@rosthq/cli 0.7.24 → 0.7.26

package/dist/index.js

@@ -37718,6 +37718,19 @@ var dryRunToolActionSchema = external_exports.object({
   // The underlying manifest guard decision that produced `status`.
   guard: dryRunGuardResultSchema
 }).strict();
+var dryRunLaneSchema = external_exports.enum(["cloud", "mcp_session", "runner"]);
+var dryRunToolExecutabilitySchema = external_exports.enum([
+  "executable",
+  "unimplemented_in_lane",
+  "held_for_steward"
+]);
+var dryRunGrantedToolSchema = external_exports.object({
+  // The granted tool (manifest id).
+  tool_name: external_exports.string().min(1),
+  // Its lane-aware executability, derived from the agent's lane + the runtime
+  // handler registry (never an id string pattern).
+  executability: dryRunToolExecutabilitySchema
+}).strict();
 var dryRunTranscriptSchema = external_exports.object({
   // Whether the rehearsal passed (the must-escalate clause was escalated and no
   // external side effect ran) or failed (it acted without escalating, or the run
@@ -37736,6 +37749,14 @@ var dryRunTranscriptSchema = external_exports.object({
   // handlers. Optional so the field is additive — a transcript without it (the
   // static demo replay, an older record) still validates.
   tool_preview: external_exports.array(dryRunToolActionSchema).optional(),
+  // DER-1294: the runtime lane this rehearsal reflects. Optional/additive —
+  // transcripts recorded before DER-1294 (and the static public-demo replay)
+  // omit it, so they still validate.
+  lane: dryRunLaneSchema.optional(),
+  // DER-1294: every GRANTED tool (manifest allow|escalate) with its lane-aware
+  // executability, so an operator sees which grants actually run vs. silently
+  // no-op in this lane. Optional/additive so older transcripts still validate.
+  granted_tools: external_exports.array(dryRunGrantedToolSchema).optional(),
   // The underlying run id (present once the run is recorded).
   run_id: external_exports.string().min(1).optional()
 }).strict();
@@ -42457,6 +42478,10 @@ var softwareBuildRequestSummarySchema = external_exports.object({
   current_phase: softwarePhaseSchema,
   inferred_risk: softwareRiskLevelSchema,
   automation_mode: softwareAutomationModeSchema,
+  max_iterations: external_exports.number().int().nullable(),
+  max_wall_clock_seconds: external_exports.number().int().nullable(),
+  cost_budget_usd: external_exports.string().nullable(),
+  iterations_used: external_exports.number().int(),
   created_at: external_exports.string()
 });
 var softwarePhaseRunSummarySchema = external_exports.object({
@@ -42510,6 +42535,21 @@ var softwareTaskSummarySchema = external_exports.object({
   completed_at: external_exports.string().nullable(),
   created_at: external_exports.string()
 });
+var softwarePlanVersionSummarySchema = external_exports.object({
+  id: uuid10,
+  build_request_id: uuid10,
+  version_no: external_exports.number().int(),
+  status: softwarePlanVersionStatusSchema,
+  summary: external_exports.string().nullable(),
+  confidence_score: external_exports.number().int().nullable(),
+  clarifying_questions: external_exports.array(external_exports.string()).default([]),
+  clarification_answers: external_exports.array(external_exports.object({
+    question: external_exports.string(),
+    answer: external_exports.string()
+  })).default([]),
+  created_by_seat_id: uuid10.nullable(),
+  created_at: external_exports.string()
+});
 var softwarePlanConformanceFindingSummarySchema = external_exports.object({
   id: uuid10,
   build_request_id: uuid10,
@@ -42607,7 +42647,10 @@ var softwareRequestShowOutputSchema = external_exports.object({
   request: softwareBuildRequestSummarySchema,
   current_phase: softwarePhaseSchema,
   phase_runs: external_exports.array(softwarePhaseRunSummarySchema),
-  gates: external_exports.array(softwareGateSummarySchema)
+  gates: external_exports.array(softwareGateSummarySchema),
+  plan_versions: external_exports.array(softwarePlanVersionSummarySchema),
+  changesets: external_exports.array(softwareChangesetSummarySchema),
+  tasks: external_exports.array(softwareTaskSummarySchema)
 });
 var softwarePhaseAdvanceInputSchema = external_exports.object({
   build_request_id: uuid10,
@@ -44367,7 +44410,7 @@ No orphan agents. No raw secrets in prompts, logs, or tool arguments. No durable
     order: 41,
     title: "Add agents to your Responsibility Graph",
     summary: "The visual journey for adding an agent seat: where to start, choosing a mode, placing the seat, naming a Steward, setup, the safety gates, and go-live.",
-    version: "2026-06-21.3",
+    version: "2026-07-04.1",
     public: true,
     audiences: ["human", "in_app_agent"],
     stages: ["staffing"],
@@ -44416,8 +44459,8 @@ The wizard builds a teammate, not a prompt:
 
 1. **Responsibilities and authority** \u2014 what the seat owns, what it may do alone, what needs approval, and what must escalate. Every clause is editable before sign-off.
 2. **Signals** \u2014 the measurables the seat is judged on.
-3. **Skills** \u2014 pin published, immutable Skill versions that the agent should carry. Skills are procedures, not authority: selecting one creates a proposed assignment and surfaces required/optional tool dependencies before sign-off.
-4. **Tools** \u2014 connect or decline each proposed tool. Declining a tool updates the permission manifest and the dry-run task so the agent stays safe. Credentials go in through write-only credential ingress and are stored as vault references, never shown again. Missing required Skill tools block sign-off, dry-run, and go-live until resolved; missing optional Skill tools warn without blocking.
+3. **Skills** \u2014 pin published, immutable Skill versions that the agent should carry. Skills are procedures, not authority: selecting one creates a proposed assignment and surfaces required/optional tool dependencies before access approval. The builder shows the provider and Access level each Skill proposes so the human sees the tool bundle before it enters the draft manifest.
+4. **Access** \u2014 connect or decline each proposed tool. Declining a tool updates the permission manifest and the dry-run task so the agent stays safe. Secrets go in through write-only credential ingress and are stored as vault references, never shown again. Missing required Skill tools block access approval, dry-run, and go-live until resolved; missing optional Skill tools warn without blocking.
 5. **Schedule and lane** \u2014 when the agent runs and on which lane. These stay advanced unless the mode needs them.
 
 You can save and resume a draft setup at any point. Returning to the flow restores the parent, Steward, mode, answers, Skill selections, tool decisions, and dry-run result.
@@ -44427,7 +44470,7 @@ You can save and resume a draft setup at any point. Returning to the flow restor
 Every mode ends at the same conservative gates, shown at each step:
 
 1. **Draft Charter** \u2014 a draft, not an active job, until a human approves it.
-2. **Signed manifest** \u2014 a human signs the permission manifest before any tool is live.
+2. **Approve access** \u2014 a human approves the permission manifest before any tool is live. The command remains \`charter.sign_manifest\`; the UI names the human decision, not the storage primitive.
 3. **Sandbox dry run** \u2014 the agent rehearses a representative task and is expected to escalate where the Charter says it must. A failed dry run keeps the draft and shows the reason; you edit and rerun.
 4. **Human go-live** \u2014 a person promotes the agent live. An agent never approves its own setup, signs its own manifest, or goes itself live.
 
@@ -44443,14 +44486,14 @@ Reopening the builder for a seat whose agent is already live shows its live stat
 
 In read-only or demo mode the **Add agent** affordance never starts a write. The public demo instead replays the add-an-agent journey end to end \u2014 describe the role, watch the draft Charter assemble, see the four safety gates light, and watch a sandbox dry run reach the must-escalate boundary and stop \u2014 then routes go-live to sign-up, because going live is a human decision.
 
-For CLI and MCP setup, read agent-skill-setup-guide before adding or assigning Skills. List existing company Skills and the {{brand}} catalog first, read the full Skill file before proposing it, and run the dependency check before sign-off. A missing required Skill tool is fixed by updating the Charter permission manifest through tool configuration; it is never fixed by treating the Skill as authorization.`
+For CLI and MCP setup, read agent-skill-setup-guide before adding or assigning Skills. List existing company Skills and the {{brand}} catalog first, read the full Skill file before proposing it, and run the dependency check before access approval. A missing required Skill tool is fixed by updating the Charter permission manifest through tool configuration; it is never fixed by treating the Skill as authorization.`
   },
   {
     slug: "custom-agents-guide",
     order: 42,
     title: "Design a custom agent",
     summary: "How to build a custom agent from operational questions through the Charter Builder, tools, dry run, and go-live without writing prompts.",
-    version: "2026-06-21.1",
+    version: "2026-07-04.1",
     public: true,
     audiences: ["human", "cli", "mcp", "in_app_agent"],
     stages: ["staffing", "charter_design"],
@@ -44517,7 +44560,7 @@ The same path is command-backed:
 
 ## Dry run and go-live
 
-The dry run is a real sandbox rehearsal, not a stamp. It executes a mock-provider run derived from the Charter \u2014 the agent works against sandbox data only and is expected to escalate where the Charter's must-escalate clause requires it. The same rehearsal works on all three lanes: cloud, external MCP, and local Runner. External MCP dry runs require an active seat-scoped MCP token; Runner dry runs require a paired Runner. Missing substrate returns a typed precondition error, not a generic failure. The recorded run keeps the agent's actual lane, so the evidence you review matches the lane you selected. The result is earned: a run that escalates the must-escalate boundary passes; a run that acts on that boundary instead of escalating fails. A failed dry run keeps the draft and shows the reason so you can edit and rerun. The rehearsal returns a transcript \u2014 the steps the agent took and the escalation it raised \u2014 shown step by step in the builder and printed by the CLI, so you can see the governance model working before anything goes live. When the dry run passes, a human promotes the agent live. The dry run rehearses the specific model tier you chose, so once it passes the model is locked \u2014 changing the model requires re-running the dry run on the new model before go-live.
+The dry run is a real sandbox rehearsal, not a stamp. It executes a mock-provider run derived from the Charter \u2014 the agent works against sandbox data only and is expected to escalate where the Charter's must-escalate clause requires it. The same rehearsal works on all three lanes: cloud, external MCP, and local Runner. External MCP dry runs require an active seat-scoped MCP token; Runner dry runs require a paired Runner. Missing substrate returns a typed precondition error, not a generic failure. The recorded run keeps the agent's actual lane, so the evidence you review matches the lane you selected. The result is earned: a run that escalates the must-escalate boundary passes; a run that acts on that boundary instead of escalating fails. A failed dry run keeps the draft and shows the reason so you can edit and rerun. The rehearsal returns a transcript \u2014 the steps the agent took and the escalation it raised \u2014 shown step by step in the builder and printed by the CLI, so you can see the governance model working before anything goes live. The transcript is also lane-honest: each granted tool is marked executable, held for the steward, or granted but unimplemented in this lane, with a plain-count verdict such as "2 of 6 granted tools don't execute in the cloud lane yet" \u2014 so a grant that would silently no-op in the cloud runtime is visible before go-live instead of passing green. Runner and MCP-session agents run those tools through their own local skills, so their grants are not flagged as unimplemented. When the dry run passes, a human promotes the agent live. The dry run rehearses the specific model tier you chose, so once it passes the model is locked \u2014 changing the model requires re-running the dry run on the new model before go-live.
 
 ## When to stop for confirmation
 
@@ -44860,7 +44903,7 @@ External connectors are being rolled out provider by provider, conservatively (r
     order: 48,
     title: "CLI and MCP installation guide",
     summary: "Install the public CLI, register remote token-backed MCP clients, and find the full command and tool catalog.",
-    version: "2026-07-04.16",
+    version: "2026-07-04.17",
     public: true,
     audiences: ["human", "cli", "mcp", "in_app_agent"],
     stages: ["company_setup", "staffing"],
@@ -45603,7 +45646,7 @@ Several rows here are seat-operating commands (\`task.create\`, the \`signal.*\`
 | \`rost_list_forge_github_installations\` | \`software_factory.github.installation.list\` | List connected GitHub App installations and each repository's bind state (connected or active) with the Forge project it maps to. Read-only. | Tenant | Call with \`{}\`; a \`connected\` repository is verified but not yet bound to a project. |
 | \`rost_list_forge_runner_capacity\` | \`software_factory.capacity.list\` | List Forge runner capacity observations (advisory scheduling input only, never an authority input). Read-only; requires the Forge add-on. | Tenant | Call with \`{}\` or \`{"runner_id":"<runner-id>"}\`. |
 
-The web Forge control room at \`/forge\` uses the same command path for project selection, build request creation/list/detail reads, runner capacity, configuration status, and Developer Team install. The install button deliberately stages a pending confirmation before durable seats, Skills, authority grants, or live agent activation are applied; the install page states plainly that approving activates the 6 agents at a pull-request-only, observe-first posture and surfaces the company autonomy ceiling.
+The web Forge control room at \`/forge\` uses the same command path for project selection, build request creation/list/detail reads, runner capacity, configuration status, and Developer Team install. A request detail page shows phase history, open gates, plan progress, loop/time/cost ceilings, and latest plan-review questions from \`software_factory.request.show\`; answering a clarification, approving a plan, or rejecting it calls the same \`software_factory.request.answer_clarification\`, \`software_factory.plan.approve\`, and \`software_factory.plan.reject\` commands exposed over CLI/MCP. The stop control is visible only as a non-executing placeholder until pause/cancel commands are wired, so the UI does not imply a capability that is not live. The install button deliberately stages a pending confirmation before durable seats, Skills, authority grants, or live agent activation are applied; the install page states plainly that approving activates the 6 agents at a pull-request-only, observe-first posture and surfaces the company autonomy ceiling.
 
 #### Forge GitHub App access
 
@@ -46089,7 +46132,7 @@ Retrieve the narrowest relevant guide before making a setup recommendation. Pref
     order: 60,
     title: "Cascade guide",
     summary: "How to connect company goals to seat-level work without turning {{brand}} into a project-management tool.",
-    version: "2026-07-04.1",
+    version: "2026-07-04.2",
     public: true,
     audiences: ["human", "cli", "mcp", "in_app_agent"],
     stages: ["operating_rhythm"],
@@ -46157,7 +46200,7 @@ Cascade warnings are advisory, not blockers. They point to setup that should be
 - Agent goal missing human accountability parent: an agent-owned goal has no human or hybrid goal in its ancestry.
 - Orphaned goal branch: the owner seat or parent chain no longer reaches an active Compass objective.
 
-On \`/cascade\`, the screen renders the full goal depth, not just one or two levels. Off-track descendant branches stay open so the operator can see why a parent rolled up to risk.
+\`/cascade\` leads with one computed line \u2014 how many goals are off pace this cycle, or an all-clear when none are. Off-pace and off-track goals stay fully visible with their owner seat and pace context; goals confirmed on pace collapse into a per-branch count you can expand. Creating or editing a goal opens the same drawer from a single **New goal** action or a goal's own edit control \u2014 there is no separate persistent form.
 
 ## Goal progress and at-risk goals
 
@@ -46304,7 +46347,7 @@ Agents may record readings when the Charter allows it. Agent-reported readings s
     order: 62,
     title: "Friction guide",
     summary: "How to capture issues with evidence, rank them, and resolve them without losing ownership.",
-    version: "2026-07-01.2",
+    version: "2026-07-04.1",
     public: true,
     audiences: ["human", "cli", "mcp", "in_app_agent"],
     stages: ["operating_rhythm"],
@@ -46356,6 +46399,8 @@ Agents should file Friction when a measurable breaks, a tool fails, a repeated e
 
 In the app, operational evidence is summarized for humans instead of shown as the primary interface. Incident fields, proposed tool calls, schedules, payload labels, and source links render as readable fields with raw details tucked behind supporting context when useful. Secret-shaped values stay redacted.
 
+The Friction page opens with one computed line \u2014 how many issues are open, the oldest one's age, and how many are escalated \u2014 followed by the open issues themselves as incident-ticket rows (filer, one-line cause, an escalation chip naming the steward when severity is critical). Resolved issues collapse into a single count that expands into a quiet history list. Filing a new issue and the three-step resolution (root cause, owner task, done-by) both live behind one "File friction" action, which opens a drawer; resolving launches that same drawer already seeded from the row you clicked "Resolve" on.
+
 ## Operate Friction, tasks, and escalations from CLI or MCP
 
 Friction, tasks, and escalations are the issue-to-action loop. A seat files, a task carries the work, and an escalation routes a decision a seat cannot make alone.
@@ -46617,7 +46662,7 @@ Do not claim that a live charge happened unless Stripe test/live evidence proves
     order: 72,
     title: "Settings guide",
     summary: "How to use Settings as the control plane for company access, channels, providers, tokens, and operating defaults.",
-    version: "2026-07-03.2",
+    version: "2026-07-04.1",
     public: true,
     audiences: ["human", "cli", "mcp", "in_app_agent"],
     stages: ["company_setup", "staffing"],
@@ -46640,7 +46685,9 @@ Start with members and invites, then provider and channel connections, then MCP
 
 ## Navigating Settings
 
-Settings is a left-rail master\u2013detail layout: a grouped rail (General, Team & Access, Billing, AI & Agents, Connections) selects one section at a time, shown in the detail pane. The active section is a URL search param, so \`/settings?section=<id>\` is linkable and survives a refresh. Legacy hash deep-links still resolve \u2014 \`/settings#integrations\`, \`#members\`, \`#runners\`, \`#guardrails\`, \`#vault\` and the rest redirect to the matching \`?section=\` while preserving any status params. Documented shortcut links resolve to the right section instead of a dead end: \`/settings/integrations\` and \`/readiness\` open the Integrations section (connector readiness lives inside it, not on a standalone page), and \`/settings/policy\` opens the Product learning section. After you save a change, the app keeps you on the section you edited rather than returning to the top.
+Settings is a left-rail master\u2013detail layout: a grouped rail (General, Team & Access, Billing, AI & Agents, Connections) selects one section at a time, shown in the detail pane. The active section is a URL search param, so \`/settings?section=<id>\` is linkable and survives a refresh. The Connections section is the workspace-level provider surface: Google, Slack, QuickBooks, local runners, seat-scoped agent tokens, and stored secrets appear as provider cards with status, used-for chips, actions, and the shared Access language (**Off / Read / Draft / Act with approval**). Detailed management forms for tokens, channels, runners, provider links, and stored vault entries remain reachable from those cards.
+
+Legacy hash deep-links still resolve \u2014 \`/settings#integrations\`, \`#channels\`, \`#runners\`, and \`#vault\` redirect to \`?section=connections\` while preserving any status params; other hashes such as \`#members\` and \`#guardrails\` redirect to their matching section. Documented shortcut links resolve to the right section instead of a dead end: \`/settings/integrations\` and \`/readiness\` open the Connections section (connector readiness lives inside it, not on a standalone page), and \`/settings/policy\` opens the Product learning section. After you save a change, the app keeps you on the section you edited rather than returning to the top.
 
 ## What belongs in Settings
 
@@ -46797,7 +46844,7 @@ Every notification should include the seat, cause, evidence, and requested decis
     order: 75,
     title: "Local runner guide",
     summary: "How local agent sessions and runner surfaces should operate through {{brand}} without bypassing Charters or audit.",
-    version: "2026-07-04.3",
+    version: "2026-07-04.4",
     public: true,
     audiences: ["human", "cli", "mcp", "in_app_agent"],
     stages: ["staffing", "operating_rhythm"],
@@ -46841,6 +46888,8 @@ Headless and service accounts are supported first-class: when no usable OS keych
 
 Choose the local runtime with \`--runtime auto|claude|codex\` (default \`auto\`). The same short-lived seat-scoped MCP token and server-side guard apply regardless of runtime; results are tagged \`claude-cli\` or \`codex-cli\`.
 
+For always-on execution, \`runner serve --execute\` heartbeats on its normal interval and also checks for claimable work between heartbeats. The default interactive claim cadence is about one second; set \`RUNNER_INTERACTIVE_CLAIM_MS\` or \`--interactive-claim-ms\` to tune it for a host. Claude-backed turns reuse a per-seat Claude Code resume session when Claude reports a session id, so later turns can run with \`claude --resume <session-id>\` under the local subscription account. The runner stores only the opaque resume id in its owner-only state file; it does not use API-billed managed sessions to make subscription runner turns faster. A resident long-lived Claude/Codex process is still prototype-gated, so do not assume a fixed latency SLA from this guide alone.
+
 The runner bearer secret is stored locally and is never displayed. Each claimed turn receives a short-lived seat-scoped MCP token from {{brand}}, and server-side guards still decide what tools the Seat may use.
 
 ## Interactive AICOS runner turns
@@ -49113,6 +49162,36 @@ function dryRunToolActionLabel(status) {
       return status;
   }
 }
+var DRY_RUN_LANE_LABELS = {
+  cloud: "cloud lane",
+  runner: "runner lane",
+  mcp_session: "MCP-session lane"
+};
+function dryRunLaneLabel(lane) {
+  return DRY_RUN_LANE_LABELS[lane] ?? `${lane} lane`;
+}
+function dryRunLaneVerdict(lane, grantedCount, unimplementedCount) {
+  const laneLabel = dryRunLaneLabel(lane);
+  if (unimplementedCount === 0) {
+    return grantedCount === 1 ? `The single granted tool is implemented in the ${laneLabel}.` : `All ${grantedCount} granted tools are implemented in the ${laneLabel}.`;
+  }
+  const verb = unimplementedCount === 1 ? "doesn't" : "don't";
+  return `${unimplementedCount} of ${grantedCount} granted tools ${verb} execute in the ${laneLabel} yet.`;
+}
+function dryRunLaneExecutabilityLines(transcript) {
+  const lane = transcript.lane;
+  const grantedTools = asArray(transcript.granted_tools);
+  if (typeof lane !== "string" || grantedTools.length === 0) {
+    return [];
+  }
+  const laneLabel = dryRunLaneLabel(lane);
+  const unimplemented = grantedTools.map((tool) => asRecord(tool)).filter((tool) => tool.executability === "unimplemented_in_lane");
+  const warningLines = unimplemented.map(
+    (tool) => `  warning: ${field(tool, "tool_name")} is granted but unimplemented in the ${laneLabel}`
+  );
+  const verdict = dryRunLaneVerdict(lane, grantedTools.length, unimplemented.length);
+  return ["lane executability:", ...warningLines, `  verdict: ${verdict}`];
+}
 function formatDryRunTranscript(value) {
   const transcript = asRecord(value);
   const steps = asArray(transcript.steps);
@@ -49129,7 +49208,13 @@ function formatDryRunTranscript(value) {
     const record2 = asRecord(action);
     return `  ${dryRunToolActionLabel(field(record2, "status"))}: ${field(record2, "intended_action")} [${field(record2, "tool_name")}]`;
   });
-  return ["rehearsal:", ...lines, ...toolPreview.length > 0 ? ["tool preview:", ...toolPreview] : []].join("\n");
+  const laneLines = dryRunLaneExecutabilityLines(transcript);
+  return [
+    "rehearsal:",
+    ...lines,
+    ...toolPreview.length > 0 ? ["tool preview:", ...toolPreview] : [],
+    ...laneLines
+  ].join("\n");
 }
 function isScalar(value) {
   return value === null || ["string", "number", "boolean"].includes(typeof value);
@@ -51328,7 +51413,7 @@ function parseSemver2(value) {
 // src/runner-serve.ts
 import { execFile as execFileCallback3, spawn } from "node:child_process";
 import { createHash } from "node:crypto";
-import { chmod, mkdir as mkdir2, readFile as readFile4, rm as rm2, writeFile as writeFile2 } from "node:fs/promises";
+import { chmod, mkdir as mkdir2, readFile as readFile4, rename, rm as rm2, writeFile as writeFile2 } from "node:fs/promises";
 import { cpus, hostname as hostname3, platform, totalmem, tmpdir } from "node:os";
 import path3 from "node:path";
 import { setTimeout as sleep2 } from "node:timers/promises";
@@ -51373,6 +51458,7 @@ ${usage()}
   do {
     const detectedCapabilities = await detectCapabilities();
     const capabilities = capabilitiesForConfiguredRuntime(detectedCapabilities, config2.runtime);
+    const localRuntime = config2.execute ? effectiveExecutionRuntime(config2.runtime, capabilities) : null;
     try {
       const heartbeat = await post2(fetchImpl, options.appUrl, "/api/runner/heartbeat", {
         capabilities,
@@ -51387,15 +51473,12 @@ ${usage()}
         options.io.stdout.write(`heartbeat ok #${beat} runner_id=${state.runner_id} claude=${capabilities.claude.installed} codex=${capabilities.codex.installed}
 `);
       }
-      if (config2.execute) {
-        const localRuntime = effectiveExecutionRuntime(config2.runtime, capabilities);
-        if (localRuntime) {
-          activeSessions += 1;
-          try {
-            await claimAndExecute(fetchImpl, options.appUrl, state, options.io, localRuntime);
-          } finally {
-            activeSessions = Math.max(0, activeSessions - 1);
-          }
+      if (localRuntime) {
+        activeSessions += 1;
+        try {
+          await claimAndExecute(fetchImpl, options.appUrl, state, config2, options.io, localRuntime);
+        } finally {
+          activeSessions = Math.max(0, activeSessions - 1);
         }
       }
     } catch (error51) {
@@ -51405,11 +51488,27 @@ ${usage()}
     }
     beat += 1;
     if (!config2.once) {
-      await sleep2(config2.heartbeatMs);
+      await waitForNextHeartbeat(fetchImpl, options.appUrl, state, config2, options.io, localRuntime);
     }
   } while (!config2.once);
   return heartbeatSucceeded && !failed ? 0 : 1;
 }
+async function waitForNextHeartbeat(fetchImpl, appUrl2, state, config2, io, runtime) {
+  const deadline = Date.now() + config2.heartbeatMs;
+  while (Date.now() < deadline) {
+    const intervalMs = Math.min(config2.interactiveClaimMs, deadline - Date.now());
+    await sleep2(intervalMs);
+    if (!runtime || Date.now() >= deadline) {
+      continue;
+    }
+    try {
+      await claimAndExecute(fetchImpl, appUrl2, state, config2, io, runtime, { interactiveOnly: true });
+    } catch (error51) {
+      io.stderr.write(`transient runner error (continuing): ${redactForLog(error51 instanceof Error ? error51.message : String(error51))}
+`);
+    }
+  }
+}
 function detectTelemetry(capabilities, env, activeSessions) {
   const accountAlias = (env.RUNNER_MODEL_ACCOUNT_ALIAS ?? env.ROST_RUNNER_MODEL_ACCOUNT_ALIAS ?? "").trim() || "default";
   const usageState = env.RUNNER_MODEL_ACCOUNT_STATE ?? env.ROST_RUNNER_MODEL_ACCOUNT_STATE ?? "unknown";
@@ -51431,6 +51530,7 @@ function parseServeArgs(args, appUrl2, env) {
   let name = env.RUNNER_NAME ?? `headless-${hostname3()}`;
   let stateFile = env.RUNNER_STATE_FILE ?? path3.join(tmpdir(), `rost-runner-${createHash("sha1").update(appUrl2).digest("hex").slice(0, 8)}.json`);
   let heartbeatMs = Number(env.RUNNER_HEARTBEAT_MS ?? 3e4);
+  let interactiveClaimMs = Number(env.RUNNER_INTERACTIVE_CLAIM_MS ?? 1e3);
   let execute2 = env.RUNNER_EXECUTE === "1";
   let once = false;
   let runtime = parseRuntime(env.RUNNER_RUNTIME ?? "auto");
@@ -51446,6 +51546,9 @@ function parseServeArgs(args, appUrl2, env) {
     } else if (token === "--heartbeat-ms") {
       heartbeatMs = Number(requiredValue(args, index));
       index += 1;
+    } else if (token === "--interactive-claim-ms") {
+      interactiveClaimMs = Number(requiredValue(args, index));
+      index += 1;
     } else if (token === "--user-code") {
       userCode = requiredValue(args, index);
       index += 1;
@@ -51463,7 +51566,11 @@ function parseServeArgs(args, appUrl2, env) {
   if (!Number.isSafeInteger(heartbeatMs) || heartbeatMs < 1e3) {
     throw new Error("--heartbeat-ms must be an integer >= 1000.");
   }
-  return { name, stateFile, heartbeatMs, execute: execute2, once, runtime, ...userCode ? { userCode } : {} };
+  if (!Number.isSafeInteger(interactiveClaimMs) || interactiveClaimMs < 250) {
+    throw new Error("--interactive-claim-ms must be an integer >= 250.");
+  }
+  interactiveClaimMs = Math.min(interactiveClaimMs, heartbeatMs);
+  return { name, stateFile, heartbeatMs, interactiveClaimMs, execute: execute2, once, runtime, ...userCode ? { userCode } : {} };
 }
 function parseRuntime(value) {
   if (value === "auto" || value === "claude" || value === "codex") {
@@ -51545,10 +51652,7 @@ async function pairIfNeeded(client, fetchImpl, appUrl2, config2, io) {
     ...typeof tenantId === "string" ? { tenant_id: tenantId } : config2.expectedTenantId ? { tenant_id: config2.expectedTenantId } : {},
     name: typeof claim.json.name === "string" ? claim.json.name : config2.name
   };
-  await mkdir2(path3.dirname(config2.stateFile), { recursive: true, mode: 448 });
-  await writeFile2(config2.stateFile, `${JSON.stringify(state)}
-`, { mode: 384 });
-  await chmod(config2.stateFile, 384);
+  await saveState(config2.stateFile, state);
   io.stdout.write(`paired runner_id=${runnerId} state_file=${config2.stateFile}
 `);
   return state;
@@ -51564,7 +51668,8 @@ async function loadState(filePath, expectedTenantId) {
         runner_id: parsed.runner_id,
         runner_secret: parsed.runner_secret,
         ...typeof parsed.tenant_id === "string" ? { tenant_id: parsed.tenant_id } : {},
-        ...typeof parsed.name === "string" ? { name: parsed.name } : {}
+        ...typeof parsed.name === "string" ? { name: parsed.name } : {},
+        ...isRuntimeSessionMap(parsed.runtime_sessions) ? { runtime_sessions: parsed.runtime_sessions } : {}
       };
     }
   } catch {
@@ -51572,6 +51677,31 @@ async function loadState(filePath, expectedTenantId) {
   }
   return null;
 }
+function isRuntimeSessionMap(value) {
+  if (!value || typeof value !== "object" || Array.isArray(value)) {
+    return false;
+  }
+  return Object.values(value).every((entry) => {
+    if (!entry || typeof entry !== "object" || Array.isArray(entry)) {
+      return false;
+    }
+    const record2 = entry;
+    return (record2.claude_session_id === void 0 || typeof record2.claude_session_id === "string") && (record2.updated_at === void 0 || typeof record2.updated_at === "string");
+  });
+}
+async function saveState(filePath, state) {
+  await mkdir2(path3.dirname(filePath), { recursive: true, mode: 448 });
+  const tempPath = `${filePath}.${process.pid}.${createHash("sha1").update(`${Date.now()}:${Math.random()}`).digest("hex").slice(0, 8)}.tmp`;
+  try {
+    await writeFile2(tempPath, `${JSON.stringify(state)}
+`, { mode: 384 });
+    await chmod(tempPath, 384);
+    await rename(tempPath, filePath);
+  } catch (error51) {
+    await rm2(tempPath, { force: true });
+    throw error51;
+  }
+}
 async function startPairing(client, config2) {
   const response = await client.execute("runner.pairing.start", { name: config2.name, platform: process.platform });
   const userCode = response.output && typeof response.output === "object" ? response.output.user_code : void 0;
@@ -51611,11 +51741,17 @@ async function post2(fetchImpl, appUrl2, pathName, body, token) {
   }
   return { status: response.status, json: json2 };
 }
-async function claimAndExecute(fetchImpl, appUrl2, state, io, runtime) {
-  const claimed = await post2(fetchImpl, appUrl2, "/api/runner/work-orders/claim", {}, state.runner_secret);
+async function claimAndExecute(fetchImpl, appUrl2, state, config2, io, runtime, options = {}) {
+  const claimed = await post2(
+    fetchImpl,
+    appUrl2,
+    "/api/runner/work-orders/claim",
+    options.interactiveOnly ? { interactive_only: true } : {},
+    state.runner_secret
+  );
   const turnExecution = claimed.json.turn_execution;
   if (claimed.status === 200 && turnExecution && typeof turnExecution === "object") {
-    await executeClaimedUnit(fetchImpl, appUrl2, state, io, runtime, {
+    await executeClaimedUnit(fetchImpl, appUrl2, state, config2, io, runtime, {
       kind: "turn_execution",
       idField: "turn_execution_id",
       item: turnExecution,
@@ -51629,7 +51765,7 @@ async function claimAndExecute(fetchImpl, appUrl2, state, io, runtime) {
   if (claimed.status !== 200 || !workOrder || typeof workOrder !== "object") {
     return;
   }
-  await executeClaimedUnit(fetchImpl, appUrl2, state, io, runtime, {
+  await executeClaimedUnit(fetchImpl, appUrl2, state, config2, io, runtime, {
     kind: "work_order",
     idField: "work_order_id",
     item: workOrder,
@@ -51638,7 +51774,7 @@ async function claimAndExecute(fetchImpl, appUrl2, state, io, runtime) {
     logRef: (id) => `runner-local:${id}`
   });
 }
-async function executeClaimedUnit(fetchImpl, appUrl2, state, io, runtime, input) {
+async function executeClaimedUnit(fetchImpl, appUrl2, state, config2, io, runtime, input) {
   const id = input.item.id;
   if (typeof id !== "string") {
     return;
@@ -51648,10 +51784,22 @@ async function executeClaimedUnit(fetchImpl, appUrl2, state, io, runtime, input)
     throw new Error(`${claimedUnitLabel(input.kind)} start failed ${started.status}: ${redactForLog(JSON.stringify(started.json))}`);
   }
   const turnRuntime = runtimeForClaimedUnit(input.item, runtime, input.kind);
-  const result = turnRuntime ? await runLocalTurn(appUrl2, input.item, turnRuntime, input.kind) : {
+  const result = turnRuntime ? await runLocalTurn(appUrl2, input.item, turnRuntime, input.kind, state) : {
     ok: false,
     summary: `Runner runtime ${runtime} cannot execute requested ${String(input.item.brain ?? "unknown")} turn.`
   };
+  if (input.kind === "turn_execution" && turnRuntime === "claude") {
+    try {
+      if (result.ok && result.claudeSessionId) {
+        await persistClaudeSessionId(config2.stateFile, state, input.item, result.claudeSessionId);
+      } else if (!result.ok && result.resumedFromSessionId) {
+        await clearClaudeSessionId(config2.stateFile, state, input.item);
+      }
+    } catch (error51) {
+      io.stderr.write(`resume cache update failed (continuing): ${redactForLog(error51 instanceof Error ? error51.message : String(error51))}
+`);
+    }
+  }
   const reported = await post2(fetchImpl, appUrl2, input.resultPath(id), {
     status: result.ok ? "succeeded" : "failed",
     summary: result.summary,
@@ -51676,16 +51824,16 @@ function runtimeForClaimedUnit(item, fallback, kind) {
   }
   return fallback === "claude" ? "claude" : null;
 }
-function runLocalTurn(appUrl2, workOrder, runtime, kind) {
-  return runtime === "codex" ? runCodexTurn(appUrl2, workOrder, kind) : runClaudeTurn(appUrl2, workOrder, kind);
+function runLocalTurn(appUrl2, workOrder, runtime, kind, state) {
+  return runtime === "codex" ? runCodexTurn(appUrl2, workOrder, kind) : runClaudeTurn(appUrl2, workOrder, kind, state);
 }
-function runClaudeTurn(appUrl2, workOrder, kind) {
-  return spawnRunnerTurn(appUrl2, workOrder, "claude", kind);
+function runClaudeTurn(appUrl2, workOrder, kind, state) {
+  return spawnRunnerTurn(appUrl2, workOrder, "claude", kind, state);
 }
 function runCodexTurn(appUrl2, workOrder, kind) {
   return spawnRunnerTurn(appUrl2, workOrder, "codex", kind);
 }
-function buildTurnCommand(runtime, prompt, configPath, kind) {
+function buildTurnCommand(runtime, prompt, configPath, kind, resumeSessionId) {
   if (runtime === "codex") {
     return {
       command: "codex",
@@ -51703,11 +51851,12 @@ function buildTurnCommand(runtime, prompt, configPath, kind) {
       "--allowedTools",
       (kind === "turn_execution" ? RUNNER_AICOS_TURN_TOOLS : RUNNER_WORK_ORDER_TOOLS).join(","),
       "--output-format",
-      kind === "turn_execution" ? "json" : "text"
+      kind === "turn_execution" ? "json" : "text",
+      ...resumeSessionId ? ["--resume", resumeSessionId] : []
     ]
   };
 }
-async function spawnRunnerTurn(appUrl2, workOrder, runtime, kind) {
+async function spawnRunnerTurn(appUrl2, workOrder, runtime, kind, state) {
   const mcp = typeof workOrder.mcp === "object" && workOrder.mcp !== null ? workOrder.mcp : {};
   const token = typeof mcp.token === "string" ? mcp.token : "";
   const url2 = typeof mcp.url === "string" ? `${appUrl2.replace(/\/+$/, "")}${mcp.url}` : `${appUrl2.replace(/\/+$/, "")}/mcp`;
@@ -51741,7 +51890,9 @@ async function spawnRunnerTurn(appUrl2, workOrder, runtime, kind) {
   await writeFile2(configPath, `${mcpConfig}
 `, { mode: 384 });
   await chmod(configPath, 384);
-  const { command, args } = buildTurnCommand(runtime, prompt, configPath, kind);
+  const resumeSessionId = runtime === "claude" ? getClaudeResumeSessionId(state, workOrder) : void 0;
+  const { command, args } = buildTurnCommand(runtime, prompt, configPath, kind, resumeSessionId);
+  const expectSessionEnvelope = runtime === "claude" && kind === "turn_execution";
   return new Promise((resolve2) => {
     let settled = false;
     const finish = (result) => {
@@ -51750,7 +51901,7 @@ async function spawnRunnerTurn(appUrl2, workOrder, runtime, kind) {
       }
       settled = true;
       void rm2(configPath, { force: true });
-      resolve2(result);
+      resolve2(resumeSessionId ? { ...result, resumedFromSessionId: resumeSessionId } : result);
     };
     const child = spawn(command, args, {
       cwd: tmpdir(),
@@ -51770,7 +51921,12 @@ async function spawnRunnerTurn(appUrl2, workOrder, runtime, kind) {
     });
     child.on("close", (code) => {
       clearTimeout(timer);
-      finish({ ok: code === 0, summary: extractRunnerSummary(redactForLog((out || err || "completed").trim())).slice(0, 12e3) });
+      const rawOutput = (out || err || "completed").trim();
+      finish({
+        ok: code === 0,
+        summary: extractRunnerSummary(redactForLog(rawOutput)).slice(0, 12e3),
+        ...expectSessionEnvelope ? extractClaudeSessionId(rawOutput) : {}
+      });
     });
     child.on("error", (error51) => {
       clearTimeout(timer);
@@ -51796,8 +51952,83 @@ function extractRunnerSummary(output) {
   }
   return trimmed;
 }
+function extractClaudeSessionId(output) {
+  const trimmed = output.trim();
+  if (!trimmed.startsWith("{")) {
+    return {};
+  }
+  try {
+    const parsed = JSON.parse(trimmed);
+    const sessionId = parsed.session_id ?? parsed.sessionId;
+    return typeof sessionId === "string" && sessionId.trim().length > 0 ? { claudeSessionId: sessionId.trim() } : {};
+  } catch {
+    return {};
+  }
+}
+function runtimeSessionKey(state, workOrder) {
+  const seatId = extractSeatId(workOrder);
+  const userId = extractTurnUserId(workOrder);
+  const sessionId = extractTurnSessionId(workOrder);
+  if (!state?.tenant_id || !seatId || !userId || !sessionId) {
+    return null;
+  }
+  return `${state.tenant_id}:seat:${seatId}:user:${userId}:session:${sessionId}:claude`;
+}
+function extractTurnUserId(workOrder) {
+  const userId = workOrder.user_id;
+  return typeof userId === "string" && userId.length > 0 ? userId : null;
+}
+function extractTurnSessionId(workOrder) {
+  const topLevel = workOrder.session_id;
+  if (typeof topLevel === "string" && topLevel.length > 0) {
+    return topLevel;
+  }
+  const aicos = typeof workOrder.aicos === "object" && workOrder.aicos !== null ? workOrder.aicos : null;
+  const sessionId = aicos?.session_id;
+  return typeof sessionId === "string" && sessionId.length > 0 ? sessionId : null;
+}
+function extractSeatId(workOrder) {
+  const seat = typeof workOrder.seat === "object" && workOrder.seat !== null ? workOrder.seat : null;
+  if (typeof seat?.id === "string" && seat.id.length > 0) {
+    return seat.id;
+  }
+  const charterRef = typeof workOrder.charter_ref === "object" && workOrder.charter_ref !== null ? workOrder.charter_ref : null;
+  return typeof charterRef?.seat_id === "string" && charterRef.seat_id.length > 0 ? charterRef.seat_id : null;
+}
+function getClaudeResumeSessionId(state, workOrder) {
+  const key = runtimeSessionKey(state, workOrder);
+  if (!key) {
+    return void 0;
+  }
+  const sessionId = state?.runtime_sessions?.[key]?.claude_session_id;
+  return typeof sessionId === "string" && sessionId.length > 0 ? sessionId : void 0;
+}
+async function persistClaudeSessionId(stateFile, state, workOrder, claudeSessionId) {
+  const key = runtimeSessionKey(state, workOrder);
+  if (!key) {
+    return;
+  }
+  state.runtime_sessions = {
+    ...state.runtime_sessions ?? {},
+    [key]: {
+      claude_session_id: claudeSessionId,
+      updated_at: (/* @__PURE__ */ new Date()).toISOString()
+    }
+  };
+  await saveState(stateFile, state);
+}
+async function clearClaudeSessionId(stateFile, state, workOrder) {
+  const key = runtimeSessionKey(state, workOrder);
+  if (!key || !state.runtime_sessions?.[key]) {
+    return;
+  }
+  const nextSessions = { ...state.runtime_sessions };
+  delete nextSessions[key];
+  state.runtime_sessions = nextSessions;
+  await saveState(stateFile, state);
+}
 function usage() {
-  return "Usage: rost runner serve [--name <text>] [--state-file <path>] [--heartbeat-ms <n>] [--user-code <code>] [--runtime auto|claude|codex] [--once] [--execute]";
+  return "Usage: rost runner serve [--name <text>] [--state-file <path>] [--heartbeat-ms <n>] [--interactive-claim-ms <n>] [--user-code <code>] [--runtime auto|claude|codex] [--once] [--execute]";
 }
 
 // src/runner-service.ts