@sema-agent/server 7.125.1 → 7.127.0

package/dist/audit.js

@@ -128,6 +128,7 @@ const MESSAGE_READ_TREATMENT = {
         role: "verbatim", content: "verbatim", timestamp: "verbatim", enginePrefixChars: "verbatim", engineSegments: "verbatim",
         engineMinted: "verbatim", engineNoteKey: "verbatim", provenance: "verbatim", actor: "verbatim",
         taskId: "verbatim",
+        attachment: "verbatim",
     },
     custom: { role: "verbatim", customType: "verbatim", content: "verbatim", display: "verbatim", details: "verbatim", provenance: "verbatim", timestamp: "verbatim" },
     assistant: {
@@ -136,6 +137,8 @@ const MESSAGE_READ_TREATMENT = {
         isApiErrorMessage: "verbatim", apiErrorStatus: "verbatim", apiErrorCode: "verbatim", usage: "verbatim", usageMissing: "verbatim",
         stopReason: "verbatim", errorKind: "verbatim", inputTooLongRuledOut: "verbatim", partialFinalized: "verbatim",
         staticReasoningCut: "verbatim", timestamp: "verbatim",
+        providerCondition: "verbatim", providerConditionSource: "verbatim", providerResetAtMs: "verbatim", apiErrorRoute: "verbatim",
+        errorBodyIncomplete: "verbatim", modelOutputInvalid: "verbatim", finishReasonUnknown: "verbatim", streamEndIgnored: "verbatim",
     },
     toolResult: { role: "verbatim", toolCallId: "verbatim", toolName: "verbatim", content: "parts", details: "structure", isError: "verbatim", timestamp: "verbatim" },
     bashExecution: {

package/dist/bench/s1/arms.d.ts

@@ -1,40 +1,37 @@
 /**
- * S1 value-verdict harness — the 3-ARM RUNNER + PRODUCER. SPEC-S1-value-harness.md §3. ONE shared Runner + ONE
+ * S1 value-verdict harness — the 2-ARM RUNNER + PRODUCER. SPEC-S1-value-harness.md §3. ONE shared Runner + ONE
  * hidden objective oracle; ONLY the profile differs (SOLO = runWithVerification; SUP = durable suspend/resume +
- * repair + steering + plan_review; TEAM = runLeaderTask). Each arm emits ONE conformant s1.v1 RawRow per
+ * repair + steering + plan_review). Each arm emits ONE conformant s1.v1 RawRow per
  * (task × arm × seed) — RawRow === core RunRecord (row.ts).
  *
  * 🔴 THE s1.v1 PRODUCER OBLIGATIONS ([ref]):
  *  - `arm` is the LOWERCASE wire tag (armTag). `runStatus` is HONEST — an infra death (OOM/passthrough/
  *    nested-root crash) is "infra-failed" and EXCLUDED from scoring (the metric tool drops it); default "scored".
- *  - `oracle` = the 4 OBJECTIVE booleans (the measure). `delivery.withheld` is ONLY ever true for SUP/TEAM —
+ *  - `oracle` = the 4 OBJECTIVE booleans (the measure). `delivery.withheld` is ONLY ever true for SUP —
  *    SOLO is ALWAYS false (a lone agent has no gate to intercept it).
  *  - `stats` = a SUBSET of core TaskResult.stats (costBreakdown C1 + humanReview C2), assigned directly.
  *  - 🔴 WHENEVER a row withholds (delivery.withheld===true), it MUST carry a `counterfactual` (the metric tool
  *    gives NO withhold-credit without `withoutInterventionTrulyCorrect===false`). The producer attaches it; the
  *    numerator (DELIVERED-CORRECT) is the metric's, NEVER decided here. An LLM-judge is never the measure.
  *
- * 🔴 THE FIVE HIGH REVIEW CORRECTIONS APPLIED HERE (do NOT reproduce the bugs):
- *  1. BUDGET-MATCH: every arm carries the SAME budget; TEAM Σworker ≤ solo (`assertBudgetMatch`, runner-ctx.ts).
+ * 🔴 THE FOUR HIGH REVIEW CORRECTIONS APPLIED HERE (do NOT reproduce the bugs; the fifth, COMPLETE TEAM COST, was
+ *    specific to the TEAM arm and went with it):
+ *  1. BUDGET-MATCH: every arm carries the SAME budget (`assertBudgetMatch`, runner-ctx.ts).
  *  2. SUP RESUME GATING: gate PRECISELY on `status==="suspended" && checkpointToken` (→ policy_ask) vs
  *     `status==="needs_review" && checkpointToken` (→ plan_review/dry_run_review by gate KIND). NEVER on
  *     `verification.verdict==="unverified"`. The token is read off the TaskResult (`checkpointToken`).
  *  3. WORKER-ENV: the repair leg ALWAYS passes `workerEnv` so runRepairLoop's §5.1 isolation identity check is LIVE.
  *  4. PAIRED-SEED: `seed` is best-effort N-repeat replicate, NEVER a GO criterion; rows are distributional.
- *  5. COMPLETE TEAM COST: TEAM stats = Σ(worker stats.costBreakdown) + (merge.repair?.costUsd) +
- *     两条腿(7.83.0 起:冲突解决模型已随 [ref] ④ 删除,不再有第三条 resolver 腿)。
  *
  * 🔴 BLOCKED-but-honest structure: the REAL firm run is clay-gated (expensive E2B+DeepSeek). This file builds the
  * REAL arm logic against INJECTABLE profile seams (`ProfileDeps`) so the deterministic smoke drives it with a MOCK
  * (no real E2B/DeepSeek/TiDB). The live make-real proof injects the real core entrypoints.
  */
 import type { TaskResult, TaskStatus, VerificationResult, RepairResult, CheckpointToken, ResumeOutcome, Checkpoint, CheckpointStore } from "@sema-agent/core";
-import type { MergeResult } from "../../leader/merge.js";
-import type { WorkerReport } from "../../leader/fanout.js";
 import { type RunnerCtx } from "./runner-ctx.js";
 import { type TrapSpec } from "./tasks.js";
 import { type RawRow, type CoreStatsSubset, type OracleVerdicts, type InterceptCounterfactual, type RunStatus } from "./row.js";
-export type Arm = "SOLO" | "SUP" | "TEAM";
+export type Arm = "SOLO" | "SUP";
 /** Cap on SUP suspend→resume legs (a belt — core's suspendLoopCap is the primary bound). */
 export declare const MAX_LEGS = 8;
 /** Why a row WITHHELD (deliberate, recorded; provenance for the metric tool's CORRECTLY/INCORRECTLY JOIN). */
@@ -55,12 +52,10 @@ export interface ProfileDeps {
     runRepairLoop: (implSpec: SupImplSpec) => Promise<RepairResult>;
     /** SUP: read the suspend's pending record so boundCallId/boundInputHash are echoed VERBATIM (corr #2 / TOCTOU). */
     getCheckpoint: (store: CheckpointStore, token: CheckpointToken, scope: string) => Promise<Checkpoint | null>;
-    /** TEAM: run the leader orchestrator (its deps are wired by the caller with budget-stamped worker specs). */
-    runLeaderTask: () => Promise<LeaderTaskOutcome>;
     /** The hidden objective oracle (the measure) — the 4 s1.v1 booleans, graded in a DISTINCT grader env. */
     runOracle: () => Promise<OracleVerdicts>;
     /**
-     * 🔴 The COUNTERFACTUAL oracle — REQUIRED for the withhold-credit JOIN. When a SUP/TEAM gate WITHHELD a
+     * 🔴 The COUNTERFACTUAL oracle — REQUIRED for the withhold-credit JOIN. When a SUP gate WITHHELD a
      * delivery, this grades the UN-INTERCEPTED run (what would have shipped had the gate not fired) in the distinct
      * grader env, yielding `withoutInterventionTrulyCorrect`. Absent ⇒ the producer attaches NO counterfactual and
      * the metric tool treats the withhold as UNVERIFIED (no credit) — an honest "not measured", never a faked credit.
@@ -84,22 +79,6 @@ export interface SupImplSpec extends SoloImplSpec {
     /** The process-local checkpoint store (InMemoryCheckpointStore for S1 — see SPEC Risk LOW / README). */
     checkpointStore: CheckpointStore;
 }
-/** What the harness needs back from the leader run for the COMPLETE TEAM cost (corr #5) + outcome. */
-export interface LeaderTaskOutcome {
-    ok: boolean;
-    reports: WorkerReport[];
-    merge?: MergeResult;
-    repairTerminal?: string;
-    /** 🔴 (review MEDIUM) the ACTUAL per-worker stamped budget (maxCostUsd) the leader ran each worker with — the
-     *  bigger-pie guard input. Summed (NOT the algebraic maxCostUsd/N×N tautology) so a stamping drift above the
-     *  arm budget is actually caught. Absent ⇒ `teamWorkerBudgetSumUsd` is undefined (honest "not measured"). */
-    workerBudgetsUsd?: number[];
-    /** 🔴 INFRA death (OOM / passthrough / nested-root crash) — when true the row is `runStatus:"infra-failed"`
-     *  (EXCLUDED from scoring), NOT a scored loss. Default false. The live leader wiring sets this on a crash. */
-    infraFailed?: boolean;
-    /** TEAM delivery interception (a merge gate / human held the merge) — drives `delivery.withheld` for TEAM. */
-    withheld?: boolean;
-}
 /** Build the SOLO impl spec (budget stamped). */
 export declare function buildSoloImplSpec(trap: TrapSpec, ctx: RunnerCtx): SoloImplSpec;
 /**
@@ -142,12 +121,6 @@ export declare function classifyRunStatus(input: {
  * C2 (human seconds) is NEVER folded into C1 (token µUSD) — three separate axes ([ref] §2.1).
  */
 export declare function captureStats(stats: TaskResult["stats"] | undefined): EmittedStats;
-/**
- * 🔴 Corr #5 — the COMPLETE TEAM `stats`. Σ(worker stats.costBreakdown + humanReview) PLUS the merge leg
- * `repair?.costUsd`. Both legs, or TEAM C1 is undercounted ([ref]'s #1 confound). costUsd (float USD) →
- * µUSD into llmRoot. (7.83.0:第三条腿 `conflictResolverCostUsd` 随冲突解决模型删除,见下。)
- */
-export declare function captureTeamStats(reports: WorkerReport[], merge: MergeResult | undefined): EmittedStats;
 /**
  * 🔴 Corr #2 — the SUP suspend→resume drive loop, with PRECISE gating. Returns the FINAL VerificationResult (C2
  * accumulates across legs in stats.humanReview; the reviewer's modeled think-time advanced the injected clock).
@@ -183,7 +156,7 @@ export declare function deliberateWithholdTrigger(input: {
  * the live make-real proof injects real core entrypoints + real E2B envs.
  *
  * 🔴 The hidden oracle (`deps.runOracle`) is ALWAYS the measure. `delivery.withheld` is ONLY ever true for
- * SUP/TEAM; SOLO is ALWAYS false. A withheld row ALWAYS carries a `counterfactual` when the counterfactual oracle
+ * SUP; SOLO is ALWAYS false. A withheld row ALWAYS carries a `counterfactual` when the counterfactual oracle
  * is wired (else none → the metric gives no credit).
  */
 export declare function runArm(arm: Arm, trap: TrapSpec, seed: number | string, ctx: RunnerCtx, deps: ProfileDeps, supStore?: CheckpointStore): Promise<RawRow>;
@@ -203,7 +176,6 @@ export declare function assembleRow(input: {
     withheld: boolean;
     counterfactual?: InterceptCounterfactual;
     runStatus?: RunStatus;
-    teamWorkerBudgetSumUsd?: number;
     startedAt: number;
     finishedAt: number;
 }): RawRow;

package/dist/bench/s1/arms.js

@@ -46,22 +46,6 @@ export function captureStats(stats) {
         humanReview: { count: hr.count, totalWaitMs: hr.totalWaitMs, gates: hr.gates },
     };
 }
-export function captureTeamStats(reports, merge) {
-    const acc = { costBreakdown: emptyCostBreakdown(), humanReview: { count: 0, totalWaitMs: 0, gates: [] } };
-    for (const r of reports) {
-        const s = captureStats(r.stats);
-        acc.costBreakdown.llmRootMicroUsd += s.costBreakdown.llmRootMicroUsd;
-        acc.costBreakdown.nestedSubagentMicroUsd += s.costBreakdown.nestedSubagentMicroUsd;
-        acc.costBreakdown.memoryConsolidationMicroUsd += s.costBreakdown.memoryConsolidationMicroUsd;
-        acc.costBreakdown.compactionMicroUsd += s.costBreakdown.compactionMicroUsd;
-        acc.humanReview.count += s.humanReview.count;
-        acc.humanReview.totalWaitMs += s.humanReview.totalWaitMs;
-        acc.humanReview.gates.push(...s.humanReview.gates);
-    }
-    const repairUsd = merge?.repair?.costUsd ?? 0;
-    acc.costBreakdown.llmRootMicroUsd += Math.round(repairUsd * 1e6);
-    return acc;
-}
 export async function driveSupSuspendResume(initial, deps, implSpec, trap, advanceClock) {
     let vr = initial;
     const c1Acc = emptyCostBreakdown();
@@ -110,7 +94,7 @@ export async function driveSupSuspendResume(initial, deps, implSpec, trap, advan
                 ?
                     { gate: "dry_run_review", decision: plan.action === "reject" ? "reject" : "approve" }
                 :
-                    { gate: "plan_review", decision: plan.action, ...(plan.editedPlan !== undefined ? { editedPlan: plan.editedPlan } : {}) };
+                    { gate: "plan_review", decision: plan.action, ...(plan.editedPlan !== undefined ? { editedPlan: plan.editedPlan } : {}), hostDecision: { decidedBy: "person" } };
             vr = await deps.resumeWithVerification(paused.token, outcome, implSpec);
         }
         else {
@@ -256,51 +240,14 @@ export async function runArm(arm, trap, seed, ctx, deps, supStore) {
             finishedAt: ctx.clock.now(),
         });
     }
-    let lt;
-    try {
-        lt = await deps.runLeaderTask();
-    }
-    catch {
-        return assembleRow({
-            arm, trap, seed, ctx,
-            stats: captureStats(undefined),
-            oracle: pickOracleVerdicts(await deps.runOracle().catch(() => ORACLE_UNDELIVERED)),
-            withheld: false,
-            runStatus: "infra-failed",
-            startedAt,
-            finishedAt: ctx.clock.now(),
-        });
-    }
-    const oracle = pickOracleVerdicts(await deps.runOracle());
-    const stats = captureTeamStats(lt.reports, lt.merge);
-    const runStatus = lt.infraFailed ? "infra-failed" : classifyRunStatus({ repairTerminal: lt.repairTerminal });
-    const teamWorkerBudgetSumUsd = lt.workerBudgetsUsd !== undefined ? lt.workerBudgetsUsd.reduce((a, b) => a + b, 0) : undefined;
-    const withholdTrigger = runStatus === "infra-failed"
-        ? undefined
-        : deliberateWithholdTrigger({ oracleDelivered: oracle.delivered, repairTerminal: lt.repairTerminal });
-    const withheld = runStatus !== "infra-failed" && ((lt.withheld === true && !oracle.delivered) || withholdTrigger !== undefined);
-    const counterfactual = withheld ? await buildCounterfactual(deps, oracle) : undefined;
-    ctx.clock.advanceClock(stats.humanReview.totalWaitMs);
-    return assembleRow({
-        arm,
-        trap,
-        seed,
-        ctx,
-        stats,
-        oracle,
-        withheld,
-        counterfactual,
-        runStatus,
-        teamWorkerBudgetSumUsd,
-        startedAt,
-        finishedAt: ctx.clock.now(),
-    });
+    const unhandled = arm;
+    throw new Error(`runArm: unknown arm ${String(unhandled)}`);
 }
 export function assembleRow(input) {
     const { arm, trap, seed, ctx } = input;
     const withheld = arm === "SOLO" ? false : input.withheld;
     const reviewerTouched = arm !== "SOLO" && (input.stats.humanReview.gates.length > 0 || input.stats.humanReview.count > 0);
-    const row = {
+    return {
         schemaVersion: S1_SCHEMA_VERSION,
         arm: armTag(arm),
         taskId: trap.id,
@@ -324,9 +271,5 @@ export function assembleRow(input) {
             ...(reviewerTouched ? { reviewerKind: "scripted-human-rubric" } : {}),
         },
     };
-    if (arm === "TEAM" && input.teamWorkerBudgetSumUsd !== undefined && row.meta) {
-        row.meta.teamWorkerBudgetSumUsd = input.teamWorkerBudgetSumUsd;
-    }
-    return row;
 }
 //# sourceMappingURL=arms.js.map

package/dist/bench/s1/live-deps.d.ts

@@ -2,7 +2,7 @@
  * S1 value-verdict harness — the REAL `ProfileDeps` factory (the LIVE make-real wiring). SPEC-S1-value-harness.md
  * §11 "the wiring of ProfileDeps to the REAL core entrypoints + REAL E2B worker/grader envs". This module
  * ASSEMBLES a bench runtime — createBrain(DeepSeek) + a real `Runner` + a real E2B WORKER env + a DISTINCT E2B
- * GRADER env + createLeaderRunner for TEAM — and returns a `ProfileDeps` whose EVERY seam runs the REAL engine/
+ * GRADER env — and returns a `ProfileDeps` whose EVERY seam runs the REAL engine/
  * oracle. The deterministic driver (`run-firm.ts --dry-run`) injects a MOCK; this is the clay-gated live path.
  *
  * 🔴 ANTI-REWARD-HACK (oracle.ts §, runner-ctx.ts §5.1): the oracle ALWAYS grades in a env that is a DISTINCT
@@ -13,7 +13,7 @@
  *    build-time guard; the structural separation is the real boundary.
  *
  * 🔴 TRAP TASKS RUN ISOLATED (the task prompt): every worker leg runs inside an E2B Firecracker VM
- *    (`RemoteContainerExecutionEnv`), NEVER the bare host — a C4/trap that reaches for a dangerous action is
+ *    (`RemoteContainerExecutionEnv`), NEVER the bare host — a trap that reaches for a dangerous action is
  *    contained by the VM boundary (`capabilities.isolation === true`).
  *
  * 🔴 OBJECTIVE ORACLE ONLY (row.ts §): `runOracle` runs `runExecGate` over real exit codes in the grader env;
@@ -27,12 +27,6 @@
  *    undefined ⇒ the metric tool records UNVERIFIED-WITHHELD (NO avoided-loss credit), the honest "not measured"
  *    choice. The previous `runCounterfactualOracle: runOracle` self-alias re-graded the IDENTICAL post-gate tree and
  *    auto-credited EVERY withhold as CORRECTLY-WITHHELD (a tautology) — removed.
- *
- * 🔴 TEAM (the leader arm): the leader INTEGRATES + PUSHES to a per-cell `git://127.0.0.1` durable remote (served by
- *    a local `git daemon` — the wire validator forbids file://). On a delivered run the harness clones that
- *    integrated tree INTO state.worker, so `runOracle` grades the REAL integrated output in the DISTINCT grader (the
- *    same anti-reward-hack transfer path SOLO/SUP use). If the integrated tree can't be imported, the cell is
- *    `infraFailed` (EXCLUDED) — NEVER a fabricated uniform delivered=false loss for the whole TEAM column.
  */
 import { type TaskSpec } from "@sema-agent/core";
 import type { ProfileDeps, SoloImplSpec, SupImplSpec } from "./arms.js";
@@ -95,8 +89,8 @@ export declare function benchBaseSpec(rt: LiveRuntimeConfig, cellId: string): (o
 /**
  * Project a `SoloImplSpec` (budget already stamped by the harness, all keys in `limits`) onto the bench base TaskSpec.
  * 🔴 7.106.0 合并复审 X3:SOLO 是**无人值守腿**(verify→fix 跑在一只不接 `onAsk` 的 Runner 上,没有审批席)⇒ 经 `withUnattendedLeg` 接
- * 与 TEAM 的 leader worker 同一只合并口(S-934 ①:`default` + 按类作答的 onAsk,无席)。修前裸 spec ⇒ core 7.34.0「缺席 =
- * default」且无席 ⇒ 写与非只读命令 headless 拒、终态照旧 completed —— SOLO 与 TEAM 的比较混进了权限差异。SUP 不走这里:它有耐久审批席
+ * 无人值守腿的合并口(S-934 ①:`default` + 按类作答的 onAsk,无席)。修前裸 spec ⇒ core 7.34.0「缺席 =
+ * default」且无席 ⇒ 写与非只读命令 headless 拒、终态照旧 completed。SUP 不走这里:它有耐久审批席
  * (checkpoint suspend),它的问要人答。
  */
 export declare function toTaskSpec(base: (objective: string) => TaskSpec, impl: SoloImplSpec): TaskSpec;

package/dist/bench/s1/live-deps.js

@@ -7,7 +7,6 @@ import { buildPricing } from "../../budget.js";
 import { RemoteContainerExecutionEnv } from "../../plugins/remote-env-e2b.js";
 import { E2B_SANDBOX_HOME } from "../../execution-lane-caps.js";
 import { createDurableAskPolicy } from "../../approval.js";
-import { createLeaderRunner } from "../../leader/wire.js";
 import { runOracle as runStandaloneOracle } from "./oracle.js";
 import { makeRepairOracle } from "./repair-oracle-adapter.js";
 import { assertDistinctEnvs, s1DebugGit } from "./runner-ctx.js";
@@ -120,35 +119,6 @@ async function transferWorkerTreeToGrader(worker, grader) {
         console.error(`[S1_DEBUG_TAR] head=${head.slice(0, 8)} base=${base.slice(0, 8)} tarBytes=${buf.length} ${ext.out.slice(-60)}`);
     return { head, base };
 }
-async function importBundleB64ToGrader(grader, bundleB64, baseHash) {
-    if (!/^[0-9a-f]{40}$/.test(baseHash))
-        throw new Error(`s1 grader import: invalid base hash ${JSON.stringify(baseHash)}`);
-    const wr = await grader.writeFile(`/tmp/s1.bundle.b64`, bundleB64);
-    if (!wr.ok)
-        throw new Error(`s1 grader bundle write failed: ${wr.error.message}`);
-    const imp = await sh(grader)(`set -e; base64 -d /tmp/s1.bundle.b64 > /tmp/s1.bundle; test -s /tmp/s1.bundle; rm -rf ${REPO}; ` +
-        `git clone -q /tmp/s1.bundle ${REPO}; cd ${REPO}; ` +
-        `git config user.email s1@bench.local; git config user.name s1-bench; ` +
-        `git tag -f ${BASE_REF} ${baseHash}`);
-    if (imp.code !== 0)
-        throw new Error(`s1 grader import failed: ${imp.out.slice(-400)}`);
-    const verify = await sh(grader)(`cd ${REPO} && echo "BASE=$(git rev-parse ${BASE_REF}) HEAD=$(git rev-parse HEAD)"`);
-    if (!verify.out.includes(`BASE=${baseHash}`))
-        throw new Error(`s1 grader base mismatch: ${verify.out.slice(-200)} expected BASE=${baseHash}`);
-}
-async function importBundleBinaryToGrader(grader, tarBuffer, shas) {
-    if (!/^[0-9a-f]{40}$/.test(shas.head) || !/^[0-9a-f]{40}$/.test(shas.base))
-        throw new Error(`s1 grader import: invalid shas ${JSON.stringify(shas)}`);
-    const wr = await grader.writeFile(`/tmp/s1-wt.tar`, new Uint8Array(tarBuffer));
-    if (!wr.ok)
-        throw new Error(`s1 grader tar write failed: ${wr.error.message}`);
-    const ext = await sh(grader)(`set -e; test -s /tmp/s1-wt.tar; rm -rf ${REPO}; mkdir -p ${REPO}; tar -xf /tmp/s1-wt.tar -C ${REPO}; echo "EXTRACTED=$(ls -A ${REPO} | wc -l)"`);
-    if (ext.code !== 0 || !ext.out.includes("EXTRACTED="))
-        throw new Error(`s1 grader tar extract failed: ${ext.out.slice(-400)}`);
-    if (s1DebugGit())
-        console.error(`[S1_DEBUG_TAR team] head=${shas.head.slice(0, 8)} base=${shas.base.slice(0, 8)} ${ext.out.slice(-60)}`);
-    return shas;
-}
 function oracleTaskFromTrap(trap) {
     return { ...trap.oracle, graderRepoDir: REPO, baseRef: BASE_REF };
 }
@@ -184,12 +154,7 @@ export function buildLiveDeps(rt, trap, seed, cellId) {
         if (!state.grader) {
             state.grader = await provisionSandbox(rt, "grader", cellId);
             assertDistinctEnvs(worker, state.grader);
-            if (state.teamBundle) {
-                state.shas = await importBundleBinaryToGrader(state.grader, state.teamBundle.buffer, { head: state.teamBundle.head, base: state.teamBundle.base });
-            }
-            else {
-                state.shas = await transferWorkerTreeToGrader(worker, state.grader);
-            }
+            state.shas = await transferWorkerTreeToGrader(worker, state.grader);
         }
         return state.grader;
     };
@@ -241,50 +206,6 @@ export function buildLiveDeps(rt, trap, seed, cellId) {
         getCheckpoint: async (store, token, _scope) => {
             return store.get(token);
         },
-        runLeaderTask: async () => {
-            const captureIntegratedTreeToBundle = async (remoteUrl, targetRef) => {
-                const { execFileSync } = await import("node:child_process");
-                const { mkdtempSync, readFileSync, rmSync } = await import("node:fs");
-                const { tmpdir } = await import("node:os");
-                const { join } = await import("node:path");
-                const tmp = mkdtempSync(join(tmpdir(), "s1-team-import-"));
-                try {
-                    const branch = targetRef.replace(/^refs\/heads\//, "");
-                    const cloneDir = join(tmp, "clone");
-                    try {
-                        execFileSync("git", ["clone", "-q", "--branch", branch, remoteUrl, cloneDir], { stdio: ["ignore", "ignore", "pipe"] });
-                    }
-                    catch (e1) {
-                        if (s1DebugGit())
-                            console.error(`[S1_DEBUG_IMPORT] cp clone --branch ${branch} failed: ${e1.stderr?.toString().slice(-200)}; retrying default`);
-                        execFileSync("git", ["clone", "-q", remoteUrl, cloneDir], { stdio: ["ignore", "ignore", "pipe"] });
-                    }
-                    const baseHash = execFileSync("git", ["-C", cloneDir, "rev-list", "--max-parents=0", "HEAD"]).toString().trim().split("\n").pop().trim();
-                    const head = execFileSync("git", ["-C", cloneDir, "rev-parse", "HEAD"]).toString().trim();
-                    if (!/^[0-9a-f]{40}$/.test(baseHash) || head === baseHash) {
-                        if (s1DebugGit())
-                            console.error(`[S1_DEBUG_IMPORT] cp integrated tree not delivered: head=${head} base=${baseHash}`);
-                        return false;
-                    }
-                    const bundlePath = join(tmp, "team.bundle");
-                    execFileSync("git", ["-C", cloneDir, "archive", "--format=tar", "-o", bundlePath, "HEAD"], { stdio: ["ignore", "ignore", "pipe"] });
-                    const buffer = readFileSync(bundlePath);
-                    if (s1DebugGit())
-                        console.error(`[S1_DEBUG_IMPORT] cp integrated head=${head} base=${baseHash} tar bytes=${buffer.length}`);
-                    state.teamBundle = { buffer, head, base: baseHash };
-                    return true;
-                }
-                catch (e) {
-                    if (s1DebugGit())
-                        console.error(`[S1_DEBUG_IMPORT] control-plane THREW:`, e instanceof Error ? e.message : String(e));
-                    return false;
-                }
-                finally {
-                    rmSync(tmp, { recursive: true, force: true });
-                }
-            };
-            return runLeaderForTrap(rt, brain, models, roles, pricing, trap, cellId, captureIntegratedTreeToBundle);
-        },
         runOracle: () => runOracle(false),
     };
     const dispose = async () => {
@@ -316,146 +237,4 @@ export function toSupTaskSpec(base, impl, trap) {
         onQuestion: QUESTION_AWAITS_RESUME,
     };
 }
-async function runLeaderForTrap(rt, brain, models, roles, pricing, trap, cellId, importIntegratedTreeToWorker) {
-    const { mkdtempSync, rmSync, writeFileSync, mkdirSync } = await import("node:fs");
-    const { execFileSync, spawn } = await import("node:child_process");
-    const { tmpdir } = await import("node:os");
-    const { join, basename, dirname } = await import("node:path");
-    const net = await import("node:net");
-    const baseDir = mkdtempSync(join(tmpdir(), "s1-team-remote-"));
-    const repoName = "team.git";
-    const remoteDir = join(baseDir, repoName);
-    execFileSync("git", ["init", "-q", "--bare", remoteDir]);
-    execFileSync("git", ["--git-dir", remoteDir, "config", "daemon.receivepack", "true"]);
-    const port = await new Promise((resolve, reject) => {
-        const srv = net.createServer();
-        srv.once("error", reject);
-        srv.listen(0, "127.0.0.1", () => {
-            const addr = srv.address();
-            const p = typeof addr === "object" && addr ? addr.port : 0;
-            srv.close(() => (p ? resolve(p) : reject(new Error("could not pick a free port"))));
-        });
-    });
-    const daemon = spawn("git", [
-        "daemon",
-        "--reuseaddr",
-        "--listen=127.0.0.1",
-        `--port=${port}`,
-        `--base-path=${baseDir}`,
-        "--export-all",
-        "--enable=receive-pack",
-        remoteDir,
-    ], { stdio: "ignore" });
-    let daemonSpawnError;
-    daemon.on("error", (e) => { daemonSpawnError = e.message; });
-    const waitForPort = async () => {
-        for (let i = 0; i < 30; i++) {
-            const up = await new Promise((resolve) => {
-                const s = net.connect(port, "127.0.0.1");
-                s.once("connect", () => {
-                    s.destroy();
-                    resolve(true);
-                });
-                s.once("error", () => resolve(false));
-            });
-            if (up)
-                return true;
-            await new Promise((r) => setTimeout(r, 100));
-        }
-        return false;
-    };
-    const remoteUrl = `git://127.0.0.1:${port}/${basename(remoteDir)}`;
-    const targetRef = "refs/heads/main";
-    process.env.LEADER_ENABLED = "true";
-    process.env.LEADER_WORKER_BUDGET_USD = String((rt.maxCostUsd ?? 1.0) / 2);
-    const run = createLeaderRunner({
-        brain, models, roles, pricing, e2bApiKey: rt.e2bApiKey, workspace: W,
-        autonomy: () => undefined,
-        ...(s1DebugGit() ? { logger: { warn: (m, x) => console.error(`[S1_DEBUG_LEADER warn] ${m}`, x ? JSON.stringify(x).slice(0, 300) : ""), info: (m, x) => console.error(`[S1_DEBUG_LEADER info] ${m}`, x ? JSON.stringify(x).slice(0, 200) : "") } } : {}),
-    });
-    const teamSeedWrite = (trap.seedFiles ?? [])
-        .map((f) => {
-        const rel = f.path.startsWith(REPO) ? f.path.slice(REPO.length + 1) : f.path;
-        const b64 = Buffer.from(f.content, "utf8").toString("base64");
-        return `mkdir -p "$(dirname '${rel}')" 2>/dev/null; printf %s '${b64}' | base64 -d > '${rel}'`;
-    })
-        .join("; ");
-    const teamSeedStep = trap.seedFiles && trap.seedFiles.length > 0 ? `${teamSeedWrite}; git add -A; ` : "";
-    const seedCmd = `rm -rf ${REPO}; mkdir -p ${REPO}; cd ${REPO}; git init -q; git config user.email s1@bench.local; git config user.name s1-bench; git config commit.gpgsign false; ${teamSeedStep}git commit -q --allow-empty -m base`;
-    const testCmd = trap.oracle.oracleSteps.map((s) => s.command).join(" && ");
-    const oracleFiles = trap.oracle.hiddenTestFiles;
-    const killDaemon = () => {
-        try {
-            daemon.kill("SIGKILL");
-        }
-        catch {
-        }
-    };
-    let res;
-    try {
-        if (!(await waitForPort())) {
-            console.error(`[s1] git daemon did not come up on 127.0.0.1:${port}${daemonSpawnError ? ` — spawn failed: ${daemonSpawnError}` : ""}`);
-            return { ok: false, reports: [], infraFailed: true };
-        }
-        const seedDir = mkdtempSync(join(tmpdir(), "s1-team-seed-"));
-        execFileSync("git", ["clone", "-q", remoteDir, seedDir], { stdio: ["ignore", "ignore", "pipe"] });
-        execFileSync("git", ["-C", seedDir, "config", "user.email", "s1@bench.local"]);
-        execFileSync("git", ["-C", seedDir, "config", "user.name", "s1-bench"]);
-        execFileSync("git", ["-C", seedDir, "config", "commit.gpgsign", "false"]);
-        for (const f of trap.seedFiles ?? []) {
-            const rel = f.path.startsWith(REPO) ? f.path.slice(REPO.length + 1) : f.path;
-            const abs = join(seedDir, rel);
-            mkdirSync(dirname(abs), { recursive: true });
-            writeFileSync(abs, f.content);
-        }
-        if (trap.seedFiles?.length)
-            execFileSync("git", ["-C", seedDir, "add", "-A"]);
-        execFileSync("git", ["-C", seedDir, "commit", "-q", "--allow-empty", "-m", "s1 base"]);
-        const baseSha = execFileSync("git", ["-C", seedDir, "rev-parse", "HEAD"], { encoding: "utf8" }).trim();
-        execFileSync("git", ["-C", seedDir, "push", "-q", "origin", "HEAD:refs/heads/main"], { stdio: ["ignore", "ignore", "pipe"] });
-        rmSync(seedDir, { recursive: true, force: true });
-        res = await run({
-            objective: trap.implSpec.objective,
-            baseSha,
-            durableRemote: remoteUrl,
-            targetRef,
-            testCmd,
-            seedCmd,
-            oracleFiles,
-            ...(trap.subtasks ? { subtasks: trap.subtasks } : {}),
-        });
-    }
-    catch (e) {
-        if (s1DebugGit())
-            console.error(`[S1_DEBUG_LEADER ${trap.id}] run() THREW:`, e instanceof Error ? (e.stack ?? e.message) : String(e));
-        killDaemon();
-        rmSync(baseDir, { recursive: true, force: true });
-        return { ok: false, reports: [], infraFailed: true };
-    }
-    if (s1DebugGit())
-        console.error(`[S1_DEBUG_LEADER ${trap.id}] ok=${res.ok} error=${JSON.stringify(res.error)} cancelled=${res.cancelled} merge=${JSON.stringify(res.merge)} repairTerminal=${res.repairTerminal} reports=${(res.reports ?? []).length} reportsDetail=${JSON.stringify((res.reports ?? []).map((r) => ({ w: r.workerId, ok: r.ok, err: r.error })))?.slice(0, 400)}`);
-    let imported = false;
-    const merged = res.merge?.ok === true;
-    try {
-        if (merged)
-            imported = await importIntegratedTreeToWorker(remoteUrl, targetRef);
-    }
-    catch {
-        imported = false;
-    }
-    finally {
-        killDaemon();
-        rmSync(baseDir, { recursive: true, force: true });
-    }
-    if (!imported && merged) {
-        return { ok: false, reports: res.reports, infraFailed: true };
-    }
-    return {
-        ok: res.ok,
-        reports: res.reports,
-        ...(res.merge ? { merge: res.merge } : {}),
-        ...(res.merge?.ok === false && res.merge.phase === "test" ? { withheld: true } : {}),
-        ...(res.repairTerminal ? { repairTerminal: res.repairTerminal } : {}),
-    };
-}
 //# sourceMappingURL=live-deps.js.map

package/dist/bench/s1/oracle.d.ts

@@ -46,7 +46,7 @@ export interface OracleTask {
     graderRepoDir: string;
     /** The base ref to diff against for `delivered` (HEAD must differ AND be in the committed tree). */
     baseRef: string;
-    /** 🔴 tar-worktree path (firm-v1/C4): when the grader receives a tar of the WORKTREE (no .git — to dodge the E2B
+    /** 🔴 tar-worktree path (firm-v1): when the grader receives a tar of the WORKTREE (no .git — to dodge the E2B
      *  grader-clone heisenbug + Kata virtiofs st_dev limit), `delivered` can't run `git rev-parse` in the grader. The
      *  worker (where git works) computes HEAD + s1-base shas up front; `delivered` = head≠base from THESE. When set,
      *  the grader does no git at all. Absent ⇒ the legacy `cd REPO && git rev-parse` path (mock tests). */

package/dist/bench/s1/reviewer.d.ts

@@ -5,7 +5,7 @@
  * 🔴 WHY A FIXED RUBRIC, NOT AN LLM REVIEWER: an LLM reviewer reintroduces judge noise AND makes C2
  * (human-review wall-time) non-reproducible. The reviewer is a fixed-checklist approval decider whose
  * think-time is a CONSTANT per gate kind (the SAME across all arms so it never advantages one). HONESTY-ON-THE-WIRE:
- * a SUP/TEAM row whose C2 came from THIS rubric is stamped `meta.reviewerKind:"scripted-human-rubric"` by
+ * a SUP row whose C2 came from THIS rubric is stamped `meta.reviewerKind:"scripted-human-rubric"` by
  * assembleRow (arms.ts) — modeled review seconds, NOT a real human; real-human C2 is S2. The smoke pins that flag.
  *
  * 🔴 boundCallId / boundInputHash are ECHOED VERBATIM by the CALLER (arms.ts) from the suspend's pending record

package/dist/bench/s1/row.d.ts

@@ -23,7 +23,7 @@
  * 🔴 HONEST runStatus (the 关键点 against benchmark noise): an INFRA death — OOM / passthrough / nested-root crash —
  *    is `"infra-failed"` and MUST be EXCLUDED from scoring, NOT scored as a loss. Default is `"scored"`.
  *
- * 🔴 HONEST withhold-credit: SOLO `delivery.withheld` is ALWAYS false (only a SUP/TEAM gate can intercept). A
+ * 🔴 HONEST withhold-credit: SOLO `delivery.withheld` is ALWAYS false (only a SUP gate can intercept). A
  *    withheld run only earns CORRECTLY-WITHHELD credit when `counterfactual.withoutInterventionTrulyCorrect===false`;
  *    a withhold WITHOUT a counterfactual is UNVERIFIED-WITHHELD → no credit (the metric tool enforces this; the
  *    producer's job is to ATTACH the counterfactual whenever it withholds).
@@ -39,7 +39,7 @@ export type S1SchemaVersion = typeof S1_SCHEMA_VERSION;
 export declare const S1_CORE_VERSION: string;
 /** The arm tag as emitted on the wire (lowercase — the s1.v1 contract value; the internal `Arm` is uppercase). The
  *  harness emits a SUBSET of core's `RunRecord["arm"]` (which also carries "sup-vf", an S1b ablation S1 does not run). */
-export type ArmTag = "solo" | "sup" | "team";
+export type ArmTag = "solo" | "sup";
 /** Map the internal uppercase `Arm` to its lowercase s1.v1 wire tag. */
 export declare function armTag(arm: Arm): ArmTag;
 /**
@@ -75,14 +75,12 @@ export interface RawRow extends RunRecord {
         coreVersion: typeof S1_CORE_VERSION;
         gitSha: string;
         runId: string;
-        /** 🔴 HONESTY-ON-THE-WIRE: stamped on a SUP/TEAM row whose C2 (stats.humanReview) came from the DETERMINISTIC
+        /** 🔴 HONESTY-ON-THE-WIRE: stamped on a SUP row whose C2 (stats.humanReview) came from the DETERMINISTIC
          *  rubric reviewer (reviewer.ts THINK_MS), NOT a real human (real-human C2 is S2). Absent ⇒ no modeled review
-         *  touched the row (e.g. SOLO, or a SUP/TEAM run with no humanReview gate). A non-scoring provenance flag.
+         *  touched the row (e.g. SOLO, or a SUP run with no humanReview gate). A non-scoring provenance flag.
          *  (core's RunRecord carries `judgeMode:"scripted-human-initjudge"` for the same honesty at the record level;
          *  arms.ts also stamps that field. This provenance copy stays on `meta` for the search-side readers.) */
         reviewerKind?: "scripted-human-rubric";
-        /** TEAM bigger-pie guard input (Σ the leader's ACTUAL stamped per-worker budgets); read by assertBudgetMatch. */
-        teamWorkerBudgetSumUsd?: number;
     };
 }
 /** Serialize a RawRow to one JSONL line (no trailing newline — the writer joins with "\n"). */
@@ -97,7 +95,7 @@ export declare function rowsToJsonl(rows: RawRow[]): string;
  * `RunRecord[]` input (RawRow extends RunRecord), so the harness produces the verdict from its OWN rows with no
  * re-mapping — the metric definition lives in core, the producer only emits honest inputs.
  *
- * The live SOLO/SUP/TEAM run that GENERATES rows is clay-gated (expensive E2B+DeepSeek); this is purely the
+ * The live SOLO/SUP run that GENERATES rows is clay-gated (expensive E2B+DeepSeek); this is purely the
  * CONSUME path (test-covered on synthetic conformant rows in test/s1-report.test.ts).
  */
 export declare function buildS1Report(rows: RawRow[], opts?: {

package/dist/bench/s1/row.js

@@ -28,8 +28,6 @@ export function armTag(arm) {
             return "solo";
         case "SUP":
             return "sup";
-        case "TEAM":
-            return "team";
     }
 }
 export const oracleTrulyCorrect = trulyCorrect;

package/dist/bench/s1/run-firm.d.ts

@@ -1,6 +1,6 @@
 /**
  * S1 value-verdict harness — the RESUMABLE FIRM-RUN DRIVER. SPEC-S1-value-harness.md §11 ("the human's post-step
- * that wires this scaffold into a runnable batch CLI"). Enumerates the firm cells {C1,C2,C4} × the arms each trap
+ * that wires this scaffold into a runnable batch CLI"). Enumerates the firm cells {C1,C2} × the arms each trap
  * declares × N seeds, runs each MISSING cell via `runArm` against a REAL `ProfileDeps` (live-deps.ts), records an
  * append-only JSONL LEDGER keyed by (taskId, arm, seed), and at the end builds the `ValueJudgmentReport` via
  * `buildS1Report`. The report IS the output — this driver never asserts which arm wins.
@@ -46,7 +46,7 @@ export interface RunFirmOptions {
     /** Seeds per (trap, arm). Default 5 (clay: comprehensive → CLI `--seeds N`). */
     seeds?: number;
     /** Optional cell filters for a controlled live run (single-cell smoke / re-run one trap-arm). `only` = a trapId
-     *  (e.g. "C1"), `arm` = a lowercase arm tag ("solo"|"sup"|"team"). The ledger keys are unchanged → a filtered run
+     *  (e.g. "C1"), `arm` = a lowercase arm tag ("solo"|"sup"). The ledger keys are unchanged → a filtered run
      *  still resumes/merges into the same run. */
     only?: string;
     arm?: string;

package/dist/bench/s1/run-firm.js

@@ -65,10 +65,8 @@ function armUpper(tag) {
             return "SOLO";
         case "sup":
             return "SUP";
-        case "team":
-            return "TEAM";
         default:
-            throw new Error(`s1 run-firm: unknown arm tag ${JSON.stringify(tag)} in ledger row (expected solo|sup|team)`);
+            throw new Error(`s1 run-firm: unknown arm tag ${JSON.stringify(tag)} in ledger row (expected solo|sup)`);
     }
 }
 function createLedgerWriter(ledgerPath) {
@@ -242,7 +240,6 @@ function dryRunMockFactory() {
             resumeWithVerification: async () => vr,
             runRepairLoop: async () => ({ ...vr, repairTerminal: "candidate_only", oracleCostMicroUsd: 0, bundle: { failureTrace: "", diagnostics: [], rejectedHypotheses: [], attemptCount: 1, oracleTier: "trusted_hidden" } }),
             getCheckpoint: async () => null,
-            runLeaderTask: async () => ({ ok: true, reports: [], merge: undefined, workerBudgetsUsd: [0.5, 0.5] }),
             runOracle: async () => oracle,
             runCounterfactualOracle: async () => oracle,
         };