consult-llm-mcp 2.5.2 → 2.5.4

package/CHANGELOG.md

@@ -0,0 +1,198 @@
+# Changelog
+
+## v2.5.4 (2026-03-07)
+
+- Rewrote server from TypeScript to Rust
+- Distributed as cross-compiled native binaries (macOS arm64/x64, Linux
+  x64/arm64) via npm with a POSIX sh launcher
+- Added `MCP_DEBUG_STDIN` env var for raw stdin transport logging
+
+## v2.5.3 (2026-03-06)
+
+- Added `CONSULT_LLM_` prefix to backend and reasoning effort env vars:
+  `CONSULT_LLM_GEMINI_BACKEND`, `CONSULT_LLM_OPENAI_BACKEND`,
+  `CONSULT_LLM_CODEX_REASONING_EFFORT`. Old unprefixed names still work with a
+  deprecation warning.
+
+## v2.5.2 (2026-03-06)
+
+- Fixed Codex CLI thread resumption failing due to unsupported `--add-dir` flag
+  in `codex exec resume`
+
+## v2.5.1 (2026-03-06)
+
+- Consult skill now queries both Gemini and Codex in parallel by default, with
+  `--gemini` and `--codex` flags for single-model consultation
+- CLI backends now receive main worktree path as additional context when running
+  inside a git worktree
+- CLI backends now detect external file directories (outside workspace) and pass
+  them to Gemini/Codex so referenced files are accessible
+- Debate skill now supports multi-round debates via `--rounds` flag (default 2,
+  max 3)
+- Reduced anchoring bias in debate/debate-vs final review phase
+- Added `install-skills` script for easy skill installation
+
+## v2.5.0 (2026-03-05)
+
+- Added gpt-5.4 model support ($2.50/$15 per million tokens)
+
+## v2.4.2 (2026-02-28)
+
+- Added gemini-3.1-pro-preview model support
+- Filter unavailable models from the tool schema based on configured API keys
+  and CLI backends, preventing errors when selecting unconfigured models
+
+## v2.4.1 (2026-02-25)
+
+- Updated tool description to tell callers not to inline file contents in the
+  prompt field, since the server reads files automatically via the `files`
+  parameter
+
+## v2.3.0 (2026-02-25)
+
+- Added Cursor CLI (`cursor-agent`) as a new executor backend
+- Replaced `GEMINI_MODE`/`OPENAI_MODE` with `GEMINI_BACKEND`/`OPENAI_BACKEND`
+  for backend routing (legacy env vars still work with deprecation warnings)
+- Added `CONSULT_LLM_EXTRA_MODELS` environment variable for adding models
+  without code changes
+- Removed gpt-5.1 and Claude models from built-in model list
+
+## v2.2.0 (2026-02-14)
+
+- Added `task_mode` parameter for adaptive system prompts with five modes:
+  `review`, `plan`, `create`, `debug`, and `general` (default)
+
+## v2.1.0 (2026-02-12)
+
+- Added multi-turn conversation support for CLI modes via `thread_id` parameter
+  - Codex CLI: uses `--json` output and `exec resume` for session continuity
+  - Gemini CLI: uses `-o json` output and `-r` flag for session resume
+  - Responses include a `[thread_id:xxx]` prefix for follow-up requests
+- Replaced generic CLI executor with dedicated Codex and Gemini executors
+- Added debate skill example (`skills/debate/SKILL.md`) showcasing multi-turn
+  conversations
+
+## v2.0.1 (2026-02-05)
+
+- Added gpt-5.3-codex model support
+
+## v2.0.0 (2026-02-04)
+
+- Log files now stored in XDG state directory
+  (`~/.local/state/consult-llm-mcp/`) instead of `~/.consult-llm-mcp/logs/`,
+  following the XDG Base Directory Specification
+
+## v1.7.2 (2026-02-04)
+
+- Extracted model definitions to dedicated module, resolving a circular
+  dependency between config and schema
+
+## v1.7.1 (2026-02-04)
+
+- Removed o3 model, succeeded by gpt-5.2 which is now the default OpenAI model
+- Documented model selection behavior and `CONSULT_LLM_ALLOWED_MODELS` usage
+
+## v1.7.0 (2026-01-29)
+
+- Added configurable system prompt path via `CONSULT_LLM_SYSTEM_PROMPT_PATH`
+  environment variable
+
+## v1.5.0 (2026-01-12)
+
+- Added gpt-5.2-codex model support
+- Added `CONSULT_LLM_ALLOWED_MODELS` environment variable to filter which models
+  are available in the tool schema
+
+## v1.4.7 (2025-12-12)
+
+- Added gpt-5.2 model support
+
+## v1.4.6 (2025-12-03)
+
+- Removed 5-minute timeout from CLI executors
+
+## v1.4.5 (2025-12-02)
+
+- Improved system prompt by removing redundant "no critical issues found"
+  statement
+- Fixed `init-prompt` command to use npx
+- Moved skill and slash command examples to separate files
+
+## v1.4.4 (2025-11-26)
+
+- Added support for gemini-3-pro-preview model
+
+## v1.4.3 (2025-11-22)
+
+- Added `--skip-git-repo-check` to Codex CLI args, fixing issues when running
+  outside git repositories
+
+## v1.4.2 (2025-11-21)
+
+- Added gpt-5.1-codex-max model support
+- Made Codex CLI reasoning effort configurable
+- Fixed Codex CLI execution
+
+## v1.4.1 (2025-11-19)
+
+- Added test suite with vitest
+- Fixed web mode file handling
+- Split server entry point for better modularity
+
+## v1.4.0 (2025-11-18)
+
+- Added Codex CLI support as a new executor backend
+- Refactored LLM execution to a functional executor pattern
+- Migrated to Zod v4 with native JSON schema generation
+- Added example Claude Code skill to README
+
+## v1.3.0 (2025-11-16)
+
+- Added web mode: copies the formatted prompt to clipboard for pasting into
+  browser-based LLMs instead of querying an API directly
+
+## v1.2.0 (2025-10-25)
+
+- Added custom system prompt support, configurable via file
+- Added environment variable documentation
+- Set up ESLint and code quality checks
+
+## v1.1.2 (2025-10-25)
+
+- Added `--version` flag to display the server version
+
+## v1.1.1 (2025-10-25)
+
+- Added Gemini CLI mode support with a dedicated system prompt and instructions
+- Enforced code file context requirement in the tool description
+- Improved code review guidance in prompts
+
+## v1.1.0 (2025-07-28)
+
+- Added Gemini CLI mode as an alternative to the Gemini API, with free quota
+  support
+- Added debug logging and configuration logging with sensitive data redaction
+
+## v1.0.5 (2025-06-29)
+
+- Improved system prompt for more concise and critical analysis
+- Updated tool description to emphasize neutral, unbiased questioning
+- Moved main prompt to the `prompt` field instead of embedding in markdown
+
+## v1.0.4 (2025-06-25)
+
+- Added system prompt to LLM queries for enhanced analysis and recommendations
+- Added `prompt` parameter to the tool
+
+## v1.0.3 (2025-06-23)
+
+- Added deepseek-reasoner and gemini-2.5-pro model options
+- Added API key validation checks for OpenAI, Gemini, and DeepSeek models
+
+## v1.0.2 (2025-06-23)
+
+- Added server version logging and version in server metadata
+
+## v1.0.1 (2025-06-23)
+
+Initial release.

package/README.md

@@ -64,7 +64,7 @@ to bring in the heavy artillery. Supports multi-turn conversations.
      -e OPENAI_API_KEY=your_openai_key \
      -e GEMINI_API_KEY=your_gemini_key \
      -e DEEPSEEK_API_KEY=your_deepseek_key \
-     -e GEMINI_BACKEND=gemini-cli \
+     -e CONSULT_LLM_GEMINI_BACKEND=gemini-cli \
      -- npx -y consult-llm-mcp
    ```
 
@@ -344,7 +344,7 @@ or use your Google AI Pro subscription.
 **Setup:**
 
 ```bash
-claude mcp add consult-llm -e GEMINI_BACKEND=gemini-cli -- npx -y consult-llm-mcp
+claude mcp add consult-llm -e CONSULT_LLM_GEMINI_BACKEND=gemini-cli -- npx -y consult-llm-mcp
 ```
 
 #### Codex CLI
@@ -359,12 +359,12 @@ Use OpenAI's Codex CLI for OpenAI models.
 **Setup:**
 
 ```bash
-claude mcp add consult-llm -e OPENAI_BACKEND=codex-cli -- npx -y consult-llm-mcp
+claude mcp add consult-llm -e CONSULT_LLM_OPENAI_BACKEND=codex-cli -- npx -y consult-llm-mcp
 ```
 
 <!-- prettier-ignore -->
 > [!TIP]
-> Set reasoning effort with `-e CODEX_REASONING_EFFORT=high`. Options:
+> Set reasoning effort with `-e CONSULT_LLM_CODEX_REASONING_EFFORT=high`. Options:
 > `none`, `minimal`, `low`, `medium`, `high`, `xhigh`.
 
 #### Cursor CLI
@@ -380,15 +380,17 @@ Use Cursor's agent CLI to route GPT and Gemini models through one tool.
 
 ```bash
 # Route GPT models through Cursor CLI
-claude mcp add consult-llm -e OPENAI_BACKEND=cursor-cli -- npx -y consult-llm-mcp
+claude mcp add consult-llm -e CONSULT_LLM_OPENAI_BACKEND=cursor-cli -- npx -y consult-llm-mcp
 
 # Route Gemini models through Cursor CLI
-claude mcp add consult-llm -e GEMINI_BACKEND=cursor-cli -- npx -y consult-llm-mcp
+claude mcp add consult-llm -e CONSULT_LLM_GEMINI_BACKEND=cursor-cli -- npx -y consult-llm-mcp
 
 # Route everything through Cursor CLI
 claude mcp add consult-llm \
-  -e OPENAI_BACKEND=cursor-cli \
-  -e GEMINI_BACKEND=cursor-cli \
+  -e CONSULT_LLM_OPENAI_BACKEND=cursor-cli \
+  -e CONSULT_LLM_GEMINI_BACKEND=cursor-cli \
+  -e CONSULT_LLM_CODEX_REASONING_EFFORT=high \
+  -e CONSULT_LLM_ALLOWED_MODELS="gemini-3-pro-preview,gpt-5.3-codex" \
   -- npx -y consult-llm-mcp
 ```
 
@@ -459,11 +461,12 @@ See the "Using web mode..." example above for a concrete transcript.
   - Options: `gpt-5.2` (default), `gpt-5.4`, `gemini-2.5-pro`,
     `gemini-3-pro-preview`, `gemini-3.1-pro-preview`, `deepseek-reasoner`,
     `gpt-5.3-codex`, `gpt-5.2-codex`
-- `GEMINI_BACKEND` - Backend for Gemini models (optional)
+- `CONSULT_LLM_GEMINI_BACKEND` - Backend for Gemini models (optional)
   - Options: `api` (default), `gemini-cli`, `cursor-cli`
-- `OPENAI_BACKEND` - Backend for OpenAI models (optional)
+- `CONSULT_LLM_OPENAI_BACKEND` - Backend for OpenAI models (optional)
   - Options: `api` (default), `codex-cli`, `cursor-cli`
-- `CODEX_REASONING_EFFORT` - Configure reasoning effort for Codex CLI (optional)
+- `CONSULT_LLM_CODEX_REASONING_EFFORT` - Configure reasoning effort for Codex
+  CLI (optional)
   - See [Codex CLI](#codex-cli) for details and available options
 - `CONSULT_LLM_EXTRA_MODELS` - Add models not in the built-in list (optional)
   - Comma-separated list, e.g., `grok-3,kimi-k2.5`
@@ -480,6 +483,11 @@ See the "Using web mode..." example above for a concrete transcript.
   (optional)
   - Overrides the default `~/.consult-llm-mcp/SYSTEM_PROMPT.md` location
   - Useful for project-specific prompts
+- `MCP_DEBUG_STDIN` - Log raw JSON-RPC messages received on stdin (optional)
+  - Set to `1` to enable
+  - Logs every message as `RAW RECV` entries and poll timing gaps as
+    `STDIN POLL` entries in `mcp.log`
+  - Useful for debugging transport-level issues
 
 ### Custom system prompt
 
@@ -777,51 +785,49 @@ forth before synthesizing and implementing. See
 
 To work on the MCP server locally and use your development version:
 
-1. Clone the repository and install dependencies:
+1. Clone the repository:
 
    ```bash
    git clone https://github.com/raine/consult-llm-mcp.git
    cd consult-llm-mcp
-   npm install
    ```
 
-2. Build the project:
+2. Build and test:
 
    ```bash
-   npm run build
+   cargo build
+   cargo test
+   just check  # format, lint, test
    ```
 
-3. Install globally from the local directory:
-
-   ```bash
-   npm link
-   ```
-
-4. Add the MCP server to Claude Code using the global command:
+3. Add the MCP server to Claude Code using your local build:
    ```bash
-   claude mcp add consult-llm -- consult-llm-mcp
+   claude mcp add consult-llm -- /path/to/consult-llm-mcp/target/debug/consult-llm-mcp
    ```
 
-Now when you make changes:
+Now when you make changes, rebuild with `cargo build` and restart Claude Code.
 
-1. Rebuild: `npm run build`
-2. Restart Claude Code to pick up the changes
+### Releasing
 
-Alternatively, you can use the dev script for development without building:
+Requires [zig](https://ziglang.org/),
+[cargo-zigbuild](https://github.com/rust-cross/cargo-zigbuild), and
+cross-compilation targets:
 
 ```bash
-claude mcp add consult-llm -- npm run dev
+brew install zig
+cargo install cargo-zigbuild
+rustup target add aarch64-apple-darwin x86_64-apple-darwin x86_64-unknown-linux-gnu aarch64-unknown-linux-gnu
 ```
 
-This runs the TypeScript source directly with `tsx`, allowing faster iteration
-without rebuilding.
-
-To unlink the global version later:
+Publish a new version:
 
 ```bash
-npm unlink -g
+scripts/publish patch  # or minor, major
 ```
 
+This cross-compiles for all platforms, stages the binaries, publishes to npm,
+then commits and tags.
+
 ## Related projects
 
 - [workmux](https://github.com/raine/workmux) — Git worktrees + tmux windows for

package/package.json

@@ -1,36 +1,7 @@
 {
   "name": "consult-llm-mcp",
-  "version": "2.5.2",
+  "version": "2.5.4",
   "description": "MCP server for consulting powerful AI models",
-  "type": "module",
-  "main": "dist/main.js",
-  "bin": {
-    "consult-llm-mcp": "./dist/main.js"
-  },
-  "scripts": {
-    "build": "./scripts/generate-version && tsc",
-    "dev": "tsx src/main.ts",
-    "start": "node dist/main.js",
-    "test": "vitest",
-    "test:run": "vitest run",
-    "format": "oxfmt --write .",
-    "check": "concurrently \"pnpm:check:*\"",
-    "check:types": "tsc --noEmit",
-    "check:format": "oxfmt --check .",
-    "check:lint": "biome lint .",
-    "check:test": "vitest run",
-    "install-global": "npm run build && npm install -g .",
-    "prepublishOnly": "pnpm run check && npm run build",
-    "publish:dry": "npm publish --dry-run"
-  },
-  "keywords": [
-    "mcp",
-    "openai",
-    "gemini",
-    "llm",
-    "ai"
-  ],
-  "author": "",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/raine/consult-llm-mcp.git"
@@ -39,28 +10,28 @@
   "bugs": {
     "url": "https://github.com/raine/consult-llm-mcp/issues"
   },
-  "files": [
-    "dist",
-    "README.md"
-  ],
-  "engines": {
-    "node": ">=18.0.0"
-  },
   "license": "MIT",
-  "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.27.1",
-    "clipboardy": "^5.3.1",
-    "openai": "^6.25.0",
-    "prettier": "^3.8.1",
-    "zod": "^4.3.6"
+  "bin": {
+    "consult-llm-mcp": "./run.sh"
   },
-  "devDependencies": {
-    "@biomejs/biome": "2.4.4",
-    "@types/node": "^25.3.0",
-    "concurrently": "^9.2.1",
-    "oxfmt": "^0.35.0",
-    "tsx": "^4.21.0",
-    "typescript": "^5.9.3",
-    "vitest": "^4.0.18"
-  }
+  "files": [
+    "run.sh",
+    "consult-llm-mcp-darwin-arm64",
+    "consult-llm-mcp-darwin-x64",
+    "consult-llm-mcp-linux-x64",
+    "consult-llm-mcp-linux-arm64",
+    "README.md",
+    "CHANGELOG.md"
+  ],
+  "os": [
+    "darwin",
+    "linux"
+  ],
+  "keywords": [
+    "mcp",
+    "openai",
+    "gemini",
+    "llm",
+    "ai"
+  ]
 }

package/run.sh

@@ -0,0 +1,36 @@
+#!/bin/sh
+set -e
+
+# Resolve symlinks to find the actual script directory (npm creates symlinks in .bin/)
+# When invoked via PATH, $0 may be just the command name without a path
+script="$0"
+case "$script" in
+*/*) ;;
+*) script="$(command -v "$script" 2>/dev/null || echo "$script")" ;;
+esac
+while [ -L "$script" ]; do
+	target="$(readlink "$script")"
+	case "$target" in
+	/*) script="$target" ;;
+	*) script="$(dirname "$script")/$target" ;;
+	esac
+done
+DIR="$(cd -P "$(dirname "$script")" && pwd)"
+
+OS="$(uname -s)"
+ARCH="$(uname -m)"
+
+case "$OS:$ARCH" in
+Darwin:arm64 | Darwin:aarch64) BIN="$DIR/consult-llm-mcp-darwin-arm64" ;;
+Darwin:x86_64) BIN="$DIR/consult-llm-mcp-darwin-x64" ;;
+Linux:x86_64 | Linux:amd64) BIN="$DIR/consult-llm-mcp-linux-x64" ;;
+Linux:aarch64 | Linux:arm64) BIN="$DIR/consult-llm-mcp-linux-arm64" ;;
+*)
+	echo "consult-llm-mcp: unsupported platform: $OS $ARCH" >&2
+	exit 1
+	;;
+esac
+
+# npm may strip executable bits on sidecar binaries during extraction
+[ -x "$BIN" ] || chmod +x "$BIN"
+exec "$BIN" "$@"

package/dist/cli.d.ts

@@ -1,2 +0,0 @@
-#!/usr/bin/env node
-export {};

package/dist/cli.js

@@ -1,39 +0,0 @@
-#!/usr/bin/env node
-import { getClientForModel } from './index.js';
-async function main() {
-    const args = process.argv.slice(2);
-    if (args.length === 0) {
-        console.error('Usage: llmtool <prompt>');
-        console.error('Environment variables:');
-        console.error('  OPENAI_API_KEY - Required for GPT models');
-        console.error('  GEMINI_API_KEY - Required for Gemini models');
-        console.error('  MODEL - Model to use (default: gpt-3.5-turbo)');
-        process.exit(1);
-    }
-    const prompt = args.join(' ');
-    const model = process.env.MODEL || 'gpt-3.5-turbo';
-    try {
-        const { client } = getClientForModel(model);
-        const completion = await client.chat.completions.create({
-            model,
-            messages: [{ role: 'user', content: prompt }],
-            temperature: 0.7,
-        });
-        const response = completion.choices[0]?.message?.content;
-        if (response) {
-            console.log(response);
-        }
-        else {
-            console.error('No response from the model');
-            process.exit(1);
-        }
-    }
-    catch (error) {
-        console.error('Error:', error instanceof Error ? error.message : String(error));
-        process.exit(1);
-    }
-}
-main().catch((error) => {
-    console.error('Fatal error:', error);
-    process.exit(1);
-});

package/dist/clipboard.d.ts

@@ -1 +0,0 @@
-export declare function copyToClipboard(text: string): Promise<void>;

package/dist/clipboard.js

@@ -1,10 +0,0 @@
-import clipboard from 'clipboardy';
-export async function copyToClipboard(text) {
-    try {
-        await clipboard.write(text);
-    }
-    catch (error) {
-        const message = error instanceof Error ? error.message : 'Unknown clipboard error';
-        throw new Error(`Failed to copy prompt to clipboard: ${message}`);
-    }
-}

package/dist/clipboard.test.d.ts

@@ -1 +0,0 @@
-export {};

package/dist/clipboard.test.js

@@ -1,21 +0,0 @@
-import { describe, it, expect, vi, afterEach } from 'vitest';
-import { copyToClipboard } from './clipboard.js';
-const writeMock = vi.hoisted(() => vi.fn());
-vi.mock('clipboardy', () => ({
-    default: {
-        write: writeMock,
-    },
-}));
-afterEach(() => {
-    writeMock.mockReset();
-});
-describe('copyToClipboard', () => {
-    it('writes text to clipboard', async () => {
-        await copyToClipboard('hello world');
-        expect(writeMock).toHaveBeenCalledWith('hello world');
-    });
-    it('wraps clipboard errors with a descriptive message', async () => {
-        writeMock.mockRejectedValueOnce(new Error('OS error'));
-        await expect(copyToClipboard('x')).rejects.toThrow('Failed to copy prompt to clipboard: OS error');
-    });
-});

package/dist/config.d.ts

@@ -1,50 +0,0 @@
-import { z } from 'zod/v4';
-export interface ProviderAvailability {
-    geminiApiKey?: string;
-    geminiBackend: string;
-    openaiApiKey?: string;
-    openaiBackend: string;
-    deepseekApiKey?: string;
-}
-export declare function filterByAvailability(models: string[], providers: ProviderAvailability): string[];
-/** Build the final model catalog from built-in + extra + allowlist filtering. */
-export declare function buildModelCatalog(builtinModels: readonly string[], extraModelsRaw?: string, allowedModelsRaw?: string): string[];
-export declare const SupportedChatModel: z.ZodEnum<{
-    [x: string]: string;
-}>;
-export type SupportedChatModel = z.infer<typeof SupportedChatModel>;
-export declare const fallbackModel: string;
-declare const Config: z.ZodObject<{
-    openaiApiKey: z.ZodOptional<z.ZodString>;
-    geminiApiKey: z.ZodOptional<z.ZodString>;
-    deepseekApiKey: z.ZodOptional<z.ZodString>;
-    defaultModel: z.ZodOptional<z.ZodEnum<{
-        [x: string]: string;
-    }>>;
-    geminiBackend: z.ZodDefault<z.ZodEnum<{
-        api: "api";
-        "gemini-cli": "gemini-cli";
-        "cursor-cli": "cursor-cli";
-    }>>;
-    openaiBackend: z.ZodDefault<z.ZodEnum<{
-        api: "api";
-        "cursor-cli": "cursor-cli";
-        "codex-cli": "codex-cli";
-    }>>;
-    codexReasoningEffort: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        minimal: "minimal";
-        low: "low";
-        medium: "medium";
-        high: "high";
-        xhigh: "xhigh";
-    }>>;
-    systemPromptPath: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>;
-type ParsedConfig = z.infer<typeof Config>;
-export type Config = ParsedConfig & {
-    allowedModels: string[];
-};
-export declare function migrateBackendEnv(newVar: string | undefined, oldVar: string | undefined, providerCliValue: string, legacyName: string, newName: string): string | undefined;
-export declare const config: Config;
-export {};