Tags

Tags give the ability to mark specific points in history as being important
  • v2.1.0

    v2.1.0 — add predictalot (foundation time-series forecasting)
    
    New service psyb0t/docker-predictalot v0.1.1 at /predictalot/.
    Five univariate quantile forecasters (chronos-2, timesfm-2.5,
    moirai-2, toto-1, sundial-base-128m) behind one wire shape, plus
    weighted ensemble. CPU + CUDA variants, mutually exclusive.
    
    Direct route only (not in LiteLLM providers). MCP enabled by default
    with 7 tools via the /mcp/ aggregator. Bearer-token auth.
    
    Lazy model loading (~1.4GB HF snapshots) with idle-unload after
    PREDICTALOT_MODEL_IDLE_TIMEOUT (default 30m).
    
    tests/test_predictalot.sh: 5 tests gated on PREDICTALOT=1 or
    PREDICTALOT_CUDA=1.
    
    Co-shipped with v2.2.0 (mailbox) on the same commit — shared edits
    to docker-compose, .env.example, Makefile, README, docs.
  • v2.0.0

    v2.0.0 — BREAKING: replace claudebox-zai with pibox-zai (psyb0t/docker-pibox v0.3.1)
    
    Swap z.ai/GLM backend from a second claudebox instance to pibox
    (pi-coding-agent in aicodebox). Service / env var / route / data dir
    renamed. LiteLLM model names: pibox-zai-glm-{4.5-air,4.7,5.1}.
    
    New surface: /files/* CRUD, /mcp opt-in via PIBOX_MCP_MODE, /run
    native agent execution, /healthz (not /health).
    
    Tests: new tests/test_pibox.sh (7 end-to-end), claudebox tests gated
    on CLAUDEBOX=1. Docs fully migrated. CHANGELOG backfilled to match
    all 47 prior git tags 1:1.
    
    Pibox instances are pinned to one Anthropic-compatible backend by
    design (init.d marker on bind-mounted volume) — future siblings
    (pibox-or, pibox-openai) run as separate services.
    
    Migration: rename CLAUDEBOX_ZAI_* env vars to PIBOX_ZAI_*, move
    .data/claudebox-zai → .data/pibox-zai (rm .init-done first), update
    model names claudebox-zai-* → pibox-zai-glm-*, update direct route
    callers /claudebox-zai/* → /pibox-zai/*.
  • v1.6.3

    v1.6.3 — docs: correct Cerebras (5 RPM / 1M TPD on 4 models) and Mistral (limits not published) free-tier claims; cross-verify Groq/OpenRouter/HF/Cohere from official docs
  • v1.6.2

    v1.6.2 — fix nginx prefix-stripping on claudebox/claudebox-zai/stealthy-auto-browse/librechat/searxng routes (variable proxy_pass needs explicit rewrite); drop dead langfuse .gitkeep
  • v1.6.1

    v1.6.1 — claudebox v1.14.0-minimal: OpenAI-compat wrapper hardening (multi-turn workspace fix, SSRF guard, finish_reason mapping, 400 on unsupported fields, reasoning_effort snake_case binding)
  • v1.6.0

    04522c9c · f ·
    v1.6.0 — security + claudebox v1.13.1 + reasoning/vision models + Piper TTS
    
    Security
    - litellm v1.83.5-nightly -> v1.83.14-stable.patch.2: fixes CVE-2026-42208
      (pre-auth SQL injection in API key verification, CVSS 9.3). Stable channel,
      cosign-signed. Post-supply-chain-incident v1.83.x line.
    - redis 7.4.8-alpine -> 7.4.9-alpine: RCE/UAF CVEs disclosed 2026-05-05.
      Applied to both `redis` and `stealthy-auto-browse-redis`.
    
    Claudebox
    - psyb0t/claudebox v1.9.0-minimal -> v1.13.1-minimal (11 versions, mostly
      cron+telegram fixes irrelevant to aigate's CLAUDE_MODE_API=1 usage).
    - v1.11.0 surfaces `opusplan` in /openai/v1/models — added claudebox-opusplan
      provider entry. claudebox-zai unchanged.
    
    New local CUDA LLMs (4)
    - local-ollama-cuda-phi4-reasoning-plus  — Phi-4 Reasoning Plus 14B+RL (~11GB)
    - local-ollama-cuda-deepseek-r1-14b      — DeepSeek-R1-Distill-Qwen-14B (~9GB)
    - local-ollama-cuda-qwen3-30b-a3b        — Qwen3 30B MoE / 3B active (~17GB)
    - local-ollama-cuda-qwen2.5vl-7b         — Qwen2.5-VL 7B vision (~5GB)
    - All four added to ollama-pull CUDA section for auto-download.
    
    New local CPU Piper TTS models (17, 16 languages)
    - Fills the multilingual TTS gap (Kokoro = English/Chinese-leaning,
      Qwen3-TTS needs CUDA and covers 10 langs).
    - Multi-speaker (5): en_US-libritts-high (904), en_GB-vctk-medium (109),
      fr_FR-mls-medium, nl_NL-mls-medium, sv_SE-nst-medium — speaker via voice:"<id>"
    - Single-speaker (12): de_DE-thorsten-high, es_ES-davefx-medium,
      it_IT-paola-medium, pl_PL-darkman-medium, pt_BR-faber-medium,
      ro_RO-mihai-medium, ru_RU-irina-medium, tr_TR-fahrettin-medium,
      uk_UA-ukrainian_tts-medium, vi_VN-vais1000-medium, zh_CN-huayan-medium,
      ar_JO-kareem-medium
    - Served by existing speaches container (lazy-downloaded), no infra changes.
    
    Fallback chains (5 chains in litellm/config/fallbacks.json)
    - phi4-reasoning-plus -> deepseek-r1-14b -> deepseek-r1-8b -> qwen3-30b-a3b ->
      qwen3-8b -> hf-deepseek-r1 -> cerebras-qwen3-235b -> cpu-qwen3-4b
    - deepseek-r1-14b -> phi4-reasoning-plus -> deepseek-r1-8b -> qwen3-30b-a3b ->
      qwen3-8b -> hf-deepseek-r1 -> cerebras-qwen3-235b -> cpu-qwen3-4b
    - deepseek-r1-8b (updated) -> deepseek-r1-14b -> phi4-reasoning-plus ->
      qwen3-8b -> llama3.1-8b -> hf-deepseek-r1 -> cpu-qwen3-4b
    - qwen3-30b-a3b -> qwen3-abliterated-16b -> qwen3-8b -> llama3.1-8b ->
      cerebras-qwen3-235b -> or-qwen3-80b -> mistral-large -> cpu-qwen3-4b
    - qwen2.5vl-7b -> gemma4-e4b -> hf-qwen3-vl-8b -> hf-qwen-vl-72b ->
      groq-llama-4-scout -> hf-llama-4-scout -> openai-gpt-4o
    
    Note: the actual git commit landed on main as "f" (04522c9) before this tag
    was cut; full release notes live in this annotated tag.
    
    No env var changes. No breaking changes.
  • v1.5.0

    v1.5.0 — Tailscale support (L4 TCPForward) + rename internal networks aigate-*
  • v1.4.2

    v1.4.2 — bump claudebox to v1.9.0-minimal
  • v1.4.1

    v1.4.1 — upgrade telethon to telethon-plus v0.2.0
    
    Drop psyb0t/telethon in favour of psyb0t/telethon-plus. Same tool set,
    same MCP endpoint, same env vars. Login command updated in docs.
  • v1.4.0

    v1.4.0 — Telethon Telegram client + claudebox v1.8.0-minimal
    
    Optional TELETHON=1 service via psyb0t/docker-telethon. REST API at
    /telethon/ and MCP server with 15 Telegram tools. Nginx rewrite fix for
    variable proxy_pass. Force-recreate on make run/restart. LLM-driven
    integration test. Claudebox bumped to v1.8.0-minimal.
  • v1.3.4

    v1.3.4 — remove hardcoded counts from docs
  • v1.3.3

    v1.3.3 — fix nginx startup crash on disabled optional services
    
    nginx crashed with "host not found in upstream" when claudebox,
    hybrids3, librechat, searxng, or browser were disabled. Fixed with
    Docker DNS resolver + variable-based proxy_pass for optional upstreams.
    Also forces container recreation on make run/restart.
  • v1.3.2

    v1.3.2 — remove remaining Langfuse references from docs
    
    Cleans up README, services-reference, docker-compose comment, and adds
    v1.3.1 changelog entry. No functional changes.
  • v1.3.1

    v1.3.1 — remove Langfuse v3 (broken S3 event pipeline with HybridS3)
    
    Langfuse v3 requires S3 ListObjectsV2 which fails against HybridS3 due
    to JS AWS SDK v3 signing the canonical URI without the endpoint path
    prefix. Removes all Langfuse services, config, nginx routing, env vars,
    and HybridS3 bucket.
  • v1.3.0

    v1.3.0 — SearXNG meta-search, search_web MCP tool, Langfuse v3 observability
    
    - SearXNG self-hosted meta-search at /searxng/ (SEARXNG=1)
    - search_web MCP tool — any function-calling model can search the web
    - Langfuse v3 LLM observability at /langfuse/ (LANGFUSE=1)
      → requires ClickHouse (added automatically with langfuse profile)
    - Langfuse traces all LiteLLM requests via success/failure callbacks
    - 29 services, 99 models, 21 tools
  • v1.2.0

    v1.2.0 — nuextract-v1.5, all CPU models on CUDA, registration default on
    
    - Add nuextract-v1.5 (structured extraction) to CPU + CUDA
    - All CPU models now available as CUDA variants
    - LibreChat registration enabled by default
    - proxq v0.9.0, rate limit 600r/m, timeout 30m
  • v1.1.1

    v1.1.1 — proxq v0.9.0, raise rate limit, 30m upstream timeout
    
    - proxq v0.9.0 fixes timeout config not applied to HTTP client
    - nginx proxq rate limit 120r/m → 600r/m
    - PROXQ_UPSTREAM_TIMEOUT default 10m → 30m
  • v1.1.0

    v1.1.0 — local model lineup overhaul
    
    CPU:
    - Add gemma4:e2b (multimodal), phi4-mini (reasoning), gemma3:4b (vision)
    - Add qwen3-embedding:0.6b
    - Drop phi3.5, nomic-embed-text
    
    CUDA:
    - Add gemma4:e4b + e2b, deepseek-coder-v2:16b, deepseek-r1:8b
    - Add qwen3-abliterated:16b, gemma4-abliterated:e4b (uncensored)
    - Drop dolphin-mistral:7b, dolphin3
    
    Fallback chains rewritten for all new models. Tests and docs updated.
  • v1.0.1

    v1.0.1 — OS memory reserve, CPU resource-manager-aware scaling
    
    - Reserve 2 GB or 5% RAM (whichever larger) for OS before service allocation
    - CPU local services (ollama, speaches, sdcpp) use max-of-active + idle
      overhead in concurrent RAM calculation, same as CUDA group
    - Add CHANGELOG.md with full release history
  • v1.0.0

    v1.0.0 — per-service CUDA flags replace global CUDA=1
    
    Breaking: CUDA=1 no longer exists. Replace with:
    - OLLAMA_CUDA=1 (GPU inference)
    - SDCPP_CUDA=1 (GPU image generation)
    - SPEACHES_CUDA=1 (GPU STT)
    - QWEN_TTS_CUDA=1 (GPU TTS)
    
    Each CUDA service is now independently toggleable.