faberun 0.25.1 → 0.26.1

package/src/engine/failover.mjs

@@ -17,6 +17,7 @@
 import { harnessCapabilities } from "../harnesses/index.mjs";
 import { isRuntimeAvailable, nextSameTierRuntime } from "./runtime-discovery.mjs";
 import { routeRuntime } from "../contract/runtime.mjs";
+import { effectiveProvider } from "../contract/provider.mjs";
 
 /** @typedef {import("../contract/index.mjs").ValidatedNode} ValidatedNode */
 /** @typedef {import("../contract/index.mjs").NodeSnapshot} NodeSnapshot */
@@ -188,11 +189,17 @@ export function routingBackoffActive(state, phase) {
  * survives both a routed hop and the revision boundary: `resetPhaseRouting`
  * clears the routing override between revisions, but never the invocations.
  *
+ * Exported so a same-provider-review mark (R20) can be taken from the worker
+ * invocation that actually ran instead of `state.runtime`, which `startJudge`
+ * overwrites with the judge's own runtime as soon as it dispatches -- a re-ask
+ * or a judge-failure retry then calls `startJudge` again with the judge still
+ * in `state.runtime`, so comparing against it compares the judge to itself.
+ *
  * @param {NodeSnapshot} state
  * @param {"worker"|"judge"} role
  * @returns {string|undefined}
  */
-function previousAttemptRuntimeId(state, role) {
+export function previousAttemptRuntimeId(state, role) {
   const found = [...(state.invocations ?? [])].reverse().find((invocation) => invocation.phase === role)?.runtimeId;
   return typeof found === "string" ? found : undefined;
 }
@@ -223,7 +230,8 @@ function admitsAffinity(contract, state, role, candidate) {
   if (availability && !isRuntimeAvailable(availability)) return false;
   if (role !== "judge") return true;
   const workerId = previousAttemptRuntimeId(state, "worker") ?? state.routing?.assignments?.worker;
-  return runtime.vendor !== (workerId ? contract.runtimes[workerId]?.vendor : null);
+  const workerRuntime = workerId ? contract.runtimes[workerId] : undefined;
+  return effectiveProvider(runtime) !== (workerRuntime ? effectiveProvider(workerRuntime) : null);
 }
 
 /**

package/src/engine/judge-gate.mjs

@@ -142,8 +142,9 @@ function proveVerification(id, proof, recorded) {
  * `--test-skip-pattern`) is judged on more than its exit code, because the
  * runner exits 0 whether its filter selected anything or not. The proof runs
  * with the TAP reporter selected through `NODE_OPTIONS` and is refused when the
- * output carries TAP's zero-plan line. Measured 2026-09-21 on node v26.8.1: a
- * filter that matches emits `1..0` zero times, a filter that matches nothing
+ * output says the filter selected nothing (`tapSelectedNothing`). Measured
+ * 2026-09-21 on node v26.8.1, in a single-file run: a filter that matches
+ * emits `1..0` zero times, a filter that matches nothing
  * emits it exactly once, a run with no filter emits it zero times, and an empty
  * suite under a matching filter emits no nested zero plan. The default reporter
  * cannot make the distinction -- both cases print identical counters, because
@@ -154,7 +155,26 @@ function proveVerification(id, proof, recorded) {
  * never sees it.
  */
 const TEST_FILTER_FLAGS = ["--test-name-pattern", "--test-skip-pattern"];
-const TAP_ZERO_PLAN = /^1\.\.0$/mu;
+const TAP_ZERO_PLAN = /^1\.\.0$/gmu;
+const TAP_TOP_LEVEL_RESULT = /^(?:not )?ok \d+ /gmu;
+
+/**
+ * Whether a filtered run's TAP says its filter selected no test. A run of
+ * several files prints, for each file the filter selected nothing in, an
+ * unindented `1..0` followed by that file's own top-level `ok` line, while a
+ * selected test prints its own top-level result with no zero plan in front of
+ * it. Measured 2026-09-26 on node v26.8.1 (AP1 of safe-to-hand-to-a-friend):
+ * reading a single `1..0` as "nothing selected" refused two proofs whose
+ * filter had selected a passing test in one of several files. Nothing was
+ * selected exactly when every top-level result is one of those file lines.
+ *
+ * @param {string} stdout
+ * @returns {boolean}
+ */
+function tapSelectedNothing(stdout) {
+  const zeroPlans = stdout.match(TAP_ZERO_PLAN)?.length ?? 0;
+  return zeroPlans > 0 && zeroPlans >= (stdout.match(TAP_TOP_LEVEL_RESULT)?.length ?? 0);
+}
 
 /**
  * The environment a filtered proof runs under: the ambient environment with
@@ -252,7 +272,7 @@ async function proveCommand(id, proof, cwd, timeoutMs) {
     child.on("close", (code, signal) => {
       // Its own detail, not an ordinary command failure: the command succeeded,
       // so what failed is that the declared filter selected nothing to prove.
-      if (code === 0 && signal === null && filters.length > 0 && TAP_ZERO_PLAN.test(stdout)) {
+      if (code === 0 && signal === null && filters.length > 0 && tapSelectedNothing(stdout)) {
         finish({
           id,
           kind: "command",

package/src/engine/judge-list.mjs

@@ -0,0 +1,206 @@
+/**
+ * R18: the judge of a gated node is the first eligible entry of a static,
+ * ordered list rather than the single strongest cross-vendor candidate
+ * `composeAssignments` otherwise ranks. Split out of `runtime-discovery.mjs`
+ * and `lifecycle.mjs` -- both already near this repository's 800-line
+ * ceiling -- and kept out of `backoff.mjs`, whose `planRoute` is a pure
+ * decision this module cannot be: picking an entry reads the durable refusal
+ * and usage-window stores.
+ */
+import { effectiveProvider } from "../contract/provider.mjs";
+import { availabilityKey, readRefusal } from "../run/availability.mjs";
+import { readUsageWindows, usageAccountOf } from "../run/usage-windows.mjs";
+import { getHarness } from "../harnesses/index.mjs";
+
+/** @typedef {import("../contract/index.mjs").ValidatedContract} ValidatedContract */
+/** @typedef {{id: string, reason: string}} JudgeSkip */
+/** @typedef {{chosen: string|null, skipped: JudgeSkip[]}} JudgeListPick */
+/** @typedef {{list: string[], chosen: string|null, skipped: JudgeSkip[]}} JudgeListState */
+
+/**
+ * A usage window at or above this spend is treated as unavailable for a judge
+ * pick, so a node's review does not race the account into a refusal the next
+ * call would hit anyway.
+ */
+export const JUDGE_USAGE_WINDOW_LIMIT_PERCENT = 90;
+
+/**
+ * The static judge list a node's assignment reads: the contract's own list
+ * wins over the machine default, exactly as `runtimeDefaults.worker` already
+ * wins over `config.worker` in `composeAssignments`.
+ *
+ * @param {{judges?: string[]}} contract
+ * @param {{judges?: string[]}|null|undefined} config
+ * @returns {string[]|undefined}
+ */
+export function resolveJudgeList(contract, config) {
+  return contract.judges ?? config?.judges ?? undefined;
+}
+
+/**
+ * Every provider a worker attempt on this node could actually run on: its own
+ * canonical provider, plus every hop of its declared fallback chain. A judge
+ * of any of these providers cannot independently review whichever one
+ * actually ran. Walked directly off `runtime.fallback` rather than through
+ * `failover.mjs`'s `synthesizedChain`, so this module does not import a
+ * runtime-discovery consumer and risk a cycle back through
+ * `run/availability.mjs`.
+ *
+ * @param {ValidatedContract} contract
+ * @param {string} workerId
+ * @returns {Set<string|undefined>}
+ */
+function reachableWorkerProviders(contract, workerId) {
+  const providers = new Set([effectiveProvider(contract.runtimes[workerId])]);
+  const seen = new Set([workerId]);
+  let current = workerId;
+  for (;;) {
+    const next = contract.runtimes[current]?.fallback;
+    if (!next || seen.has(next) || !contract.runtimes[next]) break;
+    seen.add(next);
+    providers.add(effectiveProvider(contract.runtimes[next]));
+    current = next;
+  }
+  return providers;
+}
+
+/**
+ * Is this runtime's account inside `JUDGE_USAGE_WINDOW_LIMIT_PERCENT` of a
+ * usage window closing? `null` means no window is known -- unknown must not
+ * read as over budget, exactly as an absent availability record must not read
+ * as exhausted (see `admitsAffinity` in `failover.mjs`).
+ *
+ * @param {import("../contract/index.mjs").ValidatedRuntime} runtime
+ * @param {number} now
+ * @returns {import("../harnesses/codex/usage-window.mjs").UsageWindow|null}
+ */
+function overspentWindow(runtime, now) {
+  const account = usageAccountOf(runtime);
+  if (!account) return null;
+  return readUsageWindows(account, now).find((window) => window.usedPercent > JUDGE_USAGE_WINDOW_LIMIT_PERCENT) ?? null;
+}
+
+/**
+ * The first eligible entry of an ordered judge list, and why every earlier
+ * entry was skipped: already attempted this run, a provider the worker (or
+ * its fallback chain) already runs, a recorded refusal, or an account usage
+ * window over `JUDGE_USAGE_WINDOW_LIMIT_PERCENT`.
+ *
+ * @param {ValidatedContract} contract
+ * @param {string[]} list
+ * @param {string} workerId
+ * @param {{attempted?: Iterable<string>, now?: number}} [options]
+ * @returns {JudgeListPick}
+ */
+export function selectListJudge(contract, list, workerId, options = {}) {
+  const attempted = new Set(options.attempted ?? []);
+  const now = options.now ?? Date.now();
+  const workerProviders = reachableWorkerProviders(contract, workerId);
+  /** @type {JudgeSkip[]} */
+  const skipped = [];
+  for (const id of list) {
+    if (attempted.has(id)) {
+      skipped.push({ id, reason: "already attempted this run" });
+      continue;
+    }
+    const runtime = contract.runtimes[id];
+    if (!runtime) {
+      skipped.push({ id, reason: "not declared in contract.runtimes" });
+      continue;
+    }
+    const provider = effectiveProvider(runtime);
+    if (workerProviders.has(provider)) {
+      skipped.push({ id, reason: `same provider as the worker: ${provider}` });
+      continue;
+    }
+    const key = availabilityKey({ harness: runtime.harness, model: runtime.model, executable: getHarness(runtime.harness).executable(runtime) });
+    const refusal = readRefusal(key, now);
+    if (refusal) {
+      skipped.push({ id, reason: `recorded refusal: ${refusal.reason}` });
+      continue;
+    }
+    const window = overspentWindow(runtime, now);
+    if (window) {
+      skipped.push({ id, reason: `usage window ${window.window} at ${window.usedPercent}%, above ${JUDGE_USAGE_WINDOW_LIMIT_PERCENT}%` });
+      continue;
+    }
+    // An attempted entry after the pick is still recorded: `nextListJudge`
+    // rebuilds the attempted set from these entries alone, so dropping one
+    // here would let a later hop return to a judge that already refused.
+    const after = list.slice(list.indexOf(id) + 1).filter((later) => attempted.has(later));
+    return { chosen: id, skipped: [...skipped, ...after.map((later) => ({ id: later, reason: "already attempted this run" }))] };
+  }
+  return { chosen: null, skipped };
+}
+
+/**
+ * The judge-list evidence a node's initial assignment records: the list it
+ * read, the entry it chose (or null when every entry was skipped), and why
+ * each earlier one was.
+ *
+ * @param {ValidatedContract} contract
+ * @param {string[]} list
+ * @param {string} workerId
+ * @param {number} [now]
+ * @returns {JudgeListState}
+ */
+export function initialJudgeListState(contract, list, workerId, now = Date.now()) {
+  const pick = selectListJudge(contract, list, workerId, { now });
+  return { list, chosen: pick.chosen, skipped: pick.skipped };
+}
+
+/**
+ * The next hop out of a list-driven judge that was just refused: the current
+ * choice joins the excluded set (it is the one that just failed), so the same
+ * eligibility pass never returns to it -- the hop-by-hop rule. Only an entry
+ * that actually ran and then failed belongs in that excluded set; a candidate
+ * that was skipped for cause (same provider, a recorded refusal, an overspent
+ * window) is re-checked fresh instead, so it keeps its real reason rather than
+ * being relabelled "already attempted" on every later hop. A prior hop's own
+ * "already attempted this run" entries name exactly the ids that were chosen
+ * and failed before this one, so folding those back in carries that set
+ * forward without a dedicated field.
+ *
+ * @param {ValidatedContract} contract
+ * @param {JudgeListState} judgeListState
+ * @param {string} workerId
+ * @param {number} [now]
+ * @returns {JudgeListState}
+ */
+export function nextListJudge(contract, judgeListState, workerId, now = Date.now()) {
+  const previouslyChosen = judgeListState.skipped
+    .filter((entry) => entry.reason === "already attempted this run")
+    .map((entry) => entry.id);
+  const attempted = new Set([...previouslyChosen, ...(judgeListState.chosen ? [judgeListState.chosen] : [])]);
+  // A fresh pass re-derives the whole prefix up to (and including) wherever
+  // it lands: every id the loop reaches before returning is pushed to
+  // `pick.skipped`, on the first hop as on the fifth. That is already a
+  // complete accounting of this hop, so carrying the previous hops' entries
+  // forward too would only repeat each of them once per later hop.
+  const pick = selectListJudge(contract, judgeListState.list, workerId, { attempted, now });
+  return { list: judgeListState.list, chosen: pick.chosen, skipped: pick.skipped };
+}
+
+/**
+ * Does a judge's next fallback runtime share the vendor of the worker it
+ * would be reviewing? Reachability depends on which worker runtime actually
+ * ran, which a static contract cannot know, so this is checked at the moment
+ * a fallback is about to be taken rather than at contract validation.
+ * A list-driven pick already excludes the worker's reachable providers
+ * (`selectListJudge`), so this only ever fires for the single declared
+ * `runtime.fallback` edge.
+ *
+ * @param {ValidatedContract} contract
+ * @param {string|null|undefined} workerRuntimeId
+ * @param {string} nextRuntimeId
+ * @returns {{code: string, message: string}|null}
+ */
+export function judgeFallbackVendorConflict(contract, workerRuntimeId, nextRuntimeId) {
+  const workerVendor = workerRuntimeId ? contract.runtimes[workerRuntimeId]?.vendor : undefined;
+  const judgeFallbackVendor = contract.runtimes[nextRuntimeId]?.vendor;
+  if (!workerVendor || !judgeFallbackVendor || workerVendor !== judgeFallbackVendor) return null;
+  return {
+    code: "judge_fallback_vendor_conflict",
+    message: `judge fallback runtime ${nextRuntimeId} shares vendor ${judgeFallbackVendor} with worker runtime ${workerRuntimeId}`,
+  };
+}

package/src/engine/lifecycle.mjs

@@ -35,6 +35,7 @@ import {
 import { exhaustedUntilOf } from "./runtime-discovery.mjs";
 import { getHarness, normalizeProviderAvailability } from "../harnesses/index.mjs";
 import { availabilityKey, recordRefusal, refusalOfAvailability } from "../run/availability.mjs";
+import { judgeFallbackVendorConflict, nextListJudge } from "./judge-list.mjs";
 
 import { acquire as acquireLock } from "../run/lock.mjs";
 import { invalidResultRepair, parseDiscoveryResult } from "../contract/worker-result.mjs";
@@ -641,7 +642,15 @@ export function handleProviderExhaustion(contract, runDir, node, state, role, en
   const runtime = contract.runtimes?.[current];
   const refusal = runtime ? refusalOfAvailability(normalizeProviderAvailability(runtime, envelope)) : null;
   if (runtime && refusal) recordRefusal(availabilityKey({ harness: runtime.harness, model: runtime.model, executable: getHarness(runtime.harness).executable(runtime) }), refusal);
-  const plan = planRoute(contract, node, state, role, error, current, schedule, now, exhaustedUntil);
+  // R18: a list-driven judge that is about to hop reads the durable refusal
+  // and usage-window stores here -- the impure orchestration layer, not
+  // `planRoute`'s pure decision -- so `current` joins the excluded set before
+  // the next eligible entry is chosen. A reset (wait, no hop) leaves the list
+  // untouched: it costs no runtime and must not spend a list entry either.
+  const judgeListState = role === "judge" && schedule.kind === "failover" && state.routing?.judgeList
+    ? nextListJudge(contract, state.routing.judgeList, sourceWorkerRuntime(state) ?? state.routing?.assignments?.worker ?? current, now)
+    : undefined;
+  const plan = planRoute(contract, node, state, role, error, current, schedule, now, exhaustedUntil, judgeListState ? { judgeListState } : {});
   const status = envelope.status === "failed" ? "failed" : "exhausted";
   if (plan.blocked) {
     // A caller that classified the failure itself also owns what happens when
@@ -659,17 +668,21 @@ export function handleProviderExhaustion(contract, runDir, node, state, role, en
         phase: role,
         result: state.result,
         usage: state.usage,
-        routing: { ...(state.routing ?? {}), tierExhaustion },
+        routing: { ...(state.routing ?? {}), tierExhaustion, ...(judgeListState ? { judgeList: judgeListState } : {}) },
         error: { ...plan.blocked, ...(exhaustedUntil ? { exhaustedUntil } : {}) },
       }, lock);
       void raiseNodeAttention(campaignPath, runDir, state, plan.blocked.code).catch(() => {});
     } else {
-      // Any other block reason is not this feature's evidence to keep.
+      // `judge_list_exhausted` reaches this branch (it is not
+      // `runtime_tier_exhausted`), and its evidence -- naming the judge that
+      // was just refused and why every remaining entry was ineligible -- is
+      // exactly this feature's evidence to keep; only `tierExhaustion` is not.
       clearTierExhaustion(state);
       transition(runDir, state, "exhausted", {
         phase: role,
         result: state.result,
         usage: state.usage,
+        ...(judgeListState ? { routing: { ...(state.routing ?? {}), judgeList: judgeListState } } : {}),
         error: { ...plan.blocked, ...(exhaustedUntil ? { exhaustedUntil } : {}) },
       }, lock);
     }
@@ -681,21 +694,11 @@ export function handleProviderExhaustion(contract, runDir, node, state, role, en
   // so it is refused here, and the node is parked for a human rather than
   // quietly arbitrated by the vendor it is supposed to check.
   if (role === "judge" && plan.nextRuntime !== current) {
-    const workerRuntimeId = sourceWorkerRuntime(state);
-    const workerVendor = workerRuntimeId ? contract.runtimes[workerRuntimeId]?.vendor : undefined;
-    const judgeFallbackVendor = contract.runtimes[plan.nextRuntime]?.vendor;
-    if (workerVendor && judgeFallbackVendor && workerVendor === judgeFallbackVendor) {
+    const conflict = judgeFallbackVendorConflict(contract, sourceWorkerRuntime(state), plan.nextRuntime);
+    if (conflict) {
       if (precomputed) return false;
       clearTierExhaustion(state);
-      transition(runDir, state, "blocked", {
-        phase: role,
-        result: state.result,
-        usage: state.usage,
-        error: {
-          code: "judge_fallback_vendor_conflict",
-          message: `judge fallback runtime ${plan.nextRuntime} shares vendor ${judgeFallbackVendor} with worker runtime ${workerRuntimeId}`,
-        },
-      }, lock);
+      transition(runDir, state, "blocked", { phase: role, result: state.result, usage: state.usage, error: conflict }, lock);
       return true;
     }
   }

package/src/engine/live-preflight.mjs

@@ -46,7 +46,7 @@ const LIVE_PREFLIGHT_OUTPUT_LIMIT_BYTES = 512 * 1024;
  * whole is part of the answer.
  *
  * @param {string} contractPath
- * @param {{static?: boolean, liveTimeoutSec?: number, persisted?: boolean}} [options]
+ * @param {{static?: boolean, liveTimeoutSec?: number, minimumTimeoutSec?: number, persisted?: boolean}} [options]
  * @returns {Promise<ProbeResult[]>}
  */
 export async function preflightContract(contractPath, options = {}) {
@@ -70,7 +70,7 @@ export async function preflightContract(contractPath, options = {}) {
  * once, so a caller cannot accidentally ask a different question.
  *
  * @param {{runtime: RuntimeSnapshot, requiredCapabilitySets?: import("../harnesses/index.mjs").CapabilityRequirements[]}[]} entries
- * @param {{static?: boolean, liveTimeoutSec?: number, cwd?: string}} [options]
+ * @param {{static?: boolean, liveTimeoutSec?: number, minimumTimeoutSec?: number, cwd?: string}} [options]
  * @returns {Promise<ProbeResult[]>}
  */
 export async function preflightRuntimes(entries, options = {}) {
@@ -80,7 +80,7 @@ export async function preflightRuntimes(entries, options = {}) {
   ));
   if (options.static === true) return staticChecks;
 
-  const timeoutSec = livePreflightTimeout(options.liveTimeoutSec);
+  const configuredTimeoutSec = configuredLivePreflightTimeout(options.liveTimeoutSec);
   let liveRepo;
   try {
     liveRepo = createLivePreflightRepo();
@@ -95,7 +95,7 @@ export async function preflightRuntimes(entries, options = {}) {
   }
   try {
     return await Promise.all(staticChecks.map(async (check, index) => {
-      const live = await livePreflight(runtimes[index], liveRepo, timeoutSec);
+      const live = await livePreflight(runtimes[index], liveRepo, configuredTimeoutSec ?? Math.max(options.minimumTimeoutSec ?? 0, defaultLivePreflightTimeout(runtimes[index])));
       const liveDetail = live.status === "done"
         ? `live done · usage ${formatUsage(live.usage)} · cost ${formatCost(live.costUsd)}`
         : `live ${live.status} · ${live.error?.code ?? "provider_error"}: ${redactProviderText(live.error?.message ?? "generation failed")} · usage ${formatUsage(live.usage)} · cost ${formatCost(live.costUsd)}`;
@@ -116,16 +116,54 @@ export async function preflightRuntimes(entries, options = {}) {
     rmSync(liveRepo, { recursive: true, force: true });
   }
 }
-/** @param {number|undefined} configured */
-function livePreflightTimeout(configured) {
+/**
+ * An explicit budget (option, then environment) applies to every runtime.
+ *
+ * @param {number|undefined} configured
+ * @returns {number|null}
+ */
+function configuredLivePreflightTimeout(configured) {
   const raw = configured ?? (process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC === undefined
-    ? 15
+    ? null
     : Number(process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC));
+  if (raw === null) return null;
   if (typeof raw !== "number" || !Number.isFinite(raw) || raw <= 0) {
     throw new TypeError("preflight live timeout must be a positive number of seconds");
   }
   return raw;
 }
+/**
+ * Without an explicit budget, the wait follows the runtime's reasoning effort:
+ * a model at high effort thinks before answering even a one-word prompt.
+ * Measured 2026-09-26: Opus 5.5 at `xhigh` failed every `faberun plan`
+ * preflight at 15 s and passed once the budget was raised, so the planning
+ * pipeline was refusing a runtime that was healthy.
+ *
+ * @param {RuntimeSnapshot} runtime
+ * @returns {number}
+ */
+export function defaultLivePreflightTimeout(runtime) {
+  // Keep LIVE_PREFLIGHT_CEILING_SEC equal to the largest value returned here.
+  const reasoning = /** @type {{reasoning?: unknown}} */ (runtime).reasoning;
+  if (reasoning === "xhigh" || reasoning === "max") return 180;
+  if (reasoning === "high") return 60;
+  return 15;
+}
+/**
+ * The longest a live preflight can take without an explicit budget: what a
+ * detached launcher must be willing to wait before the controller it spawned
+ * can report ready, because the dispatch gate asks before it does.
+ */
+export const LIVE_PREFLIGHT_CEILING_SEC = 180;
+/**
+ * The longest one live preflight ask may take under the current environment.
+ *
+ * @returns {number}
+ */
+export function livePreflightCeilingSec() {
+  const override = Number(process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC);
+  return process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC !== undefined && Number.isFinite(override) && override > 0 ? override : LIVE_PREFLIGHT_CEILING_SEC;
+}
 /** @returns {string} */
 function createLivePreflightRepo() {
   const directory = mkdtempSync(join(tmpdir(), "faberun-preflight-"));

package/src/engine/resume.mjs

@@ -44,6 +44,7 @@ import { validateRunMetadata, OPERATOR_ANSWER_MAX_BYTES } from "../contract/snap
 import { verificationFailureVerdict } from "./judge-gate.mjs";
 import { verificationFailureWithScope } from "../contract/scope-findings.mjs";
 import { applyRejection, applyVerificationFailure, raiseNodeAttention, settleDone } from "./settle.mjs";
+import { isHumanStepPending } from "../contract/human-step.mjs";
 
 /** @typedef {import("../repo/integrate.mjs").IntegrationResult} IntegrationResult */
 /** @typedef {import("./lifecycle.mjs").Invocation} Invocation */
@@ -139,16 +140,37 @@ export async function resumeRun(runDirPath, options = {}) {
       const answerNode = contract.nodes.find((node) => node.id === options.answer?.node);
       if (!answerNode) throw new Error(`unknown node id: ${options.answer.node}`);
       const answerState = states.get(options.answer.node);
-      if (!answerState || !isBlockedContextTerminal(answerState)) {
-        throw new Error(`node ${options.answer.node} is not blocked on missing context`);
-      }
       const answerText = readOperatorAnswer(options.answer.path);
-      const answerOverride = /** @type {import("../contract/index.mjs").ExecutionOverride} */ (/** @type {unknown} */ ({
-        kind: "operator-answer",
-        reason: `operator answered missing context for node ${options.answer.node}`,
-        text: answerText,
-      }));
-      recordExecutionOverride(runDir, answerState, answerOverride, lock);
+      if (answerNode.humanStep) {
+        if (!answerState || !isHumanStepPending(answerState)) {
+          throw new Error(`node ${options.answer.node} is not a pending human step`);
+        }
+        recordExecutionOverride(runDir, answerState, /** @type {import("../contract/index.mjs").ExecutionOverride} */ (/** @type {unknown} */ ({
+          kind: "operator-answer",
+          reason: `operator confirmed the human step for node ${options.answer.node}`,
+          text: answerText,
+        })), lock);
+        // A human step never dispatches to a provider and leaves no attempt
+        // worktree for `settleDone` to seal and integrate -- the operator's
+        // confirmation is the whole of its work, so it settles done directly,
+        // here, before `planResumeRetry` (which knows nothing of `humanStep`)
+        // ever classifies it.
+        transition(runDir, answerState, "done", {
+          phase: "complete",
+          error: null,
+          result: { status: "done", summary: boundedSummary(answerText.trim() || answerNode.humanStep.step), verification: [], artifacts: [], missingContext: [] },
+        }, lock);
+      } else {
+        if (!answerState || !isBlockedContextTerminal(answerState)) {
+          throw new Error(`node ${options.answer.node} is not blocked on missing context`);
+        }
+        const answerOverride = /** @type {import("../contract/index.mjs").ExecutionOverride} */ (/** @type {unknown} */ ({
+          kind: "operator-answer",
+          reason: `operator answered missing context for node ${options.answer.node}`,
+          text: answerText,
+        }));
+        recordExecutionOverride(runDir, answerState, answerOverride, lock);
+      }
     }
     const plan = planResumeRetry(contract, states, { node: options.node, reconcile: options.reconcile, answer: options.answer?.node });
     for (const item of plan.attention) {
@@ -164,6 +186,7 @@ export async function resumeRun(runDirPath, options = {}) {
       if (!state) continue;
       if (state.status === "done") continue;
       if (isBlockedContextTerminal(state) && node.id !== options.answer?.node) continue;
+      if (isHumanStepPending(state) && node.id !== options.answer?.node) continue;
       const action = plan.actions.get(node.id) ?? "recover";
       // Adoption before retry: an unresolved blocking review is re-judged from
       // the preserved worker result, never reset to a fresh worker attempt.
@@ -523,6 +546,20 @@ export async function resumeRun(runDirPath, options = {}) {
 export function isBlockedContextTerminal(state) {
   return state.status === "blocked" && state.error?.code === "context_missing";
 }
+/**
+ * The operator-answer result summary for a human-step node's `done`
+ * transition, bounded to the worker-result protocol's own summary ceiling
+ * (`contract/worker-result.mjs`'s `RESULT_LIMITS.summaryBytes`) since this
+ * writes the same `result.summary` field a provider's own answer would.
+ *
+ * @param {string} text
+ * @returns {string}
+ */
+function boundedSummary(text) {
+  const maxBytes = 4 * 1024;
+  if (Buffer.byteLength(text, "utf8") <= maxBytes) return text;
+  return `${Buffer.from(text, "utf8").subarray(0, maxBytes - 1).toString("utf8")}…`;
+}
 /**
  * The routing write that opens the next tier-exhaustion generation, or null
  * when this re-dispatch is not one: the generation counter bumps by exactly

package/src/engine/run-identity.mjs

@@ -507,11 +507,12 @@ export async function assertEnvironmentReady(contract, runDir, sourceIdentity) {
   // worker and no judge, and can spend no availability.
   if (blocking === null && launchMayDispatch(runDir)) {
     // measured 2026-09-22: asking four routed runtimes in parallel took about
-    // 18s, so the default budget is 60s. FABERUN_PREFLIGHT_TIMEOUT_SEC stays
-    // the operator override; preflightContract validates it, so only a valid
-    // number is lifted here.
+    // 18s, so no runtime is given less than 60s here; a runtime at high
+    // reasoning effort gets its longer default (defaultLivePreflightTimeout).
+    // FABERUN_PREFLIGHT_TIMEOUT_SEC stays the operator override for every
+    // runtime; preflightContract validates it, so only a valid number is lifted.
     const override = Number(process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC);
-    const timeoutSec = process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC !== undefined && Number.isFinite(override) && override > 0 ? override : 60;
+    const timeoutSec = process.env.FABERUN_PREFLIGHT_TIMEOUT_SEC !== undefined && Number.isFinite(override) && override > 0 ? override : undefined;
     const probes = await livePreflightProbes(contract, runDir, sourceIdentity, timeoutSec);
     const silent = probes.filter((probe) => liveSilenceCause(probe) !== null);
     writeJsonAtomic(join(runDir, "env-preflight.json"), evidence(silent.length === 0, probes));
@@ -536,6 +537,8 @@ export async function assertEnvironmentReady(contract, runDir, sourceIdentity) {
   throw Object.assign(new Error(`env_preflight_failed: ${blocking} · the run stays resumable: fix the environment and resume ${runDir}`), { code: "env_preflight_failed" });
 }
 
+/** No routed runtime is given less than this by the dispatch gate. */
+const GATE_MINIMUM_PREFLIGHT_SEC = 60;
 /**
  * The live half of the gate for one launch. Every routed runtime either holds
  * a verdict this machine recorded inside its freshness window -- reused, the
@@ -552,7 +555,7 @@ export async function assertEnvironmentReady(contract, runDir, sourceIdentity) {
  * @param {ValidatedContract} contract
  * @param {string} runDir
  * @param {SourceIdentity|undefined} sourceIdentity
- * @param {number} timeoutSec
+ * @param {number|undefined} timeoutSec
  * @returns {Promise<ProbeResult[]>}
  */
 async function livePreflightProbes(contract, runDir, sourceIdentity, timeoutSec) {
@@ -628,7 +631,7 @@ async function livePreflightProbes(contract, runDir, sourceIdentity, timeoutSec)
       };
     });
   }
-  const probes = await preflightContract(join(runDir, "contract.json"), { liveTimeoutSec: timeoutSec, persisted: true });
+  const probes = await preflightContract(join(runDir, "contract.json"), { ...(timeoutSec === undefined ? {} : { liveTimeoutSec: timeoutSec }), minimumTimeoutSec: GATE_MINIMUM_PREFLIGHT_SEC, persisted: true });
   // What this launch bought is durable from here on: every ask that reached a
   // provider -- a refusal included, an answer being an answer -- is recorded
   // under the provider's own identity. Silence and a command that never