Better error handling

.dockerignore

@@ -0,0 +1,50 @@
+# Git
+.git
+.gitignore
+
+# Dependencies (will be installed in container)
+node_modules
+
+# IDE/Editor
+.vscode
+.idea
+*.swp
+*.swo
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Logs
+*.log
+npm-debug.log*
+
+# Test coverage
+coverage
+
+# Environment files (should be provided via docker-compose/env_file)
+.env
+.env.*
+!.env.example
+
+# Data directories (mounted as volumes)
+data/
+*.json
+!package.json
+!package-lock.json
+
+# Build artifacts
+dist
+build
+
+# Documentation (not needed at runtime)
+docs/
+
+# Scripts (not needed at runtime)
+scripts/
+
+# Docker
+Dockerfile
+docker-compose.yml
+docker-entrypoint.sh
+.dockerignore

.env.example

@@ -1,31 +1,83 @@
-PORT=9766
-HOST=0.0.0.0
-DEFAULT_PROVIDER=kimi
-DEFAULT_MODEL=kimi-k2.5
-AUTH_PATH=./auth.json
-# Preferred for current GLM web app (chat.z.ai)
-GLM_TOKEN=
-# Legacy China ChatGLM backend (chatglm.cn); set GLM_BACKEND=chatglm to use this
-GLM_REFRESH_TOKEN=
-GLM_BACKEND=zai
-# Optional: full Cookie header copied from a successful chat.z.ai browser completion request (keeps cdn/acw/ssxmod anti-bot cookies)
-ZAI_COOKIE=
-# Optional: captcha_verify_param copied from Z.ai browser completion request body. It is usually a base64 string, not JSON.
-ZAI_CAPTCHA_VERIFY_PARAM=
-# Optional browser fingerprint overrides for Z.ai request query/headers
-ZAI_USER_AGENT=
-ZAI_ACCEPT_LANGUAGE=en-US
-ZAI_LANGUAGE=ru-RU
-ZAI_LANGUAGES=ru-RU,ru,en-US,en
-ZAI_TIMEZONE=Europe/Samara
-ZAI_TIMEZONE_OFFSET=-240
-KIMI_TOKEN=
-# Set to 1 for local tests without real GLM/Kimi credentials
-MOCK_PROVIDER=0
-# Optional local keys accepted by proxy; empty disables auth check
-API_KEYS=
-# Admin / account-pool behavior
-# 1 = POST/DELETE /admin/accounts persists auth.json by default; use persist:false per request to keep runtime-only
-PERSIST_ADMIN_ACCOUNTS=1
-# Milliseconds to skip an account after provider error before retrying it
-ACCOUNT_COOLDOWN_MS=60000
+PORT=9766
+HOST=0.0.0.0
+DEFAULT_PROVIDER=kimi
+DEFAULT_MODEL=kimi-k2.5
+AUTH_PATH=./auth.json
+# Preferred for current GLM web app (chat.z.ai)
+GLM_TOKEN=
+# Legacy China ChatGLM backend (chatglm.cn); set GLM_BACKEND=chatglm to use this
+GLM_REFRESH_TOKEN=
+GLM_BACKEND=zai
+# Optional: full Cookie header copied from a successful chat.z.ai browser completion request (keeps cdn/acw/ssxmod anti-bot cookies)
+ZAI_COOKIE=
+# Optional: captcha_verify_param copied from Z.ai browser completion request body. It is usually a base64 string, not JSON.
+ZAI_CAPTCHA_VERIFY_PARAM=
+# Optional browser fingerprint overrides for Z.ai request query/headers
+ZAI_USER_AGENT=
+ZAI_ACCEPT_LANGUAGE=en-US
+ZAI_LANGUAGE=ru-RU
+ZAI_LANGUAGES=ru-RU,ru,en-US,en
+ZAI_TIMEZONE=Europe/Samara
+ZAI_TIMEZONE_OFFSET=-240
+# Z.ai browser fallback (Puppeteer/CloakBrowser). See docs/browser-fallback.md
+# for what this is, and for the CPU/network-leak issue that was fixed here.
+# 1: enable per-account (or set browser_fallback:true in auth.json instead)
+# ZAI_BROWSER_FALLBACK=1
+# cloak or puppeteer (default puppeteer)
+# ZAI_BROWSER_ENGINE=puppeteer
+# false to watch it work locally; default true (headless)
+# ZAI_BROWSER_HEADLESS=true
+# How long to wait for a UI completion response before giving up:
+# ZAI_BROWSER_COMPLETION_TIMEOUT=180000
+# How often to log a "still waiting" heartbeat while waiting on the above:
+# ZAI_BROWSER_HEARTBEAT_MS=15000
+# Close the whole browser after this long with zero requests (0 disables
+# auto-close entirely). Relaunches lazily on the next request either way.
+# ZAI_BROWSER_IDLE_CLOSE_MS=600000
+# ZAI_BROWSER_IDLE_CHECK_MS=60000
+# Extra Chromium launch flags (space-separated), appended to the built-in
+# Docker-safe defaults (--disable-gpu, --disable-dev-shm-usage, etc.):
+# ZAI_BROWSER_PUPPETEER_ARGS=
+KIMI_TOKEN=
+# Debugging REASON_CHAT_MESSAGE_NOT_FOUND on Kimi turn 2+: www.kimi.com's
+# Connect-RPC Chat endpoint likely needs session/device identity headers we're
+# not confirmed to be sending correctly. kimi.js best-effort-derives some from
+# the JWT; KIMI_EXTRA_HEADERS (JSON) always overrides those guesses. Capture
+# the real ones with `npm run kimi:dump-headers` (see script header for steps).
+KIMI_LANGUAGE=en-US
+KIMI_TIMEZONE=America/Los_Angeles
+KIMI_EXTRA_HEADERS=
+# Set to 1 to log every outgoing Kimi request headers + chat_id/parent_id
+DEBUG_KIMI=0
+# Set to 1 for local tests without real GLM/Kimi credentials
+MOCK_PROVIDER=0
+# Optional local keys accepted by proxy; empty disables auth check
+API_KEYS=
+# Admin / account-pool behavior
+# 1 = POST/DELETE /admin/accounts persists auth.json by default; use persist:false per request to keep runtime-only
+PERSIST_ADMIN_ACCOUNTS=1
+# Milliseconds to skip an account after provider error before retrying it
+ACCOUNT_COOLDOWN_MS=60000
+# How long an idle conversation keeps its provider-side chat id before the
+# next turn starts a fresh chat, and how many turns one chat id chain may
+# grow to before the same reset happens. See docs/sessions.md.
+SESSION_TTL_MS=7200000
+SESSION_MAX_DEPTH=100
+
+# Model catalog / auto-discover-and-retire. See docs/model-lifecycle.md.
+# Where the discovered/retired model catalog persists across restarts
+# (leave unset for in-memory only, reset on every restart):
+# MODEL_CATALOG_PATH=/app/data/models.json
+# Consecutive model-not-found-shaped failures before a model is auto-
+# retired from GET /v1/models (it can still be requested explicitly;
+# retirement only affects what gets advertised):
+MODEL_RETIRE_AFTER_FAILURES=3
+# Extra regex patterns (JSON array of strings) to recognize a provider
+# error as a model-does-not-exist error, beyond the built-in ones in
+# src/modelCatalog.js:
+MODEL_ERROR_PATTERNS_EXTRA=
+# Kimi only: override which upstream scenario (effectively: which model
+# generation) a given model id maps to. The built-in guesses for
+# kimi-k2.6/k2.7-code/k3 are UNCONFIRMED - verify with
+# node scripts/kimi_dump_curl_headers.js and correct here if needed:
+KIMI_SCENARIOS=

.gitignore

@@ -7,3 +7,4 @@ auth.jso
 auth.*.json
 *.log
 .DS_Store
+/data/zai-browser-profile

.mimocode/commands/docker-test.md

@@ -0,0 +1,84 @@
+---
+description: Build, start, and test the FreeGLMKimiAPI Docker container with a curl request to /v1/chat/completions
+agent: main
+---
+
+# Docker Test Command
+
+Build the Docker image, start the container, and run a test request against the OpenAI-compatible endpoint.
+
+## Usage
+
+```bash
+/mimocode:command docker-test [model] [port] [--no-cache] [--mock]
+```
+
+## Parameters
+
+- `model` (optional): Model to test. Default: `glm-5-search`
+- `port` (optional): Host port mapped to container port 3364. Default: `3364`
+- `--no-cache`: Run `docker-compose build --no-cache`
+- `--mock`: Set `MOCK_PROVIDER=1` in container environment (tests without real credentials)
+
+## Procedure
+
+1. **Build the image**
+   ```bash
+   cd /home/yury/Documents/FreeGLMKimiAPI
+   docker-compose build [--no-cache]
+   ```
+
+2. **Start the container** (detached)
+   ```bash
+   docker-compose up -d
+   ```
+
+3. **Wait for server readiness** (5-10 seconds)
+   ```bash
+   sleep 8
+   ```
+
+4. **Run test request**
+   ```bash
+   curl -X POST http://localhost:${PORT:-3364}/v1/chat/completions \
+     -H "Content-Type: application/json" \
+     -d '{
+       "model": "${MODEL:-glm-5-search}",
+       "messages": [{"role": "user", "content": "hello"}],
+       "stream": false
+     }'
+   ```
+
+5. **Show container logs** (last 30 lines)
+   ```bash
+   docker-compose logs --tail=30
+   ```
+
+6. **Stop container** (optional, for clean state)
+   ```bash
+   docker-compose down
+   ```
+
+## Example Invocations
+
+```bash
+# Default test (glm-5-search on port 3364)
+/mimocode:command docker-test
+
+# Test kimi-k2.5 model
+/mimocode:command docker-test kimi-k2.5
+
+# Test with mock provider (no real API keys needed)
+/mimocode:command docker-test glm-5-search --mock
+
+# Full rebuild without cache
+/mimocode:command docker-test --no-cache
+```
+
+## Notes
+
+- The container exposes port 3364 internally; docker-compose.yml maps it to host port 3364 by default
+- VNC is available at `vnc://localhost:3365` (password: `vampir`) when `ENABLE_VNC=1`
+- Real credentials are loaded from `/home/yury/Documents/FreeGLMKimiAPI/data/auth.json` (mounted in container)
+- `ZAI_BROWSER_ENGINE=cloak` is set in Dockerfile ENV for GLM browser fallback
+- Node 22 and system chromium are pre-installed in the image

.mimocode/commands/local-test.md

@@ -0,0 +1,100 @@
+---
+description: Start FreeGLMKimiAPI server locally and test with a curl request to /v1/chat/completions
+agent: main
+---
+
+# Local Test Command
+
+Start the FreeGLMKimiAPI server locally (with optional mock provider) and run a test request against the OpenAI-compatible endpoint.
+
+## Usage
+
+```bash
+/mimocode:command local-test [model] [port] [--mock] [--bg]
+```
+
+## Parameters
+
+- `model` (optional): Model to test. Default: `glm-5-search`
+- `port` (optional): Server port. Default: `9766` (from config.js)
+- `--mock`: Set `MOCK_PROVIDER=1` to test without real API credentials
+- `--bg`: Run server in background and leave it running after test
+
+## Procedure
+
+1. **Stop any existing server on the port**
+   ```bash
+   pkill -f "node src/server.js" 2>/dev/null || true
+   sleep 1
+   ```
+
+2. **Start the server**
+   ```bash
+   cd /home/yury/Documents/FreeGLMKimiAPI
+
+   # Environment setup
+   export PUPPETEER_SKIP_CHROMIUM_DOWNLOAD=true
+   export PUPPETEER_EXECUTABLE_PATH=/usr/bin/chromium
+   export PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1
+   export CHROME_PATH=/usr/bin/chromium
+   export ZAI_BROWSER_ENGINE=cloak
+
+   if [ "${MOCK:-false}" = "true" ]; then
+     export MOCK_PROVIDER=1
+   fi
+
+   PORT=${PORT:-9766} node src/server.js &
+   SERVER_PID=$!
+   sleep 3
+   ```
+
+3. **Verify server is running**
+   ```bash
+   ps aux | grep "server.js" | grep -v grep
+   ```
+
+4. **Run test request**
+   ```bash
+   curl -X POST http://localhost:${PORT:-9766}/v1/chat/completions \
+     -H "Content-Type: application/json" \
+     -d '{
+       "model": "${MODEL:-glm-5-search}",
+       "messages": [{"role": "user", "content": "hello"}],
+       "stream": false
+     }'
+   ```
+
+5. **Stop server** (unless `--bg` specified)
+   ```bash
+   if [ "${BG:-false}" != "true" ]; then
+     kill $SERVER_PID 2>/dev/null || true
+   fi
+   ```
+
+## Example Invocations
+
+```bash
+# Default test (glm-5-search on port 9766)
+/mimocode:command local-test
+
+# Test kimi-k2.5 model
+/mimocode:command local-test kimi-k2.5
+
+# Test with mock provider (no real API keys needed)
+/mimocode:command local-test glm-5-search --mock
+
+# Start server in background and leave running
+/mimocode:command local-test --bg
+
+# Test specific port
+/mimocode:command local-test glm-5-search 9767
+```
+
+## Notes
+
+- Server runs on port 9766 by default (configured in `src/config.js`)
+- Real credentials loaded from `/home/yury/Documents/FreeGLMKimiAPI/data/auth.json`
+- `ZAI_BROWSER_ENGINE=cloak` required for GLM browser fallback to work
+- System chromium at `/usr/bin/chromium` must be installed (Node 22+)
+- First GLM browser request takes ~100s (chromium download + warmup); subsequent requests faster
+- Kimi provider works end-to-end with real credentials (~2s response)

.mimocode/plans/1786118231225-silent-harbor.md

@@ -0,0 +1,19 @@
+# Plan: Fix Identical Responses from GLM and Kimi Models
+
+## Root Cause
+The model names being sent include a `glmkimi-free/` prefix (e.g., `glmkimi-free/glm-5-search`, `glmkimi-free/kimi-k2.6-search`). The `resolveModel()` function in `src/config.js:67-73` only checks for models starting with `glm` or `kimi` (case-insensitive). With the prefix, neither condition matches, so both requests fall back to `DEFAULT_MODEL` (which is `kimi-k2.5` - a Kimi model). Both requests route to the Kimi provider, producing identical responses.
+
+## Solution
+Modify `resolveModel()` in `src/config.js` to strip known gateway prefixes before matching, OR add the prefix patterns to the matching logic.
+
+## Files to Modify
+- `src/config.js` - Update `resolveModel()` function
+
+## Implementation Approach
+Add prefix stripping logic at the start of `resolveModel()` to handle common gateway prefixes like `glmkimi-free/`. This ensures the actual model ID (`glm-5-search`, `kimi-k2.6-search`) is used for provider resolution.
+
+## Verification
+1. Send request with `model: "glmkimi-free/glm-5-search"` → should route to GLM provider (ZaiProvider since GLM_BACKEND=zai)
+2. Send request with `model: "glmkimi-free/kimi-k2.6-search"` → should route to Kimi provider
+3. Verify responses are different (from different backends)
+4. Test without prefix still works