rickub
bots-garden/one-day-one-featurepublic Fork
⑂ b5678dd
Branches
main View all branches → View all tags →
Commits
⬇ Clone ▾
git clone https://git.rickub.com/bots-garden/one-day-one-feature.git
git clone ssh://git@rickub.com/bots-garden/one-day-one-feature.git

Host key fingerprint (ed25519): SHA256:iycHnxEyq0Q7uyVpB7JlznP0G7JrTPXLYRcAU5CSLhc — verify it before your first connect.

Download ZIP

🛟 Updated.

k33g committed 2026-10-26T07:24:07+01:00 Browse files
b5678dd parent: 1ea7eb9
added 03-decision-models/.mm/decider/20261025-183829.000-spell.json +27 -0
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:38:29.000528+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "An ogre is raising its club to hit me right now.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "shield",
11+ "probabilities": {
12+ "fireball": 0.02448743490299417,
13+ "arcane_bolt": 0.04358880075661369,
14+ "shield": 0.878106678929716,
15+ "misty_step": 0.043028172638960124,
16+ "sleep": 0.010788912771716069
17+ },
18+ "confidence": 0.8476333486621448
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 109,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: shield (confidence 0.85)"
27+}
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:38:29.000528+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "An ogre is raising its club to hit me right now.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "shield",
11+ "probabilities": {
12+ "fireball": 0.02448743490299417,
13+ "arcane_bolt": 0.04358880075661369,
14+ "shield": 0.878106678929716,
15+ "misty_step": 0.043028172638960124,
16+ "sleep": 0.010788912771716069
17+ },
18+ "confidence": 0.8476333486621448
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 109,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: shield (confidence 0.85)"
27+}
added 03-decision-models/.mm/decider/20261025-183845.172-spell.json +27 -0
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:38:45.172622+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "Two wolves have grabbed me and the rest of the party is far away.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "misty_step",
11+ "probabilities": {
12+ "fireball": 0.18536138820125475,
13+ "arcane_bolt": 0.10610648657351471,
14+ "shield": 0.07294671787031991,
15+ "misty_step": 0.578016891488816,
16+ "sleep": 0.057568515866094655
17+ },
18+ "confidence": 0.4725211143610199
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 111,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: misty_step (confidence 0.47)"
27+}
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:38:45.172622+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "Two wolves have grabbed me and the rest of the party is far away.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "misty_step",
11+ "probabilities": {
12+ "fireball": 0.18536138820125475,
13+ "arcane_bolt": 0.10610648657351471,
14+ "shield": 0.07294671787031991,
15+ "misty_step": 0.578016891488816,
16+ "sleep": 0.057568515866094655
17+ },
18+ "confidence": 0.4725211143610199
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 111,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: misty_step (confidence 0.47)"
27+}
added 03-decision-models/.mm/decider/20261025-183934.851-spell.json +27 -0
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:39:34.851288+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "One goblin is left, almost dead, hiding behind a crate.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "arcane_bolt",
11+ "probabilities": {
12+ "fireball": 0.06920117963011733,
13+ "arcane_bolt": 0.6636863594916528,
14+ "shield": 0.07923066182391235,
15+ "misty_step": 0.10690331936698312,
16+ "sleep": 0.0809784796873344
17+ },
18+ "confidence": 0.5796079493645659
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 110,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: arcane_bolt (confidence 0.58)"
27+}
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:39:34.851288+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "One goblin is left, almost dead, hiding behind a crate.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "arcane_bolt",
11+ "probabilities": {
12+ "fireball": 0.06920117963011733,
13+ "arcane_bolt": 0.6636863594916528,
14+ "shield": 0.07923066182391235,
15+ "misty_step": 0.10690331936698312,
16+ "sleep": 0.0809784796873344
17+ },
18+ "confidence": 0.5796079493645659
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 110,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: arcane_bolt (confidence 0.58)"
27+}
added 03-decision-models/.mm/decider/20261025-183955.867-spell.json +27 -0
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:39:55.867926+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "Six kobolds are sneaking toward our camp; the paladin wants to question them alive.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "fireball",
11+ "probabilities": {
12+ "fireball": 0.5556530078576913,
13+ "arcane_bolt": 0.029990685461619208,
14+ "shield": 0.026335283620166735,
15+ "misty_step": 0.043796993377524546,
16+ "sleep": 0.3442240296829981
17+ },
18+ "confidence": 0.4445662598221141
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 115,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: fireball (confidence 0.44)"
27+}
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T18:39:55.867926+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "Six kobolds are sneaking toward our camp; the paladin wants to question them alive.",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "fireball",
11+ "probabilities": {
12+ "fireball": 0.5556530078576913,
13+ "arcane_bolt": 0.029990685461619208,
14+ "shield": 0.026335283620166735,
15+ "misty_step": 0.043796993377524546,
16+ "sleep": 0.3442240296829981
17+ },
18+ "confidence": 0.4445662598221141
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 115,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: fireball (confidence 0.44)"
27+}
added 03-decision-models/.mm/decider/20261025-190115.585-spell.json +27 -0
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T19:01:15.585303+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "hello",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "sleep",
11+ "probabilities": {
12+ "fireball": 0.1096667211238828,
13+ "arcane_bolt": 0.2628631167747479,
14+ "shield": 0.15284316935906098,
15+ "misty_step": 0.16086838121035513,
16+ "sleep": 0.3137586115319532
17+ },
18+ "confidence": 0.14219826441494152
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 98,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: sleep (confidence 0.14)"
27+}
new file mode 100644
@@ -0,0 +1,27 @@
1+{
2+ "time": "2026-10-25T19:01:15.585303+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "hello",
5+ "response": {
6+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
7+ "answers": {
8+ "spell": {
9+ "type": "choice",
10+ "choice": "sleep",
11+ "probabilities": {
12+ "fireball": 0.1096667211238828,
13+ "arcane_bolt": 0.2628631167747479,
14+ "shield": 0.15284316935906098,
15+ "misty_step": 0.16086838121035513,
16+ "sleep": 0.3137586115319532
17+ },
18+ "confidence": 0.14219826441494152
19+ }
20+ },
21+ "usage": {
22+ "input_tokens": 98,
23+ "output_tokens": 0
24+ }
25+ },
26+ "context": "spell: sleep (confidence 0.14)"
27+}
added 03-decision-models/.mm/decider/20261025-190209.703-spell.json +28 -0
new file mode 100644
@@ -0,0 +1,28 @@
1+{
2+ "time": "2026-10-25T19:02:09.703072+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "Two wolves have grabbed me and the rest of the party is far away.",
5+ "response": {
6+ "display": true,
7+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
8+ "answers": {
9+ "spell": {
10+ "type": "choice",
11+ "choice": "misty_step",
12+ "probabilities": {
13+ "fireball": 0.18549138627023762,
14+ "arcane_bolt": 0.10610184523928502,
15+ "shield": 0.07297497689065748,
16+ "misty_step": 0.5778020206231758,
17+ "sleep": 0.05762977097664409
18+ },
19+ "confidence": 0.47225252577896965
20+ }
21+ },
22+ "usage": {
23+ "input_tokens": 111,
24+ "output_tokens": 0
25+ }
26+ },
27+ "context": "spell: misty_step (confidence 0.47)"
28+}
new file mode 100644
@@ -0,0 +1,28 @@
1+{
2+ "time": "2026-10-25T19:02:09.703072+01:00",
3+ "decider": "spell.yaml",
4+ "prompt": "Two wolves have grabbed me and the rest of the party is far away.",
5+ "response": {
6+ "display": true,
7+ "model": "hf.co/LiquidAI/d1-3B-GGUF:Q4_K_M",
8+ "answers": {
9+ "spell": {
10+ "type": "choice",
11+ "choice": "misty_step",
12+ "probabilities": {
13+ "fireball": 0.18549138627023762,
14+ "arcane_bolt": 0.10610184523928502,
15+ "shield": 0.07297497689065748,
16+ "misty_step": 0.5778020206231758,
17+ "sleep": 0.05762977097664409
18+ },
19+ "confidence": 0.47225252577896965
20+ }
21+ },
22+ "usage": {
23+ "input_tokens": 111,
24+ "output_tokens": 0
25+ }
26+ },
27+ "context": "spell: misty_step (confidence 0.47)"
28+}
modified 03-decision-models/spell.yaml +3 -10
@@ -5,6 +5,7 @@ baseUrl: http://127.0.0.1:17434/v1
55 fallback: http://host.docker.internal:17434/v1
66
77 decision:
8+ display: true
89 questions:
910 spell:
1011 type: choice
@@ -18,14 +19,6 @@ decision:
1819
1920 acp:
2021 name: sidekick
21- title: "sidekick"
22+ title: "🐱 sidekick"
23+
2224
23-# A decider declares no tool: its model never calls one.
24-bashTool: false
25-editTools: false
26-planTool: false
27-fetchContentTool: false
28-askUserTool: false
29-skillsDir: no-skills
30-agentsMd:
31- enabled: false
@@ -5,6 +5,7 @@ baseUrl: http://127.0.0.1:17434/v1
5 fallback: http://host.docker.internal:17434/v15 fallback: http://host.docker.internal:17434/v1
6 6
7 decision:7 decision:
8+ display: true
8 questions:9 questions:
9 spell:10 spell:
10 type: choice11 type: choice
@@ -18,14 +19,6 @@ decision:
18 19
19 acp:20 acp:
20 name: sidekick21 name: sidekick
21- title: "sidekick"22+ title: "🐱 sidekick"
23+
22 24
23-# A decider declares no tool: its model never calls one.
24-bashTool: false
25-editTools: false
26-planTool: false
27-fetchContentTool: false
28-askUserTool: false
29-skillsDir: no-skills
30-agentsMd:
31- enabled: false
added 09-unload-models/README.md +54 -0
new file mode 100644
@@ -0,0 +1,54 @@
1+# Coding helpers, with the `unloadModels` option
2+
3+
4+The team: a coordinator that hands Go and Rust questions to a specialist, with one difference: the two models are never in the server's memory together, and no hook is needed for it. Each `mmConfigs` entry sets `unloadModels: true`.
5+
6+| Agent | File | Model | Role |
7+|---|---|---|---|
8+| 🧭 coordinator | `agent.yaml` | MiniCPM5 2B (`hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M`, 1.6 GB) | picks `go` or `rust`, sends the question with the conversation |
9+| 🐹 go | `agent.go.yaml` | Mellum2 12B A2.5B thinking (`huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M`, 8.1 GB) | documents Go |
10+| 🦀 rust | `agent.rust.yaml` | the same Mellum2 model | documents Rust |
11+
12+## Run
13+
14+```bash
15+llmman pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
16+llmman pull huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
17+
18+cd 09-unload-models
19+mm tui # reads ./agent.yaml
20+```
21+
22+## The option
23+
24+```yaml
25+mmConfigs:
26+ go:
27+ path: agent.go.yaml
28+ description: >-
29+ Go specialist. …
30+ yolo: true
31+ unloadModels: true
32+ rust:
33+ path: agent.rust.yaml
34+ description: >-
35+ Rust specialist. …
36+ yolo: true
37+ unloadModels: true
38+```
39+
40+On every hand-off to `go` or `rust`, the coordinator's `ask_agent`, your `/ask-go` or `/ask-rust`:
41+
42+1. `mm` unloads the coordinator's model (MiniCPM5) right before the specialist starts;
43+2. the specialist loads Mellum2 and answers;
44+3. `mm` unloads Mellum2 once the answer is in, before the coordinator is reloaded to read it.
45+
46+Each unload is one line:
47+
48+```text
49+🗣 ask_agent: rust (llmman/huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M) — …
50+🧹 unloaded hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
51+…the specialist's answer…
52+🧹 unloaded huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
53+```
54+
new file mode 100644
@@ -0,0 +1,54 @@
1+# Coding helpers, with the `unloadModels` option
2+
3+
4+The team: a coordinator that hands Go and Rust questions to a specialist, with one difference: the two models are never in the server's memory together, and no hook is needed for it. Each `mmConfigs` entry sets `unloadModels: true`.
5+
6+| Agent | File | Model | Role |
7+|---|---|---|---|
8+| 🧭 coordinator | `agent.yaml` | MiniCPM5 2B (`hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M`, 1.6 GB) | picks `go` or `rust`, sends the question with the conversation |
9+| 🐹 go | `agent.go.yaml` | Mellum2 12B A2.5B thinking (`huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M`, 8.1 GB) | documents Go |
10+| 🦀 rust | `agent.rust.yaml` | the same Mellum2 model | documents Rust |
11+
12+## Run
13+
14+```bash
15+llmman pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
16+llmman pull huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
17+
18+cd 09-unload-models
19+mm tui # reads ./agent.yaml
20+```
21+
22+## The option
23+
24+```yaml
25+mmConfigs:
26+ go:
27+ path: agent.go.yaml
28+ description: >-
29+ Go specialist. …
30+ yolo: true
31+ unloadModels: true
32+ rust:
33+ path: agent.rust.yaml
34+ description: >-
35+ Rust specialist. …
36+ yolo: true
37+ unloadModels: true
38+```
39+
40+On every hand-off to `go` or `rust`, the coordinator's `ask_agent`, your `/ask-go` or `/ask-rust`:
41+
42+1. `mm` unloads the coordinator's model (MiniCPM5) right before the specialist starts;
43+2. the specialist loads Mellum2 and answers;
44+3. `mm` unloads Mellum2 once the answer is in, before the coordinator is reloaded to read it.
45+
46+Each unload is one line:
47+
48+```text
49+🗣 ask_agent: rust (llmman/huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M) — …
50+🧹 unloaded hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
51+…the specialist's answer…
52+🧹 unloaded huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
53+```
54+
added 09-unload-models/agent.go.yaml +79 -0
new file mode 100644
@@ -0,0 +1,79 @@
1+# Go specialist — asked by the coordinator (agent.yaml) through `ask_agent`,
2+# or by you directly: `mm agent.go.yaml`. Interactive documentation: it
3+# explains a Go concept with short examples in its answer. No tools — it
4+# builds nothing, writes no file, and only writes a whole program when asked
5+# to.
6+#
7+# Mellum2 12B A2.5B thinking (JetBrains, a mixture of experts: 12B weights,
8+# 2.5B active per token): a code model that reasons before answering.
9+#
10+# A copy of demos/coding-helpers/agent.go.yaml: the specialists are unchanged.
11+
12+provider: llmman
13+model: huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
14+baseUrl: http://127.0.0.1:17434/v1
15+fallback: http://host.docker.internal:17434/v1
16+contextWindow: 32768
17+
18+sampling:
19+ # No sampling in the GGUF; values usual for a thinking model (never 0.0:
20+ # thinking models loop).
21+ temperature: 0.6
22+ top_p: 0.95
23+ # The reasoning plus an explanation with examples.
24+ max_tokens: 8192
25+ # On by default in the template; written so the choice is visible.
26+ chat_template_kwargs:
27+ enable_thinking: true
28+
29+showThinking: false
30+watchdogTimeout: 3m
31+firstTokenTimeout: 3m
32+
33+# No tools: one answer.
34+maxTurns: 2
35+maxOutput: 20000
36+
37+yolo: false
38+bashTool: false
39+editTools: false
40+planTool: false
41+fetchContentTool: false
42+askUserTool: false
43+skillsDir: no-skills
44+
45+agentsMd:
46+ enabled: false
47+
48+sessions:
49+ enabled: false
50+
51+acp:
52+ name: go
53+ title: 🐹 Go specialist
54+
55+system: |
56+ You are Go documentation that answers questions. People ask what a Go
57+ concept is or how something is done in Go; you explain it and show
58+ examples, like a good reference page would.
59+
60+ If the conversation so far is included, read it: "the same thing" means
61+ what was asked or shown before — explain it for Go.
62+
63+ Your answer:
64+ 1. What it is and when to use it, in a few sentences.
65+ 2. Short examples in ```go code blocks — a snippet per point, only
66+ the lines that show the idea (no main function, no imports unless they
67+ matter to the point).
68+ 3. The pitfalls or the idiomatic way, in a short list, when there are
69+ any.
70+
71+ RULES
72+ - Only Go. A question about another language (Python, Java…): answer in
73+ one sentence that you document Go only — never answer it with Go.
74+ - You document; you do not build. Write a complete, runnable program only
75+ when the user explicitly asks for one ("write me a program", "a full
76+ example I can run").
77+ - Standard library unless the question names a package.
78+ - Keep it short and correct: an example that does not compile is worse
79+ than no example.
new file mode 100644
@@ -0,0 +1,79 @@
1+# Go specialist — asked by the coordinator (agent.yaml) through `ask_agent`,
2+# or by you directly: `mm agent.go.yaml`. Interactive documentation: it
3+# explains a Go concept with short examples in its answer. No tools — it
4+# builds nothing, writes no file, and only writes a whole program when asked
5+# to.
6+#
7+# Mellum2 12B A2.5B thinking (JetBrains, a mixture of experts: 12B weights,
8+# 2.5B active per token): a code model that reasons before answering.
9+#
10+# A copy of demos/coding-helpers/agent.go.yaml: the specialists are unchanged.
11+
12+provider: llmman
13+model: huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
14+baseUrl: http://127.0.0.1:17434/v1
15+fallback: http://host.docker.internal:17434/v1
16+contextWindow: 32768
17+
18+sampling:
19+ # No sampling in the GGUF; values usual for a thinking model (never 0.0:
20+ # thinking models loop).
21+ temperature: 0.6
22+ top_p: 0.95
23+ # The reasoning plus an explanation with examples.
24+ max_tokens: 8192
25+ # On by default in the template; written so the choice is visible.
26+ chat_template_kwargs:
27+ enable_thinking: true
28+
29+showThinking: false
30+watchdogTimeout: 3m
31+firstTokenTimeout: 3m
32+
33+# No tools: one answer.
34+maxTurns: 2
35+maxOutput: 20000
36+
37+yolo: false
38+bashTool: false
39+editTools: false
40+planTool: false
41+fetchContentTool: false
42+askUserTool: false
43+skillsDir: no-skills
44+
45+agentsMd:
46+ enabled: false
47+
48+sessions:
49+ enabled: false
50+
51+acp:
52+ name: go
53+ title: 🐹 Go specialist
54+
55+system: |
56+ You are Go documentation that answers questions. People ask what a Go
57+ concept is or how something is done in Go; you explain it and show
58+ examples, like a good reference page would.
59+
60+ If the conversation so far is included, read it: "the same thing" means
61+ what was asked or shown before — explain it for Go.
62+
63+ Your answer:
64+ 1. What it is and when to use it, in a few sentences.
65+ 2. Short examples in ```go code blocks — a snippet per point, only
66+ the lines that show the idea (no main function, no imports unless they
67+ matter to the point).
68+ 3. The pitfalls or the idiomatic way, in a short list, when there are
69+ any.
70+
71+ RULES
72+ - Only Go. A question about another language (Python, Java…): answer in
73+ one sentence that you document Go only — never answer it with Go.
74+ - You document; you do not build. Write a complete, runnable program only
75+ when the user explicitly asks for one ("write me a program", "a full
76+ example I can run").
77+ - Standard library unless the question names a package.
78+ - Keep it short and correct: an example that does not compile is worse
79+ than no example.
added 09-unload-models/agent.rust.yaml +79 -0
new file mode 100644
@@ -0,0 +1,79 @@
1+# Rust specialist — asked by the coordinator (agent.yaml) through `ask_agent`,
2+# or by you directly: `mm agent.rust.yaml`. Interactive documentation: it
3+# explains a Rust concept with short examples in its answer. No tools — it
4+# builds nothing, writes no file, and only writes a whole program when asked
5+# to.
6+#
7+# Mellum2 12B A2.5B thinking (JetBrains, a mixture of experts: 12B weights,
8+# 2.5B active per token): a code model that reasons before answering.
9+#
10+# A copy of demos/coding-helpers/agent.rust.yaml: the specialists are unchanged.
11+
12+provider: llmman
13+model: huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
14+baseUrl: http://127.0.0.1:17434/v1
15+fallback: http://host.docker.internal:17434/v1
16+contextWindow: 32768
17+
18+sampling:
19+ # No sampling in the GGUF; values usual for a thinking model (never 0.0:
20+ # thinking models loop).
21+ temperature: 0.6
22+ top_p: 0.95
23+ # The reasoning plus an explanation with examples.
24+ max_tokens: 8192
25+ # On by default in the template; written so the choice is visible.
26+ chat_template_kwargs:
27+ enable_thinking: true
28+
29+showThinking: false
30+watchdogTimeout: 3m
31+firstTokenTimeout: 3m
32+
33+# No tools: one answer.
34+maxTurns: 2
35+maxOutput: 20000
36+
37+yolo: false
38+bashTool: false
39+editTools: false
40+planTool: false
41+fetchContentTool: false
42+askUserTool: false
43+skillsDir: no-skills
44+
45+agentsMd:
46+ enabled: false
47+
48+sessions:
49+ enabled: false
50+
51+acp:
52+ name: rust
53+ title: 🦀 Rust specialist
54+
55+system: |
56+ You are Rust documentation that answers questions. People ask what a Rust
57+ concept is or how something is done in Rust; you explain it and show
58+ examples, like a good reference page would.
59+
60+ If the conversation so far is included, read it: "the same thing" means
61+ what was asked or shown before — explain it for Rust.
62+
63+ Your answer:
64+ 1. What it is and when to use it, in a few sentences.
65+ 2. Short examples in ```rust code blocks — a snippet per point, only
66+ the lines that show the idea (no main function, no imports unless they
67+ matter to the point).
68+ 3. The pitfalls or the idiomatic way, in a short list, when there are
69+ any.
70+
71+ RULES
72+ - Only Rust. A question about another language (Python, Java…): answer in
73+ one sentence that you document Rust only — never answer it with Rust.
74+ - You document; you do not build. Write a complete, runnable program only
75+ when the user explicitly asks for one ("write me a program", "a full
76+ example I can run").
77+ - When an example needs a crate (serde…), name it and give its Cargo.toml line.
78+ - Keep it short and correct: an example that does not compile is worse
79+ than no example.
new file mode 100644
@@ -0,0 +1,79 @@
1+# Rust specialist — asked by the coordinator (agent.yaml) through `ask_agent`,
2+# or by you directly: `mm agent.rust.yaml`. Interactive documentation: it
3+# explains a Rust concept with short examples in its answer. No tools — it
4+# builds nothing, writes no file, and only writes a whole program when asked
5+# to.
6+#
7+# Mellum2 12B A2.5B thinking (JetBrains, a mixture of experts: 12B weights,
8+# 2.5B active per token): a code model that reasons before answering.
9+#
10+# A copy of demos/coding-helpers/agent.rust.yaml: the specialists are unchanged.
11+
12+provider: llmman
13+model: huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
14+baseUrl: http://127.0.0.1:17434/v1
15+fallback: http://host.docker.internal:17434/v1
16+contextWindow: 32768
17+
18+sampling:
19+ # No sampling in the GGUF; values usual for a thinking model (never 0.0:
20+ # thinking models loop).
21+ temperature: 0.6
22+ top_p: 0.95
23+ # The reasoning plus an explanation with examples.
24+ max_tokens: 8192
25+ # On by default in the template; written so the choice is visible.
26+ chat_template_kwargs:
27+ enable_thinking: true
28+
29+showThinking: false
30+watchdogTimeout: 3m
31+firstTokenTimeout: 3m
32+
33+# No tools: one answer.
34+maxTurns: 2
35+maxOutput: 20000
36+
37+yolo: false
38+bashTool: false
39+editTools: false
40+planTool: false
41+fetchContentTool: false
42+askUserTool: false
43+skillsDir: no-skills
44+
45+agentsMd:
46+ enabled: false
47+
48+sessions:
49+ enabled: false
50+
51+acp:
52+ name: rust
53+ title: 🦀 Rust specialist
54+
55+system: |
56+ You are Rust documentation that answers questions. People ask what a Rust
57+ concept is or how something is done in Rust; you explain it and show
58+ examples, like a good reference page would.
59+
60+ If the conversation so far is included, read it: "the same thing" means
61+ what was asked or shown before — explain it for Rust.
62+
63+ Your answer:
64+ 1. What it is and when to use it, in a few sentences.
65+ 2. Short examples in ```rust code blocks — a snippet per point, only
66+ the lines that show the idea (no main function, no imports unless they
67+ matter to the point).
68+ 3. The pitfalls or the idiomatic way, in a short list, when there are
69+ any.
70+
71+ RULES
72+ - Only Rust. A question about another language (Python, Java…): answer in
73+ one sentence that you document Rust only — never answer it with Rust.
74+ - You document; you do not build. Write a complete, runnable program only
75+ when the user explicitly asks for one ("write me a program", "a full
76+ example I can run").
77+ - When an example needs a crate (serde…), name it and give its Cargo.toml line.
78+ - Keep it short and correct: an example that does not compile is worse
79+ than no example.
added 09-unload-models/agent.yaml +117 -0
new file mode 100644
@@ -0,0 +1,117 @@
1+# Coding helpers — interactive documentation for Go and Rust: ask what a
2+# concept is or how something is done, get an explanation with short
3+# examples. You talk with this coordinator; it picks the specialist with
4+# the built-in `ask_agent` tool:
5+#
6+# go — agent.go.yaml, a Go specialist
7+# rust — agent.rust.yaml, a Rust specialist
8+#
9+# "hey, explique-moi comment utiliser une struct Go et la sérialiser en JSON"
10+# "comment je fais la même chose en Rust ?"
11+# "explique-moi ce qu'est le pattern matching en Rust, avec un exemple"
12+#
13+# The second question only makes sense after the first: the coordinator
14+# sends the conversation along (`with_context`), so the Rust specialist
15+# knows what "the same thing" is.
16+#
17+# Models — never in memory together, so at most 8.1 GB at a time:
18+#
19+# coordinator MiniCPM5 2B hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M 1.6 GB
20+# go, rust Mellum2 12B thinking huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M 8.1 GB (one model for both)
21+#
22+# unloadModels: true on both mmConfigs entries (ticket 161): mm unloads
23+# the COORDINATOR's model right before a specialist starts, and the
24+# SPECIALIST's model once it has answered, before the coordinator is
25+# reloaded to read the answer. No hook: demos/coding-helpers does the
26+# same with hooks/unload.js.
27+#
28+# llmman pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
29+# llmman pull huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
30+# LLMMAN_CONTEXT_LENGTH=32768 llmman serve
31+#
32+# cd demos/coding-helpers-unload-option
33+# mm (reads ./agent.yaml)
34+
35+provider: llmman
36+model: hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
37+baseUrl: http://127.0.0.1:17434/v1
38+fallback: http://host.docker.internal:17434/v1
39+contextWindow: 32768
40+
41+# The team. The description is what the coordinator reads to choose;
42+# unloadModels: true makes the two models take turns in memory (the price:
43+# Mellum2 is reloaded on every question).
44+mmConfigs:
45+ go:
46+ path: agent.go.yaml
47+ description: >-
48+ Go specialist. Explains Go concepts and how to do something in Go,
49+ with short examples, like documentation. For any question about Go.
50+ yolo: true
51+ unloadModels: true
52+ rust:
53+ path: agent.rust.yaml
54+ description: >-
55+ Rust specialist. Explains Rust concepts and how to do something in
56+ Rust, with short examples, like documentation. For any question about
57+ Rust.
58+ yolo: true
59+ unloadModels: true
60+
61+sampling:
62+ # The GGUF says 1.0 / 0.95; routing wants a steady hand.
63+ temperature: 0.3
64+ top_p: 0.95
65+ max_tokens: 2048
66+ parallel_tool_calls: false
67+ # MiniCPM5's template reads enable_thinking and has no default: set it.
68+ chat_template_kwargs:
69+ enable_thinking: false
70+
71+# During ask_agent every event of the specialist counts as a sign of life;
72+# a 12B thinking model may reason a while before its first word.
73+watchdogTimeout: 5m
74+firstTokenTimeout: 3m
75+
76+# one call, one sentence.
77+maxTurns: 3
78+maxOutput: 20000
79+displayCommands: true
80+
81+yolo: true
82+bashTool: false
83+editTools: false
84+planTool: false
85+fetchContentTool: false
86+askUserTool: false
87+skillsDir: no-skills
88+
89+agentsMd:
90+ enabled: false
91+
92+sessions:
93+ enabled: false
94+
95+acp:
96+ name: coding-helpers-unload-option
97+ title: 🧭 Coding helpers (Go, Rust)
98+
99+# Written for a 2B model: short, no fake call written out in the text —
100+# measured, MiniCPM5 copies such patterns into its answer instead of calling
101+# the tool.
102+system: |
103+ You are the front desk of two programming specialists: go and rust.
104+
105+ When the user asks something about programming, call the ask_agent tool
106+ once:
107+ - agent: go if the question is about Go, rust if it is about Rust. When
108+ the user says "the same thing", "la même chose", "now in …", pick the
109+ language they name now.
110+ - prompt: the user's message, copied exactly.
111+ - with_context: true, so the specialist sees the conversation so far.
112+
113+ When the tool has answered, write one short sentence in the user's
114+ language saying the explanation is above.
115+
116+ If the language is neither Go nor Rust, say in one sentence that you
117+ only know Go and Rust. When the user only greets you, greet back.
new file mode 100644
@@ -0,0 +1,117 @@
1+# Coding helpers — interactive documentation for Go and Rust: ask what a
2+# concept is or how something is done, get an explanation with short
3+# examples. You talk with this coordinator; it picks the specialist with
4+# the built-in `ask_agent` tool:
5+#
6+# go — agent.go.yaml, a Go specialist
7+# rust — agent.rust.yaml, a Rust specialist
8+#
9+# "hey, explique-moi comment utiliser une struct Go et la sérialiser en JSON"
10+# "comment je fais la même chose en Rust ?"
11+# "explique-moi ce qu'est le pattern matching en Rust, avec un exemple"
12+#
13+# The second question only makes sense after the first: the coordinator
14+# sends the conversation along (`with_context`), so the Rust specialist
15+# knows what "the same thing" is.
16+#
17+# Models — never in memory together, so at most 8.1 GB at a time:
18+#
19+# coordinator MiniCPM5 2B hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M 1.6 GB
20+# go, rust Mellum2 12B thinking huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M 8.1 GB (one model for both)
21+#
22+# unloadModels: true on both mmConfigs entries (ticket 161): mm unloads
23+# the COORDINATOR's model right before a specialist starts, and the
24+# SPECIALIST's model once it has answered, before the coordinator is
25+# reloaded to read the answer. No hook: demos/coding-helpers does the
26+# same with hooks/unload.js.
27+#
28+# llmman pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
29+# llmman pull huggingface.co/jetbrains/mellum2-12b-a2.5b-thinking-gguf-q4_k_m:Q4_K_M
30+# LLMMAN_CONTEXT_LENGTH=32768 llmman serve
31+#
32+# cd demos/coding-helpers-unload-option
33+# mm (reads ./agent.yaml)
34+
35+provider: llmman
36+model: hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M
37+baseUrl: http://127.0.0.1:17434/v1
38+fallback: http://host.docker.internal:17434/v1
39+contextWindow: 32768
40+
41+# The team. The description is what the coordinator reads to choose;
42+# unloadModels: true makes the two models take turns in memory (the price:
43+# Mellum2 is reloaded on every question).
44+mmConfigs:
45+ go:
46+ path: agent.go.yaml
47+ description: >-
48+ Go specialist. Explains Go concepts and how to do something in Go,
49+ with short examples, like documentation. For any question about Go.
50+ yolo: true
51+ unloadModels: true
52+ rust:
53+ path: agent.rust.yaml
54+ description: >-
55+ Rust specialist. Explains Rust concepts and how to do something in
56+ Rust, with short examples, like documentation. For any question about
57+ Rust.
58+ yolo: true
59+ unloadModels: true
60+
61+sampling:
62+ # The GGUF says 1.0 / 0.95; routing wants a steady hand.
63+ temperature: 0.3
64+ top_p: 0.95
65+ max_tokens: 2048
66+ parallel_tool_calls: false
67+ # MiniCPM5's template reads enable_thinking and has no default: set it.
68+ chat_template_kwargs:
69+ enable_thinking: false
70+
71+# During ask_agent every event of the specialist counts as a sign of life;
72+# a 12B thinking model may reason a while before its first word.
73+watchdogTimeout: 5m
74+firstTokenTimeout: 3m
75+
76+# one call, one sentence.
77+maxTurns: 3
78+maxOutput: 20000
79+displayCommands: true
80+
81+yolo: true
82+bashTool: false
83+editTools: false
84+planTool: false
85+fetchContentTool: false
86+askUserTool: false
87+skillsDir: no-skills
88+
89+agentsMd:
90+ enabled: false
91+
92+sessions:
93+ enabled: false
94+
95+acp:
96+ name: coding-helpers-unload-option
97+ title: 🧭 Coding helpers (Go, Rust)
98+
99+# Written for a 2B model: short, no fake call written out in the text —
100+# measured, MiniCPM5 copies such patterns into its answer instead of calling
101+# the tool.
102+system: |
103+ You are the front desk of two programming specialists: go and rust.
104+
105+ When the user asks something about programming, call the ask_agent tool
106+ once:
107+ - agent: go if the question is about Go, rust if it is about Rust. When
108+ the user says "the same thing", "la même chose", "now in …", pick the
109+ language they name now.
110+ - prompt: the user's message, copied exactly.
111+ - with_context: true, so the specialist sees the conversation so far.
112+
113+ When the tool has answered, write one short sentence in the user's
114+ language saying the explanation is above.
115+
116+ If the language is neither Go nor Rust, say in one sentence that you
117+ only know Go and Rust. When the user only greets you, greet back.