DmitrL-dev · Oct 12, 2026 · Oct 11, 2026 · Oct 12, 2026 · Oct 12, 2026 · Oct 12, 2026
diff --git a/.env.example b/.env.example
@@ -1,89 +1,161 @@
 # 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.
-DEEPSEEK_ENABLED=0
+
+# --- 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
+
+# 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).
-# 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.
 # This is pacing, not a guarantee of acceptable use or website quota.
 PROVIDER_MIN_INTERVAL=30
-# 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
-# Accepted replies per saved stateless conversation; integer 1..1000.
-# Raising preserves existing history and one-use continuation protection.
-DEEPSEEK_MAX_CONVERSATION_REPLIES=20
-# Optional Qwen Chat models: run `python -m qwen.auth`, then set QWEN_ENABLED=1.
-# Qwen uses a separate profile and session under session/qwen/.
-QWEN_ENABLED=0
-# Optional: before a new request, renew a locally expired Qwen access token
-# once using its existing website refresh cookie. No browser/background work,
-# no replay after refusal; failures persist a pause. See docs/HARNESS.en.md.
-QWEN_RENEW_SESSION=0
-# QWEN_PROFILE_DIR=/absolute/path/to/a/dedicated/qwen/profile
 
-# Experimental signed-in browser tabs (Safari Userscripts or a compatible manager).
-# Read browser/README.md before enabling. Each uses the web UI's selected model.
-BROWSER_BRIDGE_ENABLED=0
-GROK_ENABLED=0
-MISTRAL_ENABLED=0
-KIMI_ENABLED=0
-# Optional direct Kimi Instant transport; browser remains necessary for manual
-# initial session capture only. See docs/KIMI.md for qualification and setup.
-# Session renewal is a separate opt-in and never follows a provider refusal.
-KIMI_TRANSPORT=browser
-KIMI_RENEW_SESSION=0
-GLM_ENABLED=0
-# WEB_PROVIDER_TIMEOUT=180
+# 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
 
-# Gemini uses the official Antigravity CLI, with tools disabled.
-# Sign in with `agy`; obtain real account model slugs with `agy models`.
-GEMINI_ENABLED=0
-GEMINI_MODELS={}
-# 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.
-
-# --- Optional ---
-# DEEPSEEK_PROFILE_DIR=session/profile   # reuse an existing signed-in Chrome profile
-# HOST=127.0.0.1
-# PORT=8000
 # RATE_LIMIT_PER_MINUTE=30               # per-client-IP limit on /v1 endpoints
 
-# Sessions are loaded from cache by default. The server never logs in, refreshes a
-# browser session in the background, or replays a rejected request.
-# Missing/expired sessions return 401 login_required, or a terminal SSE error
-# after streaming headers are sent. Run the auth CLI manually.
-# Restrictions, quota and uncertain upstream failures pause that provider in
-# session/provider-pauses.json (HTTP 403; SSE error if headers are already sent).
-# Inspect: python -m providers.access status
-# After confirming normal website access manually:
-#   python -m providers.access resume qwen
-# Existing SERVER_INTERACTIVE_LOGIN / SESSION_REFRESH_ENABLED settings no longer
-# enable automatic browser capture in the server.
-
-# --- Tool-call truncation handling ---
+# --- TOOL / TOOLCALL ---
+# 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).
+# 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
+
 # DeepSeek's web output limit can cut a large tool call off mid-JSON. When that
 # happens it stops without returning partial calls. Continuations are off by
 # default. An explicit opt-in (>0) only continues a provider-confirmed output
 # limit, never an otherwise finished reply with a missing Markdown fence.
 TOOLCALL_MAX_CONTINUATIONS=0
+
 # Optional: one format correction after a normally completed invalid DeepSeek
 # tool reply, in the same conversation. Adds at most one paced request. Never
 # retries provider refusals, output limits, or uncertain dispatches. Default off.
 TOOLCALL_FORMAT_REPAIR=0
 
 # Dump the raw model reply to the log whenever a tool call can't be parsed.
-# DEBUG_TOOLCALLS=1
+# 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.
 # 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
+
diff --git a/.gitignore b/.gitignore
@@ -3,10 +3,12 @@ __pycache__/
 *.pyc
 .env
 logs/
+.tmp/
 
 # Session artifacts — contain session tokens + cookies. NEVER commit.
 session.json
 session
 *.env.*
 !.env.example
-CLAUDE.md
+CLAUDE.md
+
diff --git a/README.en.md b/README.en.md
@@ -358,8 +358,9 @@ not resume a stopped goal, replay an old reply, or reset history or counters.
 | Missing group or model | Check `.env`, the running server and `/v1/models`, then reinstall the catalog |
 
 DeepSeek defaults are **12/60 generations, 3/6 new chats and 40/200 HTTP requests
-per hour/day**, with 20 accepted replies per conversation and a 30-second
-interval. `DEEPSEEK_MAX_*` settings belong to the bridge process environment.
+per hour/day**, with a shared 20 accepted replies per conversation
+(`CONVERSATION_MAX_REPLIES`) and a 30-second interval. `DEEPSEEK_MAX_*` budget
+ceilings and `CONVERSATION_MAX_REPLIES` belong to the bridge process environment.
 Raising a local ceiling does not raise the website's allowance.
 [Configuration and error guide](docs/HARNESS.en.md#local-limits-and-errors).
 
@@ -388,14 +389,17 @@ Provider documentation retains detailed qualification results.
 
 | Provider / model ID | Browser during requests | Status |
 | --- | --- | --- |
-| DeepSeek · `deepseek-chat`, `deepseek-expert` | Not needed after manual sign-in | Small Harness tasks passed; repeated account restrictions occurred. No automatic renewal |
-| Qwen · `qwen3.8-omni-flash`, `qwen3.8-max` | Not needed after sign-in | Tested in Harness; Max reached a daily limit. [Session](docs/HARNESS.en.md#qwen-session) |
+| DeepSeek · `deepseek-chat`, `deepseek-expert` | Not needed after manual sign-in | Browserless HTTP; small Harness tasks passed; repeated account restrictions occurred. No automatic renewal |
+| Qwen · HTTP via `QWEN_MODELS` / `qwen-web` | HTTP: not needed after sign-in; `qwen-web`: connected tab | HTTP Flash/Max and the browser adapter verified; Max hit a daily limit. `qwen3.7-plus` is catalogued without a separate live check. [Session](docs/HARNESS.en.md#qwen-session) |
 | Kimi · `kimi-instant` | Not needed after capture | Experimental HTTP: reply, renewal and Harness `read` passed |
-| Kimi · `kimi-k3`, `kimi-k2.8` | Not needed after capture | Implemented, not qualified end to end: K3 refused free-user capacity; K2.8 was not attempted afterward. [Details](docs/KIMI.md) |
-| GLM · `glm-web` | Connected tab required | Previously tested with GLM-5.3-Flash; browserless access remains incomplete. [Blockers](docs/GLM.md) |
-| Kimi · `kimi-web` | Connected tab required | Older tab adapter; model selected on the website |
-| Mistral / Grok | Tab required | Experimental: Mistral continuation returned `429`; Grok capture did not complete |
-| Gemini / Antigravity | Official CLI | Request rejected by region in our test; working access not established |
+| Kimi · `kimi-k3`, `kimi-k2.8` | Not needed after capture | HTTP implemented, not qualified end to end: K3 refused free-user capacity; K2.8 was not attempted afterward. [Details](docs/KIMI.md) |
+| Kimi · `kimi-web` | Connected tab required | Browser adapter verified; large composer pastes (~4KB+, system+tools) become a TXT attachment on the site — confirmed. See `browser/README.md` and `.env.example` |
+| GLM · `glm-web` | Connected tab required | Browser adapter verified; browserless access remains incomplete. [Blockers](docs/GLM.md) |
+| Mistral · `mistral-web` | Connected tab required | Browser adapter verified — see `browser/README.md` |
+| Grok · `grok-web` | Connected tab required | Browser adapter verified — see `browser/README.md` |
+| ChatGPT · `chatgpt-web` | Connected tab required | Browser adapter verified — see `browser/README.md` |
+| Claude · `claude-web` | Connected tab required | Browser adapter verified (ProseMirror + `/api/*`) — see `browser/README.md` |
+| Gemini · `gemini-web` / Antigravity CLI | `gemini-web`: connected tab; CLI: `agy` | Browser adapter verified. CLI was rejected by region in our test |
 
 All connections start disabled. Website versions, quotas and availability may
 change. Chinese providers also have limits; the exact Qwen Max daily allowance
@@ -417,7 +421,8 @@ not clear account restrictions or quotas. Inspect pauses with
 `python -m providers.access status`.
 [Resume rules](docs/ACCOUNT_SAFETY.md).
 
-The API binds to `127.0.0.1`; `session/` and `.env` are excluded from Git.
+The API binds to `127.0.0.1` by default (local only). Set `HOST=0.0.0.0` for
+network or Docker access. `session/` and `.env` are excluded from Git.
 Do not publish them, personalized Userscripts or budget credentials.
 Supported: text and emulated tool calls. Images, audio and video are not uploaded,
 even for models named Omni.
@@ -445,6 +450,21 @@ CI covers Python 3.9/3.12, PoW WASM, shell/JavaScript and isolated browser fixtu
 Locally: `python -m unittest discover -s tests -t . -v`.
 [Real Harness CLI fixtures with mocked providers](docs/HARNESS.en.md#offline-verification).
 
+### Browser exchange safety
+
+DOM fallback is disabled: an old visible answer cannot complete a new request.
+Claude requires a network terminal marker; truncated, partial or malformed answers are
+not reported as complete. Uncertain dispatched outcomes retain pending state
+and a safety pause across restart. There is no automatic retry.
+Verified offline without live accounts. Re-export and update the userscripts
+to apply the browser changes (controller 0.3.34).
+
+Mistral accepts only a request-bound stream with a terminal trailer.
+Harness `--without-browser` excludes browser models and their aliases.
+The browser controller waits for text insertion, preserves user edits and resumes
+only explicitly bound conversations. These cases passed 28 isolated browser
+fixtures without accounts.
+
 ### Model identifiers
 
 `GET /v1/models` lists bare IDs and `provider/model` aliases, for example
@@ -480,3 +500,8 @@ will complete.
 [Tsuev/opencode-deepseek](https://github.com/Tsuev/opencode-deepseek) and
 [sums001/Deepseek-API](https://github.com/sums001/Deepseek-API).
 Use accounts under each service's terms. [License](LICENSE).
+
+Browser adapters (userscripts, tab protocol, ChatGPT / Claude / Gemini / Qwen /
+Kimi / GLM / Mistral / Grok providers and related checks) are reworked in the
+[SimolSem/opencode-deepseek](https://github.com/SimolSem/opencode-deepseek) fork
+and contributed here via pull requests.