@sema-agent/server 7.56.0 → 7.58.0

package/dist/leader/fanout.js

@@ -1,3 +1,4 @@
+import { redactErrorMessage, redactWorkerReportError } from "../observability/run-terminal-log.js";
 class Semaphore {
     max;
     active = 0;
@@ -62,10 +63,10 @@ export async function fanOut(subtasks, opts) {
                 return { ...base, status: "timeout", error: "cancelled before start" };
             if (opts.runWorker) {
                 try {
-                    return maybeCancelOnSuspend(await opts.runWorker(st, sharedAC.signal));
+                    return maybeCancelOnSuspend(redactWorkerReportError(await opts.runWorker(st, sharedAC.signal)));
                 }
                 catch (e) {
-                    return { ...base, status: cancelled ? "timeout" : "failed", error: e instanceof Error ? e.message : String(e) };
+                    return { ...base, status: cancelled ? "timeout" : "failed", error: redactErrorMessage(e instanceof Error ? e.message : String(e)) ?? "" };
                 }
             }
             let stream;
@@ -73,16 +74,16 @@ export async function fanOut(subtasks, opts) {
                 stream = opts.runner.runTaskStream({ ...st.spec, sessionId: st.sessionId, signal: sharedAC.signal });
             }
             catch (e) {
-                return { ...base, status: "failed", error: e instanceof Error ? e.message : String(e) };
+                return { ...base, status: "failed", error: redactErrorMessage(e instanceof Error ? e.message : String(e)) ?? "" };
             }
             streams.set(st.workerId, stream);
             let report;
             try {
                 const result = await stream.result();
-                report = { ...base, status: mapStatus(result.status), stats: result.stats, ...(result.errorMessage ? { error: result.errorMessage } : {}) };
+                report = { ...base, status: mapStatus(result.status), stats: result.stats, ...(result.errorMessage ? { error: redactErrorMessage(result.errorMessage) ?? "" } : {}) };
             }
             catch (e) {
-                report = { ...base, status: cancelled ? "timeout" : "failed", error: e instanceof Error ? e.message : String(e) };
+                report = { ...base, status: cancelled ? "timeout" : "failed", error: redactErrorMessage(e instanceof Error ? e.message : String(e)) ?? "" };
             }
             finally {
                 streams.delete(st.workerId);

package/dist/leader/leader.js

@@ -1,4 +1,5 @@
 import { fanOut, isDurablePauseStatus } from "./fanout.js";
+import { redactErrorMessage, redactHead, redactPatchHead } from "../observability/run-terminal-log.js";
 import { pullDiff } from "./diffout.js";
 import { mergeBranches } from "./merge.js";
 import { Runner, runWithVerification, resumeWithVerification, delimitUntrusted } from "@sema-agent/core";
@@ -58,8 +59,8 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                             deps.logger?.warn?.("worker_failed", {
                                 workerId: w.workerId, status: r.status, turns: r.stats?.turns ?? 0,
                                 ...(r.errorCode ? { errorCode: r.errorCode } : {}),
-                                ...(r.errorMessage ? { error: r.errorMessage.slice(0, 200) } : {}),
-                                ...(r.blockedReason ? { blockedReason: r.blockedReason.slice(0, 200) } : {}),
+                                ...(r.errorMessage ? { error: redactHead(r.errorMessage, 200) } : {}),
+                                ...(r.blockedReason ? { blockedReason: redactHead(r.blockedReason, 200) } : {}),
                             });
                         }
                         return {
@@ -85,7 +86,7 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                         deps.logger?.info?.("worker_resource_resume", { workerId: w.workerId, slice });
                         return resumeWithVerification(w.runner, token, { gate: "resource_limit", decision: "continue" }, { ...w.spec, sessionId: w.sessionId, signal }, w.spec.objective ?? "", cfg);
                     });
-                    deps.logger?.info?.("verify_outcome", { workerId: w.workerId, status: vr.status, verdict: vr.verification.verdict, rounds: vr.verification.rounds, finding: (vr.verification.findings[0] ?? "").slice(0, 220) });
+                    deps.logger?.info?.("verify_outcome", { workerId: w.workerId, status: vr.status, verdict: vr.verification.verdict, rounds: vr.verification.rounds, finding: redactHead(vr.verification.findings[0] ?? "", 220) });
                     const pass = vr.status === "completed" && vr.verification.verdict === "PASS";
                     const stillSuspended = vr.checkpointToken !== undefined;
                     const reportStatus = pass ? "completed" : stillSuspended ? "suspended" : "failed";
@@ -93,8 +94,8 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                         deps.logger?.warn?.("worker_failed", {
                             workerId: w.workerId, status: reportStatus, turns: vr.stats?.turns ?? 0, verdict: vr.verification.verdict,
                             ...(vr.errorCode ? { errorCode: vr.errorCode } : {}),
-                            ...(vr.errorMessage ? { error: vr.errorMessage.slice(0, 200) } : {}),
-                            ...(vr.blockedReason ? { blockedReason: vr.blockedReason.slice(0, 200) } : {}),
+                            ...(vr.errorMessage ? { error: redactHead(vr.errorMessage, 200) } : {}),
+                            ...(vr.blockedReason ? { blockedReason: redactHead(vr.blockedReason, 200) } : {}),
                         });
                     }
                     return {
@@ -105,7 +106,7 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                             : {
                                 error: vr.verification.verdict === "unverified"
                                     ? (vr.errorMessage ?? vr.blockedReason ?? `worker ${vr.status}`)
-                                    : `verify ${vr.verification.verdict}: ${(vr.verification.findings[0] ?? "").slice(0, 200)}`,
+                                    : `verify ${vr.verification.verdict}: ${redactHead(vr.verification.findings[0] ?? "", 200)}`,
                             }),
                         ...(vr.checkpointGate ? { checkpointGate: vr.checkpointGate } : {}),
                         ...(vr.errorCode ? { errorCode: vr.errorCode } : {}),
@@ -127,7 +128,7 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                 return { reports: [rr.report], cancelled: false };
             }
             catch (e) {
-                deps.logger?.warn?.("repair_loop_dispatch_failed", { workerId: solo.workerId, err: e instanceof Error ? e.message : String(e) });
+                deps.logger?.warn?.("repair_loop_dispatch_failed", { workerId: solo.workerId, err: redactHead(e instanceof Error ? e.message : String(e), 2000) });
                 return runWave([toSubtask(solo)]);
             }
         };
@@ -178,14 +179,15 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                     r.repairs = attempt;
                     if (verdict.ok)
                         return;
+                    const safeVerdictReason = redactErrorMessage(verdict.reason ?? "behaviorally incorrect") ?? "";
                     if (attempt === maxR) {
                         r.status = "failed";
-                        r.error = `module gate: ${(verdict.reason ?? "behaviorally incorrect").slice(0, 200)}`;
+                        r.error = `module gate: ${redactHead(verdict.reason ?? "behaviorally incorrect", 200)}`;
                         return;
                     }
-                    deps.logger?.info?.("module_repair", { workerId: r.workerId, attempt: attempt + 1, reason: verdict.reason });
+                    deps.logger?.info?.("module_repair", { workerId: r.workerId, attempt: attempt + 1, reason: redactHead(verdict.reason ?? "", 500) });
                     try {
-                        const rr = await w.runner.runTaskStream({ ...sub.spec, objective: deps.moduleGate.repairObjective(sub, verdict.reason ?? ""), sessionId: `${w.sessionId}-r${attempt + 1}` }).result();
+                        const rr = await w.runner.runTaskStream({ ...sub.spec, objective: deps.moduleGate.repairObjective(sub, safeVerdictReason), sessionId: `${w.sessionId}-r${attempt + 1}` }).result();
                         if (r.stats && rr?.stats)
                             r.stats = { ...r.stats, tokens: (r.stats.tokens ?? 0) + (rr.stats.tokens ?? 0), costMicroUsd: (r.stats.costMicroUsd ?? 0) + (rr.stats.costMicroUsd ?? 0) };
                     }
@@ -226,8 +228,9 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
             };
         };
         const collapseToSingle = async (trigger, reason, priorMerge) => {
-            deps.logger?.warn?.("replan_collapse_to_single", { trigger, reason: reason.slice(0, 300) });
-            const replan = { trigger, collapsed: reason.slice(0, 300), ...(redispatched ? { redispatched } : {}) };
+            const safeReason = redactHead(reason, 300);
+            deps.logger?.warn?.("replan_collapse_to_single", { trigger, reason: safeReason });
+            const replan = { trigger, collapsed: safeReason, ...(redispatched ? { redispatched } : {}) };
             const priorDiffs = [];
             await Promise.all(reports.map(async (r) => {
                 const w = byId.get(r.workerId);
@@ -249,8 +252,8 @@ export async function runLeaderTask(task, durableBaseSha, deps) {
                     task,
                     ``,
                     `--- PRIOR PARALLEL WORK — REUSE IT, do not re-derive from scratch ---`,
-                    `${priorDiffs.length} worker(s) attempted this task in parallel; their changes could NOT be integrated cleanly (${reason.slice(0, 150)}). Each worker's diff over the base is below. COPY the correct, non-overlapping parts verbatim and produce ONE coherent change for the WHOLE task, resolving only the overlaps. Treat these as a strong starting point — fix anything wrong, but do not throw the work away.`,
-                    ...priorDiffs.map((d) => delimitUntrusted(`PRIOR WORKER '${d.workerId}' (${d.status}) DIFF (untrusted)`, d.patch.slice(0, 8000))),
+                    `${priorDiffs.length} worker(s) attempted this task in parallel; their changes could NOT be integrated cleanly (${safeReason.slice(0, 150)}). Each worker's diff over the base is below. COPY the correct, non-overlapping parts verbatim and produce ONE coherent change for the WHOLE task, resolving only the overlaps. Treat these as a strong starting point — fix anything wrong, but do not throw the work away.`,
+                    ...priorDiffs.map((d) => delimitUntrusted(`PRIOR WORKER '${d.workerId}' (${d.status}) DIFF (untrusted)`, redactPatchHead(d.patch, 8000))),
                 ].join("\n");
             const soloSub = { workerId: "collapse-solo", branch: "worker/collapse-solo", files: ["."], spec: { objective: soloObjective } };
             let solo;

package/dist/leader/merge.js

@@ -1,4 +1,5 @@
 import { delimitUntrusted } from "@sema-agent/core";
+import { OVERSIZED_DIAG_PLACEHOLDER, redactErrorMessage, redactHead, redactTail } from "../observability/run-terminal-log.js";
 import { applyAndGrade, changedFilesOutOfAllowlist } from "./grader-env-factory.js";
 import { shellQuote as Q } from "../plugins/remote-shell.js";
 export async function mergeBranches(reports, baseSha, deps) {
@@ -40,7 +41,7 @@ export async function mergeBranches(reports, baseSha, deps) {
                 conflictResolverCostUsd += cr.costUsd;
                 if (!cr.resolved) {
                     deps.logger?.warn?.("merge_apply_failed", { workerId: p.workerId, exitCode: ap.exitCode });
-                    return { ok: false, phase: "apply", reason: `git apply --3way failed for ${p.workerId} (conflict)`, quarantine: [p.workerId], rej: rej.stdout, details: ap.stderr.slice(0, 800), ...(conflictResolverCostUsd > 0 ? { conflictResolverCostUsd } : {}) };
+                    return { ok: false, phase: "apply", reason: `git apply --3way failed for ${p.workerId} (conflict)`, quarantine: [p.workerId], rej: rej.stdout, details: redactHead(ap.stderr, 800), ...(conflictResolverCostUsd > 0 ? { conflictResolverCostUsd } : {}) };
                 }
                 conflictsResolved.push(p.workerId);
             }
@@ -51,14 +52,18 @@ export async function mergeBranches(reports, baseSha, deps) {
             let cost = 0;
             let round = 0;
             for (; round < integ.repair.maxRounds && test.exitCode !== 0; round++) {
-                const errorText = (test.stdout + "\n" + test.stderr).slice(-4000);
+                const errorText = redactTail(test.stdout + "\n" + test.stderr, 4000);
+                if (errorText.startsWith(OVERSIZED_DIAG_PLACEHOLDER)) {
+                    deps.logger?.warn?.("integration_repair_skipped_oversized_output", { round: round + 1, outputChars: test.stdout.length + test.stderr.length });
+                    break;
+                }
                 deps.logger?.warn?.("integration_repair_round", { round: round + 1, maxRounds: integ.repair.maxRounds });
                 try {
                     const r = await integ.repair.run({ round: round + 1, errorText, testCmd: deps.testCmd, repoDir: repo });
                     cost += r.costUsd ?? 0;
                 }
                 catch (e) {
-                    deps.logger?.warn?.("integration_repair_error", { round: round + 1, err: e instanceof Error ? e.message : String(e) });
+                    deps.logger?.warn?.("integration_repair_error", { round: round + 1, err: redactHead(e instanceof Error ? e.message : String(e), 2000) });
                     break;
                 }
                 repairMeta = { rounds: round + 1, fixed: false, costUsd: cost };
@@ -69,11 +74,11 @@ export async function mergeBranches(reports, baseSha, deps) {
             deps.logger?.info?.("integration_repair_done", { ...repairMeta });
         }
         if (test.exitCode !== 0) {
-            return { ok: false, phase: "test", reason: `integration test failed (exit ${test.exitCode})`, details: (test.stdout + "\n" + test.stderr).slice(-2000), ...(repairMeta ? { repair: repairMeta } : {}), ...(conflictResolverCostUsd > 0 ? { conflictResolverCostUsd } : {}) };
+            return { ok: false, phase: "test", reason: `integration test failed (exit ${test.exitCode})`, details: redactTail(test.stdout + "\n" + test.stderr, 2000), ...(repairMeta ? { repair: repairMeta } : {}), ...(conflictResolverCostUsd > 0 ? { conflictResolverCostUsd } : {}) };
         }
         if (deps.measureCmd) {
             const mr = await env.exec(deps.measureCmd, { cwd: repo });
-            measure = mr.exitCode === 0 ? { pass: true } : { pass: false, reason: (mr.stdout + " " + mr.stderr).slice(-200) };
+            measure = mr.exitCode === 0 ? { pass: true } : { pass: false, reason: redactTail(mr.stdout + " " + mr.stderr, 200) };
         }
         phase = "push";
         if (repairMeta)
@@ -81,11 +86,11 @@ export async function mergeBranches(reports, baseSha, deps) {
         const costTail = { ...(repairMeta ? { repair: repairMeta } : {}), ...(conflictResolverCostUsd > 0 ? { conflictResolverCostUsd } : {}) };
         const commit = await env.exec(`git add -A && git -c user.email=leader@local -c user.name=leader commit -m ${Q(`merge: integrate ${mergeable.map((r) => r.workerId).join("+")}`)} --allow-empty`, { cwd: repo });
         if (commit.exitCode !== 0) {
-            return { ok: false, phase: "push", reason: `failed to create integration commit (exit ${commit.exitCode})`, details: (commit.stdout + "\n" + commit.stderr).slice(0, 800), ...costTail };
+            return { ok: false, phase: "push", reason: `failed to create integration commit (exit ${commit.exitCode})`, details: redactHead(commit.stdout + "\n" + commit.stderr, 800), ...costTail };
         }
         const integrated = await env.exec(`git format-patch --stdout ${Q(integBase)}..HEAD`, { cwd: repo });
         if (integrated.exitCode !== 0) {
-            return { ok: false, phase: "push", reason: `failed to produce integrated patch (exit ${integrated.exitCode})`, details: integrated.stderr.slice(0, 800), ...costTail };
+            return { ok: false, phase: "push", reason: `failed to produce integrated patch (exit ${integrated.exitCode})`, details: redactHead(integrated.stderr, 800), ...costTail };
         }
         if (!integrated.stdout.trim()) {
             return { ok: false, phase: "push", reason: "integrated patch is empty (nothing applied over base) — refusing to push", quarantine: mergeable.map((r) => r.workerId), ...costTail };
@@ -148,11 +153,11 @@ async function evaluatePushHold(mergeable, observeMeasure, integratedPatch, deps
                 return { hold: { reason: `integrated patch touches out-of-allowlist / build-config paths (${offending.slice(0, 5).join(", ")}) — needs_human_oracle (§10.4(ii))`, terminal: "needs_human_oracle", candidatePatch: heldCandidate() }, measure };
             }
             const grade = await applyAndGrade(mdr.grader, integratedPatch, mdr.measureSteps);
-            measure = grade.passed ? { pass: true } : { pass: false, reason: (grade.trace ?? "measure failed in out-of-process grader").slice(-200) };
+            measure = grade.passed ? { pass: true } : { pass: false, reason: redactTail(grade.trace ?? "measure failed in out-of-process grader", 200) };
             deps.logger?.info?.("merge_measure_grader", { passed: grade.passed, applied: grade.applied });
         }
         catch (e) {
-            deps.logger?.warn?.("merge_measure_grader_error", { err: e instanceof Error ? e.message : String(e) });
+            deps.logger?.warn?.("merge_measure_grader_error", { err: redactErrorMessage(e instanceof Error ? e.message : String(e)) ?? "" });
             return { hold: { reason: `measure grader errored — needs_human_oracle (§10.4)`, terminal: "needs_human_oracle", candidatePatch: heldCandidate() }, measure };
         }
     }
@@ -160,14 +165,14 @@ async function evaluatePushHold(mergeable, observeMeasure, integratedPatch, deps
     const strongOracleUnconfirmed = deps.strongOracleSeeded === true;
     if (measureFailing || strongOracleUnconfirmed) {
         const why = measureFailing ? "measure did not pass" : "hidden held-out oracle seeded but only the compile gate cleared";
-        deps.logger?.warn?.("merge_push_held_signal", { why, measureReason: (measure?.reason ?? "").slice(0, 200) });
+        deps.logger?.warn?.("merge_push_held_signal", { why, measureReason: redactHead(measure?.reason ?? "", 200) });
         return { hold: { reason: `strong oracle unconfirmed: ${why} — candidate_only, push held (§10.3)`, terminal: "candidate_only", candidatePatch: heldCandidate() }, measure };
     }
     return { measure };
 }
 async function resolveApplyConflict(env, repo, p, rejText, applyStderr, resolver, deps) {
     let costUsd = 0;
-    const conflictText = `git apply --3way could not cleanly apply worker '${p.workerId}'.\n--- apply stderr ---\n${applyStderr.slice(0, 1500)}\n--- reject hunks (*.rej) ---\n${rejText.slice(0, 3000)}`;
+    const conflictText = `git apply --3way could not cleanly apply worker '${p.workerId}'.\n--- apply stderr ---\n${redactHead(applyStderr, 1500)}\n--- reject hunks (*.rej) ---\n${redactHead(rejText, 3000)}`;
     for (let round = 0; round < resolver.maxRounds; round++) {
         deps.logger?.warn?.("merge_conflict_repair_round", { workerId: p.workerId, round: round + 1, maxRounds: resolver.maxRounds });
         try {
@@ -175,7 +180,7 @@ async function resolveApplyConflict(env, repo, p, rejText, applyStderr, resolver
             costUsd += r.costUsd ?? 0;
         }
         catch (e) {
-            deps.logger?.warn?.("merge_conflict_repair_error", { workerId: p.workerId, err: e instanceof Error ? e.message : String(e) });
+            deps.logger?.warn?.("merge_conflict_repair_error", { workerId: p.workerId, err: redactHead(e instanceof Error ? e.message : String(e), 2000) });
             return { resolved: false, costUsd };
         }
         const check = await env

package/dist/leader/planner.js

@@ -1,3 +1,4 @@
+import { redactErrorMessage, redactHead, redactTail } from "../observability/run-terminal-log.js";
 export function economicGuard(objective, opts = {}) {
     if (opts.fanoutEnabled === false)
         return { single: true, reason: "fan-out disabled by deployment (A6)" };
@@ -124,13 +125,13 @@ export async function routePlanWithFallback(objective, runRoute, opts = {}) {
         catch (e) {
             firstErr ??= e;
             if (!firstRaw)
-                firstRaw = raw.length > 600 ? `${raw.slice(0, 300)}…[${raw.length}ch]…${raw.slice(-150)}` : raw;
+                firstRaw = raw.length > 600 ? `${redactHead(raw, 300)}…[${raw.length}ch]…${redactTail(raw, 150)}` : (redactErrorMessage(raw) ?? "");
         }
     }
     const reason = firstErr instanceof Error ? firstErr.message : String(firstErr);
     return {
         subs: singleRoute(objective),
-        collapsed: `plan rejected twice → collapse to single: ${reason.slice(0, 200)}`,
+        collapsed: `plan rejected twice → collapse to single: ${redactHead(reason, 200)}`,
         rawSample: firstRaw,
     };
 }

package/dist/leader/repair-oracle.js

@@ -1,6 +1,7 @@
 import { runExecGate, } from "@sema-agent/core";
 import { createHash } from "node:crypto";
 import { shellQuote as sq } from "../plugins/remote-shell.js";
+import { redactHead, redactTail } from "../observability/run-terminal-log.js";
 const sha = (s) => createHash("sha256").update(s).digest("hex").slice(0, 16);
 function changedFilesOf(patch) {
     const out = new Set();
@@ -28,8 +29,9 @@ export function mkRepairOracle(cfg) {
     return async (graderEnv, _evidence) => {
         const diff = await cfg.pullWorkerDiff();
         if (!diff.ok) {
-            cfg.log?.("repair_oracle_pull_failed", { error: diff.error });
-            return { tier: cfg.passTier, passed: false, flaky: false, retries: 0, trace: `pull worker diff failed: ${diff.error}` };
+            const safeErr = redactHead(diff.error ?? "", 2000);
+            cfg.log?.("repair_oracle_pull_failed", { error: safeErr });
+            return { tier: cfg.passTier, passed: false, flaky: false, retries: 0, trace: `pull worker diff failed: ${safeErr}` };
         }
         const diffHash = sha(diff.patch);
         const tampered = cfg.oraclePaths && cfg.oraclePaths.length > 0 ? touchesOraclePath(changedFilesOf(diff.patch), cfg.oraclePaths) : undefined;
@@ -51,7 +53,7 @@ export function mkRepairOracle(cfg) {
             { command: `git -C ${sq(repo)} apply --whitespace=nowarn ${sq(patchPath)}`, label: "grader-apply" },
         ]);
         if (!prep.passed) {
-            const trace = prep.steps.map((s) => `${s.label}: ${s.stderr}`).join("\n").slice(-2000);
+            const trace = redactTail(prep.steps.map((s) => `${s.label}: ${s.stderr}`).join("\n"), 2000);
             cfg.log?.("repair_oracle_prep_failed", { trace });
             cfg.onGraded?.({ diff: diff.patch, hash: diffHash, passed: false });
             return { tier: cfg.passTier, passed: false, flaky: false, retries: 0, trace: `grader prep failed:\n${trace}` };
@@ -64,7 +66,7 @@ export function mkRepairOracle(cfg) {
             const gate = await runExecGate(graderEnv, cfg.oracleSteps);
             verdicts.push(gate.passed);
             if (!gate.passed) {
-                lastTrace = gate.steps.map((s) => `${s.label}\n${s.stdout}\n${s.stderr}`).join("\n").slice(-4000);
+                lastTrace = redactTail(gate.steps.map((s) => `${s.label}\n${s.stdout}\n${s.stderr}`).join("\n"), 4000);
                 break;
             }
         }

package/dist/leader/repair-wire.js

@@ -1,5 +1,6 @@
 import { runRepairLoop, delimitUntrusted, } from "@sema-agent/core";
 import { mkRepairOracle } from "./repair-oracle.js";
+import { redactHead } from "../observability/run-terminal-log.js";
 function statusForTerminal(t) {
     return t === "fixed" || t === "candidate_only" ? "completed" : "failed";
 }
@@ -41,12 +42,13 @@ async function runRepairLoopForLeaderInner(solo, implSpec, cfg) {
             ...(cfg.onAttempt
                 ? { onAttempt: (i) => cfg.onAttempt({ attempt: i.attempt, terminal: i.terminal, oracleTier: i.oracle.tier, passed: i.oracle.passed, flaky: i.oracle.flaky }) }
                 : {}),
-            onWarn: (e) => cfg.log?.("repair_loop_warn", { warning: e.message }),
+            onWarn: (e) => cfg.log?.("repair_loop_warn", { warning: redactHead(e.message, 2000) }),
         });
     }
     catch (e) {
-        cfg.log?.("repair_loop_threw", { error: e instanceof Error ? e.message : String(e) });
-        return { report: { ...base, status: "failed", error: e instanceof Error ? e.message : String(e) }, result: undefined };
+        const safe = redactHead(e instanceof Error ? e.message : String(e), 2000);
+        cfg.log?.("repair_loop_threw", { error: safe });
+        return { report: { ...base, status: "failed", error: safe }, result: undefined };
     }
     const terminal = result.terminal;
     const report = {

package/dist/leader/wire.d.ts

@@ -1,4 +1,4 @@
-import { type Brain, type Model, type ModelRoles, type ModelPricing, type TaskSpec, type ExecutionEnvFactory, type RemoteExecutionEnv, type ToolResultStore, type SessionStore, type RunnerDeps } from "@sema-agent/core";
+import { type Brain, type Model, type ModelRoles, type ModelPricing, type TaskSpec, type ExecutionEnvFactory, type RemoteExecutionEnv, type ToolResultStore, type SessionStore, type ExecutionEnv, type RunnerDeps } from "@sema-agent/core";
 import type { CheckpointStoreFull } from "../plugins/store-backend.js";
 import { type LeaderDeps, type LeaderResult } from "./leader.js";
 import { type DeploymentPostureSeats } from "../boot/runner-deps.js";
@@ -146,6 +146,35 @@ export interface LeaderWireConfig {
 }
 /** The env surface wire needs (both the E2B and k8s adapters implement core's RemoteExecutionEnv). */
 type WireEnv = Pick<RemoteExecutionEnv, "exec" | "writeFile">;
+/**
+ * [ref] —— leader 自铸沙盒的**回收闸**(集成沙盒 / repair-loop grader 各一只)。
+ *
+ * core 契约:`ExecutionEnv` 基面**没有** `destroy`;只有实现了 `RemoteExecutionEnv` 的适配器才有
+ * (`core/remote-env.d.ts:426`),而「这只 env 到底有没有」是**运行期结构事实**,core 自己也只用结构探测
+ * `hasDestroy()`(:487,runtask/prepare-task 全部 destroy 站点都先过它)。修前 leader 三处把它写成编译期
+ * 断言(`as … & { destroy(): Promise<unknown> }`),于是 TS 对「手里其实是个 **Promise**」完全无感——
+ * `env.destroy()` 抛 `env.destroy is not a function`,还把**原始错误盖掉**([ref] 现场)。
+ *
+ * 🔴 缺席臂 = **fail-closed,当场响亮拒**([ref]:执行车道禁静默兜底;codex r2-[medium] 采纳)。理由:
+ *  · core 之所以容忍「工厂产物没有 destroy」,是因为工厂**可以**交回一只**共享静态** env(`withWorktreeIsolation`
+ *    的非隔离臂就是这一形)——那种 env 本来就不该被谁拆掉。但 leader 这两只是**专用、临时、只属于它自己**
+ *    的沙盒:唯一的拆机口就是本闭包(`mergeBranches` / `runRepairLoopForLeader` 在 `finally` 里调它)。
+ *    收不动 = 一只计费中的沙盒必然留到后端自己的 timeout,零上行收益 ⇒ 没有「合法兜底」可言。
+ *  · 静默 no-op 反而比**修前**更松:修前同步工厂路径上,一只没有 destroy 的 env 会当场抛 TypeError。
+ *    本闸只是把那个抛**提前到 provision 时刻**,并换成一句点名沙盒与成因的话——不是新增严格度。
+ *  · 当前三条已接线的 leader 车道交回的 env 都实现了 destroy(e2b `RemoteContainerExecutionEnv`、
+ *    k8s、host `RemoteHostExecutionEnv`——`hostExecutionEnvFactory` 恒 `new RemoteHostExecutionEnv`),
+ *    所以这条臂是**防御性**的:它真响的那天,意思是有人接了一条新工厂车道而 leader 收不动它的沙盒。
+ */
+/** 冲突解决模型的目标文本——纯函数(第十三轮重扫 [medium]:wire 内联时无法钉「补丁先脱」;抽出来直接钉)。
+ *  补丁走补丁头窗(头 256KB 扫再切 6000;解决者仍有工作树里的冲突标记),conflictText 由 merge.ts 先脱后切。 */
+export declare function buildConflictResolverObjective(args: {
+    workerId: string;
+    repoDir: string;
+    patch: string;
+    conflictText: string;
+}): string;
+export declare function createEnvReaper(env: ExecutionEnv, what: string): () => Promise<void>;
 /**
  * Resume-safe worker staging (the [ref] F2(a) factory-env path). Runs adapter-side before the agent's
  * first workspace-touching call: seed the base repo at the worker branch + plant the upload script (so the

package/dist/leader/wire.js

@@ -3,7 +3,7 @@ import { execFileSync } from "node:child_process";
 import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { tmpdir } from "node:os";
-import { Runner, combinePolicies, createDurableQuestionPolicy, QUESTION_AWAITS_RESUME, delimitUntrusted, } from "@sema-agent/core";
+import { Runner, combinePolicies, createDurableQuestionPolicy, QUESTION_AWAITS_RESUME, delimitUntrusted, hasDestroy, } from "@sema-agent/core";
 import { RemoteContainerExecutionEnv } from "../plugins/remote-env-e2b.js";
 import { createDurableAskPolicy } from "../approval.js";
 import { presignS3Url } from "../plugins/s3-presign.js";
@@ -14,6 +14,7 @@ import { attachRepairLoopDeps } from "./repair-wire.js";
 import { routePlanWithFallback, validateSubtasks } from "./planner.js";
 import { createEngineNoticeSeat } from "../boot/runner-deps.js";
 import { currentPrincipal } from "../observability/principal-context.js";
+import { redactHead, redactPatchHead } from "../observability/run-terminal-log.js";
 const REPAIR_LOOP_WIRED = true;
 const sh = (env) => async (cmd) => {
     const r = await env.exec(`bash -lc ${JSON.stringify(cmd)}`);
@@ -31,6 +32,24 @@ const asIntegrationEnv = (env) => ({
     writeFile: async (p, c) => { const r = await env.writeFile(p, c); if (!r.ok)
         throw r.error; },
 });
+export function buildConflictResolverObjective(args) {
+    const { workerId, repoDir, patch, conflictText } = args;
+    return [
+        `A parallel worker '${workerId}' ported a module on its own branch, but applying its patch onto the already-integrated tree at ${repoDir} produced a MERGE CONFLICT (git apply --3way). The OTHER workers' patches already applied cleanly — integrate THIS worker's changes too, resolving the overlap.`,
+        `Resolve EVERY conflict in the working tree: open each file containing conflict markers (<<<<<<< / ======= / >>>>>>>) and merge BOTH sides' real intent (keep both workers' behaviour — never drop one side just to make it apply). Then apply any rejected hunks recorded in *.rej files by hand, and DELETE every *.rej and *.orig file.`,
+        `When done, NO conflict markers and NO *.rej files may remain under ${repoDir}. Verify yourself before finishing: grep -rIlE '^(<<<<<<<|>>>>>>>)' . ; find . -name '*.rej'  — BOTH must print nothing.`,
+        `HARD RULES: do NOT git commit / git add / git push (the coordinator squashes at the end); do NOT modify build configuration; do NOT call AskUserQuestion — decide autonomously. Preserve both sides' ported semantics.`,
+        ``,
+        delimitUntrusted("CONFLICTING WORKER PATCH (untrusted)", redactPatchHead(patch, 6000)),
+        delimitUntrusted("CONFLICT DETAILS (untrusted)", conflictText),
+    ].join("\n");
+}
+export function createEnvReaper(env, what) {
+    if (!hasDestroy(env)) {
+        throw new Error(`${what}: the execution-env factory returned an env with no destroy() — the leader OWNS this dedicated sandbox and this closure is its only teardown, so it would leak (and keep billing) until the backend's own timeout reaps it. Deploy a factory whose env implements RemoteExecutionEnv.destroy for the leader lane.`);
+    }
+    return async () => void (await env.destroy());
+}
 export async function stageWorkerEnv(env, opts) {
     const { seedCmd, repo, branch, uploadScriptPath, uploadScript } = opts;
     const probe = await env.exec(`bash -lc ${JSON.stringify(`if [ -f ${uploadScriptPath} ]; then exit 0; fi; ` +
@@ -197,15 +216,7 @@ export function createLeaderRunner(cfg) {
                 }
                 try {
                     const runner = new Runner({ brain: cfg.brain, models: cfg.models, roles: cfg.roles, pricing: cfg.pricing, ...onNoticeSeat, ...governanceSeat, ...deploymentPostureSeat, ...tracerSeat, ...usageWindowSeat, ...promptSourceSeat, ...onErrorSeat, executionEnv: rawEnv, rootPath: repoDir });
-                    const objective = [
-                        `A parallel worker '${workerId}' ported a module on its own branch, but applying its patch onto the already-integrated tree at ${repoDir} produced a MERGE CONFLICT (git apply --3way). The OTHER workers' patches already applied cleanly — integrate THIS worker's changes too, resolving the overlap.`,
-                        `Resolve EVERY conflict in the working tree: open each file containing conflict markers (<<<<<<< / ======= / >>>>>>>) and merge BOTH sides' real intent (keep both workers' behaviour — never drop one side just to make it apply). Then apply any rejected hunks recorded in *.rej files by hand, and DELETE every *.rej and *.orig file.`,
-                        `When done, NO conflict markers and NO *.rej files may remain under ${repoDir}. Verify yourself before finishing: grep -rIlE '^(<<<<<<<|>>>>>>>)' . ; find . -name '*.rej'  — BOTH must print nothing.`,
-                        `HARD RULES: do NOT git commit / git add / git push (the coordinator squashes at the end); do NOT modify build configuration; do NOT call AskUserQuestion — decide autonomously. Preserve both sides' ported semantics.`,
-                        ``,
-                        delimitUntrusted("CONFLICTING WORKER PATCH (untrusted)", patch.slice(0, 6000)),
-                        delimitUntrusted("CONFLICT DETAILS (untrusted)", conflictText),
-                    ].join("\n");
+                    const objective = buildConflictResolverObjective({ workerId, repoDir, patch, conflictText });
                     const res = await runner
                         .runTaskStream({
                         objective,
@@ -290,7 +301,7 @@ export function createLeaderRunner(cfg) {
                     onQuestion: QUESTION_AWAITS_RESUME,
                 };
                 await d.checkpointStore.putCtx(sessionId, { body: { objective: sub.spec.objective } }).catch((e) => {
-                    cfg.logger?.warn?.("leader_putctx_failed", { workerId: sub.workerId, err: String(e) });
+                    cfg.logger?.warn?.("leader_putctx_failed", { workerId: sub.workerId, err: redactHead(e instanceof Error ? e.message : String(e), 2000) });
                 });
             }
             let resourceSpec = {};
@@ -354,7 +365,7 @@ export function createLeaderRunner(cfg) {
                         brain: cfg.brain, models: cfg.models, roles: cfg.roles, pricing: cfg.pricing, ...onNoticeSeat, ...governanceSeat, ...deploymentPostureSeat, ...tracerSeat, ...usageWindowSeat, ...promptSourceSeat, ...onErrorSeat,
                         ...(cfg.toolResultStore ? { toolResultStore: cfg.toolResultStore } : {}),
                         ...(cfg.sessionStore ? { sessionStore: cfg.sessionStore } : {}),
-                        executionEnvFactory: (ctx) => withStaging(envFactory(ctx), stage, async (e) => {
+                        executionEnvFactory: async (ctx) => withStaging(await envFactory(ctx), stage, async (e) => {
                             await sh(e)(`cd ${repo} && git add -A && git -c user.name='leader-belt' -c user.email='belt@leader' commit -qm 'belt: auto-commit uncommitted worker output' || true; sh ${W}/.leader-upload.sh`);
                         }),
                     })),
@@ -380,16 +391,21 @@ export function createLeaderRunner(cfg) {
         const provisionIntegrationSandbox = async () => {
             if (cfg.envFactory) {
                 const integSessionId = `leader-integ-${Date.now()}`;
-                const env = cfg.envFactory({ sessionId: integSessionId, placementRootSessionId: integSessionId });
+                const env = await cfg.envFactory({ sessionId: integSessionId, placementRootSessionId: integSessionId });
+                const reap = createEnvReaper(env, "leader integration sandbox");
                 try {
                     await sh(env)(body.seedCmd);
                     await injectOracles(env);
                 }
                 catch (e) {
-                    await Promise.resolve(env.destroy()).catch(() => { });
+                    await reap().catch(() => { });
                     throw e;
                 }
-                return { env: asIntegrationEnv(env), destroy: () => Promise.resolve(env.destroy()).then(() => { }), ...(mkRepair(env, repo, body.oracleFiles ?? []) ? { repair: mkRepair(env, repo, body.oracleFiles ?? []) } : {}), ...(mkConflictResolver(env, repo, body.oracleFiles ?? []) ? { conflictResolver: mkConflictResolver(env, repo, body.oracleFiles ?? []) } : {}) };
+                const rawEnv = env;
+                const oracleFiles = body.oracleFiles ?? [];
+                const repair = mkRepair(rawEnv, repo, oracleFiles);
+                const conflictResolver = mkConflictResolver(rawEnv, repo, oracleFiles);
+                return { env: asIntegrationEnv(env), destroy: reap, ...(repair ? { repair } : {}), ...(conflictResolver ? { conflictResolver } : {}) };
             }
             if (!cfg.e2bApiKey)
                 throw new Error("leader wire: no envFactory and no e2bApiKey");
@@ -427,7 +443,7 @@ export function createLeaderRunner(cfg) {
             let env;
             if (cfg.envFactory) {
                 const graderSessionId = `leader-grader-${Date.now()}`;
-                env = cfg.envFactory({ sessionId: graderSessionId, placementRootSessionId: graderSessionId });
+                env = await cfg.envFactory({ sessionId: graderSessionId, placementRootSessionId: graderSessionId });
             }
             else {
                 if (!cfg.e2bApiKey)
@@ -439,6 +455,7 @@ export function createLeaderRunner(cfg) {
                 }
                 env = e2b;
             }
+            const reapGrader = createEnvReaper(env, "leader repair-loop grader");
             let integBase;
             try {
                 await sh(env)(body.seedCmd);
@@ -446,11 +463,11 @@ export function createLeaderRunner(cfg) {
                 integBase = await sh(env)(`cd ${repo} && git rev-parse HEAD`);
             }
             catch (e) {
-                await Promise.resolve(env.destroy()).catch(() => { });
+                await reapGrader().catch(() => { });
                 throw e;
             }
             const grader = {
-                env: env,
+                env,
                 writeFile: async (p, c) => { const r = await env.writeFile(p, c); if (!r.ok)
                     throw r.error; },
                 repoDir: repo,
@@ -475,7 +492,7 @@ export function createLeaderRunner(cfg) {
             return {
                 runner: solo.runner,
                 grader,
-                graderDestroy: () => Promise.resolve(env.destroy()).then(() => { }),
+                graderDestroy: reapGrader,
                 pullWorkerDiff,
                 oracleSteps,
                 passTier,