@agent-compose/sdk 0.5.9 → 0.6.0

package/src/step-invocation/invoker.ts

@@ -24,7 +24,7 @@
 import { randomBytes } from "node:crypto";
 import { z } from "zod";
 import { SandboxUnavailableError } from "../sandbox-errors.js";
-import type { SandboxProvider } from "../types/sandbox.js";
+import type { SandboxProvider, SandboxBackgroundProcess } from "../types/sandbox.js";
 import { RUNNER_COMMAND, STEP_ENV, stepResultLinePrefix, stepPauseLinePrefix, requestContextPath, stepInputPath, stepResultFilePath } from "./protocol.js";
 import { StepPauseRequestSchema } from "./types.js";
 import type { StepRequest, StepResult } from "./types.js";
@@ -196,40 +196,34 @@ export interface InvokeStepOptions {
   onStderr?: (line: string) => void;
 }
 
-export async function invokeStep<TOutput = unknown>(
-  sandbox: SandboxProvider,
-  request: StepRequest,
-  opts?: InvokeStepOptions,
-): Promise<StepResult<TOutput>> {
-  const resultToken = randomBytes(16).toString("hex");
-
-  await Promise.all([
-    sandbox.files.write(stepInputPath(request.stepIndex),      JSON.stringify(request.input)),
-    sandbox.files.write(requestContextPath(request.stepIndex), JSON.stringify(request.requestContext)),
-  ]);
-
-  const envs = {
-    ...(opts?.envs ?? {}),
-    ...buildStepEnvs({
-      runId:       request.runId,
-      stepIndex:   request.stepIndex,
-      resultToken,
-      isResume:    opts?.isResume,
-    }),
-  };
+/** A step running as a background command (ADR-0028). The activity races its
+ *  `wait()` against a server pause request; on a pause it freezes the VM
+ *  (`pauseProcess`) and persists `runnerPid` + `resultToken` so the resume
+ *  activity can `reconnectStep(...)` to the SAME process — no re-run. */
+export interface RunningStep<TOutput = unknown> {
+  /** OS pid of the background runner inside the VM — the reconnect handle. */
+  runnerPid: number;
+  /** Per-invocation token keying the durable result file the runner writes. */
+  resultToken: string;
+  /** Await the runner's exit and classify its output into a `StepResult`. */
+  wait(): Promise<StepResult<TOutput>>;
+}
 
-  // The sandbox emits stdout as raw chunks, not lines. Buffer between
-  // emissions so a `console.log` split across two chunks (or a partial
-  // trailing line) is delivered to onStdout/onStderr as one logical line.
-  // Sentinel lines (carrying `resultToken`) are filtered out so callers
-  // never see protocol bytes in user-log capture. Both the result and
-  // pause prefixes are built from the same shared helpers `parseStepResult`
-  // uses, so the filter and the parser can't drift apart.
+/** Buffer the sandbox's raw stdout/stderr chunks into whole lines, filtering
+ *  the protocol sentinel out of `onStdout` so user-log capture never sees
+ *  protocol bytes. Shared by the foreground (`invokeStep`), background
+ *  (`launchStep`), and resume (`reconnectStep`) paths so the filter and the
+ *  result parser can't drift. Returns the per-stream chunk handlers (undefined
+ *  when the caller passed no sink) plus a `flush` for trailing newline-less
+ *  output. */
+function makeStreamSplitters(
+  opts: Pick<InvokeStepOptions, "onStdout" | "onStderr"> | undefined,
+  resultToken: string,
+): { onStdout?: (chunk: string) => void; onStderr?: (chunk: string) => void; flush: () => void } {
   const resultSentinel = stepResultLinePrefix(resultToken);
   const pauseSentinel  = stepPauseLinePrefix(resultToken);
-  const isSentinel = (line: string) =>
-    line.startsWith(resultSentinel) || line.startsWith(pauseSentinel);
-  const makeLineSplitter = (sink: ((line: string) => void) | undefined, filterSentinel: boolean) => {
+  const isSentinel = (line: string) => line.startsWith(resultSentinel) || line.startsWith(pauseSentinel);
+  const make = (sink: ((line: string) => void) | undefined, filterSentinel: boolean) => {
     if (!sink) return { onChunk: undefined, flush: () => {} };
     let buf = "";
     return {
@@ -243,10 +237,6 @@ export async function invokeStep<TOutput = unknown>(
           sink(line);
         }
       },
-      // Drain any remaining buffered output that ended without a newline.
-      // Called after `sandbox.commands.run` resolves so a runner that
-      // exits with `process.stdout.write("final")` (no trailing \n)
-      // doesn't silently drop its last line.
       flush: () => {
         if (buf.length === 0) return;
         const line = buf;
@@ -256,16 +246,102 @@ export async function invokeStep<TOutput = unknown>(
       },
     };
   };
-  const stdoutSplitter = makeLineSplitter(opts?.onStdout, true);
-  const stderrSplitter = makeLineSplitter(opts?.onStderr, false);
+  const stdout = make(opts?.onStdout, true);
+  const stderr = make(opts?.onStderr, false);
+  return {
+    ...(stdout.onChunk ? { onStdout: stdout.onChunk } : {}),
+    ...(stderr.onChunk ? { onStderr: stderr.onChunk } : {}),
+    flush: () => { stdout.flush(); stderr.flush(); },
+  };
+}
+
+/** Turn the runner's terminal output into a kinded `StepResult`. stdout is the
+ *  fast path; the runner also persists the sentinel to a durable token-keyed
+ *  file, so when stdout carries no sentinel (providers drop the tail of heavy
+ *  streams; a reconnect captures only post-resume output) we read the file
+ *  back. Shared by all three invocation paths. */
+async function classifyRunnerOutcome<TOutput>(
+  sandbox: SandboxProvider,
+  output: { stdout: string; stderr: string; exitCode: number },
+  resultToken: string,
+): Promise<StepResult<TOutput>> {
+  const parsed = parseStepResult<TOutput>(output.stdout, resultToken);
+  if (parsed) return parsed;
+
+  // No sentinel on stdout — read the durable token-keyed file the runner wrote
+  // BEFORE its stdout emit. A runner that exited 1 after a user-step throw
+  // still wrote `{ok:false, kind:"user-step"}`, which beats a generic exit.
+  const fileRead = await sandbox.commands.run(
+    `cat ${stepResultFilePath(resultToken)} 2>/dev/null || true`,
+  ).catch(() => null);
+  if (fileRead?.stdout) {
+    const fromFile = parseStepResult<TOutput>(fileRead.stdout, resultToken);
+    if (fromFile) return fromFile;
+  }
+
+  // Distinguish "runner exited badly before emitting" (runner-exit) from
+  // "runner exited cleanly but didn't speak the protocol" (protocol).
+  if (output.exitCode !== 0) {
+    const stderrTail = output.stderr.trim().slice(-2000);
+    const stdoutTail = output.stdout.trim().slice(-2000);
+    const details = [
+      stderrTail ? `stderr:\n${stderrTail}` : "",
+      stdoutTail ? `stdout:\n${stdoutTail}` : "",
+    ].filter(Boolean).join("\n");
+    return {
+      ok: false,
+      error: {
+        kind:     "runner-exit",
+        message:  `runner subprocess exited ${output.exitCode} before emitting a step result${details ? `\n${details}` : ""}`,
+        exitCode: output.exitCode,
+      },
+    };
+  }
+  const tail = output.stdout.trim().slice(-500);
+  return {
+    ok: false,
+    error: {
+      kind: "protocol",
+      message: `no tokenised step result on stdout or in the result file (runner exited 0 without emitting)${tail ? `\nstdout tail:\n${tail}` : ""}`,
+    },
+  };
+}
+
+/** Write the step's input + request-context files and build the runner env.
+ *  Shared by the foreground and background launch paths. Returns the
+ *  per-invocation `resultToken` + the env map. */
+async function prepareStepLaunch(
+  sandbox: SandboxProvider,
+  request: StepRequest,
+  opts?: InvokeStepOptions,
+): Promise<{ resultToken: string; envs: Record<string, string> }> {
+  const resultToken = randomBytes(16).toString("hex");
+  await Promise.all([
+    sandbox.files.write(stepInputPath(request.stepIndex),      JSON.stringify(request.input)),
+    sandbox.files.write(requestContextPath(request.stepIndex), JSON.stringify(request.requestContext)),
+  ]);
+  const envs = {
+    ...(opts?.envs ?? {}),
+    ...buildStepEnvs({ runId: request.runId, stepIndex: request.stepIndex, resultToken, isResume: opts?.isResume }),
+  };
+  return { resultToken, envs };
+}
+
+export async function invokeStep<TOutput = unknown>(
+  sandbox: SandboxProvider,
+  request: StepRequest,
+  opts?: InvokeStepOptions,
+): Promise<StepResult<TOutput>> {
+  const { resultToken, envs } = await prepareStepLaunch(sandbox, request, opts);
+  const splitters = makeStreamSplitters(opts, resultToken);
 
   let result: { stdout: string; stderr: string; exitCode: number };
   try {
     result = await sandbox.commands.run(RUNNER_COMMAND, {
       envs,
       timeoutMs: 0,
-      ...(stdoutSplitter.onChunk ? { onStdout: stdoutSplitter.onChunk } : {}),
-      ...(stderrSplitter.onChunk ? { onStderr: stderrSplitter.onChunk } : {}),
+      ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
+      ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
     });
   } catch (e) {
     // Typed sandbox-infrastructure failure: the engine's recovery contract
@@ -286,55 +362,90 @@ export async function invokeStep<TOutput = unknown>(
       exitCode: ce.exitCode ?? ce.result?.exitCode ?? 1,
     };
   }
-  stdoutSplitter.flush();
-  stderrSplitter.flush();
-
-  const parsed = parseStepResult<TOutput>(result.stdout, resultToken);
-  if (parsed) return parsed;
+  splitters.flush();
+  return classifyRunnerOutcome<TOutput>(sandbox, result, resultToken);
+}
 
-  // No sentinel on stdout. The runner also persists the sentinel line to a
-  // token-keyed file before emitting — providers drop the tail of heavy
-  // stdout streams, so read the durable copy back. Tried before exit-code
-  // classification: a runner that exited 1 after a user-step throw still
-  // wrote the file, and its `{ok:false, kind:"user-step"}` beats a generic
-  // runner-exit error. A runner killed before emitting wrote no file and
-  // falls through.
-  const fileRead = await sandbox.commands.run(
-    `cat ${stepResultFilePath(resultToken)} 2>/dev/null || true`,
-  ).catch(() => null);
-  if (fileRead?.stdout) {
-    const fromFile = parseStepResult<TOutput>(fileRead.stdout, resultToken);
-    if (fromFile) return fromFile;
+/**
+ * Launch the step runner as a BACKGROUND command (ADR-0028) and return a handle
+ * the activity drives: it races `wait()` against a server pause request and, on
+ * a pause, freezes the VM (`pauseProcess`) and persists `runnerPid` +
+ * `resultToken` so `reconnectStep` can continue the SAME process — no re-run.
+ * Requires a provider with `commands.runBackground` (E2B); the foreground
+ * `invokeStep` is the path for providers without it (Vercel).
+ */
+export async function launchStep<TOutput = unknown>(
+  sandbox: SandboxProvider,
+  request: StepRequest,
+  opts?: InvokeStepOptions,
+): Promise<RunningStep<TOutput>> {
+  if (!sandbox.commands.runBackground) {
+    throw new Error("launchStep requires a provider with background-command support (commands.runBackground)");
   }
+  const { resultToken, envs } = await prepareStepLaunch(sandbox, request, opts);
+  const splitters = makeStreamSplitters(opts, resultToken);
+  const proc = await sandbox.commands.runBackground(RUNNER_COMMAND, {
+    envs,
+    timeoutMs: 0,
+    ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
+    ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
+  });
+  return {
+    runnerPid:   proc.pid,
+    resultToken,
+    async wait() {
+      const result = await proc.wait();
+      splitters.flush();
+      return classifyRunnerOutcome<TOutput>(sandbox, result, resultToken);
+    },
+  };
+}
 
-  // No tokenised sentinel on stdout. Distinguish "runner exited badly
-  // before emitting" (runner-exit) from "runner exited cleanly but didn't
-  // speak the protocol" (protocol) — the exit code is the evidence.
-  if (result.exitCode !== 0) {
-    const stderrTail = result.stderr.trim().slice(-2000);
-    const stdoutTail = result.stdout.trim().slice(-2000);
-    const details = [
-      stderrTail ? `stderr:\n${stderrTail}` : "",
-      stdoutTail ? `stdout:\n${stdoutTail}` : "",
-    ].filter(Boolean).join("\n");
-    return {
-      ok: false,
-      error: {
-        kind:     "runner-exit",
-        message:  `runner subprocess exited ${result.exitCode} before emitting a step result${details ? `\n${details}` : ""}`,
-        exitCode: result.exitCode,
-      },
-    };
+/**
+ * Resume a previously-paused background runner (ADR-0028) and return a
+ * `RunningStep` handle — uniform with `launchStep` so the activity can race
+ * `wait()` against a fresh pause request (a resumed step can pause again). After
+ * the workflow reconnects the suspended VM (`Sandbox.connect` auto-resumes it),
+ * this re-attaches to the still-running runner by `runnerPid`; `wait()` awaits
+ * its exit (event-driven — no polling) and classifies the output (the durable
+ * result file is authoritative on this path). If the runner already exited
+ * during resume, the re-attach fails and `wait()` classifies from the file.
+ */
+export async function reconnectStep<TOutput = unknown>(
+  sandbox: SandboxProvider,
+  resume: { runnerPid: number; resultToken: string; stepIndex: number },
+  opts?: Pick<InvokeStepOptions, "onStdout" | "onStderr">,
+): Promise<RunningStep<TOutput>> {
+  if (!sandbox.commands.connectProcess) {
+    throw new Error("reconnectStep requires a provider with background-command support (commands.connectProcess)");
+  }
+  const { runnerPid, resultToken } = resume;
+  const splitters = makeStreamSplitters(opts, resultToken);
+  let proc: SandboxBackgroundProcess | null = null;
+  try {
+    proc = await sandbox.commands.connectProcess(runnerPid, {
+      ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
+      ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
+    });
+  } catch (e) {
+    // A sandbox-infrastructure failure must propagate intact (recovery
+    // contract). Anything else means the runner already exited during resume
+    // (we re-attached too late) — its durable result file is written, so the
+    // handle's wait() classifies from the file below.
+    if (e instanceof SandboxUnavailableError) throw e;
   }
-  // Carry the stdout tail: "no tokenised step result" alone is useless to
-  // an operator — the tail usually shows whether the runner finished its
-  // work (output truncated by the provider) or never got there.
-  const tail = result.stdout.trim().slice(-500);
+  const attached = proc;
   return {
-    ok: false,
-    error: {
-      kind: "protocol",
-      message: `no tokenised step result on stdout or in the result file (runner exited 0 without emitting)${tail ? `\nstdout tail:\n${tail}` : ""}`,
+    runnerPid,
+    resultToken,
+    async wait() {
+      let output: { stdout: string; stderr: string; exitCode: number } = { stdout: "", stderr: "", exitCode: 0 };
+      if (attached) {
+        try { output = await attached.wait(); }
+        catch (e) { if (e instanceof SandboxUnavailableError) throw e; }
+      }
+      splitters.flush();
+      return classifyRunnerOutcome<TOutput>(sandbox, output, resultToken);
     },
   };
 }

package/src/types/runtime.ts

@@ -32,6 +32,13 @@ export interface RuntimeOptions {
   /** Agent id and label for processor context / adapter logs. */
   agentId?: string;
   iteration?: number;
+  /** The platform manual (file conventions, connectors & access, how to pause).
+   *  `agent()` builds it per-run (`buildAgentContextDoc`) and threads it here so
+   *  a runtime that supports a system-prompt append (the claude runtime) injects
+   *  it directly — instead of relying on the agent to `cat` the on-disk
+   *  AGENTS.md/CLAUDE.md, which the Agent SDK doesn't auto-load and which can
+   *  fail to write on a read-only/degraded working dir. */
+  agentManual?: string;
   /** The run's pause boundary, threaded from the agent loop so a runtime-driven
    *  pre-tool gate (e.g. the ACP `session/request_permission` path through
    *  `gateToolCall`) can raise a human-approval `ctx.pause`. The runtime binds it

package/src/types/sandbox.ts

@@ -25,6 +25,30 @@ export interface SandboxCommandResult {
   stderr: string;
 }
 
+/** A long-running command launched in the background (ADR-0028). Unlike
+ *  `commands.run` (which awaits completion on one connection), a background
+ *  command keeps running inside the VM independent of the launching
+ *  connection: it survives `pauseProcess()`/resume and is re-attachable by
+ *  `pid` after a fresh `Sandbox.connect`. This is what lets the platform
+ *  freeze an agent mid-turn for a human-in-the-loop pause and continue the
+ *  SAME process on resume — no re-run.
+ *
+ *  Implemented ONLY by process-resume-capable providers (E2B); the presence
+ *  of `commands.runBackground` IS the capability flag, paired with
+ *  `pauseProcess`. Providers without it leave both undefined and pause via
+ *  `snapshot()` + re-run instead. */
+export interface SandboxBackgroundProcess {
+  /** OS pid inside the VM — the durable handle used to reconnect after a
+   *  pause/resume cycle via `commands.connectProcess(pid)`. */
+  pid: number;
+  /** Resolve when the process exits, with its buffered result. Live output
+   *  streams to the `onStdout`/`onStderr` passed at launch / connect time.
+   *  Does NOT throw on a non-zero exit — the result carries `exitCode`. */
+  wait(): Promise<SandboxCommandResult>;
+  /** Force-terminate the process. */
+  kill(): Promise<void>;
+}
+
 /** A spawned long-lived command with a writable stdin and readable stdout,
  *  exposed as byte web-streams. Unlike `commands.run` (which buffers to
  *  completion and exposes stdout only via an `onStdout` callback), a duplex
@@ -85,6 +109,16 @@ export interface SandboxProvider {
      *  view). The vercel/e2b providers (server→sandbox) leave it undefined; an
      *  ACP caller that finds it absent falls back to the JSONL transport. */
     spawnDuplex?(cmd: string, opts?: SandboxSpawnDuplexOptions): SandboxDuplexProcess;
+    /** Launch a command in the background and return immediately with a
+     *  reconnectable handle (ADR-0028). The process survives the launching
+     *  connection dropping AND a `pauseProcess()`/resume cycle. OPTIONAL —
+     *  only process-resume providers (E2B) implement it; its presence (paired
+     *  with `pauseProcess`) is the native-pause capability flag. */
+    runBackground?(cmd: string, opts?: SandboxCommandRunOptions): Promise<SandboxBackgroundProcess>;
+    /** Re-attach to a background command by `pid` after a fresh
+     *  `Sandbox.connect` (the resume half of `runBackground`). OPTIONAL,
+     *  E2B-only. Throws if no process with that pid is running. */
+    connectProcess?(pid: number, opts?: Pick<SandboxCommandRunOptions, "onStdout" | "onStderr" | "timeoutMs">): Promise<SandboxBackgroundProcess>;
   };
   files: {
     write(path: string, content: string): Promise<void>;
@@ -99,6 +133,16 @@ export interface SandboxProvider {
    *  be omitted when the provider doesn't expose it; the server stores
    *  `null` for missing values rather than estimating. */
   snapshot?(): Promise<{ snapshotId: string; sizeBytes?: number }>;
+  /** Suspend the live VM in place and return a handle to resume it (ADR-0027).
+   *  Present ONLY on process-resume-capable providers (E2B via `sandbox.pause()`,
+   *  returning the sandbox id; resume is `Sandbox.connect(handle)`, which
+   *  auto-resumes the paused VM). Unlike `snapshot()` — which captures an FS
+   *  image, kills the origin, and re-runs the step from a fresh sandbox — a
+   *  process-resume pause FREEZES the live process (zero compute) and continues
+   *  it exactly where it blocked. The presence of this method IS the capability
+   *  flag: providers without native VM-suspend leave it undefined and fall back
+   *  to `snapshot()` + re-run. */
+  pauseProcess?(): Promise<{ resumeHandle: string }>;
   /** Replace the live sandbox's egress policy in place — so the server can
    *  push a freshly resolved policy (with re-minted connector access tokens)
    *  before each step instead of relying on the policy baked at create.

package/dist/agent/local-pause-request.d.ts

@@ -1,49 +0,0 @@
-/**
- * Local pause-request marker — the in-sandbox bridge from `agentc pause` to
- * the agent loop's snapshot-release pause boundary.
- *
- * `agentc pause` runs as a grandchild subprocess of the runner (the agent CLI
- * shells out to it). It cannot throw a `PauseSignal` into the loop and the
- * in-memory steer flag (`signalSteerPending`) lives in a different process, so
- * the only reliable channel is the shared sandbox filesystem. The CLI writes a
- * durable marker here; the agent loop consumes it at the end of the turn
- * (alongside the `needs_input` self-pause) and stages a real `ctx.pause` the
- * next boundary takes — which snapshots the sandbox, releases the activity
- * (compute stops), and parks the workflow. On resume the human's answer is
- * delivered as the agent's next user turn.
- *
- * This is the ONLY place the marker path + shape are defined — both the CLI
- * (writer) and the SDK loop (reader) import it, so the two halves can never
- * drift. Like the rest of the state-dir, the layout is a wire protocol between
- * the runner and the next subprocess invocation (ADR-0006). The marker is
- * consumed BEFORE the snapshot, so it never needs to survive a pause.
- *
- * Scoped by `agentId` so concurrent `agent()` calls sharing one sandbox each
- * see only their own request. When the id is unavailable (older agent env that
- * doesn't inject `AGENT_COMPOSE_AGENT_ID`) both sides fall back to a single
- * unscoped slot — correct for the common single-agent step, and the only case
- * where an unscoped marker can be ambiguous (two anonymous agents) is one the
- * old block-poll CLI couldn't handle either.
- */
-/** A choice offered to the human. A bare string is shorthand for
- *  `{ label, value }` with both equal — exactly what the dashboard's
- *  `readOptions` accepts. */
-export type PauseOption = string | {
-    label: string;
-    value: string;
-};
-/** What `agentc pause` records for the loop to turn into a `ctx.pause`. */
-export interface LocalPauseRequest {
-    /** The question shown to the human (becomes the pause `reason`). */
-    reason: string;
-    /** Optional offered choices, rendered as buttons in the dashboard. */
-    options?: PauseOption[];
-}
-/** Write the marker atomically (tmp → rename) so the loop never reads a
- *  partial file mid-write. Called by `agentc pause`. */
-export declare function writeLocalPauseRequest(agentId: string | undefined | null, req: LocalPauseRequest): void;
-/** Take-once read: returns + deletes this agent's pending pause request, else
- *  null. Checks the agent-scoped slot first, then the unscoped fallback. The
- *  loop calls this once per turn; a malformed marker is dropped (deleted +
- *  null) rather than wedging the loop. */
-export declare function consumeLocalPauseRequest(agentId: string | undefined | null): LocalPauseRequest | null;

package/src/agent/local-pause-request.ts

@@ -1,90 +0,0 @@
-/**
- * Local pause-request marker — the in-sandbox bridge from `agentc pause` to
- * the agent loop's snapshot-release pause boundary.
- *
- * `agentc pause` runs as a grandchild subprocess of the runner (the agent CLI
- * shells out to it). It cannot throw a `PauseSignal` into the loop and the
- * in-memory steer flag (`signalSteerPending`) lives in a different process, so
- * the only reliable channel is the shared sandbox filesystem. The CLI writes a
- * durable marker here; the agent loop consumes it at the end of the turn
- * (alongside the `needs_input` self-pause) and stages a real `ctx.pause` the
- * next boundary takes — which snapshots the sandbox, releases the activity
- * (compute stops), and parks the workflow. On resume the human's answer is
- * delivered as the agent's next user turn.
- *
- * This is the ONLY place the marker path + shape are defined — both the CLI
- * (writer) and the SDK loop (reader) import it, so the two halves can never
- * drift. Like the rest of the state-dir, the layout is a wire protocol between
- * the runner and the next subprocess invocation (ADR-0006). The marker is
- * consumed BEFORE the snapshot, so it never needs to survive a pause.
- *
- * Scoped by `agentId` so concurrent `agent()` calls sharing one sandbox each
- * see only their own request. When the id is unavailable (older agent env that
- * doesn't inject `AGENT_COMPOSE_AGENT_ID`) both sides fall back to a single
- * unscoped slot — correct for the common single-agent step, and the only case
- * where an unscoped marker can be ambiguous (two anonymous agents) is one the
- * old block-poll CLI couldn't handle either.
- */
-
-import { existsSync, mkdirSync, readFileSync, renameSync, rmSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import { randomBytes } from "node:crypto";
-
-import { getStateDir } from "../pause/state-dir.js";
-
-/** A choice offered to the human. A bare string is shorthand for
- *  `{ label, value }` with both equal — exactly what the dashboard's
- *  `readOptions` accepts. */
-export type PauseOption = string | { label: string; value: string };
-
-/** What `agentc pause` records for the loop to turn into a `ctx.pause`. */
-export interface LocalPauseRequest {
-  /** The question shown to the human (becomes the pause `reason`). */
-  reason:   string;
-  /** Optional offered choices, rendered as buttons in the dashboard. */
-  options?: PauseOption[];
-}
-
-const UNSCOPED = "_unscoped";
-
-const requestsDir = () => join(getStateDir(), "pause-requests");
-const markerPath  = (agentId: string | undefined | null) =>
-  join(requestsDir(), `${slug(agentId) || UNSCOPED}.json`);
-
-/** Keep the agentId filename-safe. Agent ids are `step<idx>-agent-<n>` shaped,
- *  but defend against anything exotic so the marker can never escape the dir. */
-function slug(agentId: string | undefined | null): string {
-  return (agentId ?? "").replace(/[^a-zA-Z0-9_.-]/g, "_");
-}
-
-/** Write the marker atomically (tmp → rename) so the loop never reads a
- *  partial file mid-write. Called by `agentc pause`. */
-export function writeLocalPauseRequest(agentId: string | undefined | null, req: LocalPauseRequest): void {
-  mkdirSync(requestsDir(), { recursive: true });
-  const final = markerPath(agentId);
-  const tmp   = `${final}.${randomBytes(6).toString("hex")}.tmp`;
-  writeFileSync(tmp, JSON.stringify(req), "utf8");
-  renameSync(tmp, final);
-}
-
-/** Take-once read: returns + deletes this agent's pending pause request, else
- *  null. Checks the agent-scoped slot first, then the unscoped fallback. The
- *  loop calls this once per turn; a malformed marker is dropped (deleted +
- *  null) rather than wedging the loop. */
-export function consumeLocalPauseRequest(agentId: string | undefined | null): LocalPauseRequest | null {
-  const paths = [...new Set([markerPath(agentId), markerPath(null)])]; // dedupe when agentId is absent
-  for (const path of paths) {
-    if (!existsSync(path)) continue;
-    try {
-      const raw = JSON.parse(readFileSync(path, "utf8")) as unknown;
-      rmSync(path, { force: true });
-      if (raw && typeof raw === "object" && typeof (raw as LocalPauseRequest).reason === "string" && (raw as LocalPauseRequest).reason.trim().length > 0) {
-        const r = raw as LocalPauseRequest;
-        return { reason: r.reason.trim(), ...(Array.isArray(r.options) && r.options.length > 0 ? { options: r.options } : {}) };
-      }
-    } catch {
-      rmSync(path, { force: true }); // unreadable / partial → drop it
-    }
-  }
-  return null;
-}