humanish 0.106.0 → 0.106.1

package/dist/routes/computer-use/participant-model.js

@@ -3,7 +3,7 @@
 // finished session. The desktop side stays with the lane runner.
 import { startClaudeSession } from "../../actors/local-agent/claude-session.js";
 import { createLocalAgentProvider } from "../../actors/local-agent/cli.js";
-import { DEFAULT_OPENAI_CU_MODEL } from "../../actors/computer-use/openai-provider.js";
+import { pricedModel } from "../../lab/plan-base.js";
 import { estimateActorCostForExecution, round6 } from "../../run/pricing.js";
 import { redactText } from "../../evidence/redaction.js";
 import { createRestrictedCodexParticipant } from "../../actors/codex/restricted-participant.js";
@@ -14,7 +14,7 @@ import { hollowCompletion } from "../../run/judge.js";
 import { legacyHookSpecOf } from "./legacy-lane-spec.js";
 /** Starts the lane's model: a caller's provider, the operator's Codex, a Claude session, or none. */
 export async function startParticipantModel(spec, deps, executor) {
-    const { config, env } = deps;
+    const { config, env, brain } = deps;
     if (deps.hooks.buildProvider) {
         return {
             provider: await deps.hooks.buildProvider({
@@ -26,7 +26,8 @@ export async function startParticipantModel(spec, deps, executor) {
             }),
         };
     }
-    if (deps.localAgent === "codex") {
+    const localAgent = brain.kind === "local-agent" ? brain.agent : undefined;
+    if (localAgent === "codex") {
         // Hosted local-agent studies use the same native participant engine as local desktops.
         // Operator auth deliberately retains the operator's Codex home, config and supported auth
         // stores instead of applying the isolated restricted-account profile used by local studies.
@@ -35,13 +36,13 @@ export async function startParticipantModel(spec, deps, executor) {
             ...(spec.planned.limits.reasoningEffort === undefined
                 ? {}
                 : { reasoningEffort: spec.planned.limits.reasoningEffort }),
-            ...(config.actors[0]?.model === undefined ? {} : { model: config.actors[0].model }),
+            ...(brain.declaredModel === undefined ? {} : { model: brain.declaredModel }),
             ...(executor.speechEnabled === true ? { speechEnabled: true } : {}),
             session: { env },
         });
         return { provider: codexParticipant.provider, codexParticipant };
     }
-    if (deps.localAgent === "claude") {
+    if (localAgent === "claude") {
         // One session for the whole run, like the codex thread above (#520). The one-shot
         // provider (createLocalAgentProvider) spawned `claude -p` per turn, and every turn
         // started with no memory of the last. HUMANISH_LOCAL_AGENT_ONE_SHOT=1 keeps that path
@@ -58,7 +59,7 @@ export async function startParticipantModel(spec, deps, executor) {
                     ...(spec.planned.limits.reasoningEffort === undefined
                         ? {}
                         : { reasoningEffort: spec.planned.limits.reasoningEffort }),
-                    ...(config.actors[0]?.model === undefined ? {} : { model: config.actors[0].model }),
+                    ...(brain.declaredModel === undefined ? {} : { model: brain.declaredModel }),
                 }),
             };
         }
@@ -66,7 +67,7 @@ export async function startParticipantModel(spec, deps, executor) {
             ...(spec.planned.limits.reasoningEffort === undefined
                 ? {}
                 : { reasoningEffort: spec.planned.limits.reasoningEffort }),
-            ...(config.actors[0]?.model === undefined ? {} : { model: config.actors[0].model }),
+            ...(brain.declaredModel === undefined ? {} : { model: brain.declaredModel }),
         });
         return { provider: claudeSession.provider, claudeSession };
     }
@@ -80,7 +81,7 @@ export function participantSessionOptions(spec, deps, ready, provider, writeScre
     // injected pure per-turn estimator keyed on the resolved model. Preflight already refused a
     // cap on an unpriced model, so the estimate is measurable whenever a cap is in force. The
     // model id here matches provider.version (openai-responses-cu resolves the default when unset).
-    const capModelId = config.actors[0]?.model ?? DEFAULT_OPENAI_CU_MODEL;
+    const capModelId = pricedModel(deps.brain);
     const maxUsd = config.execution?.caps?.maxUsd;
     return {
         instructions: inbox
@@ -94,7 +95,7 @@ export function participantSessionOptions(spec, deps, ready, provider, writeScre
         ...(provider === undefined ? {} : { provider: provider }),
         openai: {
             apiKey: deps.openaiApiKey,
-            ...(config.actors[0]?.model ? { model: config.actors[0].model } : {}),
+            ...(deps.brain.declaredModel ? { model: deps.brain.declaredModel } : {}),
             // Per-LANE, not per-actor: two lanes at different efforts is the control this exists for.
             ...(spec.planned.limits.reasoningEffort === undefined
                 ? {}

package/dist/routes/computer-use/plan.js

@@ -6,10 +6,11 @@ import { resolveAutomaticAnalysis } from "../../analysis/automatic-config.js";
 import { actorRegistry, isCuaActorDescriptor, } from "../../actors/registry.js";
 import { isHttpUrl, isLoopbackUrl } from "../../lab/parse/subject.js";
 import { subjectStateInvalidReason } from "../../lab/parse/subject-state.js";
-import { brainOf, capsOf, desktopRequirements, planBase, provisionedSubject, } from "../../lab/plan-base.js";
+import { brainOf, callerBrainOf, capsOf, desktopRequirements, planBase, provisionedSubject, } from "../../lab/plan-base.js";
 import { computerUseParticipants, declaredTargets } from "../../lab/plan-participants.js";
 import { MAX_CUA_LANES } from "../../lab/routing.js";
 import { cloneTargetValidationReason, cuaLaneValidationReason, desktopMediaValidationReason, outputTokenLimitValidationReason, receivingEmailValidationReason, scenarioCapsValidationReason, taskProtocolValidationReason, } from "../../lab/validation.js";
+import { desktopCliProductReason } from "../../lab/composition-rules.js";
 import { isLocalBrowserLab } from "../../substrates/local/runtime-config.js";
 import { boundedConcurrency, defaultSessionTimeoutMs, resolveParticipantSandboxMs, } from "./participant-runs.js";
 import { MAX_SANDBOX_MS } from "../../substrates/e2b/lifetime.js";
@@ -236,9 +237,10 @@ export function planComputerUseLab(config, input) {
     if (hooks.buildExecutor !== undefined &&
         (source === "clone" || source === "local-tree" || source === "desktop-cli"))
         return refuse("in-scope", "HUMANISH_CUA_LAB_SUBJECT_INVALID", `cuaHooks.buildExecutor drives subject.appUrl in this process, and a ${source} subject needs the hosted desktop the in-process route never creates. Use an app-url or local-app subject with buildExecutor, or remove buildExecutor to run on a hosted desktop.`, actor);
-    // The parser refuses this too; without a product the desktop study fails later.
-    if (source === "desktop-cli" && config.subject.product?.name === undefined)
-        return refuse("in-scope", "HUMANISH_CUA_LAB_SUBJECT_INVALID", "desktop-cli subjects need `subject.product.name` — the CLI the participant is being asked to use.", actor);
+    // The parser's rule; without a product the desktop study fails later.
+    const productReason = desktopCliProductReason(config);
+    if (productReason)
+        return refuse("in-scope", "HUMANISH_CUA_LAB_SUBJECT_INVALID", productReason, actor);
     const participants = computerUseParticipants(config, input.countOverride);
     if (participants.length > MAX_CUA_LANES)
         return refuse("after-personas", "HUMANISH_CUA_LAB_FANOUT_INVALID", `Computer-use fan-out is capped at ${MAX_CUA_LANES} participants (resolved ${participants.length}); N concurrent paid desktops is real spend.`, actor);
@@ -261,7 +263,7 @@ export function planComputerUseLab(config, input) {
     const runner = hooks.buildExecutor !== undefined
         ? {
             desktop: "in-process",
-            brain: { kind: "caller" },
+            brain: callerBrainOf(config),
             participants: [first],
             subject: source === "local-app" ? { kind: "local-app", appUrl } : appUrlSubject,
         }

package/dist/routes/computer-use/preflight.d.ts

@@ -1,6 +1,5 @@
-import type { ComputerUsePlan } from "../../lab/plan-types.js";
+import type { Brain, ComputerUsePlan } from "../../lab/plan-types.js";
 import type { LabCommsExternal } from "../../lab/types.js";
-import { type LocalAgentId } from "../../actors/local-agent/cli.js";
 import type { CuaActorLabErrorCode, CuaActorLabHooks } from "./types.js";
 /**
  * The first reason a live computer-use run cannot start on this machine: missing keys, a missing
@@ -10,14 +9,12 @@ import type { CuaActorLabErrorCode, CuaActorLabHooks } from "./types.js";
  */
 export declare function liveCuaRejection(args: {
     caps: ComputerUsePlan["caps"];
-    /** The declared participant model; a spend cap needs its price. */
-    model: string | undefined;
+    /** The plan's brain: whether a local agent drives the participant, and the model a cap prices. */
+    brain: Brain;
     hooks: CuaActorLabHooks;
     env: Record<string, string | undefined>;
     openaiApiKey: string;
     e2bApiKey: string;
-    localAgentRoute: boolean;
-    preferredLocalAgent: LocalAgentId;
     subjectEnvNames: string[];
     externalCommsConfig: LabCommsExternal | undefined;
 }): Promise<{

package/dist/routes/computer-use/preflight.js

@@ -1,4 +1,4 @@
-import { DEFAULT_OPENAI_CU_MODEL } from "../../actors/computer-use/openai-provider.js";
+import { pricedModel } from "../../lab/plan-base.js";
 import { detectLocalAgents } from "../../actors/local-agent/cli.js";
 import { localAgentRefusal } from "../../actors/local-agent/readiness.js";
 import { describeMissingKeys } from "../../keys/key-resolution.js";
@@ -19,9 +19,12 @@ const LOCAL_AGENT_REFUSAL_CODES = {
  * already paid for a desktop.
  */
 export async function liveCuaRejection(args) {
-    const { caps, model, hooks, env, openaiApiKey, e2bApiKey, localAgentRoute, preferredLocalAgent, subjectEnvNames, externalCommsConfig, } = args;
+    const { caps, brain, hooks, env, openaiApiKey, e2bApiKey, subjectEnvNames, externalCommsConfig } = args;
+    // The operator's own signed-in coding agent is the brain, so there is no provider key to ask
+    // for. E2B is still required: the persona needs a machine.
+    const localAgent = brain.kind === "local-agent" ? brain.agent : undefined;
     const missingKeys = [
-        ...(openaiApiKey || localAgentRoute || hooks.buildProvider ? [] : ["OPENAI_API_KEY"]),
+        ...(openaiApiKey || localAgent || hooks.buildProvider ? [] : ["OPENAI_API_KEY"]),
         ...(e2bApiKey || participantDesktopOf(hooks) !== undefined ? [] : ["E2B_API_KEY"]),
     ];
     if (missingKeys.length > 0) {
@@ -42,10 +45,10 @@ export async function liveCuaRejection(args) {
             message: `Live computer-use labs need ${missingKeys.join(" and ")} in the environment (values are never persisted). ${describeMissingKeys(missingKeys, env)}${suggestion}`,
         };
     }
-    if (localAgentRoute && !hooks.buildProvider) {
+    if (localAgent && !hooks.buildProvider) {
         // Refuse HERE, before a sandbox exists. "codex is not installed" discovered after the
         // machine is paid for is the same information delivered at the worst possible moment.
-        const refusal = await localAgentRefusal({ agent: preferredLocalAgent, env, caps });
+        const refusal = await localAgentRefusal({ agent: localAgent, env, caps });
         if (refusal)
             return { code: LOCAL_AGENT_REFUSAL_CODES[refusal.kind], message: refusal.message };
     }
@@ -63,11 +66,12 @@ export async function liveCuaRejection(args) {
     // uncapped: an unenforceable cap is more dangerous than none. The operator adds a rate to
     // src/run/pricing.ts (the honest place) or removes the cap.
     if (caps.maxUsd !== undefined || caps.maxTotalUsd !== undefined) {
-        const capModelId = (model ?? DEFAULT_OPENAI_CU_MODEL).trim().toLowerCase();
+        const model = pricedModel(brain);
+        const capModelId = model.trim().toLowerCase();
         if (!MODEL_RATES[capModelId]) {
             return {
                 code: "HUMANISH_CUA_LAB_UNPRICED_CAP",
-                message: `execution.caps declares a spend cap (maxUsd/maxTotalUsd) but src/run/pricing.ts has no rate for model "${model ?? DEFAULT_OPENAI_CU_MODEL}"; add a rate or remove the cap — an unenforceable cap is refused rather than run uncapped.`,
+                message: `execution.caps declares a spend cap (maxUsd/maxTotalUsd) but src/run/pricing.ts has no rate for model "${model}"; add a rate or remove the cap — an unenforceable cap is refused rather than run uncapped.`,
             };
         }
     }

package/dist/routes/computer-use/route.d.ts

@@ -26,8 +26,8 @@ export declare function admitComputerUsePlan(plan: ComputerUsePlan, input: Compu
  * Run a computer-use plan: its local checks, then the run. The run scope in runAdmittedCuaRun
  * finalizes any status record the run opened, on every exit, so a test or library caller does not
  * leave the 5 s status cadence writing into a directory something else is deleting (an unrelated
- * ENOTEMPTY). Participants come from the plan. `config` is still read for values the plan does not
- * carry yet (the actor's model and local agent, persona ids, comms settings, the per-participant
- * cap) and by the compatibility hooks, which take the whole config.
+ * ENOTEMPTY). Participants and the brain (model and local agent) come from the plan. `config` is
+ * still read for values the plan does not carry yet (persona ids, comms settings, the
+ * per-participant cap) and by the compatibility hooks, which take the whole config.
  */
 export declare function runComputerUsePlan(plan: ComputerUsePlan, input: ComputerUseRunInput, config: LabConfig): Promise<CuaActorLabResult>;

package/dist/routes/computer-use/route.js

@@ -107,9 +107,9 @@ function cuaOutcome(result) {
  * Run a computer-use plan: its local checks, then the run. The run scope in runAdmittedCuaRun
  * finalizes any status record the run opened, on every exit, so a test or library caller does not
  * leave the 5 s status cadence writing into a directory something else is deleting (an unrelated
- * ENOTEMPTY). Participants come from the plan. `config` is still read for values the plan does not
- * carry yet (the actor's model and local agent, persona ids, comms settings, the per-participant
- * cap) and by the compatibility hooks, which take the whole config.
+ * ENOTEMPTY). Participants and the brain (model and local agent) come from the plan. `config` is
+ * still read for values the plan does not carry yet (persona ids, comms settings, the
+ * per-participant cap) and by the compatibility hooks, which take the whole config.
  */
 export async function runComputerUsePlan(plan, input, config) {
     const admission = await admitCuaRun(plan, input, config);

package/dist/routes/computer-use/run-lanes.js

@@ -2,7 +2,7 @@
 // receiving, the lanes themselves, and the drain of an operator-hosted inbox.
 import { prepareReceivingRun } from "../../comms/receiving-runtime.js";
 import { liveObserverResult } from "../../observer/live.js";
-import { DEFAULT_OPENAI_CU_MODEL } from "../../actors/computer-use/openai-provider.js";
+import { pricedModel } from "../../lab/plan-base.js";
 import { runAllCuaParticipants } from "./lanes.js";
 import { startLiveTraceFlush } from "./live-flush.js";
 import { drainExternalComms } from "./external-comms.js";
@@ -13,7 +13,7 @@ import {} from "./types.js";
  * the participant outcomes and the warnings the finish records.
  */
 export async function runLabParticipants(setup, participants) {
-    const { plan, input, config, cwd, streams, descriptor, subjectRoute, run } = setup;
+    const { plan, input, cwd, streams, descriptor, subjectRoute, run } = setup;
     const { participantRuns, participantPlan, scrubKnownValues, bundleBase } = setup;
     const { env, knownSecretValues, deps, liveTrace, externalComms, inProgress, fail } = participants;
     const { dryRun } = plan;
@@ -53,7 +53,7 @@ export async function runLabParticipants(setup, participants) {
         const liveFlush = startLiveTraceFlush({
             bundle: inProgressBundle,
             participantRuns: participantRuns,
-            model: config.actors[0]?.model ?? DEFAULT_OPENAI_CU_MODEL,
+            model: pricedModel(plan.runner.brain),
             write: (bundle) => run.writeSnapshot(bundle),
         });
         liveTrace.flush = liveFlush.flush;

package/dist/routes/computer-use/setup.d.ts

@@ -1,5 +1,4 @@
 import { type RunScope } from "../../run/run.js";
-import { type LocalAgentId } from "../../actors/local-agent/cli.js";
 import type { CuaActorDescriptor } from "../../actors/registry.js";
 import type { LabCommsEmail, LabCommsExternal, LabConfig } from "../../lab/types.js";
 import { type RunRerunLineage } from "../../run/bundle.js";
@@ -125,8 +124,6 @@ export declare function admitCuaRun(plan: ComputerUsePlan, input: ComputerUseRun
         scrubKnownValues: (text: string) => string;
         publicRepo: string | undefined;
         hasGithubToken: boolean;
-        localAgentRoute: boolean;
-        preferredLocalAgent: LocalAgentId;
         localTreeArchive: LocalTreeArchive | undefined;
         localTreeArchiveBuffer: ArrayBuffer | undefined;
     };

package/dist/routes/computer-use/setup.js

@@ -6,7 +6,6 @@ import { realpath } from "node:fs/promises";
 import path from "node:path";
 import {} from "../../run/run.js";
 import { externalInboxUrl } from "../../comms/sandbox-catch.js";
-import {} from "../../actors/local-agent/cli.js";
 import { redactText, scrubLiterals, toErrorMessage } from "../../evidence/redaction.js";
 import { buildRunSource } from "../../run/bundle.js";
 import { assertPreparedSelectedOutputDirectory, prepareContainedOutputDirectory, prepareSelectedOutputDirectory, } from "../../run/contained-output.js";
@@ -127,25 +126,16 @@ export async function admitCuaRun(plan, input, config) {
     const redactRepoLabel = plan.residual.policies?.redactRepos ?? subjectEnvNames.includes("GITHUB_TOKEN");
     const publicRepo = cloneRoute && subjectRepo ? (redactRepoLabel ? "repo-01" : subjectRepo) : undefined;
     const hasGithubToken = subjectEnvNames.includes("GITHUB_TOKEN");
-    // The operator's own signed-in coding agent is the brain, so there is no provider key to ask
-    // for — the entire point of the actor. E2B is still required: the persona needs a machine.
-    const localAgentRoute = descriptor.id === "local-agent";
-    // Which local CLI, from its OWN field: `model` means the model, so that "Claude Code running
-    // Opus" is sayable. Preflight below refuses when the chosen one is missing or signed out — that
-    // news is worthless after a sandbox is paid for.
-    const preferredLocalAgent = config.actors[0]?.localAgent ?? "codex";
     // Key-gating is route-aware: the in-process route uses the caller's OWN model + executor, and
     // the local-agent route uses a CLI the operator has already signed in to.
     if (!dryRun && !inProcessRoute) {
         const rejection = await liveCuaRejection({
             caps: plan.caps,
-            model: config.actors[0]?.model,
+            brain: plan.runner.brain,
             hooks,
             env,
             openaiApiKey,
             e2bApiKey,
-            localAgentRoute,
-            preferredLocalAgent,
             subjectEnvNames,
             externalCommsConfig,
         });
@@ -195,8 +185,6 @@ export async function admitCuaRun(plan, input, config) {
             scrubKnownValues,
             publicRepo,
             hasGithubToken,
-            localAgentRoute,
-            preferredLocalAgent,
             localTreeArchive,
             localTreeArchiveBuffer,
         },
@@ -312,7 +300,7 @@ export async function startCuaRun(plan, input, admitted, scope) {
 /** The lane deps every lane reads: the route, keys, timeouts, scrubber, budget and hooks. */
 function cuaParticipantDeps(plan, input, admitted, run) {
     const { config, dryRun, hooks, streams, env, descriptor, runSession, participantCount } = admitted;
-    const { localAgentRoute, preferredLocalAgent, hasGithubToken, localTreeArchiveBuffer } = admitted;
+    const { hasGithubToken, localTreeArchiveBuffer } = admitted;
     const { openaiApiKey, e2bApiKey, scrubKnownValues } = admitted;
     const { externalCommsConfig, externalCommsEmail } = admitted;
     const { appUrl, cloneRoute, desktopCliRoute, localTreeRoute, serve, subjectRepo } = admitted.subjectRoute;
@@ -327,7 +315,7 @@ function cuaParticipantDeps(plan, input, admitted, run) {
         config,
         descriptor,
         appUrl,
-        ...(localAgentRoute ? { localAgent: preferredLocalAgent } : {}),
+        brain: plan.runner.brain,
         cloneRoute,
         desktopCliRoute,
         localTreeRoute,

package/dist/routes/computer-use/types.d.ts

@@ -14,9 +14,8 @@ import type { CuaActorSessionOptions } from "../../actors/computer-use/actor.js"
 import { type E2BDesktopModule, type E2BDesktopSandbox } from "../../substrates/e2b/sdk.js";
 import { type DesktopResourceObservation } from "../../substrates/e2b/desktop-resources.js";
 import { type DetachedTimers } from "../../substrates/detached.js";
-import type { ComputerUsePlan } from "../../lab/plan-types.js";
+import type { Brain, ComputerUsePlan } from "../../lab/plan-types.js";
 import { type LabCommsEmail, type LabConfig, type LabSubjectServe } from "../../lab/types.js";
-import { type LocalAgentId } from "../../actors/local-agent/cli.js";
 import { renderObserver, type ObserverResult } from "../../observer/render.js";
 import { type RunLabProvenance } from "../../run/status.js";
 import { type RunBundle, type RunRerunLineage, type RunScorerProvenance, type RunSubjectProvenance, type RunSubjectStateStepRecord } from "../../run/bundle.js";
@@ -398,8 +397,8 @@ export interface CuaParticipantDeps {
     config: LabConfig;
     descriptor: CuaActorDescriptor;
     appUrl: string;
-    /** When set, the computer-use brain is this locally-signed-in CLI instead of a keyed API. */
-    localAgent?: LocalAgentId;
+    /** The plan's brain: the model and, for a local agent, which signed-in CLI drives the participant. */
+    brain: Brain;
     cloneRoute: boolean;
     /** desktop-cli (#495): a CLI studied at a desktop. Nothing is cloned and no browser is opened. */
     desktopCliRoute?: boolean;

package/dist/routes/computer-use/types.js

@@ -7,7 +7,6 @@ import {} from "../../substrates/e2b/sdk.js";
 import {} from "../../substrates/e2b/desktop-resources.js";
 import {} from "../../substrates/detached.js";
 import {} from "../../lab/types.js";
-import {} from "../../actors/local-agent/cli.js";
 import { renderObserver } from "../../observer/render.js";
 import {} from "../../run/status.js";
 import {} from "../../run/bundle.js";

package/dist/routes/scripted/subject.js

@@ -137,7 +137,8 @@ export class ScriptedSubject {
         this.killed = reading.released;
         if (reading.warning)
             warnings.push(reading.warning);
-        // Without a kill method nothing ran, so there is no teardown time to record.
+        // Without a kill method, or in E2B debug mode, no kill reached E2B, so there is no teardown
+        // time to record.
         if (released.status !== "unconfirmed" || released.reason !== "release_unavailable")
             this.tornDownAtMs = now();
     }

package/dist/routes/shared-world/provisioned.js

@@ -345,7 +345,8 @@ class SubjectPlane {
             this.subjectKilled = reading.released;
             if (reading.warning)
                 warnings.push(reading.warning);
-            // Without a kill method nothing ran, so there is no teardown time to record.
+            // Without a kill method, or in E2B debug mode, no kill reached E2B, so there is no teardown
+            // time to record.
             if (released.status !== "unconfirmed" || released.reason !== "release_unavailable")
                 this.subjectTornDownAtMs = this.ctx.now();
         }

package/dist/routes/shared-world/seats.js

@@ -1,6 +1,6 @@
 // Builds each participant's actor spec and mission, resolves its entry URL, derives the session
 // and sandbox time budgets, and records a follower that never received the host's lobby code.
-import { DEFAULT_OPENAI_CU_MODEL } from "../../actors/computer-use/openai-provider.js";
+import { pricedModel } from "../../lab/plan-base.js";
 import { scrubPersonaBrief } from "../../lab/persona.js";
 import { resolveParticipant } from "../../run/participant.js";
 import { attachObserverRuntimeStreamUrls } from "../../observer/render.js";
@@ -111,7 +111,7 @@ export function startParticipantFlush(ctx, live, bundle) {
     live.flush = startLiveTraceFlush({
         bundle,
         participantRuns: ctx.actorSpecs,
-        model: ctx.config.actors[0]?.model ?? DEFAULT_OPENAI_CU_MODEL,
+        model: pricedModel(ctx.plan.brain),
         write: (snapshot) => ctx.run.writeSnapshot(snapshot),
     });
 }
@@ -176,7 +176,7 @@ export function participantRunDeps(ctx, live, scrubKnownValues) {
         scrubKnownValues,
         runSession: ctx.runSession,
         // A local-agent brain runs each seat on the operator's signed-in agent, as on computer use.
-        ...(ctx.plan.brain.kind === "local-agent" ? { localAgent: ctx.plan.brain.agent } : {}),
+        brain: ctx.plan.brain,
         ...(receiving ? { receiving } : {}),
         now: ctx.now,
         hooks: runtimeStreamHooks(ctx.hooks, live),

package/dist/routes/terminal/live-sandbox.d.ts

@@ -47,6 +47,7 @@ export declare class LiveTerminalSandbox {
     private module;
     private allocation;
     private startupCleanup;
+    private startupCleanupDetail;
     private createdAtMs;
     private tornDownAtMs;
     private resources;

package/dist/routes/terminal/live-sandbox.js

@@ -38,6 +38,7 @@ export class LiveTerminalSandbox {
     module;
     allocation;
     startupCleanup;
+    startupCleanupDetail;
     // The sandbox's billed span (acquired to cleanup) and size price its compute time.
     createdAtMs;
     tornDownAtMs;
@@ -389,8 +390,10 @@ export class LiveTerminalSandbox {
     recordSessionError(error) {
         const { sanitize } = this.inputs;
         const { recordLifecycle } = this.inputs.recorder;
-        if (error instanceof E2BDesktopStartupError)
+        if (error instanceof E2BDesktopStartupError) {
             this.startupCleanup = error.cleanup;
+            this.startupCleanupDetail = error.cleanupDetail;
+        }
         this.error = sanitize(toErrorMessage(error));
         this.status = "failed";
         this.completionReason = "harness_error";
@@ -406,6 +409,7 @@ export class LiveTerminalSandbox {
             allocation: this.allocation,
             sandboxModule: this.module,
             ...(this.startupCleanup === undefined ? {} : { startupCleanup: this.startupCleanup }),
+            startupCleanupDetail: this.startupCleanupDetail,
             requestTimeoutMs,
             sanitize,
             recordLifecycle,

package/dist/routes/terminal/sandbox.d.ts

@@ -17,6 +17,8 @@ export declare function teardownSandbox(args: {
     /** For the Sandbox.getInfo(id) re-check only. */
     sandboxModule: E2BDesktopModule | undefined;
     startupCleanup?: E2BDesktopStartupError["cleanup"];
+    /** Why the startup guard's cleanup is unconfirmed, when it knows. */
+    startupCleanupDetail?: string | undefined;
     requestTimeoutMs: number;
     sanitize: (text: string) => string;
     recordLifecycle: (event: string, message: string) => void;

package/dist/routes/terminal/sandbox.js

@@ -1,5 +1,6 @@
 import { toErrorMessage } from "../../evidence/redaction.js";
 import { E2BDesktopStartupError, isSandboxNotFoundError, } from "../../substrates/e2b/sdk.js";
+import { releaseUnavailableDetail } from "../../substrates/e2b/sandbox.js";
 /**
  * Tear the sandbox down and PROVE it BY EXACT ID -- NEVER Sandbox.list (humanish must never
  * enumerate the operator's E2B account; see docs/principles/invariants-and-defaults.md). The
@@ -11,7 +12,7 @@ import { E2BDesktopStartupError, isSandboxNotFoundError, } from "../../substrate
  * fails closed on an unproven teardown.
  */
 export async function teardownSandbox(args) {
-    const { allocation, sandboxModule, startupCleanup, requestTimeoutMs, sanitize, recordLifecycle, warnings, } = args;
+    const { allocation, sandboxModule, startupCleanup, startupCleanupDetail, requestTimeoutMs, sanitize, recordLifecycle, warnings, } = args;
     if (allocation === undefined || !sandboxModule) {
         // create() can reject AFTER its constructor acquired a handle. The default loader retains
         // that authority and reclaims it before rejecting; the lane itself never receives its ID.
@@ -21,7 +22,7 @@ export async function teardownSandbox(args) {
             return { killed: true, remaining: 0, reason };
         }
         const reason = startupCleanup === "unconfirmed"
-            ? "desktop startup guard could not confirm cleanup of its acquired sandbox; provider timeout remains the backstop"
+            ? `desktop startup guard could not confirm cleanup of its acquired sandbox${startupCleanupDetail === undefined ? "" : ` (${startupCleanupDetail})`}; provider timeout remains the backstop`
             : "create did not return a sandbox; the participant has no acquired handle and cannot establish allocation or cleanup";
         recordLifecycle("terminal-lab.cleanup.unconfirmed", reason);
         return { killed: false, remaining: -1, reason };
@@ -30,11 +31,7 @@ export async function teardownSandbox(args) {
     const released = await allocation.close();
     if (released.status !== "released") {
         if (released.status === "unconfirmed" && released.reason === "release_unavailable") {
-            return {
-                killed: false,
-                remaining: -1,
-                reason: "installed @e2b/desktop SDK does not expose Sandbox.kill; server-side kill-on-timeout will reclaim the sandbox",
-            };
+            return { killed: false, remaining: -1, reason: releaseUnavailableDetail(released.detail) };
         }
         if (released.status === "unconfirmed" && released.reason === "release_failed") {
             const sanitizedError = sanitize(toErrorMessage(released.error));

package/dist/run/reclaim.d.ts

@@ -21,7 +21,7 @@ export interface ReclaimResult {
     outcomes: ReclaimOutcome[];
     warnings: string[];
     error?: {
-        code: "HUMANISH_RECLAIM_RUN_NOT_FOUND" | "HUMANISH_RECLAIM_MODULE_UNAVAILABLE" | "HUMANISH_RECLAIM_RECEIPTS_UNREADABLE";
+        code: "HUMANISH_RECLAIM_RUN_NOT_FOUND" | "HUMANISH_RECLAIM_E2B_DEBUG" | "HUMANISH_RECLAIM_MODULE_UNAVAILABLE" | "HUMANISH_RECLAIM_RECEIPTS_UNREADABLE";
         message: string;
     };
 }

package/dist/run/reclaim.js

@@ -3,7 +3,7 @@
 // the managed run dir, kills by id, and never lists the account: an account-wide operation once
 // destroyed unrelated infrastructure. Without it, orphaned sandboxes run until the server-side
 // create-time TTL.
-import { loadE2BDesktopModule } from "../substrates/e2b/sdk.js";
+import { E2B_DEBUG_ENV, e2bDebugMode, loadE2BDesktopModule, } from "../substrates/e2b/sdk.js";
 import { containedPathAbsent, readContainedRegularFile, writeContainedOutputFile, } from "./contained-output.js";
 import { discardPreflightJournal, listPreflightJournals, preflightReclaimDecision, } from "./preflight-receipts.js";
 import { resolveRunPath } from "./locate.js";
@@ -23,6 +23,8 @@ export async function reclaimRunSandboxes(cwd, runInput, hooks = {}) {
         outcomes: [],
         warnings,
     };
+    if (e2bDebugMode())
+        return { ...base, ok: false, error: E2B_DEBUG_REFUSAL };
     const runPaths = await resolveRunPath(cwd, runInput);
     if (!runPaths) {
         return {
@@ -79,6 +81,8 @@ export async function reclaimPreflightSandboxes(cwd, hooks = {}) {
     let receiptCount = 0;
     let unreadable = false;
     const base = { schema: RECLAIM_RESULT_SCHEMA, cwd, runId: "preflight", warnings };
+    if (e2bDebugMode())
+        return { ...base, ok: false, receiptCount, outcomes, error: E2B_DEBUG_REFUSAL };
     const journals = await listPreflightJournals(cwd);
     if (journals.length === 0) {
         warnings.push("No preflight journals in .humanish/preflight: every probe's teardown was confirmed, or no probe ran here. Nothing to reclaim by id.");
@@ -120,6 +124,12 @@ export async function reclaimPreflightSandboxes(cwd, hooks = {}) {
     }
     return { ...base, ok: !unreadable && allGone(outcomes), receiptCount, outcomes };
 }
+// In debug mode the SDK returns true from Sandbox.kill without contacting E2B, so every receipt
+// would read killed and preflight journals would be discarded for sandboxes that may still run.
+const E2B_DEBUG_REFUSAL = {
+    code: "HUMANISH_RECLAIM_E2B_DEBUG",
+    message: `${E2B_DEBUG_ENV}=true puts the E2B SDK in debug mode, where Sandbox.kill reports success without contacting E2B. Nothing was killed and the receipts were kept; unset ${E2B_DEBUG_ENV} and run reclaim again.`,
+};
 const UNREADABLE_MESSAGE = "sandbox-receipts.ndjson is present but could not be read safely (not a single regular file, or a read error). Nothing was killed and the receipts were kept; check the file, then run reclaim again.";
 /** Kill every sandbox one receipts journal records, once per sandbox, by exact id. */
 async function reclaimJournal(root, hooks) {

package/dist/substrates/desktop-session.d.ts

@@ -10,6 +10,8 @@ export type DesktopReleaseResult = {
     status: "unconfirmed";
     reason: "release_unavailable" | "invalid_result" | "release_failed";
     error?: unknown;
+    /** Why release is unavailable, when it is not the missing kill method. */
+    detail?: string;
 };
 export interface DesktopSession {
     readonly resourceId: string;

package/dist/substrates/e2b/sandbox.d.ts

@@ -48,6 +48,8 @@ export declare function readE2BRelease(result: DesktopReleaseResult, options: {
     scrub: (text: string) => string;
     costSpan?: boolean;
 }): E2BReleaseReading;
+/** A release_unavailable result as one lowercase line, for routes and reclaim that report it. */
+export declare function releaseUnavailableDetail(detail: string | undefined): string;
 /** Reclaim's persisted outcome; run/reclaim.ts writes these strings. */
 export type E2BSandboxDestroyOutcome = {
     state: "killed" | "already-gone";