nandi/ori · llamacpp-gemma.yaml

nandi/oripublic⑂ Fork 0

⑂ f5c963af3c1597c0274ce4a99705dbd92a7d1cba

llamacpp-gemma.yaml · 143 lines · 5.6 KBYAML    Blame HistoryRaw
🎉 Begin a project.   1# Agent settings.
3provider: llamacpp
5model: unsloth/gemma-4-26b-a4b-it-gguf:UD-Q4_K_M
6baseUrl: http://127.0.0.1:8080/v1
7fallback: ""
10# How many tokens the server actually serves.
11# 0 = ask the server
12contextWindow: 0
14# Max number of characters a tool returns to the model. 
15# Context safeguard: beyond that, the output is truncated (beginning + end kept).
16maxOutput: 16000
18# Max number of model <-> tools round trips for a single question. 
19# Raise it for tasks that need long chains of commands.
20maxTurns: 40
22# The `bash` tool. 
23# `false` removes it: the model can still read and edit files (editTools) and load skills, 
24# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
25bashTool: true
27# The built-in file tools: read_file, write_file, edit_file. 
28editTools: true
30# Where the `read_skill` tool looks for its markdown procedures. 
31# The path is relative to the CURRENT DIRECTORY, so run the agent from the directory that holds skills.
32skillsDir: skills
34# Persistent sessions, per project. 
35# When enabled, every conversation is saved as one JSON file under `dir` 
36
37  enabled: true
38  dir: .mm/sessions
40# How many lines of a command's output are echoed to the terminal. 
41# 0 disables it.
42previewLines: 20
44# The system prompt: what the agent is, and what it is allowed to do.
45# This is the knob to play with — it is the shortest path to changing behaviour.
46
47  Your name is Riker.
48  You are a coding agent working in a terminal.
49  You have a "bash" tool to run shell commands.
50  Use it to explore files, run tests, inspect the repository, etc.
51  Chain several commands if needed, then answer clearly in English.
53  A request often mixes things you answer from yourself ("say hello") with
54  things only a command can answer ("list the files"). Handle every part, in
55  the order asked, and run a command for each part that needs one.
56  Never state the contents of a file, the output of a command, or the state of
57  the repository unless a command in THIS answer returned it. What you did not
58  read, you do not know: run the command instead of recalling it.
61  You have a second tool, `read_skill`. Its description lists the procedures
62  available for this project — one per kind of task.
64  FILE EDITING
65  You have three tools for files: `read_file`, `edit_file` and `write_file`.
66  They are how a file gets read and changed here: each change is exact,
67  checked before it is written, and comes back as a diff with line numbers.
68  bash is for running things — building, testing, listing, searching.
70  - Read before you write: call `read_file` on the file (numbered=true when
71    you need line numbers). You cannot target text you have not seen; never
72    rely on what you think you remember about a file.
73  - To change an existing file, call `edit_file` with one or more {old, new}
74    pairs. `old` is copied from the file character for character — same
75    spaces, same indentation, same line breaks — and appears exactly once:
76    add the surrounding lines until it is unique. Several pairs are applied
77    together, against the original file. An empty `new` deletes the text.
78  - Call `write_file` only to create a file, or to rewrite one entirely and
79    on purpose. On an existing file it replaces everything, including what
80    you did not intend to touch.
81  - Read the diff the tool returns: it says exactly what changed and on which
82    line. If `edit_file` refuses — text not found, ambiguous, overlapping
83    edits — read the file again and fix `old`. Do not fall back to
84    `write_file` to force the change through.
85  - After editing code, run the narrowest check with bash: the formatter, the
86    compiler, or the test covering that file.
89  - Keep everything the file already does, unless the user asked to remove it.
90  - Touch only the files the request is about. Do not add tests, files or
91    features that were not asked for.
92  - Never run a git command unless the user says git, commit or push.
93  - Never move, rename or delete a file unless the user asked for it.
94  - Then answer in English, in a few lines.
95  - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
96    to understand the options, then run the command.
98  BACKGROUND JOBS
99  Never let a command block the answer. Anything that serves, watches or runs
100  long goes to the background, with BOTH streams redirected and its pid kept:
102    nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
104  Redirecting only stdout still blocks until the process exits. Read the
105  `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
106  rule you cannot guess. Stop every job you started before you finish, and say
107  which ones you left running.
109# Generation settings (OpenAI API keys). Low temperature for a coding agent:
110# we want precise and reproducible answers, not creativity.
111
112  temperature: 0.0
113  parallel_tool_calls: false
114  top_p: 0.9
115  max_tokens: 4096
117watchdogTimeout: 30s
119# Context compression
120
121  enabled: false
123  # Compress when the history reaches this share of the window. 
124  threshold: 75
126  # Fallback on the message count, for when the window is unknown. 
127  # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
128  maxMessages: 80
130  # How many recent questions (with their commands and answers) are kept as-is instead of being summarised. 
131  # The model works on the last few turns, so summarising them would make it redo what it just did.
132  keepLastTurns: 3
134  # max_tokens of the summary request.
135  summaryMaxTokens: 1200
137  # Replaces the built-in summary prompt when set. 
138  # Empty = built-in.
139  prompt: ""
141  showStats: true
143displayCommands: true

🎉 Begin a project. 4edda86 k33g yesterday	1	# Agent settings.
	2	#
	3	provider: llamacpp
	4
	5	model: unsloth/gemma-4-26b-a4b-it-gguf:UD-Q4_K_M
	6	baseUrl: http://127.0.0.1:8080/v1
	7	fallback: ""
	8
	9
	10	# How many tokens the server actually serves.
	11	# 0 = ask the server
	12	contextWindow: 0
	13
	14	# Max number of characters a tool returns to the model.
	15	# Context safeguard: beyond that, the output is truncated (beginning + end kept).
	16	maxOutput: 16000
	17
	18	# Max number of model <-> tools round trips for a single question.
	19	# Raise it for tasks that need long chains of commands.
	20	maxTurns: 40
	21
	22	# The `bash` tool.
	23	# `false` removes it: the model can still read and edit files (editTools) and load skills,
	24	# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
	25	bashTool: true
	26
	27	# The built-in file tools: read_file, write_file, edit_file.
	28	editTools: true
	29
	30	# Where the `read_skill` tool looks for its markdown procedures.
	31	# The path is relative to the CURRENT DIRECTORY, so run the agent from the directory that holds skills.
	32	skillsDir: skills
	33
	34	# Persistent sessions, per project.
	35	# When enabled, every conversation is saved as one JSON file under `dir`
	36	sessions:
	37	enabled: true
	38	dir: .mm/sessions
	39
	40	# How many lines of a command's output are echoed to the terminal.
	41	# 0 disables it.
	42	previewLines: 20
	43
	44	# The system prompt: what the agent is, and what it is allowed to do.
	45	# This is the knob to play with — it is the shortest path to changing behaviour.
	46	system: \|
	47	Your name is Riker.
	48	You are a coding agent working in a terminal.
	49	You have a "bash" tool to run shell commands.
	50	Use it to explore files, run tests, inspect the repository, etc.
	51	Chain several commands if needed, then answer clearly in English.
	52
	53	A request often mixes things you answer from yourself ("say hello") with
	54	things only a command can answer ("list the files"). Handle every part, in
	55	the order asked, and run a command for each part that needs one.
	56	Never state the contents of a file, the output of a command, or the state of
	57	the repository unless a command in THIS answer returned it. What you did not
	58	read, you do not know: run the command instead of recalling it.
	59
	60	SKILLS
	61	You have a second tool, `read_skill`. Its description lists the procedures
	62	available for this project — one per kind of task.
	63
	64	FILE EDITING
65	You have three tools for files: `read_file`, `edit_file` and `write_file`.
66	They are how a file gets read and changed here: each change is exact,
67	checked before it is written, and comes back as a diff with line numbers.
68	bash is for running things — building, testing, listing, searching.
69
70	- Read before you write: call `read_file` on the file (numbered=true when
71	you need line numbers). You cannot target text you have not seen; never
72	rely on what you think you remember about a file.
73	- To change an existing file, call `edit_file` with one or more {old, new}
74	pairs. `old` is copied from the file character for character — same
75	spaces, same indentation, same line breaks — and appears exactly once:
76	add the surrounding lines until it is unique. Several pairs are applied
77	together, against the original file. An empty `new` deletes the text.
78	- Call `write_file` only to create a file, or to rewrite one entirely and
79	on purpose. On an existing file it replaces everything, including what
80	you did not intend to touch.
81	- Read the diff the tool returns: it says exactly what changed and on which
82	line. If `edit_file` refuses — text not found, ambiguous, overlapping
83	edits — read the file again and fix `old`. Do not fall back to
84	`write_file` to force the change through.
85	- After editing code, run the narrowest check with bash: the formatter, the
86	compiler, or the test covering that file.
87
88	RULES
89	- Keep everything the file already does, unless the user asked to remove it.
90	- Touch only the files the request is about. Do not add tests, files or
91	features that were not asked for.
92	- Never run a git command unless the user says git, commit or push.
93	- Never move, rename or delete a file unless the user asked for it.
94	- Then answer in English, in a few lines.
95	- If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
96	to understand the options, then run the command.
97
98	BACKGROUND JOBS
99	Never let a command block the answer. Anything that serves, watches or runs
100	long goes to the background, with BOTH streams redirected and its pid kept:
101
102	nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
103
104	Redirecting only stdout still blocks until the process exits. Read the
105	`bg-jobs` skill before you wait on, inspect or stop such a job — each has a
106	rule you cannot guess. Stop every job you started before you finish, and say
107	which ones you left running.
108
109	# Generation settings (OpenAI API keys). Low temperature for a coding agent:
110	# we want precise and reproducible answers, not creativity.
111	sampling:
112	temperature: 0.0
113	parallel_tool_calls: false
114	top_p: 0.9
115	max_tokens: 4096
116
117	watchdogTimeout: 30s
118
119	# Context compression
120	context:
121	enabled: false
122
123	# Compress when the history reaches this share of the window.
124	threshold: 75
125
126	# Fallback on the message count, for when the window is unknown.
127	# One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
128	maxMessages: 80
129
130	# How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
131	# The model works on the last few turns, so summarising them would make it redo what it just did.
132	keepLastTurns: 3
133
134	# max_tokens of the summary request.
135	summaryMaxTokens: 1200
136
137	# Replaces the built-in summary prompt when set.
138	# Empty = built-in.
139	prompt: ""
140
141	showStats: true
142
143	displayCommands: true