fix(runtime): clean PR-open task state and quote prompts safely

apps/api/src/services/persistent-agent-pool-service.ts

@@ -19,6 +19,7 @@ import type { ContainerHandle, ContainerSpec, ExecSession } from "@optio/shared"
 import { PersistentAgentPodLifecycle, parseIntEnv, type RepoImageConfig } from "@optio/shared";
 import { logger } from "../logger.js";
 import { resolveImage } from "./repo-pool-service.js";
+import { buildEnvExports } from "../utils/pod-env.js";
 
 const POD_PROVISION_TIMEOUT_MS = parseIntEnv("OPTIO_PERSISTENT_AGENT_POD_PROVISION_MS", 120_000);
 
@@ -242,17 +243,11 @@ export async function execTurnInPod(
     name: pod.podName,
   };
 
-  const envJson = JSON.stringify({ ...env, OPTIO_PERSISTENT_AGENT_TURN_ID: turnId });
-  const envB64 = Buffer.from(envJson).toString("base64");
-
   const script = [
     "set -e",
-    `eval $(echo '${envB64}' | base64 -d | python3 -c "`,
-    `import json, sys, shlex`,
-    `env = json.load(sys.stdin)`,
-    `for k, v in env.items():`,
-    `    print(f'export {k}={shlex.quote(v)}')`,
-    `")`,
+    // Env values (including the prompt) are embedded as inert single-quoted
+    // exports — see buildEnvExports.
+    ...buildEnvExports({ ...env, OPTIO_PERSISTENT_AGENT_TURN_ID: turnId }),
     `for i in $(seq 1 120); do [ -f /workspace/.ready ] && break; sleep 1; done`,
     `[ -f /workspace/.ready ] || { echo "[optio] ERROR: pod not ready after 120s"; exit 1; }`,
     `mkdir -p /workspace/turns/${turnId}`,

apps/api/src/services/repo-pool-service.test.ts

@@ -1,4 +1,7 @@
 import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { mkdtempSync, readFileSync, readdirSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
 
 // ── Mocks ───────────────────────────────────────────────────────────
 
@@ -102,6 +105,7 @@ import {
   deleteNetworkPolicy,
   killOrphanedAgentInPod,
   parseJsonEnv,
+  execTaskInRepoPod,
 } from "./repo-pool-service.js";
 
 // ── resolveImage ────────────────────────────────────────────────────
@@ -1170,3 +1174,72 @@ describe("getOrCreateRepoPod — service account propagation", () => {
     expect(spec.serviceAccountName).toBeUndefined();
   });
 });
+
+// ── execTaskInRepoPod — env injection safety ────────────────────────
+
+describe("execTaskInRepoPod", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("embeds a hostile prompt inertly — nothing executes before the agent command", async () => {
+    // Regression for the Phase 4F symptoms ("n: command not found",
+    // "optio/task-*: No such file or directory"): prompts with markdown
+    // backticks, $HOME, wildcards, and literal newlines must reach the pod
+    // byte-for-byte without the shell interpreting any of it.
+    const hostilePrompt = [
+      "Fix the `pr_opened` handling.",
+      "",
+      "1. Inspect $HOME and run `git status`.",
+      "2. Ignore branches named optio/task-* entirely.",
+      "3. Don't rewrite 'quoted' text.",
+      "```bash",
+      "touch injected-from-fenced-block",
+      "```",
+      "$(touch injected-from-substitution)",
+    ].join("\n");
+
+    mockRuntimeExec.mockResolvedValueOnce({ stdin: { write: vi.fn() } });
+    const pod = {
+      id: "pod-1",
+      repoUrl: "https://github.com/org/repo",
+      podName: "optio-repo-org-repo-0",
+      podId: "k8s-pod-1",
+      state: "ready",
+    };
+
+    await execTaskInRepoPod(pod as any, "task-1", [`echo "[optio] agent"`], {
+      OPTIO_PROMPT: hostilePrompt,
+      OPTIO_REPO_BRANCH: "main",
+    });
+
+    const execCall = mockRuntimeExec.mock.calls[0];
+    expect(execCall[1][0]).toBe("bash");
+    expect(execCall[1][1]).toBe("-c");
+    const script: string = execCall[1][2];
+    expect(script).not.toContain("eval $(");
+
+    // Run the script prefix (set -e + env exports) through real bash and
+    // verify the prompt round-trips exactly with no side effects.
+    const lines = script.split("\n");
+    const readyIdx = lines.findIndex((l) => l.includes("Waiting for repo to be ready"));
+    expect(readyIdx).toBeGreaterThan(0);
+    const prefix = lines.slice(0, readyIdx).join("\n");
+
+    const { execFileSync } =
+      await vi.importActual<typeof import("node:child_process")>("node:child_process");
+    const dir = mkdtempSync(join(tmpdir(), "repo-pool-env-"));
+    try {
+      execFileSync("bash", ["-c", `${prefix}\nprintf '%s' "$OPTIO_PROMPT" > prompt-out`], {
+        cwd: dir,
+        encoding: "utf8",
+        stdio: ["ignore", "pipe", "pipe"],
+      });
+      expect(readFileSync(join(dir, "prompt-out"), "utf8")).toBe(hostilePrompt);
+      // No canary files — prompt contents never executed
+      expect(readdirSync(dir)).toEqual(["prompt-out"]);
+    } finally {
+      rmSync(dir, { recursive: true, force: true });
+    }
+  });
+});

apps/api/src/services/repo-pool-service.ts

@@ -29,6 +29,7 @@ import {
 } from "./envoy-sidecar.js";
 import { parseIntEnv } from "@optio/shared";
 import { withSpan } from "../telemetry/spans.js";
+import { buildEnvExports } from "../utils/pod-env.js";
 
 const IDLE_TIMEOUT_MS = parseIntEnv("OPTIO_REPO_POD_IDLE_MS", 600000); // 10 min default
 const REPO_INIT_TIMEOUT_MS = parseIntEnv("OPTIO_REPO_INIT_TIMEOUT_MS", 120000); // 2 min default
@@ -902,13 +903,13 @@ export async function execTaskInRepoPod(
         .set({ worktreeState: "active", lastPodId: pod.id, updatedAt: new Date() })
         .where(eq(tasks.id, taskId));
 
-      // Build the exec command
-      const envJson = JSON.stringify({
+      // Build the exec command. Env values (including the task prompt) are
+      // embedded as inert single-quoted exports — see buildEnvExports.
+      const envExports = buildEnvExports({
         ...env,
         OPTIO_TASK_ID: taskId,
         REPO_INIT_TIMEOUT_SECS: String(Math.ceil(REPO_INIT_TIMEOUT_MS / 1000)),
       });
-      const envB64 = Buffer.from(envJson).toString("base64");
       const runToken = randomUUID();
 
       // Build worktree setup commands based on whether we're resetting or creating fresh
@@ -964,12 +965,7 @@ export async function execTaskInRepoPod(
 
       const script = [
         "set -e",
-        `eval $(echo '${envB64}' | base64 -d | python3 -c "`,
-        `import json, sys, shlex`,
-        `env = json.load(sys.stdin)`,
-        `for k, v in env.items():`,
-        `    print(f'export {k}={shlex.quote(v)}')`,
-        `")`,
+        ...envExports,
         `echo "[optio] Waiting for repo to be ready..."`,
         `for i in $(seq 1 \${REPO_INIT_TIMEOUT_SECS}); do [ -f /workspace/.ready ] && break; sleep 1; done`,
         `[ -f /workspace/.ready ] || { echo "[optio] ERROR: repo not ready after \${REPO_INIT_TIMEOUT_SECS}s (increase OPTIO_REPO_INIT_TIMEOUT_MS to extend)"; exit 1; }`,

apps/api/src/services/task-service.test.ts

@@ -35,6 +35,9 @@ vi.mock("../db/schema.js", () => ({
 }));
 
 vi.mock("./event-bus.js", () => ({ publishEvent: vi.fn() }));
+vi.mock("../workers/webhook-worker.js", () => ({
+  enqueueWebhookEvent: vi.fn().mockResolvedValue(undefined),
+}));
 vi.mock("../logger.js", () => ({
   logger: { info: vi.fn(), warn: vi.fn(), error: vi.fn() },
 }));
@@ -133,6 +136,34 @@ describe("transitionTask", () => {
     );
   });
 
+  it("clears stale errorMessage and resultSummary when a PR is detected (pr_opened)", async () => {
+    // Regression: the agent can exit non-zero after opening a valid PR.
+    // updateTaskResult persists e.g. "Exit code: 1" before PR detection runs,
+    // so the pr_opened transition must wipe those stale fields — a task with
+    // an open PR must not look like a failed task.
+    const task = {
+      id: "t1",
+      state: "running",
+      startedAt: new Date(),
+      ticketSource: null,
+      errorMessage: "Exit code: 1",
+      resultSummary: "Agent exited with code 1",
+    };
+    vi.mocked(db.select().from(undefined as any).where).mockResolvedValueOnce([task]);
+    vi.mocked(db as any).returning.mockResolvedValueOnce([
+      { ...task, state: "pr_opened", errorMessage: null, resultSummary: null },
+    ]);
+    vi.mocked(db.insert(undefined as any).values).mockResolvedValueOnce(undefined as any);
+    await transitionTask("t1", TaskState.PR_OPENED, "pr_detected", "https://github.com/o/r/pull/1");
+    expect(db.update(undefined as any).set).toHaveBeenCalledWith(
+      expect.objectContaining({
+        state: TaskState.PR_OPENED,
+        errorMessage: null,
+        resultSummary: null,
+      }),
+    );
+  });
+
   it("throws StateRaceError when atomic update returns 0 rows", async () => {
     const task = { id: "t1", state: "queued", startedAt: null };
     vi.mocked(db.select().from(undefined as any).where)

apps/api/src/services/task-service.ts

@@ -250,7 +250,10 @@ export async function transitionTask(
     updateFields.completedAt = new Date();
   }
   // Clear error fields on successful completion (PR merged after prior errors)
-  if (toState === TaskState.COMPLETED) {
+  // and on PR-open: the agent can exit non-zero after opening a valid PR, in
+  // which case updateTaskResult has already persisted e.g. "Exit code: 1".
+  // A task with an open PR is not completed, but it must not look failed.
+  if (toState === TaskState.COMPLETED || toState === TaskState.PR_OPENED) {
     updateFields.errorMessage = null;
     updateFields.resultSummary = null;
   }

apps/api/src/services/workflow-pool-service.test.ts

@@ -462,8 +462,9 @@ describe("execRunInPod", () => {
     const execCall = mockRuntimeExec.mock.calls[0];
     expect(execCall[1][0]).toBe("bash");
     expect(execCall[1][1]).toBe("-c");
-    // The script should contain the base64-encoded env
-    expect(execCall[1][2]).toContain("base64");
+    // Env vars are embedded as inert single-quoted exports (no eval/word-splitting)
+    expect(execCall[1][2]).toContain("export MY_VAR='hello'");
+    expect(execCall[1][2]).not.toContain("eval $(");
     // And cd into per-run working directory
     expect(execCall[1][2]).toContain("/workspace/runs/run-1");
   });

apps/api/src/services/workflow-pool-service.ts

@@ -12,6 +12,7 @@ import {
 import { logger } from "../logger.js";
 import { resolveImage } from "./repo-pool-service.js";
 import { getWorkloadManager, isStatefulSetEnabled } from "./k8s-workload-service.js";
+import { buildEnvExports } from "../utils/pod-env.js";
 
 const IDLE_TIMEOUT_MS = parseIntEnv("OPTIO_WORKFLOW_POD_IDLE_MS", 600000); // 10 min default
 
@@ -396,17 +397,11 @@ export async function execRunInPod(
     })
     .where(eq(workflowPods.id, pod.id));
 
-  const envJson = JSON.stringify({ ...env, OPTIO_WORKFLOW_RUN_ID: runId });
-  const envB64 = Buffer.from(envJson).toString("base64");
-
   const script = [
     "set -e",
-    `eval $(echo '${envB64}' | base64 -d | python3 -c "`,
-    `import json, sys, shlex`,
-    `env = json.load(sys.stdin)`,
-    `for k, v in env.items():`,
-    `    print(f'export {k}={shlex.quote(v)}')`,
-    `")`,
+    // Env values (including the prompt) are embedded as inert single-quoted
+    // exports — see buildEnvExports.
+    ...buildEnvExports({ ...env, OPTIO_WORKFLOW_RUN_ID: runId }),
     `echo "[optio] Waiting for workflow pod to be ready..."`,
     `for i in $(seq 1 120); do [ -f /workspace/.ready ] && break; sleep 1; done`,
     `[ -f /workspace/.ready ] || { echo "[optio] ERROR: workflow pod not ready after 120s"; exit 1; }`,

apps/api/src/utils/pod-env.test.ts

@@ -0,0 +1,108 @@
+import { describe, it, expect } from "vitest";
+import { execFileSync } from "node:child_process";
+import { mkdtempSync, readFileSync, readdirSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { shellSingleQuote, buildEnvExports } from "./pod-env.js";
+
+/**
+ * Regression payload for the Phase 4F shell-quoting bug: a realistic task
+ * prompt with markdown backticks, `$HOME`, wildcard text like `optio/task-*`,
+ * literal newlines, and single quotes. It also carries canary commands — if
+ * any of them run, the env injection leaked prompt content to the shell.
+ */
+const HOSTILE_PROMPT = [
+  "Fix the `pr_opened` handling before broader autonomous use.",
+  "",
+  "Steps:",
+  "1. Inspect $HOME and run `git status` in the worktree.",
+  "2. Don't touch branches named optio/task-* — they belong to other agents.",
+  "3. Preserve 'single-quoted' text exactly as written.",
+  "```bash",
+  "touch injected-from-fenced-block",
+  "```",
+  "$(touch injected-from-substitution)",
+  "`touch injected-from-backquotes`",
+  "rm -rf $HOME/should-never-expand",
+].join("\n");
+
+function runBash(script: string, cwd: string): string {
+  return execFileSync("bash", ["-c", script], {
+    cwd,
+    encoding: "utf8",
+    stdio: ["ignore", "pipe", "pipe"],
+  });
+}
+
+describe("shellSingleQuote", () => {
+  it("wraps plain values in single quotes", () => {
+    expect(shellSingleQuote("hello")).toBe("'hello'");
+  });
+
+  it("escapes embedded single quotes", () => {
+    expect(shellSingleQuote("don't")).toBe("'don'\\''t'");
+  });
+
+  it("handles empty strings", () => {
+    expect(shellSingleQuote("")).toBe("''");
+  });
+});
+
+describe("buildEnvExports", () => {
+  it("emits one export statement per env entry", () => {
+    expect(buildEnvExports({ A: "1", B_2: "two" })).toEqual(["export A='1'", "export B_2='two'"]);
+  });
+
+  it("rejects env names bash would not accept as identifiers", () => {
+    expect(() => buildEnvExports({ "BAD-NAME": "x" })).toThrow(/Invalid environment variable/);
+    expect(() => buildEnvExports({ "PATH; touch pwned": "x" })).toThrow(
+      /Invalid environment variable/,
+    );
+    expect(() => buildEnvExports({ "1LEADING": "x" })).toThrow(/Invalid environment variable/);
+  });
+
+  it("round-trips a hostile prompt through bash without executing its contents", () => {
+    const dir = mkdtempSync(join(tmpdir(), "pod-env-"));
+    try {
+      const script = [
+        "set -e",
+        ...buildEnvExports({
+          OPTIO_PROMPT: HOSTILE_PROMPT,
+          OPTIO_TASK_ID: "task-1",
+        }),
+        `printf '%s' "$OPTIO_PROMPT" > prompt-out`,
+        `printf '%s' "$OPTIO_TASK_ID" > task-id-out`,
+      ].join("\n");
+
+      // Throws on non-zero exit — e.g. "command not found" from a prompt
+      // line leaking to the shell under `set -e`.
+      const stdout = runBash(script, dir);
+      expect(stdout).toBe("");
+
+      // Exact round-trip: newlines, backticks, $HOME, globs, quotes intact.
+      expect(readFileSync(join(dir, "prompt-out"), "utf8")).toBe(HOSTILE_PROMPT);
+      expect(readFileSync(join(dir, "task-id-out"), "utf8")).toBe("task-1");
+
+      // No canary commands executed — only our two output files exist.
+      expect(readdirSync(dir).sort()).toEqual(["prompt-out", "task-id-out"]);
+    } finally {
+      rmSync(dir, { recursive: true, force: true });
+    }
+  });
+
+  it("keeps values with only quotes and whitespace intact", () => {
+    const dir = mkdtempSync(join(tmpdir(), "pod-env-"));
+    try {
+      const value = `  '  "  \t  '' \n `;
+      const script = [
+        "set -e",
+        ...buildEnvExports({ TRICKY: value }),
+        `printf '%s' "$TRICKY" > out`,
+      ].join("\n");
+      runBash(script, dir);
+      expect(readFileSync(join(dir, "out"), "utf8")).toBe(value);
+    } finally {
+      rmSync(dir, { recursive: true, force: true });
+    }
+  });
+});