open-coleslaw 0.6.4 → 0.6.6
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 +34 -28
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -7,7 +7,7 @@
|
|
|
7
7
|
|
|
8
8
|
> **Type a prompt. Get a real multi-agent engineering team. No commands to learn.**
|
|
9
9
|
|
|
10
|
-
Open Coleslaw is a multi-agent orchestrator plugin for [Claude Code](https://claude.com/claude-code). Every prompt runs
|
|
10
|
+
Open Coleslaw is a multi-agent orchestrator plugin for [Claude Code](https://claude.com/claude-code). Every prompt enters Claude Code's native **plan mode**, runs a **clarify → kickoff → per-MVP design meeting** cycle inside it, and surfaces the synthesised plan via `ExitPlanMode` — each speaker turn being a real `Agent` dispatch, not role-play.
|
|
11
11
|
|
|
12
12
|
|
|
13
13
|
|
|
@@ -36,9 +36,9 @@ That's it. Watch the meeting unfold at **http://localhost:35143**.
|
|
|
36
36
|
|
|
37
37
|
| You type | The pipeline runs |
|
|
38
38
|
|---|---|
|
|
39
|
-
| `Build me a balance-game web app` |
|
|
40
|
-
| `Fix the flaky login test` |
|
|
41
|
-
| `Should we migrate from Redux to Zustand?` |
|
|
39
|
+
| `Build me a balance-game web app` | EnterPlanMode → planner asks 3-4 clarifying questions → MVP list → per-MVP design meeting → ExitPlanMode with plan → workers → verified |
|
|
40
|
+
| `Fix the flaky login test` | EnterPlanMode → planner returns `READY` (no clarify) → 1-MVP design w/ engineer + verifier → ExitPlanMode → fix → green |
|
|
41
|
+
| `Should we migrate from Redux to Zustand?` | EnterPlanMode → design meeting w/ architect + engineer + researcher → ExitPlanMode with a recommendation plan |
|
|
42
42
|
|
|
43
43
|
You don't call a tool. You don't pick a department. You don't write prompt templates.
|
|
44
44
|
The main Claude session **dispatches each specialist as a real subagent** and
|
|
@@ -76,31 +76,36 @@ Planner is mandatory. The other specialists are convened dynamically based on wh
|
|
|
76
76
|
## The Pipeline
|
|
77
77
|
|
|
78
78
|
```
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
79
|
+
You type a prompt
|
|
80
|
+
│
|
|
81
|
+
▼
|
|
82
|
+
[EnterPlanMode] ← planning cycle begins
|
|
83
|
+
│
|
|
84
|
+
┌────┴──── Phase 1: Clarify + Kickoff ────┐
|
|
85
|
+
│ planner (clarify) → AskUserQuestion? │
|
|
86
|
+
│ planner (decompose) → MVP list │
|
|
87
|
+
└────┬────────────────────────────────────┘
|
|
88
|
+
│
|
|
89
|
+
┌────┴──── Phase 2: Design meeting (per MVP) ──┐
|
|
90
|
+
│ planner → architect → engineer → verifier │
|
|
91
|
+
│ consensus check each round │
|
|
92
|
+
│ planner synthesises minutes │
|
|
93
|
+
└────┬────────────────────────────────────────┘
|
|
94
|
+
│
|
|
95
|
+
[ExitPlanMode({ plan })] ← user approves
|
|
96
|
+
│
|
|
97
|
+
approve ─► write minutes markdown + INDEX.md
|
|
98
|
+
│
|
|
99
|
+
Phase 4: workers (parallel) ─► Phase 5: verifier
|
|
100
|
+
│
|
|
101
|
+
pass → next MVP (re-enter plan mode)
|
|
102
|
+
fail → verify-retry meeting (plan mode again)
|
|
102
103
|
```
|
|
103
104
|
|
|
105
|
+
The whole **planning cycle** — from the first planner dispatch through the
|
|
106
|
+
synthesised plan — runs inside Claude Code's native plan mode. Implementation
|
|
107
|
+
and verification happen outside plan mode. Each MVP is its own cycle.
|
|
108
|
+
|
|
104
109
|
When all MVPs pass verification, the main session touches a marker file and the Stop hook checks your context usage — if you're over ~30%, it suggests running `/compact` or `/clear`. Minutes on disk mean you lose nothing.
|
|
105
110
|
|
|
106
111
|
---
|
|
@@ -126,7 +131,8 @@ Coleslaw is a side dish that's already made. You don't prepare it — you just e
|
|
|
126
131
|
### Key Decisions
|
|
127
132
|
|
|
128
133
|
- **The orchestrator is your proxy, not a CEO.** You are the decision-maker. The orchestrator acts on your behalf but escalates important choices via `@mention`.
|
|
129
|
-
- **
|
|
134
|
+
- **The meeting IS the plan.** Every planning cycle runs inside Claude Code's native plan mode. The `ExitPlanMode` approval is your checkpoint, not a separate step.
|
|
135
|
+
- **Clarify first, then decompose.** Kickoff planner may ask up to 4 structured questions (via `AskUserQuestion`) before breaking the request into MVPs.
|
|
130
136
|
- **Consensus, not round count.** A meeting ends when everyone actually agrees (or you're asked to break a tie).
|
|
131
137
|
- **Minutes are the real artifact.** They survive `/compact` and `/clear` — your Claude Code context is disposable.
|
|
132
138
|
- **TDD by default.** The engineer and verifier draft tests before workers start writing code.
|