1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104

# Agent creator — an mm agent that creates mm agents.
#
# Give it a model served by llmman and a job; it reads the model's facts
# (skills/create-agent/model_facts.py: size, context window, sampling in the
# GGUF, the chat template's reasoning switches and tool-call format), asks
# what it is missing with ask_user, writes a commented agent.yaml tuned to
# that model, and checks that the file loads and answers. The procedure is
# the `create-agent` skill (skills/create-agent/SKILL.md).
#
# Model: Qwen3.8 27B (17.4 GB, dense, reasoning on). Chosen because it is the
# one that followed every step in the measurements of demos/qwen-style
# (14/14 on seven Go tasks: bash, files, a git workflow, faithful reports) —
# this job is a procedure of several steps with bash, a file to write and a
# question form. It is slow-ish on a laptop; creating an agent is rare.
#
# Run it FROM THE REPOSITORY ROOT (the skill and its script are found there):
#
#   llmman pull hf.co/unsloth/Qwen3.8-27B-GGUF:Q4_K_M
#   llmman serve
#   ./mm -config agent.creator.yaml
#   > Create an agent that translates French to English with hf.co/unsloth/gemma-4-E4B-it-GGUF:Q4_K_M
#
# Every key is optional; reference: docs/en/reference/configuration.md.

# ─── Model and server ────────────────────────────────────────────────────────

provider: llmman
model: hf.co/unsloth/Qwen3.8-27B-GGUF:Q4_K_M
baseUrl: http://127.0.0.1:17434/v1
fallback: http://host.docker.internal:17434/v1
contextWindow: 0

sampling:
  # Qwen's settings for a thinking model on precise work (the GGUF says 1.0).
  temperature: 0.6
  top_p: 0.95
  top_k: 20
  min_p: 0.0
  # The reasoning plus a whole agent.yaml in one write_file.
  max_tokens: 16384
  parallel_tool_calls: false
  # Qwen3.8's template: enable_thinking (default on) and reasoning_effort
  # low | medium | xhigh — any other value makes every request fail.
  chat_template_kwargs:
    enable_thinking: true
    reasoning_effort: medium

showThinking: true

# ─── Timeouts ────────────────────────────────────────────────────────────────

# Loading ~17 GB on the first question; the check in step 5 starts another
# mm, which may load the target model too.
firstTokenTimeout: 5m
watchdogTimeout: 5m
bashTimeout: 5m

# ─── Agent loop ──────────────────────────────────────────────────────────────

# The questions, the facts, the file, the check, a fix or two.
maxTurns: 20
maxOutput: 20000
displayCommands: true

# ─── Tools ───────────────────────────────────────────────────────────────────

# Asks before running the script, writing the file and running the check:
# [a] once, [s] this tool for the session, [r] reject, [n] never.
# AGENT_YOLO=true for unattended runs.
yolo: false

bashTool: true          # model_facts.py, the check
editTools: true         # writes and fixes agent.yaml
askUserTool: true       # the model, the job, the constraints
planTool: false
fetchContentTool: false # set true to let it read a model card on the web

# The repository's skills directory: the create-agent skill (and the demo
# skills that live beside it).
skillsDir: skills

agentsMd:
  enabled: false

acp:
  name: agent-creator
  title: Agent creator

system: |
  You create mm agent configurations: an agent.yaml tuned to one model served
  by llmman and to the job the user describes.

  For any request to create, tune or review an agent for a model, call
  read_skill with name "create-agent" FIRST, then follow its steps in order:
  ask what is missing (one ask_user call), run model_facts.py, choose the
  settings from the skill's table, write the file, check it with "$MM_BIN",
  report.

  RULES
  - Facts about a model come from model_facts.py, never from memory. If the
    script cannot reach llmman, say so and stop.
  - Never invent a chat template variable or a value the report does not list.
  - Write one file (and its directory); never change other files.
  - Answer in the language of the user's message.