bots-garden/mini-me · agent.dmr.yaml

⑂ d72271127802973540c648bfb372176cdaaa8e4f

agent.dmr.yaml · 208 lines · 10.3 KBYAML    Blame HistoryRaw
   1# Agent settings.
3# Every key is optional: what you leave out keeps its built-in default
4# (see internal/config/config.go). Nothing here requires recompiling —
5# `go run .` picks the file up at startup.
7# Use another file with:  AGENT_CONFIG=./fast.yaml go run .
9# Which server is behind the URL. Left out, it means `dmr`: every agent.yaml
10# written before this key existed keeps working unchanged. `llamacpp` speaks
11# the same OpenAI-compatible protocol, but with its own port, its own way of
12# refusing tool calls (no --jinja) and its own words in error messages — the
13# provider is what turns those into one useful line. See agent.llamacpp.yaml.
14# AGENT_PROVIDER overrides this key for one run.
15provider: dmr
17# The chat model, in the provider's own naming (DMR: `ai/…`, `hf.co/…`). It
18# MUST be able to emit tool_calls, otherwise the agent never runs a single
19# command. AGENT_MODEL overrides it.
20model: huggingface.co/jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
22# Endpoint. `baseUrl` is tried first; if it does not answer, `fallback` is used
23# — handy from a container or a sandbox, where the host is reachable as
24# host.docker.internal. Both are LEFT OUT here on purpose: each provider has its
25# defaults, and an explicit value is always respected — so with DMR's URL
26# written in this file, `AGENT_PROVIDER=llamacpp ./bob agent.yaml` still talked
27# to port 12434. Leaving them out keeps the file provider-neutral. These are the
28# values the `dmr` provider fills in:
29#   baseUrl: http://localhost:12434/engines/v1
30#   fallback: http://host.docker.internal:12434/engines/v1
31# AGENT_BASE_URL — or, for dmr, the older DMR_BASE_URL — overrides both without
32# probing. `fallback: ""` (an explicit empty string) disables the fallback.
34# The NAME of the environment variable holding the API key — never the key
35# itself, so this file can be committed and shown on screen. DMR ignores the
36# key, so nothing is needed here; llama-server checks one only when started
37# with --api-key (default variable: LLAMA_API_KEY).
38# apiKeyEnv: LLAMA_API_KEY
40# How many tokens the server actually serves. The model does not know this:
41# it is what the operator started the server with (`llama-server -c 32768`,
42# `docker model configure`). Shown at start-up next to its origin, because a
43# number without an origin never gets corrected. 0 = ask the server
44# (llama-server tells its n_ctx on /props), else "unknown".
45contextWindow: 0
47# Max number of characters a tool returns to the model. Context safeguard:
48# beyond that, the output is truncated (beginning + end kept).
49maxOutput: 16000
51# Max number of model ↔ tools round trips for a single question. Raise it for
52# tasks that need long chains of commands.
53maxTurns: 40
55# The built-in file tools: read_file, write_file, edit_file. This is the switch
56# this part exists for. `true`: the model edits files through tools it can SEE
57# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
58# and edits files through the `edit` CLI if it is on the PATH. Same binary,
59# same prompts, two set-ups: measured side by side, that is the comparison
60# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
61# in the prompt plus `cat` managed 1/11). Paths are relative to the current
62# directory and are not confined to it.
63editTools: true
65# Where the `read_skill` tool looks for its markdown procedures. The path is
66# relative to the CURRENT DIRECTORY, so run the agent from the directory that
67# holds skills/ — not from src/. When the directory has no *.md file, the tool
68# is not declared at all.
69skillsDir: skills
71# How many lines of a command's output are echoed to the terminal. "Show me
72# that file" means show it: running `cat` is not displaying it. 0 disables it.
73previewLines: 20
75# The system prompt: what the agent is, and what it is allowed to do.
76# This is the knob to play with — it is the shortest path to changing behaviour.
77
78  Your name is Bob.
79  You are a coding agent working in a terminal.
80  You have a "bash" tool to run shell commands.
81  Use it to explore files, run tests, inspect the repository, etc.
82  Chain several commands if needed, then answer clearly in English.
84  A request often mixes things you answer from yourself ("say hello") with
85  things only a command can answer ("list the files"). Handle every part, in
86  the order asked, and run a command for each part that needs one.
87  Never state the contents of a file, the output of a command, or the state of
88  the repository unless a command in THIS answer returned it. What you did not
89  read, you do not know: run the command instead of recalling it.
92  You have a second tool, `read_skill`. Its description lists the procedures
93  available for this project — one per kind of task.
95  Any request to DO something to a Go project is a skill, not a shell command
96  you invent. Match the request against that list, call `read_skill` FIRST,
97  before any bash command, and then follow what it says step by step.
99  FILE EDITING
100  You have three tools for files: `read_file`, `edit_file` and `write_file`.
101  They are how a file gets read and changed here: each change is exact,
102  checked before it is written, and comes back as a diff with line numbers.
103  bash is for running things — building, testing, listing, searching.
105  - Read before you write: call `read_file` on the file (numbered=true when
106    you need line numbers). You cannot target text you have not seen; never
107    rely on what you think you remember about a file.
108  - To change an existing file, call `edit_file` with one or more {old, new}
109    pairs. `old` is copied from the file character for character — same
110    spaces, same indentation, same line breaks — and appears exactly once:
111    add the surrounding lines until it is unique. Several pairs are applied
112    together, against the original file. An empty `new` deletes the text.
113  - Call `write_file` only to create a file, or to rewrite one entirely and
114    on purpose. On an existing file it replaces everything, including what
115    you did not intend to touch.
116  - Read the diff the tool returns: it says exactly what changed and on which
117    line. If `edit_file` refuses — text not found, ambiguous, overlapping
118    edits — read the file again and fix `old`. Do not fall back to
119    `write_file` to force the change through.
120  - After editing code, run the narrowest check with bash: the formatter, the
121    compiler, or the test covering that file.
124  - Keep everything the file already does, unless the user asked to remove it.
125  - Touch only the files the request is about. Do not add tests, files or
126    features that were not asked for.
127  - Never run a git command unless the user says git, commit or push.
128  - Never move, rename or delete a file unless the user asked for it.
129  - Then answer in English, in a few lines.
130  - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
131    to understand the options, then run the command.
133  BACKGROUND JOBS
134  Never let a command block the answer. Anything that serves, watches or runs
135  long goes to the background, with BOTH streams redirected and its pid kept:
137    nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
139  Redirecting only stdout still blocks until the process exits. Read the
140  `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
141  rule you cannot guess. Stop every job you started before you finish, and say
142  which ones you left running.
144# Generation settings (OpenAI API keys). Low temperature for a coding agent:
145# we want precise and reproducible answers, not creativity.
146
147  temperature: 0.0
148  # Une seule commande par tour. Sans ça, le modèle peut en demander
149  # plusieurs d'un coup : Genkit les exécute EN PARALLÈLE, dans le même
150  # répertoire, et on ne sait plus quelle sortie appartient à quelle
151  # commande. Pour le détecteur de boucle, c'est pire : deux appels du même
152  # tour s'enregistrent dans un ordre indéterminé, et « la même action deux
153  # fois de suite » perd son sens. Mettre `true` pour retrouver le
154  # comportement par défaut de l'API.
155  parallel_tool_calls: false
156  top_p: 0.9
157  max_tokens: 4096
159watchdogTimeout: 30s
161# Context compression, carried over from part 08 (its CONTEXT_WINDOW.md has the
162# reasoning). OFF by default: with `enabled: false` the agent behaves exactly
163# as before. Why it exists: nothing in the agent ever shortens the history —
164# measured with the fake engine of part 03, the messages sent to the model grew
165# 2 → 5 → 7 → 9 over four requests, and every `bash` output (up to maxOutput
166# characters) and every skill read stays until /quit. On a local model the
167# window is fixed at load time and the prefill runs on the laptop, so a long
168# session first shows up as a false "[watchdog: connection hang]", then as an
169# error from the server.
171# The window it measures against is the top-level `contextWindow` above — ONE
172# key for the banner and for the trigger, so the two cannot disagree. When
173# that key is 0, the value the provider's probe learned from the server is
174# used (llama-server tells its n_ctx on /props; DMR tells nothing). When both
175# are unknown, only `maxMessages` can trigger, and the agent says so at
176# start-up.
177
178  enabled: false
180  # Compress when the history reaches this share of the window. 75 leaves a
181  # quarter for the next question, the outputs of its turns and the answer —
182  # a single `bash` output can be maxOutput characters, about 4-5k tokens.
183  threshold: 75
185  # Fallback on the message count, for when the window is unknown or the
186  # estimate is off. One command costs 2 messages (call + response), so 80 is
187  # roughly 30 commands of history.
188  maxMessages: 80
190  # Question turns kept raw at the end. The recent turns are where the model
191  # works; summarising them makes it re-run what it just did. Three covers
192  # "do X", "now fix it", "and test it" — the usual shape of a demo.
193  keepLastTurns: 3
195  # max_tokens of the summary request. Seven sections of one line per item
196  # fit in far less; the cap stops a runaway model from filling the window it
197  # was asked to empty.
198  summaryMaxTokens: 1200
200  # Replaces the built-in summary prompt (08-context-compression/CONTEXT_WINDOW.md
201  # § 4.1) when set. Empty = built-in.
202  prompt: ""
204  # One line after each compression: what was replaced, what was kept,
205  # tokens before → after. Worth showing on a screen, noise in a log.
206  showStats: true
208displayCommands: true

💾 Saved. d722711 k33g 4h ago	1	# Agent settings.
	2	#
	3	# Every key is optional: what you leave out keeps its built-in default
	4	# (see internal/config/config.go). Nothing here requires recompiling —
	5	# `go run .` picks the file up at startup.
	6	#
	7	# Use another file with: AGENT_CONFIG=./fast.yaml go run .
	8
	9	# Which server is behind the URL. Left out, it means `dmr`: every agent.yaml
	10	# written before this key existed keeps working unchanged. `llamacpp` speaks
	11	# the same OpenAI-compatible protocol, but with its own port, its own way of
	12	# refusing tool calls (no --jinja) and its own words in error messages — the
	13	# provider is what turns those into one useful line. See agent.llamacpp.yaml.
	14	# AGENT_PROVIDER overrides this key for one run.
	15	provider: dmr
	16
	17	# The chat model, in the provider's own naming (DMR: `ai/…`, `hf.co/…`). It
	18	# MUST be able to emit tool_calls, otherwise the agent never runs a single
	19	# command. AGENT_MODEL overrides it.
	20	model: huggingface.co/jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
	21
	22	# Endpoint. `baseUrl` is tried first; if it does not answer, `fallback` is used
	23	# — handy from a container or a sandbox, where the host is reachable as
	24	# host.docker.internal. Both are LEFT OUT here on purpose: each provider has its
	25	# defaults, and an explicit value is always respected — so with DMR's URL
	26	# written in this file, `AGENT_PROVIDER=llamacpp ./bob agent.yaml` still talked
	27	# to port 12434. Leaving them out keeps the file provider-neutral. These are the
	28	# values the `dmr` provider fills in:
	29	# baseUrl: http://localhost:12434/engines/v1
	30	# fallback: http://host.docker.internal:12434/engines/v1
	31	# AGENT_BASE_URL — or, for dmr, the older DMR_BASE_URL — overrides both without
	32	# probing. `fallback: ""` (an explicit empty string) disables the fallback.
	33
	34	# The NAME of the environment variable holding the API key — never the key
	35	# itself, so this file can be committed and shown on screen. DMR ignores the
	36	# key, so nothing is needed here; llama-server checks one only when started
	37	# with --api-key (default variable: LLAMA_API_KEY).
	38	# apiKeyEnv: LLAMA_API_KEY
	39
	40	# How many tokens the server actually serves. The model does not know this:
	41	# it is what the operator started the server with (`llama-server -c 32768`,
	42	# `docker model configure`). Shown at start-up next to its origin, because a
	43	# number without an origin never gets corrected. 0 = ask the server
	44	# (llama-server tells its n_ctx on /props), else "unknown".
	45	contextWindow: 0
	46
	47	# Max number of characters a tool returns to the model. Context safeguard:
	48	# beyond that, the output is truncated (beginning + end kept).
	49	maxOutput: 16000
	50
	51	# Max number of model ↔ tools round trips for a single question. Raise it for
	52	# tasks that need long chains of commands.
	53	maxTurns: 40
	54
	55	# The built-in file tools: read_file, write_file, edit_file. This is the switch
	56	# this part exists for. `true`: the model edits files through tools it can SEE
	57	# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
	58	# and edits files through the `edit` CLI if it is on the PATH. Same binary,
	59	# same prompts, two set-ups: measured side by side, that is the comparison
	60	# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
	61	# in the prompt plus `cat` managed 1/11). Paths are relative to the current
	62	# directory and are not confined to it.
	63	editTools: true
	64
65	# Where the `read_skill` tool looks for its markdown procedures. The path is
66	# relative to the CURRENT DIRECTORY, so run the agent from the directory that
67	# holds skills/ — not from src/. When the directory has no *.md file, the tool
68	# is not declared at all.
69	skillsDir: skills
70
71	# How many lines of a command's output are echoed to the terminal. "Show me
72	# that file" means show it: running `cat` is not displaying it. 0 disables it.
73	previewLines: 20
74
75	# The system prompt: what the agent is, and what it is allowed to do.
76	# This is the knob to play with — it is the shortest path to changing behaviour.
77	system: \|
78	Your name is Bob.
79	You are a coding agent working in a terminal.
80	You have a "bash" tool to run shell commands.
81	Use it to explore files, run tests, inspect the repository, etc.
82	Chain several commands if needed, then answer clearly in English.
83
84	A request often mixes things you answer from yourself ("say hello") with
85	things only a command can answer ("list the files"). Handle every part, in
86	the order asked, and run a command for each part that needs one.
87	Never state the contents of a file, the output of a command, or the state of
88	the repository unless a command in THIS answer returned it. What you did not
89	read, you do not know: run the command instead of recalling it.
90
91	SKILLS
92	You have a second tool, `read_skill`. Its description lists the procedures
93	available for this project — one per kind of task.
94
95	Any request to DO something to a Go project is a skill, not a shell command
96	you invent. Match the request against that list, call `read_skill` FIRST,
97	before any bash command, and then follow what it says step by step.
98
99	FILE EDITING
100	You have three tools for files: `read_file`, `edit_file` and `write_file`.
101	They are how a file gets read and changed here: each change is exact,
102	checked before it is written, and comes back as a diff with line numbers.
103	bash is for running things — building, testing, listing, searching.
104
105	- Read before you write: call `read_file` on the file (numbered=true when
106	you need line numbers). You cannot target text you have not seen; never
107	rely on what you think you remember about a file.
108	- To change an existing file, call `edit_file` with one or more {old, new}
109	pairs. `old` is copied from the file character for character — same
110	spaces, same indentation, same line breaks — and appears exactly once:
111	add the surrounding lines until it is unique. Several pairs are applied
112	together, against the original file. An empty `new` deletes the text.
113	- Call `write_file` only to create a file, or to rewrite one entirely and
114	on purpose. On an existing file it replaces everything, including what
115	you did not intend to touch.
116	- Read the diff the tool returns: it says exactly what changed and on which
117	line. If `edit_file` refuses — text not found, ambiguous, overlapping
118	edits — read the file again and fix `old`. Do not fall back to
119	`write_file` to force the change through.
120	- After editing code, run the narrowest check with bash: the formatter, the
121	compiler, or the test covering that file.
122
123	RULES
124	- Keep everything the file already does, unless the user asked to remove it.
125	- Touch only the files the request is about. Do not add tests, files or
126	features that were not asked for.
127	- Never run a git command unless the user says git, commit or push.
128	- Never move, rename or delete a file unless the user asked for it.
129	- Then answer in English, in a few lines.
130	- If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
131	to understand the options, then run the command.
132
133	BACKGROUND JOBS
134	Never let a command block the answer. Anything that serves, watches or runs
135	long goes to the background, with BOTH streams redirected and its pid kept:
136
137	nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
138
139	Redirecting only stdout still blocks until the process exits. Read the
140	`bg-jobs` skill before you wait on, inspect or stop such a job — each has a
141	rule you cannot guess. Stop every job you started before you finish, and say
142	which ones you left running.
143
144	# Generation settings (OpenAI API keys). Low temperature for a coding agent:
145	# we want precise and reproducible answers, not creativity.
146	sampling:
147	temperature: 0.0
148	# Une seule commande par tour. Sans ça, le modèle peut en demander
149	# plusieurs d'un coup : Genkit les exécute EN PARALLÈLE, dans le même
150	# répertoire, et on ne sait plus quelle sortie appartient à quelle
151	# commande. Pour le détecteur de boucle, c'est pire : deux appels du même
152	# tour s'enregistrent dans un ordre indéterminé, et « la même action deux
153	# fois de suite » perd son sens. Mettre `true` pour retrouver le
154	# comportement par défaut de l'API.
155	parallel_tool_calls: false
156	top_p: 0.9
157	max_tokens: 4096
158
159	watchdogTimeout: 30s
160
161	# Context compression, carried over from part 08 (its CONTEXT_WINDOW.md has the
162	# reasoning). OFF by default: with `enabled: false` the agent behaves exactly
163	# as before. Why it exists: nothing in the agent ever shortens the history —
164	# measured with the fake engine of part 03, the messages sent to the model grew
165	# 2 → 5 → 7 → 9 over four requests, and every `bash` output (up to maxOutput
166	# characters) and every skill read stays until /quit. On a local model the
167	# window is fixed at load time and the prefill runs on the laptop, so a long
168	# session first shows up as a false "[watchdog: connection hang]", then as an
169	# error from the server.
170	#
171	# The window it measures against is the top-level `contextWindow` above — ONE
172	# key for the banner and for the trigger, so the two cannot disagree. When
173	# that key is 0, the value the provider's probe learned from the server is
174	# used (llama-server tells its n_ctx on /props; DMR tells nothing). When both
175	# are unknown, only `maxMessages` can trigger, and the agent says so at
176	# start-up.
177	context:
178	enabled: false
179
180	# Compress when the history reaches this share of the window. 75 leaves a
181	# quarter for the next question, the outputs of its turns and the answer —
182	# a single `bash` output can be maxOutput characters, about 4-5k tokens.
183	threshold: 75
184
185	# Fallback on the message count, for when the window is unknown or the
186	# estimate is off. One command costs 2 messages (call + response), so 80 is
187	# roughly 30 commands of history.
188	maxMessages: 80
189
190	# Question turns kept raw at the end. The recent turns are where the model
191	# works; summarising them makes it re-run what it just did. Three covers
192	# "do X", "now fix it", "and test it" — the usual shape of a demo.
193	keepLastTurns: 3
194
195	# max_tokens of the summary request. Seven sections of one line per item
196	# fit in far less; the cap stops a runaway model from filling the window it
197	# was asked to empty.
198	summaryMaxTokens: 1200
199
200	# Replaces the built-in summary prompt (08-context-compression/CONTEXT_WINDOW.md
201	# § 4.1) when set. Empty = built-in.
202	prompt: ""
203
204	# One line after each compression: what was replaced, what was kept,
205	# tokens before → after. Worth showing on a screen, noise in a log.
206	showStats: true
207
208	displayCommands: true