llm-cli-gateway 1.17.3 → 1.17.5

package/dist/index.js

@@ -35,7 +35,6 @@ import { printDoctorJson } from "./doctor.js";
 import { registerValidationTools } from "./validation-tools.js";
 import { assertUpstreamCliArgs, assertUpstreamCliEnv, buildUpstreamContractReport, } from "./upstream-contracts.js";
 import { entrypointFileURL } from "./entrypoint-url.js";
-// Simple logger that writes to stderr (stdout is used for MCP protocol)
 const logger = {
     info: (message, ...args) => {
         console.error(`[INFO] ${new Date().toISOString()} - ${message}`, ...args);
@@ -94,10 +93,6 @@ function logOptimizationTokens(kind, correlationId, original, optimized) {
     const reduction = originalTokens === 0 ? 0 : ((originalTokens - optimizedTokens) / originalTokens) * 100;
     logger.info(`[${correlationId}] ${kind} tokens ${originalTokens} → ${optimizedTokens} (${reduction.toFixed(1)}% reduction)`);
 }
-// Sync-to-async deadline: if a sync tool's CLI call hasn't finished within this
-// window, the tool returns a deferred async job reference instead of blocking
-// until the MCP client's tool-call timeout fires (~60s in many runtimes).
-// Configurable via SYNC_DEADLINE_MS env var. Set to 0 to disable (pure sync).
 const SYNC_DEADLINE_MS = (() => {
     const env = process.env.SYNC_DEADLINE_MS;
     if (env !== undefined) {
@@ -105,11 +100,8 @@ const SYNC_DEADLINE_MS = (() => {
         if (Number.isFinite(parsed) && parsed >= 0)
             return parsed;
     }
-    return 45_000; // 45s default — safely under the 60s MCP client cap
+    return 45_000;
 })();
-//──────────────────────────────────────────────────────────────────────────────
-// Skills loader — reads .agents/skills/*/SKILL.md at startup
-//──────────────────────────────────────────────────────────────────────────────
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 const SKILLS_DIR = join(__dirname, "..", ".agents", "skills");
@@ -124,7 +116,6 @@ function packageVersion() {
             return parsed.version || "unknown";
         }
         catch {
-            // Try next candidate.
         }
     }
     return "unknown";
@@ -137,24 +128,19 @@ function loadSkills() {
             const skillPath = join(SKILLS_DIR, dir.name, "SKILL.md");
             try {
                 const content = readFileSync(skillPath, "utf-8");
-                // Extract description from YAML frontmatter
                 const descMatch = content.match(/^---[\s\S]*?description:\s*(.+?)$/m);
                 const description = descMatch?.[1]?.trim() || dir.name;
                 skills.push({ name: dir.name, content, description });
             }
             catch {
-                // Skill file missing or unreadable — skip silently
             }
         }
     }
     catch {
-        // Skills directory missing — not fatal
     }
     return skills;
 }
 const loadedSkills = loadSkills();
-// L1: Compact server instructions (~200 tokens) — injected into every client's
-// system prompt at connection time. Covers key patterns + pointers to L2 resources.
 const SERVER_INSTRUCTIONS = `llm-cli-gateway: Multi-LLM orchestration via MCP.
 
 Tools: claude_request, codex_request, gemini_request, grok_request, mistral_request (sync) | *_request_async (async)
@@ -175,17 +161,11 @@ ${loadedSkills.map(s => `- skills://${s.name} — ${s.description}`).join("\n")}
 function newGatewayMcpServer() {
     return new McpServer({ name: "llm-cli-gateway", version: "1.0.0" }, { instructions: SERVER_INSTRUCTIONS });
 }
-// Global state (initialized asynchronously)
 let sessionManager;
 let db = null;
 const performanceMetrics = new PerformanceMetrics();
 let resourceProvider;
 let flightRecorder = null;
-// Resolved persistence config — single source of truth for the async-job backend.
-// Driven by ~/.llm-cli-gateway/config.toml (+ deprecated env-var overrides).
-// When backend = "none", the JobStore is null AND *_request_async tools are not
-// registered (see createGatewayServer), making silent in-memory loss
-// structurally impossible.
 let persistenceConfig = null;
 let cacheAwarenessConfig = null;
 let jobStore = null;
@@ -231,47 +211,9 @@ function getApprovalManager(runtimeLogger = logger) {
     return approvalManager;
 }
 const MCP_SERVER_ENUM = z.enum(CLAUDE_MCP_SERVER_NAMES);
-/**
- * Phase 4 slice δ — shared Zod fragments for `maxTurns` / `maxPrice`.
- *
- * Both flags reach the upstream CLIs as decimal-formatted argv strings via
- * `String(N)`. `z.number().int().positive()` alone lets values past
- * `Number.MAX_SAFE_INTEGER` through, after which `String(1e21)` emits
- * scientific notation that Grok and Vibe both reject. The bounds below
- * (safe-integer cap + 10000 ceiling for turns; finite + 10000 USD ceiling
- * for price) guarantee a lossless decimal stringification AND a sane
- * upper bound — no plausible single agent loop exceeds 10k turns or 10k USD.
- */
 export const MAX_TURNS_SCHEMA = z.number().int().positive().safe().max(10_000);
-// Token budgets can legitimately exceed the agent-turn cap by orders of
-// magnitude. Keep a finite operational guardrail while avoiding the 10k turn
-// ceiling that would make large-context Vibe sessions unusable.
 export const MAX_TOKENS_SCHEMA = z.number().int().positive().safe().max(100_000_000);
-// `.min(1e-6)` keeps the value in JS's decimal-stringify range:
-// String(1e-6) === "0.000001" but String(1e-7) === "1e-7", which both
-// upstream CLIs would reject. 1µUSD per request is fine-grained enough
-// for any plausible budget-cap use.
 export const MAX_PRICE_SCHEMA = z.number().positive().finite().min(1e-6).max(10_000);
-/**
- * Slice λ: shared worktree directive for all 10 `*_request` / `*_request_async`
- * tools. `true` creates a fresh worktree under `<repoRoot>/.worktrees/<uuid>`
- * branched from HEAD. `{ name?, ref? }` lets the caller supply a sanitized
- * name and/or git ref (default ref: HEAD).
- *
- * Lifecycle is gateway-owned: the gateway pre-creates the worktree via
- * `git worktree add`, then spawns the child CLI with `cwd: <worktree-path>`.
- * No `-w` / `--worktree` flag is ever emitted to the underlying CLI. When
- * the request carries a sessionId and the session already has a worktree,
- * that worktree is reused. On session_delete or TTL eviction the gateway
- * runs `git worktree remove --force`.
- *
- * Tool response: when a worktree was used, the successful response stdout
- * is prefixed with `[gateway] worktree=<absolute-path>\n` so callers can
- * parse/use the path without a schema change (slice λ §1.d).
- *
- * NOTE: callers should `.gitignore` the `.worktrees/` directory in their
- * repo (the gateway does NOT auto-gitignore — see slice λ spec Q4).
- */
 export const WORKTREE_SCHEMA = z
     .union([
     z.boolean(),
@@ -296,9 +238,6 @@ export const WORKTREE_SCHEMA = z
     "path. NOTE: callers should `.gitignore` the `.worktrees/` " +
     "directory in their repo (the gateway does NOT auto-gitignore — " +
     "see slice λ spec Q4).");
-// U22: Session-provider enum extended to five providers. The storage layer's
-// CLI_TYPES already includes "mistral"; the MCP-tool layer mirrors that here so
-// session_create / session_list / session_clear_all accept the fifth provider.
 export const SESSION_PROVIDER_VALUES = ["claude", "codex", "gemini", "grok", "mistral"];
 export const SESSION_PROVIDER_ENUM = z.enum(SESSION_PROVIDER_VALUES);
 let activeServer = null;
@@ -308,13 +247,10 @@ export function resolveGatewayServerRuntime(deps = {}, options = {}) {
     const runtimeSessionManager = deps.sessionManager ?? sessionManager;
     const runtimePerformanceMetrics = deps.performanceMetrics ??
         (options.isolateState ? new PerformanceMetrics() : performanceMetrics);
-    // Resolve flight recorder BEFORE async manager so isolateState managers
-    // can be wired with the same recorder instance the runtime exposes.
     const runtimeFlightRecorder = deps.flightRecorder ?? getFlightRecorder(runtimeLogger);
     const runtimeAsyncJobManager = deps.asyncJobManager ??
         (options.isolateState
-            ? // Factory-created test/HTTP session servers must not mark another instance's
-                // durable jobs orphaned. Stdio startup injects the process-global manager.
+            ?
                 newAsyncJobManager(runtimePerformanceMetrics, runtimeLogger, null, runtimeFlightRecorder)
             : getAsyncJobManager(runtimeLogger));
     const runtimeApprovalManager = deps.approvalManager ??
@@ -337,15 +273,12 @@ export function resolveGatewayServerRuntime(deps = {}, options = {}) {
         cacheAwareness: deps.cacheAwareness ?? getCacheAwarenessConfig(runtimeLogger),
     };
 }
-// Per-CLI idle timeouts: kill process if no stdout/stderr activity for this duration.
-// Claude idle timeout only applies in stream-json mode (with --include-partial-messages).
-// In text/json mode, Claude produces no output until done, so idle timeout would false-positive.
 const CLI_IDLE_TIMEOUTS = {
-    claude: 600_000, // 10 minutes — only used when outputFormat=stream-json
-    codex: 600_000, // 10 minutes — Codex streams stderr progress
-    gemini: 600_000, // 10 minutes — Gemini streams stdout in real-time
-    grok: 600_000, // 10 minutes — Grok streams stderr/stdout activity in headless mode
-    mistral: 600_000, // 10 minutes — Vibe streams stdout/stderr in headless mode
+    claude: 600_000,
+    codex: 600_000,
+    gemini: 600_000,
+    grok: 600_000,
+    mistral: 600_000,
 };
 function resolveIdleTimeout(cli, override) {
     if (override !== undefined)
@@ -353,41 +286,7 @@ function resolveIdleTimeout(cli, override) {
     return CLI_IDLE_TIMEOUTS[cli];
 }
 const SYNC_POLL_INTERVAL_MS = 1_000;
-/**
- * Start an async job and poll until completion or deadline.
- * Returns the job result if it finishes in time, or a deferral marker.
- */
-async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, forceRefresh, runtime = resolveGatewayServerRuntime(), env, onComplete, 
-/**
- * Slice 1.5: when the sync handler has already written a logStart row
- * keyed on `corrId`, pass these so the manager can write logComplete
- * (with usage extraction) when the underlying async job terminates —
- * even if the sync handler returned a deferred response.
- * `writeFlightStart` is NEVER true on this path: the sync handler is
- * always the upstream logStart writer.
- */
-flightRecorderEntry, extractUsage, 
-/**
- * Slice κ: optional stdin payload piped to the child CLI. Currently
- * only Claude's `--input-format stream-json` path sets this. Threaded
- * through both the direct-execute fallback (SYNC_DEADLINE_MS===0) and
- * the AsyncJobManager spawn path, and participates in the dedup key.
- */
-stdin, 
-/**
- * Slice λ: optional working directory for the spawned child process,
- * derived from a gateway-owned git worktree. Threaded to both the
- * direct-execute fallback (`executeCli({ cwd })`) and the
- * AsyncJobManager dedup-aware spawn path
- * (`startJobWithDedup({ cwd })`). `cwd` also participates in the
- * dedup key (see async-job-manager.buildRequestKey) so two requests
- * with identical argv in different worktrees do not collide.
- */
-cwd) {
-    // U26 fix: ownership of onComplete is a contract. Once this function returns
-    // OR throws, the caller MUST consider onComplete consumed — i.e. it has
-    // either been run, or the AsyncJobManager has taken ownership of it. The
-    // caller never needs to reclaim.
+async function awaitJobOrDefer(cli, args, corrId, idleTimeoutMs, outputFormat, forceRefresh, runtime = resolveGatewayServerRuntime(), env, onComplete, flightRecorderEntry, extractUsage, stdin, cwd) {
     let onCompleteOwnedByCaller = onComplete !== undefined;
     const consumeOnComplete = () => {
         if (!onCompleteOwnedByCaller || !onComplete)
@@ -409,8 +308,6 @@ cwd) {
         throw err;
     }
     if (SYNC_DEADLINE_MS === 0) {
-        // Disabled — fall through to direct execution.
-        // Note: direct execution bypasses dedup. forceRefresh is implied.
         const command = cli === "mistral" ? "vibe" : cli;
         try {
             return await executeCli(command, args, {
@@ -422,8 +319,6 @@ cwd) {
             });
         }
         finally {
-            // Direct-execution path completes inline; release per-request resources
-            // (e.g. outputSchema temp files) here.
             consumeOnComplete();
         }
     }
@@ -437,22 +332,12 @@ cwd) {
             env,
             stdin,
             onComplete,
-            // Sync-deferred path: the upstream sync handler already wrote
-            // logStart for this corrId, so writeFlightStart stays false. The
-            // manager still writes logComplete on terminal state (which UPDATEs
-            // the sync handler's row), closing the previously-orphaned
-            // sync-deferred case.
             flightRecorderEntry,
             extractUsage,
         });
-        // Handoff succeeded: AsyncJobManager owns onComplete (it'll fire via
-        // fireOnComplete on terminal status, or run inline immediately for dedup).
         onCompleteOwnedByCaller = false;
     }
     catch (err) {
-        // Spawn or pre-spawn failure inside AsyncJobManager. The record was never
-        // registered, so onComplete will never be called by the manager. Reclaim
-        // here so the temp file is not leaked.
         consumeOnComplete();
         throw err;
     }
@@ -464,7 +349,6 @@ cwd) {
     while (Date.now() < deadline) {
         const snapshot = runtime.asyncJobManager.getJobSnapshot(job.id);
         if (snapshot && snapshot.status !== "running") {
-            // Job finished within deadline — extract result
             const result = runtime.asyncJobManager.getJobResult(job.id);
             if (!result) {
                 return { stdout: "", stderr: "Job result unavailable", code: 1 };
@@ -477,13 +361,6 @@ cwd) {
         }
         await new Promise(resolve => setTimeout(resolve, SYNC_POLL_INTERVAL_MS));
     }
-    // Deadline exceeded — return deferral.
-    // R2 Codex-Unit-B F1: hand FR-complete ownership to the manager. Until
-    // this call, the manager skips writeFlightComplete on terminal so the
-    // sync handler's safeFlightComplete (with rich approvalDecision /
-    // optimizationApplied metadata) wins for sync-inline completions. From
-    // here on the sync handler returns deferred and will NOT write
-    // safeFlightComplete, so the manager must.
     runtime.asyncJobManager.armFlightCompleteForDeferral(job.id);
     runtime.logger.info(`[${corrId}] ${cli} sync deadline exceeded (${SYNC_DEADLINE_MS}ms), deferring to async job ${job.id}`);
     return {
@@ -517,27 +394,6 @@ function buildDeferredToolResponse(deferred, sessionId) {
         ],
     };
 }
-/**
- * Slice λ: resolve a request's worktree directive into a spawn cwd.
- *
- * - `worktreeOpt` is the Zod-validated input value (boolean |
- *   `{ name?, ref? }` | undefined).
- * - When the request has a session AND the session already has a
- *   `metadata.worktreePath`, that path is reused (resume semantics).
- *   The reused path is returned without touching git; if the directory
- *   was externally removed between requests, the next CLI invocation
- *   will surface the error naturally.
- * - When no reusable worktree exists, `createWorktree` runs; on success
- *   the new path is written to `session.metadata` (only when a session
- *   exists — request-scoped worktrees do NOT persist).
- * - Returns `{}` when `worktreeOpt` is undefined/false (preserves
- *   pre-λ behaviour at non-worktree call sites).
- * - Errors propagate as `WorktreeError`/`Error`; the caller wraps them
- *   in a `createErrorResponse` envelope. Do NOT swallow.
- *
- * Spec: docs/plans/slice-lambda.spec.md §"Implementation surface to
- * verify" §5.
- */
 export async function resolveWorktreeForRequest(worktreeOpt, sessionId, runtime) {
     if (!worktreeOpt)
         return {};
@@ -566,30 +422,13 @@ export async function resolveWorktreeForRequest(worktreeOpt, sessionId, runtime)
     }
     return { cwd: handle.path, worktreePath: handle.path };
 }
-/**
- * Slice λ §1.d: response-envelope shape decision for `worktreePath`.
- *
- * We surface the worktree path inline as a stdout prefix
- * (`[gateway] worktree=<absolute-path>\n`) rather than as a
- * structuredContent field or JSON wrapper. Rationale:
- *   - zero schema change across all 10 tools and their downstream parsers
- *   - matches how other slice features (session warnings, cache_state
- *     aggregates) surface side-channel metadata today
- *   - callers that want the path can split on the first newline; callers
- *     that don't care see a single ignorable header line
- *
- * Use `formatWorktreePrefix(resolution.worktreePath)` once per tool, at
- * the moment a successful response is constructed.
- */
 export function formatWorktreePrefix(worktreePath) {
     return worktreePath ? `[gateway] worktree=${worktreePath}\n` : "";
 }
-// Helper function for standardized error responses
 function createErrorResponse(cli, code, stderr, correlationId, error) {
     let errorMessage = `Error executing ${cli} CLI`;
     const isLaunchExit = code === 127 || code === -4058;
     if (error) {
-        // Command not found or spawn error
         errorMessage += `:\n${error.message}`;
         if (error.message.includes("ENOENT")) {
             errorMessage += `\n\nThe '${cli}' command was not found. Please ensure ${cli} CLI is installed and in your PATH.`;
@@ -597,12 +436,10 @@ function createErrorResponse(cli, code, stderr, correlationId, error) {
         logger.error(`[${correlationId || "unknown"}] ${cli} CLI execution failed:`, error.message);
     }
     else if (code === 124) {
-        // Wall-clock timeout
         errorMessage += `: Command timed out\n${stderr}`;
         logger.error(`[${correlationId || "unknown"}] ${cli} CLI timed out`);
     }
     else if (code === 125) {
-        // Idle timeout (stuck process)
         errorMessage += `: Process killed due to inactivity\n${stderr}`;
         logger.error(`[${correlationId || "unknown"}] ${cli} CLI killed due to inactivity`);
     }
@@ -611,7 +448,6 @@ function createErrorResponse(cli, code, stderr, correlationId, error) {
         logger.error(`[${correlationId || "unknown"}] ${cli} CLI failed to launch`);
     }
     else if (code !== 0) {
-        // Other non-zero exit code
         errorMessage += ` (exit code ${code}):\n${stderr}`;
         logger.error(`[${correlationId || "unknown"}] ${cli} CLI failed with exit code ${code}`);
     }
@@ -634,14 +470,7 @@ function createErrorResponse(cli, code, stderr, correlationId, error) {
         },
     };
 }
-export function extractUsageAndCost(cli, output, outputFormat, 
-/**
- * Optional context for off-stdout telemetry sources. Today only Mistral
- * uses this — its meta.json lives on disk keyed by sessionId. Threading
- * this in keeps the closure built by `buildAsyncFlightRecorderHandoff`
- * primitives-only (no `params`/`prep` retention on AsyncJobRecord).
- */
-ctx) {
+export function extractUsageAndCost(cli, output, outputFormat, ctx) {
     if (cli === "claude" && outputFormat === "stream-json") {
         const parsed = parseStreamJson(output);
         if (!parsed.usage) {
@@ -679,29 +508,12 @@ ctx) {
             cacheReadTokens: parsed.usage.cache_read_tokens,
         };
     }
-    // Mistral/Vibe: usage/cost live on disk in `~/.vibe/logs/session/<id>/meta.json`
-    // (Phase 4 slice β). Best-effort: if we don't know the sessionId (fresh
-    // session whose Vibe-assigned UUID we never observed) or the file is
-    // missing/malformed, the parser returns `{}` and the FR row simply lacks
-    // usage data — matching pre-slice behaviour. No stdout fallback exists.
     if (cli === "mistral") {
         return parseVibeMetaJson(ctx?.home ?? homedir(), ctx?.sessionId);
     }
     return {};
 }
-/**
- * Slice 1.5: build the async-job-manager's FR payload from a prep object
- * (which every prepare*Request returns), plus the bound CLI and output
- * format primitives needed by extractUsageAndCost. Returning the closure
- * separately means it captures `cliName` and `fmt` ONLY — never `params`
- * or `prep` — so retention on AsyncJobRecord is O(constant).
- */
 function buildAsyncFlightRecorderHandoff(cliName, prep, sessionId, outputFormat) {
-    // Extract primitives BEFORE building the closure — capturing `prep` or
-    // `params` directly would pin large attachments / promptParts on the
-    // AsyncJobRecord for JOB_TTL_MS. Phase 4 slice β: `sid` and `home` are
-    // primitives too, threaded through so the Mistral branch of
-    // extractUsageAndCost can read `~/.vibe/logs/session/<id>/meta.json`.
     const cli = cliName;
     const fmt = outputFormat;
     const sid = sessionId;
@@ -795,11 +607,7 @@ function resolveClaudeMcpConfig(operation, correlationId, requestedMcpServers, s
     }
     return { config: mcpConfig };
 }
-//──────────────────────────────────────────────────────────────────────────────
-// MCP Resources
-//──────────────────────────────────────────────────────────────────────────────
 function registerBaseResources(server, runtime) {
-    // Register skill resources (L2: full docs, read on demand)
     for (const skill of loadedSkills) {
         server.registerResource(`skill-${skill.name}`, `skills://${skill.name}`, {
             title: skill.name,
@@ -816,7 +624,6 @@ function registerBaseResources(server, runtime) {
         }));
     }
     runtime.logger.info(`Registered ${loadedSkills.length} skill resources`);
-    // Register all sessions resource
     server.registerResource("all-sessions", "sessions://all", {
         title: "📋 All Sessions",
         description: "All conversation sessions across CLIs",
@@ -826,7 +633,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Claude sessions resource
     server.registerResource("claude-sessions", "sessions://claude", {
         title: "🤖 Claude Sessions",
         description: "Claude conversation sessions",
@@ -836,7 +642,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Codex sessions resource
     server.registerResource("codex-sessions", "sessions://codex", {
         title: "💻 Codex Sessions",
         description: "Codex conversation sessions",
@@ -846,7 +651,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Gemini sessions resource
     server.registerResource("gemini-sessions", "sessions://gemini", {
         title: "✨ Gemini Sessions",
         description: "Gemini conversation sessions",
@@ -856,7 +660,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Grok sessions resource
     server.registerResource("grok-sessions", "sessions://grok", {
         title: "⚡ Grok Sessions",
         description: "Grok conversation sessions",
@@ -866,7 +669,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Mistral sessions resource
     server.registerResource("mistral-sessions", "sessions://mistral", {
         title: "🌬 Mistral Sessions",
         description: "Mistral Vibe conversation sessions",
@@ -876,7 +678,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Claude models resource
     server.registerResource("claude-models", "models://claude", {
         title: "🧠 Claude Models",
         description: "Claude models and capabilities",
@@ -886,7 +687,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Codex models resource
     server.registerResource("codex-models", "models://codex", {
         title: "🔧 Codex Models",
         description: "Codex models and capabilities",
@@ -896,7 +696,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Gemini models resource
     server.registerResource("gemini-models", "models://gemini", {
         title: "🌟 Gemini Models",
         description: "Gemini models and capabilities",
@@ -906,7 +705,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Grok models resource
     server.registerResource("grok-models", "models://grok", {
         title: "⚡ Grok Models",
         description: "Grok models and capabilities",
@@ -916,7 +714,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register Mistral models resource
     server.registerResource("mistral-models", "models://mistral", {
         title: "🌬 Mistral Models",
         description: "Mistral Vibe models and capabilities",
@@ -926,7 +723,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Register performance metrics resource
     server.registerResource("performance-metrics", "metrics://performance", {
         title: "📈 Performance Metrics",
         description: "Request counts, latency, success/failure rates",
@@ -936,11 +732,6 @@ function registerBaseResources(server, runtime) {
         const contents = await runtime.resourceProvider.readResource(uri.href);
         return { contents: contents ? [contents] : [] };
     });
-    // Cache-state resources (slice 2). Static URI for global, templated for
-    // session/{id} and prefix/{hash}. All three return tokens/hashes/aggregates
-    // ONLY — never raw prompt or response text. The structural guarantee is in
-    // the SessionCacheStats / PrefixCacheStats / GlobalCacheStats types
-    // themselves: those shapes have no prompt/response/system/task fields.
     server.registerResource("cache-state-global", "cache_state://global", {
         title: "💾 Cache State (Global)",
         description: "Aggregate cache hit/miss/savings across all CLIs in the flight recorder. Tokens/hashes only — no prompt text.",
@@ -999,11 +790,6 @@ function registerBaseResources(server, runtime) {
         };
     });
 }
-/**
- * Slice 1: validate the prompt / promptParts mutex at the prep boundary and
- * return either an error response or the resolved input. The exact error
- * messages are part of the public contract — tests assert them verbatim.
- */
 function resolvePromptOrPartsForPrep(args) {
     const hasPrompt = typeof args.prompt === "string" && args.prompt.length > 0;
     const hasParts = args.promptParts !== undefined;
@@ -1045,7 +831,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
     const assembledPrompt = inputResolution.assembledPrompt;
     const stablePrefixHash = inputResolution.stablePrefixHash;
     const stablePrefixTokens = inputResolution.stablePrefixTokens;
-    // Review integrity check on raw prompt (before optimization)
     const reviewIntegrity = checkReviewIntegrity({
         prompt: assembledPrompt,
         allowedTools: params.allowedTools,
@@ -1058,13 +843,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
             score: reviewIntegrity.totalScore,
         });
     }
-    // Rec #5 (slice κ): refuse the optimizePrompt + cacheControl combo
-    // before running optimization. Optimization rewrites the assembled
-    // prompt text the flight-recorder logs, but the κ stdin payload is
-    // built from raw `promptParts` content blocks — letting both run
-    // produces a FR row whose `prompt` no longer matches what Claude
-    // actually received, AND any optimisation-driven text change would
-    // silently break Anthropic prefix-cache reuse on the next call.
     const ccEarly = params.promptParts?.cacheControl;
     const cacheControlRequestedEarly = !!(ccEarly &&
         (ccEarly.system || ccEarly.tools || ccEarly.context));
@@ -1088,7 +866,7 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
         approvalDecision = runtime.approvalManager.decide({
             cli: "claude",
             operation: params.operation,
-            prompt: assembledPrompt, // Use raw assembled prompt for review-context detection, not optimized
+            prompt: assembledPrompt,
             bypassRequested: params.dangerouslySkipPermissions,
             fullAuto: false,
             requestedMcpServers,
@@ -1102,18 +880,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
             return createApprovalDeniedResponse(params.operation, approvalDecision);
         }
     }
-    // Rec #2 (slice κ): auto-emit `cache_control` when the caller passes
-    // `promptParts` whose stable prefix exceeds the per-model minimum,
-    // the caller has NOT explicitly set `cacheControl`, the gateway
-    // config has opted in (`[cache_awareness].emit_anthropic_cache_control`),
-    // and outputFormat is stream-json. Auto-emit marks the LAST non-empty
-    // stable block (context → tools → system priority — the rightmost
-    // stable block covers the widest prefix). Skipped when optimizePrompt
-    // is on (same rec #5 desync risk).
-    //
-    // The 1h ttl is forced regardless of `anthropic_ttl_seconds`: 5m
-    // breakpoints from caller content are rejected by Anthropic once
-    // Claude Code's own 1h-marked session-wrap blocks land ahead of them.
     let autoEmittedCacheControlBlock = null;
     if (!cacheControlRequestedEarly &&
         runtime.cacheAwareness.emitAnthropicCacheControl &&
@@ -1124,9 +890,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
         const threshold = minStableTokensForModel(runtime.cacheAwareness, resolvedModel ?? "default");
         if (stablePrefixTokens >= threshold) {
             const pp = params.promptParts;
-            // Rightmost non-empty stable block — its cache_control breakpoint
-            // covers everything above it in the message (the API matches
-            // breakpoints in order).
             if (pp.context && pp.context.length > 0)
                 autoEmittedCacheControlBlock = "context";
             else if (pp.tools && pp.tools.length > 0)
@@ -1141,12 +904,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
             }
         }
     }
-    // Rec #4: warn when promptParts has a cacheable stable prefix but no
-    // cache_control breakpoint is being emitted (neither explicit nor
-    // auto). Either the caller forgot to set `cacheControl` or
-    // `[cache_awareness].emit_anthropic_cache_control` is off — both
-    // leave the stable prefix bytes unreused across calls, defeating the
-    // point of using `promptParts`.
     const warnings = [];
     if (!cacheControlRequestedEarly &&
         autoEmittedCacheControlBlock === null &&
@@ -1168,13 +925,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
             });
         }
     }
-    // Slice κ: switch from the legacy positional `-p <prompt>` emission
-    // to `claude -p --input-format stream-json` and feed a JSON
-    // content-blocks payload via stdin. Non-κ callers (no cacheControl,
-    // or cacheControl with all flags false) take the existing positional
-    // path bit-for-bit. The κ path activates on EITHER an explicit caller
-    // opt-in (`cacheControlRequestedEarly`) OR a gateway-driven auto-emit
-    // (`autoEmittedCacheControlBlock`).
     const cacheControlRequested = cacheControlRequestedEarly || autoEmittedCacheControlBlock !== null;
     let stdinPayload;
     let cacheControlBlocks;
@@ -1182,9 +932,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
         if (params.outputFormat !== "stream-json") {
             return createErrorResponse(params.operation, 1, "", corrId, new Error("promptParts.cacheControl requires outputFormat: 'stream-json' (slice κ pipes the cache_control blocks over --input-format stream-json; text/json output formats cannot carry the required NDJSON usage events)."));
         }
-        // promptParts is non-null whenever cacheControlRequested is true
-        // (explicit opt-in lives in PromptParts; auto-emit guard requires
-        // promptParts to be defined).
         const effectiveParts = autoEmittedCacheControlBlock !== null
             ? {
                 ...params.promptParts,
@@ -1216,11 +963,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
             args.push("--output-format", "json");
         }
         else if (params.outputFormat === "stream-json") {
-            // Claude CLI 2.x rejects `--print --output-format stream-json` without
-            // `--verbose`: "When using --print, --output-format=stream-json requires
-            // --verbose". --verbose only affects what claude logs to stderr; the
-            // stream-json stdout payload is unchanged, so the gateway's NDJSON
-            // parser is unaffected.
             args.push("--output-format", "stream-json", "--include-partial-messages", "--verbose");
         }
     }
@@ -1251,7 +993,6 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
             args.push("--strict-mcp-config");
         }
     }
-    // U25: Claude high-impact features (agent, agents, fork, system-prompt, budget, effort, …)
     let validatedAgents;
     if (params.agents && Object.keys(params.agents).length > 0) {
         const result = validateClaudeAgentsMap(params.agents);
@@ -1273,6 +1014,10 @@ export function prepareClaudeRequest(params, runtime = resolveGatewayServerRunti
         fallbackModel: params.fallbackModel,
         jsonSchema: params.jsonSchema,
         addDir: params.addDir,
+        noSessionPersistence: params.noSessionPersistence,
+        settingSources: params.settingSources,
+        settings: params.settings,
+        tools: params.tools,
     }));
     return {
         corrId,
@@ -1305,7 +1050,6 @@ export function prepareCodexRequest(params, runtime = resolveGatewayServerRuntim
     const assembledPrompt = inputResolution.assembledPrompt;
     const stablePrefixHash = inputResolution.stablePrefixHash;
     const stablePrefixTokens = inputResolution.stablePrefixTokens;
-    // Review integrity check on raw prompt (before optimization)
     const reviewIntegrity = checkReviewIntegrity({ prompt: assembledPrompt });
     if (reviewIntegrity.violations.length > 0) {
         runtime.logger.info(`[${corrId}] Review integrity violations detected: ${reviewIntegrity.violations.map(v => v.type).join(", ")}`, {
@@ -1326,7 +1070,7 @@ export function prepareCodexRequest(params, runtime = resolveGatewayServerRuntim
         approvalDecision = runtime.approvalManager.decide({
             cli: "codex",
             operation: params.operation,
-            prompt: assembledPrompt, // Use raw assembled prompt for review-context detection, not optimized
+            prompt: assembledPrompt,
             bypassRequested: params.dangerouslyBypassApprovalsAndSandbox,
             fullAuto: params.fullAuto,
             requestedMcpServers,
@@ -1338,9 +1082,6 @@ export function prepareCodexRequest(params, runtime = resolveGatewayServerRuntim
             return createApprovalDeniedResponse(params.operation, approvalDecision);
         }
     }
-    // Resume mode: codex exec resume <SESSION_ID|--last> [flags] PROMPT
-    // Note: `codex exec resume` does NOT accept sandbox policy flags; the original
-    // session's approval policy is inherited. We silently drop fullAuto on resume.
     let sessionPlan;
     try {
         sessionPlan = resolveCodexSessionArgs({
@@ -1361,9 +1102,6 @@ export function prepareCodexRequest(params, runtime = resolveGatewayServerRuntim
     }
     if (resolvedModel)
         args.push("--model", resolvedModel);
-    // Codex sandbox / approval: resolve modern flags + legacy fullAuto shorthand.
-    // `codex exec resume` rejects all of these (the original session's policy is
-    // inherited), so we only emit them when starting a NEW session.
     const sandboxFlags = resolveCodexSandboxFlags({
         sandboxMode: params.sandboxMode,
         askForApproval: params.askForApproval,
@@ -1379,26 +1117,12 @@ export function prepareCodexRequest(params, runtime = resolveGatewayServerRuntim
     if (params.dangerouslyBypassApprovalsAndSandbox) {
         args.push("--dangerously-bypass-approvals-and-sandbox");
     }
-    // U23 fix: emit `--json` when the caller asked for JSON output so the
-    // codex-json-parser actually receives JSONL events. This is what makes
-    // extractUsageAndCost() reachable from the tool surface; without it, the
-    // U23 parser is dead code.
     if (params.outputFormat === "json") {
         args.push("--json");
     }
     args.push("--skip-git-repo-check");
-    // U26: High-impact feature flags. `--search` is retained as a compatibility
-    // input but current `codex exec` no longer accepts it, so the helper warns
-    // and emits no argv. `--profile` is accepted for new sessions only. The other
-    // flags here are accepted on resume per `codex exec resume --help` and are
-    // emitted in both branches.
     let highImpactCleanup;
     if (sessionPlan.mode === "new") {
-        // Phase 4 slice ζ: emit working-dir and add-dir on new sessions only.
-        // Both flags are listed in CODEX_RESUME_FILTERED_FLAGS — resume inherits
-        // the original session's cwd and writable-dir policy, so emitting them
-        // on resume would be silently stripped (wasteful + misleading on argv
-        // logs). Gating here mirrors `--search` / `--sandbox`.
         if (params.workingDir) {
             args.push("-C", params.workingDir);
         }
@@ -1481,7 +1205,6 @@ export function prepareGeminiRequest(params, runtime = resolveGatewayServerRunti
     const assembledPrompt = inputResolution.assembledPrompt;
     const stablePrefixHash = inputResolution.stablePrefixHash;
     const stablePrefixTokens = inputResolution.stablePrefixTokens;
-    // Review integrity check on raw prompt (before optimization)
     const reviewIntegrity = checkReviewIntegrity({
         prompt: assembledPrompt,
         allowedTools: params.allowedTools,
@@ -1505,8 +1228,8 @@ export function prepareGeminiRequest(params, runtime = resolveGatewayServerRunti
         approvalDecision = runtime.approvalManager.decide({
             cli: "gemini",
             operation: params.operation,
-            prompt: assembledPrompt, // Use raw assembled prompt for review-context detection, not optimized
-            bypassRequested: params.approvalMode === "yolo",
+            prompt: assembledPrompt,
+            bypassRequested: params.approvalMode === "yolo" || params.yolo === true,
             fullAuto: false,
             requestedMcpServers,
             allowedTools: params.allowedTools,
@@ -1519,8 +1242,6 @@ export function prepareGeminiRequest(params, runtime = resolveGatewayServerRunti
         }
     }
     const effectiveApprovalMode = params.approvalStrategy === "mcp_managed" ? "yolo" : params.approvalMode;
-    // U27: Validate high-impact policy paths and prepend attachment tokens
-    // BEFORE the `-p` pair is emitted, preserving the U21 ordering invariant.
     const highImpact = prepareGeminiHighImpactFlags({
         sandbox: params.sandbox,
         policyFiles: params.policyFiles,
@@ -1537,15 +1258,14 @@ export function prepareGeminiRequest(params, runtime = resolveGatewayServerRunti
             return createErrorResponse(params.operation, 1, "", corrId, err instanceof Error ? err : new Error(String(err)));
         }
     }
-    // U21: Emit the prompt via -p/--prompt rather than as a positional argument.
-    // Positional prompts depend on Gemini's TTY/mode-detection heuristics; -p is
-    // the documented non-interactive flag and is robust against future CLI mode
-    // changes.
     const args = ["-p", effectivePrompt];
     if (resolvedModel)
         args.push("--model", resolvedModel);
     if (effectiveApprovalMode)
         args.push("--approval-mode", effectiveApprovalMode);
+    if (params.yolo && effectiveApprovalMode !== "yolo") {
+        args.push("--yolo");
+    }
     if (params.allowedTools && params.allowedTools.length > 0) {
         sanitizeCliArgValues(params.allowedTools, "allowedTools");
         params.allowedTools.forEach(tool => args.push("--allowed-tools", tool));
@@ -1558,26 +1278,13 @@ export function prepareGeminiRequest(params, runtime = resolveGatewayServerRunti
         sanitizeCliArgValues(params.includeDirs, "includeDirs");
         params.includeDirs.forEach(dir => args.push("--include-directories", dir));
     }
-    // U27 high-impact flags (-s / --policy / --admin-policy) appended after the
-    // existing flag set so positional ordering relative to `-p` is preserved.
     args.push(...highImpact.args);
-    // U23 fix: emit `-o json` when the caller asked for JSON output. The Gemini
-    // JSON parser is otherwise unreachable from the tool surface and the
-    // structured usageMetadata is silently dropped.
-    //
-    // Phase 4 slice ε: same wiring for `-o stream-json` (NDJSON event stream).
-    // Gemini already streams stdout in real-time so the existing 10-minute
-    // idle timeout (CLI_IDLE_TIMEOUTS.gemini) covers both modes without
-    // adjustment — unlike Claude, no `--include-partial-messages` companion
-    // flag is required because Gemini emits assistant `delta` events as part
-    // of the default stream-json shape.
     if (params.outputFormat === "json") {
         args.push("-o", "json");
     }
     else if (params.outputFormat === "stream-json") {
         args.push("-o", "stream-json");
     }
-    // Phase 4 slice γ: opt-in trust-prompt bypass for fresh workspaces.
     if (params.skipTrust) {
         args.push("--skip-trust");
     }
@@ -1608,7 +1315,6 @@ export function prepareGrokRequest(params, runtime = resolveGatewayServerRuntime
     const assembledPrompt = inputResolution.assembledPrompt;
     const stablePrefixHash = inputResolution.stablePrefixHash;
     const stablePrefixTokens = inputResolution.stablePrefixTokens;
-    // Review integrity check on raw prompt (before optimization)
     const reviewIntegrity = checkReviewIntegrity({
         prompt: assembledPrompt,
         allowedTools: params.allowedTools,
@@ -1633,7 +1339,7 @@ export function prepareGrokRequest(params, runtime = resolveGatewayServerRuntime
         approvalDecision = runtime.approvalManager.decide({
             cli: "grok",
             operation: params.operation,
-            prompt: assembledPrompt, // Use raw assembled prompt for review-context detection, not optimized
+            prompt: assembledPrompt,
             bypassRequested: Boolean(params.alwaysApprove) || params.permissionMode === "bypassPermissions",
             fullAuto: false,
             requestedMcpServers,
@@ -1694,6 +1400,12 @@ export function prepareGrokRequest(params, runtime = resolveGatewayServerRuntime
             args.push("--deny", rule);
         }
     }
+    if (params.compactionMode) {
+        args.push("--compaction-mode", params.compactionMode);
+    }
+    if (params.compactionDetail) {
+        args.push("--compaction-detail", params.compactionDetail);
+    }
     return {
         corrId,
         effectivePrompt,
@@ -1762,9 +1474,6 @@ export function prepareMistralRequest(params, runtime = resolveGatewayServerRunt
             return createApprovalDeniedResponse(params.operation, approvalDecision);
         }
     }
-    // Under mcp_managed, force --agent auto-approve so the approval gate's
-    // verdict carries through to the CLI invocation (mirrors Grok's --always-approve
-    // forcing under mcp_managed).
     const effectivePermissionMode = params.approvalStrategy === "mcp_managed"
         ? "auto-approve"
         : (params.permissionMode ?? "auto-approve");
@@ -1773,8 +1482,6 @@ export function prepareMistralRequest(params, runtime = resolveGatewayServerRunt
         resolvedModel,
         outputFormat: params.outputFormat,
         permissionMode: effectivePermissionMode,
-        effort: params.effort,
-        reasoningEffort: params.reasoningEffort,
         allowedTools: params.allowedTools,
         disallowedTools: params.disallowedTools,
         trust: params.trust,
@@ -1813,15 +1520,6 @@ function selectMistralRecoveryModel(failedModel) {
     ].filter((model) => Boolean(model && model !== failedModel));
     return candidates.find(model => model !== "local");
 }
-/**
- * Phase 4 slice δ post-review: pure helper extracted from
- * `handleMistralRequest` so the retry-path arg-preservation invariants
- * (trust + maxTurns + maxPrice from slices γ/δ) are unit-testable
- * without mocking awaitJobOrDefer. Any param the wrapper threads into
- * the FIRST `buildMistralCliInvocation` call MUST also be threaded
- * through here, or a fresh-workspace / budgeted run can degrade on
- * the second attempt.
- */
 export function buildMistralRetryPrep(params, recoveryModel) {
     return buildMistralCliInvocation({
         prompt: params.effectivePrompt,
@@ -1830,8 +1528,6 @@ export function buildMistralRetryPrep(params, recoveryModel) {
         permissionMode: params.approvalStrategy === "mcp_managed"
             ? "auto-approve"
             : (params.permissionMode ?? "auto-approve"),
-        effort: params.effort,
-        reasoningEffort: params.reasoningEffort,
         allowedTools: params.allowedTools,
         disallowedTools: params.disallowedTools,
         trust: params.trust,
@@ -1844,13 +1540,11 @@ export function buildMistralRetryPrep(params, recoveryModel) {
 }
 function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionId, prep, durationMs, resumable, outputFormat, warnings) {
     let finalStdout = stdout;
-    // Skip response optimization for JSON output to prevent corrupting structured data
     if (optimizeResponse && outputFormat !== "json") {
         const optimized = optimizeResponseText(finalStdout);
         logOptimizationTokens("response", corrId, finalStdout, optimized);
         finalStdout = optimized;
     }
-    // Append review integrity warnings to response text (skip for JSON output to avoid corruption)
     if (prep.reviewIntegrity &&
         prep.reviewIntegrity.violations.length > 0 &&
         outputFormat !== "json") {
@@ -1867,9 +1561,6 @@ function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionId, prep
             correlationId: corrId,
             sessionId: sessionId || null,
             durationMs,
-            // Phase 4 slice β: thread sessionId + home so the Mistral branch of
-            // extractUsageAndCost can read `~/.vibe/logs/session/<dir>/meta.json`.
-            // Other CLIs ignore the ctx (their usage source is stdout).
             ...extractUsageAndCost(cli, stdout, outputFormat, { sessionId, home: homedir() }),
             exitCode: 0,
             retryCount: 0,
@@ -1899,12 +1590,6 @@ function buildCliResponse(cli, stdout, optimizeResponse, corrId, sessionId, prep
     }
     return response;
 }
-/**
- * Slice 3 helper: compute the cache_ttl_expiring_soon warning for a
- * claude session, if the feature is enabled, the session has prior cache
- * writes, and ttlRemainingMs is below the threshold (30s by default).
- * Returns null when no warning applies.
- */
 function maybeBuildCacheTtlWarning(args) {
     if (args.cli !== "claude")
         return null;
@@ -1933,7 +1618,6 @@ function resolveHandlerRuntime(deps) {
     if (deps.runtime)
         return deps.runtime;
     const asyncDeps = deps;
-    // Older HandlerDeps callers may not provide `warn`; default-route to `info`.
     const depLogger = deps.logger;
     const normalizedLogger = {
         info: depLogger.info,
@@ -1969,6 +1653,7 @@ export async function handleGeminiRequest(deps, params) {
         adminPolicyFiles: params.adminPolicyFiles,
         attachments: params.attachments,
         skipTrust: params.skipTrust,
+        yolo: params.yolo,
     }, runtime);
     if (!("args" in prep))
         return prep;
@@ -1986,8 +1671,6 @@ export async function handleGeminiRequest(deps, params) {
     }, runtime);
     deps.logger.info(`[${corrId}] gemini_request invoked with model=${prep.resolvedModel || "default"}, approvalMode=${params.approvalMode}, prompt length=${prep.effectivePrompt.length}`);
     try {
-        // Gemini CLI 0.43 supports `--resume`, but not a supported fresh
-        // `--session-id` flag. Fresh sessions emit no session flag.
         const sessionPlan = resolveGeminiSessionPlan({
             sessionId: params.sessionId,
             resumeLatest: params.resumeLatest,
@@ -2005,7 +1688,6 @@ export async function handleGeminiRequest(deps, params) {
         }
         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, worktreeResolution.cwd);
-        // Deferred — job still running, return async reference
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, effectiveSessionIdHint);
         }
@@ -2026,9 +1708,6 @@ export async function handleGeminiRequest(deps, params) {
             return createErrorResponse("gemini", code, stderr, corrId);
         }
         wasSuccessful = true;
-        // Post-success session I/O for explicit resume flows. Fresh Gemini sessions
-        // are owned by the CLI because the current CLI has no supported fresh
-        // session-id flag the gateway can inject.
         let effectiveSessionId = effectiveSessionIdHint;
         if (effectiveSessionId) {
             const existing = await deps.sessionManager.getSession(effectiveSessionId);
@@ -2111,19 +1790,18 @@ export async function handleGeminiRequestAsync(deps, params) {
         adminPolicyFiles: params.adminPolicyFiles,
         attachments: params.attachments,
         skipTrust: params.skipTrust,
+        yolo: params.yolo,
     }, runtime);
     if (!("args" in prep))
         return prep;
     const { corrId, args, requestedMcpServers, approvalDecision } = prep;
     try {
-        // Gemini CLI 0.43 supports `--resume`, but fresh sessions emit no session flag.
         const sessionPlan = resolveGeminiSessionPlan({
             sessionId: params.sessionId,
             resumeLatest: params.resumeLatest,
             createNewSession: params.createNewSession,
         });
         args.push(...sessionPlan.args);
-        // Pre-start session I/O (async handlers: prevent orphaned jobs)
         let effectiveSessionId = sessionPlan.resumed ? params.sessionId : undefined;
         if (effectiveSessionId) {
             const existing = await deps.sessionManager.getSession(effectiveSessionId);
@@ -2146,13 +1824,8 @@ export async function handleGeminiRequestAsync(deps, params) {
         catch (err) {
             return createErrorResponse("gemini_request_async", 1, "", corrId, err);
         }
-        // Start job only after all session I/O succeeds. U23: forward outputFormat
-        // so AsyncJobManager records it in the durable store (the manager also
-        // surfaces it in the snapshot).
         assertUpstreamCliArgs("gemini", args);
         assertUpstreamCliEnv("gemini", undefined);
-        // Slice 1.5: pure async path — no upstream safeFlightStart, so the
-        // manager owns both logStart and logComplete for this corrId.
         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);
         deps.logger.info(`[${corrId}] gemini_request_async started job ${job.id}`);
@@ -2210,6 +1883,8 @@ export async function handleGrokRequest(deps, params) {
         systemPromptOverride: params.systemPromptOverride,
         allow: params.allow,
         deny: params.deny,
+        compactionMode: params.compactionMode,
+        compactionDetail: params.compactionDetail,
     }, runtime);
     if (!("args" in prep))
         return prep;
@@ -2227,7 +1902,6 @@ export async function handleGrokRequest(deps, params) {
     }, runtime);
     deps.logger.info(`[${corrId}] grok_request invoked with model=${prep.resolvedModel || "default"}, permissionMode=${params.permissionMode}, prompt length=${prep.effectivePrompt.length}`);
     try {
-        // Session arg planning (pure, no I/O)
         const sessionResult = resolveGrokSessionArgs({
             sessionId: params.sessionId,
             resumeLatest: params.resumeLatest,
@@ -2243,7 +1917,6 @@ export async function handleGrokRequest(deps, params) {
         }
         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);
-        // Deferred — job still running, return async reference
         if (isDeferredResponse(result)) {
             return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
         }
@@ -2264,7 +1937,6 @@ export async function handleGrokRequest(deps, params) {
             return createErrorResponse("grok", code, stderr, corrId);
         }
         wasSuccessful = true;
-        // Post-success session I/O (sync handlers: no phantom sessions on CLI failure)
         let effectiveSessionId = sessionResult.effectiveSessionId;
         if (sessionResult.userProvidedSession && effectiveSessionId) {
             const existing = await deps.sessionManager.getSession(effectiveSessionId);
@@ -2350,19 +2022,19 @@ export async function handleGrokRequestAsync(deps, params) {
         systemPromptOverride: params.systemPromptOverride,
         allow: params.allow,
         deny: params.deny,
+        compactionMode: params.compactionMode,
+        compactionDetail: params.compactionDetail,
     }, runtime);
     if (!("args" in prep))
         return prep;
     const { corrId, args, requestedMcpServers, approvalDecision } = prep;
     try {
-        // Session arg planning (pure, no I/O)
         const sessionResult = resolveGrokSessionArgs({
             sessionId: params.sessionId,
             resumeLatest: params.resumeLatest,
             createNewSession: params.createNewSession,
         });
         args.push(...sessionResult.resumeArgs);
-        // Pre-start session I/O (async handlers: prevent orphaned jobs)
         let effectiveSessionId = sessionResult.effectiveSessionId;
         if (sessionResult.userProvidedSession && effectiveSessionId) {
             const existing = await deps.sessionManager.getSession(effectiveSessionId);
@@ -2389,7 +2061,6 @@ export async function handleGrokRequestAsync(deps, params) {
         catch (err) {
             return createErrorResponse("grok_request_async", 1, "", corrId, err);
         }
-        // Start job only after all session I/O succeeds
         assertUpstreamCliArgs("grok", args);
         assertUpstreamCliEnv("grok", undefined);
         const grokAsyncFrHandoff = buildAsyncFlightRecorderHandoff("grok", prep, effectiveSessionId, params.outputFormat);
@@ -2431,8 +2102,6 @@ export async function handleMistralRequest(deps, params) {
         model: params.model,
         outputFormat: params.outputFormat,
         permissionMode: params.permissionMode,
-        effort: params.effort,
-        reasoningEffort: params.reasoningEffort,
         allowedTools: params.allowedTools,
         disallowedTools: params.disallowedTools,
         approvalStrategy: params.approvalStrategy,
@@ -2488,8 +2157,6 @@ 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];
-                // Reuse the FR handoff built above — the retry preserves corrId,
-                // so the manager's logComplete still updates the original row.
                 result = await awaitJobOrDefer("mistral", retryArgs, corrId, resolveIdleTimeout("mistral", params.idleTimeoutMs), params.outputFormat, true, runtime, retryPrep.env, undefined, mistralFrHandoff.flightRecorderEntry, mistralFrHandoff.extractUsage, undefined, worktreeResolution.cwd);
                 if (isDeferredResponse(result)) {
                     return buildDeferredToolResponse(result, sessionResult.effectiveSessionId);
@@ -2582,8 +2249,6 @@ export async function handleMistralRequestAsync(deps, params) {
         model: params.model,
         outputFormat: params.outputFormat,
         permissionMode: params.permissionMode,
-        effort: params.effort,
-        reasoningEffort: params.reasoningEffort,
         allowedTools: params.allowedTools,
         disallowedTools: params.disallowedTools,
         approvalStrategy: params.approvalStrategy,
@@ -2702,11 +2367,6 @@ export async function handleCodexRequestAsync(deps, params) {
     if (!("args" in prep))
         return prep;
     const { corrId, args, requestedMcpServers, approvalDecision } = prep;
-    // U26 fix: outputSchema temp-file ownership. The cleanup callable lives in
-    // exactly one place at a time: this scope until startJob succeeds, then
-    // AsyncJobManager (via onComplete → persistComplete → fireOnComplete) once
-    // the job is registered. Any code path that fails to hand it off MUST run
-    // it locally.
     const prepCleanup = "cleanup" in prep && typeof prep.cleanup === "function" ? prep.cleanup : undefined;
     let prepCleanupOwnedHere = prepCleanup !== undefined;
     const runPrepCleanupLocally = () => {
@@ -2721,7 +2381,6 @@ export async function handleCodexRequestAsync(deps, params) {
         }
     };
     try {
-        // Pre-start session I/O (async handlers: prevent orphaned jobs)
         let effectiveSessionId = params.sessionId;
         if (!params.createNewSession && !params.sessionId) {
             const activeSession = await deps.sessionManager.getActiveSession("codex");
@@ -2740,9 +2399,6 @@ export async function handleCodexRequestAsync(deps, params) {
             const newSession = await deps.sessionManager.createSession("codex", "Codex Session");
             effectiveSessionId = newSession.id;
         }
-        // Slice λ: resolve worktree directive after session I/O so resume reuse
-        // can read metadata.worktreePath. A pre-startJob failure here means
-        // prepCleanup is still owned locally; run it before returning.
         let worktreeResolution = {};
         try {
             worktreeResolution = await resolveWorktreeForRequest(params.worktree, effectiveSessionId, runtime);
@@ -2751,22 +2407,15 @@ export async function handleCodexRequestAsync(deps, params) {
             runPrepCleanupLocally();
             return createErrorResponse("codex_request_async", 1, "", corrId, err);
         }
-        // Start job only after all session I/O succeeds. If startJob throws before
-        // registering the record, ownership stays here and we run it in the catch.
         assertUpstreamCliArgs("codex", args);
         assertUpstreamCliEnv("codex", undefined);
         const codexAsyncFrHandoff = buildAsyncFlightRecorderHandoff("codex", prep, effectiveSessionId, params.outputFormat);
         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);
-            // Handoff succeeded: AsyncJobManager will fire prepCleanup on terminal
-            // status. Release our local ownership claim so the catch path doesn't
-            // double-fire.
             prepCleanupOwnedHere = false;
         }
         catch (startErr) {
-            // startJob never stored the record → manager won't call onComplete. We
-            // still own the cleanup; let the outer catch run it.
             throw startErr;
         }
         deps.logger.info(`[${corrId}] codex_request_async started job ${job.id}`);
@@ -2793,42 +2442,15 @@ export async function handleCodexRequestAsync(deps, params) {
         };
     }
     catch (error) {
-        // Pre-start failure: either session I/O threw, or startJob threw before
-        // registering the record. In either case the manager will NOT fire
-        // prepCleanup, so we must run it here.
         runPrepCleanupLocally();
         return createErrorResponse("codex_request_async", 1, "", corrId, error);
     }
 }
-//──────────────────────────────────────────────────────────────────────────────
-// Claude Code Tool
-//──────────────────────────────────────────────────────────────────────────────
 export function createGatewayServer(deps = {}) {
     const runtime = resolveGatewayServerRuntime(deps, { isolateState: true });
     const { sessionManager, asyncJobManager, approvalManager, performanceMetrics, logger, persistence, flightRecorder, cacheAwareness, } = runtime;
-    // `flightRecorder` is destructured into closure scope so the session_get
-    // handler (see ~line 5590) has the FlightRecorderQuery read capability
-    // available without re-resolving runtime. Slice 2 will populate the
-    // `cacheState` field of session_get's response from this read surface.
-    // `cacheAwareness` is the loaded [cache_awareness] block (config.ts).
     void flightRecorder;
     void cacheAwareness;
-    // Structural invariant: tools register iff ALL THREE conditions hold:
-    //   (1) persistence.backend !== "none"  — the operator/config has not
-    //       explicitly disabled durable persistence;
-    //   (2) persistence.asyncJobsEnabled === true — the derived opt-in flag
-    //       agrees (loadPersistenceConfig sets this iff backend is one of
-    //       sqlite/postgres/memory);
-    //   (3) asyncJobManager.hasStore() === true — the runtime manager
-    //       actually has a store attached (isolate-mode runtimes use null).
-    //
-    // Each guard closes a distinct re-entry path for the silent-loss footgun:
-    //   - Without (1), a caller can inject {backend:'none', asyncJobsEnabled:true}
-    //     and re-advertise the async tools while reporting backend='none' in
-    //     llm_process_health — exactly contradicting SPEC CLAIM 4f.
-    //   - Without (2), config that opts out is ignored.
-    //   - Without (3), a null-store manager (isolate-mode / HTTP per-session)
-    //     accepts registrations that have nowhere to persist results.
     const asyncJobsEnabled = persistence.backend !== "none" && persistence.asyncJobsEnabled && asyncJobManager.hasStore();
     const server = newGatewayMcpServer();
     registerBaseResources(server, runtime);
@@ -2865,7 +2487,6 @@ export function createGatewayServer(deps = {}) {
             .enum(CLAUDE_PERMISSION_MODES)
             .optional()
             .describe("Claude --permission-mode: default|acceptEdits|plan|auto|dontAsk|bypassPermissions. `default` is a no-op (no flag emitted)."),
-        // U25 — Claude high-impact features
         agent: z
             .string()
             .optional()
@@ -2905,7 +2526,6 @@ export function createGatewayServer(deps = {}) {
             .boolean()
             .optional()
             .describe("Claude --exclude-dynamic-system-prompt-sections: trim dynamic context blocks from the system prompt."),
-        // Phase 4 slice η — Claude reliability + structured-output parity
         fallbackModel: z
             .string()
             .min(1)
@@ -2915,11 +2535,28 @@ export function createGatewayServer(deps = {}) {
             .union([z.string(), z.record(z.string(), z.unknown())])
             .optional()
             .describe("Claude --json-schema: JSON Schema literal (NOT a path) constraining structured output. Object values are JSON.stringify-d; string values are passed verbatim. Use with outputFormat='json'."),
-        // Phase 4 slice ζ — Claude additional-workspace-dirs parity
         addDir: z
             .array(z.string())
             .optional()
             .describe("Claude --add-dir: additional directories the CLI is allowed to read/write beyond the process cwd. Each entry is emitted as its own --add-dir instance."),
+        noSessionPersistence: z
+            .boolean()
+            .optional()
+            .describe("Claude --no-session-persistence: do not write this session to disk (ephemeral one-shot runs; mirrors codex --ephemeral)."),
+        settingSources: z
+            .string()
+            .min(1)
+            .optional()
+            .describe("Claude --setting-sources: comma-separated setting sources to load (user|project|local) for reproducible/isolated headless runs."),
+        settings: z
+            .string()
+            .min(1)
+            .optional()
+            .describe("Claude --settings: path to a settings JSON file or a JSON literal of additional settings. Powerful: settings can define hooks/permissions/model; passed verbatim."),
+        tools: z
+            .array(z.string())
+            .optional()
+            .describe('Claude --tools: restrict the available built-in tool set (distinct from allowedTools permission gating). Pass [""] to disable all tools.'),
         worktree: WORKTREE_SCHEMA.optional(),
         approvalStrategy: z
             .enum(["legacy", "mcp_managed"])
@@ -2951,7 +2588,7 @@ export function createGatewayServer(deps = {}) {
             .boolean()
             .default(false)
             .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
-    }, async ({ prompt, promptParts, model, outputFormat, sessionId, continueSession, createNewSession, allowedTools, disallowedTools, dangerouslySkipPermissions, permissionMode, agent, agents, forkSession, systemPrompt, appendSystemPrompt, maxBudgetUsd, maxTurns, effort, excludeDynamicSystemPromptSections, fallbackModel, jsonSchema, addDir, 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, worktree, approvalStrategy, approvalPolicy, mcpServers, strictMcpConfig, correlationId, optimizePrompt, optimizeResponse, 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)."));
@@ -2984,24 +2621,18 @@ export function createGatewayServer(deps = {}) {
             fallbackModel,
             jsonSchema,
             addDir,
+            noSessionPersistence,
+            settingSources,
+            settings,
+            tools,
         }, runtime);
         if (!("args" in prep))
             return prep;
         const { corrId, args } = prep;
         let durationMs = 0;
         let wasSuccessful = false;
-        // Session resolution happens BEFORE safeFlightStart so that:
-        //   (1) the TTL warning reads the PRIOR session's lastWriteAt
-        //       rather than the row about to be inserted (codex-r1/F1).
-        //   (2) the flight-recorder row is tagged with effectiveSessionId
-        //       (the session the CLI will actually resume), not the raw
-        //       user-provided sessionId.
         let effectiveSessionId = sessionId;
         let useContinue = continueSession;
-        // Guard the active-session lookup: in some test harnesses the
-        // sessionManager is undefined; the original try-catch wrapped this
-        // block, so we replicate that tolerance here. Failure leaves
-        // effectiveSessionId as the user-provided sessionId.
         let activeSession = null;
         try {
             activeSession = await sessionManager.getActiveSession("claude");
@@ -3016,16 +2647,11 @@ export function createGatewayServer(deps = {}) {
         if (!useContinue && effectiveSessionId && activeSession?.id === effectiveSessionId) {
             useContinue = true;
         }
-        // Slice 3: if the resolved session has a near-expiry Anthropic
-        // cache breakpoint, attach a structured warning (NOT a hard error)
-        // to the response. Computed BEFORE safeFlightStart so the current
-        // row does not skew lastRequestAt.
         const ttlWarning = maybeBuildCacheTtlWarning({
             runtime,
             sessionId: effectiveSessionId,
             cli: "claude",
         });
-        // Rec #4: include any prep-time warnings (e.g. cacheable_prefix_uncached).
         const warnings = [
             ...(ttlWarning ? [ttlWarning] : []),
             ...(prep.warnings ?? []),
@@ -3049,8 +2675,6 @@ export function createGatewayServer(deps = {}) {
                 args.push("--session-id", effectiveSessionId);
                 await sessionManager.updateSessionUsage(effectiveSessionId);
             }
-            // Slice λ: resolve worktree directive into spawn cwd. Done after
-            // session resolution so resume reuse can read metadata.worktreePath.
             let worktreeResolution = {};
             try {
                 worktreeResolution = await resolveWorktreeForRequest(worktree, effectiveSessionId, runtime);
@@ -3058,11 +2682,9 @@ export function createGatewayServer(deps = {}) {
             catch (err) {
                 return createErrorResponse("claude_request", 1, "", corrId, err);
             }
-            // Idle timeout only for stream-json (text/json produce no output until done)
             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);
-            // Deferred — job still running, return async reference
             if (isDeferredResponse(result)) {
                 return buildDeferredToolResponse(result, effectiveSessionId);
             }
@@ -3080,9 +2702,6 @@ export function createGatewayServer(deps = {}) {
                     errorMessage: stderr || `Exit code ${code}`,
                     status: "failed",
                 }, runtime);
-                // Slice 3: attach any computed warnings to the error response so
-                // the caller still sees cache_ttl_expiring_soon when the CLI
-                // happens to fail for an unrelated reason.
                 const errResp = createErrorResponse("claude", code, stderr, corrId);
                 if (warnings.length > 0) {
                     errResp.warnings = warnings;
@@ -3090,7 +2709,6 @@ export function createGatewayServer(deps = {}) {
                 return errResp;
             }
             wasSuccessful = true;
-            // If we used a session ID and it's not tracked yet, create a session record
             if (effectiveSessionId) {
                 const existingSession = await sessionManager.getSession(effectiveSessionId);
                 if (!existingSession) {
@@ -3098,7 +2716,6 @@ export function createGatewayServer(deps = {}) {
                 }
             }
             logger.info(`[${corrId}] claude_request completed successfully in ${durationMs}ms`);
-            // Parse stream-json NDJSON output to extract result text
             if (outputFormat === "stream-json") {
                 const parsed = parseStreamJson(stdout);
                 if (parsed.costUsd !== null) {
@@ -3165,9 +2782,6 @@ export function createGatewayServer(deps = {}) {
             performanceMetrics.recordRequest("claude", finalizedDurationMs, wasSuccessful);
         }
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Codex Tool
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("codex_request", {
         prompt: z
             .string()
@@ -3232,14 +2846,10 @@ export function createGatewayServer(deps = {}) {
             .boolean()
             .default(false)
             .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
-        // U23: emit `--json` so the codex-json-parser surfaces input/output/cache
-        // tokens (and any cost) through extractUsageAndCost. Without "json", the
-        // parser is unreachable and Codex usage is never reported.
         outputFormat: z
             .enum(["text", "json"])
             .default("text")
             .describe("Codex output format. `json` emits --json (JSONL events) so token usage and cost are parsed and reported in the flight recorder. `text` is the default."),
-        // U26: high-impact feature flags. All optional.
         outputSchema: z
             .union([z.string(), z.record(z.string(), z.unknown())])
             .optional()
@@ -3269,7 +2879,6 @@ export function createGatewayServer(deps = {}) {
             .boolean()
             .optional()
             .describe("Codex --ignore-rules: skip project rule files for this run."),
-        // Phase 4 slice ζ — Codex working-dir + add-dir parity (new sessions only).
         workingDir: z
             .string()
             .min(1)
@@ -3327,15 +2936,7 @@ export function createGatewayServer(deps = {}) {
             stablePrefixTokens: prep.stablePrefixTokens ?? undefined,
         }, runtime);
         logger.info(`[${corrId}] codex_request invoked with model=${prep.resolvedModel || "default"}, fullAuto=${fullAuto}, prompt length=${prep.effectivePrompt.length}`);
-        // U26 fix: pass the outputSchema cleanup to awaitJobOrDefer, which
-        // guarantees the cleanup runs exactly once — inline for direct
-        // execution, on terminal status for the job-backed path (sync
-        // completion or deferred). The outer finally MUST NOT clean again.
         const prepCleanup = "cleanup" in prep && typeof prep.cleanup === "function" ? prep.cleanup : undefined;
-        // Slice λ: resolve worktree directive into spawn cwd. Codex has no
-        // in-handler session resolution prior to spawn (session lookup is
-        // lazy via `codex exec resume`), so the user-supplied sessionId is
-        // the only reuse key.
         let worktreeResolution = {};
         try {
             worktreeResolution = await resolveWorktreeForRequest(worktree, sessionId, runtime);
@@ -3346,8 +2947,6 @@ export function createGatewayServer(deps = {}) {
         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);
-            // Deferred — job still running, return async reference. Cleanup
-            // ownership belongs to AsyncJobManager via onComplete.
             if (isDeferredResponse(result)) {
                 return buildDeferredToolResponse(result, sessionId);
             }
@@ -3368,7 +2967,6 @@ export function createGatewayServer(deps = {}) {
                 return createErrorResponse("codex", code, stderr, corrId);
             }
             wasSuccessful = true;
-            // Track session usage
             let effectiveSessionId = sessionId;
             if (!createNewSession && !sessionId) {
                 const activeSession = await sessionManager.getActiveSession("codex");
@@ -3430,12 +3028,8 @@ export function createGatewayServer(deps = {}) {
         finally {
             const finalizedDurationMs = Math.max(0, durationMs || Date.now() - startTime);
             performanceMetrics.recordRequest("codex", finalizedDurationMs, wasSuccessful);
-            // Cleanup is owned by awaitJobOrDefer's contract; nothing to do here.
         }
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // U26: codex_fork_session — `codex fork <SESSION_ID|--last> <prompt>`
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("codex_fork_session", {
         prompt: z
             .string()
@@ -3472,8 +3066,6 @@ export function createGatewayServer(deps = {}) {
         const startTime = Date.now();
         let durationMs = 0;
         let wasSuccessful = false;
-        // Enforce mutual exclusion at tool boundary (Zod records the params but
-        // the SDK's `.tool(...)` does not accept top-level refines).
         if (sessionId && forkLast) {
             return createErrorResponse("codex_fork_session", 1, "", corrId, new Error("sessionId and forkLast are mutually exclusive"));
         }
@@ -3489,11 +3081,6 @@ export function createGatewayServer(deps = {}) {
         }
         const cliInfo = getCliInfo();
         const resolvedModel = resolveModelAlias("codex", model, cliInfo);
-        // Compose argv: forkArgs already starts with `fork`. Inject model and
-        // sandbox/approval flags BEFORE the positional <sessionId|--last> +
-        // prompt to keep them as flags rather than positionals. forkArgs layout
-        // is either ["fork", "--last", prompt] or ["fork", sessionId, prompt];
-        // we splice flags right after "fork".
         const flagSegment = [];
         if (resolvedModel)
             flagSegment.push("--model", resolvedModel);
@@ -3530,9 +3117,6 @@ export function createGatewayServer(deps = {}) {
             performanceMetrics.recordRequest("codex", finalizedDurationMs, wasSuccessful);
         }
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Gemini Tool
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("gemini_request", {
         prompt: z
             .string()
@@ -3583,11 +3167,6 @@ export function createGatewayServer(deps = {}) {
             .boolean()
             .default(false)
             .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
-        // U23: emit `-o json` to extract token usage via parseGeminiJson. Default
-        // remains text so existing callers see no behavior change. Phase 4 slice
-        // ε adds `stream-json` (NDJSON event stream parsed by
-        // parseGeminiStreamJson — `init`/`message`/`result` lines, idle-timeout
-        // semantics covered by Gemini's existing real-time stdout streaming).
         outputFormat: z
             .enum(["text", "json", "stream-json"])
             .default("text")
@@ -3600,8 +3179,12 @@ export function createGatewayServer(deps = {}) {
             .boolean()
             .default(false)
             .describe("Emit `--skip-trust` so Gemini trusts the workspace for this session and skips the interactive trust prompt (Phase 4 slice γ). Required for headless runs in fresh workspaces."),
+        yolo: z
+            .boolean()
+            .optional()
+            .describe("Emit `--yolo` to auto-approve all actions. Equivalent to approvalMode 'yolo'; routed through the same approval gate. Under mcp_managed the gate still decides."),
         worktree: WORKTREE_SCHEMA.optional(),
-    }, async ({ prompt, promptParts, model, sessionId, resumeLatest, createNewSession, approvalMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, includeDirs, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, outputFormat, sandbox, policyFiles, adminPolicyFiles, attachments, skipTrust, worktree, }) => {
+    }, async ({ prompt, promptParts, model, sessionId, resumeLatest, createNewSession, approvalMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, includeDirs, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, outputFormat, sandbox, policyFiles, adminPolicyFiles, attachments, skipTrust, yolo, worktree, }) => {
         return handleGeminiRequest({ sessionManager, logger, runtime }, {
             prompt,
             promptParts,
@@ -3626,12 +3209,10 @@ export function createGatewayServer(deps = {}) {
             adminPolicyFiles,
             attachments,
             skipTrust,
+            yolo,
             worktree,
         });
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Grok Tool
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("grok_request", {
         prompt: z
             .string()
@@ -3702,13 +3283,11 @@ export function createGatewayServer(deps = {}) {
             .default(false)
             .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
         maxTurns: MAX_TURNS_SCHEMA.optional().describe("Grok `--max-turns N`: cap on agent-loop iterations for cost / latency control (Phase 4 slice δ). Bounded to safe integers ≤ 10000."),
-        // Phase 4 slice ζ — Grok working-directory parity.
         workingDir: z
             .string()
             .min(1)
             .optional()
             .describe("Grok --cwd <DIR>: working directory for this invocation. Lets headless callers run Grok against a directory other than the gateway process's cwd."),
-        // Phase 4 slice θ — Grok HIGH parity (sandbox, rules, system-prompt-override, allow, deny).
         sandbox: z
             .string()
             .min(1)
@@ -3732,8 +3311,16 @@ export function createGatewayServer(deps = {}) {
             .array(z.string())
             .optional()
             .describe('Grok --deny <RULE>: permission deny rules. Each entry is emitted as its own --deny instance (per `grok --help`: "Repeat to add multiple rules").'),
+        compactionMode: z
+            .enum(["summary", "transcript", "segments"])
+            .optional()
+            .describe("Grok --compaction-mode: summary (default; no pointer) | transcript (points at the raw transcript) | segments (persists per-segment markdown to grep). Sets GROK_COMPACTION_MODE."),
+        compactionDetail: z
+            .enum(["none", "minimal", "balanced", "verbose"])
+            .optional()
+            .describe("Grok --compaction-detail: verbatim segment detail (none|minimal|balanced|verbose, default verbose). Only affects `--compaction-mode segments`. Sets GROK_COMPACTION_DETAIL."),
         worktree: WORKTREE_SCHEMA.optional(),
-    }, 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, worktree, }) => {
+    }, 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, worktree, }) => {
         return handleGrokRequest({ sessionManager, logger, runtime }, {
             prompt,
             promptParts,
@@ -3763,12 +3350,11 @@ export function createGatewayServer(deps = {}) {
             systemPromptOverride,
             allow,
             deny,
+            compactionMode,
+            compactionDetail,
             worktree,
         });
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Mistral Vibe Tool
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("mistral_request", {
         prompt: z
             .string()
@@ -3798,11 +3384,6 @@ export function createGatewayServer(deps = {}) {
             .enum(MISTRAL_AGENT_MODES)
             .optional()
             .describe("Vibe agent mode (default|plan|accept-edits|auto-approve|chat|explore|lean). Defaults to auto-approve for programmatic use."),
-        effort: z
-            .enum(["low", "medium", "high", "xhigh", "max"])
-            .optional()
-            .describe("Vibe effort level"),
-        reasoningEffort: z.string().optional().describe("Reasoning effort for reasoning models"),
         approvalStrategy: z
             .enum(["legacy", "mcp_managed"])
             .default("legacy")
@@ -3844,7 +3425,6 @@ export function createGatewayServer(deps = {}) {
         maxTurns: MAX_TURNS_SCHEMA.optional().describe("Vibe `--max-turns N`: cap the agent-loop iteration count (programmatic mode only, Phase 4 slice δ). Bounded to safe integers ≤ 10000."),
         maxPrice: MAX_PRICE_SCHEMA.optional().describe("Vibe `--max-price DOLLARS`: interrupt the session when cumulative cost crosses this cap (programmatic mode only, Phase 4 slice δ). Bounded to finite values ≤ 10000 USD."),
         maxTokens: MAX_TOKENS_SCHEMA.optional().describe("Vibe `--max-tokens N`: cap cumulative prompt + completion tokens for the session (programmatic mode only). Bounded to safe integers ≤ 100000000."),
-        // Phase 4 slice ζ — Vibe working-directory + additional-dirs parity.
         workingDir: z
             .string()
             .min(1)
@@ -3855,7 +3435,7 @@ export function createGatewayServer(deps = {}) {
             .optional()
             .describe("Vibe --add-dir <DIR>: additional writable workspace directories. Each entry is emitted as its own --add-dir instance (Vibe states this flag may be specified multiple times)."),
         worktree: WORKTREE_SCHEMA.optional(),
-    }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, permissionMode, effort, reasoningEffort, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, optimizeResponse, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, worktree, }) => {
+    }, 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, worktree, }) => {
         return handleMistralRequest({ sessionManager, logger, runtime }, {
             prompt,
             promptParts,
@@ -3865,8 +3445,6 @@ export function createGatewayServer(deps = {}) {
             resumeLatest,
             createNewSession,
             permissionMode,
-            effort,
-            reasoningEffort,
             approvalStrategy,
             approvalPolicy,
             mcpServers,
@@ -3886,16 +3464,6 @@ export function createGatewayServer(deps = {}) {
             worktree,
         });
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Async Long-Running Job Tools (No Time-Bound LLM Execution)
-    //
-    // STRUCTURAL INVARIANT: these tools are only registered when a real job
-    // store is attached (`persistence.asyncJobsEnabled === true`). When the
-    // operator has configured `[persistence].backend = "none"`, none of the
-    // *_request_async / llm_job_* tools exist in the MCP tool list at all —
-    // orchestrating agents get a clean "tool not found" signal at connect
-    // time instead of silent in-memory loss after the 1-hour TTL.
-    //──────────────────────────────────────────────────────────────────────────────
     if (asyncJobsEnabled) {
         server.tool("claude_request_async", {
             prompt: z
@@ -3929,7 +3497,6 @@ export function createGatewayServer(deps = {}) {
                 .enum(CLAUDE_PERMISSION_MODES)
                 .optional()
                 .describe("Claude --permission-mode: default|acceptEdits|plan|auto|dontAsk|bypassPermissions. `default` is a no-op."),
-            // U25 — Claude high-impact features
             agent: z
                 .string()
                 .optional()
@@ -3969,7 +3536,6 @@ export function createGatewayServer(deps = {}) {
                 .boolean()
                 .optional()
                 .describe("Claude --exclude-dynamic-system-prompt-sections: trim dynamic context blocks from the system prompt."),
-            // Phase 4 slice η — Claude reliability + structured-output parity
             fallbackModel: z
                 .string()
                 .min(1)
@@ -3979,11 +3545,28 @@ export function createGatewayServer(deps = {}) {
                 .union([z.string(), z.record(z.string(), z.unknown())])
                 .optional()
                 .describe("Claude --json-schema: JSON Schema literal (NOT a path) constraining structured output. Object values are JSON.stringify-d; string values are passed verbatim. Use with outputFormat='json'."),
-            // Phase 4 slice ζ — Claude additional-workspace-dirs parity
             addDir: z
                 .array(z.string())
                 .optional()
                 .describe("Claude --add-dir: additional directories the CLI is allowed to read/write beyond the process cwd. Each entry is emitted as its own --add-dir instance."),
+            noSessionPersistence: z
+                .boolean()
+                .optional()
+                .describe("Claude --no-session-persistence: do not write this session to disk (ephemeral one-shot runs; mirrors codex --ephemeral)."),
+            settingSources: z
+                .string()
+                .min(1)
+                .optional()
+                .describe("Claude --setting-sources: comma-separated setting sources to load (user|project|local) for reproducible/isolated headless runs."),
+            settings: z
+                .string()
+                .min(1)
+                .optional()
+                .describe("Claude --settings: path to a settings JSON file or a JSON literal of additional settings. Powerful: settings can define hooks/permissions/model; passed verbatim."),
+            tools: z
+                .array(z.string())
+                .optional()
+                .describe('Claude --tools: restrict the available built-in tool set (distinct from allowedTools permission gating). Pass [""] to disable all tools.'),
             worktree: WORKTREE_SCHEMA.optional(),
             approvalStrategy: z
                 .enum(["legacy", "mcp_managed"])
@@ -4014,7 +3597,7 @@ export function createGatewayServer(deps = {}) {
                 .boolean()
                 .default(false)
                 .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
-        }, async ({ prompt, promptParts, model, outputFormat, sessionId, continueSession, createNewSession, allowedTools, disallowedTools, dangerouslySkipPermissions, permissionMode, agent, agents, forkSession, systemPrompt, appendSystemPrompt, maxBudgetUsd, maxTurns, effort, excludeDynamicSystemPromptSections, fallbackModel, jsonSchema, addDir, 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, worktree, approvalStrategy, approvalPolicy, mcpServers, strictMcpConfig, correlationId, optimizePrompt, 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)."));
             }
@@ -4046,12 +3629,15 @@ export function createGatewayServer(deps = {}) {
                 fallbackModel,
                 jsonSchema,
                 addDir,
+                noSessionPersistence,
+                settingSources,
+                settings,
+                tools,
             }, runtime);
             if (!("args" in prep))
                 return prep;
             const { corrId, args, requestedMcpServers, mcpConfig, approvalDecision } = prep;
             try {
-                // Session management (before job start for async)
                 let effectiveSessionId = sessionId;
                 let useContinue = continueSession;
                 const activeSession = await sessionManager.getActiveSession("claude");
@@ -4075,14 +3661,11 @@ export function createGatewayServer(deps = {}) {
                         await sessionManager.createSession("claude", "Claude Session", effectiveSessionId);
                     }
                 }
-                // Slice 3: TTL warning on resume (async path too).
                 const ttlWarning = maybeBuildCacheTtlWarning({
                     runtime,
                     sessionId: effectiveSessionId,
                     cli: "claude",
                 });
-                // Slice λ: resolve worktree directive after session metadata is
-                // settled so resume reuse can read metadata.worktreePath.
                 let worktreeResolution = {};
                 try {
                     worktreeResolution = await resolveWorktreeForRequest(worktree, effectiveSessionId, runtime);
@@ -4090,7 +3673,6 @@ export function createGatewayServer(deps = {}) {
                 catch (err) {
                     return createErrorResponse("claude_request_async", 1, "", corrId, err);
                 }
-                // Idle timeout only for stream-json (text/json produce no output until done)
                 const effectiveIdleTimeout = outputFormat === "stream-json"
                     ? resolveIdleTimeout("claude", idleTimeoutMs)
                     : undefined;
@@ -4116,8 +3698,6 @@ export function createGatewayServer(deps = {}) {
                 if (worktreeResolution.worktreePath) {
                     asyncResponse.worktreePath = worktreeResolution.worktreePath;
                 }
-                // Rec #4: include any prep-time warnings (e.g.
-                // cacheable_prefix_uncached) alongside ttlWarning.
                 const mergedWarnings = [
                     ...(ttlWarning ? [ttlWarning] : []),
                     ...(prep.warnings ?? []),
@@ -4201,12 +3781,10 @@ export function createGatewayServer(deps = {}) {
                 .boolean()
                 .default(false)
                 .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
-            // U23: emit `--json` to enable JSONL event-stream parsing for token usage.
             outputFormat: z
                 .enum(["text", "json"])
                 .default("text")
                 .describe("Codex output format. `json` emits --json (JSONL events) for token usage extraction."),
-            // U26: high-impact feature flags. All optional.
             outputSchema: z
                 .union([z.string(), z.record(z.string(), z.unknown())])
                 .optional()
@@ -4221,7 +3799,6 @@ export function createGatewayServer(deps = {}) {
             images: z.array(z.string()).optional().describe("Codex -i <path>: image attachments."),
             ignoreUserConfig: z.boolean().optional().describe("Codex --ignore-user-config."),
             ignoreRules: z.boolean().optional().describe("Codex --ignore-rules."),
-            // Phase 4 slice ζ — Codex working-dir + add-dir parity (new sessions only).
             workingDir: z
                 .string()
                 .min(1)
@@ -4318,11 +3895,6 @@ export function createGatewayServer(deps = {}) {
                 .boolean()
                 .default(false)
                 .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
-            // U23: emit `-o json` to extract token usage via parseGeminiJson. Default
-            // remains text so existing callers see no behavior change. Phase 4 slice
-            // ε adds `stream-json` (NDJSON event stream parsed by
-            // parseGeminiStreamJson — `init`/`message`/`result` lines, idle-timeout
-            // semantics covered by Gemini's existing real-time stdout streaming).
             outputFormat: z
                 .enum(["text", "json", "stream-json"])
                 .default("text")
@@ -4335,8 +3907,12 @@ export function createGatewayServer(deps = {}) {
                 .boolean()
                 .default(false)
                 .describe("Emit `--skip-trust` so Gemini trusts the workspace for this session and skips the interactive trust prompt (Phase 4 slice γ). Required for headless runs in fresh workspaces."),
+            yolo: z
+                .boolean()
+                .optional()
+                .describe("Emit `--yolo` to auto-approve all actions. Equivalent to approvalMode 'yolo'; routed through the same approval gate. Under mcp_managed the gate still decides."),
             worktree: WORKTREE_SCHEMA.optional(),
-        }, async ({ prompt, promptParts, model, sessionId, resumeLatest, createNewSession, approvalMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, includeDirs, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, outputFormat, sandbox, policyFiles, adminPolicyFiles, attachments, skipTrust, worktree, }) => {
+        }, async ({ prompt, promptParts, model, sessionId, resumeLatest, createNewSession, approvalMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, includeDirs, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, outputFormat, sandbox, policyFiles, adminPolicyFiles, attachments, skipTrust, yolo, worktree, }) => {
             return handleGeminiRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -4360,6 +3936,7 @@ export function createGatewayServer(deps = {}) {
                 adminPolicyFiles,
                 attachments,
                 skipTrust,
+                yolo,
                 worktree,
             });
         });
@@ -4432,13 +4009,11 @@ export function createGatewayServer(deps = {}) {
                 .default(false)
                 .describe("Bypass dedup and force a fresh CLI run even if a recent identical request exists"),
             maxTurns: MAX_TURNS_SCHEMA.optional().describe("Grok `--max-turns N`: cap on agent-loop iterations for cost / latency control (Phase 4 slice δ). Bounded to safe integers ≤ 10000."),
-            // Phase 4 slice ζ — Grok working-directory parity.
             workingDir: z
                 .string()
                 .min(1)
                 .optional()
                 .describe("Grok --cwd <DIR>: working directory for this invocation. Lets headless callers run Grok against a directory other than the gateway process's cwd."),
-            // Phase 4 slice θ — Grok HIGH parity (sandbox, rules, system-prompt-override, allow, deny).
             sandbox: z
                 .string()
                 .min(1)
@@ -4462,8 +4037,16 @@ export function createGatewayServer(deps = {}) {
                 .array(z.string())
                 .optional()
                 .describe("Grok --deny <RULE>: permission deny rules. Each entry → its own --deny instance."),
+            compactionMode: z
+                .enum(["summary", "transcript", "segments"])
+                .optional()
+                .describe("Grok --compaction-mode: summary (default) | transcript | segments. Sets GROK_COMPACTION_MODE."),
+            compactionDetail: z
+                .enum(["none", "minimal", "balanced", "verbose"])
+                .optional()
+                .describe("Grok --compaction-detail: segment verbatim detail (none|minimal|balanced|verbose, default verbose). Only affects segments mode. Sets GROK_COMPACTION_DETAIL."),
             worktree: WORKTREE_SCHEMA.optional(),
-        }, 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, worktree, }) => {
+        }, 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, worktree, }) => {
             return handleGrokRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -4492,6 +4075,8 @@ export function createGatewayServer(deps = {}) {
                 systemPromptOverride,
                 allow,
                 deny,
+                compactionMode,
+                compactionDetail,
                 worktree,
             });
         });
@@ -4524,11 +4109,6 @@ export function createGatewayServer(deps = {}) {
                 .enum(MISTRAL_AGENT_MODES)
                 .optional()
                 .describe("Vibe agent mode (default|plan|accept-edits|auto-approve|chat|explore|lean). Defaults to auto-approve for programmatic use."),
-            effort: z
-                .enum(["low", "medium", "high", "xhigh", "max"])
-                .optional()
-                .describe("Vibe effort level"),
-            reasoningEffort: z.string().optional().describe("Reasoning effort for reasoning models"),
             approvalStrategy: z
                 .enum(["legacy", "mcp_managed"])
                 .default("legacy")
@@ -4569,7 +4149,6 @@ export function createGatewayServer(deps = {}) {
             maxTurns: MAX_TURNS_SCHEMA.optional().describe("Vibe `--max-turns N`: cap the agent-loop iteration count (programmatic mode only, Phase 4 slice δ). Bounded to safe integers ≤ 10000."),
             maxPrice: MAX_PRICE_SCHEMA.optional().describe("Vibe `--max-price DOLLARS`: interrupt the session when cumulative cost crosses this cap (programmatic mode only, Phase 4 slice δ). Bounded to finite values ≤ 10000 USD."),
             maxTokens: MAX_TOKENS_SCHEMA.optional().describe("Vibe `--max-tokens N`: cap cumulative prompt + completion tokens for the session (programmatic mode only). Bounded to safe integers ≤ 100000000."),
-            // Phase 4 slice ζ — Vibe working-directory + additional-dirs parity.
             workingDir: z
                 .string()
                 .min(1)
@@ -4580,7 +4159,7 @@ export function createGatewayServer(deps = {}) {
                 .optional()
                 .describe("Vibe --add-dir <DIR>: additional writable workspace directories. Each entry is emitted as its own --add-dir instance."),
             worktree: WORKTREE_SCHEMA.optional(),
-        }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, permissionMode, effort, reasoningEffort, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, worktree, }) => {
+        }, async ({ prompt, promptParts, model, outputFormat, sessionId, resumeLatest, createNewSession, permissionMode, approvalStrategy, approvalPolicy, mcpServers, allowedTools, disallowedTools, correlationId, optimizePrompt, idleTimeoutMs, forceRefresh, trust, maxTurns, maxPrice, maxTokens, workingDir, addDir, worktree, }) => {
             return handleMistralRequestAsync({ sessionManager, asyncJobManager, logger, runtime }, {
                 prompt,
                 promptParts,
@@ -4590,8 +4169,6 @@ export function createGatewayServer(deps = {}) {
                 resumeLatest,
                 createNewSession,
                 permissionMode,
-                effort,
-                reasoningEffort,
                 approvalStrategy,
                 approvalPolicy,
                 mcpServers,
@@ -4667,7 +4244,6 @@ export function createGatewayServer(deps = {}) {
                     isError: true,
                 };
             }
-            // Parse stream-json output for Claude async jobs
             const outputFormat = asyncJobManager.getJobOutputFormat(jobId);
             let parsed;
             if (outputFormat === "stream-json" && result.stdout) {
@@ -4727,14 +4303,7 @@ export function createGatewayServer(deps = {}) {
                 ],
             };
         });
-    } // end if (asyncJobsEnabled)
-    // Read back any persisted request (sync OR async) by its correlation id.
-    // Registered unconditionally — it reads the flight recorder, which is
-    // independent of async-job persistence. Every sync/async response echoes
-    // its id in `structuredContent.correlationId`; pass that id here to recover
-    // the persisted prompt/response after the inline result is gone. With flight
-    // recording disabled (LLM_GATEWAY_LOGS_DB=none → NoopFlightRecorder) the
-    // query yields no rows and this returns the "not found" shape.
+    }
     server.tool("llm_request_result", {
         correlationId: z
             .string()
@@ -4805,9 +4374,6 @@ export function createGatewayServer(deps = {}) {
             ],
         };
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Approval Audit Tools
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("approval_list", {
         limit: z
             .number()
@@ -4835,9 +4401,6 @@ export function createGatewayServer(deps = {}) {
             ],
         };
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // List Models Tool
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("list_models", {
         cli: z
             .preprocess(value => (value === "" || value === null ? undefined : value), z.enum(["claude", "codex", "gemini", "grok", "mistral"]).optional())
@@ -4916,9 +4479,6 @@ export function createGatewayServer(deps = {}) {
             };
         }
     });
-    //──────────────────────────────────────────────────────────────────────────────
-    // Session Management Tools
-    //──────────────────────────────────────────────────────────────────────────────
     server.tool("session_create", {
         cli: SESSION_PROVIDER_ENUM.describe("CLI type (claude|codex|gemini|grok|mistral)"),
         description: z.string().optional().describe("Session description"),
@@ -5094,15 +4654,6 @@ export function createGatewayServer(deps = {}) {
                 };
             }
             const activeSession = await sessionManager.getActiveSession(session.cli);
-            // Slice 2: project a compact cacheState view from the flight
-            // recorder at read time. NOT persisted on the Session interface
-            // (sessions.json stays content-free per the project invariant).
-            // The field is OMITTED entirely (not null, not empty object) when
-            // the session has zero rows in the flight recorder so the response
-            // stays compact for fresh sessions.
-            //
-            // Slice 3: include ttlRemainingMs derived from the gateway's
-            // configured TTL policy. Null for non-claude sessions.
             let cacheState;
             try {
                 const stats = computeSessionCacheStats(flightRecorder, session.id);
@@ -5171,16 +4722,8 @@ export function createGatewayServer(deps = {}) {
     });
     return server;
 }
-//──────────────────────────────────────────────────────────────────────────────
-// Async Initialization
-//──────────────────────────────────────────────────────────────────────────────
 async function initializeSessionManager() {
     const config = loadConfig();
-    // Slice λ: file-backed sessions get a cleanup hook that tears down any
-    // git worktrees recorded on session.metadata.worktreePath. PG-backed
-    // sessions skip the hook (multi-tenant deployments don't necessarily
-    // own a single filesystem); revisit if/when worktree support extends
-    // there.
     const worktreeCleanupHook = createWorktreeSessionCleanupHook(logger);
     if (config.database) {
         logger.info("Initializing PostgreSQL session manager");
@@ -5198,9 +4741,6 @@ async function initializeSessionManager() {
     }
     resourceProvider = new ResourceProvider(sessionManager, performanceMetrics, getFlightRecorder(logger), getCacheAwarenessConfig(logger));
 }
-//──────────────────────────────────────────────────────────────────────────────
-// Health Check Resource (only if using PostgreSQL)
-//──────────────────────────────────────────────────────────────────────────────
 function registerHealthResource(server) {
     if (db) {
         server.registerResource("health", "health://status", {
@@ -5221,7 +4761,6 @@ function registerHealthResource(server) {
         });
         logger.info("Health check resource registered");
     }
-    // Process health resource (always available, not dependent on DB)
     server.registerResource("process-health", "metrics://process-health", {
         title: "Process Health",
         description: "Async job health (CPU, memory, zombie detection)",
@@ -5240,13 +4779,9 @@ function registerHealthResource(server) {
     });
     logger.info("Process health resource registered");
 }
-//──────────────────────────────────────────────────────────────────────────────
-// Graceful Shutdown
-//──────────────────────────────────────────────────────────────────────────────
 async function shutdown(signal) {
     logger.info(`Received ${signal}, shutting down gracefully...`);
     try {
-        // Kill all active process groups (SIGTERM → wait 3s → SIGKILL)
         await killAllProcessGroups();
         logger.info("All process groups terminated");
         if (activeHttpGateway) {
@@ -5276,9 +4811,6 @@ async function shutdown(signal) {
 }
 process.on("SIGTERM", () => shutdown("SIGTERM"));
 process.on("SIGINT", () => shutdown("SIGINT"));
-//──────────────────────────────────────────────────────────────────────────────
-// Server Startup
-//──────────────────────────────────────────────────────────────────────────────
 async function main() {
     startWindowsBootstrapperSelfHeal();
     const args = process.argv.slice(2);
@@ -5342,7 +4874,6 @@ async function main() {
         process.env.MCP_TRANSPORT ||
         "stdio";
     logger.info(`Starting llm-cli-gateway MCP server with ${transportMode} transport`);
-    // Initialize session manager first
     await initializeSessionManager();
     const serverDeps = {
         sessionManager,
@@ -5369,14 +4900,11 @@ async function main() {
     activeServer = createGatewayServer({
         ...serverDeps,
     });
-    // Register health check resource if using PostgreSQL
     registerHealthResource(activeServer);
     const transport = new StdioServerTransport();
     await activeServer.connect(transport);
     logger.info("llm-cli-gateway MCP server connected and ready");
 }
-// Guard: only auto-start when run directly (not imported for testing)
-// Resolve symlinks so `llm-cli-gateway` (npm-linked bin) matches import.meta.url
 const __entryUrl = entrypointFileURL(process.argv[1]);
 if (__entryUrl === import.meta.url) {
     main().catch(error => {