Claude Code now talks to the raw /anthropic path of the Eltex proxy,
which forwards requests byte-for-byte to the Z.ai Anthropic endpoint,
the same one ai-glm uses directly:
- images work through Z.ai's server-side analyze_image (Read png ->
server_tool_use analyze_image -> correct answer, same as ai-glm);
the regular /v1/messages route converts to the OpenAI endpoint,
where Z.ai rejects images with 400;
- effort reaches Z.ai: low has no thinking, then medium < high < max;
- prompt cache: ~99% of input is a cache read from the second turn,
same as ai-glm.
The key check and model pick still use the proxy root /v1/models.
The model is glm-5.3 in every picker slot with [1m] for the 1M
window; the anth-glm-5.3 / zai-glm-5.3 / flash names are gone.
Effort stays max by default via CLI --effort.
Checked on 20.10 on llm-t2.eltex.loc:4000; admins are moving that
config to llm.eltex.loc:4000, which stays the default URL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ai-qwen runs Claude Code through Alibaba Cloud Model Studio's
Anthropic-compatible Token Plan endpoint
(token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic). The shared
key is built into the generated script, so it starts right away without
asking for a key.
qwen3.8-max (the smartest model the key has) is in opus/sonnet and for
subagents, --model opus on each start; qwen3.8-flash is in haiku for
background tasks. Measured on 20.10: both models accept up to 983616
input tokens (CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000, autocompact fires
before the API limit) and see images. Qwen validates
output_config.effort (bogus -> 400), Max shows no trend from low to
max, Flash thinks ~2x longer on max; the launcher starts on --effort max
like ai-mimo, EFFORT_MAPPING.md records the measurements.
setup.sh status, statusline brand color, README, QUICK_START and
AGENTS.md list the new launcher.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Admins now give every key three models. ai-eltex maps them to the
picker slots:
- opus (default): anth-glm-5.3 - glm-5.3 on the Z.ai Anthropic
endpoint: images via analyze_image, effort works, 1M window;
- sonnet: zai-glm-5.3 - glm-5.3 on the OpenAI endpoint, no images or
effort, 200K. Two keys get it as zai-glm-5.3-2-2 (zai-glm-5.3 is 403
for them), so the name is matched exactly or by prefix;
- haiku: zai-glm-5.3-flash - fast, sees images itself, no effort,
200K; Claude Code also uses this slot for background tasks.
A missing zai/flash model falls back to the default one.
Model name is anth-glm-5.3 again: with anth-glm-5.3-5 the auto-logout
would drop every saved key, since no key has that name now.
Context window is per model instead of a global
CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000: [1m] only on anth-glm-5.3
(Claude Code strips the suffix before the API call), 200K for the
others, otherwise a long flash/zai session would hit "Prompt exceeds
max length" instead of auto-compacting.
Checked on all 6 pilot keys (19.10): default model reads a png via
Read + analyze_image; sonnet/haiku answer with 200K windows; effort
low on anth-glm-5.3 = 68 output tokens (max ~3500).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Z.ai maps xhigh to native high, so this trims the thinking budget for
everyday ai-glm launches; AI_EFFORT=max still forces the old behavior.
Co-Authored-By: Claude Code <noreply@anthropic.com>
ai-mimo runs Claude Code through Xiaomi's Anthropic-compatible API
(Token Plan or pay-as-you-go). The key's endpoint is found by probing
token-plan-sgp, token-plan-ams and api.xiaomimimo.com (a key only works
in its own region) and is saved to mimo_base_url; MIMO_BASE_URL
overrides. Unreachable endpoints are skipped, token-plan-cn is left out
of the probe since it times out from our network, and a 402 on the
pay-as-you-go endpoint explains that a tp- Token Plan key is needed.
Always the smartest model: mimo-v2.6-pro in every slot and for
subagents, --model opus and --effort max on each start, 128000 output
tokens (Claude Code's ceiling), real 1M context window. Measured on
13.10: MiMo ignores output_config.effort (even "bogus" is accepted, no
length trend from low to max) and treats thinking adaptive like
enabled, so reasoning is always on; EFFORT_MAPPING.md records this.
/logout-mimo is registered as a command and handled by the logout hook
(removes mimo_key and mimo_base_url). setup.sh status, statusline brand
color, README, QUICK_START and AGENTS.md list the new launcher.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ai-eltex-qwen: new launcher for the Eltex proxy on Qwen models
(Coder-Senior, then Flash-Next), own key eltex_qwen_key, config dir and
/logout-eltex-qwen hook branch, 200K window.
Effort: ai-glm and ai-eltex start on max via CLI --effort (AI_EFFORT
overrides), CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000 so long GLM answers are
not cut at 32000. EFFORT_MAPPING notes the proxy passes effort for
anth-glm-5.3 since 09.10.
ai-eltex model pick: no longer falls back to the first /v1/models entry,
which twice picked something unusable (zai-glm-5.3 routed to flash, and
the access-group name ZAI-Test-Group -> 400). Without anth-glm-5.3 the
launcher exits with a clear message; ELTEX_MODEL still overrides.
/logout-eltex: Claude Code answered "Unknown command" and never called
the logout hook (the built-in /logout bypasses hooks too). ai-eltex now
registers commands/logout-eltex.md in its config dir, so the hook
receives the command and deletes eltex_key (checked end to end).
ru-bypass: resolve exception hosts via ahostsv4 first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ai-eltex runs Claude Code through the Eltex LiteLLM proxy
(http://llm.eltex.loc:4000, team GLM subscription). The personal key
from admins is checked via /v1/models and stored in
~/.config/ai-setup/eltex_key; the model is taken from the key's
allowed list (ELTEX_MODEL / ELTEX_BASE_URL override). The proxy
ignores effort/thinking, so no effort forcing. /logout-eltex and
/logout in ai-eltex drop the key; setup.sh status shows it.
ru-bypass: hosts block now has an end marker and only that block is
rewritten, manual entries outside it survive; llm.eltex.loc/llm2 added
since the VPN DNS does not resolve it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Z.ai Anthropic endpoint reads output_config.effort and accepts all 5
Claude Code levels, but GLM 5.3 has only low/high/max natively:
medium -> low, xhigh -> high (measured, not documented by Z.ai).
ai-glm now starts on xhigh instead of max.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
glm-5.3 доступна на api.z.ai (вышла 2026-08-14, 1M контекст): opus/sonnet =
glm-5.3, haiku = glm-5.3[1m], subagent = glm-5.3, health-check ключа тоже
на 5.3. Секция OpenRouter остаётся на z-ai/glm-5.2 — там 5.3 ещё нет.
Co-Authored-By: Claude <noreply@anthropic.com>
Генерация нативного лаунчера ai-cursor для Cursor CLI (agent) с
автоустановкой через cursor.com/install, интеграцией proxychains4 в
VLESS-режиме, обновлением ~/.cursorrules и флагом --force (--yolo).
Очистка старого ai-cursor из ~/.local/bin при нестандартном BIN_DIR.
Добавлены тест генерации и упоминания в EFFORT_MAPPING/QUICK_START/README.
Co-Authored-By: Claude <noreply@anthropic.com>
Новый лаунчер ai-glm запускает Claude Code через официальный Anthropic-совместимый
API Zhipu AI (api.z.ai). Модели: glm-5.2 (opus/sonnet), glm-5.2[1m] (haiku).
Обновлены документация, статус, тесты и EFFORT_MAPPING.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
EFFORT_MAPPING, statusline и _apply_effort переведены на нативное
хранение уровня в settings.json лаунчера. Убран CLAUDE_CODE_EFFORT_LEVEL
(он блокировал /effort внутри сессии) и кэш ~/.cache/ai-setup/effort_*,
который никто не читал и который врал относительно реального уровня.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Update ai-openrouter models in ai-setup.sh output and header
- Fix effort persistence wording (statusbar -> launcher)
- Unify KS_EXCEPTIONS -> KILL_SWITCH_EXCEPTIONS in docs and setup.sh prompt
- Clarify menu item 7 only updates AI launchers
- Add IPv6 note after ks-off.sh
- Scope UFW before.rules insertion to *filter section
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Раньше все ai-* лаунчеры делили один ~/.claude и общий settings.json, из-за
чего кастомная модель (openai/gpt-5.5) из ai-openrouter протекала в пикер
ai-claude. Теперь каждый сторонний провайдер изолирован в своём
CLAUDE_CONFIG_DIR (~/.config/ai-setup/cfg/<launcher>) - свои settings.json и
.claude.json, ноль протечек. ai-claude остаётся на ~/.claude (нативный логин).
Пикеры /model приведены к требуемому виду:
- ai-deepseek: только DeepSeek V4 Pro (opus) и DeepSeek V4 Flash (haiku),
дефолт Pro; через availableModels + ANTHROPIC_DEFAULT_*_MODEL_NAME
- ai-kimi: только Kimi K2.6 (opus)
- ai-claude: только нативные модели Claude
Общие skills и CLAUDE.md шарятся симлинком из ~/.claude.
Persistence effort - гибрид:
- low/medium/high/xhigh живут нативно в settings.json лаунчера, /effort
внутри сессии работает свободно и уровень сохраняется
- max нельзя сохранить в settings.json (session-only), поэтому он
восстанавливается через CLAUDE_CODE_EFFORT_LEVEL; в такой max-сессии
/effort залочен (ограничение Claude Code), выход - AI_EFFORT=<lvl> ai-*
Текущий уровень ловит статусбар в ~/.cache/ai-setup/effort_<launcher>.
Удалён устаревший effort-save-hook (заменён нативным persistence + гибридом),
почищен из ~/.claude/settings.json и осиротевший кэш model_*.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Fix effort-proxy direction: GPT-5.5 natively supports xhigh but not max,
so map max→xhigh (was incorrectly xhigh→max)
- Add EFFORT_MAPPING.md with full mapping table for all providers:
Anthropic, GPT-5.5, DeepSeek V4, Kimi K2.6, Gemini 3.x
- Update proxy docstring and comments
