self-evolve-framework 1.3.0 → 1.4.0
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/package.json +1 -1
- package/template/skills/ponytail/SKILL.md +16 -0
- package/template/skills/ponytail/scripts/hooks/claude-codex-hooks.json +44 -0
- package/template/skills/ponytail/scripts/hooks/copilot-hooks.json +21 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-activate.js +91 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-config.js +122 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-instructions.js +94 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-mode-tracker.js +55 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-runtime.js +68 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-statusline.ps1 +21 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-statusline.sh +12 -0
- package/template/skills/ponytail/scripts/hooks/ponytail-subagent.js +22 -0
- package/template/skills/ponytail/scripts/mcp/README.md +46 -0
- package/template/skills/ponytail/scripts/mcp/index.js +48 -0
- package/template/skills/ponytail/scripts/mcp/instructions.js +26 -0
- package/template/skills/ponytail/scripts/mcp/package.json +13 -0
- package/template/skills/ponytail/scripts/mcp/test/instructions.test.js +22 -0
- package/template/skills/skillopt-sleep/SKILL.md +48 -34
- package/template/skills/skillopt-sleep/scripts/python/__init__.py +20 -0
- package/template/skills/skillopt-sleep/scripts/python/__main__.py +343 -0
- package/template/skills/skillopt-sleep/scripts/python/backend.py +1371 -0
- package/template/skills/skillopt-sleep/scripts/python/budget.py +75 -0
- package/template/skills/skillopt-sleep/scripts/python/config.py +162 -0
- package/template/skills/skillopt-sleep/scripts/python/consolidate.py +238 -0
- package/template/skills/skillopt-sleep/scripts/python/cycle.py +291 -0
- package/template/skills/skillopt-sleep/scripts/python/dream.py +138 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/__init__.py +1 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/gbrain_bench.py +119 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/personas.py +86 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/report.py +132 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/run_experiment.py +178 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/run_gbrain.py +209 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/run_transfer.py +155 -0
- package/template/skills/skillopt-sleep/scripts/python/experiments/sweep.py +164 -0
- package/template/skills/skillopt-sleep/scripts/python/gate.py +50 -0
- package/template/skills/skillopt-sleep/scripts/python/harvest.py +304 -0
- package/template/skills/skillopt-sleep/scripts/python/harvest_codex.py +253 -0
- package/template/skills/skillopt-sleep/scripts/python/harvest_sources.py +41 -0
- package/template/skills/skillopt-sleep/scripts/python/judges.py +84 -0
- package/template/skills/skillopt-sleep/scripts/python/llm_miner.py +134 -0
- package/template/skills/skillopt-sleep/scripts/python/memory.py +129 -0
- package/template/skills/skillopt-sleep/scripts/python/mine.py +312 -0
- package/template/skills/skillopt-sleep/scripts/python/replay.py +146 -0
- package/template/skills/skillopt-sleep/scripts/python/rollout.py +153 -0
- package/template/skills/skillopt-sleep/scripts/python/scheduler.py +138 -0
- package/template/skills/skillopt-sleep/scripts/python/slow_update.py +142 -0
- package/template/skills/skillopt-sleep/scripts/python/staging.py +103 -0
- package/template/skills/skillopt-sleep/scripts/python/state.py +96 -0
- package/template/skills/skillopt-sleep/scripts/python/tasks_file.py +81 -0
- package/template/skills/skillopt-sleep/scripts/python/types.py +146 -0
- package/template/skills/skillopt-sleep/scripts/shell/__init__.py +0 -0
- package/template/skills/skillopt-sleep/scripts/shell/eval_only.py +466 -0
- package/template/skills/skillopt-sleep/scripts/shell/materialize_searchqa.py +148 -0
- package/template/skills/skillopt-sleep/scripts/shell/run_alfworld.sh +60 -0
- package/template/skills/skillopt-sleep/scripts/shell/run_searchqa.sh +40 -0
- package/template/skills/skillopt-sleep/scripts/shell/run_spreadsheetbench.sh +39 -0
- package/template/skills/skillopt-sleep/scripts/shell/train.py +556 -0
|
@@ -0,0 +1,48 @@
|
|
|
1
|
+
#!/usr/bin/env node
|
|
2
|
+
// Ponytail MCP server: serves the lazy-senior-dev ruleset over stdio as a
|
|
3
|
+
// prompt (user-invoked) and a tool (for hosts that pull context via tools).
|
|
4
|
+
// It does NOT replace the always-on adapters; it's the clean option for hosts
|
|
5
|
+
// whose only injection point is the prompt menu (see #70).
|
|
6
|
+
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
|
|
7
|
+
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
|
|
8
|
+
import { z } from "zod";
|
|
9
|
+
|
|
10
|
+
import { MODES, buildInstructions, resolveMode } from "./instructions.js";
|
|
11
|
+
|
|
12
|
+
const server = new McpServer({ name: "ponytail", version: "0.1.0" });
|
|
13
|
+
|
|
14
|
+
const modeArg = z
|
|
15
|
+
.enum(MODES)
|
|
16
|
+
.optional()
|
|
17
|
+
.describe("Ponytail intensity: lite, full, or ultra. Omit for the configured default.");
|
|
18
|
+
|
|
19
|
+
server.registerPrompt(
|
|
20
|
+
"ponytail",
|
|
21
|
+
{
|
|
22
|
+
title: "Ponytail mode",
|
|
23
|
+
description: "Lazy senior dev instructions: YAGNI, stdlib first, the smallest correct change.",
|
|
24
|
+
argsSchema: { mode: modeArg },
|
|
25
|
+
},
|
|
26
|
+
({ mode }) => ({
|
|
27
|
+
messages: [{ role: "user", content: { type: "text", text: buildInstructions(mode) } }],
|
|
28
|
+
}),
|
|
29
|
+
);
|
|
30
|
+
|
|
31
|
+
server.registerTool(
|
|
32
|
+
"ponytail_instructions",
|
|
33
|
+
{
|
|
34
|
+
title: "Ponytail instructions",
|
|
35
|
+
description: "Return the Ponytail ruleset for the given intensity (lite, full, or ultra).",
|
|
36
|
+
inputSchema: { mode: modeArg },
|
|
37
|
+
outputSchema: { mode: z.string(), instructions: z.string() },
|
|
38
|
+
annotations: { readOnlyHint: true, openWorldHint: false },
|
|
39
|
+
},
|
|
40
|
+
({ mode }) => {
|
|
41
|
+
const resolvedMode = resolveMode(mode);
|
|
42
|
+
const instructions = buildInstructions(resolvedMode);
|
|
43
|
+
const structuredContent = { mode: resolvedMode, instructions };
|
|
44
|
+
return { content: [{ type: "text", text: instructions }], structuredContent };
|
|
45
|
+
},
|
|
46
|
+
);
|
|
47
|
+
|
|
48
|
+
await server.connect(new StdioServerTransport());
|
|
@@ -0,0 +1,26 @@
|
|
|
1
|
+
// Pure instruction selection for the Ponytail MCP server. No MCP/SDK imports,
|
|
2
|
+
// so this stays unit-testable on its own. Reuses the same builder the Claude
|
|
3
|
+
// hooks and Pi extension use, so every host emits identical rules.
|
|
4
|
+
import { createRequire } from "node:module";
|
|
5
|
+
|
|
6
|
+
const require = createRequire(import.meta.url);
|
|
7
|
+
const { getPonytailInstructions } = require("../hooks/ponytail-instructions.js");
|
|
8
|
+
const { getDefaultMode, normalizeMode } = require("../hooks/ponytail-config.js");
|
|
9
|
+
|
|
10
|
+
// The three intensities the server offers. "off" has no instructions to serve.
|
|
11
|
+
export const MODES = ["lite", "full", "ultra"];
|
|
12
|
+
|
|
13
|
+
// Resolve a requested mode to a runtime intensity. Unknown, empty, or "off"
|
|
14
|
+
// falls back to the configured default, then to "full".
|
|
15
|
+
// ponytail: keep the surface to these three; "off"/"review" aren't served here.
|
|
16
|
+
export function resolveMode(requested) {
|
|
17
|
+
const asked = normalizeMode(requested);
|
|
18
|
+
if (asked && asked !== "off") return asked;
|
|
19
|
+
|
|
20
|
+
const fallback = normalizeMode(getDefaultMode());
|
|
21
|
+
return fallback && fallback !== "off" ? fallback : "full";
|
|
22
|
+
}
|
|
23
|
+
|
|
24
|
+
export function buildInstructions(requested) {
|
|
25
|
+
return getPonytailInstructions(resolveMode(requested));
|
|
26
|
+
}
|
|
@@ -0,0 +1,13 @@
|
|
|
1
|
+
{
|
|
2
|
+
"name": "ponytail-mcp",
|
|
3
|
+
"version": "4.8.3",
|
|
4
|
+
"description": "MCP server that serves Ponytail's lazy-senior-dev instructions as a prompt and a tool.",
|
|
5
|
+
"private": true,
|
|
6
|
+
"type": "module",
|
|
7
|
+
"license": "MIT",
|
|
8
|
+
"scripts": { "test": "node --test ./test/*.test.js" },
|
|
9
|
+
"dependencies": {
|
|
10
|
+
"@modelcontextprotocol/sdk": "^1.26.0",
|
|
11
|
+
"zod": "^3.23.0"
|
|
12
|
+
}
|
|
13
|
+
}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
import assert from "node:assert/strict";
|
|
2
|
+
import test from "node:test";
|
|
3
|
+
|
|
4
|
+
import { MODES, resolveMode, buildInstructions } from "../instructions.js";
|
|
5
|
+
|
|
6
|
+
test("resolveMode keeps valid intensities", () => {
|
|
7
|
+
for (const mode of MODES) assert.equal(resolveMode(mode), mode);
|
|
8
|
+
});
|
|
9
|
+
|
|
10
|
+
test("resolveMode falls back to a runtime intensity for off/unknown/empty", () => {
|
|
11
|
+
// PONYTAIL_DEFAULT_MODE could be anything in CI, so just assert the contract:
|
|
12
|
+
// never returns "off", "review", or junk — always one of the served modes.
|
|
13
|
+
for (const input of ["off", "review", "nonsense", "", undefined, null]) {
|
|
14
|
+
assert.ok(MODES.includes(resolveMode(input)), `resolveMode(${input}) must be a served mode`);
|
|
15
|
+
}
|
|
16
|
+
});
|
|
17
|
+
|
|
18
|
+
test("buildInstructions returns the ruleset tagged with the resolved mode", () => {
|
|
19
|
+
const text = buildInstructions("ultra");
|
|
20
|
+
assert.match(text, /PONYTAIL MODE ACTIVE/);
|
|
21
|
+
assert.match(text, /ultra/);
|
|
22
|
+
});
|
|
@@ -1,16 +1,26 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: skillopt-sleep
|
|
3
|
-
description: 离线自我进化引擎。分析历史会话、发现重复模式、优化项目 memory
|
|
3
|
+
description: 离线自我进化引擎。分析历史会话、发现重复模式、优化项目 memory 和规则。底层调用 Python 执行引擎(微软 SkillOpt)。
|
|
4
4
|
---
|
|
5
5
|
|
|
6
6
|
# SkillOpt-Sleep — 自我进化引擎
|
|
7
7
|
|
|
8
|
-
>
|
|
8
|
+
> 执行型 skill:AI 提供指令编排,底层 `python -m skillopt_sleep` 驱动 Harvest → Mine → Replay → Consolidate → Stage → Adopt 六阶段工作流。
|
|
9
9
|
|
|
10
|
-
##
|
|
10
|
+
## 工作流程
|
|
11
|
+
|
|
12
|
+
### 前置条件
|
|
13
|
+
|
|
14
|
+
每次会话首次运行前,确保引擎可用:
|
|
15
|
+
|
|
16
|
+
```bash
|
|
17
|
+
cd .codebuddy/skills/skillopt-sleep/scripts/python
|
|
18
|
+
pip install -r requirements.txt 2>/dev/null || echo "依赖已安装"
|
|
19
|
+
```
|
|
11
20
|
|
|
12
21
|
### 阶段 1:Harvest — 收集信息
|
|
13
|
-
|
|
22
|
+
|
|
23
|
+
收集项目数据:
|
|
14
24
|
- 项目文件结构、代码模式
|
|
15
25
|
- 构建/编译错误记录(从 `.codebuddy/memory/`)
|
|
16
26
|
- 已知问题列表
|
|
@@ -19,50 +29,54 @@ AI 从以下来源收集项目数据:
|
|
|
19
29
|
- 前端文件的设计质量报告(如果安装了 Impeccable skill,调用 `impeccable detect` 检查项目前端设计质量)
|
|
20
30
|
|
|
21
31
|
### 阶段 2:Mine — 挖掘模式
|
|
22
|
-
|
|
32
|
+
|
|
33
|
+
分析数据,识别:
|
|
23
34
|
- **重复错误**:同一类 build/lint 错误出现次数
|
|
24
35
|
- **代码异嗅**:循环依赖、过度耦合、重复代码
|
|
25
36
|
- **缺失约束**:重复发生但无规则防范的问题
|
|
26
37
|
- **优化机会**:可简化的逻辑、可合并的文件
|
|
27
|
-
- **★
|
|
28
|
-
- **★
|
|
38
|
+
- **★ 核心边界**:模块间契约接口、关键用户路径和依赖方
|
|
39
|
+
- **★ 设计异嗅**:Impeccable 检测报告中的 P0/P1 反复问题模式
|
|
40
|
+
|
|
41
|
+
### 阶段 3-6:执行引擎运行
|
|
29
42
|
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
-
|
|
34
|
-
|
|
43
|
+
```bash
|
|
44
|
+
python -m .codebuddy.skills.skillopt-sleep.scripts.python \
|
|
45
|
+
--project "$(pwd)" \
|
|
46
|
+
--mode <dry-run|run|adopt> \
|
|
47
|
+
--backend claude
|
|
48
|
+
```
|
|
49
|
+
|
|
50
|
+
- `--mode dry-run` → 阶段 1-3,只输出分析报告,不写文件
|
|
51
|
+
- `--mode run` → 阶段 1-5,输出暂存清单到 `.codebuddy/memory/skillopt-staging/`
|
|
52
|
+
- `--mode adopt` → 阶段 6,采纳暂存清单中的建议
|
|
53
|
+
|
|
54
|
+
### 手动模式(无 Python 引擎时)
|
|
55
|
+
|
|
56
|
+
如果 Python 环境不可用,AI 按以下指令模拟执行:
|
|
35
57
|
|
|
36
|
-
|
|
58
|
+
#### Consolidate — 合并优化
|
|
37
59
|
生成改进建议,只接受通过 Held-out 门控的比例:
|
|
38
60
|
- **新增规则**:重复 ≥2 次且无规则防范的问题
|
|
39
61
|
- **更新规则**:现有规则描述模糊或过时
|
|
40
62
|
- **删除规则**:不再相关或从未被触发的规则
|
|
41
63
|
- **优化 memory**:合并重复的错误模式记录
|
|
42
|
-
- **★ 边界标记**:涉及核心功能路径的删除/修改建议,加 ⚠️
|
|
64
|
+
- **★ 边界标记**:涉及核心功能路径的删除/修改建议,加 ⚠️ 标签
|
|
43
65
|
|
|
44
|
-
|
|
45
|
-
|
|
66
|
+
#### Stage — 暂存
|
|
67
|
+
所有建议写入 `.codebuddy/memory/skillopt-staging/`,不修改活文件。
|
|
46
68
|
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
```
|
|
50
|
-
📋 待采用清单
|
|
51
|
-
✅ 新增 rules/promise-chain-check.mdc → 防止未捕获 async 错误
|
|
52
|
-
✅ 更新 rules/self-evolve.mdc → 增加 cargo check 验证步骤
|
|
53
|
-
❌ 删除 rules/old-node-format.mdc → 节点格式已统一
|
|
54
|
-
```
|
|
69
|
+
#### Adopt — 采用
|
|
70
|
+
列出建议由用户逐条决定是否采纳。
|
|
55
71
|
|
|
56
|
-
##
|
|
72
|
+
## 触发
|
|
57
73
|
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
skillopt-sleep run
|
|
61
|
-
skillopt-sleep
|
|
62
|
-
|
|
74
|
+
| 对话输入 | 作用 |
|
|
75
|
+
|----------|------|
|
|
76
|
+
| `skillopt-sleep dry-run` | 执行阶段 1-3(有引擎则调 Python,无则 AI 模拟) |
|
|
77
|
+
| `skillopt-sleep run` | 执行阶段 1-5 |
|
|
78
|
+
| `skillopt-sleep adopt` | 采纳建议 |
|
|
63
79
|
|
|
64
|
-
##
|
|
80
|
+
## 内核
|
|
65
81
|
|
|
66
|
-
|
|
67
|
-
- **周五总结**:输入 `skillopt-sleep run` 生成周改进提案
|
|
68
|
-
- **出现相同错误 2 次后**:不用调,self-evolve rule 会自动触发单条规则创建
|
|
82
|
+
执行引擎来自 [microsoft/SkillOpt](https://github.com/microsoft/SkillOpt) — 由微软研究院开源的 skill 文档训练框架。
|
|
@@ -0,0 +1,20 @@
|
|
|
1
|
+
"""SkillOpt-Sleep — nightly offline self-evolution for a local Claude agent.
|
|
2
|
+
|
|
3
|
+
A Claude Code plugin engine that gives a user's agent a "sleep cycle":
|
|
4
|
+
harvest the day's real session transcripts, mine recurring tasks, replay
|
|
5
|
+
them offline, and consolidate short-term experience into long-term memory
|
|
6
|
+
(CLAUDE.md) and skills (SKILL.md) behind a SkillOpt validation gate.
|
|
7
|
+
|
|
8
|
+
Synthesizes three ideas:
|
|
9
|
+
* SkillOpt — validation-gated bounded text optimization (this repo)
|
|
10
|
+
* Dreams — offline memory consolidation, input never mutated
|
|
11
|
+
* Sleep — short-term experience -> long-term competence, offline
|
|
12
|
+
|
|
13
|
+
Public entry points:
|
|
14
|
+
* skillopt_sleep.cli — `python -m skillopt_sleep ...`
|
|
15
|
+
* skillopt_sleep.cycle.run_sleep_cycle(...)
|
|
16
|
+
"""
|
|
17
|
+
from __future__ import annotations
|
|
18
|
+
|
|
19
|
+
__all__ = ["__version__"]
|
|
20
|
+
__version__ = "0.1.0"
|
|
@@ -0,0 +1,343 @@
|
|
|
1
|
+
"""SkillOpt-Sleep — command-line interface.
|
|
2
|
+
|
|
3
|
+
python -m skillopt_sleep run # full cycle: harvest->mine->replay->gate->stage
|
|
4
|
+
python -m skillopt_sleep dry-run # same but report only, no staging/adopt
|
|
5
|
+
python -m skillopt_sleep status # show state + latest staged proposal
|
|
6
|
+
python -m skillopt_sleep adopt # apply the latest staged proposal (with backup)
|
|
7
|
+
python -m skillopt_sleep harvest # just print what would be mined (debug)
|
|
8
|
+
|
|
9
|
+
Common flags:
|
|
10
|
+
--project PATH project to evolve (default: cwd)
|
|
11
|
+
--scope all|invoked harvest scope (default: invoked)
|
|
12
|
+
--max-sessions N cap transcript sessions per run
|
|
13
|
+
--max-tasks N cap mined tasks per run
|
|
14
|
+
--target-skill-path PATH explicit live SKILL.md to stage/adopt
|
|
15
|
+
--tasks-file PATH reviewed TaskRecord JSON file to replay instead of harvesting
|
|
16
|
+
--backend mock|claude|codex|copilot
|
|
17
|
+
--source claude|codex|auto
|
|
18
|
+
--model NAME
|
|
19
|
+
--lookback-hours N
|
|
20
|
+
--auto-adopt
|
|
21
|
+
--json machine-readable output
|
|
22
|
+
"""
|
|
23
|
+
from __future__ import annotations
|
|
24
|
+
|
|
25
|
+
import argparse
|
|
26
|
+
import json
|
|
27
|
+
import os
|
|
28
|
+
import sys
|
|
29
|
+
from typing import Any, Dict
|
|
30
|
+
|
|
31
|
+
from skillopt_sleep.config import load_config
|
|
32
|
+
from skillopt_sleep.cycle import run_sleep_cycle
|
|
33
|
+
from skillopt_sleep.harvest_sources import harvest_for_config
|
|
34
|
+
from skillopt_sleep.mine import mine
|
|
35
|
+
from skillopt_sleep.staging import adopt as adopt_staging
|
|
36
|
+
from skillopt_sleep.staging import latest_staging
|
|
37
|
+
from skillopt_sleep.state import SleepState
|
|
38
|
+
from skillopt_sleep.tasks_file import load_tasks_file, make_tasks_payload, write_tasks_file
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
def _read_text(path: str) -> str:
|
|
42
|
+
try:
|
|
43
|
+
with open(path, encoding="utf-8") as f:
|
|
44
|
+
return f.read()
|
|
45
|
+
except Exception:
|
|
46
|
+
return ""
|
|
47
|
+
|
|
48
|
+
|
|
49
|
+
def _report_payload(rep, outcome) -> Dict[str, Any]:
|
|
50
|
+
return {
|
|
51
|
+
"night": rep.night,
|
|
52
|
+
"accepted": rep.accepted,
|
|
53
|
+
"gate_action": rep.gate_action,
|
|
54
|
+
"no_edits_reason": getattr(rep, "no_edits_reason", ""),
|
|
55
|
+
"baseline": rep.baseline_score,
|
|
56
|
+
"candidate": rep.candidate_score,
|
|
57
|
+
"n_tasks": rep.n_tasks,
|
|
58
|
+
"n_sessions": rep.n_sessions,
|
|
59
|
+
"n_accepted_edits": len(rep.edits),
|
|
60
|
+
"n_rejected_edits": len(rep.rejected_edits),
|
|
61
|
+
"edits": [e.__dict__ for e in rep.edits],
|
|
62
|
+
"rejected_edits": [e.__dict__ for e in rep.rejected_edits],
|
|
63
|
+
"notes": rep.notes,
|
|
64
|
+
"staging_dir": outcome.staging_dir,
|
|
65
|
+
"adopted": outcome.adopted,
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
|
|
69
|
+
def _add_common(p: argparse.ArgumentParser) -> None:
|
|
70
|
+
p.add_argument("--project", default="")
|
|
71
|
+
p.add_argument("--scope", default="", choices=["", "all", "invoked"])
|
|
72
|
+
p.add_argument("--backend", default="", choices=["", "mock", "claude", "codex", "copilot"])
|
|
73
|
+
p.add_argument("--model", default="")
|
|
74
|
+
p.add_argument("--codex-path", default="", help="path to the real @openai/codex binary")
|
|
75
|
+
p.add_argument("--claude-home", default="", help="override ~/.claude (also isolates state)")
|
|
76
|
+
p.add_argument("--codex-home", default="", help="override ~/.codex for archived session harvest")
|
|
77
|
+
p.add_argument("--source", default="", choices=["", "claude", "codex", "auto"],
|
|
78
|
+
help="session transcript source")
|
|
79
|
+
p.add_argument("--lookback-hours", type=int, default=None,
|
|
80
|
+
help="harvest window in hours; 0 = scan full history")
|
|
81
|
+
p.add_argument("--edit-budget", type=int, default=0)
|
|
82
|
+
p.add_argument("--max-sessions", type=int, default=0,
|
|
83
|
+
help="cap harvested sessions before mining; default derives from max tasks")
|
|
84
|
+
p.add_argument("--max-tasks", type=int, default=0,
|
|
85
|
+
help="cap mined tasks for this run")
|
|
86
|
+
p.add_argument("--target-skill-path", default="",
|
|
87
|
+
help="explicit live SKILL.md path to evolve/stage/adopt")
|
|
88
|
+
p.add_argument("--tasks-file", default="",
|
|
89
|
+
help="reviewed TaskRecord JSON file to replay instead of harvesting")
|
|
90
|
+
p.add_argument("--progress", action="store_true",
|
|
91
|
+
help="print phase progress to stderr")
|
|
92
|
+
p.add_argument("--auto-adopt", action="store_true")
|
|
93
|
+
p.add_argument("--json", action="store_true")
|
|
94
|
+
|
|
95
|
+
|
|
96
|
+
def _cfg_from_args(args, task_meta: Dict[str, Any] | None = None) -> Any:
|
|
97
|
+
overrides: Dict[str, Any] = {}
|
|
98
|
+
if args.project:
|
|
99
|
+
overrides["invoked_project"] = os.path.abspath(args.project)
|
|
100
|
+
overrides["projects"] = "invoked"
|
|
101
|
+
if args.scope:
|
|
102
|
+
overrides["projects"] = args.scope
|
|
103
|
+
if args.backend:
|
|
104
|
+
overrides["backend"] = args.backend
|
|
105
|
+
if args.model:
|
|
106
|
+
overrides["model"] = args.model
|
|
107
|
+
if getattr(args, "codex_path", ""):
|
|
108
|
+
overrides["codex_path"] = os.path.abspath(args.codex_path)
|
|
109
|
+
if getattr(args, "claude_home", ""):
|
|
110
|
+
overrides["claude_home"] = os.path.abspath(args.claude_home)
|
|
111
|
+
if getattr(args, "codex_home", ""):
|
|
112
|
+
overrides["codex_home"] = os.path.abspath(args.codex_home)
|
|
113
|
+
if getattr(args, "source", ""):
|
|
114
|
+
overrides["transcript_source"] = args.source
|
|
115
|
+
lh = getattr(args, "lookback_hours", None)
|
|
116
|
+
if lh is not None: # --lookback-hours was explicitly passed (0 = full history)
|
|
117
|
+
overrides["lookback_hours"] = lh
|
|
118
|
+
if getattr(args, "edit_budget", 0):
|
|
119
|
+
overrides["edit_budget"] = args.edit_budget
|
|
120
|
+
if getattr(args, "max_sessions", 0):
|
|
121
|
+
overrides["max_sessions_per_night"] = args.max_sessions
|
|
122
|
+
if getattr(args, "max_tasks", 0):
|
|
123
|
+
overrides["max_tasks_per_night"] = args.max_tasks
|
|
124
|
+
target_skill_path = getattr(args, "target_skill_path", "")
|
|
125
|
+
if not target_skill_path and task_meta:
|
|
126
|
+
target_skill_path = str(task_meta.get("target_skill_path") or "")
|
|
127
|
+
if target_skill_path:
|
|
128
|
+
path = os.path.expanduser(target_skill_path)
|
|
129
|
+
if args.project and not os.path.isabs(path):
|
|
130
|
+
path = os.path.join(os.path.abspath(args.project), path)
|
|
131
|
+
overrides["target_skill_path"] = os.path.abspath(path)
|
|
132
|
+
if getattr(args, "progress", False):
|
|
133
|
+
overrides["progress"] = True
|
|
134
|
+
if getattr(args, "auto_adopt", False):
|
|
135
|
+
overrides["auto_adopt"] = True
|
|
136
|
+
return load_config(**overrides)
|
|
137
|
+
|
|
138
|
+
|
|
139
|
+
def cmd_run(args, dry: bool = False) -> int:
|
|
140
|
+
task_meta: Dict[str, Any] = {}
|
|
141
|
+
tasks = None
|
|
142
|
+
if getattr(args, "tasks_file", ""):
|
|
143
|
+
# Load once before config so target_skill_path can default from metadata.
|
|
144
|
+
tasks, task_meta = load_tasks_file(args.tasks_file)
|
|
145
|
+
cfg = _cfg_from_args(args, task_meta=task_meta)
|
|
146
|
+
if getattr(args, "tasks_file", ""):
|
|
147
|
+
tasks, task_meta = load_tasks_file(
|
|
148
|
+
args.tasks_file,
|
|
149
|
+
holdout_fraction=cfg.get("holdout_fraction", 0.34),
|
|
150
|
+
seed=cfg.get("seed", 42),
|
|
151
|
+
)
|
|
152
|
+
if cfg.get("backend", "mock") != "mock" and task_meta.get("reviewed") is not True:
|
|
153
|
+
print(
|
|
154
|
+
"[sleep] refusing real-backend replay from an unreviewed tasks file; "
|
|
155
|
+
"inspect/redact it and set \"reviewed\": true first",
|
|
156
|
+
file=sys.stderr,
|
|
157
|
+
)
|
|
158
|
+
return 2
|
|
159
|
+
outcome = run_sleep_cycle(cfg, seed_tasks=tasks, dry_run=dry)
|
|
160
|
+
rep = outcome.report
|
|
161
|
+
if args.json:
|
|
162
|
+
payload = _report_payload(rep, outcome)
|
|
163
|
+
if task_meta:
|
|
164
|
+
payload["tasks_file"] = task_meta.get("tasks_file", "")
|
|
165
|
+
payload["tasks_reviewed"] = task_meta.get("reviewed", False)
|
|
166
|
+
print(json.dumps(payload, ensure_ascii=False, indent=2))
|
|
167
|
+
else:
|
|
168
|
+
print(f"[sleep] night {rep.night}: {rep.n_sessions} sessions -> {rep.n_tasks} tasks")
|
|
169
|
+
print(f"[sleep] held-out {rep.baseline_score:.3f} -> {rep.candidate_score:.3f} "
|
|
170
|
+
f"=> {rep.gate_action} (accepted={rep.accepted})")
|
|
171
|
+
for e in rep.edits:
|
|
172
|
+
print(f" + [{e.target}/{e.op}] {e.content}")
|
|
173
|
+
if rep.rejected_edits:
|
|
174
|
+
print("[sleep] rejected by gate:")
|
|
175
|
+
for e in rep.rejected_edits:
|
|
176
|
+
print(f" - [{e.target}/{e.op}] {e.content}")
|
|
177
|
+
if outcome.staging_dir:
|
|
178
|
+
print(f"[sleep] staged: {outcome.staging_dir}")
|
|
179
|
+
if not outcome.adopted:
|
|
180
|
+
print("[sleep] review it, then: python -m skillopt_sleep adopt")
|
|
181
|
+
if outcome.adopted:
|
|
182
|
+
print(f"[sleep] auto-adopted: {', '.join(outcome.adopted_paths)}")
|
|
183
|
+
return 0
|
|
184
|
+
|
|
185
|
+
|
|
186
|
+
def cmd_status(args) -> int:
|
|
187
|
+
cfg = _cfg_from_args(args)
|
|
188
|
+
state = SleepState.load(cfg.state_path)
|
|
189
|
+
project = cfg.get("invoked_project") or os.getcwd()
|
|
190
|
+
latest = latest_staging(project)
|
|
191
|
+
info = {
|
|
192
|
+
"night": state.night,
|
|
193
|
+
"state_path": cfg.state_path,
|
|
194
|
+
"project": project,
|
|
195
|
+
"history_tail": state.data.get("history", [])[-5:],
|
|
196
|
+
"latest_staging": latest,
|
|
197
|
+
"slow_memory_chars": len(state.slow_memory),
|
|
198
|
+
}
|
|
199
|
+
if args.json:
|
|
200
|
+
print(json.dumps(info, ensure_ascii=False, indent=2))
|
|
201
|
+
else:
|
|
202
|
+
print(f"[sleep] nights so far: {state.night}")
|
|
203
|
+
print(f"[sleep] project: {project}")
|
|
204
|
+
if latest:
|
|
205
|
+
print(f"[sleep] latest staged proposal: {latest}")
|
|
206
|
+
rp = os.path.join(latest, "report.md")
|
|
207
|
+
if os.path.exists(rp):
|
|
208
|
+
with open(rp) as f:
|
|
209
|
+
print("\n" + f.read())
|
|
210
|
+
else:
|
|
211
|
+
print("[sleep] no staged proposals yet.")
|
|
212
|
+
return 0
|
|
213
|
+
|
|
214
|
+
|
|
215
|
+
def cmd_adopt(args) -> int:
|
|
216
|
+
cfg = _cfg_from_args(args)
|
|
217
|
+
project = cfg.get("invoked_project") or os.getcwd()
|
|
218
|
+
target = args.staging or latest_staging(project)
|
|
219
|
+
if not target or not os.path.isdir(target):
|
|
220
|
+
print("[sleep] nothing to adopt (no staging dir).")
|
|
221
|
+
return 1
|
|
222
|
+
updated = adopt_staging(target)
|
|
223
|
+
print(f"[sleep] adopted from {target}")
|
|
224
|
+
for p in updated:
|
|
225
|
+
print(f" -> {p}")
|
|
226
|
+
if not updated:
|
|
227
|
+
print("[sleep] (proposal contained no accepted changes)")
|
|
228
|
+
return 0
|
|
229
|
+
|
|
230
|
+
|
|
231
|
+
def cmd_harvest(args) -> int:
|
|
232
|
+
cfg = _cfg_from_args(args)
|
|
233
|
+
session_limit = cfg.get("max_sessions_per_night", 0) or cfg.get("max_tasks_per_night", 40) * 3
|
|
234
|
+
target_skill_path = cfg.managed_skill_path() if cfg.get("target_skill_path", "") else ""
|
|
235
|
+
target_skill_text = _read_text(target_skill_path) if target_skill_path else ""
|
|
236
|
+
max_tasks = cfg.get("max_tasks_per_night", 40)
|
|
237
|
+
candidate_limit = max_tasks
|
|
238
|
+
if cfg.get("target_task_filter", True) and target_skill_text:
|
|
239
|
+
candidate_limit = max(max_tasks, max_tasks * 3)
|
|
240
|
+
digests = harvest_for_config(cfg, limit=session_limit)
|
|
241
|
+
tasks = mine(
|
|
242
|
+
digests,
|
|
243
|
+
max_tasks=max_tasks,
|
|
244
|
+
candidate_limit=candidate_limit,
|
|
245
|
+
holdout_fraction=cfg.get("holdout_fraction", 0.34),
|
|
246
|
+
seed=cfg.get("seed", 42),
|
|
247
|
+
target_skill_text=target_skill_text,
|
|
248
|
+
target_skill_path=target_skill_path,
|
|
249
|
+
)
|
|
250
|
+
payload = make_tasks_payload(
|
|
251
|
+
tasks,
|
|
252
|
+
project=cfg.get("invoked_project") or os.getcwd(),
|
|
253
|
+
transcript_source=cfg.get("transcript_source", ""),
|
|
254
|
+
n_sessions=len(digests),
|
|
255
|
+
target_skill_path=target_skill_path,
|
|
256
|
+
)
|
|
257
|
+
output_path = ""
|
|
258
|
+
if getattr(args, "output", ""):
|
|
259
|
+
output_path = write_tasks_file(args.output, payload)
|
|
260
|
+
if args.json:
|
|
261
|
+
json_payload = dict(payload)
|
|
262
|
+
if output_path:
|
|
263
|
+
json_payload["output"] = output_path
|
|
264
|
+
print(json.dumps(json_payload, ensure_ascii=False, indent=2))
|
|
265
|
+
else:
|
|
266
|
+
print(f"[sleep] {len(digests)} sessions -> {len(tasks)} tasks")
|
|
267
|
+
if output_path:
|
|
268
|
+
print(f"[sleep] wrote reviewed-task draft: {output_path}")
|
|
269
|
+
for t in tasks:
|
|
270
|
+
print(f" [{t.split}/{t.outcome}] {t.intent[:90]}")
|
|
271
|
+
return 0
|
|
272
|
+
|
|
273
|
+
|
|
274
|
+
def cmd_schedule(args) -> int:
|
|
275
|
+
from skillopt_sleep.scheduler import schedule, list_scheduled
|
|
276
|
+
cfg = _cfg_from_args(args)
|
|
277
|
+
project = cfg.get("invoked_project") or os.getcwd()
|
|
278
|
+
ok, msg = schedule(project, backend=cfg.get("backend", "mock"),
|
|
279
|
+
hour=args.hour, minute=args.minute,
|
|
280
|
+
extra=("--auto-adopt" if getattr(args, "auto_adopt", False) else ""))
|
|
281
|
+
print("[sleep] " + msg)
|
|
282
|
+
cur = list_scheduled()
|
|
283
|
+
if cur:
|
|
284
|
+
print("[sleep] currently scheduled:")
|
|
285
|
+
for ln in cur:
|
|
286
|
+
print(" " + ln[:140])
|
|
287
|
+
return 0 if ok else 1
|
|
288
|
+
|
|
289
|
+
|
|
290
|
+
def cmd_unschedule(args) -> int:
|
|
291
|
+
from skillopt_sleep.scheduler import unschedule
|
|
292
|
+
cfg = _cfg_from_args(args)
|
|
293
|
+
project = cfg.get("invoked_project") or os.getcwd()
|
|
294
|
+
ok, msg = unschedule(project, all_projects=getattr(args, "all", False))
|
|
295
|
+
print("[sleep] " + msg)
|
|
296
|
+
return 0 if ok else 1
|
|
297
|
+
|
|
298
|
+
|
|
299
|
+
def main(argv=None) -> int:
|
|
300
|
+
parser = argparse.ArgumentParser(prog="skillopt_sleep", description="SkillOpt-Sleep nightly self-evolution")
|
|
301
|
+
sub = parser.add_subparsers(dest="cmd", required=True)
|
|
302
|
+
|
|
303
|
+
p_run = sub.add_parser("run", help="run a full sleep cycle")
|
|
304
|
+
_add_common(p_run)
|
|
305
|
+
p_dry = sub.add_parser("dry-run", help="harvest+mine+replay, report only")
|
|
306
|
+
_add_common(p_dry)
|
|
307
|
+
p_status = sub.add_parser("status", help="show state + latest proposal")
|
|
308
|
+
_add_common(p_status)
|
|
309
|
+
p_adopt = sub.add_parser("adopt", help="apply latest staged proposal")
|
|
310
|
+
_add_common(p_adopt)
|
|
311
|
+
p_adopt.add_argument("--staging", default="", help="specific staging dir")
|
|
312
|
+
p_harvest = sub.add_parser("harvest", help="debug: show mined tasks")
|
|
313
|
+
_add_common(p_harvest)
|
|
314
|
+
p_harvest.add_argument("--output", default="", help="write mined tasks JSON for review")
|
|
315
|
+
p_sched = sub.add_parser("schedule", help="install a nightly cron entry for this project")
|
|
316
|
+
_add_common(p_sched)
|
|
317
|
+
p_sched.add_argument("--hour", type=int, default=3)
|
|
318
|
+
p_sched.add_argument("--minute", type=int, default=17)
|
|
319
|
+
p_unsched = sub.add_parser("unschedule", help="remove the nightly cron entry")
|
|
320
|
+
_add_common(p_unsched)
|
|
321
|
+
p_unsched.add_argument("--all", action="store_true", help="remove all managed entries")
|
|
322
|
+
|
|
323
|
+
args = parser.parse_args(argv)
|
|
324
|
+
if args.cmd == "run":
|
|
325
|
+
return cmd_run(args, dry=False)
|
|
326
|
+
if args.cmd == "dry-run":
|
|
327
|
+
return cmd_run(args, dry=True)
|
|
328
|
+
if args.cmd == "status":
|
|
329
|
+
return cmd_status(args)
|
|
330
|
+
if args.cmd == "adopt":
|
|
331
|
+
return cmd_adopt(args)
|
|
332
|
+
if args.cmd == "harvest":
|
|
333
|
+
return cmd_harvest(args)
|
|
334
|
+
if args.cmd == "schedule":
|
|
335
|
+
return cmd_schedule(args)
|
|
336
|
+
if args.cmd == "unschedule":
|
|
337
|
+
return cmd_unschedule(args)
|
|
338
|
+
parser.print_help()
|
|
339
|
+
return 2
|
|
340
|
+
|
|
341
|
+
|
|
342
|
+
if __name__ == "__main__":
|
|
343
|
+
sys.exit(main())
|