llm-cli-gateway 2.10.0 → 2.11.1

package/dist/index.js

@@ -18,7 +18,11 @@ import { createWorktree, createWorktreeSessionCleanupHook, } from "./worktree-ma
 import { ResourceProvider } from "./resources.js";
 import { PerformanceMetrics } from "./metrics.js";
 import { estimateTokens, optimizePrompt as optimizePromptText, optimizeResponse as optimizeResponseText, } from "./optimizer.js";
-import { loadConfig, loadPersistenceConfig, loadCacheAwarenessConfig, loadProvidersConfig, defaultGatewayConfigPath, isXaiProviderEnabled, minStableTokensForModel, } from "./config.js";
+import { loadConfig, loadPersistenceConfig, loadCacheAwarenessConfig, loadProvidersConfig, loadAcpConfig, defaultGatewayConfigPath, isXaiProviderEnabled, enabledApiProviders, minStableTokensForModel, } from "./config.js";
+import { runAcpRequest } from "./acp/runtime.js";
+import { isAcpError } from "./acp/errors.js";
+import { createApiProvider, runApiRequest, apiProviderBreakerState, } from "./api-provider.js";
+import { prepareApiRequest, apiProviderCatalogEntry, ApiModelNotAllowedError, } from "./api-request.js";
 import { createXaiResponse, XaiApiError, } from "./xai-api-provider.js";
 import { checkHealth } from "./health.js";
 import { clearModelRegistryCache, getAvailableCliInfo, getCliInfo, resolveModelAlias, } from "./model-registry.js";
@@ -28,7 +32,7 @@ import { createJobStore } from "./job-store.js";
 import { ApprovalManager, bypassAllowedByOperator, } from "./approval-manager.js";
 import { checkReviewIntegrity } from "./review-integrity.js";
 import { buildClaudeMcpConfig, CLAUDE_MCP_SERVER_NAMES, } from "./claude-mcp-config.js";
-import { resolveGrokSessionArgs, resolveMistralSessionArgs, resolveCodexSessionArgs, sanitizeCliArgValues, prepareMistralRequest as buildMistralCliInvocation, MISTRAL_AGENT_MODES, GATEWAY_SESSION_PREFIX, resolveClaudePermissionFlags, resolveCodexSandboxFlags, CLAUDE_PERMISSION_MODES, GEMINI_APPROVAL_MODES, CODEX_SANDBOX_MODES, CODEX_ASK_FOR_APPROVAL_MODES, CLAUDE_EFFORT_LEVELS, prepareClaudeHighImpactFlags, validateClaudeAgentsMap, prepareCodexHighImpactFlags, prepareCodexForkRequest, CODEX_CONFIG_OVERRIDES_SCHEMA, resolveGeminiSessionPlan, GEMINI_HIGH_IMPACT_PARAMS_SCHEMA, } from "./request-helpers.js";
+import { resolveGrokSessionArgs, resolveMistralSessionArgs, resolveCodexSessionArgs, sanitizeCliArgValues, prepareMistralRequest as buildMistralCliInvocation, GATEWAY_SESSION_PREFIX, resolveClaudePermissionFlags, resolveCodexSandboxFlags, CLAUDE_PERMISSION_MODES, GEMINI_APPROVAL_MODES, CODEX_SANDBOX_MODES, CODEX_ASK_FOR_APPROVAL_MODES, CLAUDE_EFFORT_LEVELS, prepareClaudeHighImpactFlags, validateClaudeAgentsMap, prepareCodexHighImpactFlags, prepareCodexForkRequest, CODEX_CONFIG_OVERRIDES_SCHEMA, resolveGeminiSessionPlan, GEMINI_HIGH_IMPACT_PARAMS_SCHEMA, } from "./request-helpers.js";
 import { createFlightRecorder } from "./flight-recorder.js";
 import { resolvePromptInput, PromptPartsSchema, assembleClaudeCacheBlocks, } from "./prompt-parts.js";
 import { computeSessionCacheStats, computeTtlRemaining, readPersistedRequest, PERSISTED_REQUEST_DEFAULT_MAX_CHARS, } from "./cache-stats.js";
@@ -183,6 +187,7 @@ let flightRecorder = null;
 let persistenceConfig = null;
 let cacheAwarenessConfig = null;
 let providersConfig = null;
+let acpConfig = null;
 let jobStore = null;
 let jobStoreInitialized = false;
 let asyncJobManager = null;
@@ -195,6 +200,10 @@ function getPersistenceConfig(runtimeLogger = logger) {
     persistenceConfig ??= loadPersistenceConfig(runtimeLogger);
     return persistenceConfig;
 }
+function getAcpConfig(runtimeLogger = logger) {
+    acpConfig ??= loadAcpConfig(runtimeLogger);
+    return acpConfig;
+}
 function getCacheAwarenessConfig(runtimeLogger = logger) {
     cacheAwarenessConfig ??= loadCacheAwarenessConfig(runtimeLogger);
     return cacheAwarenessConfig;
@@ -229,7 +238,11 @@ function getApprovalManager(runtimeLogger = logger) {
     approvalManager ??= new ApprovalManager(undefined, runtimeLogger);
     return approvalManager;
 }
-const MCP_SERVER_ENUM = z.enum(CLAUDE_MCP_SERVER_NAMES);
+function mcpServerEnum() {
+    return CLAUDE_MCP_SERVER_NAMES.length > 0
+        ? z.enum(CLAUDE_MCP_SERVER_NAMES)
+        : z.string();
+}
 const CLI_TYPE_ENUM = z.enum(CLI_TYPES);
 export const MAX_TURNS_SCHEMA = z.number().int().positive().safe().max(10_000);
 const GROK_GENERATED_SHAPE = deriveZodShapeFromGeneration(UPSTREAM_CLI_CONTRACTS.grok, GROK_FLAG_GENERATION);
@@ -299,6 +312,7 @@ export function resolveGatewayServerRuntime(deps = {}, options = {}) {
         persistence: deps.persistence ?? getPersistenceConfig(runtimeLogger),
         cacheAwareness: deps.cacheAwareness ?? getCacheAwarenessConfig(runtimeLogger),
         providers: deps.providers ?? getProvidersConfig(runtimeLogger),
+        acpConfig: deps.acpConfig ?? getAcpConfig(runtimeLogger),
         workspaces: deps.workspaces ?? loadWorkspaceRegistry(runtimeLogger),
     };
 }
@@ -317,6 +331,44 @@ function resolveIdleTimeout(cli, override) {
         return override;
     return CLI_IDLE_TIMEOUTS[cli];
 }
+export async function runAcpTransport(deps, params) {
+    const runtime = resolveHandlerRuntime(deps);
+    const operation = `${params.provider}_request`;
+    const corrId = params.correlationId ?? randomUUID();
+    const prompt = (params.prompt ?? "").trim();
+    if (!prompt) {
+        return createErrorResponse(operation, 1, "prompt is required and cannot be empty", corrId);
+    }
+    try {
+        const result = await runAcpRequest({
+            config: runtime.acpConfig,
+            sessionManager: runtime.sessionManager,
+            approvalManager: runtime.approvalManager,
+            flightRecorder: runtime.flightRecorder,
+            logger: runtime.logger,
+        }, {
+            provider: params.provider,
+            prompt,
+            model: params.model,
+            sessionId: params.sessionId,
+            correlationId: corrId,
+        });
+        return {
+            content: [
+                {
+                    type: "text",
+                    text: `[gateway] transport=acp session=${result.gatewaySessionId}\n${result.text}`,
+                },
+            ],
+        };
+    }
+    catch (err) {
+        if (isAcpError(err)) {
+            return createErrorResponse(operation, 1, err.userMessage, corrId);
+        }
+        return createErrorResponse(operation, 1, "", corrId, err);
+    }
+}
 const SYNC_POLL_INTERVAL_MS = 1_000;
 async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, forceRefresh, runtime = resolveGatewayServerRuntime(), env, onComplete, flightRecorderEntry, extractUsage, stdin, cwd) {
     let onCompleteOwnedByCaller = onComplete !== undefined;
@@ -406,6 +458,79 @@ async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, f
         message: `Execution exceeded sync deadline (${SYNC_DEADLINE_MS}ms). Poll with llm_job_status, collect with llm_job_result.`,
     };
 }
+async function awaitApiJobOrDefer(provider, apiRequest, corrId, runtime = resolveGatewayServerRuntime(), onComplete, flightRecorderEntry, extractUsage) {
+    let onCompleteOwnedByCaller = onComplete !== undefined;
+    const consumeOnComplete = () => {
+        if (!onCompleteOwnedByCaller || !onComplete)
+            return;
+        onCompleteOwnedByCaller = false;
+        try {
+            onComplete();
+        }
+        catch (err) {
+            runtime.logger.error(`awaitApiJobOrDefer onComplete (${provider.name}) threw`, err);
+        }
+    };
+    const deferralAvailable = runtime.persistence.backend !== "none" &&
+        runtime.persistence.asyncJobsEnabled &&
+        runtime.asyncJobManager.hasStore();
+    if (SYNC_DEADLINE_MS === 0 || !deferralAvailable) {
+        try {
+            const result = await runApiRequest(provider, apiRequest, runtime.logger);
+            return { stdout: result.text, stderr: "", code: 0 };
+        }
+        catch (err) {
+            return { stdout: "", stderr: err.message, code: 1 };
+        }
+        finally {
+            consumeOnComplete();
+        }
+    }
+    let outcome;
+    try {
+        outcome = runtime.asyncJobManager.startHttpJob({
+            provider,
+            apiRequest,
+            correlationId: corrId,
+            onComplete,
+            flightRecorderEntry,
+            extractUsage,
+        });
+        onCompleteOwnedByCaller = false;
+    }
+    catch (err) {
+        consumeOnComplete();
+        throw err;
+    }
+    const job = outcome.snapshot;
+    if (outcome.deduped) {
+        runtime.logger.info(`[${corrId}] api request deduped onto job ${job.id} (original corrId=${outcome.originalCorrelationId})`);
+    }
+    const deadline = Date.now() + SYNC_DEADLINE_MS;
+    while (Date.now() < deadline) {
+        const snapshot = runtime.asyncJobManager.getJobSnapshot(job.id);
+        if (snapshot && snapshot.status !== "running") {
+            const result = runtime.asyncJobManager.getJobResult(job.id);
+            if (!result)
+                return { stdout: "", stderr: "Job result unavailable", code: 1 };
+            return {
+                stdout: result.stdout,
+                stderr: result.stderr || result.error || "",
+                code: result.exitCode ?? 1,
+            };
+        }
+        await new Promise(resolve => setTimeout(resolve, SYNC_POLL_INTERVAL_MS));
+    }
+    runtime.asyncJobManager.armFlightCompleteForDeferral(job.id);
+    runtime.logger.info(`[${corrId}] ${provider.name} sync deadline exceeded (${SYNC_DEADLINE_MS}ms), deferring to async job ${job.id}`);
+    return {
+        deferred: true,
+        jobId: job.id,
+        cli: provider.name,
+        correlationId: corrId,
+        message: `Execution exceeded sync deadline (${SYNC_DEADLINE_MS}ms). Poll with llm_job_status, collect with llm_job_result.`,
+    };
+}
 function isDeferredResponse(result) {
     return "deferred" in result && result.deferred === true;
 }
@@ -795,7 +920,7 @@ function createApprovalDeniedResponse(operation, decision) {
 }
 function normalizeMcpServers(mcpServers) {
     if (!mcpServers || mcpServers.length === 0) {
-        return ["sqry"];
+        return [];
     }
     return [...new Set(mcpServers)];
 }
@@ -1552,9 +1677,6 @@ export function prepareGeminiRequest(params, runtime = resolveGatewayServerRunti
     if (params.allowedTools && params.allowedTools.length > 0) {
         return unsupported("allowedTools", "agy has no non-interactive allowed-tools flag");
     }
-    if (requestedMcpServers.length > 0) {
-        return unsupported("mcpServers", "agy has no non-interactive allowed MCP server allowlist flag");
-    }
     if (params.outputFormat && params.outputFormat !== "text") {
         return unsupported("outputFormat", "agy print mode currently emits text only");
     }
@@ -2182,6 +2304,120 @@ export async function handleGrokApiRequest(deps, params) {
         runtime.performanceMetrics.recordRequest("grok-api", durationMs || Math.max(0, Date.now() - startTime), wasSuccessful);
     }
 }
+function buildApiProviderCall(providerRuntime, params) {
+    const apiRequest = prepareApiRequest(providerRuntime, {
+        prompt: params.prompt ?? "",
+        system: params.system,
+        model: params.model,
+        maxOutputTokens: params.maxOutputTokens,
+        temperature: params.temperature,
+        topP: params.topP,
+        reasoningEffort: params.reasoningEffort,
+        timeoutMs: params.timeoutMs,
+    });
+    const provider = createApiProvider(providerRuntime.name, providerRuntime.kind);
+    return { provider, apiRequest };
+}
+function buildApiSuccessResponse(text, corrId, providerName) {
+    return {
+        content: [{ type: "text", text }],
+        structuredContent: {
+            response: text,
+            correlationId: corrId,
+            cli: providerName,
+            exitCode: 0,
+        },
+    };
+}
+export async function handleApiProviderRequest(runtimeArg, providerRuntime, params) {
+    const toolName = `api_${providerRuntime.name}_request`;
+    const corrId = params.correlationId ?? randomUUID();
+    const startTime = Date.now();
+    let wasSuccessful = false;
+    try {
+        if (!params.prompt || params.prompt.trim().length === 0) {
+            return createErrorResponse(toolName, 1, "prompt is required and cannot be empty", corrId);
+        }
+        const { provider, apiRequest } = buildApiProviderCall(providerRuntime, params);
+        const result = await awaitApiJobOrDefer(provider, apiRequest, corrId, runtimeArg);
+        if (isDeferredResponse(result))
+            return buildDeferredToolResponse(result);
+        if (result.code !== 0) {
+            return createErrorResponse(toolName, result.code, result.stderr, corrId);
+        }
+        wasSuccessful = true;
+        return buildApiSuccessResponse(result.stdout, corrId, providerRuntime.name);
+    }
+    catch (err) {
+        if (err instanceof ApiModelNotAllowedError) {
+            return createErrorResponse(toolName, 1, err.message, corrId, err);
+        }
+        return createErrorResponse(toolName, 1, "", corrId, err);
+    }
+    finally {
+        runtimeArg.performanceMetrics.recordRequest(providerRuntime.name, Math.max(0, Date.now() - startTime), wasSuccessful);
+    }
+}
+export function handleApiProviderRequestAsync(runtimeArg, providerRuntime, params) {
+    const toolName = `api_${providerRuntime.name}_request_async`;
+    const corrId = params.correlationId ?? randomUUID();
+    try {
+        if (!params.prompt || params.prompt.trim().length === 0) {
+            return createErrorResponse(toolName, 1, "prompt is required and cannot be empty", corrId);
+        }
+        const { provider, apiRequest } = buildApiProviderCall(providerRuntime, params);
+        const outcome = runtimeArg.asyncJobManager.startHttpJob({
+            provider,
+            apiRequest,
+            correlationId: corrId,
+            writeFlightStart: true,
+        });
+        return buildDeferredToolResponse({
+            deferred: true,
+            jobId: outcome.snapshot.id,
+            cli: providerRuntime.name,
+            correlationId: corrId,
+            message: outcome.deduped
+                ? `Deduped onto existing job ${outcome.snapshot.id}. Poll with llm_job_status.`
+                : `Started async job ${outcome.snapshot.id}. Poll with llm_job_status, collect with llm_job_result.`,
+        });
+    }
+    catch (err) {
+        if (err instanceof ApiModelNotAllowedError) {
+            return createErrorResponse(toolName, 1, err.message, corrId, err);
+        }
+        return createErrorResponse(toolName, 1, "", corrId, err);
+    }
+}
+const ApiReasoningEffortSchema = z.enum(["none", "low", "medium", "high"]);
+export function registerApiProviderTools(server, runtime, providers, asyncJobsEnabled) {
+    const registered = [];
+    const inputSchema = {
+        prompt: z.string().min(1).max(100000).optional().describe("Prompt text for the API provider"),
+        system: z.string().max(100000).optional().describe("Optional system instruction"),
+        model: z
+            .string()
+            .min(1)
+            .optional()
+            .describe("Model id; defaults to the provider default_model"),
+        correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
+        maxOutputTokens: z.number().int().positive().max(100000000).optional(),
+        temperature: z.number().finite().min(0).max(2).optional(),
+        topP: z.number().finite().min(0).max(1).optional(),
+        reasoningEffort: ApiReasoningEffortSchema.optional(),
+        timeoutMs: z.number().int().min(30_000).max(3_600_000).optional(),
+    };
+    for (const providerRuntime of enabledApiProviders(providers)) {
+        const name = providerRuntime.name;
+        server.tool(`api_${name}_request`, `Run a request against the "${name}" API provider (kind: ${providerRuntime.kind}) synchronously. Registered only when [providers.${name}] is configured and enabled.`, inputSchema, { title: `${name} API request`, readOnlyHint: false, openWorldHint: true }, async (params) => handleApiProviderRequest(runtime, providerRuntime, params));
+        registered.push(`api_${name}_request`);
+        if (asyncJobsEnabled) {
+            server.tool(`api_${name}_request_async`, `Start an async request against the "${name}" API provider; returns a jobId to poll with llm_job_status.`, inputSchema, { title: `${name} API request (async)`, readOnlyHint: false, openWorldHint: true }, async (params) => handleApiProviderRequestAsync(runtime, providerRuntime, params));
+            registered.push(`api_${name}_request_async`);
+        }
+    }
+    return registered;
+}
 function maybeBuildCacheTtlWarning(args) {
     if (args.cli !== "claude")
         return null;
@@ -2467,6 +2703,15 @@ export async function handleGeminiRequestAsync(deps, params) {
     }
 }
 export async function handleGrokRequest(deps, params) {
+    if (params.transport === "acp") {
+        return runAcpTransport(deps, {
+            provider: "grok",
+            prompt: params.prompt,
+            model: params.model,
+            sessionId: params.sessionId,
+            correlationId: params.correlationId,
+        });
+    }
     const runtime = resolveHandlerRuntime(deps);
     const startTime = Date.now();
     const prep = prepareGrokRequest({
@@ -2761,7 +3006,222 @@ export async function handleGrokRequestAsync(deps, params) {
         return createErrorResponse("grok_request_async", 1, "", corrId, error);
     }
 }
+export function prepareDevinRequest(params, _runtime) {
+    const corrId = params.correlationId ?? randomUUID();
+    let prompt = (params.prompt ?? "").trim();
+    if (!prompt) {
+        return createErrorResponse(params.operation, 1, "prompt is required and cannot be empty", corrId);
+    }
+    if (params.optimizePrompt)
+        prompt = optimizePromptText(prompt);
+    const resolvedModel = resolveModelAlias("devin", params.model, getCliInfo());
+    const args = ["-p", prompt];
+    if (resolvedModel)
+        args.push("--model", resolvedModel);
+    if (params.permissionMode)
+        args.push("--permission-mode", params.permissionMode);
+    if (params.promptFile)
+        args.push("--prompt-file", params.promptFile);
+    return {
+        corrId,
+        effectivePrompt: prompt,
+        resolvedModel,
+        requestedMcpServers: [],
+        approvalDecision: null,
+        args,
+        stablePrefixHash: null,
+        stablePrefixTokens: null,
+    };
+}
+export async function handleDevinRequest(deps, params) {
+    if (params.transport === "acp") {
+        return runAcpTransport(deps, {
+            provider: "devin",
+            prompt: params.prompt,
+            model: params.model,
+            sessionId: params.sessionId,
+            correlationId: params.correlationId,
+        });
+    }
+    const runtime = resolveHandlerRuntime(deps);
+    const startTime = Date.now();
+    const prep = prepareDevinRequest({
+        prompt: params.prompt,
+        model: params.model,
+        permissionMode: params.permissionMode,
+        promptFile: params.promptFile,
+        correlationId: params.correlationId,
+        optimizePrompt: params.optimizePrompt,
+        operation: "devin_request",
+    }, runtime);
+    if (!("args" in prep))
+        return prep;
+    const { corrId, args } = prep;
+    let durationMs = 0;
+    let wasSuccessful = false;
+    safeFlightStart({
+        correlationId: corrId,
+        cli: "devin",
+        model: prep.resolvedModel || "default",
+        prompt: prep.effectivePrompt,
+        sessionId: params.sessionId,
+    }, runtime);
+    try {
+        const sessionResult = resolveGrokSessionArgs({
+            sessionId: params.sessionId,
+            resumeLatest: params.resumeLatest,
+            createNewSession: params.createNewSession,
+        });
+        if (sessionResult.userProvidedSession) {
+            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);
+        if (isDeferredResponse(result)) {
+            return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
+        }
+        const { stdout, stderr, code } = result;
+        durationMs = Math.max(0, Date.now() - startTime);
+        if (code !== 0) {
+            safeFlightComplete(corrId, {
+                response: stderr || "",
+                durationMs,
+                retryCount: 0,
+                circuitBreakerState: "closed",
+                optimizationApplied: false,
+                exitCode: code,
+                errorMessage: stderr || `Exit code ${code}`,
+                status: "failed",
+            }, runtime);
+            return createErrorResponse("devin", code, stderr, corrId);
+        }
+        wasSuccessful = true;
+        let effectiveSessionId = sessionResult.effectiveSessionId;
+        if (sessionResult.userProvidedSession && effectiveSessionId) {
+            const existing = await deps.sessionManager.getSession(effectiveSessionId);
+            if (!existing) {
+                try {
+                    await deps.sessionManager.createSession("devin", "Devin Session", effectiveSessionId);
+                }
+                catch {
+                    const rechecked = await deps.sessionManager.getSession(effectiveSessionId);
+                    if (!rechecked)
+                        throw new Error(`Failed to create or find session ${effectiveSessionId}`);
+                }
+            }
+            await deps.sessionManager.updateSessionUsage(effectiveSessionId);
+        }
+        else if (!params.createNewSession && !effectiveSessionId) {
+            const newSession = await deps.sessionManager.createSession("devin", "Devin Session", `${GATEWAY_SESSION_PREFIX}${randomUUID()}`);
+            effectiveSessionId = newSession.id;
+        }
+        const response = buildCliResponse("devin", stdout, params.optimizeResponse ?? false, corrId, effectiveSessionId, prep, durationMs, sessionResult.userProvidedSession);
+        safeFlightComplete(corrId, {
+            response: stdout,
+            durationMs,
+            retryCount: 0,
+            circuitBreakerState: "closed",
+            optimizationApplied: params.optimizePrompt || (params.optimizeResponse ?? false),
+            exitCode: 0,
+            status: "completed",
+        }, runtime);
+        return response;
+    }
+    catch (error) {
+        const elapsedMs = Math.max(0, Date.now() - startTime);
+        safeFlightComplete(corrId, {
+            response: "",
+            durationMs: elapsedMs,
+            retryCount: 0,
+            circuitBreakerState: "closed",
+            optimizationApplied: false,
+            exitCode: 1,
+            errorMessage: error.message,
+            status: "failed",
+        }, runtime);
+        return createErrorResponse("devin", 1, "", corrId, error);
+    }
+    finally {
+        runtime.performanceMetrics.recordRequest("devin", Math.max(0, durationMs || Date.now() - startTime), wasSuccessful);
+    }
+}
+export async function handleDevinRequestAsync(deps, params) {
+    const runtime = resolveHandlerRuntime(deps);
+    const prep = prepareDevinRequest({
+        prompt: params.prompt,
+        model: params.model,
+        permissionMode: params.permissionMode,
+        promptFile: params.promptFile,
+        correlationId: params.correlationId,
+        optimizePrompt: params.optimizePrompt,
+        operation: "devin_request_async",
+    }, runtime);
+    if (!("args" in prep))
+        return prep;
+    const { corrId, args } = prep;
+    try {
+        const sessionResult = resolveGrokSessionArgs({
+            sessionId: params.sessionId,
+            resumeLatest: params.resumeLatest,
+            createNewSession: params.createNewSession,
+        });
+        if (sessionResult.userProvidedSession) {
+            await getExistingSessionForProvider(deps.sessionManager, sessionResult.effectiveSessionId, "devin");
+        }
+        args.push(...sessionResult.resumeArgs);
+        let effectiveSessionId = sessionResult.effectiveSessionId;
+        if (sessionResult.userProvidedSession && effectiveSessionId) {
+            const existing = await deps.sessionManager.getSession(effectiveSessionId);
+            if (!existing) {
+                try {
+                    await deps.sessionManager.createSession("devin", "Devin Session", effectiveSessionId);
+                }
+                catch {
+                    const rechecked = await deps.sessionManager.getSession(effectiveSessionId);
+                    if (!rechecked)
+                        throw new Error(`Failed to create or find session ${effectiveSessionId}`);
+                }
+            }
+            await deps.sessionManager.updateSessionUsage(effectiveSessionId);
+        }
+        else if (!params.createNewSession && !effectiveSessionId) {
+            const newSession = await deps.sessionManager.createSession("devin", "Devin Session", `${GATEWAY_SESSION_PREFIX}${randomUUID()}`);
+            effectiveSessionId = newSession.id;
+        }
+        assertUpstreamCliArgs("devin", args);
+        assertUpstreamCliEnv("devin", undefined);
+        const devinAsyncFrHandoff = buildAsyncFlightRecorderHandoff("devin", prep, effectiveSessionId, undefined);
+        const job = deps.asyncJobManager.startJob("devin", args, corrId, undefined, resolveIdleTimeout("devin", params.idleTimeoutMs), undefined, params.forceRefresh, undefined, undefined, devinAsyncFrHandoff.flightRecorderEntry, devinAsyncFrHandoff.extractUsage, true);
+        deps.logger.info(`[${corrId}] devin_request_async started job ${job.id}`);
+        return {
+            content: [
+                {
+                    type: "text",
+                    text: JSON.stringify({
+                        success: true,
+                        job,
+                        sessionId: effectiveSessionId || null,
+                        resumable: sessionResult.userProvidedSession,
+                    }, null, 2),
+                },
+            ],
+        };
+    }
+    catch (error) {
+        return createErrorResponse("devin_request_async", 1, "", corrId, error);
+    }
+}
 export async function handleMistralRequest(deps, params) {
+    if (params.transport === "acp") {
+        return runAcpTransport(deps, {
+            provider: "mistral",
+            prompt: params.prompt,
+            model: params.model,
+            sessionId: params.sessionId,
+            correlationId: params.correlationId,
+        });
+    }
     const runtime = resolveHandlerRuntime(deps);
     const startTime = Date.now();
     const prep = prepareMistralRequest({
@@ -3157,8 +3617,15 @@ export function createGatewayServer(deps = {}) {
     const asyncJobsEnabled = persistence.backend !== "none" && persistence.asyncJobsEnabled && asyncJobManager.hasStore();
     const server = newGatewayMcpServer(asyncJobsEnabled, grokApiToolsEnabled);
     registerBaseResources(server, runtime);
-    registerValidationTools(server, { asyncJobManager });
+    registerValidationTools(server, {
+        asyncJobManager,
+        apiProviders: enabledApiProviders(providers),
+    });
     registerWorkspaceTools(server, runtime);
+    const apiProviderTools = registerApiProviderTools(server, runtime, providers, asyncJobsEnabled);
+    if (apiProviderTools.length > 0) {
+        runtime.logger.info(`Registered API provider tools: ${apiProviderTools.join(", ")}`);
+    }
     if (grokApiToolsEnabled) {
         server.tool("grok_api_request", "Run an xAI Grok API request synchronously through the Responses API. Requires exactly one of prompt or promptParts. Registered only when [providers.xai] is configured and its API-key env var is present.", {
             prompt: z
@@ -3352,10 +3819,7 @@ export function createGatewayServer(deps = {}) {
             .enum(["strict", "balanced", "permissive"])
             .optional()
             .describe("Approval policy override"),
-        mcpServers: z
-            .array(MCP_SERVER_ENUM)
-            .default(["sqry"])
-            .describe("MCP servers exposed to Claude"),
+        mcpServers: z.array(mcpServerEnum()).default([]).describe("MCP servers exposed to Claude"),
         strictMcpConfig: z
             .boolean()
             .default(false)
@@ -3626,8 +4090,8 @@ export function createGatewayServer(deps = {}) {
             .optional()
             .describe("Approval policy override"),
         mcpServers: z
-            .array(MCP_SERVER_ENUM)
-            .default(["sqry"])
+            .array(mcpServerEnum())
+            .default([])
             .describe("MCP server names for approval tracking (Codex manages its own MCP config)"),
         sessionId: z
             .string()
@@ -3994,9 +4458,9 @@ export function createGatewayServer(deps = {}) {
             .optional()
             .describe("Approval policy override"),
         mcpServers: z
-            .array(MCP_SERVER_ENUM)
+            .array(mcpServerEnum())
             .default([])
-            .describe("Unsupported for Antigravity CLI; non-empty values are rejected"),
+            .describe("MCP server names accepted for approval tracking only; Antigravity manages its own MCP configuration."),
         allowedTools: z
             .array(z.string())
             .optional()
@@ -4082,6 +4546,10 @@ export function createGatewayServer(deps = {}) {
             .describe("Prompt text for Grok (mutually exclusive with promptParts)"),
         promptParts: PromptPartsSchema.optional().describe("Cache-aware structured prompt: { system?, tools?, context?, task }. Mutually exclusive with prompt. Stable parts hash into cache_state for prefix-discipline tracking."),
         model: z.string().optional().describe("Model name or alias (e.g. grok-build, latest)"),
+        transport: z
+            .enum(["cli", "acp"])
+            .default("cli")
+            .describe("Transport: 'cli' (default) runs the Grok CLI; 'acp' routes through `grok agent stdio` when [acp].enabled and the provider's runtime_enabled are set (fails closed otherwise)."),
         ...GROK_GENERATED_SHAPE,
         sessionId: z
             .string()
@@ -4109,8 +4577,8 @@ export function createGatewayServer(deps = {}) {
             .optional()
             .describe("Approval policy override"),
         mcpServers: z
-            .array(MCP_SERVER_ENUM)
-            .default(["sqry"])
+            .array(mcpServerEnum())
+            .default([])
             .describe("MCP server names for approval tracking (Grok manages its own MCP config via `grok mcp`)"),
         correlationId: z.string().optional().describe("Request trace ID (auto if omitted)"),
         optimizePrompt: z.boolean().default(false).describe("Optimize prompt before execution"),
@@ -4146,11 +4614,12 @@ 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, optimizeResponse, 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, workspace, worktree, }) => {
+    }, async ({ prompt, promptParts, model, transport, outputFormat, sessionId, resumeLatest, createNewSession, alwaysApprove, permissionMode, effort, reasoningEffort, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, optimizeResponse, 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, workspace, worktree, }) => {
         return handleGrokRequest({ sessionManager, logger, runtime }, {
             prompt,
             promptParts,
             model,
+            transport,
             outputFormat,
             sessionId,
             resumeLatest,
@@ -4201,6 +4670,72 @@ export function createGatewayServer(deps = {}) {
             worktree,
         });
     });
+    server.tool("devin_request", "Run a Cognition Devin CLI request synchronously (auto-defers to a pollable job past the sync deadline when async jobs are enabled; otherwise runs to completion). Headless print mode (`devin -p`).", {
+        prompt: z
+            .string()
+            .min(1, "Prompt cannot be empty")
+            .max(100000, "Prompt too long (max 100k chars)")
+            .optional()
+            .describe("Prompt text for Devin CLI"),
+        model: z.string().optional().describe("Model name or alias (e.g. opus, latest)"),
+        transport: z
+            .enum(["cli", "acp"])
+            .default("cli")
+            .describe("Transport: 'cli' (default) runs the Devin CLI; 'acp' routes through `devin acp` when [acp].enabled and the provider's runtime_enabled are set (fails closed otherwise)."),
+        permissionMode: z
+            .enum(["auto", "smart", "dangerous"])
+            .optional()
+            .describe("Devin CLI permission mode (--permission-mode). auto auto-approves read-only tools; smart additionally auto-runs actions a fast model judges safe; dangerous auto-approves all."),
+        promptFile: z
+            .string()
+            .optional()
+            .describe("Load the initial prompt from a file (--prompt-file)"),
+        sessionId: z
+            .string()
+            .optional()
+            .describe("Devin session ID to resume (emits --resume <id>; use resumeLatest for --continue)"),
+        resumeLatest: z
+            .boolean()
+            .default(false)
+            .describe("Resume the most recent Devin session in cwd (--continue)"),
+        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"),
+        optimizeResponse: z.boolean().default(false).describe("Optimize response output"),
+        idleTimeoutMs: z
+            .number()
+            .int()
+            .min(30_000)
+            .max(3_600_000)
+            .optional()
+            .describe("Idle timeout in ms (min 30s, max 1h, omit=CLI default)"),
+        forceRefresh: z
+            .boolean()
+            .default(false)
+            .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
+    }, {
+        title: "Devin CLI request",
+        readOnlyHint: false,
+        destructiveHint: true,
+        idempotentHint: false,
+        openWorldHint: true,
+    }, async ({ prompt, model, transport, permissionMode, promptFile, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, }) => {
+        return handleDevinRequest({ sessionManager, logger, runtime }, {
+            prompt,
+            model,
+            transport,
+            permissionMode,
+            promptFile,
+            sessionId,
+            resumeLatest,
+            createNewSession,
+            correlationId,
+            optimizePrompt,
+            optimizeResponse,
+            idleTimeoutMs,
+            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.", {
         prompt: z
             .string()
@@ -4213,6 +4748,10 @@ export function createGatewayServer(deps = {}) {
             .string()
             .optional()
             .describe("Model alias (e.g. mistral-medium-3.5, latest). Resolved alias is injected via VIBE_ACTIVE_MODEL env var; Vibe has no --model flag."),
+        transport: z
+            .enum(["cli", "acp"])
+            .default("cli")
+            .describe("Transport: 'cli' (default) runs the Vibe CLI; 'acp' routes through `vibe-acp` when [acp].enabled and the provider's runtime_enabled are set (fails closed otherwise)."),
         outputFormat: z
             .enum(["text", "plain", "json", "streaming", "stream-json"])
             .optional()
@@ -4227,9 +4766,9 @@ export function createGatewayServer(deps = {}) {
             .describe("Resume most recent Vibe session in cwd (--continue)"),
         createNewSession: z.boolean().default(false).describe("Force new session"),
         permissionMode: z
-            .enum(MISTRAL_AGENT_MODES)
+            .string()
             .optional()
-            .describe("Vibe agent mode (default|plan|accept-edits|auto-approve|chat|explore|lean). 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 auto-approve for programmatic use."),
         approvalStrategy: z
             .enum(["legacy", "mcp_managed"])
             .default("legacy")
@@ -4239,8 +4778,8 @@ export function createGatewayServer(deps = {}) {
             .optional()
             .describe("Approval policy override"),
         mcpServers: z
-            .array(MCP_SERVER_ENUM)
-            .default(["sqry"])
+            .array(mcpServerEnum())
+            .default([])
             .describe("MCP server names for approval tracking (Vibe manages its own MCP config via `vibe mcp`)"),
         allowedTools: z
             .array(z.string())
@@ -4288,11 +4827,12 @@ 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, optimizeResponse, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, workspace, worktree, }) => {
+    }, async ({ prompt, promptParts, model, transport, outputFormat, sessionId, resumeLatest, createNewSession, permissionMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, workspace, worktree, }) => {
         return handleMistralRequest({ sessionManager, logger, runtime }, {
             prompt,
             promptParts,
             model,
+            transport,
             outputFormat,
             sessionId,
             resumeLatest,
@@ -4437,10 +4977,7 @@ export function createGatewayServer(deps = {}) {
                 .enum(["strict", "balanced", "permissive"])
                 .optional()
                 .describe("Approval policy override"),
-            mcpServers: z
-                .array(MCP_SERVER_ENUM)
-                .default(["sqry"])
-                .describe("MCP servers exposed to Claude"),
+            mcpServers: z.array(mcpServerEnum()).default([]).describe("MCP servers exposed to Claude"),
             strictMcpConfig: z
                 .boolean()
                 .default(false)
@@ -4630,8 +5167,8 @@ export function createGatewayServer(deps = {}) {
                 .optional()
                 .describe("Approval policy override"),
             mcpServers: z
-                .array(MCP_SERVER_ENUM)
-                .default(["sqry"])
+                .array(mcpServerEnum())
+                .default([])
                 .describe("MCP server names for approval tracking (Codex manages its own MCP config)"),
             sessionId: z
                 .string()
@@ -4756,9 +5293,9 @@ export function createGatewayServer(deps = {}) {
                 .optional()
                 .describe("Approval policy override"),
             mcpServers: z
-                .array(MCP_SERVER_ENUM)
+                .array(mcpServerEnum())
                 .default([])
-                .describe("Unsupported for Antigravity CLI; non-empty values are rejected"),
+                .describe("MCP server names accepted for approval tracking only; Antigravity manages its own MCP configuration."),
             allowedTools: z
                 .array(z.string())
                 .optional()
@@ -4877,8 +5414,8 @@ export function createGatewayServer(deps = {}) {
                 .optional()
                 .describe("Approval policy override"),
             mcpServers: z
-                .array(MCP_SERVER_ENUM)
-                .default(["sqry"])
+                .array(mcpServerEnum())
+                .default([])
                 .describe("MCP server names for approval tracking (Grok manages its own MCP config via `grok mcp`)"),
             allowedTools: z
                 .array(z.string())
@@ -5071,6 +5608,65 @@ export function createGatewayServer(deps = {}) {
                 worktree,
             });
         });
+        server.tool("devin_request_async", "Start a Cognition Devin CLI request as a durable background job. Poll with llm_job_status, collect with llm_job_result.", {
+            prompt: z
+                .string()
+                .min(1, "Prompt cannot be empty")
+                .max(100000, "Prompt too long (max 100k chars)")
+                .optional()
+                .describe("Prompt text for Devin CLI"),
+            model: z.string().optional().describe("Model name or alias (e.g. opus, latest)"),
+            permissionMode: z
+                .enum(["auto", "smart", "dangerous"])
+                .optional()
+                .describe("Devin CLI permission mode (--permission-mode). auto, smart, or dangerous."),
+            promptFile: z
+                .string()
+                .optional()
+                .describe("Load the initial prompt from a file (--prompt-file)"),
+            sessionId: z
+                .string()
+                .optional()
+                .describe("Devin session ID to resume (--resume <id>; use resumeLatest for --continue)"),
+            resumeLatest: z
+                .boolean()
+                .default(false)
+                .describe("Resume the most recent Devin session in cwd (--continue)"),
+            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"),
+            idleTimeoutMs: z
+                .number()
+                .int()
+                .min(30_000)
+                .max(3_600_000)
+                .optional()
+                .describe("Idle timeout in ms (min 30s, max 1h, omit=CLI default)"),
+            forceRefresh: z
+                .boolean()
+                .default(false)
+                .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
+        }, {
+            title: "Devin CLI request (async)",
+            readOnlyHint: false,
+            destructiveHint: true,
+            idempotentHint: false,
+            openWorldHint: true,
+        }, async ({ prompt, model, permissionMode, promptFile, sessionId, resumeLatest, createNewSession, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, }) => {
+            return handleDevinRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
+                prompt,
+                model,
+                permissionMode,
+                promptFile,
+                sessionId,
+                resumeLatest,
+                createNewSession,
+                correlationId,
+                optimizePrompt,
+                idleTimeoutMs,
+                forceRefresh,
+            });
+        });
         server.tool("mistral_request_async", "Start a Mistral Vibe CLI request as a durable background job. Poll with llm_job_status, collect with llm_job_result.", {
             prompt: z
                 .string()
@@ -5097,9 +5693,9 @@ export function createGatewayServer(deps = {}) {
                 .describe("Resume most recent Vibe session in cwd (--continue)"),
             createNewSession: z.boolean().default(false).describe("Force new session"),
             permissionMode: z
-                .enum(MISTRAL_AGENT_MODES)
+                .string()
                 .optional()
-                .describe("Vibe agent mode (default|plan|accept-edits|auto-approve|chat|explore|lean). 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 auto-approve for programmatic use."),
             approvalStrategy: z
                 .enum(["legacy", "mcp_managed"])
                 .default("legacy")
@@ -5109,8 +5705,8 @@ export function createGatewayServer(deps = {}) {
                 .optional()
                 .describe("Approval policy override"),
             mcpServers: z
-                .array(MCP_SERVER_ENUM)
-                .default(["sqry"])
+                .array(mcpServerEnum())
+                .default([])
                 .describe("MCP server names for approval tracking (Vibe manages its own MCP config via `vibe mcp`)"),
             allowedTools: z
                 .array(z.string())
@@ -5439,6 +6035,11 @@ export function createGatewayServer(deps = {}) {
                     defaultModel: null,
                     mode: "disabled",
                 },
+            apiProviders: enabledApiProviders(providers).map(p => ({
+                ...apiProviderCatalogEntry(p),
+                baseUrl: p.baseUrl,
+                breakerState: apiProviderBreakerState(p.name),
+            })),
             sources: providers.sources,
         };
         return {