.env.example

# Copy this file to .env. Sign-in is done by hand in a browser window
# (run `python -m deepseek.auth` or `python -m qwen.auth`), so no credentials are stored here.
# All providers are opt-in. DeepSeek web automation may restrict your account.

# --- Common ---
# Default listen address is 127.0.0.1 (local services only).
# Use 0.0.0.0 for access from the network or Docker containers.
# HOST=0.0.0.0
# PORT=8000

# Transports (how a provider answers requests):
# browser — signed-in site tab + userscripts; model id is <provider>-web; the
#           model is chosen in the site UI. Needs BROWSER_BRIDGE_ENABLED=1, a
#           paired tab, and browser/README.md. Used by Grok, Mistral, GLM,
#           ChatGPT/Claude (browser only); Qwen/Gemini/Kimi when *_TRANSPORT=browser.
# http    — direct website API after a manual auth CLI (no tab during requests).
#           Qwen default; optional for Kimi. DeepSeek has no *_TRANSPORT flag
#           and always uses this style after `python -m deepseek.auth`.
# cli     — Gemini default: official Antigravity CLI (`agy`) + GEMINI_MODELS.
# Enable a provider with <PROVIDER>_ENABLED=1. Dual-transport providers set
# <PROVIDER>_TRANSPORT (qwen: http|browser, gemini: cli|browser, kimi: browser|http).

# Shared signed-in browser tab bridge via a userscript browser extension:
# Tampermonkey for Chrome, Violentmonkey for Chrome, or Userscripts for Safari.
# Read browser/README.md before enabling. Required by browser transports
# (Qwen, Gemini, Grok, Mistral, GLM, ChatGPT, Claude, Kimi).
BROWSER_BRIDGE_ENABLED=0

# Provider error blocking (default 1). When 1, durable pause and unresolved
# pending attempts block future requests until `python -m providers.access resume <provider>`.
# When 0, stored provider pauses/pending blocks are ignored and new failures are
# not latched; this does not disable provider quotas, local budgets, or chat-history
# consistency checks, and may retry uncertain requests / duplicate provider work.
PROVIDER_ACCESS_PAUSE=1

# How long to wait for one provider answer, in seconds (1–18000). Default 600.
# Used by browser tabs, Kimi HTTP and Gemini CLI (after a browser stream starts,
# the same budget applies).
# PROVIDER_RESPONSE_TIMEOUT=18000

# Minimum seconds between starts of attempts/continuations (durable across processes).
# This is pacing, not a guarantee of acceptable use or website quota.
PROVIDER_MIN_INTERVAL=30

# Accepted replies per saved stateless conversation (DeepSeek + web providers,
# including Qwen browser); integer 1..1000. Raising preserves existing history
# and one-use continuation protection.
CONVERSATION_MAX_REPLIES=20

# Repeated initial prompts with exactly identical history produce the same cache key.
# The bridge does not open a new provider chat while a matching link is retained; it may
# reuse that link if still unused, or reject the repeat when already consumed.
# On startup, optionally delete closed/consumed links so a repeated initial prompt
# can start a new chat: 0 (default) keeps them; 1 deletes rows with used=1 only;
# unused links (used=0) remain. The same prompt may then create a new upstream chat,
# and its later continuation may bind to the wrong chat and fail.
# For a complete manual cleanup of all chat-cache tables, run:
#   python -m server.conversation_admin -clear
CONVERSATION_CLEAR_USED_ON_STARTUP=0

# Opt-in: one automatic context migration per failed request, in the SAME
# Harness goal. Only an empty, explicitly cleared context refusal qualifies.
# Private checkpoints in session/context-checkpoints; no counter reset/replay.
# Existing paused requests need separate recovery and are not auto-resumed.
# Also migrate before dispatch when CONVERSATION_MAX_REPLIES is reached.
DEEPSEEK_CONTEXT_ROLLOVER=0

# RATE_LIMIT_PER_MINUTE=30               # per-client-IP limit on /v1 endpoints

# --- TOOL / TOOLCALL ---
# Optional raw per-chat traces (prompts, tool calls/results and replies) in ./llog/.
# Files can contain private project data. Disabled by default.
TOOLCALL_LOG=0

# How OpenAI `tools` are flattened into the provider prompt.
# full  — full JSON Schema + tool_calls rules on every turn.
# light — recommended; full block on the first turn of a provider chat; short
#         name/format reminder on continuations (smaller follow-ups).
# medium — like light, but the continuation reminder also lists each tool's
#         argument shape (field names, nesting, `?` optional; no descriptions), so
#         a model does not guess arguments from memory. Longer follow-up prompts.
# none  — like light on the first turn; no tools block on continuations.
# HTTP transports (DeepSeek, Qwen API, Kimi direct, Gemini CLI): default full.
# Browser tab bridge (grok/mistral/glm/chatgpt/claude/kimi/qwen/gemini): default light.
TOOL_INSTRUCTIONS_HTTP=full
TOOL_INSTRUCTIONS_BROWSER=light

# Requested tool envelope for DeepSeek only: json (default) or dsml (opt-in).
# DSML avoids JSON-escaped shell/file strings; the same complete-batch/schema
# checks apply. This never changes history or authorizes repair/replay.
DEEPSEEK_TOOL_PROTOCOL=json
# Opt-in DeepThink for DeepSeek tool requests that omit thinking. Explicit
# thinking=false still wins. Restart the bridge; no retries or history reset.
DEEPSEEK_TOOL_THINKING=0

# Provider output limit can cut a tool call off mid-JSON. Continuations are off
# by default. An opt-in (>0) continues only when the whole reply starts with an
# open tool_calls fence. Never continues a normal stop that merely omitted a
# Markdown fence.
TOOLCALL_MAX_CONTINUATIONS=0

# Optional: one format correction after a normally completed erroneous tool reply.
# DeepSeek only. Default off.
TOOLCALL_FORMAT_REPAIR=0

# Dump the raw model reply to the log whenever a tool call can't be parsed.
# TOOLCALL_DEBUG=1

# Save the last invalid tool reply privately under session/, without a history request.
# May contain private source or secrets; disabled by default. Do not publish the file.
# Also keeps exact DeepSeek reply ID + request hash for private recovery review.
# With DEEPSEEK_STREAM_DIAGNOSTICS=1, failed tool replies retain the original
# Maximum retained raw captures per directory (1..32, default 4). Never evicts.
DEEPSEEK_STREAM_CAPTURE_FILES=4
# response stream in session/failed-tool-streams/ (4 bounded files, no eviction).
# TOOLCALL_DIAGNOSTICS=1

# --- DeepSeek ---
DEEPSEEK_ENABLED=0
# TRANSPORT
# http only after `python -m deepseek.auth` (no DEEPSEEK_TRANSPORT).
# DeepSeek also spaces HTTP dispatches by up to 3 seconds and enforces durable
# default rolling budgets: 12 generations/hour, 60/day; 3 new chats/hour, 6/day;
# 40 HTTP requests/hour, 200/day. These are local ceilings, not safe quotas.
# Configurable DeepSeek ceilings. Integers 1..2000; day >= hour for each kind.
# Restart to apply. Existing usage is preserved; website quotas are independent.
DEEPSEEK_MAX_GENERATIONS_PER_HOUR=12
DEEPSEEK_MAX_GENERATIONS_PER_DAY=60
DEEPSEEK_MAX_CHATS_PER_HOUR=3
DEEPSEEK_MAX_CHATS_PER_DAY=6
DEEPSEEK_MAX_HTTP_PER_HOUR=40
DEEPSEEK_MAX_HTTP_PER_DAY=200
# DEEPSEEK_PROFILE_DIR=session/profile   # reuse an existing signed-in Chrome profile
# Private failed SSE bytes (may contain private source/response text):
# up to 4 files, 8 MiB decoded bytes each; no automatic deletion or replay.
# DEEPSEEK_STREAM_DIAGNOSTICS=1

# --- Qwen ---
QWEN_ENABLED=0
# TRANSPORT
# http    — `python -m qwen.auth` + QWEN_MODELS (default).
# browser — model qwen-web; needs BROWSER_BRIDGE_ENABLED=1 and a paired tab.
QWEN_TRANSPORT=http
# Public OpenAI id → website wire model (HTTP transport). Default if unset:
# qwen3.8-omni-flash, qwen3.8-max, qwen3.7-plus.
QWEN_MODELS={"qwen3.8-omni-flash":"qwen3.8-omni-flash","qwen3.8-max":"qwen3.8-max","qwen3.7-plus":"qwen3.7-plus"}
# Optional HTTP-only: renew a locally expired access token once via the website
# refresh cookie. No browser/background work, no replay after refusal; failures
# persist a pause. Ignored for browser transport. See docs/HARNESS.en.md.
QWEN_RENEW_SESSION=0
# QWEN_PROFILE_DIR=/absolute/path/to/a/dedicated/qwen/profile

# --- Gemini ---
GEMINI_ENABLED=0
# TRANSPORT
# cli     — sign in with `agy`; put slugs from `agy models` into GEMINI_MODELS (default).
# browser — gemini-web on AI Mode (google.com/search?udm=50) or gemini.google.com/app.
GEMINI_TRANSPORT=cli
# Example catalog current as of 2026-10-11, from `agy models` (account/region may differ).
# GEMINI_MODELS={"gemini-3.8-flash-high":"gemini-3.8-flash-high","gemini-3.8-flash-medium":"gemini-3.8-flash-medium","gemini-3.8-flash-low":"gemini-3.8-flash-low","gemini-3.7-flash-high":"gemini-3.7-flash-high","gemini-3.7-flash-medium":"gemini-3.7-flash-medium","gemini-3.7-flash-low":"gemini-3.7-flash-low","gemini-3.6-flash-high":"gemini-3.6-flash-high","gemini-3.6-flash-medium":"gemini-3.6-flash-medium","gemini-3.6-flash-low":"gemini-3.6-flash-low","gemini-3.1-pro-high":"gemini-3.1-pro-high","gemini-3.1-pro-low":"gemini-3.1-pro-low","gemini-claude-sonnet-4-6":"claude-sonnet-4-6","gemini-claude-opus-4-6-thinking":"claude-opus-4-6-thinking","gemini-gpt-oss-120b-medium":"gpt-oss-120b-medium"}
# ANTIGRAVITY_BIN=agy
# For a broken IPv6 socket only, providers.antigravity_ipv4 wraps the official
# CLI using IPv4 on your existing route, without changing system/VPN settings.

# --- Grok ---
GROK_ENABLED=0
# TRANSPORT
# browser only (no GROK_TRANSPORT). Model: grok-web.
# Needs BROWSER_BRIDGE_ENABLED=1 and a paired tab on grok.com.

# --- Mistral ---
MISTRAL_ENABLED=0
# TRANSPORT
# browser only (no MISTRAL_TRANSPORT). Model: mistral-web.
# Needs BROWSER_BRIDGE_ENABLED=1 and a paired tab on chat.mistral.ai.

# --- GLM ---
GLM_ENABLED=0
# TRANSPORT
# browser only (no GLM_TRANSPORT). Model: glm-web.
# Needs BROWSER_BRIDGE_ENABLED=1 and a paired tab on chat.z.ai.

# --- ChatGPT ---
CHATGPT_ENABLED=0
# TRANSPORT
# browser only (no CHATGPT_TRANSPORT). Model: chatgpt-web.
# Needs BROWSER_BRIDGE_ENABLED=1 and a paired tab on chatgpt.com.

# --- Claude ---
CLAUDE_ENABLED=0
# TRANSPORT
# browser only (no CLAUDE_TRANSPORT). Model: claude-web.
# Needs BROWSER_BRIDGE_ENABLED=1 and a paired signed-in tab on https://claude.ai/;
# select the model in the site UI.

# --- Kimi ---
KIMI_ENABLED=0
# TRANSPORT
# browser — kimi-web (default); large pastes ~4KB+ may become a TXT attachment —
#           keep prompts shorter, or use http. Needs BROWSER_BRIDGE_ENABLED=1.
# http    — Instant after manual session capture; browser still needed once for
#           capture. See docs/KIMI.md. Renewal never follows a refusal.
KIMI_TRANSPORT=browser
KIMI_RENEW_SESSION=0


# --- /v1 API access (required) ---
# Every /v1 request needs Authorization: Bearer <token from session/api-token>.
# python -m server.api_auth path prints the file path (never the token).
# BRIDGE_API_TOKEN_FILE=/absolute/path/to/api-token
# Temporary migration only: admit tokenless loopback, non-browser requests.
# API_AUTH=legacy-loopback
# Non-loopback HOST is refused unless explicitly allowed (token mode only):
# API_ALLOW_PUBLIC_BIND=1
# API_ALLOWED_HOSTS=bridge.lan