# WM HOST OVERLAY — full-file copy of common/.config/aichat/config.yaml plus the `wm` client
# (this box's own llama.cpp router at 127.0.0.1:11343, see wm/.config/llamacpp/README.md) and the
# default model flipped to the duskadiy client: wm is not on the home LAN, so the `local` endpoint
# (192.168.0.204) is unreachable here (kept below for VPN use). Stowed with --override so it
# shadows the common file on wm only. When the common config changes, mirror the change here.
# see https://github.com/sigoden/aichat/blob/main/config.example.yaml
keybindings: vi
editor: nvim
model: duskadiy:Qwen3.6-35B-A3B-Thinking

# Sessions: persist REPL sessions and keep more history before summarizing.
# The default compress_threshold (4000) summarizes far too early for 24k+ windows.
save_session: true
compress_threshold: 16000

# REPL prompts show live context usage (needs max_input_tokens, set per model below)
left_prompt:  '{color.green}{?session {session}{?role /}}{role}{color.cyan}{?rag @{rag}}{color.reset}> '
right_prompt: '{color.purple}{?session {consume_tokens}/{max_input_tokens} }{color.reset}'

# NOTE: temperature/top_p are intentionally unset — the LAN server applies tuned
# per-model sampling via --jinja (e.g. GLM 0.6/0.95); a global value would clobber it.

# max_input_tokens = real ctx-size (from the router) minus output headroom.
clients:
# This box's own GTX 1080 Ti server (`llamaserver` abbr → 127.0.0.1:11343; presets + benched
# speeds in wm/.config/llamacpp/README.md). Model id = config.ini section name.
- type: openai-compatible
  name: wm
  api_base: http://127.0.0.1:11343/v1
  models:
  - name: Qwen3.5-9B-Q4_K_M
    max_input_tokens: 14000            # ctx 16384 · 33 t/s — daily driver, instant answers
    supports_vision: true
  - name: Qwen3.5-9B-Thinking
    max_input_tokens: 14000            # ctx 16384 · same file, reasoning on (budget 8k)
    supports_vision: true
  - name: Qwen3.5-9B-MTP
    max_input_tokens: 7000             # ctx 8192 · 57 t/s (MTP, +0.75 GB VRAM) — loads on a QUIET desktop only
    supports_vision: true
  - name: gemma-4-E4B-it-UD-Q4_K_XL
    max_input_tokens: 30000            # ctx 32768 · 85 t/s (MTP) — fastest; small vision model, long docs
    supports_vision: true
  - name: Mellum2.1-12B-A2.5B-Thinking-IQ4_XS
    max_input_tokens: 14000            # ctx 16384 · 63 t/s (87 quiet) — fast reasoning coder, prefill > 1k t/s
  - name: Ornith-1.5-35B-A3B-IQ4_XS
    max_input_tokens: 14000            # ctx 16384 · ~10 t/s (MTP, 47% acc) — agentic coder, thinking
  - name: Ornith-1.5-35B-A3B-instant
    max_input_tokens: 14000            # ctx 16384 · 14–21 t/s busy / 31 quiet (MTP) — same file, thinking off: agent loops
  - name: gpt-oss-20b
    max_input_tokens: 30000            # ctx 32768 · 17 t/s — reasoning + tools (TTFT = thinking)
  - name: gpt-oss-20b-low
    max_input_tokens: 30000            # ctx 32768 · reasoning effort low, snappy
# LAN access (fast; only reachable on the home network)
- type: openai-compatible
  name: local
  api_base: http://192.168.0.204:11343/v1
  models: &lan_models
  # Speeds are benched decode t/s (clean-night sweeps) — see fl/.config/llamacpp/README.md roster.
  - name: Qwen3-Coder-30B-Instruct-UD-Q3_K_XL
    max_input_tokens: 30000            # ctx 32768 · ~30 t/s — main agent coder
  - name: Qwen3-Coder-Next-UD-IQ3_XXS
    max_input_tokens: 128000           # ctx 131072 · ~16 t/s — long sessions (128k)
  - name: Qwen3.6-35B-A3B-MTP-UD-IQ3_XXS
    max_input_tokens: 22000            # ctx 24576 · ~36 t/s — daily driver
    supports_vision: true
  - name: Qwen3.6-35B-A3B-Thinking
    max_input_tokens: 22000            # ctx 24576 · ~39 t/s — hard problems
    supports_vision: true
  - name: Qwen3.5-9B-UD-Q6_K_XL
    max_input_tokens: 30000            # ctx 32768 · ~32 t/s — quick tasks
    supports_vision: true
  - name: Qwen3.8-27B-UD-IQ3_XXS
    max_input_tokens: 22000            # ctx 24576 · ~15 t/s — hybrid reasoner, quiet desktop only
    supports_vision: true
  - name: gemma-4-26B-A4B-it-UD-IQ4_XS
    max_input_tokens: 22000            # ctx 24576 · ~34 t/s — quality generalist, best vision
    supports_vision: true
  - name: gemma-4-E4B-it-UD-Q8_K_XL
    max_input_tokens: 62000            # ctx 65536 · ~57 t/s — fast generalist, long docs
    supports_vision: true
  - name: GLM-4.7-Flash-UD-Q4_K_XL
    max_input_tokens: 22000            # ctx 24576 · ~21 t/s — quality coder
  - name: GLM-4.7-Flash-UD-Q3_K_XL
    max_input_tokens: 22000            # ctx 24576 · ~28 t/s — fast coder (3-bit)
  - name: Ornith-1.5-35B-A3B-IQ3_XXS
    max_input_tokens: 30000            # ctx 32768 · ~32 t/s — agentic coder, thinking
    supports_vision: true
  - name: gpt-oss-20b
    max_input_tokens: 62000            # ctx 65536 · ~38 t/s — fast reasoning + tools
  - name: gpt-oss-20b-low
    max_input_tokens: 62000            # ctx 65536 · ~37 t/s — snappy answers
# Remote access via duskadiy.com (reachable from anywhere; requires an API key —
# $DUSKADIY_API_KEY, from the gitignored conf.d/secrets.fish). The LAN endpoint above is keyless.
# Same server, same model ids — reuse the list above via a YAML anchor.
- type: openai-compatible
  name: duskadiy
  api_base: https://llm.duskadiy.com/api/v1
  models: *lan_models