llm-cli-gateway 1.17.3 → 1.17.5

package/dist/model-registry.js

@@ -14,9 +14,6 @@ const FALLBACK_INFO = {
         modelOrder: ["opus", "sonnet", "haiku"],
     },
     codex: {
-        // U26: gpt-5.5 is the bundled fallback default. Config/env overrides still
-        // win (applyCodexOverrides runs after). Older aliases are retained in the
-        // models map so callers that still pass `gpt-5.3-codex` resolve cleanly.
         description: "OpenAI's Codex CLI - best for code execution in sandboxed environments",
         models: {
             "gpt-5.5": "Latest Codex frontier model. Best for: most Codex tasks (default since U26)",
@@ -36,8 +33,6 @@ const FALLBACK_INFO = {
         },
     },
     grok: {
-        // No hardcoded `defaultModel`. Let Grok CLI pick its own built-in default
-        // unless an explicit value is found via env vars in applyGrokOverrides.
         description: "xAI's Grok Build CLI - best for agentic coding tasks via xAI's Grok models",
         models: {
             "grok-build": "Default Grok model for code/agentic tasks. Best for: most Grok build sessions",
@@ -45,10 +40,6 @@ const FALLBACK_INFO = {
         modelOrder: ["grok-build"],
     },
     mistral: {
-        // Mistral Vibe selects the active model via VIBE_ACTIVE_MODEL; there is no
-        // `--model` flag. Do not set a bundled default here: Vibe's own default and
-        // user config move independently of this gateway. The model list is only a
-        // low-confidence recovery set for stale config/model-not-found failures.
         description: "Mistral AI's Vibe CLI - agentic coding via Mistral models (model selection via VIBE_ACTIVE_MODEL env var)",
         models: {
             "mistral-medium-3.5": "Vibe coding model alias observed in Vibe 2.x defaults. Used only when discovery/config requires an explicit VIBE_ACTIVE_MODEL.",
@@ -99,9 +90,6 @@ export function resolveModelAlias(cli, model, info) {
     const normalized = trimmed.toLowerCase();
     const cliInfo = info[cli];
     if (normalized === "default" || normalized === "latest") {
-        // If no default is configured, return undefined so the CLI picks its own
-        // built-in default. Avoids passing the literal string "default"/"latest"
-        // as a model name to the CLI.
         return cliInfo.defaultModel;
     }
     const alias = resolveConfiguredAlias(cliInfo, normalized);
@@ -393,7 +381,6 @@ function applyMistralOverrides(info) {
     addEnvModels(info, "MISTRAL_MODELS");
     addEnvAliases(info, "mistral", "MISTRAL_MODEL_ALIASES");
     addGlobalEnvAliases(info, "mistral");
-    // Vibe uses VIBE_ACTIVE_MODEL instead of a CLI flag. Explicit env values win.
     const envDefault = process.env.MISTRAL_DEFAULT_MODEL || process.env.VIBE_ACTIVE_MODEL;
     if (envDefault) {
         const source = process.env.MISTRAL_DEFAULT_MODEL

package/dist/pricing.d.ts

@@ -1,54 +1,8 @@
-/**
- * Per-model pricing for cache-savings estimation.
- *
- * `priced_as_of` is the date these numbers were last refreshed. The
- * gateway's doctor surfaces this so operators can see when the table is
- * stale — pricing is an ESTIMATE, not a billing number.
- *
- * Pricing units: USD per 1M tokens.
- *
- * Anthropic source: <https://platform.claude.com/docs/en/about-claude/pricing>
- *   - Sonnet 4.x / Sonnet 3.5: $3 input / $15 output.
- *   - Opus 4.5+ / Mythos Preview: $15 input / $75 output.
- *   - Opus 4 / 4.1 (deprecated): same as 4.5+.
- *   - Haiku 4.5: $1 input / $5 output.
- *   - Haiku 3.5 (Vertex-only): $0.80 input / $4 output.
- *
- * Cache pricing multipliers (Anthropic):
- *   - cache write 5-min TTL: 1.25× base input.
- *   - cache write 1-hour TTL: 2× base input.
- *   - cache read: 0.10× base input (90% savings).
- *
- * Codex / OpenAI: GPT-5.4 input ~$1.25 / output $10 per 1M (approx; OpenAI
- * does not publish a stable per-CLI table). Cached input ~50% of base.
- *
- * Gemini, Grok, Mistral: pricing varies by model and is not surfaced in
- * gateway today. Returns 0 for unknown.
- */
 export interface PricePerMillion {
     inputUsd: number;
     outputUsd: number;
-    /** Multiplier on inputUsd for a cache HIT (read). Anthropic: 0.10. */
     cacheReadMultiplier: number;
 }
 export declare const PRICING_AS_OF = "2026-05-26";
-/**
- * Look up pricing by (cli, model) name. Best-effort; unknown models return
- * ZEROED pricing so estimated_savings_usd in aggregates falls back to 0
- * rather than throwing OR over-reporting savings on an unpriced model.
- *
- * Recognised model families:
- *   - claude: model name contains "sonnet" | "opus" | "haiku".
- *   - codex: model name contains "gpt-5" or "o3" (current OpenAI families).
- *
- * Anything outside these explicit matches returns ZERO. This is a
- * deliberate conservative choice — we'd rather under-report savings on
- * an unrecognised model than over-report on one whose actual pricing we
- * don't know. Update this table when a new model family ships.
- */
 export declare function getPricing(cli: "claude" | "codex" | "gemini" | "grok" | "mistral", model: string): PricePerMillion;
-/**
- * Estimate USD saved by `cacheReadTokens` being served from cache instead
- * of fresh input. Returns 0 for zero cache reads or unknown pricing.
- */
 export declare function estimateCacheSavingsUsd(cli: "claude" | "codex" | "gemini" | "grok" | "mistral", model: string, cacheReadTokens: number): number;

package/dist/pricing.js

@@ -1,30 +1,3 @@
-/**
- * Per-model pricing for cache-savings estimation.
- *
- * `priced_as_of` is the date these numbers were last refreshed. The
- * gateway's doctor surfaces this so operators can see when the table is
- * stale — pricing is an ESTIMATE, not a billing number.
- *
- * Pricing units: USD per 1M tokens.
- *
- * Anthropic source: <https://platform.claude.com/docs/en/about-claude/pricing>
- *   - Sonnet 4.x / Sonnet 3.5: $3 input / $15 output.
- *   - Opus 4.5+ / Mythos Preview: $15 input / $75 output.
- *   - Opus 4 / 4.1 (deprecated): same as 4.5+.
- *   - Haiku 4.5: $1 input / $5 output.
- *   - Haiku 3.5 (Vertex-only): $0.80 input / $4 output.
- *
- * Cache pricing multipliers (Anthropic):
- *   - cache write 5-min TTL: 1.25× base input.
- *   - cache write 1-hour TTL: 2× base input.
- *   - cache read: 0.10× base input (90% savings).
- *
- * Codex / OpenAI: GPT-5.4 input ~$1.25 / output $10 per 1M (approx; OpenAI
- * does not publish a stable per-CLI table). Cached input ~50% of base.
- *
- * Gemini, Grok, Mistral: pricing varies by model and is not surfaced in
- * gateway today. Returns 0 for unknown.
- */
 export const PRICING_AS_OF = "2026-05-26";
 const ANTHROPIC_SONNET = {
     inputUsd: 3,
@@ -44,7 +17,6 @@ const ANTHROPIC_HAIKU = {
 const OPENAI_GPT5 = {
     inputUsd: 1.25,
     outputUsd: 10,
-    // OpenAI prompt-caching: cached input tokens billed at 50% of base.
     cacheReadMultiplier: 0.5,
 };
 const ZERO = {
@@ -52,20 +24,6 @@ const ZERO = {
     outputUsd: 0,
     cacheReadMultiplier: 0,
 };
-/**
- * Look up pricing by (cli, model) name. Best-effort; unknown models return
- * ZEROED pricing so estimated_savings_usd in aggregates falls back to 0
- * rather than throwing OR over-reporting savings on an unpriced model.
- *
- * Recognised model families:
- *   - claude: model name contains "sonnet" | "opus" | "haiku".
- *   - codex: model name contains "gpt-5" or "o3" (current OpenAI families).
- *
- * Anything outside these explicit matches returns ZERO. This is a
- * deliberate conservative choice — we'd rather under-report savings on
- * an unrecognised model than over-report on one whose actual pricing we
- * don't know. Update this table when a new model family ships.
- */
 export function getPricing(cli, model) {
     const lower = model.toLowerCase();
     if (cli === "claude") {
@@ -84,17 +42,12 @@ export function getPricing(cli, model) {
     }
     return ZERO;
 }
-/**
- * Estimate USD saved by `cacheReadTokens` being served from cache instead
- * of fresh input. Returns 0 for zero cache reads or unknown pricing.
- */
 export function estimateCacheSavingsUsd(cli, model, cacheReadTokens) {
     if (cacheReadTokens <= 0)
         return 0;
     const p = getPricing(cli, model);
     if (p.inputUsd === 0)
         return 0;
-    // Savings = (fresh-input-cost) - (cache-read-cost) = inputUsd × (1 - mult)
     const savedPerToken = (p.inputUsd * (1 - p.cacheReadMultiplier)) / 1_000_000;
     return cacheReadTokens * savedPerToken;
 }

package/dist/process-monitor.d.ts

@@ -1,7 +1,3 @@
-/**
- * On-demand process health monitoring via /proc (Linux).
- * Gracefully degrades on non-Linux platforms.
- */
 import type { Logger } from "./logger.js";
 export interface ProcessHealth {
     pid: number;
@@ -20,25 +16,15 @@ export interface JobHealth {
     isZombie: boolean;
     runningForMs: number;
 }
-/**
- * Parse /proc/[pid]/stat safely.
- * The `comm` field (field 2) is in parentheses and may contain spaces,
- * so we find the LAST ')' and parse remaining fields from there.
- */
 export declare function parseProcStat(content: string): {
     state: string;
     utime: number;
     stime: number;
 } | null;
-/**
- * Parse VmRSS from /proc/[pid]/status.
- * Returns RSS in kilobytes (already in kB in /proc/[pid]/status).
- */
 export declare function parseVmRss(content: string): number | null;
 export declare class ProcessMonitor {
     private prevSamples;
     constructor(_logger?: Logger);
-    /** Clear all cached CPU samples */
     reset(): void;
     sampleProcess(pid: number): ProcessHealth;
     checkJobHealth(jobs: {
@@ -48,6 +34,5 @@ export declare class ProcessMonitor {
         pid: number | null;
         startedAt: string;
     }[]): JobHealth[];
-    /** Clean up stale samples for PIDs that no longer exist */
     cleanupSamples(activePids: Set<number>): void;
 }

package/dist/process-monitor.js

@@ -1,20 +1,10 @@
-/**
- * On-demand process health monitoring via /proc (Linux).
- * Gracefully degrades on non-Linux platforms.
- */
 import { readFileSync } from "fs";
-/**
- * Parse /proc/[pid]/stat safely.
- * The `comm` field (field 2) is in parentheses and may contain spaces,
- * so we find the LAST ')' and parse remaining fields from there.
- */
 export function parseProcStat(content) {
     const lastParen = content.lastIndexOf(")");
     if (lastParen === -1)
         return null;
-    const afterComm = content.slice(lastParen + 2); // skip ") "
+    const afterComm = content.slice(lastParen + 2);
     const fields = afterComm.split(" ");
-    // fields[0] = state, fields[11] = utime (14-3), fields[12] = stime (15-3)
     if (fields.length < 13)
         return null;
     const utime = parseInt(fields[11], 10);
@@ -27,22 +17,14 @@ export function parseProcStat(content) {
         stime,
     };
 }
-/**
- * Parse VmRSS from /proc/[pid]/status.
- * Returns RSS in kilobytes (already in kB in /proc/[pid]/status).
- */
 export function parseVmRss(content) {
     const match = content.match(/^VmRSS:\s+(\d+)\s+kB$/m);
     return match ? parseInt(match[1], 10) : null;
 }
-/**
- * Read total system CPU jiffies from /proc/stat.
- * Used to normalize per-process CPU into a percentage.
- */
 function getTotalCpuJiffies() {
     try {
         const content = readFileSync("/proc/stat", "utf-8");
-        const cpuLine = content.split("\n")[0]; // "cpu  user nice system idle ..."
+        const cpuLine = content.split("\n")[0];
         const fields = cpuLine.split(/\s+/).slice(1).map(Number);
         return fields.reduce((a, b) => a + b, 0);
     }
@@ -51,16 +33,13 @@ function getTotalCpuJiffies() {
     }
 }
 export class ProcessMonitor {
-    // Previous samples for CPU delta calculation
     prevSamples = new Map();
     constructor(_logger) { }
-    /** Clear all cached CPU samples */
     reset() {
         this.prevSamples.clear();
     }
     sampleProcess(pid) {
         const now = new Date().toISOString();
-        // 1. Existence check
         let alive = false;
         try {
             process.kill(pid, 0);
@@ -77,12 +56,10 @@ export class ProcessMonitor {
                     sampledAt: now,
                 };
             }
-            // EPERM = process exists but we can't signal it
             if (err.code === "EPERM") {
                 alive = true;
             }
         }
-        // 2. Parse /proc/[pid]/stat for state + CPU ticks
         let state = null;
         let cpuPercent = null;
         try {
@@ -90,7 +67,6 @@ export class ProcessMonitor {
             const parsed = parseProcStat(statContent);
             if (parsed) {
                 state = parsed.state;
-                // CPU delta calculation
                 const totalJiffies = getTotalCpuJiffies();
                 const prev = this.prevSamples.get(pid);
                 if (prev && totalJiffies !== null) {
@@ -100,7 +76,6 @@ export class ProcessMonitor {
                         cpuPercent = (processJiffiesDelta / totalJiffiesDelta) * 100;
                     }
                 }
-                // Store for next delta
                 if (totalJiffies !== null) {
                     this.prevSamples.set(pid, {
                         utime: parsed.utime,
@@ -112,16 +87,13 @@ export class ProcessMonitor {
             }
         }
         catch {
-            // /proc not available (non-Linux) — degrade gracefully
         }
-        // 3. Parse /proc/[pid]/status for VmRSS
         let memoryRssKb = null;
         try {
             const statusContent = readFileSync(`/proc/${pid}/status`, "utf-8");
             memoryRssKb = parseVmRss(statusContent);
         }
         catch {
-            // Non-Linux or process exited between checks
         }
         return { pid, alive, state, cpuPercent, memoryRssKb, sampledAt: now };
     }
@@ -151,7 +123,6 @@ export class ProcessMonitor {
             };
         });
     }
-    /** Clean up stale samples for PIDs that no longer exist */
     cleanupSamples(activePids) {
         for (const pid of this.prevSamples.keys()) {
             if (!activePids.has(pid)) {

package/dist/prompt-parts.d.ts

@@ -9,20 +9,6 @@ export interface PromptParts {
     tools?: string;
     context?: string;
     task: string;
-    /**
-     * Slice κ (Claude only): per-block opt-in to Anthropic `cache_control`
-     * breakpoints. Setting `system: true` (or tools/context) marks that
-     * block with `cache_control: {type:"ephemeral", ttl:"1h"}` in the
-     * stream-json payload the gateway pipes to `claude --input-format
-     * stream-json`. The `task` block is NEVER marked (it's the volatile
-     * tail). Empty parts are silently skipped even if their flag is true.
-     *
-     * Constraint: callers MUST also pass `outputFormat:"stream-json"` —
-     * mixing cacheControl with text/json output returns an error response.
-     * `ttl` is hard-coded to `"1h"` because Claude Code injects its own
-     * 1h-marked system blocks ahead of caller content and Anthropic
-     * rejects a 1h block after a 5m block.
-     */
     cacheControl?: PromptPartsCacheControl;
 }
 export declare const PromptPartsSchema: z.ZodObject<{
@@ -35,32 +21,32 @@ export declare const PromptPartsSchema: z.ZodObject<{
         tools: z.ZodOptional<z.ZodBoolean>;
         context: z.ZodOptional<z.ZodBoolean>;
     }, "strict", z.ZodTypeAny, {
-        system?: boolean | undefined;
         tools?: boolean | undefined;
+        system?: boolean | undefined;
         context?: boolean | undefined;
     }, {
-        system?: boolean | undefined;
         tools?: boolean | undefined;
+        system?: boolean | undefined;
         context?: boolean | undefined;
     }>>;
 }, "strip", z.ZodTypeAny, {
     task: string;
-    system?: string | undefined;
     tools?: string | undefined;
+    system?: string | undefined;
     context?: string | undefined;
     cacheControl?: {
-        system?: boolean | undefined;
         tools?: boolean | undefined;
+        system?: boolean | undefined;
         context?: boolean | undefined;
     } | undefined;
 }, {
     task: string;
-    system?: string | undefined;
     tools?: string | undefined;
+    system?: string | undefined;
     context?: string | undefined;
     cacheControl?: {
-        system?: boolean | undefined;
         tools?: boolean | undefined;
+        system?: boolean | undefined;
         context?: boolean | undefined;
     } | undefined;
 }>;
@@ -98,15 +84,4 @@ export interface AssembleClaudeCacheBlocksResult {
     payload: ClaudeStreamJsonUserMessage;
     markedBlockCount: number;
 }
-/**
- * Slice κ: build the Claude `--input-format stream-json` payload from
- * a `PromptParts`. Each non-empty part becomes one content block in
- * `system → tools → context → task` order; parts whose name is `true`
- * in `cacheControl` get `cache_control: {type:"ephemeral", ttl:"1h"}`.
- *
- * Empty parts are skipped (no zero-byte blocks) — a true flag on an
- * empty part is silently a no-op and not counted in `markedBlockCount`.
- * The `task` block is never marked, even if a caller accidentally
- * tries (the schema doesn't expose `task` in `cacheControl`).
- */
 export declare function assembleClaudeCacheBlocks(parts: PromptParts): AssembleClaudeCacheBlocksResult;

package/dist/prompt-parts.js

@@ -48,17 +48,6 @@ export function resolvePromptInput(input) {
         stablePrefixTokens: null,
     };
 }
-/**
- * Slice κ: build the Claude `--input-format stream-json` payload from
- * a `PromptParts`. Each non-empty part becomes one content block in
- * `system → tools → context → task` order; parts whose name is `true`
- * in `cacheControl` get `cache_control: {type:"ephemeral", ttl:"1h"}`.
- *
- * Empty parts are skipped (no zero-byte blocks) — a true flag on an
- * empty part is silently a no-op and not counted in `markedBlockCount`.
- * The `task` block is never marked, even if a caller accidentally
- * tries (the schema doesn't expose `task` in `cacheControl`).
- */
 export function assembleClaudeCacheBlocks(parts) {
     const blocks = [];
     let markedBlockCount = 0;

package/dist/provider-status.d.ts

@@ -30,12 +30,4 @@ export interface GeminiAuthStatus {
     status: "present" | "not_found";
     methods: GeminiAuthMethods;
 }
-/**
- * U27: Detect Gemini auth across all supported methods.
- * Returns "present" if ANY of:
- *   - OAuth credential file present (~/.gemini/oauth_creds.json, etc.)
- *   - GEMINI_API_KEY env var set and non-empty
- *   - GOOGLE_API_KEY env var set and non-empty
- *   - GOOGLE_CLOUD_PROJECT set AND GOOGLE_GENAI_USE_VERTEXAI=true
- */
 export declare function geminiAuthStatus(env?: NodeJS.ProcessEnv, home?: string): GeminiAuthStatus;

package/dist/provider-status.js

@@ -12,8 +12,6 @@ const VERSION_ARGS = {
     grok: ["--version"],
     mistral: ["--version"],
 };
-// Mistral Vibe ships as the `vibe` binary (PyPI package mistral-vibe); the gateway
-// uses `mistral` as the provider key but invokes `vibe` on the shell.
 export const PROVIDER_COMMANDS = {
     claude: "claude",
     codex: "codex",
@@ -130,7 +128,6 @@ function inferLoginStatus(provider, exitCode, output) {
                 return "not_authenticated";
         }
         catch {
-            // Fall through to text heuristics.
         }
     }
     if (/not\s+(logged|signed|authenticated)\s*in|unauthenticated|login required|not authorized/i.test(output)) {
@@ -158,14 +155,6 @@ function loginCheckDetail(provider, status, exitCode) {
         return `${provider} login check exited non-zero without exposing credential material.`;
     return `${provider} login check completed, but the output did not clearly indicate login state.`;
 }
-/**
- * U27: Detect Gemini auth across all supported methods.
- * Returns "present" if ANY of:
- *   - OAuth credential file present (~/.gemini/oauth_creds.json, etc.)
- *   - GEMINI_API_KEY env var set and non-empty
- *   - GOOGLE_API_KEY env var set and non-empty
- *   - GOOGLE_CLOUD_PROJECT set AND GOOGLE_GENAI_USE_VERTEXAI=true
- */
 export function geminiAuthStatus(env = process.env, home = homedir()) {
     const candidates = [
         join(home, ".gemini", "oauth_creds.json"),