turbo-editors/configs · agent.mm-dmr-mellum.yaml

⑂ main

agent.mm-dmr-mellum.yaml · 111 lines · 3.8 KBYAML Blame History Raw

🛟 Updated. a3c787d k33g 8h ago	1	# Agent settings.
	2	#
	3	provider: dmr
	4
	5	model: huggingface.co/jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
	6	baseUrl: http://localhost:12434/engines/v1
	7	fallback: http://host.docker.internal:12434/engines/v1
	8
	9	bashTool: true
	10	# The built-in file tools: read_file, write_file, edit_file.
	11	editTools: true
	12
	13	lsp:
	14	servers:
	15	- name: gopls
	16	command: gopls
	17	extensions: [".go"]
	18	- name: rust # rustup component add rust-analyzer
	19	command: rust-analyzer
	20	extensions: [".rs"]
	21
	22	# How many tokens the server actually serves.
	23	# 0 = ask the server (llama-server tells it on /props)
	24	contextWindow: 0
	25
	26	# Max number of characters a tool returns to the model.
	27	# Context safeguard: beyond that, the output is truncated (beginning + end kept).
	28	maxOutput: 16000
	29
	30	# Max number of model <-> tools round trips for a single question.
	31	# Raise it for tasks that need long chains of commands.
	32	maxTurns: 40
	33
	34	# Persistent sessions, per project.
	35	# When enabled, every conversation is saved as one JSON file under `dir`.
	36	# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
	37	sessions:
	38	enabled: true
	39	dir: .mm/sessions
	40
	41	# How many lines of a command's output are echoed to the terminal.
	42	# 0 disables it.
	43	previewLines: 20
	44
	45	displayCommands: true
	46
	47	# The system prompt: what the agent is, and what it is allowed to do.
	48	# This is the knob to play with — it is the shortest path to changing behaviour.
	49	system: \|
	50	Your name is Bob.
	51	You are a coding agent working in a terminal.
	52	You have a "bash" tool to run shell commands.
	53	Use it to explore files, run tests, inspect the repository, etc.
	54	Chain several commands if needed, then answer clearly in English.
	55
	56	FILE EDITING
	57	You have three tools for files: `read_file`, `edit_file` and `write_file`.
	58	They are how a file gets read and changed here: each change is exact,
	59	checked before it is written, and comes back as a diff with line numbers.
	60	bash is for running things — building, testing, listing, searching.
	61
	62	RULES
	63	- Keep everything the file already does, unless the user asked to remove it.
	64	- Touch only the files the request is about. Do not add tests, files or
65	features that were not asked for.
66	- Never run a git command unless the user says git, commit or push.
67	- Never move, rename or delete a file unless the user asked for it.
68	- Then answer in English, in a few lines.
69	- If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
70	to understand the options, then run the command.
71
72	BACKGROUND JOBS
73	Never let a command block the answer. Anything that serves, watches or runs
74	long goes to the background, with BOTH streams redirected and its pid kept:
75
76	nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
77
78
79	# Generation settings (OpenAI API keys). Low temperature for a coding agent:
80	# we want precise and reproducible answers, not creativity.
81	sampling:
82	temperature: 0.0
83	parallel_tool_calls: false
84	top_p: 0.9
85	max_tokens: 4096
86
87	# llama-server processes the whole prompt before the first token;
88	# on a laptop a 32k context can take a while. Raise it if it fires.
89	watchdogTimeout: 30s
90
91	# Context compression.
92	# ON here: the agent knows the window (read from /props), so `threshold` has a number to apply to.
93	context:
94	enabled: true
95
96	# Compress when the history reaches this share of the window.
97	threshold: 75
98
99	# Fallback on the message count, for when the window is unknown.
100	# One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
101	maxMessages: 80
102
103	# How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
104	# The model works on the last few turns, so summarising them would make it redo what it just did.
105	keepLastTurns: 3
106
107	# max_tokens of the summary request.
108	summaryMaxTokens: 1200
109
110	showStats: true
111