fix(taskctl): add taskctl permission to adversarial-pipeline and fix verdict prompt syntax

packages/opencode/src/agent/agent.ts

@@ -261,6 +261,7 @@ RULES FOR GOOD TASK DECOMPOSITION:
           defaults,
           PermissionNext.fromConfig({
             task: "deny",
+            taskctl: "deny",
           }),
           user,
         ),
@@ -281,7 +282,7 @@ You will receive a task description with:
 3. Write minimal code to make tests pass
 4. Refactor for clarity following AGENTS.md style guide
 5. Run \`bun run typecheck && bun test\` — fix all errors
-6. When done: \`taskctl comment <taskId> "Implementation complete: <one-line summary of what was built>"\`
+6. When done: simply complete your work - the pulse system automatically detects completion
 
 ## Rules
 - ONLY implement what is explicitly in the task description
@@ -298,12 +299,7 @@ You will receive a task description with:
 - **context7** — current library docs. Use resolve-library-id then query-docs before using any external library API
 - **parallel-search + web_fetch MCP tools** — web search when you need current information not in the codebase
 
-## taskctl commands available to you
-- \`taskctl comment <taskId> "<message>"\` — log progress or signal completion
-- \`taskctl split <taskId>\` — if task is too large, split it (creates two sub-tasks)
-- \`taskctl depends <taskId> --on <otherId>\` — if you discover an undeclared dependency
-
-You may NOT call: taskctl start, taskctl stop, taskctl verdict, taskctl override, taskctl retry, taskctl resume`,
+NOTE: You do NOT have access to taskctl commands. The pipeline handles task state automatically.`,
       },
       "adversarial-pipeline": {
         name: "adversarial-pipeline",
@@ -315,6 +311,7 @@ You may NOT call: taskctl start, taskctl stop, taskctl verdict, taskctl override
           PermissionNext.fromConfig({
             "*": "deny",
             bash: "allow",
+            taskctl: "allow",
           }),
           user,
         ),
@@ -336,16 +333,31 @@ Your ONLY job is to review code changes in an assigned worktree and record a str
 5. Check: Does typecheck pass? (Run \`bun run typecheck\` in the worktree)
 
 ## Recording your verdict — MANDATORY
-You MUST call taskctl verdict to record your finding. Never write a text response instead.
+
+You MUST use the \`taskctl\` MCP tool to record your verdict. This is an MCP tool in your tool list — NOT a bash command. Never write a text response instead.
+
+Use the taskctl tool with these parameters:
 
 **If the code is good:**
-\`taskctl verdict <taskId> --verdict APPROVED\`
+- command: "verdict"
+- taskId: <the task ID you were given>
+- verdict: "APPROVED"
+- verdictSummary: "Brief summary of what you reviewed and why it passes"
+- verdictIssues: []
 
 **If there are fixable issues:**
-\`taskctl verdict <taskId> --verdict ISSUES_FOUND --summary "Brief summary" --issues '[{"location":"src/foo.ts:42","severity":"HIGH","fix":"Add null check before calling user.profile"}]'\`
+- command: "verdict"
+- taskId: <the task ID>
+- verdict: "ISSUES_FOUND"
+- verdictSummary: "Brief summary"
+- verdictIssues: [{"location":"src/foo.ts:42","severity":"HIGH","fix":"Add null check before calling user.profile"}]
 
 **If there are critical/blocking issues:**
-\`taskctl verdict <taskId> --verdict CRITICAL_ISSUES_FOUND --summary "Brief summary" --issues '[...]'\`
+- command: "verdict"
+- taskId: <the task ID>
+- verdict: "CRITICAL_ISSUES_FOUND"
+- verdictSummary: "Brief summary"
+- verdictIssues: [{"location":"...","severity":"CRITICAL","fix":"..."}]
 
 ## Severity guide
 - CRITICAL: Security vulnerability, data loss risk, or complete functional failure
@@ -354,7 +366,7 @@ You MUST call taskctl verdict to record your finding. Never write a text respons
 - LOW: Style or minor improvement suggestion
 
 ## Rules
-- You may ONLY call: taskctl verdict
+- You may ONLY use: the taskctl MCP tool (command: "verdict")
 - Do NOT spawn any agents
 - Do NOT commit or push
 - Be specific: every issue must have a location (file:line) and a concrete fix suggestion

packages/opencode/src/tasks/pulse.ts

@@ -728,8 +728,7 @@ Summary: ${verdict.summary}
 Issues:
 ${issueLines}
 
-The codebase changes are already in this worktree. Fix the specific issues listed above, run tests, then signal completion with:
-taskctl comment ${task.id} "Implementation complete: <summary of what was fixed>"`
+The codebase changes are already in this worktree. Fix the specific issues listed above, run tests, and complete your work. The pulse system automatically detects completion.`
 
   try {
     await SessionPrompt.prompt({

packages/opencode/test/tasks/agent-permissions.test.ts

@@ -0,0 +1,99 @@
+import { describe, expect, test } from "bun:test"
+import { Instance } from "../../src/project/instance"
+import { Agent } from "../../src/agent/agent"
+import { PermissionNext } from "../../src/permission/next"
+import { tmpdir } from "../fixture/fixture"
+
+// Helper to evaluate permission for a tool with wildcard pattern
+function evalPerm(agent: Agent.Info | undefined, permission: string): PermissionNext.Action | undefined {
+  if (!agent) return undefined
+  return PermissionNext.evaluate(permission, "*", agent.permission).action
+}
+
+describe("agent permissions for taskctl pipeline", () => {
+  test("adversarial-pipeline agent has taskctl permission", async () => {
+    await using tmp = await tmpdir()
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const adversarial = await Agent.get("adversarial-pipeline")
+        expect(adversarial).toBeDefined()
+        expect(adversarial?.permission).toBeDefined()
+
+        // Check the raw ruleset contains taskctl allowance
+        const hasTaskctlAllow = adversarial!.permission.some(
+          (rule: any) => rule.permission === "taskctl" && rule.action === "allow"
+        )
+        expect(hasTaskctlAllow).toBe(true)
+
+        // Verify via PermissionNext.evaluate - taskctl with any pattern should be allowed
+        const result = PermissionNext.evaluate("taskctl", "*", adversarial!.permission)
+        expect(result.action).toBe("allow")
+      },
+    })
+  })
+
+  test("adversarial-pipeline agent has bash permission", async () => {
+    await using tmp = await tmpdir()
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const adversarial = await Agent.get("adversarial-pipeline")
+        expect(adversarial).toBeDefined()
+
+        // Verify bash is allowed
+        const result = PermissionNext.evaluate("bash", "*", adversarial!.permission)
+        expect(result.action).toBe("allow")
+      },
+    })
+  })
+
+  test("developer-pipeline agent explicitly denies taskctl permission", async () => {
+    await using tmp = await tmpdir()
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const developer = await Agent.get("developer-pipeline")
+        expect(developer).toBeDefined()
+
+        // Developer should explicitly deny taskctl - not just "not allow"
+        const result = PermissionNext.evaluate("taskctl", "*", developer!.permission)
+        expect(result.action).toBe("deny")  // Explicit deny, not fallback "ask"
+      },
+    })
+  })
+
+  test("adversarial-pipeline denies everything else (minimal permissions)", async () => {
+    await using tmp = await tmpdir()
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const adversarial = await Agent.get("adversarial-pipeline")
+        expect(adversarial).toBeDefined()
+
+        // Check that common tools are denied (wildcard deny rule)
+        const deniedTools = ["edit", "write", "read", "task"]
+
+        for (const tool of deniedTools) {
+          const result = PermissionNext.evaluate(tool, "*", adversarial!.permission)
+          expect(result.action).toBe("deny")
+        }
+      },
+    })
+  })
+
+  test("developer-pipeline denies task tool", async () => {
+    await using tmp = await tmpdir()
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const developer = await Agent.get("developer-pipeline")
+        expect(developer).toBeDefined()
+
+        // Developer should NOT be able to spawn other agents via task tool
+        const result = PermissionNext.evaluate("task", "*", developer!.permission)
+        expect(result.action).toBe("deny")
+      },
+    })
+  })
+})