llm-cli-gateway 2.14.1 → 2.16.0

package/dist/index.js

@@ -3,7 +3,7 @@ import { McpServer, ResourceTemplate } from "@modelcontextprotocol/sdk/server/mc
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 import { randomUUID } from "crypto";
 import { createRequire } from "module";
-import { existsSync, mkdirSync, readFileSync, readdirSync, renameSync, unlinkSync, writeFileSync, chmodSync, } from "fs";
+import { existsSync, mkdirSync, readFileSync, renameSync, unlinkSync, writeFileSync, chmodSync, } from "fs";
 import { basename, dirname, isAbsolute, join, relative } from "path";
 import { fileURLToPath } from "url";
 import { z } from "zod/v3";
@@ -21,7 +21,9 @@ import { ResourceProvider } from "./resources.js";
 import { generateResourceDescriptors, generateProviderAcpDescriptors, } from "./provider-surface-generator.js";
 import { PerformanceMetrics } from "./metrics.js";
 import { estimateTokens, optimizePrompt as optimizePromptText, optimizeResponse as optimizeResponseText, } from "./optimizer.js";
-import { loadConfig, loadPersistenceConfig, loadCacheAwarenessConfig, loadProvidersConfig, loadAcpConfig, loadAdminConfig, loadLimitsConfig, defaultGatewayConfigPath, isXaiProviderEnabled, enabledApiProviders, minStableTokensForModel, } from "./config.js";
+import { compressDisplayText } from "./compressor/index.js";
+import { loadConfig, loadPersistenceConfig, loadCacheAwarenessConfig, loadProvidersConfig, loadAcpConfig, loadAdminConfig, loadLimitsConfig, loadCompressionConfig, loadSkillsConfig, defaultGatewayConfigPath, isXaiProviderEnabled, enabledApiProviders, minStableTokensForModel, } from "./config.js";
+import { loadGatewaySkills } from "./skill-loader.js";
 import { runAcpRequest } from "./acp/runtime.js";
 import { isAcpError } from "./acp/errors.js";
 import { redactSecrets } from "./secret-redaction.js";
@@ -184,27 +186,13 @@ function packageVersion() {
     }
     return "unknown";
 }
-function loadSkills() {
-    const skills = [];
-    try {
-        const dirs = readdirSync(SKILLS_DIR, { withFileTypes: true }).filter(d => d.isDirectory());
-        for (const dir of dirs) {
-            const skillPath = join(SKILLS_DIR, dir.name, "SKILL.md");
-            try {
-                const content = readFileSync(skillPath, "utf-8");
-                const descMatch = content.match(/^---[\s\S]*?description:\s*(.+?)$/m);
-                const description = descMatch?.[1]?.trim() || dir.name;
-                skills.push({ name: dir.name, content, description });
-            }
-            catch {
-            }
-        }
-    }
-    catch {
-    }
-    return skills;
-}
-const loadedSkills = loadSkills();
+const skillsConfig = loadSkillsConfig(logger);
+const loadedSkills = loadGatewaySkills({
+    bundledSkillsDir: SKILLS_DIR,
+    configuredPaths: skillsConfig.paths,
+    envSkillsPath: undefined,
+    logger,
+});
 export function buildServerInstructions(asyncJobsEnabled, grokApiToolsEnabled = false) {
     const asyncToolsNote = asyncJobsEnabled ? " | *_request_async (async)" : "";
     const apiToolsNote = grokApiToolsEnabled ? ", grok_api_request" : "";
@@ -241,6 +229,7 @@ let resourceProvider;
 let flightRecorder = null;
 let persistenceConfig = null;
 let cacheAwarenessConfig = null;
+let compressionConfig = null;
 let providersConfig = null;
 let acpConfig = null;
 let adminConfig = null;
@@ -273,6 +262,10 @@ function getCacheAwarenessConfig(runtimeLogger = logger) {
     cacheAwarenessConfig ??= loadCacheAwarenessConfig(runtimeLogger);
     return cacheAwarenessConfig;
 }
+function getCompressionConfig(runtimeLogger = logger) {
+    compressionConfig ??= loadCompressionConfig(runtimeLogger);
+    return compressionConfig;
+}
 function getProvidersConfig(runtimeLogger = logger) {
     providersConfig ??= loadProvidersConfig(runtimeLogger);
     return providersConfig;
@@ -481,6 +474,7 @@ export function resolveGatewayServerRuntime(deps = {}, options = {}) {
         logger: runtimeLogger,
         persistence: deps.persistence ?? getPersistenceConfig(runtimeLogger),
         cacheAwareness: deps.cacheAwareness ?? getCacheAwarenessConfig(runtimeLogger),
+        compression: deps.compression ?? getCompressionConfig(runtimeLogger),
         providers: deps.providers ?? getProvidersConfig(runtimeLogger),
         acpConfig: deps.acpConfig ?? getAcpConfig(runtimeLogger),
         adminConfig: deps.adminConfig ?? getAdminConfig(runtimeLogger),
@@ -558,7 +552,7 @@ export async function runAcpTransport(deps, params) {
     }
 }
 const SYNC_POLL_INTERVAL_MS = 1_000;
-async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, forceRefresh, runtime = resolveGatewayServerRuntime(), env, onComplete, flightRecorderEntry, extractUsage, stdin, cwd) {
+async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, forceRefresh, runtime = resolveGatewayServerRuntime(), env, onComplete, flightRecorderEntry, extractUsage, stdin, cwd, compressResponse) {
     let onCompleteOwnedByCaller = onComplete !== undefined;
     const consumeOnComplete = () => {
         if (!onCompleteOwnedByCaller || !onComplete)
@@ -618,6 +612,7 @@ async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, f
             onComplete,
             flightRecorderEntry,
             extractUsage,
+            compressResponse,
         });
         onCompleteOwnedByCaller = false;
     }
@@ -1179,7 +1174,7 @@ export function extractUsageAndCost(cli, output, outputFormat, ctx) {
     }
     return {};
 }
-function buildAsyncFlightRecorderHandoff(cliName, prep, sessionId, outputFormat) {
+function buildAsyncFlightRecorderHandoff(cliName, prep, sessionId, outputFormat, optimizationApplied = false) {
     const cli = cliName;
     const fmt = outputFormat;
     const sid = sessionId;
@@ -1189,6 +1184,7 @@ function buildAsyncFlightRecorderHandoff(cliName, prep, sessionId, outputFormat)
             model: prep.resolvedModel || "default",
             prompt: prep.effectivePrompt,
             sessionId,
+            optimizationApplied,
             stablePrefixHash: prep.stablePrefixHash ?? undefined,
             stablePrefixTokens: prep.stablePrefixTokens ?? undefined,
             cacheControlBlocks: prep.cacheControlBlocks,
@@ -1213,6 +1209,32 @@ function safeFlightComplete(correlationId, result, runtime = resolveGatewayServe
         runtime.logger.error("Flight recorder logComplete failed", error);
     }
 }
+export function resolveEffectiveCompression(compression, input) {
+    const requested = input.compressResponse ?? compression.enabled;
+    if (!requested)
+        return false;
+    if (input.outputFormat === "json")
+        return false;
+    if (input.outputSchemaDeclared)
+        return false;
+    return true;
+}
+function safeRecordCompression(correlationId, compression, runtime) {
+    if (!compression)
+        return;
+    try {
+        runtime.flightRecorder.recordCompressionTelemetry(correlationId, {
+            route: compression.route,
+            transforms: compression.transforms,
+            originalChars: compression.originalChars,
+            compressedChars: compression.compressedChars,
+            estimatedTokensSaved: compression.estimatedTokensSaved,
+        });
+    }
+    catch (error) {
+        runtime.logger.error("Flight recorder recordCompressionTelemetry failed", error);
+    }
+}
 function createApprovalDeniedResponse(operation, decision) {
     return {
         content: [
@@ -2223,6 +2245,14 @@ export function prepareGrokRequest(params, runtime = resolveGatewayServerRuntime
         stablePrefixTokens,
     };
 }
+export function resolveMistralAgentMode(approvalStrategy, callerMode) {
+    if (approvalStrategy === "mcp_managed") {
+        return bypassAllowedByOperator() ? "auto-approve" : "accept-edits";
+    }
+    if (callerMode !== undefined)
+        return callerMode;
+    return bypassAllowedByOperator() ? "auto-approve" : "accept-edits";
+}
 export function prepareMistralRequest(params, runtime = resolveGatewayServerRuntime()) {
     const corrId = params.correlationId || randomUUID();
     const cliInfo = getCliInfo();
@@ -2279,11 +2309,7 @@ export function prepareMistralRequest(params, runtime = resolveGatewayServerRunt
             return createApprovalDeniedResponse(params.operation, approvalDecision);
         }
     }
-    const effectivePermissionMode = params.approvalStrategy === "mcp_managed"
-        ? bypassAllowedByOperator()
-            ? "auto-approve"
-            : "accept-edits"
-        : (params.permissionMode ?? "auto-approve");
+    const effectivePermissionMode = resolveMistralAgentMode(params.approvalStrategy, params.permissionMode);
     const prep = buildMistralCliInvocation({
         prompt: effectivePrompt,
         resolvedModel,
@@ -2332,11 +2358,7 @@ export function buildMistralRetryPrep(params, recoveryModel) {
         prompt: params.effectivePrompt,
         resolvedModel: recoveryModel,
         outputFormat: params.outputFormat,
-        permissionMode: params.approvalStrategy === "mcp_managed"
-            ? bypassAllowedByOperator()
-                ? "auto-approve"
-                : "accept-edits"
-            : (params.permissionMode ?? "auto-approve"),
+        permissionMode: resolveMistralAgentMode(params.approvalStrategy, params.permissionMode),
         allowedTools: params.allowedTools,
         disallowedTools: params.disallowedTools,
         trust: params.trust,
@@ -2347,7 +2369,7 @@ export function buildMistralRetryPrep(params, recoveryModel) {
         addDir: params.addDir,
     });
 }
-export function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionId, prep, durationMs, resumable, outputFormat, warnings) {
+export function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionId, prep, durationMs, resumable, outputFormat, warnings, compressResponse = false) {
     let finalStdout = stdout;
     if (cli === "codex" && outputFormat !== "json") {
         finalStdout = codexDisplayText(stdout);
@@ -2360,6 +2382,20 @@ export function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionI
         logOptimizationTokens("response", corrId, finalStdout, optimized);
         finalStdout = optimized;
     }
+    let compression;
+    if (compressResponse && outputFormat !== "json") {
+        const compressed = compressDisplayText(finalStdout, {
+            provider: cli,
+            direction: "outbound",
+            outputFormat,
+            outputSchemaDeclared: false,
+            lossless: true,
+        });
+        if (compressed.text !== finalStdout) {
+            compression = compressed;
+            finalStdout = compressed.text;
+        }
+    }
     if (prep.reviewIntegrity &&
         prep.reviewIntegrity.violations.length > 0 &&
         outputFormat !== "json") {
@@ -2435,6 +2471,7 @@ export function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionI
                 missing: prep.mcpConfig.missing,
             }
             : { requested: prep.requestedMcpServers },
+        ...(compression ? { compression } : {}),
     };
     if (sessionId) {
         response.sessionId = sessionId;
@@ -3188,8 +3225,13 @@ export async function handleGeminiRequest(deps, params) {
         catch (err) {
             return createErrorResponse("gemini_request", 1, "", corrId, err);
         }
-        const geminiFrHandoff = buildAsyncFlightRecorderHandoff("gemini", prep, params.sessionId, params.outputFormat);
-        const result = await awaitJobOrDefer("gemini", args, corrId, resolveIdleTimeout("gemini", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, undefined, undefined, geminiFrHandoff.flightRecorderEntry, geminiFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const geminiFrHandoff = buildAsyncFlightRecorderHandoff("gemini", prep, params.sessionId, params.outputFormat, params.optimizePrompt);
+        const result = await awaitJobOrDefer("gemini", args, corrId, resolveIdleTimeout("gemini", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, undefined, undefined, geminiFrHandoff.flightRecorderEntry, geminiFrHandoff.extractUsage, undefined, worktreeResolution.cwd, effectiveCompress);
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, effectiveSessionIdHint);
         }
@@ -3226,7 +3268,8 @@ export async function handleGeminiRequest(deps, params) {
             await deps.sessionManager.updateSessionUsage(effectiveSessionId);
         }
         deps.logger.info(`[${corrId}] gemini_request completed successfully in ${durationMs}ms`);
-        const response = buildCliResponse("gemini", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, userProvidedSession, params.outputFormat);
+        const response = buildCliResponse("gemini", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, userProvidedSession, params.outputFormat, undefined, effectiveCompress);
+        safeRecordCompression(corrId, response.compression, runtime);
         if (worktreeResolution.worktreePath) {
             const first = response.content[0];
             if (first && first.type === "text") {
@@ -3341,8 +3384,13 @@ export async function handleGeminiRequestAsync(deps, params) {
         }
         assertUpstreamCliArgs("gemini", args);
         assertUpstreamCliEnv("gemini", undefined);
-        const geminiAsyncFrHandoff = buildAsyncFlightRecorderHandoff("gemini", prep, effectiveSessionId, params.outputFormat);
-        const job = deps.asyncJobManager.startJob("gemini", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("gemini", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, undefined, geminiAsyncFrHandoff.flightRecorderEntry, geminiAsyncFrHandoff.extractUsage, true);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const geminiAsyncFrHandoff = buildAsyncFlightRecorderHandoff("gemini", prep, effectiveSessionId, params.outputFormat, params.optimizePrompt);
+        const job = deps.asyncJobManager.startJob("gemini", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("gemini", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, undefined, geminiAsyncFrHandoff.flightRecorderEntry, geminiAsyncFrHandoff.extractUsage, true, undefined, effectiveCompress);
         deps.logger.info(`[${corrId}] gemini_request_async started job ${job.id}`);
         const asyncResponse = {
             success: true,
@@ -3471,8 +3519,13 @@ export async function handleGrokRequest(deps, params) {
         catch (err) {
             return createErrorResponse("grok_request", 1, "", corrId, err);
         }
-        const grokFrHandoff = buildAsyncFlightRecorderHandoff("grok", prep, params.sessionId, params.outputFormat);
-        const result = await awaitJobOrDefer("grok", args, corrId, resolveIdleTimeout("grok", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, undefined, undefined, grokFrHandoff.flightRecorderEntry, grokFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const grokFrHandoff = buildAsyncFlightRecorderHandoff("grok", prep, params.sessionId, params.outputFormat, params.optimizePrompt);
+        const result = await awaitJobOrDefer("grok", args, corrId, resolveIdleTimeout("grok", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, undefined, undefined, grokFrHandoff.flightRecorderEntry, grokFrHandoff.extractUsage, undefined, worktreeResolution.cwd, effectiveCompress);
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
         }
@@ -3513,7 +3566,8 @@ export async function handleGrokRequest(deps, params) {
             effectiveSessionId = newSession.id;
         }
         deps.logger.info(`[${corrId}] grok_request completed successfully in ${durationMs}ms`);
-        const response = buildCliResponse("grok", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, params.outputFormat);
+        const response = buildCliResponse("grok", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, params.outputFormat, undefined, effectiveCompress);
+        safeRecordCompression(corrId, response.compression, runtime);
         if (worktreeResolution.worktreePath) {
             const first = response.content[0];
             if (first && first.type === "text") {
@@ -3654,8 +3708,13 @@ export async function handleGrokRequestAsync(deps, params) {
         }
         assertUpstreamCliArgs("grok", args);
         assertUpstreamCliEnv("grok", undefined);
-        const grokAsyncFrHandoff = buildAsyncFlightRecorderHandoff("grok", prep, effectiveSessionId, params.outputFormat);
-        const job = deps.asyncJobManager.startJob("grok", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("grok", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, undefined, grokAsyncFrHandoff.flightRecorderEntry, grokAsyncFrHandoff.extractUsage, true);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const grokAsyncFrHandoff = buildAsyncFlightRecorderHandoff("grok", prep, effectiveSessionId, params.outputFormat, params.optimizePrompt);
+        const job = deps.asyncJobManager.startJob("grok", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("grok", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, undefined, grokAsyncFrHandoff.flightRecorderEntry, grokAsyncFrHandoff.extractUsage, true, undefined, effectiveCompress);
         deps.logger.info(`[${corrId}] grok_request_async started job ${job.id}`);
         const asyncResponse = {
             success: true,
@@ -3783,8 +3842,13 @@ export async function handleDevinRequest(deps, params) {
             await getExistingSessionForProvider(deps.sessionManager, sessionResult.effectiveSessionId, "devin");
         }
         args.push(...sessionResult.resumeArgs);
-        const devinFrHandoff = buildAsyncFlightRecorderHandoff("devin", prep, params.sessionId, undefined);
-        const result = await awaitJobOrDefer("devin", args, corrId, resolveIdleTimeout("devin", params.idleTimeoutMs), undefined, params.forceRefresh, runtime, undefined, undefined, devinFrHandoff.flightRecorderEntry, devinFrHandoff.extractUsage);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: undefined,
+            outputSchemaDeclared: false,
+        });
+        const devinFrHandoff = buildAsyncFlightRecorderHandoff("devin", prep, params.sessionId, undefined, params.optimizePrompt);
+        const result = await awaitJobOrDefer("devin", args, corrId, resolveIdleTimeout("devin", params.idleTimeoutMs), undefined, params.forceRefresh, runtime, undefined, undefined, devinFrHandoff.flightRecorderEntry, devinFrHandoff.extractUsage, undefined, undefined, effectiveCompress);
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
         }
@@ -3823,7 +3887,8 @@ export async function handleDevinRequest(deps, params) {
             const newSession = await deps.sessionManager.createSession("devin", "Devin Session", `${GATEWAY_SESSION_PREFIX}${randomUUID()}`);
             effectiveSessionId = newSession.id;
         }
-        const response = buildCliResponse("devin", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession);
+        const response = buildCliResponse("devin", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, undefined, undefined, effectiveCompress);
+        safeRecordCompression(corrId, response.compression, runtime);
         safeFlightComplete(corrId, {
             response: stdout,
             durationMs,
@@ -3903,8 +3968,13 @@ export async function handleDevinRequestAsync(deps, params) {
         }
         assertUpstreamCliArgs("devin", args);
         assertUpstreamCliEnv("devin", undefined);
-        const devinAsyncFrHandoff = buildAsyncFlightRecorderHandoff("devin", prep, effectiveSessionId, undefined);
-        const job = deps.asyncJobManager.startJob("devin", args, corrId, undefined, resolveIdleTimeout("devin", params.idleTimeoutMs), undefined, params.forceRefresh, undefined, undefined, devinAsyncFrHandoff.flightRecorderEntry, devinAsyncFrHandoff.extractUsage, true);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: undefined,
+            outputSchemaDeclared: false,
+        });
+        const devinAsyncFrHandoff = buildAsyncFlightRecorderHandoff("devin", prep, effectiveSessionId, undefined, params.optimizePrompt);
+        const job = deps.asyncJobManager.startJob("devin", args, corrId, undefined, resolveIdleTimeout("devin", params.idleTimeoutMs), undefined, params.forceRefresh, undefined, undefined, devinAsyncFrHandoff.flightRecorderEntry, devinAsyncFrHandoff.extractUsage, true, undefined, effectiveCompress);
         deps.logger.info(`[${corrId}] devin_request_async started job ${job.id}`);
         return {
             content: [
@@ -4102,8 +4172,13 @@ export async function handleCursorRequest(deps, params) {
                 args[workspaceFlagIndex + 1] = worktreeResolution.workspace.cwd;
             }
         }
-        const cursorFrHandoff = buildAsyncFlightRecorderHandoff("cursor", prep, params.sessionId, params.outputFormat);
-        const result = await awaitJobOrDefer("cursor", args, corrId, resolveIdleTimeout("cursor", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, undefined, undefined, cursorFrHandoff.flightRecorderEntry, cursorFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const cursorFrHandoff = buildAsyncFlightRecorderHandoff("cursor", prep, params.sessionId, params.outputFormat, params.optimizePrompt);
+        const result = await awaitJobOrDefer("cursor", args, corrId, resolveIdleTimeout("cursor", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, undefined, undefined, cursorFrHandoff.flightRecorderEntry, cursorFrHandoff.extractUsage, undefined, worktreeResolution.cwd, effectiveCompress);
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
         }
@@ -4142,7 +4217,8 @@ export async function handleCursorRequest(deps, params) {
             const newSession = await deps.sessionManager.createSession("cursor", "Cursor Session", `${GATEWAY_SESSION_PREFIX}${randomUUID()}`);
             effectiveSessionId = newSession.id;
         }
-        const response = buildCliResponse("cursor", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, params.outputFormat);
+        const response = buildCliResponse("cursor", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, params.outputFormat, undefined, effectiveCompress);
+        safeRecordCompression(corrId, response.compression, runtime);
         safeFlightComplete(corrId, {
             response: stdout,
             durationMs,
@@ -4244,8 +4320,13 @@ export async function handleCursorRequestAsync(deps, params) {
         }
         assertUpstreamCliArgs("cursor", args);
         assertUpstreamCliEnv("cursor", undefined);
-        const cursorAsyncFrHandoff = buildAsyncFlightRecorderHandoff("cursor", prep, effectiveSessionId, params.outputFormat);
-        const job = deps.asyncJobManager.startJob("cursor", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("cursor", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, undefined, cursorAsyncFrHandoff.flightRecorderEntry, cursorAsyncFrHandoff.extractUsage, true);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const cursorAsyncFrHandoff = buildAsyncFlightRecorderHandoff("cursor", prep, effectiveSessionId, params.outputFormat, params.optimizePrompt);
+        const job = deps.asyncJobManager.startJob("cursor", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("cursor", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, undefined, cursorAsyncFrHandoff.flightRecorderEntry, cursorAsyncFrHandoff.extractUsage, true, undefined, effectiveCompress);
         deps.logger.info(`[${corrId}] cursor_request_async started job ${job.id}`);
         return {
             content: [
@@ -4312,7 +4393,7 @@ export async function handleMistralRequest(deps, params) {
         stablePrefixHash: prep.stablePrefixHash ?? undefined,
         stablePrefixTokens: prep.stablePrefixTokens ?? undefined,
     }, runtime);
-    deps.logger.info(`[${corrId}] mistral_request invoked with model=${prep.resolvedModel || "default"}, permissionMode=${params.permissionMode || "auto-approve"}, prompt length=${prep.effectivePrompt.length}`);
+    deps.logger.info(`[${corrId}] mistral_request invoked with model=${prep.resolvedModel || "default"}, permissionMode=${resolveMistralAgentMode(params.approvalStrategy, params.permissionMode)}, prompt length=${prep.effectivePrompt.length}`);
     try {
         const sessionResult = resolveMistralSessionArgs({
             sessionId: params.sessionId,
@@ -4338,8 +4419,13 @@ export async function handleMistralRequest(deps, params) {
         catch (err) {
             return createErrorResponse("mistral_request", 1, "", corrId, err);
         }
-        const mistralFrHandoff = buildAsyncFlightRecorderHandoff("mistral", prep, params.sessionId, params.outputFormat);
-        let result = await awaitJobOrDefer("mistral", args, corrId, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, mistralEnv, undefined, mistralFrHandoff.flightRecorderEntry, mistralFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const mistralFrHandoff = buildAsyncFlightRecorderHandoff("mistral", prep, params.sessionId, params.outputFormat, params.optimizePrompt);
+        let result = await awaitJobOrDefer("mistral", args, corrId, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, runtime, mistralEnv, undefined, mistralFrHandoff.flightRecorderEntry, mistralFrHandoff.extractUsage, undefined, worktreeResolution.cwd, effectiveCompress);
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
         }
@@ -4349,7 +4435,7 @@ export async function handleMistralRequest(deps, params) {
                 deps.logger.info(`[${corrId}] mistral_request detected stale Vibe model selection; retrying once with ${recoveryModel}`);
                 const retryPrep = buildMistralRetryPrep({ ...params, effectivePrompt: prep.effectivePrompt }, recoveryModel);
                 const retryArgs = [...retryPrep.args, ...sessionResult.resumeArgs];
-                result = await awaitJobOrDefer("mistral", retryArgs, corrId, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, true, runtime, retryPrep.env, undefined, mistralFrHandoff.flightRecorderEntry, mistralFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
+                result = await awaitJobOrDefer("mistral", retryArgs, corrId, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, true, runtime, retryPrep.env, undefined, mistralFrHandoff.flightRecorderEntry, mistralFrHandoff.extractUsage, undefined, worktreeResolution.cwd, effectiveCompress);
                 if (isDeferredResponse(result)) {
                     return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
                 }
@@ -4394,7 +4480,8 @@ export async function handleMistralRequest(deps, params) {
             effectiveSessionId = newSession.id;
         }
         deps.logger.info(`[${corrId}] mistral_request completed successfully in ${durationMs}ms`);
-        const response = buildCliResponse("mistral", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, params.outputFormat);
+        const response = buildCliResponse("mistral", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession, params.outputFormat, undefined, effectiveCompress);
+        safeRecordCompression(corrId, response.compression, runtime);
         if (worktreeResolution.worktreePath) {
             const first = response.content[0];
             if (first && first.type === "text") {
@@ -4502,8 +4589,13 @@ export async function handleMistralRequestAsync(deps, params) {
         }
         assertUpstreamCliArgs("mistral", args);
         assertUpstreamCliEnv("mistral", mistralEnv);
-        const mistralAsyncFrHandoff = buildAsyncFlightRecorderHandoff("mistral", prep, effectiveSessionId, params.outputFormat);
-        const job = deps.asyncJobManager.startJob("mistral", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, mistralEnv, undefined, mistralAsyncFrHandoff.flightRecorderEntry, mistralAsyncFrHandoff.extractUsage, true);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: false,
+        });
+        const mistralAsyncFrHandoff = buildAsyncFlightRecorderHandoff("mistral", prep, effectiveSessionId, params.outputFormat, params.optimizePrompt);
+        const job = deps.asyncJobManager.startJob("mistral", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, mistralEnv, undefined, mistralAsyncFrHandoff.flightRecorderEntry, mistralAsyncFrHandoff.extractUsage, true, undefined, effectiveCompress);
         deps.logger.info(`[${corrId}] mistral_request_async started job ${job.id}`);
         const asyncResponse = {
             success: true,
@@ -4632,10 +4724,15 @@ export async function handleCodexRequestAsync(deps, params) {
         }
         assertUpstreamCliArgs("codex", args);
         assertUpstreamCliEnv("codex", undefined);
-        const codexAsyncFrHandoff = buildAsyncFlightRecorderHandoff("codex", prep, effectiveSessionId, params.outputFormat);
+        const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+            compressResponse: params.compressResponse,
+            outputFormat: params.outputFormat,
+            outputSchemaDeclared: params.outputSchema !== undefined,
+        });
+        const codexAsyncFrHandoff = buildAsyncFlightRecorderHandoff("codex", prep, effectiveSessionId, params.outputFormat, params.optimizePrompt);
         let job;
         try {
-            job = deps.asyncJobManager.startJob("codex", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("codex", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, prepCleanup, codexAsyncFrHandoff.flightRecorderEntry, codexAsyncFrHandoff.extractUsage, true);
+            job = deps.asyncJobManager.startJob("codex", args, corrId, worktreeResolution.cwd, resolveIdleTimeout("codex", params.idleTimeoutMs), params.outputFormat, params.forceRefresh, undefined, prepCleanup, codexAsyncFrHandoff.flightRecorderEntry, codexAsyncFrHandoff.extractUsage, true, undefined, effectiveCompress);
             prepCleanupOwnedHere = false;
         }
         catch (startErr) {
@@ -4896,6 +4993,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -4913,7 +5014,7 @@ export function createGatewayServer(deps = {}) {
         destructiveHint: true,
         idempotentHint: false,
         openWorldHint: true,
-    }, async ({ prompt, promptParts, model, outputFormat, sessionId, continueSession, createNewSession, allowedTools, disallowedTools, dangerouslySkipPermissions, permissionMode, agent, agents, forkSession, systemPrompt, appendSystemPrompt, maxBudgetUsd, maxTurns, effort, excludeDynamicSystemPromptSections, fallbackModel, jsonSchema, addDir, noSessionPersistence, settingSources, settings, tools, includeHookEvents, replayUserMessages, systemPromptFile, appendSystemPromptFile, name, pluginDir, pluginUrl, safeMode, bare, debug, debugFile, workspace, worktree, approvalStrategy, approvalPolicy, mcpServers, strictMcpConfig, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, }) => {
+    }, async ({ prompt, promptParts, model, outputFormat, sessionId, continueSession, createNewSession, allowedTools, disallowedTools, dangerouslySkipPermissions, permissionMode, agent, agents, forkSession, systemPrompt, appendSystemPrompt, maxBudgetUsd, maxTurns, effort, excludeDynamicSystemPromptSections, fallbackModel, jsonSchema, addDir, noSessionPersistence, settingSources, settings, tools, includeHookEvents, replayUserMessages, systemPromptFile, appendSystemPromptFile, name, pluginDir, pluginUrl, safeMode, bare, debug, debugFile, workspace, worktree, approvalStrategy, approvalPolicy, mcpServers, strictMcpConfig, correlationId, optimizePrompt, optimizeResponse, compressResponse, idleTimeoutMs, forceRefresh, }) => {
         const startTime = Date.now();
         if (systemPrompt !== undefined && appendSystemPrompt !== undefined) {
             return createErrorResponse("claude", 1, "", correlationId, new Error("systemPrompt and appendSystemPrompt are mutually exclusive; use one or the other (not both)."));
@@ -5033,8 +5134,13 @@ export function createGatewayServer(deps = {}) {
                 return createErrorResponse("claude_request", 1, "", corrId, err);
             }
             const effectiveIdleTimeout = outputFormat === "stream-json" ? resolveIdleTimeout("claude", idleTimeoutMs) : undefined;
-            const claudeSyncFrHandoff = buildAsyncFlightRecorderHandoff("claude", prep, effectiveSessionId, outputFormat);
-            const result = await awaitJobOrDefer("claude", args, corrId, effectiveIdleTimeout, outputFormat, forceRefresh, runtime, undefined, undefined, claudeSyncFrHandoff.flightRecorderEntry, claudeSyncFrHandoff.extractUsage, prep.stdinPayload, worktreeResolution.cwd);
+            const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+                compressResponse,
+                outputFormat,
+                outputSchemaDeclared: false,
+            });
+            const claudeSyncFrHandoff = buildAsyncFlightRecorderHandoff("claude", prep, effectiveSessionId, outputFormat, optimizePrompt);
+            const result = await awaitJobOrDefer("claude", args, corrId, effectiveIdleTimeout, outputFormat, forceRefresh, runtime, undefined, undefined, claudeSyncFrHandoff.flightRecorderEntry, claudeSyncFrHandoff.extractUsage, prep.stdinPayload, worktreeResolution.cwd, effectiveCompress);
             if (isDeferredResponse(result)) {
                 const deferred = buildDeferredToolResponse(result, effectiveSessionId);
                 if (warnings.length > 0) {
@@ -5091,7 +5197,8 @@ export function createGatewayServer(deps = {}) {
                     providerSessionId: parsed.sessionId ?? undefined,
                     stopReason: parsed.stopReason ?? undefined,
                 }, runtime);
-                const streamResponse = buildCliResponse("claude", parsed.text, optimizeResponse, corrId, effectiveSessionId, prep, durationMs, undefined, outputFormat, warnings);
+                const streamResponse = buildCliResponse("claude", parsed.text, optimizeResponse, corrId, effectiveSessionId, prep, durationMs, undefined, outputFormat, warnings, effectiveCompress);
+                safeRecordCompression(corrId, streamResponse.compression, runtime);
                 if (worktreeResolution.worktreePath) {
                     const first = streamResponse.content[0];
                     if (first && first.type === "text") {
@@ -5112,7 +5219,8 @@ export function createGatewayServer(deps = {}) {
                 providerSessionId: claudeMeta.sessionId,
                 stopReason: claudeMeta.stopReason,
             }, runtime);
-            const nonStreamResponse = buildCliResponse("claude", stdout, optimizeResponse, corrId, effectiveSessionId, prep, durationMs, undefined, outputFormat, warnings);
+            const nonStreamResponse = buildCliResponse("claude", stdout, optimizeResponse, corrId, effectiveSessionId, prep, durationMs, undefined, outputFormat, warnings, effectiveCompress);
+            safeRecordCompression(corrId, nonStreamResponse.compression, runtime);
             if (worktreeResolution.worktreePath) {
                 const first = nonStreamResponse.content[0];
                 if (first && first.type === "text") {
@@ -5194,6 +5302,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -5258,7 +5370,7 @@ export function createGatewayServer(deps = {}) {
         destructiveHint: true,
         idempotentHint: false,
         openWorldHint: true,
-    }, async ({ prompt, promptParts, model, fullAuto, sandboxMode, askForApproval, useLegacyFullAutoFlag, dangerouslyBypassApprovalsAndSandbox, approvalStrategy, approvalPolicy, mcpServers, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, outputFormat, outputSchema, search, profile, configOverrides, ephemeral, images, ignoreUserConfig, ignoreRules, workingDir, addDir, enable, disable, strictConfig, oss, localProvider, color, outputLastMessage, dangerouslyBypassHookTrust, workspace, worktree, }) => {
+    }, async ({ prompt, promptParts, model, fullAuto, sandboxMode, askForApproval, useLegacyFullAutoFlag, dangerouslyBypassApprovalsAndSandbox, approvalStrategy, approvalPolicy, mcpServers, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, optimizeResponse, compressResponse, idleTimeoutMs, forceRefresh, outputFormat, outputSchema, search, profile, configOverrides, ephemeral, images, ignoreUserConfig, ignoreRules, workingDir, addDir, enable, disable, strictConfig, oss, localProvider, color, outputLastMessage, dangerouslyBypassHookTrust, workspace, worktree, }) => {
         const startTime = Date.now();
         const prep = prepareCodexRequest({
             prompt,
@@ -5336,8 +5448,13 @@ export function createGatewayServer(deps = {}) {
             return createErrorResponse("codex_request", 1, "", corrId, err);
         }
         try {
-            const codexSyncFrHandoff = buildAsyncFlightRecorderHandoff("codex", prep, sessionId, outputFormat);
-            const result = await awaitJobOrDefer("codex", args, corrId, resolveIdleTimeout("codex", idleTimeoutMs), outputFormat, forceRefresh, runtime, undefined, prepCleanup, codexSyncFrHandoff.flightRecorderEntry, codexSyncFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
+            const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+                compressResponse,
+                outputFormat,
+                outputSchemaDeclared: outputSchema !== undefined,
+            });
+            const codexSyncFrHandoff = buildAsyncFlightRecorderHandoff("codex", prep, sessionId, outputFormat, optimizePrompt);
+            const result = await awaitJobOrDefer("codex", args, corrId, resolveIdleTimeout("codex", idleTimeoutMs), outputFormat, forceRefresh, runtime, undefined, prepCleanup, codexSyncFrHandoff.flightRecorderEntry, codexSyncFrHandoff.extractUsage, undefined, worktreeResolution.cwd, effectiveCompress);
             if (isDeferredResponse(result)) {
                 return buildDeferredToolResponse(result, sessionId);
             }
@@ -5405,7 +5522,8 @@ export function createGatewayServer(deps = {}) {
                 providerSessionId: codexMeta.sessionId,
                 stopReason: codexMeta.stopReason,
             }, runtime);
-            const codexResponse = buildCliResponse("codex", stdout, optimizeResponse, corrId, effectiveSessionId, prep, durationMs, undefined, outputFormat);
+            const codexResponse = buildCliResponse("codex", stdout, optimizeResponse, corrId, effectiveSessionId, prep, durationMs, undefined, outputFormat, undefined, effectiveCompress);
+            safeRecordCompression(corrId, codexResponse.compression, runtime);
             if (worktreeResolution.worktreePath) {
                 const first = codexResponse.content[0];
                 if (first && first.type === "text") {
@@ -5598,6 +5716,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -5726,6 +5848,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -5882,6 +6008,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -5982,6 +6112,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -6024,7 +6158,7 @@ export function createGatewayServer(deps = {}) {
             forceRefresh,
         });
     });
-    server.tool("mistral_request", "Run a Mistral Vibe CLI request synchronously (when async jobs are enabled, auto-defers to a pollable job past the sync deadline; otherwise runs to completion). Requires exactly one of prompt or promptParts. Defaults to --agent auto-approve (unattended tool execution); pass permissionMode plan or accept-edits for safer runs.", {
+    server.tool("mistral_request", "Run a Mistral Vibe CLI request synchronously (when async jobs are enabled, auto-defers to a pollable job past the sync deadline; otherwise runs to completion). Requires exactly one of prompt or promptParts. Defaults to --agent accept-edits (auto-accepts file edits; dangerous ops such as shell stay gated); pass permissionMode auto-approve (or set LLM_GATEWAY_APPROVAL_ALLOW_BYPASS) for unattended tool execution.", {
         prompt: z
             .string()
             .min(1, "Prompt cannot be empty")
@@ -6056,7 +6190,7 @@ export function createGatewayServer(deps = {}) {
         permissionMode: z
             .string()
             .optional()
-            .describe("Vibe --agent name. Builtins: default|plan|accept-edits|auto-approve; Vibe also accepts install-gated builtins (e.g. lean) and custom agents from ~/.vibe/agents, so any name is passed through. Defaults to auto-approve for programmatic use."),
+            .describe("Vibe --agent name. Builtins: default|plan|accept-edits|auto-approve; Vibe also accepts install-gated builtins (e.g. lean) and custom agents from ~/.vibe/agents, so any name is passed through. Defaults to accept-edits; pass auto-approve (or set LLM_GATEWAY_APPROVAL_ALLOW_BYPASS) for unattended tool execution."),
         approvalStrategy: z
             .enum(["legacy", "mcp_managed"])
             .default("legacy")
@@ -6080,6 +6214,10 @@ export function createGatewayServer(deps = {}) {
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
         optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        compressResponse: z
+            .boolean()
+            .optional()
+            .describe("Compress the response display text via the native compressor (default: [compression].enabled in config.toml, off unless opted in). Skipped for structured output."),
         idleTimeoutMs: z
             .number()
             .int()
@@ -6276,6 +6414,10 @@ export function createGatewayServer(deps = {}) {
                 .describe("Restrict Claude to provided MCP config only"),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -6293,7 +6435,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, promptParts, model, outputFormat, sessionId, continueSession, createNewSession, allowedTools, disallowedTools, dangerouslySkipPermissions, permissionMode, agent, agents, forkSession, systemPrompt, appendSystemPrompt, maxBudgetUsd, maxTurns, effort, excludeDynamicSystemPromptSections, fallbackModel, jsonSchema, addDir, noSessionPersistence, settingSources, settings, tools, includeHookEvents, replayUserMessages, systemPromptFile, appendSystemPromptFile, name, pluginDir, pluginUrl, safeMode, bare, debug, debugFile, workspace, worktree, approvalStrategy, approvalPolicy, mcpServers, strictMcpConfig, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, }) => {
+        }, async ({ prompt, promptParts, model, outputFormat, sessionId, continueSession, createNewSession, allowedTools, disallowedTools, dangerouslySkipPermissions, permissionMode, agent, agents, forkSession, systemPrompt, appendSystemPrompt, maxBudgetUsd, maxTurns, effort, excludeDynamicSystemPromptSections, fallbackModel, jsonSchema, addDir, noSessionPersistence, settingSources, settings, tools, includeHookEvents, replayUserMessages, systemPromptFile, appendSystemPromptFile, name, pluginDir, pluginUrl, safeMode, bare, debug, debugFile, workspace, worktree, approvalStrategy, approvalPolicy, mcpServers, strictMcpConfig, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, }) => {
             if (systemPrompt !== undefined && appendSystemPrompt !== undefined) {
                 return createErrorResponse("claude", 1, "", correlationId, new Error("systemPrompt and appendSystemPrompt are mutually exclusive; use one or the other (not both)."));
             }
@@ -6392,8 +6534,13 @@ export function createGatewayServer(deps = {}) {
                     : undefined;
                 assertUpstreamCliArgs("claude", args);
                 assertUpstreamCliEnv("claude", undefined);
-                const claudeAsyncFrHandoff = buildAsyncFlightRecorderHandoff("claude", prep, effectiveSessionId, outputFormat);
-                const job = asyncJobManager.startJob("claude", args, corrId, worktreeResolution.cwd, effectiveIdleTimeout, outputFormat, forceRefresh, undefined, undefined, claudeAsyncFrHandoff.flightRecorderEntry, claudeAsyncFrHandoff.extractUsage, true, prep.stdinPayload);
+                const effectiveCompress = resolveEffectiveCompression(runtime.compression, {
+                    compressResponse,
+                    outputFormat,
+                    outputSchemaDeclared: false,
+                });
+                const claudeAsyncFrHandoff = buildAsyncFlightRecorderHandoff("claude", prep, effectiveSessionId, outputFormat, optimizePrompt);
+                const job = asyncJobManager.startJob("claude", args, corrId, worktreeResolution.cwd, effectiveIdleTimeout, outputFormat, forceRefresh, undefined, undefined, claudeAsyncFrHandoff.flightRecorderEntry, claudeAsyncFrHandoff.extractUsage, true, prep.stdinPayload, effectiveCompress);
                 logger.info(`[${corrId}] claude_request_async started job ${job.id}, outputFormat=${outputFormat}`);
                 const asyncResponse = {
                     success: true,
@@ -6487,6 +6634,10 @@ export function createGatewayServer(deps = {}) {
             createNewSession: z.boolean().default(false).describe("Force a fresh session (no resume)"),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -6536,7 +6687,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, promptParts, model, fullAuto, sandboxMode, askForApproval, useLegacyFullAutoFlag, dangerouslyBypassApprovalsAndSandbox, approvalStrategy, approvalPolicy, mcpServers, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, outputFormat, outputSchema, search, profile, configOverrides, ephemeral, images, ignoreUserConfig, ignoreRules, workingDir, addDir, enable, disable, strictConfig, oss, localProvider, color, outputLastMessage, dangerouslyBypassHookTrust, workspace, worktree, }) => {
+        }, async ({ prompt, promptParts, model, fullAuto, sandboxMode, askForApproval, useLegacyFullAutoFlag, dangerouslyBypassApprovalsAndSandbox, approvalStrategy, approvalPolicy, mcpServers, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, outputFormat, outputSchema, search, profile, configOverrides, ephemeral, images, ignoreUserConfig, ignoreRules, workingDir, addDir, enable, disable, strictConfig, oss, localProvider, color, outputLastMessage, dangerouslyBypassHookTrust, workspace, worktree, }) => {
             return handleCodexRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -6554,6 +6705,7 @@ export function createGatewayServer(deps = {}) {
                 createNewSession,
                 correlationId,
                 optimizePrompt,
+                compressResponse,
                 idleTimeoutMs,
                 forceRefresh,
                 outputFormat,
@@ -6627,6 +6779,10 @@ export function createGatewayServer(deps = {}) {
                 LOCAL_INCLUDE_DIRS_FIELD_SUFFIX),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -6676,7 +6832,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, promptParts, model, sessionId, resumeLatest, createNewSession, approvalMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, includeDirs, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, outputFormat, sandbox, policyFiles, adminPolicyFiles, attachments, skipTrust, yolo, project, newProject, printTimeout, workspace, worktree, }) => {
+        }, async ({ prompt, promptParts, model, sessionId, resumeLatest, createNewSession, approvalMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, includeDirs, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, outputFormat, sandbox, policyFiles, adminPolicyFiles, attachments, skipTrust, yolo, project, newProject, printTimeout, workspace, worktree, }) => {
             return handleGeminiRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -6692,6 +6848,7 @@ export function createGatewayServer(deps = {}) {
                 includeDirs,
                 correlationId,
                 optimizePrompt,
+                compressResponse,
                 idleTimeoutMs,
                 forceRefresh,
                 outputFormat,
@@ -6765,6 +6922,10 @@ export function createGatewayServer(deps = {}) {
                 .describe("Disallowed built-in tools (passed as --disallowed-tools comma list)"),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -6906,7 +7067,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, alwaysApprove, permissionMode, effort, reasoningEffort, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, maxTurns, workingDir, sandbox, rules, systemPromptOverride, allow, deny, compactionMode, compactionDetail, agent, bestOfN, check, disableWebSearch, todoGate, verbatim, agents, promptFile, promptJson, single, experimentalMemory, noAltScreen, noMemory, noPlan, noSubagents, oauth, restoreCode, leaderSocket, nativeWorktree, worktreeRef, forkSession, jsonSchema, workspace, worktree, }) => {
+        }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, alwaysApprove, permissionMode, effort, reasoningEffort, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, maxTurns, workingDir, sandbox, rules, systemPromptOverride, allow, deny, compactionMode, compactionDetail, agent, bestOfN, check, disableWebSearch, todoGate, verbatim, agents, promptFile, promptJson, single, experimentalMemory, noAltScreen, noMemory, noPlan, noSubagents, oauth, restoreCode, leaderSocket, nativeWorktree, worktreeRef, forkSession, jsonSchema, workspace, worktree, }) => {
             return handleGrokRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -6926,6 +7087,7 @@ export function createGatewayServer(deps = {}) {
                 disallowedTools,
                 correlationId,
                 optimizePrompt,
+                compressResponse,
                 idleTimeoutMs,
                 forceRefresh,
                 maxTurns,
@@ -7007,6 +7169,10 @@ export function createGatewayServer(deps = {}) {
             createNewSession: z.boolean().default(false).describe("Force a new session"),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -7024,7 +7190,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, model, permissionMode, promptFile, config, sandbox, exportSession, respectWorkspaceTrust, agentConfig, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, }) => {
+        }, async ({ prompt, model, permissionMode, promptFile, config, sandbox, exportSession, respectWorkspaceTrust, agentConfig, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, }) => {
             return handleDevinRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 model,
@@ -7040,6 +7206,7 @@ export function createGatewayServer(deps = {}) {
                 createNewSession,
                 correlationId,
                 optimizePrompt,
+                compressResponse,
                 idleTimeoutMs,
                 forceRefresh,
             });
@@ -7099,6 +7266,10 @@ export function createGatewayServer(deps = {}) {
                 .describe("Approval policy when approvalStrategy is mcp_managed: strict|balanced|permissive (default balanced). Ignored under legacy strategy."),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -7116,7 +7287,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, model, mode, outputFormat, force, autoReview, sandbox, trust, workspace, addDir, sessionId, resumeLatest, createNewSession, approvalStrategy, approvalPolicy, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, }) => {
+        }, async ({ prompt, model, mode, outputFormat, force, autoReview, sandbox, trust, workspace, addDir, sessionId, resumeLatest, createNewSession, approvalStrategy, approvalPolicy, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, }) => {
             return handleCursorRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 model,
@@ -7135,6 +7306,7 @@ export function createGatewayServer(deps = {}) {
                 approvalPolicy,
                 correlationId,
                 optimizePrompt,
+                compressResponse,
                 idleTimeoutMs,
                 forceRefresh,
             });
@@ -7167,7 +7339,7 @@ export function createGatewayServer(deps = {}) {
             permissionMode: z
                 .string()
                 .optional()
-                .describe("Vibe --agent name. Builtins: default|plan|accept-edits|auto-approve; Vibe also accepts install-gated builtins (e.g. lean) and custom agents from ~/.vibe/agents, so any name is passed through. Defaults to auto-approve for programmatic use."),
+                .describe("Vibe --agent name. Builtins: default|plan|accept-edits|auto-approve; Vibe also accepts install-gated builtins (e.g. lean) and custom agents from ~/.vibe/agents, so any name is passed through. Defaults to accept-edits; pass auto-approve (or set LLM_GATEWAY_APPROVAL_ALLOW_BYPASS) for unattended tool execution."),
             approvalStrategy: z
                 .enum(["legacy", "mcp_managed"])
                 .default("legacy")
@@ -7190,6 +7362,10 @@ export function createGatewayServer(deps = {}) {
                 .describe("Accepted for caller parity; Vibe has no deny-list flag, so values are ignored (a warning is logged)."),
             correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
             optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
+            compressResponse: z
+                .boolean()
+                .optional()
+                .describe("Compress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled)."),
             idleTimeoutMs: z
                 .number()
                 .int()
@@ -7227,7 +7403,7 @@ export function createGatewayServer(deps = {}) {
             destructiveHint: true,
             idempotentHint: false,
             openWorldHint: true,
-        }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, permissionMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, workspace, worktree, }) => {
+        }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, permissionMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, compressResponse, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, workspace, worktree, }) => {
             return handleMistralRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -7244,6 +7420,7 @@ export function createGatewayServer(deps = {}) {
                 disallowedTools,
                 correlationId,
                 optimizePrompt,
+                compressResponse,
                 idleTimeoutMs,
                 forceRefresh,
                 trust,
@@ -7337,6 +7514,23 @@ export function createGatewayServer(deps = {}) {
                 result.stdout) {
                 result.stdout = codexDisplayText(result.stdout);
             }
+            const compressJob = asyncJobManager.getJobCompressResponse(jobId) && outputFormat !== "json";
+            if (compressJob && result.stdout) {
+                if (outputFormat === "stream-json" && parsed) {
+                    result.stdout = parsed.text;
+                }
+                const compressed = compressDisplayText(result.stdout, {
+                    provider: asyncJobManager.getJobCli(jobId) ?? "unknown",
+                    direction: "outbound",
+                    outputFormat,
+                    outputSchemaDeclared: false,
+                    lossless: true,
+                });
+                if (compressed.text !== result.stdout) {
+                    result.stdout = compressed.text;
+                    safeRecordCompression(result.correlationId, compressed, runtime);
+                }
+            }
             if (callerIsRemote()) {
                 const leakedId = result.providerSessionId;
                 delete result.providerSessionId;
@@ -7359,7 +7553,7 @@ export function createGatewayServer(deps = {}) {
                             ...(parsed
                                 ? {
                                     parsed: {
-                                        text: parsed.text,
+                                        ...(compressJob ? {} : { text: parsed.text }),
                                         costUsd: parsed.costUsd,
                                         usage: parsed.usage,
                                         model: parsed.model,
@@ -7367,7 +7561,7 @@ export function createGatewayServer(deps = {}) {
                                     },
                                 }
                                 : {}),
-                        }, null, 2),
+                        }, null, compressJob ? 0 : 2),
                     },
                 ],
             };