pipeline-moe 0.1.0 → 0.1.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +44 -6
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -1,9 +1,14 @@
|
|
|
1
1
|
# Pipeline-MoE
|
|
2
2
|
|
|
3
|
-
A
|
|
4
|
-
|
|
3
|
+
**Think frontier, run local.** A multi-agent chat room over N stateful
|
|
4
|
+
[`pi`](https://github.com/earendil-works/pi) sessions, one shared workspace, and
|
|
5
|
+
your own GPU.
|
|
6
|
+
|
|
7
|
+
Orchestrates **N stateful `pi` agent sessions** (one per persona, different
|
|
5
8
|
system prompts + tool sets) over a **shared workspace**, routes `@mentions`
|
|
6
9
|
through a **serial queue**, and streams everything to its clients over **SSE**.
|
|
10
|
+
Plan with a frontier model, execute on a local 27B, swap any agent to cloud
|
|
11
|
+
when the project demands it.
|
|
7
12
|
|
|
8
13
|
Two clients ship with it, built on a shared framework-agnostic core
|
|
9
14
|
(`@pipeline-moe/client-core`):
|
|
@@ -16,6 +21,33 @@ Two clients ship with it, built on a shared framework-agnostic core
|
|
|
16
21
|
Local-first by policy: cloud providers are hidden and rejected unless you
|
|
17
22
|
explicitly opt in with `PIPELINE_ALLOW_CLOUD=1`.
|
|
18
23
|
|
|
24
|
+
### Example Lineup — Tiered Orchestration
|
|
25
|
+
|
|
26
|
+
| Tier | Persona | Model | Where |
|
|
27
|
+
|---|---|---|---|
|
|
28
|
+
| Frontier | planner | Claude Opus 4.6 | API |
|
|
29
|
+
| Frontier | auditor | Claude Sonnet 5 | API |
|
|
30
|
+
| Local | builder, tester, scribe, fetcher, scout | Qwopus 3.6 27B (Q5_K_M) | llama-server, RTX 3090 |
|
|
31
|
+
|
|
32
|
+
The planner and auditor need calibrated reasoning and judgment — they decide
|
|
33
|
+
*what* to build and *whether* it's correct. The execution agents need to follow
|
|
34
|
+
instructions, use tools, and write code — a well-distilled local 27B does this
|
|
35
|
+
at 50 tok/s with zero API cost.
|
|
36
|
+
|
|
37
|
+
If a project exceeds the local model's capabilities, swap a single agent up to
|
|
38
|
+
frontier (e.g. builder → Opus 4.8) without touching the rest of the lineup.
|
|
39
|
+
|
|
40
|
+
### Presets
|
|
41
|
+
|
|
42
|
+
| Preset | Builder | Planner | Auditor | Rest | Cost |
|
|
43
|
+
|---|---|---|---|---|---|
|
|
44
|
+
| `local-default` | Qwopus 27B | Qwopus 27B | Qwopus 27B | Qwopus 27B | $0 |
|
|
45
|
+
| `cloud-sprint` | Qwopus 27B | Opus 4.6 | Sonnet 5 | Qwopus 27B | ~$$ |
|
|
46
|
+
| `cloud-heavy` | Opus 4.8 | Opus 4.6 | Sonnet 5 | Qwopus 27B | ~$$$ |
|
|
47
|
+
|
|
48
|
+
You pay for frontier only on the roles that need it, only when the project
|
|
49
|
+
justifies it. The rest runs free on your GPU.
|
|
50
|
+
|
|
19
51
|
```
|
|
20
52
|
TUI (Ink) Web UI (React + Vite)
|
|
21
53
|
╰──────── @pipeline-moe/client-core ────────╯
|
|
@@ -25,7 +57,10 @@ Express backend ──► Registry of pi AgentSession instances
|
|
|
25
57
|
serial queue scout / builder / auditor / scribe / tester …
|
|
26
58
|
routing @mentions │ each = createAgentSession(persona, tools)
|
|
27
59
|
workspace diff ▼
|
|
28
|
-
|
|
60
|
+
┌────────┴────────┐
|
|
61
|
+
llama-server Cloud APIs
|
|
62
|
+
(any local GGUF) (Anthropic, etc.)
|
|
63
|
+
:5000 on opt-in
|
|
29
64
|
```
|
|
30
65
|
|
|
31
66
|
Each participant is a real `pi` `AgentSession`: it keeps its **own conversation
|
|
@@ -37,9 +72,11 @@ receipt**.
|
|
|
37
72
|
|
|
38
73
|
## Run
|
|
39
74
|
|
|
40
|
-
**From npm** (
|
|
75
|
+
**From npm** (no clone needed):
|
|
41
76
|
```bash
|
|
42
|
-
npx pipeline-moe serve
|
|
77
|
+
npx pipeline-moe serve # server: API + bundled web UI on :5300
|
|
78
|
+
npm i -g @pipeline-moe/tui # terminal client
|
|
79
|
+
pmoe # connect (defaults to localhost:5300)
|
|
43
80
|
```
|
|
44
81
|
|
|
45
82
|
**From source:**
|
|
@@ -56,7 +93,8 @@ bash start.sh # full launch: llama-server → backend → web UI, health-ga
|
|
|
56
93
|
# (set LLAMA_SCRIPT to your llama-server launch script)
|
|
57
94
|
```
|
|
58
95
|
|
|
59
|
-
|
|
96
|
+
|
|
97
|
+
**Terminal client (from source):**
|
|
60
98
|
```bash
|
|
61
99
|
npm -C packages/tui start # connect to localhost:5300
|
|
62
100
|
npm -C packages/tui start -- --server http://host:5300 --room default
|
package/package.json
CHANGED