@deksden-com/dd-flow-cli 0.9.0-beta.9 → 0.9.0-beta.91

package/dist/harness-runtime/lib/delegation-instructions.d.mts

@@ -0,0 +1,10 @@
+export type DelegationWork = { work_id: string; start_command: string; launch_policy?: string };
+export type DelegationInput = { harness: string; stage: string; capacity: number; works?: DelegationWork[] };
+export function delegationAdapter(harness: string): string | null;
+export function delegationContract(harness: string): string;
+export function workerDelegationTask(input: { workId: string; startCommand: string; launchPolicy?: string }): string;
+export function renderDelegationInstructions(input: DelegationInput): string;
+export function renderCapacityInstructions(input: { harness: string; maximum: number }): string;
+export function renderWaitInstructions(input: { harness: string; stage: string }): string;
+export function renderAdapterPolicy(input: { harness: string }): string;
+export function renderRecoveryProbeInstructions(input: { harness: string; command: string }): string;

package/dist/harness-runtime/lib/delegation-instructions.mjs

@@ -0,0 +1,130 @@
+const ADAPTERS = {
+  codex: {
+    name: "Codex",
+    launch: "This Session uses the Codex multi-agent v2 API in the `collaboration` namespace. In code mode call `tools.collaboration__spawn_agent` once per assignment with exactly `task_name`, `message`, and `fork_turns: \"none\"` for fresh context. Keep the current model and reasoning by omitting overrides. Do not create a root Session or run the child's command yourself.",
+    wait: "The spawn result is a handle, not completion. In code mode use `tools.collaboration__wait_agent` and native completion messages until every launched child is terminal. A mailbox update is not necessarily final. A wait timeout leaves the same child running: continue waiting without relaunch or interruption. Retain terminal results and product Work receipts. Codex v2 owns idle-agent residency; do not interrupt a completed child as cleanup.",
+    partial: "If a spawn is refused, stop issuing new spawns and retain every successful child handle. Wait for every already-launched child to become terminal; silence is not completion. Do not interrupt a running child to free capacity, and do not interrupt completed children as cleanup. Report completed/launched assignments, assignments not launched and the exact refusal. Return after this partial wave; the controller will issue remaining Work when capacity is available. Do not replay an ambiguous spawn."
+  },
+  zcode: {
+    name: "ZCode",
+    launch: "Use the native `Agent` tool for every assignment in this current session. Use `TaskOutput` (or the tool's native wait operation) to await the returned child; never call a shell, create a new root session, or ask a child to create descendants.",
+    wait: "Wait on the returned native Agent task handles with `TaskOutput` (or the native wait operation) until all direct children settle. A missing or quiet result is not permission to launch a replacement.",
+  },
+  droid: {
+    name: "Droid",
+    launch: "Use the native `Task` tool for every assignment with `subagent_type: \"dd-flow-worker\"` and no complexity override. Do not use Missions, a root Session, or a shell command as a delegation substitute.",
+    wait: "Wait for each native Task result. Do not interrupt or replace a quiet Task; the coordinator remains responsible for the Stage.",
+  },
+  grok: {
+    name: "Grok",
+    launch: "Use the native `spawn_subagent` tool for each assignment in this current session. Keep the returned child handles and wait with `get_command_or_subagent_output`; do not invoke external roots or shell-based substitutes.",
+    wait: "Use `get_command_or_subagent_output` for every returned child handle before continuing. Never infer completion from silence or replace an unavailable child.",
+  },
+  opencode: {
+    name: "OpenCode",
+    launch: "Use the model-facing native `task` tool with description, prompt, and an available subagent_type that can execute the assignment. Put the assignment message in prompt. Omit task_id to create a fresh child. `subtask` is an API message part, not the model tool name. Use foreground completion, or the exposed background completion mechanism when available. Do not create a new root Session or use an external runner.",
+    wait: "Await each native task result or its background completion notification. Retain its returned task handle; never synthesize a Session ID or relaunch a quiet child."
+  },
+  antigravity: {
+    name: "Antigravity",
+    launch: "Use the harness's native subagent action in this current conversation for each assignment. Do not create a root Session, use an external runner, or ask a child to create descendants; the adapter will observe the native `subagent_info.subagents` records.",
+    wait: "Wait for the native child result and let the adapter observe its conversation ID and terminal status. Never invent or synthesize IDs, infer completion from text, or replace a quiet child.",
+  }
+};
+
+export function delegationAdapter(harness) {
+  const aliases = { codex: "codex", "codex-desktop": "codex", zcode: "zcode", "zcode-acp": "zcode", droid: "droid", "droid-cli": "droid", grok: "grok", "grok-acp": "grok", opencode: "opencode", "opencode-server": "opencode", agy: "antigravity", antigravity: "antigravity", "antigravity-cli": "antigravity" };
+  return Object.hasOwn(aliases, harness) ? aliases[harness] : null;
+}
+
+function requireAdapter(harness) {
+  const key = delegationAdapter(harness);
+  if (!key) throw Object.assign(new Error(`Unsupported delegation adapter: ${harness}`), { code: "delegation_contract_unsupported" });
+  return ADAPTERS[key];
+}
+
+export function delegationContract(harness) {
+  const key = delegationAdapter(harness);
+  requireAdapter(harness);
+  return `dd-flow/delegation/${key}@${key === "codex" ? 2 : 1}`;
+}
+
+export const lifecycleRetryInstruction = "If the CLI rejects the call with effect=no_effect, recoverable=true and retry_command, execute that exact retry_command in this same Session after correcting the reported input. Never invent IDs. If the native shell tool rejects the launch before the CLI starts, correct only that tool-call error and repeat the same exact command; do not change its arguments. A hook/transport failure is an infrastructure error: stop and report it without retrying. For any other failure, stop and report the exact error; do not run status as a substitute, call a hook handler, retry blindly, or finish the Work.";
+
+export function workerDelegationTask({ workId, startCommand, launchPolicy }) {
+  if (!workId || !startCommand) throw new Error("ready fan-out Work lacks its exact start command");
+  const context = launchPolicy === "fresh_agent_required"
+    ? "This Work requires empty, non-inherited conversation context; the coordinator must select the native fresh-child mode before launching it."
+    : "Use inherited context only through the harness's declared native mechanism.";
+  return [
+    `Complete one already-declared Work: ${workId}.`,
+    context,
+    "Your first technical action must be this exact standalone lifecycle command:",
+    startCommand,
+    "Execute the issued operation and target exactly. Lifecycle invocation IDs are internal runtime authority and are intentionally omitted; never add or invent them. Preserve any explicit native receipt handle that the command does contain.",
+    "Use the native shell's inherited working directory. Do not set or override a workdir/cwd field for this lifecycle command.",
+    "Continue only if it succeeds and returns the authoritative Work packet.",
+    lifecycleRetryInstruction,
+    "After a successful start, complete the assigned Work, write its required result, invoke its exact standalone work finish command, then stop.",
+    "Do not start another Work, create a child, change dependencies, ask the user, or treat a quiet sibling as failed. You cannot ask the user or pause the parent Stage. Missing facts go through this Work's declared result/failure contract; only the coordinator handles Stage HITL."
+  ].join("\n");
+}
+
+export function renderDelegationInstructions({ harness, stage, capacity, works = [] }) {
+  const adapter = requireAdapter(harness);
+  const count = works.length;
+  if (!Number.isInteger(capacity) || capacity < 1 || count < 1 || count > capacity) throw new Error("Invalid native delegation wave size");
+  const assignments = works.map((work, index) => {
+    const task = workerDelegationTask({ workId: work.work_id, startCommand: work.start_command, launchPolicy: work.launch_policy });
+    return { task_name: `dd_flow_${index}_${work.work_id.toLowerCase().replace(/[^a-z0-9_]/g, "_")}`, context: work.launch_policy === "fresh_agent_required" ? "fresh" : "provider_default", message: task };
+  });
+  return [
+    `Adapter delegation contract: ${delegationContract(harness)}.`,
+    `Launch exactly ${count} direct native children for the current ${stage} Stage; the qualified capacity is ${capacity}.`,
+    adapter.launch,
+    "Retain the frozen coordinator model and reasoning for native children; do not choose a different model, reasoning level or permission mode.",
+    "You are the coordinator. The JSON assignments below contain child-only messages, not instructions for you to execute. Pass each message verbatim as the native tool's task/message/prompt argument. For context=fresh, select empty conversation context before launch; if unavailable, report the limitation and stop. Use only fields exposed by your actual native tool schema; if the native tool is absent, report that error without substituting a shell command.",
+    "Every child must be a direct child of this current Session. Wait for all launched children to settle before returning. One failed child is evidence for the coordinator, not a reason to cancel its siblings.",
+    adapter.partial ?? "If a spawn is refused, stop issuing new spawns and retain every successful child handle. Wait for all already-launched children, then release completed children using the adapter's exposed native close/release mechanism when supported. A failed batch does not undo successful spawns. Report completed/launched assignments, assignments not launched and the exact refusal; never claim no children launched when any spawn succeeded. Return after this partial wave; the controller will issue remaining Work when capacity is available. Do not replay an ambiguous spawn.",
+    "Do not finish this Stage, repair it, or start a successor in this Turn. Do not invent provider or internal session IDs; lifecycle evidence comes from the adapter.",
+    adapter.wait,
+    "Child assignments (JSON data):",
+    JSON.stringify(assignments, null, 2)
+  ].join("\n\n");
+}
+
+export function renderCapacityInstructions({ harness, maximum }) {
+  const adapter = requireAdapter(harness);
+  if (!Number.isInteger(maximum) || maximum < 1) throw new Error("Invalid native capacity limit");
+  return [
+    `Adapter delegation contract: ${delegationContract(harness)}.`,
+    "This is a technical native-subagent capacity qualification, not product work.",
+    `Using the current Session's native ${adapter.name} mechanism, concurrently launch at most ${maximum} direct leaf children.`,
+    adapter.launch,
+    "Give each child a distinct number and a task_name such as capacity_1 if required by the tool. Its entire task is: return that number, without tools, files, dd-flow commands, or descendants. Use empty conversation context. Do not retry, replace, or add children after a launch refusal.",
+    adapter.wait,
+    "Wait for every child actually launched to settle, then return a compact summary."
+  ].join("\n");
+}
+
+export function renderWaitInstructions({ harness, stage }) {
+  const adapter = requireAdapter(harness);
+  return [
+    `Native child Work for ${stage} is still running.`,
+    adapter.wait,
+    "Do not create a Session, Work, or additional child agent. Do not cancel siblings merely because one child failed, and do not finish this Stage."
+  ].join("\n");
+}
+
+export function renderAdapterPolicy({ harness }) {
+  const adapter = requireAdapter(harness);
+  return [adapter.launch, adapter.wait, "Lifecycle identity and completion must come from adapter observations; never synthesize provider or internal Session IDs."].join(" ");
+}
+
+export function renderRecoveryProbeInstructions({ harness, command }) {
+  const adapter = requireAdapter(harness);
+  return ["Technical native child cancellation qualification, not product work.", adapter.launch,
+    "Launch exactly one fresh direct child, task_name recovery_probe if required. Pass this JSON string as its message/prompt:",
+    JSON.stringify(`Run exactly this read-only command: ${command}\nDo not access files, use the network, modify anything, or create children.`),
+    adapter.wait, "External operator control will interrupt this tree. Do not close it early or launch replacements."].join("\n");
+}

package/dist/harness-runtime/lib/dispatch-fence.mjs

@@ -0,0 +1,14 @@
+const generations = new WeakMap();
+
+// A cancellation invalidates preparation already in flight, while a later
+// explicit prompt captures a fresh generation and can resume normally.
+export function captureDispatchGuard(owner) {
+  const generation = generations.get(owner) ?? 0;
+  return () => {
+    if (generation !== (generations.get(owner) ?? 0)) throw Object.assign(new Error("Prompt cancelled before native dispatch"), { code: "operation_cancelled" });
+  };
+}
+
+export function cancelPendingDispatch(owner) {
+  generations.set(owner, (generations.get(owner) ?? 0) + 1);
+}

package/dist/harness-runtime/lib/driver-recovery.mjs

@@ -0,0 +1,129 @@
+import { processSnapshot } from "./process-snapshot.mjs";
+import { readdir, readFile, mkdir, writeFile } from "node:fs/promises";
+import path from "node:path";
+import { inspectDaemonOperation } from "./daemon-operations.mjs";
+import { writeJsonAtomic } from "./runner-events.mjs";
+import { reportedError } from "./operation-errors.mjs";
+import { ObservationClock } from "./observation-clock.mjs";
+import { runtimeProcess } from "./managed-daemon.mjs";
+
+/** Explicit same-Session restart retains native stores and the old daemon
+ * receipt; it does not turn a terminal directory into a new execution. */
+export async function authorizeRetainedDaemonResume(root, previous, sessionId, config) {
+  const sessions = [...(previous.sessions ?? []), ...(previous.root_session_id ? [previous.root_session_id] : [])];
+  if (previous.shutdown_state !== "clean" || previous.active_tree === true || typeof sessionId !== "string" || !sessions.some(session => (typeof session === "string" ? session : session.provider_session_id) === sessionId)) throw Object.assign(new Error("terminal daemon restart requires a retained, settled Session identity"), { code: "daemon_state_terminal" });
+  for (const key of ["cwd", "projectRoot", "ddFlowHome", "ddFlowBin", "provider", "model", "reasoning", "mode", "variant", "agent", "bin", "noFlow"]) {
+    if ((previous.config?.[key] ?? null) !== (config[key] ?? null)) throw Object.assign(new Error(`retained daemon configuration changed: ${key}`), { code: "daemon_config_mismatch" });
+  }
+  await assertDaemonReplaceable(root);
+  await reconcileDriverReplies(root);
+  if (!/^[a-zA-Z0-9-]+$/.test(previous.daemon_id ?? "")) throw Object.assign(new Error("original daemon identity is invalid"), { code: "daemon_state_terminal" });
+  const directory = path.join(root, "daemon-history"); await mkdir(directory, { recursive: true, mode: 0o700 });
+  const file = path.join(directory, `${previous.daemon_id}.json`), bytes = `${JSON.stringify(previous, null, 2)}\n`;
+  try { await writeFile(file, bytes, { flag: "wx", mode: 0o600 }); }
+  catch (error) { if (error.code !== "EEXIST" || await readFile(file, "utf8") !== bytes) throw error; }
+}
+
+/** A dead bridge alone does not prove its provider tree was stopped. Require
+ * the adapter's durable clean-shutdown receipt as well, for every bridge. */
+export async function recoverySettlement(driversRoot) {
+  const daemons = [];
+  const registries = new Map();
+  async function visit(directory) {
+    const entries = await readdir(directory, { withFileTypes: true });
+    const daemonRoot = entries.some(entry => entry.name === "daemon.json" && entry.isFile());
+    for (const entry of entries) {
+      const file = path.join(directory, entry.name);
+      if (entry.isSymbolicLink()) throw Object.assign(new Error("driver evidence contains a symbolic link"), { code: "recovery_settlement_unconfirmed" });
+      // Native stores contain package/auth symlinks, not adapter daemon
+      // receipts. Do not traverse their contents as a daemon inventory.
+      if (entry.isDirectory()) {
+        if (daemonRoot && ["gemini", "grok-home", "home", "xdg-data", "xdg-cache", "xdg-config"].includes(entry.name)) continue;
+        await visit(file);
+      }
+      else if (entry.name === "daemon.json") {
+        const state = JSON.parse(await readFile(file, "utf8"));
+        if (state.shutdown_state !== "clean" || state.active_tree === true) throw Object.assign(new Error("provider tree has no clean shutdown receipt"), { code: "recovery_settlement_unconfirmed" });
+        await assertDaemonReplaceable(directory);
+        await reconcileDriverReplies(directory);
+        if (state.config?.resourceHome && state.resource_process?.id) {
+          const key = state.config.resourceHome;
+          if (!registries.has(key)) registries.set(key, await runtimeProcess(state.config, "status"));
+          const processes = registries.get(key)?.processes;
+          if (!Array.isArray(processes) || !processes.some(record => record.id === state.resource_process.id)) throw Object.assign(new Error("managed daemon registration is missing"), { code: "recovery_settlement_unconfirmed" });
+          const attemptRoot = path.dirname(path.resolve(driversRoot));
+          for (const record of processes.filter(record => record.id === state.resource_process.id || [record.stdout_path, record.stderr_path].some(file => typeof file === "string" && path.resolve(file).startsWith(`${attemptRoot}${path.sep}`)))) {
+            if (!["stopped", "failed"].includes(record.state)) throw Object.assign(new Error("a retained managed process has not settled"), { code: "recovery_settlement_unconfirmed", details: { process_id: record.id, state: record.state } });
+            if (Number.isInteger(record.pid) && record.pid > 0) {
+              try { process.kill(record.pid, 0); }
+              catch (error) { if (error.code === "ESRCH") continue; }
+              if (typeof record.pid_started_at === "string" && record.pid_started_at.trim()) {
+                const observed = (await processSnapshot()).find(item => item.pid === record.pid);
+                if (!observed || record.pid_started_at.trim().replace(/\s+/g, " ") !== observed.started) continue;
+              }
+              throw Object.assign(new Error("a retained managed process is still live or unconfirmed"), { code: "recovery_settlement_unconfirmed", details: { process_id: record.id } });
+            }
+          }
+        }
+        daemons.push({ state_file: file, daemon_id: state.daemon_id, shutdown_state: state.shutdown_state, pid: state.pid });
+      }
+    }
+  }
+  await visit(driversRoot);
+  if (!daemons.length) throw Object.assign(new Error("no provider shutdown evidence found"), { code: "recovery_settlement_unconfirmed" });
+  return { settled: true, reason: "observed_clean_adapter_shutdown", observed_at: new Date().toISOString(), daemons };
+}
+
+export async function assertDaemonReplaceable(root) {
+  let state;
+  try { state = JSON.parse(await readFile(path.join(root, "daemon.json"), "utf8")); }
+  catch (error) { if (error.code === "ENOENT") return; throw error; }
+  if (!Number.isInteger(state.pid) || state.pid <= 0) throw Object.assign(new Error("original daemon ownership is unknown"), { code: "operation_observation_lost", details: { state_dir: root } });
+  const pids = new Set([state.pid, state.provider_pid, ...(state.owned_processes ?? []).map(process => process.pid)].filter(pid => Number.isInteger(pid) && pid > 0));
+  for (const pid of pids) {
+    try { process.kill(pid, 0); }
+    catch (error) { if (error.code === "ESRCH") continue; }
+    const retained = state.owned_processes?.filter(record => record.pid === pid && typeof record.started === "string" && record.started.trim());
+    if (retained?.length) {
+      const observed = (await processSnapshot()).find(record => record.pid === pid);
+      if (!observed || retained.every(record => record.started !== observed.started)) continue;
+    }
+    throw Object.assign(new Error("original daemon/provider is live or its ownership is unconfirmed; do not create a competing bridge"), { code: "operation_observation_lost", details: { state_dir: root, pid } });
+  }
+}
+
+/** Observe the original operation; never re-send its parameters. */
+export async function recoverDriverReply(root, id, { timeoutMs = 30_000, pollMs = 250 } = {}) {
+  const clock = new ObservationClock({ timeoutMs });
+  do {
+    let receipt;
+    try { receipt = await inspectDaemonOperation(root, id); }
+    catch (error) { if (error.code !== "operation_not_found") throw error; }
+    if (receipt?.state === "completed") return receipt.result;
+    if (receipt?.state === "failed") throw reportedError(receipt.error, "provider operation failed");
+    if (clock.sample()) break;
+    await new Promise(resolve => setTimeout(resolve, pollMs));
+  } while (true);
+  throw Object.assign(new Error("original provider operation has no confirmed reply; reconcile it before starting another Turn"), {
+    code: "operation_observation_lost", details: { operation_id: id, state_dir: root }
+  });
+}
+
+/** Crash recovery barrier. Import late replies into the client ledger before
+ * any subsequent productive call, even when a new runner process is used. */
+export async function reconcileDriverReplies(root) {
+  const directory = path.join(root, "client-operations");
+  let entries;
+  try { entries = await readdir(directory); } catch (error) { if (error.code === "ENOENT") return; throw error; }
+  for (const entry of entries.filter(name => name.endsWith(".json"))) {
+    const file = path.join(directory, entry), record = JSON.parse(await readFile(file, "utf8"));
+    if (["completed", "failed", "interrupted"].includes(record.state)) continue;
+    let receipt;
+    try { receipt = await inspectDaemonOperation(root, record.operation_id); }
+    catch (error) { if (error.code !== "operation_not_found") throw error; }
+    if (!["completed", "failed", "interrupted"].includes(receipt?.state)) throw Object.assign(new Error("a previous driver request has an unknown outcome; no new productive request was sent"), {
+      code: "operation_observation_lost", details: { operation_id: record.operation_id, state_dir: root }
+    });
+    await writeJsonAtomic(file, { ...record, state: receipt.state, result: receipt.result, error: receipt.error, reconciled_at: new Date().toISOString() });
+  }
+}

package/dist/harness-runtime/lib/droid-observation.mjs

@@ -0,0 +1,66 @@
+import { createHash } from "node:crypto";
+import { open, readFile, readdir, stat } from "node:fs/promises";
+import path from "node:path";
+
+/** Cache native filenames, then read only a bounded header and small settings.
+ * New child discovery may walk the directory once; heartbeats never replay it. */
+export class DroidMetadataReader {
+  constructor(factory) { this.factory = factory; this.files = new Map(); this.cache = new Map(); }
+  async locate(id) {
+    if (this.files.has(id)) return this.files.get(id);
+    const visit = async directory => {
+      for (const entry of await readdir(directory, { withFileTypes: true }).catch(error => { if (error.code === "ENOENT") return []; throw error; })) {
+        const file = path.join(directory, entry.name);
+        if (entry.isDirectory()) await visit(file);
+        else if (entry.isFile() && entry.name.endsWith(".jsonl")) this.files.set(entry.name.slice(0, -6), file);
+      }
+    };
+    await visit(path.join(this.factory, "sessions"));
+    return this.files.get(id) ?? null;
+  }
+  async read(id) {
+    const file = await this.locate(id); if (!file) return null;
+    const settingsFile = file.replace(/\.jsonl$/, ".settings.json");
+    const info = await stat(settingsFile).catch(error => { if (error.code === "ENOENT") return null; throw error; });
+    if (!info || info.size > 1_048_576) return null;
+    const identity = `${info.ino}:${info.mtimeMs}:${info.size}`;
+    if (this.cache.get(id)?.identity === identity) return this.cache.get(id).value;
+    const handle = await open(file, "r"); const buffer = Buffer.alloc(65_536);
+    let header;
+    try { const { bytesRead } = await handle.read(buffer, 0, buffer.length, 0); header = JSON.parse(buffer.subarray(0, bytesRead).toString("utf8").split("\n")[0]); }
+    finally { await handle.close(); }
+    if (header.type !== "session_start" || header.id !== id) throw Object.assign(new Error("native metadata Session identity differs"), { code: "session_identity_mismatch" });
+    let bytes, settings;
+    try { bytes = await readFile(settingsFile, "utf8"); settings = JSON.parse(bytes); }
+    catch (error) { if (error instanceof SyntaxError || error.code === "ENOENT") return null; throw error; }
+    const value = { header, settings, transcript_path: file, source: { path: settingsFile, sha256: createHash("sha256").update(bytes).digest("hex") } };
+    this.cache.set(id, { identity, value }); return value;
+  }
+}
+
+export function parseDroidRoutingLine(line) {
+  if (!line.includes("[Model-Router] Model transition | Context: ")) return null;
+  let context; try { context = JSON.parse(line.split(" | Context: ").slice(1).join(" | Context: ")); } catch { return null; }
+  if (context.slot !== "main" || typeof context.sessionId !== "string" || typeof context.modelId !== "string") return null;
+  return { session_id: context.sessionId, model: context.modelId, previous_model: context.previousModelId ?? null,
+    reason: typeof context.reason === "string" ? context.reason : null, native_timestamp: /^\[([^\]]+)\]/.exec(line)?.[1] ?? null,
+    event_sha256: createHash("sha256").update(line).digest("hex") };
+}
+
+/** Keep the cursor in daemon state. Bound every read, tolerate incomplete lines
+ * and restart from zero after truncation/rotation of this execution's own log. */
+export async function readDroidRoutingLog(file, cursor = {}) {
+  let handle;
+  try { handle = await open(file, "r"); } catch (error) { if (error.code === "ENOENT") return { cursor, events: [] }; throw error; }
+  try {
+    const info = await handle.stat();
+    const offset = info.ino === cursor.ino && info.size >= (cursor.offset ?? 0) ? cursor.offset ?? 0 : 0;
+    const buffer = Buffer.alloc(Math.min(1_048_576, Math.max(0, info.size - offset)));
+    const { bytesRead } = await handle.read(buffer, 0, buffer.length, offset);
+    const end = buffer.subarray(0, bytesRead).lastIndexOf(10) + 1;
+    const lines = buffer.subarray(0, end).toString("utf8").split("\n");
+    // Persist offsets only, never a partial native line that may contain secrets.
+    return { cursor: { ino: info.ino, offset: offset + (end || (bytesRead === buffer.length && bytesRead === 1_048_576 ? bytesRead : 0)) }, events: lines.map(parseDroidRoutingLine).filter(Boolean) };
+
+  } finally { await handle.close(); }
+}

package/dist/harness-runtime/lib/managed-daemon.mjs

@@ -0,0 +1,260 @@
+import { lstat, readFile, readdir, mkdir } from "node:fs/promises";
+import { randomUUID } from "node:crypto";
+import path from "node:path";
+import { commandJson } from "./process-json.mjs";
+import { errorRecord } from "./operation-errors.mjs";
+import { writeJsonAtomic } from "./runner-events.mjs";
+
+const leases = new Map();
+
+/** Freeze resource ownership independently of whether Subject hooks are enabled. */
+export function bindRuntimeOwner(config, env = process.env) {
+  const raw = Object.hasOwn(config, "runtime_owner") ? config.runtime_owner === null ? undefined : JSON.stringify(config.runtime_owner) : env.DD_FLOW_RUNTIME_OWNER;
+  if (raw === undefined) return config.runtime_owner = null;
+  let owner;
+  try { if (raw.length > 16_384) throw new Error(); owner = JSON.parse(raw); }
+  catch { throw Object.assign(new Error("runtime ownership descriptor is invalid"), { code: "runtime_owner_invalid" }); }
+  const fields = ["schema_id", "owner_id", "role", "project_id", "run_id", "work_id", "operation_id", "state_dir", "dd_flow_home", "dd_flow_bin", "resource_home", "adapter_executable", "budget"];
+  const text = value => typeof value === "string" && value.trim().length > 0;
+  if (!owner || typeof owner !== "object" || Array.isArray(owner) || Object.keys(owner).some(key => !fields.includes(key))
+    || owner.schema_id !== "dd-flow/runtime-owner@1" || !text(owner.owner_id) || !text(owner.operation_id)
+    || !["subject", "worker", "merge", "judge", "probe"].includes(owner.role)
+    || ["state_dir", "dd_flow_home", "dd_flow_bin", "resource_home"].some(key => !text(owner[key]) || !path.isAbsolute(owner[key]))
+    || (owner.adapter_executable !== undefined && (!text(owner.adapter_executable) || !path.isAbsolute(owner.adapter_executable)))
+    || ["project_id", "run_id", "work_id"].some(key => owner[key] !== undefined && !text(owner[key]))
+    || (["subject", "worker", "merge"].includes(owner.role) && (!text(owner.project_id) || !text(owner.run_id)))
+    || (["worker", "merge"].includes(owner.role) && !text(owner.work_id))
+    || (["judge", "probe"].includes(owner.role) && owner.work_id !== undefined)) {
+    throw Object.assign(new Error("runtime ownership descriptor does not match its role and scope"), { code: "runtime_owner_invalid" });
+  }
+  if (owner.budget !== undefined) {
+    const budget = owner.budget, names = ["codex-desktop", "zcode-acp", "grok-acp", "antigravity-cli", "opencode-server", "droid-cli"];
+    if (!budget || typeof budget !== "object" || Array.isArray(budget) || Object.keys(budget).some(key => !["schema_id", "scope_id", "per_harness"].includes(key)) || budget.schema_id !== "dd-flow/runtime-budget@1" || !/^[A-Za-z0-9][A-Za-z0-9._:-]{0,127}$/.test(budget.scope_id ?? "") || !budget.per_harness || typeof budget.per_harness !== "object" || Array.isArray(budget.per_harness) || Object.entries(budget.per_harness).some(([name, limit]) => !names.includes(name) || !Number.isSafeInteger(limit) || limit < 1 || limit > 1024)) throw Object.assign(new Error("runtime owner budget is invalid"), { code: "runtime_budget_invalid" });
+  }
+  config.runtime_owner = Object.fromEntries(fields.filter(key => owner[key] !== undefined).map(key => [key, owner[key]]));
+  return config.runtime_owner;
+}
+
+/** Reject rebinding before provider setup, including the already-running path. */
+export async function prepareRuntimeOwner(config, stateDir, env = process.env) {
+  const owner = bindRuntimeOwner(config, env);
+  if (owner && path.resolve(owner.state_dir) !== path.resolve(stateDir)) {
+    throw Object.assign(new Error("runtime owner belongs to another state directory"), { code: "runtime_owner_mismatch" });
+  }
+  let previous;
+  try { previous = JSON.parse(await readFile(path.join(stateDir, "daemon.json"), "utf8")); }
+  catch (error) { if (error.code === "ENOENT") return owner; throw error; }
+  const retained = bindRuntimeOwner({ runtime_owner: previous?.config?.runtime_owner ?? null }, {});
+  if (JSON.stringify(retained) !== JSON.stringify(owner)) {
+    throw Object.assign(new Error("retained daemon ownership differs; use explicit recovery instead of rebinding its sessions"), { code: "runtime_owner_mismatch" });
+  }
+  return owner;
+}
+
+export async function assertDaemonOwnership(stateDir, dispatch = {}, assertCurrent = () => {}) {
+  for (const lease of leases.values()) if (lease.error) throw lease.error;
+  const state = JSON.parse(await readFile(path.join(stateDir, "daemon.json"), "utf8"));
+  const owner = bindRuntimeOwner({ runtime_owner: state.config?.runtime_owner ?? null }, {});
+  if (!owner) return;
+  const record = state.resource_process;
+  if (owner.state_dir !== stateDir || !record?.id || !record.lease_token) throw Object.assign(new Error("owned daemon has no confirmed resource binding"), { code: "process_ownership_unknown" });
+  for (;;) {
+    assertCurrent();
+    const result = await runtimeProcess(state.config, "admission", { id: record.id, "lease-token": record.lease_token, owner: owner.owner_id, "state-dir": stateDir, "daemon-id": state.daemon_id ?? state.config?.daemonId, ...dispatch });
+    assertCurrent();
+    if (result?.ok === true && result.admitted === true && result.process_id === record.id) return;
+    if (result?.ok !== true || result.reason !== "provider_capacity" || result.process_id !== record.id) throw Object.assign(new Error("productive runtime admission was not confirmed"), { code: "process_ownership_unknown" });
+    // The daemon, not its disposable socket client, owns this queue. Recheck
+    // the RUN fence and process lease on every admission attempt.
+    await new Promise(resolve => setTimeout(resolve, 250));
+  }
+}
+
+/** Snapshot terminal operations BEFORE observing native idle. A later Turn
+ * cannot be released by an older observation of the same Session. */
+export async function budgetObservationScope(stateDir, sessionId) {
+  const state = await budgetState(stateDir);
+  if (!state?.config?.runtime_owner?.budget || !sessionId) return null;
+  const daemonId = state.daemon_id ?? state.config?.daemonId, operationIds = [];
+  for (const name of await readdir(path.join(stateDir, "operations")).catch(error => { if (error.code === "ENOENT") return []; throw error; })) {
+    const directory = path.join(stateDir, "operations", name);
+    try {
+      const requested = JSON.parse(await readFile(path.join(directory, "requested.json"), "utf8"));
+      const result = JSON.parse(await readFile(path.join(directory, "result.json"), "utf8"));
+      if (requested.daemon_id === daemonId && requested.session_id === sessionId && ["completed", "failed", "interrupted"].includes(result.state) && ["session.prompt", "session.start", "session.resume"].includes(requested.operation)) operationIds.push(requested.operation_id);
+    } catch (error) { if (error.code !== "ENOENT") throw error; }
+  }
+  return { stateDir, state, daemonId, sessionId, operationIds };
+}
+
+export async function releaseBudgetObservation(scope, observation) {
+  if (!scope || !scope.operationIds.length || observation?.settled !== true) return;
+  const id = observation.provider_session_id ?? observation.session_id ?? observation.adapter_session_id;
+  if (id !== scope.sessionId) throw Object.assign(new Error("budget observation belongs to another Session"), { code: "session_identity_mismatch" });
+  const directory = path.join(scope.stateDir, "budget-observations");
+  await mkdir(directory, { recursive: true, mode: 0o700 });
+  const file = path.join(directory, `${randomUUID()}.json`);
+  await writeJsonAtomic(file, { daemon_id: scope.daemonId, session_id: scope.sessionId, settled: true, operation_ids: scope.operationIds, observed_at: new Date().toISOString() });
+  await releaseBudget(scope.state, { "observation-file": file });
+}
+
+export async function completeBudgetOperation(stateDir, request, result, observe) {
+  const state = await budgetState(stateDir);
+  if (!state?.config?.runtime_owner?.budget) return { state: "not_required" };
+  if (["session.create", "session.fork"].includes(request.operation)) {
+    await releaseBudget(state, { "operation-id": request.id });
+    return { state: "settled", operation_ids: [request.id], source: "creation_receipt" };
+  }
+  const sessionId = request.params?.sessionId;
+  // A productive daemon receipt is bound to this exact dispatch. Reuse its
+  // complete tree proof instead of immediately asking the provider again.
+  if (["session.prompt", "session.start", "session.resume"].includes(request.operation)
+    && result?.settled === true
+    && typeof sessionId === "string"
+    && (result.provider_session_id ?? result.session_id ?? result.adapter_session_id) === sessionId) {
+    const scope = { stateDir, state, daemonId: state.daemon_id ?? state.config?.daemonId, sessionId, operationIds: [request.id] };
+    await releaseBudgetObservation(scope, result);
+    return { state: "settled", operation_ids: [request.id], source: "operation_receipt" };
+  }
+  // A known live tree is not a reason to poll it again. Preserve capacity and
+  // let the next explicit reconciliation obtain fresh proof.
+  if (result?.settled === false) return { state: "pending", reason: "tree_unsettled" };
+  const scope = await budgetObservationScope(stateDir, request.params?.sessionId);
+  if (!scope?.operationIds.length) return { state: "not_required" };
+  // Reconciliation requires a fresh, explicit proof.  Treating an arbitrary
+  // productive reply as proof would silently release a live turn.
+  if (!observe) return { state: "pending", reason: "settlement_evidence_required" };
+  const observation = await observe();
+  if (observation?.settled !== true) return { state: "pending", reason: "tree_unsettled" };
+  await releaseBudgetObservation(scope, observation);
+  return { state: "settled", operation_ids: scope.operationIds, source: "reconciliation_observation" };
+}
+
+async function budgetState(stateDir) {
+  try { return JSON.parse(await readFile(path.join(stateDir, "daemon.json"), "utf8")); }
+  catch (error) { if (error.code === "ENOENT") return null; throw error; }
+}
+
+async function releaseBudget(state, options) {
+  const record = state.resource_process, owner = state.config.runtime_owner;
+  const receipt = await runtimeProcess(state.config, "release-turn", { id: record.id, "lease-token": record.lease_token, owner: owner.owner_id, "daemon-id": state.daemon_id ?? state.config?.daemonId, ...options });
+  if (receipt?.ok !== true) throw Object.assign(new Error("runtime budget release was not confirmed"), { code: "runtime_budget_release_unconfirmed" });
+}
+
+export async function runtimeProcess(config, action, options = {}) {
+  const owner = config.runtime_owner;
+  const resourceHome = owner?.resource_home ?? config.resourceHome;
+  const ddFlowHome = owner?.dd_flow_home ?? config.ddFlowHome;
+  const ddFlowBin = owner?.dd_flow_bin ?? config.ddFlowBin;
+  if (!resourceHome || !ddFlowHome || !ddFlowBin) return null;
+  const args = ["runtime", "process", action];
+  for (const [key, value] of Object.entries(options)) if (value !== null && value !== undefined) args.push(`--${key}`, String(value));
+  return await commandJson(ddFlowBin, args, {
+    cwd: config.cwd,
+    env: { ...(config.env ?? {}), DD_FLOW_HOME: ddFlowHome, DD_FLOW_RESOURCE_HOME: resourceHome }
+  });
+}
+
+export async function registerDaemonProcess(config, input) {
+  const owner = bindRuntimeOwner(config);
+  const harness = ({ "codex-daemon": "codex-desktop", "zcode-daemon": "zcode-acp", "grok-daemon": "grok-acp", "agy-daemon": "antigravity-cli", "opencode-daemon": "opencode-server", "droid-daemon": "droid-cli" })[input.kind];
+  const scope = owner ? { owner: owner.owner_id, "project-id": owner.project_id, run: owner.run_id, work: owner.work_id, role: owner.role, "state-dir": owner.state_dir, "adapter-executable": owner.adapter_executable, ...(owner.budget ? { "budget-json": JSON.stringify(owner.budget), harness } : {}), ...(owner.project_id && owner.run_id || owner.budget ? { "dispatch-barrier": "run-generation@1" } : {}), operation: `${owner.operation_id}:${input.operation ?? input.kind}` } : {};
+  const result = await runtimeProcess(config, "register", { ...input, ...scope, "owner-pid": process.pid, "lease-ms": 300_000 });
+  return result?.process ? { id: result.process.id, lease_token: result.process.lease_token, kind: input.kind } : null;
+}
+
+export async function confirmDaemonProcess(config, record, child) {
+  if (!record) return;
+  if (!child?.pid) {
+    await finishDaemonProcess(config, record, "failed", "daemon_spawn_missing_pid");
+    return;
+  }
+  await runtimeProcess(config, "confirm", { id: record.id, "lease-token": record.lease_token, pid: child.pid, "process-group-id": child.pid, "owner-pid": record.kind?.endsWith("-daemon") ? child.pid : process.pid, "lease-ms": 300_000 });
+}
+
+export async function heartbeatDaemonProcess(config, record) {
+  if (!record?.id || !record.lease_token) return;
+  let lease = leases.get(record.id);
+  if (!lease) {
+    lease = { pending: null, error: null, timer: null, lost: false };
+    leases.set(record.id, lease);
+    lease.timer = setInterval(() => { void heartbeatDaemonProcess(config, record).catch(() => {}); }, 60_000);
+    lease.timer.unref();
+  }
+  if (lease.lost) throw lease.error;
+  if (lease.pending) return await lease.pending;
+  lease.pending = (async () => {
+    try {
+      const result = await runtimeProcess(config, "heartbeat", { id: record.id, "lease-token": record.lease_token, "lease-ms": 300_000 });
+      if (result?.ok !== true) {
+        lease.lost = true; clearInterval(lease.timer);
+        throw Object.assign(new Error("managed process lease is no longer owned; new productive operations are blocked"), { code: "process_lease_lost", details: { process_id: record.id } });
+      }
+      lease.error = null;
+    } catch (cause) {
+      lease.error = cause.code === "process_lease_lost" ? cause : Object.assign(new Error("cannot confirm managed process ownership"), { code: "process_ownership_unknown", cause, details: { process_id: record.id } });
+      throw lease.error;
+    }
+  })();
+  try { await lease.pending; } finally { lease.pending = null; }
+}
+
+export async function finishDaemonProcess(config, process, state = "stopped", reason = "daemon_stopped") {
+  if (process?.id && process.lease_token) {
+    const lease = leases.get(process.id); if (lease) clearInterval(lease.timer);
+    const result = await runtimeProcess(config, "finish", { id: process.id, "lease-token": process.lease_token, state, reason });
+    if (result?.ok !== true) throw Object.assign(new Error("managed process finish was not accepted"), { code: "process_lease_lost", details: { process_id: process.id } });
+    leases.delete(process.id);
+  }
+}
+
+/** Preserve the startup error and retain the lease/resources if cleanup cannot
+ * prove termination. A failed cleanup is not a terminal managed process. */
+export async function cleanupFailedStart(config, record, child, original) {
+  try {
+    await stopProcessGroup(child);
+    await finishDaemonProcess(config, record, "failed", original.code ?? "startup_failed");
+  } catch (cleanup) {
+    original.details = { ...original.details, cleanup_failed: errorRecord(cleanup), process_id: record?.id ?? null, pid: child?.pid ?? null };
+  }
+}
+
+export async function stopProcessGroup(child, graceMs = 1_000) {
+  if (!child?.pid) return;
+  // Harness bridges are detached on Unix, so their own process group is the
+  // smallest reliable unit of cleanup. Windows has no equivalent negative-PID
+  // signal operation; there the direct child is the safe best effort.
+  const target = process.platform === "win32" ? child.pid : -child.pid;
+  const alive = () => { try { process.kill(target, 0); return true; } catch (error) { if (error.code === "ESRCH") return false; if (error.code === "EPERM") return true; throw error; } };
+  if (!alive()) return;
+  if ((child.exitCode !== undefined && child.exitCode !== null) || (child.signalCode !== undefined && child.signalCode !== null)) {
+    // The leader can exit a few ticks before its helpers. Observe their normal
+    // shutdown before failing ownership; never signal a leaderless group.
+    const deadline = Date.now() + Math.max(1_000, graceMs);
+    while (alive() && Date.now() < deadline) await new Promise(resolve => setTimeout(resolve, 25));
+    if (!alive()) return;
+    throw Object.assign(new Error("process group remains after its leader exited; ownership cannot be proven from this child handle"), { code: "process_group_ownership_unknown", details: { pid: child.pid } });
+  }
+  process.kill(target, "SIGTERM");
+  await new Promise((resolve) => setTimeout(resolve, graceMs));
+  if (!alive()) return;
+  if (child.exitCode != null || child.signalCode != null) return stopProcessGroup(child, graceMs);
+  process.kill(target, "SIGKILL");
+  // Grace controls escalation, not the OS acknowledgement budget. A caller
+  // testing immediate escalation must still allow the kernel to reap children.
+  const deadline = Date.now() + Math.max(1_000, graceMs);
+  while (alive() && Date.now() < deadline) await new Promise((resolve) => setTimeout(resolve, 25));
+  if (alive()) throw Object.assign(new Error(`process group ${child.pid} did not stop`), { code: "process_group_stop_incomplete" });
+}
+
+/** A daemon acknowledgement is not a completed stop until its endpoint closes. */
+export async function confirmDaemonStopped(request, socket, timeoutMs = 30_000) {
+  const result = await request();
+  const deadline = Date.now() + timeoutMs;
+  while (Date.now() < deadline) {
+    try { await lstat(socket); }
+    catch (error) { if (error.code === "ENOENT") return result; throw error; }
+    await new Promise((resolve) => setTimeout(resolve, 25));
+  }
+  throw Object.assign(new Error("daemon acknowledged stop but its socket remained live"), { code: "daemon_stop_incomplete", details: { socket } });
+}