@elvatis_com/openclaw-cli-bridge-elvatis 1.9.1 → 2.0.0

package/.ai/handoff/STATUS.md

@@ -1,55 +1,58 @@
 # STATUS — openclaw-cli-bridge-elvatis
 
-## Current Version: 1.7.3
+## Current Version: 1.9.2
 
-## All 4 Providers Available — persistent Chromium profiles
-| Provider | Status | Models | Login Cmd | Cookie Expiry |
-|---|---|---|---|---|
-| Grok | ✅ | web-grok/grok-3, grok-3-fast, grok-3-mini, grok-3-mini-fast | /grok-login | ~178d |
-| Gemini | ✅ | web-gemini/gemini-2-5-pro, gemini-2-5-flash, gemini-3-pro, gemini-3-flash | /gemini-login | ~398d |
-| Claude.ai | ✅ | web-claude/claude-sonnet, claude-opus, claude-haiku | /claude-login | ~364d |
-| ChatGPT | ✅ | web-chatgpt/gpt-4o, gpt-4o-mini, gpt-4.1, o3, o4-mini, gpt-5, gpt-5-mini | /chatgpt-login | ~6d (re-run /chatgpt-login to refresh) |
+- **npm:** @elvatis_com/openclaw-cli-bridge-elvatis (not yet published to npm)
+- **ClawHub:** openclaw-cli-bridge-elvatis@1.9.2
+- **GitHub:** https://github.com/elvatis/openclaw-cli-bridge-elvatis/releases/tag/v1.9.2
+
+## CLI Model Token Limits (corrected in v1.9.2)
+| Model | Context Window | Max Output |
+|---|---|---|
+| Claude Opus 4.6 (CLI) | 1,000,000 | 128,000 |
+| Claude Sonnet 4.6 (CLI) | 1,000,000 | 64,000 |
+| Claude Haiku 4.5 (CLI) | 200,000 | 64,000 |
+| Gemini 2.5 Pro (CLI) | 1,048,576 | 65,535 |
+| Gemini 2.5 Flash (CLI) | 1,048,576 | 65,535 |
+| Gemini 3 Pro Preview (CLI) | 1,048,576 | 65,536 |
+| Gemini 3 Flash Preview (CLI) | 1,048,576 | 65,536 |
+
+## Architecture
+- **Proxy server:** `http://127.0.0.1:31337/v1` (OpenAI-compatible)
+- **OpenClaw connects via** `vllm` provider with `api: openai-completions`
+- **CLI models** (`cli-claude/*`, `cli-gemini/*`): plain text completions only — NO tool/function call support
+- **Web-session models** (`web-grok/*`, `web-gemini/*`): browser-based, require `/xxx-login`
+- **Codex models** (`openai-codex/*`): OAuth auth bridge
+- **BitNet** (`local-bitnet/*`): local CPU inference
+
+## Tool Support Limitation
+CLI models explicitly reject tool/function call requests (HTTP 400):
+```
+Model cli-claude/claude-opus-4-6 does not support tool/function calls.
+Use a native API model (e.g. github-copilot/gpt-5-mini) for agents that need tools.
+```
+This is by design — CLI tools output plain text only.
+
+## All 4 Browser Providers
+| Provider | Models | Login Cmd | Profile Dir |
+|---|---|---|---|
+| Grok | web-grok/grok-3, grok-3-fast, grok-3-mini, grok-3-mini-fast | /grok-login | ~/.openclaw/grok-profile/ |
+| Gemini | web-gemini/gemini-2-5-pro, gemini-2-5-flash, gemini-3-pro, gemini-3-flash | /gemini-login | ~/.openclaw/gemini-profile/ |
+| Claude | web-claude/* (removed in v1.6.x) | /claude-login | ~/.openclaw/claude-profile/ |
+| ChatGPT | web-chatgpt/* (removed in v1.6.x) | /chatgpt-login | ~/.openclaw/chatgpt-profile/ |
 
 ## Stats
-- 22 total models (6 CLI + 16 web-session)
-- 96/96 tests green (8 test files)
-- All 4 providers use launchPersistentContext — sessions survive gateway restarts
-- /bridge-status shows cookie-based status (independent of in-memory context)
-
-## Architecture: Browser Lifecycle
-- **Profile dirs:** `~/.openclaw/{grok,gemini,claude,chatgpt}-profile/`
-- **On plugin start:** startup restore attempts headless reconnect from saved profiles (5s delay)
-- **On /xxx-login:** headed browser, user logs in, cookies baked to profile automatically
-- **On request (no in-memory ctx):** proxy lazy-launches persistent context on first request
-- **On /xxx-logout:** closes context + deletes profile + clears expiry file
-- **bridge-status:** uses cookie expiry files as source of truth (not in-memory state)
-  - ✅ active — browser connected and verified
-  - 🟡 logged in, browser not loaded — cookies valid, lazy-loads on first request
-  - 🔴 session expired — needs /xxx-login
-  - ⚪ never logged in
-
-## Cookie Expiry Tracking (fixed in v1.7.3)
-All 4 providers now track the **longest-lived** auth cookie instead of the shortest:
-- Claude: `sessionKey` (~1 year) — was `__cf_bm` (Cloudflare, ~30 min) causing false alerts
-- ChatGPT: longest of `__Secure-next-auth.session-token` / `_puid` / `oai-did`
-- Gemini: longest of `__Secure-1PSID` / `__Secure-3PSID` / `SID`
-- Grok: longest of `sso` / `sso-rw`
-
-## Release History
-- v1.7.3 (2026-03-13): Fix cookie expiry tracking — use longest-lived auth cookie for all 4 providers
-- v1.7.2 (2026-03-13): Cookie-first startup restore (skip fragile browser selector check)
-- v1.7.1 (2026-03-13): /status HTML dashboard at :31337
-- v1.7.0 (2026-03-13): Startup restore timeout fix, auto-relogin, keep-alive verification, vitest suite
-- v1.6.1 (2026-03-13): Fix /bridge-status — use cookie expiry as source of truth
-- v1.6.0 (2026-03-13): Persistent Chromium profiles for all 4 providers (Claude web + ChatGPT)
-- v1.5.1 (2026-03-12): Fix hardcoded plugin version
-- v1.5.0 (2026-03-12): Remove /claude-login and /chatgpt-login (pre-v1.6.0 interim)
-- v1.4.0 (2026-03-12): Persistent browser fallback for Claude/Gemini/ChatGPT (no CDP required)
-- v1.3.5 (2026-03-12): Startup restore guard (SIGUSR1 OOM fix)
-- v1.3.0 (2026-03-11): Browser auto-reconnect after gateway restart
-- v1.0.0 (2026-03-11): All 4 providers headless — 96/96 tests
-
-## Next Steps
-- /chatgpt-login should be re-run soon (~6d left on _puid cookie)
-- Gemini model switching via UI (2.5 Pro vs Flash vs 3)
-- Context-window management for long conversations
+- 22+ total models (7 CLI + 5 Codex + 4 Grok + 4 Gemini + 1 BitNet)
+- Persistent Chromium profiles survive gateway restarts
+- /bridge-status shows cookie-based status
+
+## Release History (recent)
+- v1.9.2 (2026-03-15): Fix maxTokens/contextWindow for all CLI_MODELS (were 8192, now correct per vendor specs)
+- v1.9.1: Previous stable
+- v1.7.3 (2026-03-13): Fix cookie expiry tracking
+- v1.7.0 (2026-03-13): Startup restore timeout fix, auto-relogin, vitest suite
+- v1.6.0 (2026-03-13): Persistent Chromium profiles for all 4 providers
+
+## Known Issues
+- CLI models cannot do tool calls (by design — plain text proxy)
+- Opus via CLI proxy may halluzinate XML tool-call tags when maxTokens was too low (fixed in v1.9.2)

package/CONTRIBUTING.md

@@ -57,6 +57,24 @@ clawhub publish "$TMPDIR" --slug openclaw-cli-bridge-elvatis --version X.Y.Z \
 
 ---
 
+## 🚨 Doku-Regel (PFLICHT)
+
+**Wenn ein Feature hinzukommt, geändert oder entfernt wird → SOFORT in ALLEN Doku-Dateien aktualisieren.**
+
+Gilt für: `README.md`, `SKILL.md`, `CONTRIBUTING.md`, `openclaw.plugin.json`, ClawHub, npm, GitHub.
+
+**Konkret für Slash Commands:**
+- Neuer Command (`/cli-xyz`) → sofort in README Utility-Tabelle + ASCII-Beispielblock eintragen
+- Command entfernt → sofort aus README + SKILL.md entfernen, nicht beim nächsten Release
+- Command umbenannt → beide Stellen gleichzeitig ändern
+
+**Tested Badge:**
+- Neues Feature getestet → sofort `✅ Tested` Badge in README setzen
+
+Kein "machen wir beim nächsten Release" — immer sofort, im gleichen Commit.
+
+---
+
 ## Versionsstellen — alle prüfen vor Release
 
 ```bash

package/README.md

@@ -2,7 +2,7 @@
 
 > OpenClaw plugin that bridges locally installed AI CLIs (Codex, Gemini, Claude Code) as model providers — with slash commands for instant model switching, restore, health testing, and model listing.
 
-**Current version:** `1.9.1`
+**Current version:** `1.9.2`
 
 ---
 
@@ -70,6 +70,7 @@ All commands use gateway-level `commands.allowFrom` for authorization (`requireA
 | `/cli-back` | Restore the model active **before** the last `/cli-*` switch |
 | `/cli-test [model]` | One-shot proxy health check — **does NOT switch your active model** |
 | `/cli-list` | Show all registered CLI bridge models with commands |
+| `/cli-help` | Full reference card — CLI/Codex/Web/BitNet sections, expiry info, quick examples, dashboard links |
 
 **`/cli-back` details:**
 - Before every `/cli-*` switch the current model is saved to `~/.openclaw/cli-bridge-state.json`
@@ -217,7 +218,8 @@ openclaw models auth login --provider openai-codex
   Utility
     /cli-back            Restore previous model
     /cli-test [model]    Health check (no model switch)
-    /cli-list            This overview
+    /cli-list            All models with slash commands + dashboard URL
+    /cli-help            Full reference card (sections, expiry, examples, links)
 
   Proxy: 127.0.0.1:31337
 ```
@@ -356,6 +358,15 @@ npm run ci          # lint + typecheck + test
 
 ## Changelog
 
+### v1.9.2
+- **fix:** Correct `maxTokens` and `contextWindow` for all CLI_MODELS — were hardcoded to 8192 output tokens
+  - Claude Opus 4.6: 1M context / 128k output (was 200k/8k)
+  - Claude Sonnet 4.6: 1M context / 64k output (was 200k/8k)
+  - Claude Haiku 4.5: 200k context / 64k output (was 200k/8k)
+  - Gemini 2.5 Pro/Flash: 1M context / 65k output (was 1M/8k)
+  - Gemini 3 Pro/Flash Preview: 1M context / 65k output (was 1M/8k)
+  - Web-session Gemini models: same corrections
+
 ### v1.9.1
 - **feat:** Full slash command mapping on status page — all models now show their /cli-* command
 - **fix:** Register missing slash commands: /cli-codex-spark, /cli-codex52, /cli-codex-mini, /cli-gemini3-flash (documented but never registered)

package/openclaw.plugin.json

@@ -2,7 +2,7 @@
   "id": "openclaw-cli-bridge-elvatis",
   "slug": "openclaw-cli-bridge-elvatis",
   "name": "OpenClaw CLI Bridge",
-  "version": "1.9.1",
+  "version": "2.0.0",
   "license": "MIT",
   "description": "Phase 1: openai-codex auth bridge. Phase 2: local HTTP proxy routing model calls through gemini/claude CLIs (vllm provider).",
   "providers": [
@@ -38,4 +38,4 @@
       }
     }
   }
-}
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@elvatis_com/openclaw-cli-bridge-elvatis",
-  "version": "1.9.1",
+  "version": "2.0.0",
   "description": "Bridges gemini, claude, and codex CLI tools as OpenClaw model providers. Reads existing CLI auth without re-login.",
   "type": "module",
   "openclaw": {

package/src/cli-runner.ts

@@ -1,19 +1,25 @@
 /**
  * cli-runner.ts
  *
- * Spawns CLI subprocesses (gemini, claude) and captures their output.
- * Input: OpenAI-format messages → formatted prompt string → CLI stdin.
+ * Spawns CLI subprocesses (gemini, claude, codex, opencode, pi) and captures their output.
+ * Input: OpenAI-format messages → formatted prompt string → CLI stdin (or CLI arg).
  *
- * Both Gemini and Claude receive the prompt via stdin to avoid:
- *   - E2BIG (arg list too long) for large conversation histories
- *   - Gemini agentic mode (triggered by @file syntax + workspace cwd)
+ * Prompt delivery:
+ *   - Gemini/Claude/Codex receive the prompt via stdin to avoid E2BIG and agentic mode.
+ *   - OpenCode receives the prompt as a CLI argument (`opencode run "prompt"`).
+ *   - Pi receives the prompt via `-p "prompt"` flag.
  *
- * Gemini is always spawned with cwd = tmpdir() so it doesn't scan the
- * workspace and enter agentic mode.
+ * Workdir isolation:
+ *   - Gemini: defaults to tmpdir() (prevents agentic workspace scanning).
+ *   - Claude/Codex: defaults to homedir().
+ *   - OpenCode/Pi: defaults to homedir().
+ *   - All runners accept an explicit `workdir` override via RouteOptions.
  */
 
-import { spawn } from "node:child_process";
+import { spawn, execSync } from "node:child_process";
 import { tmpdir, homedir } from "node:os";
+import { existsSync } from "node:fs";
+import { join } from "node:path";
 import { ensureClaudeToken, refreshClaudeToken } from "./claude-auth.js";
 
 /** Max messages to include in the prompt sent to the CLI. */
@@ -198,6 +204,41 @@ export function runCli(
   });
 }
 
+/**
+ * Spawn a CLI with the prompt delivered as a CLI argument (not stdin).
+ * Used by OpenCode which expects `opencode run "prompt"`.
+ */
+export function runCliWithArg(
+  cmd: string,
+  args: string[],
+  timeoutMs = 120_000,
+  opts: RunCliOptions = {}
+): Promise<CliRunResult> {
+  const cwd = opts.cwd ?? homedir();
+
+  return new Promise((resolve, reject) => {
+    const proc = spawn(cmd, args, {
+      timeout: timeoutMs,
+      env: buildMinimalEnv(),
+      cwd,
+    });
+
+    let stdout = "";
+    let stderr = "";
+
+    proc.stdout.on("data", (d: Buffer) => { stdout += d.toString(); });
+    proc.stderr.on("data", (d: Buffer) => { stderr += d.toString(); });
+
+    proc.on("close", (code) => {
+      resolve({ stdout: stdout.trim(), stderr: stderr.trim(), exitCode: code ?? 0 });
+    });
+
+    proc.on("error", (err) => {
+      reject(new Error(`Failed to spawn '${cmd}': ${err.message}`));
+    });
+  });
+}
+
 // ──────────────────────────────────────────────────────────────────────────────
 // Gemini CLI
 // ──────────────────────────────────────────────────────────────────────────────
@@ -215,17 +256,20 @@ export function runCli(
  * Gemini CLI: -p "" triggers headless mode; stdin content is the actual prompt
  * (per Gemini docs: "prompt is appended to input on stdin (if any)").
  *
- * cwd = tmpdir() — neutral empty-ish dir, prevents workspace context scanning.
+ * cwd = tmpdir() by default — neutral empty-ish dir, prevents workspace context scanning.
+ * Override with explicit workdir.
  */
 export async function runGemini(
   prompt: string,
   modelId: string,
-  timeoutMs: number
+  timeoutMs: number,
+  workdir?: string
 ): Promise<string> {
   const model = stripPrefix(modelId);
   // -p "" = headless mode trigger; actual prompt arrives via stdin
   const args = ["-m", model, "-p", ""];
-  const result = await runCli("gemini", args, prompt, timeoutMs, { cwd: tmpdir() });
+  const cwd = workdir ?? tmpdir();
+  const result = await runCli("gemini", args, prompt, timeoutMs, { cwd });
 
   // Filter out [WARN] lines from stderr (Gemini emits noisy permission warnings)
   const cleanStderr = result.stderr
@@ -248,11 +292,13 @@ export async function runGemini(
 /**
  * Run Claude Code CLI in headless mode with prompt delivered via stdin.
  * Strips the model prefix ("cli-claude/claude-opus-4-6" → "claude-opus-4-6").
+ * cwd = homedir() by default. Override with explicit workdir.
  */
 export async function runClaude(
   prompt: string,
   modelId: string,
-  timeoutMs: number
+  timeoutMs: number,
+  workdir?: string
 ): Promise<string> {
   // Proactively refresh OAuth token if it's about to expire (< 5 min remaining).
   // No-op for API-key users.
@@ -267,7 +313,8 @@ export async function runClaude(
     "--model", model,
   ];
 
-  const result = await runCli("claude", args, prompt, timeoutMs);
+  const cwd = workdir ?? homedir();
+  const result = await runCli("claude", args, prompt, timeoutMs, { cwd });
 
   // On 401: attempt one token refresh + retry before giving up.
   if (result.exitCode !== 0 && result.stdout.length === 0) {
@@ -275,7 +322,7 @@ export async function runClaude(
     if (stderr.includes("401") || stderr.includes("Invalid authentication credentials") || stderr.includes("authentication_error")) {
       // Refresh and retry once
       await refreshClaudeToken();
-      const retry = await runCli("claude", args, prompt, timeoutMs);
+      const retry = await runCli("claude", args, prompt, timeoutMs, { cwd });
       if (retry.exitCode !== 0 && retry.stdout.length === 0) {
         const retryStderr = retry.stderr || "(no output)";
         if (retryStderr.includes("401") || retryStderr.includes("authentication_error") || retryStderr.includes("Invalid authentication credentials")) {
@@ -294,6 +341,97 @@ export async function runClaude(
   return result.stdout;
 }
 
+// ──────────────────────────────────────────────────────────────────────────────
+// Codex CLI
+// ──────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Ensure the workdir is a git repository. Codex CLI requires a git repo.
+ * If the directory exists but is not a git repo, run `git init`.
+ */
+function ensureGitRepo(dir: string): void {
+  if (!existsSync(join(dir, ".git"))) {
+    execSync("git init", { cwd: dir, stdio: "ignore" });
+  }
+}
+
+/**
+ * Run Codex CLI in non-interactive mode with prompt via stdin.
+ * cwd = homedir() by default. Override with explicit workdir.
+ * Auto-initializes git if workdir is not already a git repo.
+ */
+export async function runCodex(
+  prompt: string,
+  modelId: string,
+  timeoutMs: number,
+  workdir?: string
+): Promise<string> {
+  const model = stripPrefix(modelId);
+  const args = ["--model", model, "--quiet", "--full-auto"];
+  const cwd = workdir ?? homedir();
+
+  // Codex requires a git repo in the working directory
+  ensureGitRepo(cwd);
+
+  const result = await runCli("codex", args, prompt, timeoutMs, { cwd });
+
+  if (result.exitCode !== 0 && result.stdout.length === 0) {
+    throw new Error(`codex exited ${result.exitCode}: ${result.stderr || "(no output)"}`);
+  }
+
+  return result.stdout || result.stderr;
+}
+
+// ──────────────────────────────────────────────────────────────────────────────
+// OpenCode CLI
+// ──────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Run OpenCode CLI. Prompt is passed as a CLI argument: `opencode run "prompt"`.
+ * cwd = homedir() by default. Override with explicit workdir.
+ */
+export async function runOpenCode(
+  prompt: string,
+  _modelId: string,
+  timeoutMs: number,
+  workdir?: string
+): Promise<string> {
+  const args = ["run", prompt];
+  const cwd = workdir ?? homedir();
+  const result = await runCliWithArg("opencode", args, timeoutMs, { cwd });
+
+  if (result.exitCode !== 0 && result.stdout.length === 0) {
+    throw new Error(`opencode exited ${result.exitCode}: ${result.stderr || "(no output)"}`);
+  }
+
+  return result.stdout || result.stderr;
+}
+
+// ──────────────────────────────────────────────────────────────────────────────
+// Pi CLI
+// ──────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Run Pi CLI in non-interactive mode: `pi -p "prompt"`.
+ * cwd = homedir() by default. Override with explicit workdir.
+ */
+export async function runPi(
+  prompt: string,
+  _modelId: string,
+  timeoutMs: number,
+  workdir?: string
+): Promise<string> {
+  const args = ["-p", prompt];
+  const cwd = workdir ?? homedir();
+  const result = await runCliWithArg("pi", args, timeoutMs, { cwd });
+
+  if (result.exitCode !== 0 && result.stdout.length === 0) {
+    throw new Error(`pi exited ${result.exitCode}: ${result.stderr || "(no output)"}`);
+  }
+
+  return result.stdout || result.stderr;
+}
+
 // ──────────────────────────────────────────────────────────────────────────────
 // Model allowlist (T-103)
 // ──────────────────────────────────────────────────────────────────────────────
@@ -319,6 +457,16 @@ export const DEFAULT_ALLOWED_CLI_MODELS: ReadonlySet<string> = new Set([
   // Aliases (map to preview variants internally)
   "cli-gemini/gemini-3-pro",   // alias → gemini-3-pro-preview
   "cli-gemini/gemini-3-flash", // alias → gemini-3-flash-preview
+  // Codex CLI
+  "openai-codex/gpt-5.3-codex",
+  "openai-codex/gpt-5.3-codex-spark",
+  "openai-codex/gpt-5.2-codex",
+  "openai-codex/gpt-5.4",
+  "openai-codex/gpt-5.1-codex-mini",
+  // OpenCode CLI
+  "opencode/default",
+  // Pi CLI
+  "pi/default",
 ]);
 
 /** Normalize model aliases to their canonical CLI model names. */
@@ -341,12 +489,20 @@ export interface RouteOptions {
    * Defaults to DEFAULT_ALLOWED_CLI_MODELS.
    */
   allowedModels?: ReadonlySet<string> | null;
+  /**
+   * Working directory for the CLI subprocess.
+   * Overrides the per-runner default (tmpdir for gemini, homedir for others).
+   */
+  workdir?: string;
 }
 
 /**
  * Route a chat completion to the correct CLI based on model prefix.
- *   cli-gemini/<id>  → gemini CLI
- *   cli-claude/<id>  → claude CLI
+ *   cli-gemini/<id>      → gemini CLI
+ *   cli-claude/<id>      → claude CLI
+ *   openai-codex/<id>    → codex CLI
+ *   opencode/<id>        → opencode CLI
+ *   pi/<id>              → pi CLI
  *
  * Enforces DEFAULT_ALLOWED_CLI_MODELS by default (T-103).
  * Pass `allowedModels: null` to skip the allowlist check.
@@ -379,11 +535,14 @@ export async function routeToCliRunner(
   // Resolve aliases (e.g. gemini-3-pro → gemini-3-pro-preview) after allowlist check
   const resolved = normalizeModelAlias(normalized);
 
-  if (resolved.startsWith("cli-gemini/")) return runGemini(prompt, resolved, timeoutMs);
-  if (resolved.startsWith("cli-claude/")) return runClaude(prompt, resolved, timeoutMs);
+  if (resolved.startsWith("cli-gemini/"))   return runGemini(prompt, resolved, timeoutMs, opts.workdir);
+  if (resolved.startsWith("cli-claude/"))   return runClaude(prompt, resolved, timeoutMs, opts.workdir);
+  if (resolved.startsWith("openai-codex/")) return runCodex(prompt, resolved, timeoutMs, opts.workdir);
+  if (resolved.startsWith("opencode/"))     return runOpenCode(prompt, resolved, timeoutMs, opts.workdir);
+  if (resolved.startsWith("pi/"))           return runPi(prompt, resolved, timeoutMs, opts.workdir);
 
   throw new Error(
-    `Unknown CLI bridge model: "${model}". Use "vllm/cli-gemini/<model>" or "vllm/cli-claude/<model>".`
+    `Unknown CLI bridge model: "${model}". Use "vllm/cli-gemini/<model>", "vllm/cli-claude/<model>", "openai-codex/<model>", "opencode/<model>", or "pi/<model>".`
   );
 }
 

package/src/grok-client.ts

@@ -51,7 +51,7 @@ const STABLE_INTERVAL_MS = 500; // ms between stability checks
 
 function resolveModel(m?: string): string {
   const clean = (m ?? "grok-3").replace("web-grok/", "");
-  const allowed = ["grok-3", "grok-3-fast", "grok-3-mini", "grok-3-mini-fast"];
+  const allowed = ["grok-4", "grok-3", "grok-3-fast", "grok-3-mini", "grok-3-mini-fast"];
   return allowed.includes(clean) ? clean : "grok-3";
 }