@duckmind/dm-linux-x64 0.65.5 → 0.65.6

package/extensions/dm-subagents/src/extension/tool-activation.js

@@ -0,0 +1,157 @@
+import * as fs from "node:fs";
+import * as path from "node:path";
+import { Type } from "typebox";
+import { DM_CODING_AGENT_PACKAGE_ROOT_ENV } from "../shared/utils.js";
+import { DM_CODING_AGENT_PACKAGE, resolveInstalledPiPackageRoot, resolveRunningPiPackageRoot } from "../runs/shared/dm-spawn.js";
+const LOADER_NAME = "subagents_enable";
+const SUBAGENT_NAME = "subagent";
+const MINIMUM_DYNAMIC_TOOLS_VERSION = [0, 86, 1];
+const UNSUPPORTED_HOST_MESSAGE = "Dynamic tool activation requires DM 0.86.1 or newer";
+let warnedUnsupportedHost = false;
+export function unsupportedDynamicToolsReason(dm, deps = {}) {
+  if (typeof dm.getAllTools !== "function" || typeof dm.getActiveTools !== "function" || typeof dm.setActiveTools !== "function")
+    return UNSUPPORTED_HOST_MESSAGE;
+  const probe = probeHostPiVersion(deps);
+  if ("reason" in probe)
+    return probe.reason;
+  return supportsMinimumVersion(probe.version) ? undefined : `${UNSUPPORTED_HOST_MESSAGE} (detected ${probe.version} in ${probe.root})`;
+}
+function probeHostPiVersion(deps) {
+  const runningRoot = resolveRunningPiPackageRoot(deps);
+  if (runningRoot && "reason" in runningRoot)
+    return runningRoot;
+  if (runningRoot)
+    return readHostPiManifest(runningRoot.root, runningRoot.source === "DM_SUBAGENTS_DM_CODING_AGENT_PACKAGE_ROOT", deps.readFileSync, deps.platform);
+  const installedRoot = resolveInstalledPiPackageRoot();
+  return { reason: installedRoot ? `Could not verify the running DM installation; ${installedRoot} is not confirmed to be the host that owns this session` : "Could not locate the running DM installation to verify dynamic tool support" };
+}
+function readHostPiManifest(root, fromOverride, readFileSync = (filePath, encoding) => fs.readFileSync(filePath, encoding), platform = process.platform) {
+  const manifestPath = (platform === "win32" ? path.win32 : path.posix).join(root, "package.json");
+  let source;
+  try {
+    source = readFileSync(manifestPath, "utf-8");
+  } catch (error) {
+    return { reason: `Could not read the DM package manifest at ${manifestPath}: ${errorMessage(error)}` };
+  }
+  let parsed;
+  try {
+    parsed = JSON.parse(source);
+  } catch (error) {
+    return { reason: `Invalid DM package manifest at ${manifestPath}: ${errorMessage(error)}` };
+  }
+  if (!isRecord(parsed) || parsed.name !== DM_CODING_AGENT_PACKAGE) {
+    return { reason: `${manifestPath} is not ${DM_CODING_AGENT_PACKAGE}${fromOverride ? ` (${DM_CODING_AGENT_PACKAGE_ROOT_ENV} override)` : ""}` };
+  }
+  return typeof parsed.version === "string" && parsed.version ? { version: parsed.version, root } : { reason: `The DM package manifest at ${manifestPath} has no version` };
+}
+function errorMessage(error) {
+  return error instanceof Error ? error.message : String(error);
+}
+function isRecord(value) {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+export function supportsMinimumVersion(version) {
+  const parsed = version.split(".").slice(0, 3).map(Number);
+  if (parsed.length !== 3 || parsed.some((part) => !Number.isInteger(part) || part < 0))
+    return false;
+  for (let index = 0;index < 3; index++) {
+    const part = parsed[index];
+    const minimum = MINIMUM_DYNAMIC_TOOLS_VERSION[index];
+    if (part !== minimum)
+      return part > minimum;
+  }
+  return true;
+}
+function hasNativeToolSelection(messages) {
+  return messages.some((message) => message.role === "system" && (Object.hasOwn(message, "toolsAdded") || Object.hasOwn(message, "toolsRemoved")));
+}
+function setSelection(dm, includeSubagent) {
+  const active = dm.getActiveTools();
+  const next = includeSubagent ? [...active, SUBAGENT_NAME] : active.filter((name) => name !== SUBAGENT_NAME);
+  if (!next.includes(LOADER_NAME))
+    next.push(LOADER_NAME);
+  dm.setActiveTools([...new Set(next)]);
+}
+function applyRecordedSelection(dm, ctx) {
+  const available = dm.getAllTools();
+  if (!Array.isArray(available) || !Array.isArray(dm.getActiveTools()))
+    return;
+  if (!available.some((tool) => tool.name === LOADER_NAME))
+    return;
+  const messages = ctx.sessionManager.buildSessionContext().messages;
+  if (hasNativeToolSelection(messages)) {
+    let selected = false;
+    for (const message of messages) {
+      if (message.role !== "system")
+        continue;
+      if (message.toolsRemoved?.some((tool) => tool.name === SUBAGENT_NAME))
+        selected = false;
+      if (message.toolsAdded?.some((tool) => tool.name === SUBAGENT_NAME))
+        selected = true;
+    }
+    setSelection(dm, selected);
+    return;
+  }
+  setSelection(dm, messages.length > 0 && dm.getActiveTools().includes(SUBAGENT_NAME));
+}
+export function registerSubagentToolActivation(dm, options) {
+  const unsupportedReason = unsupportedDynamicToolsReason(dm);
+  if (unsupportedReason) {
+    if (!warnedUnsupportedHost) {
+      warnedUnsupportedHost = true;
+      console.warn(`[dm-subagents] ${unsupportedReason}; keeping subagent eagerly available.`);
+    }
+    return;
+  }
+  const parameters = Type.Object({}, { additionalProperties: false });
+  const loader = {
+    name: LOADER_NAME,
+    label: "Enable Subagents",
+    description: "Enable dm-subagents delegation and management tools without launching work. Call when delegation is authorized by the current request or applicable user/project instructions, or when managing existing runs. Direct execution is the default; complexity alone never authorizes delegation. Full tools are available on the next model request.",
+    promptSnippet: "dm-subagents is installed. For authorized specialist, independent-review, or parallel work, call subagents_enable, then subagent. Authorization must come from the current request or applicable instructions; complexity alone is not authorization.",
+    parameters,
+    async execute() {
+      if (!dm.getAllTools().some((tool) => tool.name === SUBAGENT_NAME))
+        return {
+          isError: true,
+          content: [{ type: "text", text: "Cannot enable unavailable tools: subagent." }],
+          details: { unavailable: [SUBAGENT_NAME] }
+        };
+      try {
+        setSelection(dm, true);
+      } catch (error) {
+        return {
+          isError: true,
+          content: [{ type: "text", text: `Activation failed: ${error instanceof Error ? error.message : String(error)}` }],
+          details: { missing: [SUBAGENT_NAME] }
+        };
+      }
+      if (!dm.getActiveTools().includes(SUBAGENT_NAME))
+        return {
+          isError: true,
+          content: [{ type: "text", text: "Activation failed: subagent." }],
+          details: { missing: [SUBAGENT_NAME] }
+        };
+      const advertised = await options.advertisedPrompt();
+      return {
+        content: [{ type: "text", text: `Enabled: subagent. On the next model request, call subagent({action:"list",capabilities:true}) for current capabilities.${advertised ? `
+
+${advertised}` : ""}` }],
+        details: { enabled: [SUBAGENT_NAME] }
+      };
+    }
+  };
+  dm.registerTool(loader);
+  dm.on("session_start", (_event, ctx) => applyRecordedSelection(dm, ctx));
+  dm.on("session_tree", (_event, ctx) => applyRecordedSelection(dm, ctx));
+  dm.on("before_agent_start", (event) => {
+    const available = dm.getAllTools();
+    if (!Array.isArray(available) || !available.some((tool) => tool.name === LOADER_NAME))
+      return;
+    const selectedTools = event.systemPromptOptions.selectedTools ??= [...dm.getActiveTools()];
+    if (!selectedTools.includes(LOADER_NAME))
+      selectedTools.push(LOADER_NAME);
+    if (!dm.getActiveTools().includes(LOADER_NAME))
+      dm.setActiveTools([...dm.getActiveTools(), LOADER_NAME]);
+  });
+}

package/extensions/dm-subagents/src/extension/tool-description.js

@@ -3,87 +3,39 @@ import * as path from "node:path";
 import { getAgentDir, getProjectConfigDir } from "../shared/utils.js";
 const CUSTOM_TOOL_DESCRIPTION_FILE = "subagent-tool-description.md";
 const CUSTOM_TOOL_DESCRIPTION_MAX_BYTES = 50 * 1024;
-const EXTERNAL_CLI_RUNNER_GUIDANCE = "External CLI agents (codex-exec, codex-exec-writer, claude-code, claude-code-writer, cursor-agent, cursor-agent-writer) use their own runner contract and do not support native DM child options such as model override, structured output, acceptance/agent contract, tool budget, fast mode, fork context, skills, or native DM tools unless the runner explicitly implements them.";
-const WORKFLOW_RESUME_KEY_GUIDANCE = "Each workflow key identifies one result lane: use a new stable workflow key for every distinct retained resume pass; same-key calls are reused only when launch parameters are identical, and incompatible parameters are rejected.";
-const WORKFLOW_OUTPUT_BINDING_GUIDANCE = "For durable workflow child files, set output on runs.run/runs.all; task filename prose is not an output declaration, and return the child's outputReference, outputPathMapping, or artifactPaths instead of inventing a literal path.";
-const WORKFLOW_LANES_GUIDANCE = "For bounded parallel sequential chains, use runs.lanes([{key,stages:[{key,agent,task},{key,resume:'previous',task},...]}]); first stages run together, later stages sequence per lane, and the bounded board reports lane-local failures. Only an explicit structuredOutput.verdict === 'blocked' blocks a successful stage; reviewer prose is not parsed.";
-const WORKFLOW_SCRIPT_PORTABILITY_GUIDANCE = "workflowScript rejects nested async function, arrow, and method helpers; use top-level await, plain helper functions that return runs.run(...), or explicit Promise chains instead.";
-const WORKFLOW_HOST_GUIDANCE = "For one non-interactive operator-owned command, await runs.host(key,{kind:'command',command,timeoutMs,output?,role?,provider?}). runs.host has no per-step cwd: commands and relative output paths use the workflow cwd; set cwd on the outer subagent request instead (for example, {cwd:'/path/to/worktree',workflowScript:'...'}), or put a trusted directory change in the command (for example, 'cd /path/to/worktree && npm test'). v1 supports only command steps; output is bounded and command failure fails the workflow.";
-const AGENT_CAPABILITY_GUIDANCE = 'For capability selection, use { action: "list", capabilities: true } for compact prompt-free rows.';
-export const DEFAULT_SUBAGENT_TOOL_DESCRIPTION = `Delegate to configured subagents. For execution, omit action and use {agent, task?} for one child, workflowScript for inline orchestration, or workflowScriptPath to load a script from the request cwd. The script inputs are mutually exclusive. Use action:'validate' with either script input to check it without launching children. For multi-step or parallel work, make exactly one top-level subagent call with async:true; launch children only inside that workflow and do not make another top-level call for them. Use runs.run('key',{agent,task}) for one child, await runs.all([{key:'a',agent:'reviewer',task:'...'},{key:'b',agent:'reviewer',task:'...'}]) for ordinary parallel children, and read its ordered array result with indexes, destructuring, or .map(...), not by key property. ${WORKFLOW_SCRIPT_PORTABILITY_GUIDANCE} ${WORKFLOW_LANES_GUIDANCE} ${WORKFLOW_HOST_GUIDANCE} ${EXTERNAL_CLI_RUNNER_GUIDANCE} Use action only for management/control. Use guide or the dm-subagents skill for advanced workflow details.`;
-export const SUBAGENT_TOOL_PROMPT_SNIPPET = "Delegate to subagents; orchestrate in one workflowScript call.";
-export const SUBAGENT_TOOL_PROMPT_GUIDELINES = [
-  `Use subagent only when delegation is needed. Before executing, call { action: "list" } and run only executable, non-disabled agents. ${AGENT_CAPABILITY_GUIDANCE}`,
-  "Omit action for execution. Use { agent, task? } only for one child; use workflowScript for multi-step or parallel work.",
-  "workflowScript means exactly one top-level subagent tool call with async:true. Inside it, use runs.run/runs.all to launch children; do not make another top-level subagent call for those children.",
-  WORKFLOW_SCRIPT_PORTABILITY_GUIDANCE,
-  WORKFLOW_LANES_GUIDANCE,
-  WORKFLOW_HOST_GUIDANCE,
-  WORKFLOW_RESUME_KEY_GUIDANCE,
-  WORKFLOW_OUTPUT_BINDING_GUIDANCE,
-  "For ordinary parallel work, use await runs.all([{key,agent,task}, ...]); it resolves to an ordered array, not a key map, so use results[0], destructuring, or results.map(...), not results.<key>. Do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout and each must later be observed with direct await, Promise.race, or Promise.all.",
-  "Keep one writer per cwd/worktree unless writers run in isolated worktrees.",
-  `To pass an explicit model to a child, first call { action: "models" } and copy an exact provider/id (e.g. provider/model-id); bare ids resolve only when unique in the registry, and agent names (gpt-pro, advisor) are not model ids. Set per-run thinking with a suffix on the model string (e.g. provider/model-id:high; off/minimal/low/medium/high/xhigh/max); the suffix wins over the agent's thinking default. The thinking field only applies to action='watchdog.configure' and is ignored on dispatch.`,
-  EXTERNAL_CLI_RUNNER_GUIDANCE,
-  "Use guide or the dm-subagents skill for advanced scheduling, missions, steering, and retention."
-];
+const AGENT_SELECTION_GUIDANCE = 'First call {action:"list",capabilities:true}: executable, non-disabled agents only; external-cli requires runner.available === true. Passive PATH/PATHEXT/X_OK is not authentication/version/launch proof; preflight is authoritative.';
+const SUBAGENT_FAILURE_RECOVERY_GUIDANCE = "Workflow, child launch, prompt runtime, extension load or child tooling failure is a lane infrastructure blocker. Stop; report exact failure, run/status and repo/cwd/worktree/branch/ref; verify clean worktree or capture partial diff before same-protocol retry or asking the owner. Never silently switch to interactive_shell, dm -ne, Codex/Claude/Cursor CLI or foreground/external mode: governed-workflow fallback requires explicit owner approval, not DM core's generic dm -ne hint. Explicit foreground/CLI requests and work outside that protocol remain valid.";
 export const SUBAGENT_SAFETY_GUIDANCE = `SAFETY-CRITICAL SUBAGENT GUIDANCE:
-• Use { action: "list" } before execution and only run executable/non-disabled agents.
-• Keep execution and management separate: omit action for structured single-child or workflowScript execution; use action only for management/control.
-• Async/background runs are the normal default unless config sets asyncByDefault:false; set async:true explicitly when async behavior matters. Use async:false only when the parent must block until completion. Async mode still shows progress. Final reviews and gate checks stay async; needing a result is not a blocking reason. After an async launch, continue independent work only until its next dependency barrier; consume the result before work that depends on it. Ordinary async subagents notify this session natively, so return control and do not call bg_wait merely to get a completion wake. Do not sleep or poll status just to wait; use bg_wait only for provider, detached, or other background work without a native notification when this turn must receive its result.
-• ${WORKFLOW_RESUME_KEY_GUIDANCE}
-• ${WORKFLOW_OUTPUT_BINDING_GUIDANCE}
-• ${WORKFLOW_HOST_GUIDANCE}
-• Ordinary child subagents are not orchestrators. Only explicitly configured fanout children may use the child-safe subagent tool, still bounded by depth/session limits.
-• Oracle/advisor consultations should use supervisor dialogue for material unknowns when available; request one-shot only when desired.
-• Keep one writer for the same cwd/worktree. Use fresh-context read-only reviewers for independent review, then have the parent synthesize and apply fixes.
-• Async runs expose asyncId/asyncDir with status.json, events.jsonl, output logs, status via { action: "status", id }, and lifecycle diagnostics via { action: "debug.run", id }. Include output paths and residual risks when reporting results.`;
-export const FULL_SUBAGENT_TOOL_DESCRIPTION = `Run one child with { agent, task? }; use { workflowScript } for inline orchestration or { workflowScriptPath } to load it from the request cwd. The script inputs are mutually exclusive. Omit action for execution. Use action only for management/control actions.
-
-EXECUTION:
-• ${EXTERNAL_CLI_RUNNER_GUIDANCE}
-• Before executing, use { action: "list" } and run only executable/non-disabled configured agents.
-• When passing an explicit model to a child (on the call or a runs.run/runs.all item), first call { action: "models" } and copy an exact provider/id; bare ids resolve only when unique in the registry, and agent names (e.g. gpt-pro, advisor) are not model ids. Set per-run thinking with a suffix on the model string (e.g. provider/id:high; off/minimal/low/medium/high/xhigh/max); the suffix wins over the agent's thinking default. The thinking field only applies to action='watchdog.configure' and is ignored on dispatch.
-• SINGLE CHILD: { agent:"worker", task:"..." }. This structured form starts exactly one direct child. Fields such as model, context, cwd, worktree, output, budgets, acceptance, and async apply to that child. Do not combine agent/task with action, workflowScript, or workflowScriptPath.
-• WORKFLOW SCRIPT: { workflowScript: "return runs.run('main', {agent:'worker', task:'...'})" }. Use stable-key runs.run for one child and await runs.all([{key,agent,task}, ...]) for ordinary parallel children. runs.all resolves to an ordered array, not a key map, so use results[0], array destructuring, or results.map((result) => result.output), not results.<key>. Do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout and each must later be observed with direct await, Promise.race, or Promise.all. Ordinary JavaScript provides sequence, branching, filtering, retries, and aggregation. workflowScript is an ordinary JavaScript statement body, so use an explicit return for a useful result. Use top-level await, plain helper functions, or explicit Promise chains; nested async function, arrow, and method helpers are rejected. For task text with Markdown fences or shell blocks, build quoted lines instead of nesting raw template literals: \`const task=["Run:","\`\`\`bash","npm test","\`\`\`"].join("\\n")\`. Scripts normally start async unless config sets asyncByDefault:false; set async:true explicitly when async behavior matters. Pass async:false only when the parent must block until completion, never for final reviews or gates. Same-repo blocking workflows default to a live in-chat card; explicit live-card requires same-repository async:false, so async workflows should omit chatProgress or use auto/off. Workflow-level child controls default onto each runs.run launch, and explicit child fields override them. Use {action:"children.list"} to list recent retained workflow children with resumable/not-resumable reasons. Resume only rows reported resumable. For a simple follow-up or implementation challenge, use {action:"resume", id:"run-id", message:"..."}. Resume keeps the stored agent/model/tool contract. If no resumable child is listed, launch a same-role fallback challenge and label it as fallback. Inside workflowScript, continue one with runs.run(key, {resume:"run-id", task:"follow-up"}); workflow resumes wait for completed output, and loops must continue from each latest returned runId. Await runs.steer(key, message, {mode?, index?, ackTimeoutMs?}) to guide a prior keyed child without exposing its run id; receipts are queued, delivered, missed, or failed. Always await or return runs.steer. For repository mutation lanes, set worktree:true on the workflow or individual runs.run/runs.all item for managed isolation; each parallel child gets a separate worktree and handoff artifact. A workflow usageBudget is enforced once across the workflow. Available globals are runs.run, runs.all, runs.steer, runs.status, runs.ref/refs, emit, console, and standard JavaScript only. Workflows get async state.get(key) and state.set(key, JSONValue) through their automatic or explicit mission; mission:false workflows do not have a state global. Scripts cannot access filesystem, shell, arbitrary DM tools, or host globals.
-• ${WORKFLOW_LANES_GUIDANCE}
-• FILE SCRIPT: { workflowScriptPath:"workflows/review.js" }. Relative paths resolve against the request cwd. The host reads the file before the filesystem-free workflow sandbox starts. Do not combine this field with workflowScript.
-• Sequential example: { workflowScript: "const a = await runs.run('analyze', {agent:'agent-a', task:'Analyze the request'}); return (await runs.run('plan', {agent:'agent-b', task:'Plan from: '+a.output})).output" }
-• Parallel example: { workflowScript: "const [a,b] = await runs.all([{key:'correctness',agent:'agent-a',task:'Review correctness'},{key:'tests',agent:'agent-b',task:'Review tests'}]); return {correctness:a.output,tests:b.output}" }
-• Optional context is "fresh", "fork", or "profile". profile requires the selected agent's declared defaultContext and ignores config defaultSubagentContext. Explicit fresh/fork wins. When omitted, config defaultSubagentContext wins over agent defaultContext. Config forkContext can summarize transcript overflow with stable recovery refs before spawn without adding another public context value. timeoutMs/maxRuntimeMs apply to foreground and async workflows; foreground workflows default to 30 minutes and async workflows have no default timeout. Omit acceptance for reviewer/read-only calls; evidence levels end at verified, and acceptance.review.required requests independent writer review.
-• Durable mission attachment is automatic by default. Use missionId to attach an existing mission, mission:{...} to override auto-create, or mission:false for ephemeral work. A mission object needs exactly one non-empty title or summary; objective and labels are optional. goal may only be true and requires budget:{tokens}.
-
-MANAGEMENT / CONTROL (use action; omit execution fields):
-• validate checks workflowScript or workflowScriptPath syntax and statically decidable structure without launching children. list, get, models, guide, children.list, create, update, delete, eject, disable, enable, reset, status, debug.run, doctor, grant-spawn-budget, worktree.discard, worktree.cleanup (plan-only), lane.status, lane.recordMerge, lane.recordSupersession, refine/refine.show/refine.rollback, mission.create/list/show/update/resolve-decision/attach-run/close, inspector.open/status/close, project.open/status/close, and watchdog actions remain available. Use {action:"guide", topic:"overview"} for packaged current-version help; topics are overview, workflows, agents, missions, observability, tool-reference, configuration, models, watchdog, and extension-api.
-• status, interrupt, stop, resume, and steer manage live or persisted runs. Use status view:"fleet" for an overview or view:"transcript" with id and optional index to tail output.
-• Create durable project schedules with { action:"schedule.create", id?, name?, at:"+10m" | ISO, workflowScript:"return runs.run('main', {agent:'worker', task:'...'})" }, or use workflowScriptPath instead. Manage them with schedule.list/show/history/pause/resume/run/run-due/delete. This first slice supports fixed intervals; calendar schedules and schedule mission attachment are deferred.
+• Direct parent execution is the default. Invoke subagents only when delegation is authorized by the operator's current request or applicable user/project instructions; task size, complexity, risk, tool-call count, or recipe fit do not independently authorize delegation.
+• ${AGENT_SELECTION_GUIDANCE}
+• ${SUBAGENT_FAILURE_RECOVERY_GUIDANCE}
+• Omit action for execution. For an authorized delegated multi-step/parallel workflow: exactly one top-level subagent workflow call with async:true; children launch only inside it.
+• Async follows asyncByDefault (normally true); async:false only to block the parent, not for final reviews/gates. Consume results at dependency barriers. Native async completion wakes this session: return control, no sleep/poll or bg_wait merely for a wake. bg_wait is for provider/detached work without native notification needing a same-turn result.
+• Ordinary child subagents are not orchestrators; only configured fanout within depth/session limits. For an authorized delegated workflow, keep one writer per cwd/worktree and isolate concurrent writers. Use fresh-context read-only reviewers when independent review was requested, then parent synthesis/fixes. Oracle/advisor unknowns use supervisor dialogue; one-shot only when requested.
+• Bind durable output on runs.run/runs.all, not task filename prose; return actual outputReference/outputPathMapping/artifactPaths, evidence and residual risks.
+• children.list is workflow-only, not an exhaustive list of direct native children: resume only resumable rows. When an intended child's exact run id is known, inspect it with {action:"status",id}; if status identifies the candidate, attempt {action:"resume",id,message}. Resume authoritatively checks eligibility, may reject it, and otherwise detaches a follow-up/challenge with the stored agent/model/tool contract. Use a labeled same-role fallback only when no known candidate exists or resume rejects eligibility. Scripts await runs.run(newKey,{resume:runId,task}); continue from latest returned runId. Each distinct resume pass needs a new stable key; same-key reuse requires identical launch parameters.
+• Named resources own authority; raw workflowScript/workflowScriptPath cannot use runs.host. Granted commands/relative outputs use workflow cwd, never per-step cwd.
+• Inspect asyncId/asyncDir (status.json, events.jsonl, logs) with status/debug.run; control with interrupt/stop/resume/steer. Read {action:"guide",topic:"tool-reference"} for controls/evidence gates.`;
+const EXECUTION_GUIDANCE = `Delegate one child with {agent,task?}; otherwise choose exactly one of {workflowScript,args?}, {workflowScriptPath,args?} or {workflow,args}. agent/task exclude workflow inputs; task excludes action. agent may target management actions. action is management/control; validate accepts either script without launching. workflowScriptPath loads from request cwd before sandbox execution.
+Scripts: JavaScript statement bodies with explicit return, top-level await, plain helpers/Promise chains; nested async function/arrow/method helpers are rejected. Await runs.run('key',{agent,task}) before .output; await runs.all([{key,agent,task},...]) for an ordered array, not a key map. Observe every stored run promise with direct await, Promise.race or Promise.all. Await/return runs.steer(key,message,options?) for a prior key, never raw run ids; queued/delivered/missed/failed receipts are not compliance proof.
+Before advanced orchestration (runs.lanes, rolling fanout, mission state, handoffs), read {action:"guide",topic:"workflows"} or the dm-subagents skill. Raw-script sandboxes add deeply frozen args; all sandboxes provide runs, emit, console, JavaScript and enabled mission state, with no filesystem/shell/DM tools/host globals. External CLI agents support native options only when their runner declares them; read guide tool-reference before passing model, structured output, acceptance/agentContract, tool budget, fast, fork context or skills/tools.
+Model override: first call {action:"models"}; copy exact provider/id, not agent names. Thinking uses model suffix, not watchdog-only thinking.
+Named resources: {workflow:'review',args:{task:'...'}} or {workflow:'run-ci',args:{command:'npm test'}}. Raw scripts also accept bounded plain-data args; raw-script args persist as evidence, so never include secrets. worktree:true requires clean source; baseRef defaults to HEAD at allocation or a supported named ref, never full 40/64-character commit IDs or revision expressions.`;
+export const DEFAULT_SUBAGENT_TOOL_DESCRIPTION = `${EXECUTION_GUIDANCE}
 
 ${SUBAGENT_SAFETY_GUIDANCE}`;
-export const COMPACT_SUBAGENT_TOOL_DESCRIPTION = `Run one child with { agent, task? }; use { workflowScript } for inline orchestration or { workflowScriptPath } to load it from the request cwd. The script inputs are mutually exclusive. Omit action for execution. Use action only for management/control actions.
-
-EXECUTE:
-• ${EXTERNAL_CLI_RUNNER_GUIDANCE}
-• Call { action:"list" } first and use only executable/non-disabled agents.
-• Passing an explicit model? Call {action:"models"} first and copy an exact provider/id; bare ids resolve only when unique in the registry; agent names (e.g. gpt-pro, advisor) are not model ids. Per-run thinking is a suffix on the model string (provider/id:high; off/minimal/low/medium/high/xhigh/max), and the suffix wins over the agent's thinking default; the thinking field only applies to action='watchdog.configure' and is ignored on dispatch.
-• SINGLE {agent:"worker",task:"..."} starts exactly one direct child. Fields apply to that child. Do not combine agent/task with action, workflowScript, or workflowScriptPath.
-• SCRIPT {workflowScript:"return runs.run('main', {agent:'worker', task:'...'})"}. Use stable-key runs.run for one child and await runs.all([{key,agent,task}, ...]) for ordinary parallel work. runs.all resolves to an ordered array, not a key map; use results[0], destructuring, or results.map(...), not results.<key>. Do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout and each must later be observed with direct await, Promise.race, or Promise.all. Await runs.steer(key,message,options?) to guide a prior keyed child; it returns queued, delivered, missed, or failed and never accepts a raw run id. Always await or return steering calls. Use {action:"children.list"} for recent retained workflow children and resume only rows reported resumable. Use {action:"resume",id:"run-id",message:"..."} for a simple follow-up or challenge; resume keeps the stored agent/model/tool contract. If none is resumable, launch a same-role fallback challenge and label it as fallback. Inside workflowScript use runs.run(key,{resume:"run-id",task:"follow-up"}) when the script must wait for completion and continue from the latest returned runId. Workflows get async state.get/state.set through their automatic or explicit mission; mission:false does not. Scripts are ordinary JavaScript statement bodies; use explicit return for a useful result. Use top-level await, plain helper functions, or explicit Promise chains; nested async function, arrow, and method helpers are rejected. For task text with Markdown fences or shell blocks, build quoted lines instead of nesting raw template literals: \`const task=["Run:","\`\`\`bash","npm test","\`\`\`"].join("\\n")\`. Use JavaScript for sequence, branching, retries, and aggregation. For repository mutation lanes, use worktree:true on the workflow or runs.run/runs.all item for managed isolation. Scripts normally start async unless config sets asyncByDefault:false; set async:true explicitly when async behavior matters. async:false blocks the parent until completion and auto-enables a same-repo live chat card unless chatProgress is off; explicit live-card requires same-repository async:false, so async workflows should omit chatProgress or use auto/off.
-• ${WORKFLOW_LANES_GUIDANCE}
-• FILE SCRIPT {workflowScriptPath:"workflows/review.js"} loads the script on the host relative to the request cwd before sandbox execution. Do not combine it with workflowScript.
-• Example: {workflowScript:"const [a,b]=await runs.all([{key:'a',agent:'agent-a',task:'Implement A',worktree:true},{key:'b',agent:'agent-b',task:'Implement B',worktree:true}]); return [a.output,b.output]"}
-• context can be fresh, fork, or profile. profile requires the selected agent's declared defaultContext and ignores defaultSubagentContext. Explicit fresh/fork wins; omitted context follows defaultSubagentContext before agent defaultContext. Config forkContext can summarize transcript overflow with stable recovery refs before spawn without adding another public context value. timeoutMs/maxRuntimeMs apply to foreground and async workflows; foreground workflows default to 30 minutes and async workflows have no default timeout. Omit acceptance for reviewer/read-only calls.
-
-MANAGE / CONTROL:
-• Use action without execution fields for list/get/models/guide/authoring, refine/refine.show/refine.rollback, mission, watchdog, status, interrupt, stop, resume, steer, worktree.cleanup (mode:'plan' only), script-only scheduling, diagnostics, and other management actions. guide reads shipped current-version docs by topic.
-• A mission object needs exactly one non-empty title or summary; objective and labels are optional. goal may only be true and requires budget:{tokens}.
+export const SUBAGENT_TOOL_PROMPT_SNIPPET = "For operator-requested delegation, use subagents; compose multi-child work in one workflow call.";
+export const SUBAGENT_TOOL_PROMPT_GUIDELINES = [
+  "Do not invoke subagents unless the operator requested delegation directly or through applicable instructions."
+];
+export const COMPACT_SUBAGENT_TOOL_DESCRIPTION = DEFAULT_SUBAGENT_TOOL_DESCRIPTION;
+export const FULL_SUBAGENT_TOOL_DESCRIPTION = `${DEFAULT_SUBAGENT_TOOL_DESCRIPTION}
 
-ASYNC / SAFETY:
-• Omitted async follows asyncByDefault config; set async:true explicitly when async behavior matters. Continue independent work only until its next dependency barrier; consume the result before work that depends on it. Ordinary async subagents notify this session natively, so return control and do not call bg_wait merely to get a completion wake. Do not sleep or poll merely to wait; use bg_wait only for provider, detached, or other background work without a native notification when this turn must receive its result.
-• ${WORKFLOW_RESUME_KEY_GUIDANCE}
-• ${WORKFLOW_OUTPUT_BINDING_GUIDANCE}
-• ${WORKFLOW_HOST_GUIDANCE}
-• Ordinary children are not orchestrators. Keep one writer per cwd/worktree and use fresh read-only reviewers for independent checks.
-• Oracle/advisor consultations use available supervisor dialogue for material unknowns; request one-shot when desired.
-• Status and artifacts live under asyncId/asyncDir with status.json, events.jsonl, output logs, and {action:"status",id:"..."}.`;
+WORKFLOW DETAILS:
+• runs.lanes([{key,stages:[{key,agent,task},{key,resume:'previous',task}]}]) runs first stages together, later stages sequentially per lane. Failures stay lane-local; only explicit structuredOutput.verdict === 'blocked' blocks a successful stage, never reviewer prose.
+• Workflow child controls default onto runs.run/runs.all items; child fields override them. worktree:true isolates each child and returns handoff artifacts. usageBudget is shared across the workflow; already-running children are not stopped.
+• Missions auto-attach unless mission:false; await state.get(key)/state.set(key,JSONValue) requires a mission. See guide topic missions. Omit acceptance for reviewer/read-only calls; acceptance.review.required requests independent writer review.
+• Management discovery: list/get/models/guide; create/update/delete/eject/disable/enable/reset/refine; mission.*, schedule.*, watchdog.*, inspector.*, project.*, lane.status/recordMerge/recordSupersession; worktree.discard and plan-only worktree.cleanup; doctor and grant-spawn-budget. Use guide topics agents, missions, observability, tool-reference, configuration, models, watchdog or extension-api for exact action fields. Schedules take script inputs, not direct children; recipes live in the missions guide.`;
 function isToolDescriptionMode(value) {
   return value === "full" || value === "compact" || value === "custom";
 }
@@ -175,7 +127,7 @@ function loadCustomToolDescription(options) {
   return;
 }
 function withMandatorySafetyGuidance(description) {
-  const customDescription = description.split(SUBAGENT_SAFETY_GUIDANCE).map((part) => part.trim()).filter(Boolean).join(`
+  const customDescription = description.split(SUBAGENT_SAFETY_GUIDANCE).flatMap((part) => part.split(SUBAGENT_FAILURE_RECOVERY_GUIDANCE)).map((part) => part.trim()).filter(Boolean).join(`
 
 `);
   return customDescription ? `${customDescription}

package/extensions/dm-subagents/src/inspectors/actions.js

@@ -0,0 +1,145 @@
+import * as fs from "node:fs";
+import * as path from "node:path";
+import { fileURLToPath } from "node:url";
+import { readMissionBinding } from "../missions/lifecycle.js";
+import { listMissions, missionRecordPath, resolveMissionStoreLocation } from "../missions/store.js";
+import { resolveAuthorityDecision } from "../policy/authority.js";
+import { DIRS } from "../shared/types.js";
+import { readStatus } from "../shared/utils.js";
+import { resolveSubagentRunId } from "../runs/background/run-id-resolver.js";
+import { resolveNodeExecutable } from "../shared/node-executable.js";
+import { encodeSessionRoots } from "./session-roots-codec.js";
+import { formatShellCommand } from "./shell-command.js";
+export { INSPECTOR_ACTIONS } from "./types.js";
+function result(text, isError = false) {
+  const response = {
+    content: [{ type: "text", text }],
+    details: { mode: "management", results: [] }
+  };
+  if (isError)
+    response.isError = true;
+  return response;
+}
+function pathWithin(base, candidate) {
+  const resolvedBase = path.resolve(base);
+  const resolvedCandidate = path.resolve(candidate);
+  return resolvedCandidate === resolvedBase || resolvedCandidate.startsWith(`${resolvedBase}${path.sep}`);
+}
+function trustedDir(dir, deps) {
+  try {
+    if (fs.lstatSync(dir).isSymbolicLink() || !fs.statSync(dir).isDirectory())
+      return false;
+    const real = fs.realpathSync(dir);
+    if ([...deps.state?.asyncJobs.values() ?? [], ...deps.state?.fleetJobs?.values() ?? []].some((job) => {
+      try {
+        return fs.realpathSync(job.asyncDir) === real;
+      } catch {
+        return false;
+      }
+    }))
+      return true;
+    const root = deps.asyncDirRoot ?? DIRS.async;
+    return fs.existsSync(root) && pathWithin(root, dir) && pathWithin(fs.realpathSync(root), real);
+  } catch {
+    return false;
+  }
+}
+function resolveTarget(params, deps) {
+  const requested = params.id ?? params.runId;
+  let runId;
+  let asyncDir;
+  if (params.dir) {
+    asyncDir = path.resolve(params.dir);
+    if (!trustedDir(asyncDir, deps))
+      return { error: `Async run directory '${asyncDir}' is outside trusted run roots.` };
+    const status = readStatus(asyncDir);
+    if (!status)
+      return { error: `No async run status found in '${asyncDir}'.` };
+    if (requested && requested !== status.runId && !status.runId.startsWith(requested))
+      return { error: `Run '${requested}' does not match status run '${status.runId}'.` };
+    runId = status.runId;
+  } else {
+    if (!requested)
+      return { error: "Inspector actions require id or dir." };
+    try {
+      const found = resolveSubagentRunId(requested, { state: deps.state, asyncDirRoot: deps.asyncDirRoot ?? DIRS.async, resultsDir: deps.resultsDir ?? DIRS.results });
+      if (!found)
+        return { error: `No subagent run found for '${requested}'.` };
+      if (found.kind !== "async" || !found.location.asyncDir)
+        return { error: `Run '${found.id}' is not an inspectable async run with lifecycle artifacts.` };
+      runId = found.id;
+      asyncDir = found.location.asyncDir;
+    } catch (cause) {
+      return { error: cause instanceof Error ? cause.message : String(cause) };
+    }
+  }
+  const status = readStatus(asyncDir);
+  if (!status)
+    return { error: `No lifecycle status exists for async run '${runId}'.` };
+  if (params.index !== undefined && (params.index < 0 || params.index >= (status.steps?.length ?? 0)))
+    return { error: `Async run '${runId}' has ${status.steps?.length ?? 0} children. Index ${params.index} is out of range.` };
+  const target = { runId, asyncDir, status: { cwd: status.cwd, state: status.state, steps: status.steps } };
+  if (params.index !== undefined)
+    target.index = params.index;
+  return target;
+}
+function missionFor(target, deps) {
+  try {
+    const binding = readMissionBinding(target.asyncDir);
+    if (binding)
+      return { id: binding.missionId, path: missionRecordPath(binding.location, binding.missionId) };
+    const location = resolveMissionStoreLocation(deps.missions ? { projectRoot: deps.cwd, config: deps.missions } : { projectRoot: deps.cwd });
+    const mission = listMissions(location).records.find((record) => record.runs.some((run) => run.runId === target.runId));
+    return mission ? { id: mission.id, path: missionRecordPath(location, mission.id) } : undefined;
+  } catch {
+    return;
+  }
+}
+function launchFor(target, deps) {
+  const mission = missionFor(target, deps);
+  const runnerPath = deps.runnerPath ?? fileURLToPath(new URL("../../inspector-runner.mjs", import.meta.url));
+  const job = deps.state?.asyncJobs.get(target.runId) ?? deps.state?.fleetJobs?.get(target.runId);
+  const sessionRoots = [...new Set([...deps.sessionRoots ?? deps.state?.trustedSessionRoots ?? [], ...job?.sessionRoot ? [job.sessionRoot] : []])];
+  const allowSteer = resolveAuthorityDecision({ action: "steerRun", policy: deps.authorityPolicy }) === "auto";
+  const allowStop = resolveAuthorityDecision({ action: "stopRun", policy: deps.authorityPolicy }) === "auto";
+  const argv = [runnerPath, "--async-dir", target.asyncDir, "--run-id", target.runId, "--allow-steer", String(allowSteer), "--allow-stop", String(allowStop), "--session-roots", encodeSessionRoots(sessionRoots)];
+  if (target.index !== undefined)
+    argv.push("--index", String(target.index));
+  if (mission)
+    argv.push("--mission-path", mission.path);
+  const executable = resolveNodeExecutable();
+  const launch = { executable, argv, displayCommand: formatShellCommand(executable, argv), allowSteer, allowStop, sessionRoots };
+  if (mission)
+    launch.mission = mission;
+  return launch;
+}
+export async function handleInspectorAction(action, params, deps) {
+  const target = resolveTarget(params, deps);
+  if ("error" in target)
+    return result(target.error, true);
+  const context = {
+    cwd: deps.cwd,
+    signal: deps.signal,
+    env: deps.env ?? process.env,
+    target
+  };
+  if (deps.now)
+    context.now = deps.now;
+  if (action === "inspector.command")
+    return result(launchFor(target, deps).displayCommand);
+  const plugins = deps.plugins ?? [];
+  if (action === "inspector.open") {
+    for (const plugin of plugins) {
+      if (await plugin.available(context))
+        return plugin.open(context, launchFor(target, deps), params);
+    }
+    return result("No inspector plugin is available. Start a supported inspector host, or use inspector.command for a standalone command.", true);
+  }
+  const owner = plugins.find((plugin) => plugin.owns(context));
+  if (!owner)
+    return result(`No inspector plugin owns this binding for async run ${target.runId}.`);
+  if (action === "inspector.status") {
+    return owner.status ? owner.status(context) : result(`Inspector plugin '${owner.name}' does not support status for async run ${target.runId}.`, true);
+  }
+  return owner.close ? owner.close(context) : result(`Inspector plugin '${owner.name}' does not support close for async run ${target.runId}.`, true);
+}

package/extensions/dm-subagents/src/inspectors/ghostty/actions.js

@@ -0,0 +1,60 @@
+import { execFile } from "node:child_process";
+export const GHOSTTY_APPLESCRIPT = `on run argv
+  set launchCommand to item 1 of argv
+  set launchCwd to item 2 of argv
+  set shouldFocus to item 3 of argv
+  tell application "Ghostty"
+    set sourceWindow to front window
+    set sourceTab to selected tab of sourceWindow
+    set sourceTerminal to focused terminal of sourceTab
+    set surfaceConfiguration to new surface configuration
+    set initial working directory of surfaceConfiguration to launchCwd
+    set command of surfaceConfiguration to launchCommand
+    set newTerminal to split sourceTerminal direction right with configuration surfaceConfiguration
+    if shouldFocus is "true" then
+      focus newTerminal
+    else
+      focus sourceTerminal
+    end if
+    return id of newTerminal
+  end tell
+end run`;
+function defaultRunner(args, options) {
+  return new Promise((resolve, reject) => {
+    execFile("/usr/bin/osascript", [...args], options, (error, stdout, stderr) => {
+      if (error) {
+        const failure = error instanceof Error ? error : new Error(String(error));
+        reject(failure);
+        return;
+      }
+      resolve({ stdout, stderr });
+    });
+  });
+}
+function result(text, isError = false) {
+  const response = {
+    content: [{ type: "text", text }],
+    details: { mode: "management", results: [] }
+  };
+  if (isError)
+    response.isError = true;
+  return response;
+}
+const failureHint = "Ghostty inspector requires Ghostty 1.3+ and Automation permission for osascript.";
+export async function openGhosttyInspector(context, launch, params, runner = defaultRunner) {
+  try {
+    const output = await runner(["-e", GHOSTTY_APPLESCRIPT, "--", launch.displayCommand, context.target.status.cwd ?? context.cwd, params.focus === true ? "true" : "false"], {
+      signal: context.signal,
+      timeout: 15000,
+      encoding: "utf8",
+      maxBuffer: 64 * 1024
+    });
+    const id = output.stdout.trim();
+    if (!id)
+      return result(`Ghostty inspector error: Ghostty returned an empty terminal id. ${failureHint}`, true);
+    return result(`Opened read-only Ghostty inspector terminal ${id} for async run ${context.target.runId}. Status and close are unavailable because this plugin writes no binding.`);
+  } catch (cause) {
+    const message = cause instanceof Error ? cause.message : String(cause);
+    return result(`Ghostty inspector error: ${message}. ${failureHint}`, true);
+  }
+}

package/extensions/dm-subagents/src/inspectors/ghostty/plugin.js

@@ -0,0 +1,17 @@
+import { openGhosttyInspector } from "./actions.js";
+const GHOSTTY_BUNDLE_ID = "com.mitchellh.ghostty";
+export function createGhosttyInspectorPlugin(deps = {}) {
+  const platform = deps.platform ?? process.platform;
+  return {
+    name: "ghostty",
+    available: (context) => {
+      if (platform !== "darwin")
+        return false;
+      if (context.env.TERM_PROGRAM?.toLowerCase() !== "ghostty")
+        return false;
+      return context.env.__CFBundleIdentifier?.trim() === GHOSTTY_BUNDLE_ID;
+    },
+    owns: () => false,
+    open: (context, launch, params) => openGhosttyInspector(context, launch, params, deps.runner)
+  };
+}