, never runs them",
+    (resourceType: AiResourceType) => {
+      const persona: string = buildResourceSuggestPersona({
+        resource: readyResource(resourceType),
+      });
+
+      expect(persona).toContain("REMEDIATION PLANNING");
+      expect(persona).toContain(
+        "NOTHING you propose executes until a human approves it",
+      );
+      expect(persona).toContain("propose_remediation_commands");
+      expect(persona).toContain("stepType ResourceCommand");
+      expect(persona).toContain(
+        ResourceCommandPolicy.getWriteCommandGuide(resourceType),
+      );
+      expect(persona).toContain(getResourceUndoExamples(resourceType));
+      expect(persona).not.toContain("execute_remediation_command");
+    },
+  );
+});
+
+describe("the resource round's question", () => {
+  it("names the resource and its programs, and the mode's promise", () => {
+    const automatic: string = buildResourceQuestion({
+      resource: readyResource(AiResourceType.Host),
+      mode: "FullAuto",
+    });
+    expect(automatic).toContain('Host "web-1"');
+    expect(automatic).toContain("systemctl");
+    expect(automatic).toContain(
+      "a riskier fix never runs without a human's one-click approval",
+    );
+    expect(automatic).toContain("remediate now");
+
+    const bypass: string = buildResourceQuestion({
+      resource: readyResource(AiResourceType.DockerHost, {
+        aiRemediationMode: ResourceAiRemediationMode.BypassApproval,
+      }),
+      mode: "FullAuto",
+    });
+    expect(bypass).toContain("its operator bypassed approvals");
+    expect(bypass).toContain(RESOURCE_ALWAYS_ASKS_SUMMARY);
+    expect(bypass).toContain(
+      "the round becomes a proposal if the hourly circuit breaker trips or another unattended round holds the resource",
+    );
+
+    const suggest: string = buildResourceQuestion({
+      resource: readyResource(AiResourceType.CephCluster),
+      mode: "Suggest",
+    });
+    expect(suggest).toContain('Ceph cluster "web-1"');
+    expect(suggest).toContain("compose a plan for human approval");
+    expect(suggest).not.toContain("remediate now");
+  });
+});
+
+describe("the remediation tool's command examples are what they claim", () => {
+  /*
+   * RemediationCommandToolkit's ResourceCommand schema offers these as
+   * examples of a change; each must parse as one argv and be a SafeWrite
+   * for the type it is meant for.
+   */
+  it.each(SAFE_EXAMPLES)(
+    "%s: %s tokenizes as one argv",
+    (_resourceType: AiResourceType, command: string) => {
+      expect(tokenizeResourceCommand(command).argv).toBeDefined();
+    },
+  );
+});