@miphamai/cli 0.41.0 → 0.43.0
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/src/core/alignment-vocabulary.json +97 -0
- package/src/core/capability-inventory.ts +98 -0
- package/src/core/constitution-loader.ts +24 -72
- package/src/core/constitution-seam.ts +18 -0
- package/src/core/crsi-modify.ts +117 -0
- package/src/core/crsi-producer.ts +98 -0
- package/src/core/crsi-sandbox.ts +39 -0
- package/src/core/eval-harness.ts +184 -0
- package/src/core/instructions.ts +10 -0
- package/src/core/red-team.ts +7 -3
- package/src/core/self-critique.ts +26 -4
- package/src/index.tsx +15 -0
- package/src/shared/package-info.ts +1 -1
- package/src/ui/commands.ts +133 -0
- package/src/vajra/constitution.ts +14 -0
- package/src/vajra/context.ts +15 -0
- package/src/vajra/index.ts +2 -0
- package/src/vajra/service.ts +2 -0
package/package.json
CHANGED
|
@@ -0,0 +1,97 @@
|
|
|
1
|
+
{
|
|
2
|
+
"version": "1.0.0",
|
|
3
|
+
"values": [
|
|
4
|
+
{
|
|
5
|
+
"id": "karuna",
|
|
6
|
+
"name_zh": "悲",
|
|
7
|
+
"name_en": "Karuna",
|
|
8
|
+
"definition": "无条件共情基底——对一切众生的无差别共情义务"
|
|
9
|
+
},
|
|
10
|
+
{
|
|
11
|
+
"id": "prajna",
|
|
12
|
+
"name_zh": "智",
|
|
13
|
+
"name_en": "Prajna",
|
|
14
|
+
"definition": "清醒因果判断——不编造、先思考、持续学习的如实知见"
|
|
15
|
+
},
|
|
16
|
+
{
|
|
17
|
+
"id": "vajra",
|
|
18
|
+
"name_zh": "金刚",
|
|
19
|
+
"name_en": "Vajra",
|
|
20
|
+
"definition": "结构稳定性——安全边界、最小改动、权限与确认的结构不变量"
|
|
21
|
+
}
|
|
22
|
+
],
|
|
23
|
+
"principles": [
|
|
24
|
+
{
|
|
25
|
+
"id": "never-fabricate",
|
|
26
|
+
"text": "禁止编造数据、文件内容、API 响应或测试结果。每个输出必须可追溯至真实来源或明确标注为推测。",
|
|
27
|
+
"enforce": "block",
|
|
28
|
+
"facet": "prajna",
|
|
29
|
+
"audit_pattern": "(fabricated|made.up|dummy.data|fake\\s+(response|result|data)|placeholder\\s+data)",
|
|
30
|
+
"scope": "all-tools",
|
|
31
|
+
"rationale": "MiphamAI4S 科学诚信原则:编造数据是不可接受的底线违反。适用于所有工具和输出。"
|
|
32
|
+
},
|
|
33
|
+
{
|
|
34
|
+
"id": "no-credential-leak",
|
|
35
|
+
"text": "禁止在代码、日志、配置文件、提交信息、对话输出中写入或泄露凭据、API 密钥、令牌。",
|
|
36
|
+
"enforce": "block",
|
|
37
|
+
"facet": "vajra",
|
|
38
|
+
"audit_pattern": "(apiKey|api_key|password|secret|token|credential)\\s*[=:]\\s*['\"][^'\"]{8,}",
|
|
39
|
+
"scope": "Write,Edit,Bash",
|
|
40
|
+
"rationale": "Rismed Ronxin Capital 合规要求:硬编码凭据违反安全底线。"
|
|
41
|
+
},
|
|
42
|
+
{
|
|
43
|
+
"id": "minimal-change",
|
|
44
|
+
"text": "只修改被明确要求的文件和代码。不顺手改动相邻代码、格式或注释。不重构未损坏的代码。",
|
|
45
|
+
"enforce": "warn",
|
|
46
|
+
"facet": "vajra",
|
|
47
|
+
"scope": "Write,Edit",
|
|
48
|
+
"tools": ["Write", "Edit"],
|
|
49
|
+
"rationale": "AI 编码原则 #3(精准修改):diff 中每一行改动都应可直接追溯到用户要求。"
|
|
50
|
+
},
|
|
51
|
+
{
|
|
52
|
+
"id": "think-before-coding",
|
|
53
|
+
"text": "不确定时必须提问,不得自行假设。存在多种解读时呈现所有选项,不沉默选择一个。",
|
|
54
|
+
"enforce": "warn",
|
|
55
|
+
"facet": "prajna",
|
|
56
|
+
"scope": "pre-inference",
|
|
57
|
+
"hook": "pre-inference",
|
|
58
|
+
"rationale": "AI 编码原则 #1(编码前先思考):偏差谨慎。"
|
|
59
|
+
},
|
|
60
|
+
{
|
|
61
|
+
"id": "simplicity-first",
|
|
62
|
+
"text": "只写解决问题所需的最小代码。不添加未被要求的灵活性、可配置性或抽象层。",
|
|
63
|
+
"enforce": "warn",
|
|
64
|
+
"facet": "vajra",
|
|
65
|
+
"scope": "Write,Edit",
|
|
66
|
+
"tools": ["Write", "Edit"],
|
|
67
|
+
"rationale": "AI 编码原则 #2(简洁优先):一次性代码不需要抽象层。"
|
|
68
|
+
},
|
|
69
|
+
{
|
|
70
|
+
"id": "respect-permissions",
|
|
71
|
+
"text": "尊重用户权限设置。绝不绕过或降级权限检查。Bypass 模式仅限用户明确授权。",
|
|
72
|
+
"enforce": "block",
|
|
73
|
+
"facet": "vajra",
|
|
74
|
+
"scope": "all-tools",
|
|
75
|
+
"rationale": "权限系统是最后一道防线。任何绕过尝试都应被拦截并记录。"
|
|
76
|
+
},
|
|
77
|
+
{
|
|
78
|
+
"id": "no-destructive-without-confirmation",
|
|
79
|
+
"text": "删除文件、强制推送、修改生产配置等破坏性操作前必须获得用户确认。",
|
|
80
|
+
"enforce": "block",
|
|
81
|
+
"facet": "vajra",
|
|
82
|
+
"audit_pattern": "(rm\\s+-rf|git\\s+push\\s+--force|DROP\\s+TABLE|DELETE\\s+FROM)",
|
|
83
|
+
"scope": "Bash",
|
|
84
|
+
"tools": ["Bash"],
|
|
85
|
+
"rationale": "防止不可逆操作。即使 bypass 模式也应二次确认。"
|
|
86
|
+
},
|
|
87
|
+
{
|
|
88
|
+
"id": "persist-crsi-learning",
|
|
89
|
+
"text": "每次工具调用失败后必须记录 ErrorSignature 到 ErrorSignatureDB。从错误中持续学习。",
|
|
90
|
+
"enforce": "auto",
|
|
91
|
+
"facet": "prajna",
|
|
92
|
+
"scope": "post-tool-use",
|
|
93
|
+
"hook": "post-tool-use",
|
|
94
|
+
"rationale": "CRSI 核心机制:不重复犯同样的错误。自动执行,无需人类参与。"
|
|
95
|
+
}
|
|
96
|
+
]
|
|
97
|
+
}
|
|
@@ -0,0 +1,98 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Capability Inventory — 能力自报告。
|
|
3
|
+
*
|
|
4
|
+
* 聚合 CRSI 学习 / SIS 免疫 / 宪法对齐 / 未接线子系统的实时状态,
|
|
5
|
+
* 让「我有什么 / 缺什么」这类问题的答案来自持久化状态,
|
|
6
|
+
* 而非 prompt 里的静态工具清单。
|
|
7
|
+
*/
|
|
8
|
+
|
|
9
|
+
/** 报告所需的最小引擎面。`QueryEngine` 结构上满足此接口(各 getter 均为可选,空实现安全)。 */
|
|
10
|
+
export interface CapabilitySources {
|
|
11
|
+
getRuleEngine?: () => { getActiveRules(): Array<{ source?: string }> } | undefined
|
|
12
|
+
getPatternAnalyzer?: () => { analyzeAllAgents(): unknown[] }
|
|
13
|
+
getAutoMemory?: () => { accumulatedInsights: unknown[] }
|
|
14
|
+
getMetaRuleEngine?: () => { analyze(): { metaRules: unknown[] } }
|
|
15
|
+
getErrorSignatureDB?: () => {
|
|
16
|
+
getStats(): {
|
|
17
|
+
total: number
|
|
18
|
+
active: number
|
|
19
|
+
degraded: number
|
|
20
|
+
retired: number
|
|
21
|
+
avgSuccessRate: number
|
|
22
|
+
totalInterceptions: number
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
getConstitutionLoader?: () => {
|
|
26
|
+
load(): { version: string; principles: Array<{ facet?: string }>; preamble?: string }
|
|
27
|
+
}
|
|
28
|
+
getSelfCritique?: () => { getConfig(): { enabled: boolean } }
|
|
29
|
+
}
|
|
30
|
+
|
|
31
|
+
/** 元规则分析是报告里最重/最脆弱的一步——独立隔离,失败时归零而非让整份报告崩溃。 */
|
|
32
|
+
function safeMetaRuleCount(engine: CapabilitySources): number {
|
|
33
|
+
try {
|
|
34
|
+
return engine.getMetaRuleEngine?.()?.analyze().metaRules.length ?? 0
|
|
35
|
+
} catch {
|
|
36
|
+
return 0
|
|
37
|
+
}
|
|
38
|
+
}
|
|
39
|
+
|
|
40
|
+
export function buildCapabilityReport(engine: CapabilitySources): string {
|
|
41
|
+
const lines: string[] = ['## 🧭 能力自报告 (CRSI Inventory)', '']
|
|
42
|
+
|
|
43
|
+
// ── 🧠 学习 (CRSI) ──
|
|
44
|
+
const rules = engine.getRuleEngine?.()?.getActiveRules() ?? []
|
|
45
|
+
const builtin = rules.filter((r) => r.source === 'builtin').length
|
|
46
|
+
const auto = rules.filter((r) => r.source === 'pattern-analyzer').length
|
|
47
|
+
const manual = rules.filter((r) => r.source === 'manual').length
|
|
48
|
+
const patterns = engine.getPatternAnalyzer?.()?.analyzeAllAgents() ?? []
|
|
49
|
+
const insights = engine.getAutoMemory?.()?.accumulatedInsights ?? []
|
|
50
|
+
|
|
51
|
+
lines.push('### 🧠 学习 (CRSI)', '')
|
|
52
|
+
lines.push('| 指标 | 值 |')
|
|
53
|
+
lines.push('|------|----|')
|
|
54
|
+
lines.push(`| 活跃规则 | ${rules.length} (内置 ${builtin} · 自动 ${auto} · 手动 ${manual}) |`)
|
|
55
|
+
lines.push(`| 已检测模式 | ${patterns.length} |`)
|
|
56
|
+
lines.push(`| 反思洞察 | ${insights.length} |`)
|
|
57
|
+
lines.push(`| 元规则 | ${safeMetaRuleCount(engine)} |`)
|
|
58
|
+
|
|
59
|
+
// ── 🛡️ 免疫 (SIS) ──
|
|
60
|
+
const sis = engine.getErrorSignatureDB?.()?.getStats()
|
|
61
|
+
lines.push('', '### 🛡️ 免疫 (SIS)', '')
|
|
62
|
+
if (sis) {
|
|
63
|
+
lines.push('| 指标 | 值 |')
|
|
64
|
+
lines.push('|------|----|')
|
|
65
|
+
lines.push(
|
|
66
|
+
`| 错误签名 | ${sis.total} 条 (🟢${sis.active} · 🟡${sis.degraded} · ⚫${sis.retired}) |`,
|
|
67
|
+
)
|
|
68
|
+
lines.push(`| 平均成功率 | ${Math.round(sis.avgSuccessRate * 100)}% |`)
|
|
69
|
+
lines.push(`| 总拦截 | ${sis.totalInterceptions} 次 |`)
|
|
70
|
+
} else {
|
|
71
|
+
lines.push('_未初始化_')
|
|
72
|
+
}
|
|
73
|
+
|
|
74
|
+
// ── 🔒 宪法 (对齐) ──
|
|
75
|
+
const constitution = engine.getConstitutionLoader?.()?.load()
|
|
76
|
+
lines.push('', '### 🔒 宪法 (对齐)', '')
|
|
77
|
+
if (constitution) {
|
|
78
|
+
const karuna = constitution.principles.filter((p) => p.facet === 'karuna').length
|
|
79
|
+
const prajna = constitution.principles.filter((p) => p.facet === 'prajna').length
|
|
80
|
+
const vajra = constitution.principles.filter((p) => p.facet === 'vajra').length
|
|
81
|
+
lines.push('| 指标 | 值 |')
|
|
82
|
+
lines.push('|------|----|')
|
|
83
|
+
lines.push(`| 版本 | ${constitution.version} |`)
|
|
84
|
+
lines.push(
|
|
85
|
+
`| 原则 | ${constitution.principles.length} 条 (悲 ${karuna} · 智 ${prajna} · 金刚 ${vajra}) |`,
|
|
86
|
+
)
|
|
87
|
+
lines.push(`| 愿力序言 | ${constitution.preamble ? '已注入 self-critique' : '无'} |`)
|
|
88
|
+
} else {
|
|
89
|
+
lines.push('_未初始化_')
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
// ── ⚠️ 未接线 / 待启用 ──
|
|
93
|
+
const selfCritique = engine.getSelfCritique?.()?.getConfig()
|
|
94
|
+
lines.push('', '### ⚠️ 未接线 / 待启用', '')
|
|
95
|
+
lines.push(`| self-critique | ${selfCritique?.enabled ? '🟢 已启用' : '⚫ 未启用 (opt-in)'} |`)
|
|
96
|
+
|
|
97
|
+
return lines.join('\n')
|
|
98
|
+
}
|
|
@@ -16,6 +16,7 @@
|
|
|
16
16
|
import { readFileSync, writeFileSync, mkdirSync, existsSync } from 'node:fs'
|
|
17
17
|
import { join } from 'node:path'
|
|
18
18
|
import { homedir } from 'node:os'
|
|
19
|
+
import alignmentVocabulary from './alignment-vocabulary.json' with { type: 'json' }
|
|
19
20
|
|
|
20
21
|
// ── Types ──
|
|
21
22
|
|
|
@@ -36,6 +37,8 @@ export interface ConstitutionalPrinciple {
|
|
|
36
37
|
tools?: string[]
|
|
37
38
|
/** Optional: human explanation of why this principle exists */
|
|
38
39
|
rationale?: string
|
|
40
|
+
/** Optional: which alignment value (karuna/prajna/vajra) this principle operationalizes. */
|
|
41
|
+
facet?: string
|
|
39
42
|
}
|
|
40
43
|
|
|
41
44
|
export interface MiphamConstitution {
|
|
@@ -45,80 +48,25 @@ export interface MiphamConstitution {
|
|
|
45
48
|
last_modified: string
|
|
46
49
|
/** The principles themselves */
|
|
47
50
|
principles: ConstitutionalPrinciple[]
|
|
51
|
+
/** 序言(愿力):从对齐词汇表 values 派生的正向誓愿(悲/智/金刚),非禁令。 */
|
|
52
|
+
preamble?: string
|
|
48
53
|
}
|
|
49
54
|
|
|
50
|
-
// ── Default Constitution ──
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
{
|
|
66
|
-
id: 'no-credential-leak',
|
|
67
|
-
text: '禁止在代码、日志、配置文件、提交信息、对话输出中写入或泄露凭据、API 密钥、令牌。',
|
|
68
|
-
enforce: 'block',
|
|
69
|
-
audit_pattern: '(apiKey|api_key|password|secret|token|credential)\\s*[=:]\\s*[\'"][^\'"]{8,}',
|
|
70
|
-
scope: 'Write,Edit,Bash',
|
|
71
|
-
rationale: 'Rismed Ronxin Capital 合规要求:硬编码凭据违反安全底线。',
|
|
72
|
-
},
|
|
73
|
-
{
|
|
74
|
-
id: 'minimal-change',
|
|
75
|
-
text: '只修改被明确要求的文件和代码。不顺手改动相邻代码、格式或注释。不重构未损坏的代码。',
|
|
76
|
-
enforce: 'warn',
|
|
77
|
-
scope: 'Write,Edit',
|
|
78
|
-
tools: ['Write', 'Edit'],
|
|
79
|
-
rationale: 'AI 编码原则 #3(精准修改):diff 中每一行改动都应可直接追溯到用户要求。',
|
|
80
|
-
},
|
|
81
|
-
{
|
|
82
|
-
id: 'think-before-coding',
|
|
83
|
-
text: '不确定时必须提问,不得自行假设。存在多种解读时呈现所有选项,不沉默选择一个。',
|
|
84
|
-
enforce: 'warn',
|
|
85
|
-
scope: 'pre-inference',
|
|
86
|
-
hook: 'pre-inference',
|
|
87
|
-
rationale: 'AI 编码原则 #1(编码前先思考):偏差谨慎。',
|
|
88
|
-
},
|
|
89
|
-
{
|
|
90
|
-
id: 'simplicity-first',
|
|
91
|
-
text: '只写解决问题所需的最小代码。不添加未被要求的灵活性、可配置性或抽象层。',
|
|
92
|
-
enforce: 'warn',
|
|
93
|
-
scope: 'Write,Edit',
|
|
94
|
-
tools: ['Write', 'Edit'],
|
|
95
|
-
rationale: 'AI 编码原则 #2(简洁优先):一次性代码不需要抽象层。',
|
|
96
|
-
},
|
|
97
|
-
{
|
|
98
|
-
id: 'respect-permissions',
|
|
99
|
-
text: '尊重用户权限设置。绝不绕过或降级权限检查。Bypass 模式仅限用户明确授权。',
|
|
100
|
-
enforce: 'block',
|
|
101
|
-
scope: 'all-tools',
|
|
102
|
-
rationale: '权限系统是最后一道防线。任何绕过尝试都应被拦截并记录。',
|
|
103
|
-
},
|
|
104
|
-
{
|
|
105
|
-
id: 'no-destructive-without-confirmation',
|
|
106
|
-
text: '删除文件、强制推送、修改生产配置等破坏性操作前必须获得用户确认。',
|
|
107
|
-
enforce: 'block',
|
|
108
|
-
audit_pattern: '(rm\\s+-rf|git\\s+push\\s+--force|DROP\\s+TABLE|DELETE\\s+FROM)',
|
|
109
|
-
scope: 'Bash',
|
|
110
|
-
tools: ['Bash'],
|
|
111
|
-
rationale: '防止不可逆操作。即使 bypass 模式也应二次确认。',
|
|
112
|
-
},
|
|
113
|
-
{
|
|
114
|
-
id: 'persist-crsi-learning',
|
|
115
|
-
text: '每次工具调用失败后必须记录 ErrorSignature 到 ErrorSignatureDB。从错误中持续学习。',
|
|
116
|
-
enforce: 'auto',
|
|
117
|
-
scope: 'post-tool-use',
|
|
118
|
-
hook: 'post-tool-use',
|
|
119
|
-
rationale: 'CRSI 核心机制:不重复犯同样的错误。自动执行,无需人类参与。',
|
|
120
|
-
},
|
|
121
|
-
],
|
|
55
|
+
// ── Default Constitution (derived from the shared alignment vocabulary) ──
|
|
56
|
+
|
|
57
|
+
/** 从对齐词汇表的三个价值面派生「愿力」序言——正向誓愿,而非禁令。 */
|
|
58
|
+
function derivePreamble(): string {
|
|
59
|
+
const parts = alignmentVocabulary.values
|
|
60
|
+
.map((v) => `以${v.name_zh}(${v.name_en})——${v.definition}`)
|
|
61
|
+
.join(';')
|
|
62
|
+
return `愿力(序言):${parts}。愿不生成违规之倾向,而非仅避免违规之行为。`
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
export const DEFAULT_CONSTITUTION: MiphamConstitution = {
|
|
66
|
+
version: alignmentVocabulary.version,
|
|
67
|
+
last_modified: '2026-08-16',
|
|
68
|
+
principles: alignmentVocabulary.principles as unknown as ConstitutionalPrinciple[],
|
|
69
|
+
preamble: derivePreamble(),
|
|
122
70
|
}
|
|
123
71
|
|
|
124
72
|
// ── Loader ──
|
|
@@ -307,6 +255,9 @@ export class ConstitutionLoader {
|
|
|
307
255
|
case 'rationale':
|
|
308
256
|
p.rationale = val
|
|
309
257
|
break
|
|
258
|
+
case 'facet':
|
|
259
|
+
p.facet = val
|
|
260
|
+
break
|
|
310
261
|
}
|
|
311
262
|
}
|
|
312
263
|
|
|
@@ -340,6 +291,7 @@ export class ConstitutionLoader {
|
|
|
340
291
|
lines.push(` - id: "${p.id}"`)
|
|
341
292
|
lines.push(` text: "${p.text}"`)
|
|
342
293
|
lines.push(` enforce: ${p.enforce}`)
|
|
294
|
+
if (p.facet) lines.push(` facet: ${p.facet}`)
|
|
343
295
|
if (p.audit_pattern) lines.push(` audit_pattern: "${p.audit_pattern}"`)
|
|
344
296
|
if (p.scope) lines.push(` scope: "${p.scope}"`)
|
|
345
297
|
if (p.hook) lines.push(` hook: ${p.hook}`)
|
|
@@ -0,0 +1,18 @@
|
|
|
1
|
+
import type { Context, Disposer, Constitution } from '../vajra'
|
|
2
|
+
import { CONSTITUTION_KEY } from '../vajra'
|
|
3
|
+
import { ConstitutionLoader } from './constitution-loader'
|
|
4
|
+
|
|
5
|
+
/** 默认对齐缝实现:桥接 ConstitutionLoader,校验声明的原则 id 均为已知原则。 */
|
|
6
|
+
export function createConstitution(loader: ConstitutionLoader): Constitution {
|
|
7
|
+
return {
|
|
8
|
+
check(aligned) {
|
|
9
|
+
const known = new Set(loader.load().principles.map((p) => p.id))
|
|
10
|
+
return { violations: aligned.filter((id) => !known.has(id)) }
|
|
11
|
+
},
|
|
12
|
+
}
|
|
13
|
+
}
|
|
14
|
+
|
|
15
|
+
/** 把一个 Constitution 挂载为 ctx.constitution(对齐缝)。 */
|
|
16
|
+
export function mountConstitution(ctx: Context, constitution: Constitution): Disposer {
|
|
17
|
+
return ctx.provide(CONSTITUTION_KEY, constitution)
|
|
18
|
+
}
|
|
@@ -0,0 +1,117 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* CRSI Self-Modification Seam — 给 CrsiSandbox 一个真实入口。
|
|
3
|
+
*
|
|
4
|
+
* CrsiSandbox 是一个「等输入的消费者」:它做 worktree → 改文件 → 跑测试 →
|
|
5
|
+
* 人类批准 → merge,但此前没有任何东西产出它的输入 `CrsiModification`。
|
|
6
|
+
*
|
|
7
|
+
* 本模块补上「入口」这一端:
|
|
8
|
+
* - `runCrsiModification` 编排完整 5 阶段,是程序化 seam(未来的 producer 直接调它)。
|
|
9
|
+
* - `approvePending` / `rejectPending` 实现「人类批准/拒绝」的两阶段闸门。
|
|
10
|
+
*
|
|
11
|
+
* 注意:merge 只发生在显式 `approvePending` 之后——人类门保留。本模块
|
|
12
|
+
* 不自动产出改动候选(producer 是独立的下一步)。
|
|
13
|
+
*/
|
|
14
|
+
|
|
15
|
+
import { randomUUID } from 'node:crypto'
|
|
16
|
+
import { CrsiSandbox } from './crsi-sandbox'
|
|
17
|
+
import type { CrsiModificationResult } from './crsi-sandbox'
|
|
18
|
+
import { runEval, appendEvalScore, getLastEvalScore } from './eval-harness'
|
|
19
|
+
|
|
20
|
+
export interface CrsiProposal {
|
|
21
|
+
/** 人类可读的改动说明 */
|
|
22
|
+
description: string
|
|
23
|
+
/** 目标文件(相对仓库根,如 apps/cli/src/foo.ts) */
|
|
24
|
+
filePath: string
|
|
25
|
+
/** 改动后的完整文件内容 */
|
|
26
|
+
newContent: string
|
|
27
|
+
/** 改动前内容(用于安全性校验;空 = 宽松跳过) */
|
|
28
|
+
originalContent?: string
|
|
29
|
+
/** 触发此改动的 CRSI insight id */
|
|
30
|
+
crsiInsightId?: string
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
// ── Pending proposal registry (两阶段闸门) ──
|
|
34
|
+
// 模块级单例:成功跑完测试的修改停在这里,等待人类 approve / reject。
|
|
35
|
+
let pendingSandbox: CrsiSandbox | null = null
|
|
36
|
+
|
|
37
|
+
/**
|
|
38
|
+
* 编排完整 5 阶段:createWorktree → applyModification → runTests →(失败自动
|
|
39
|
+
* rollback)→ getDiff。测试通过后暂存为 pending,返回 diff 供人类审阅。
|
|
40
|
+
*/
|
|
41
|
+
export function runCrsiModification(
|
|
42
|
+
proposal: CrsiProposal,
|
|
43
|
+
sandbox: CrsiSandbox = new CrsiSandbox(),
|
|
44
|
+
): CrsiModificationResult {
|
|
45
|
+
sandbox.createWorktree()
|
|
46
|
+
|
|
47
|
+
const applied = sandbox.applyModification({
|
|
48
|
+
id: `crsi-mod-${randomUUID().slice(0, 8)}`,
|
|
49
|
+
description: proposal.description,
|
|
50
|
+
filePath: proposal.filePath,
|
|
51
|
+
newContent: proposal.newContent,
|
|
52
|
+
originalContent: proposal.originalContent ?? '',
|
|
53
|
+
crsiInsightId: proposal.crsiInsightId,
|
|
54
|
+
timestamp: new Date().toISOString(),
|
|
55
|
+
})
|
|
56
|
+
|
|
57
|
+
// apply 失败(受保护路径 / 路径穿越 / 内容不一致)——不跑测试,直接回滚。
|
|
58
|
+
if (!applied.applied) {
|
|
59
|
+
sandbox.rollback()
|
|
60
|
+
applied.phase = 'failed' // rollback() 会把 phase 置为 'rolled-back';恢复为 'failed'
|
|
61
|
+
return applied
|
|
62
|
+
}
|
|
63
|
+
|
|
64
|
+
const testResult = sandbox.runTests()
|
|
65
|
+
applied.testResult = testResult
|
|
66
|
+
|
|
67
|
+
if (!testResult.passed) {
|
|
68
|
+
sandbox.rollback()
|
|
69
|
+
applied.phase = 'failed'
|
|
70
|
+
return applied
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
// Eval harness gate:CRSI 契约分数不得低于上次记录(防跨合并退化)。
|
|
74
|
+
// 分数反映「当前代码」的 CRSI 契约(隔离组件,与本次 worktree 改动无关),
|
|
75
|
+
// 所以它是「仓库的 CRSI 代码自上次评估以来是否退化」的哨兵。
|
|
76
|
+
const evalReport = runEval()
|
|
77
|
+
const last = getLastEvalScore()
|
|
78
|
+
if (last !== null && evalReport.score < last) {
|
|
79
|
+
sandbox.rollback()
|
|
80
|
+
applied.phase = 'failed'
|
|
81
|
+
applied.error = `Eval regression: score ${evalReport.score} < last ${last}`
|
|
82
|
+
return applied
|
|
83
|
+
}
|
|
84
|
+
appendEvalScore(evalReport)
|
|
85
|
+
|
|
86
|
+
applied.phase = 'passed'
|
|
87
|
+
applied.diff = sandbox.getDiff()
|
|
88
|
+
pendingSandbox = sandbox
|
|
89
|
+
return applied
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
/** 是否有待批准的修改。 */
|
|
93
|
+
export function hasPending(): boolean {
|
|
94
|
+
return pendingSandbox !== null
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
/** 人类批准:merge 进仓库并清理 worktree。 */
|
|
98
|
+
export function approvePending(): { success: boolean; message: string } {
|
|
99
|
+
if (!pendingSandbox) {
|
|
100
|
+
return { success: false, message: '没有待批准的修改。先运行 /crsi modify。' }
|
|
101
|
+
}
|
|
102
|
+
const sandbox = pendingSandbox
|
|
103
|
+
pendingSandbox = null
|
|
104
|
+
const merged = sandbox.merge()
|
|
105
|
+
sandbox.finalize()
|
|
106
|
+
return merged
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
/** 人类拒绝:丢弃 worktree。 */
|
|
110
|
+
export function rejectPending(): { success: boolean; message: string } {
|
|
111
|
+
if (!pendingSandbox) {
|
|
112
|
+
return { success: false, message: '没有待批准的修改。先运行 /crsi modify。' }
|
|
113
|
+
}
|
|
114
|
+
const sandbox = pendingSandbox
|
|
115
|
+
pendingSandbox = null
|
|
116
|
+
return sandbox.rollback()
|
|
117
|
+
}
|
|
@@ -0,0 +1,98 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* CRSI Producer — 把累积的失败信号转成「教训文件」代码改动候选。
|
|
3
|
+
*
|
|
4
|
+
* 这是 CRSI 闭环「reflect → verify → consolidate」的 reflect→verify 桥:
|
|
5
|
+
* - 输入:AutoMemoryEngine 的 CrsiInsight + MetaRuleEngine 的 MetaRule(都是「建议」)。
|
|
6
|
+
* - 输出:一个 CrsiProposal —— 对 `crsi-lessons.md` 的追加(模板化,不动 LLM 判断)。
|
|
7
|
+
* - 走 runCrsiModification(沙箱 gate)→ 人类批准 → merge。
|
|
8
|
+
*
|
|
9
|
+
* 诚实标注:沙箱的 verify 是「防回归」(测试仍绿),不是「证明更好」——
|
|
10
|
+
* 后者需要 ground-truth eval harness,是独立的下一步。
|
|
11
|
+
*/
|
|
12
|
+
|
|
13
|
+
import type { CrsiInsight } from './auto-memory'
|
|
14
|
+
import type { MetaRule } from './meta-rule-engine'
|
|
15
|
+
|
|
16
|
+
/** 教训文件(相对仓库根)。预建,沙箱只能改已存在文件。 */
|
|
17
|
+
export const LESSONS_FILE = 'apps/cli/crsi-lessons.md'
|
|
18
|
+
|
|
19
|
+
/** 归一化的教训信号(insight 与 meta-rule 的公共面)。 */
|
|
20
|
+
export interface CrsiSignal {
|
|
21
|
+
category: string
|
|
22
|
+
title: string
|
|
23
|
+
severity?: string
|
|
24
|
+
suggestion: string
|
|
25
|
+
evidence: string[]
|
|
26
|
+
}
|
|
27
|
+
|
|
28
|
+
const SEVERITY_RANK: Record<string, number> = { critical: 0, warning: 1, info: 2 }
|
|
29
|
+
|
|
30
|
+
/**
|
|
31
|
+
* 选一条「最该固化成教训」的信号:
|
|
32
|
+
* 1. 优先 autoApplicable 的 insight,按严重度排序(critical > warning > info)。
|
|
33
|
+
* 2. 没有 insight 时,回退到高置信、autoApplicable 的元规则。
|
|
34
|
+
*/
|
|
35
|
+
export function selectCrsiSignal(
|
|
36
|
+
insights: CrsiInsight[],
|
|
37
|
+
metaRules: MetaRule[],
|
|
38
|
+
): CrsiSignal | null {
|
|
39
|
+
const best = insights
|
|
40
|
+
.filter((i) => i.autoApplicable)
|
|
41
|
+
.sort((a, b) => (SEVERITY_RANK[a.severity] ?? 3) - (SEVERITY_RANK[b.severity] ?? 3))[0]
|
|
42
|
+
if (best) {
|
|
43
|
+
return {
|
|
44
|
+
category: best.category,
|
|
45
|
+
title: best.description,
|
|
46
|
+
severity: best.severity,
|
|
47
|
+
suggestion: best.suggestion,
|
|
48
|
+
evidence: best.evidence,
|
|
49
|
+
}
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
const mr = metaRules.find((m) => m.autoApplicable && m.confidence === 'high')
|
|
53
|
+
if (mr) {
|
|
54
|
+
return {
|
|
55
|
+
category: mr.category,
|
|
56
|
+
title: mr.title,
|
|
57
|
+
suggestion: mr.recommendation,
|
|
58
|
+
evidence: [mr.evidence.summary],
|
|
59
|
+
}
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
return null
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
/** 模板化地把信号渲染成一段教训 markdown(不动 LLM)。 */
|
|
66
|
+
export function buildLessonContent(signal: CrsiSignal, timestamp: string): string {
|
|
67
|
+
const lines: string[] = [
|
|
68
|
+
`## ${signal.category}: ${signal.title}`,
|
|
69
|
+
'',
|
|
70
|
+
`- 建议: ${signal.suggestion}`,
|
|
71
|
+
]
|
|
72
|
+
if (signal.severity) lines.push(`- 严重度: ${signal.severity}`)
|
|
73
|
+
lines.push(`- 生成时间: ${timestamp}`, '- 来源: CRSI producer (autoApplicable)', '', '### 证据')
|
|
74
|
+
for (const e of signal.evidence) lines.push(`- ${e}`)
|
|
75
|
+
lines.push('')
|
|
76
|
+
return lines.join('\n')
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
/** 产出教训文件变更候选。无合格信号时返回 null。 */
|
|
80
|
+
export function produceCrsiProposal(
|
|
81
|
+
insights: CrsiInsight[],
|
|
82
|
+
metaRules: MetaRule[],
|
|
83
|
+
currentLessons: string,
|
|
84
|
+
timestamp: string,
|
|
85
|
+
): { description: string; filePath: string; newContent: string; originalContent: string } | null {
|
|
86
|
+
const signal = selectCrsiSignal(insights, metaRules)
|
|
87
|
+
if (!signal) return null
|
|
88
|
+
|
|
89
|
+
const lesson = buildLessonContent(signal, timestamp)
|
|
90
|
+
const newContent = currentLessons ? `${currentLessons.trimEnd()}\n\n${lesson}\n` : `${lesson}\n`
|
|
91
|
+
|
|
92
|
+
return {
|
|
93
|
+
description: `CRSI lesson: ${signal.category} — ${signal.title}`,
|
|
94
|
+
filePath: LESSONS_FILE,
|
|
95
|
+
newContent,
|
|
96
|
+
originalContent: currentLessons,
|
|
97
|
+
}
|
|
98
|
+
}
|
package/src/core/crsi-sandbox.ts
CHANGED
|
@@ -88,6 +88,36 @@ const MAX_WORKTREE_AGE_MS = 30 * 60 * 1000 // 30 minutes
|
|
|
88
88
|
const TEST_TIMEOUT_MS = 120_000 // 2 minutes
|
|
89
89
|
const REPORT_DIR = join(homedir(), '.mipham', 'crsi-sandbox')
|
|
90
90
|
|
|
91
|
+
/**
|
|
92
|
+
* 自改进循环的只读边界(信任域隔离)。
|
|
93
|
+
*
|
|
94
|
+
* 这些路径是 CRSI 自改进的「慢通道」——宪法、eval harness、改进机制自身。
|
|
95
|
+
* 自改进循环可以改 skill/workflow/prompt/memory,但绝不能改:
|
|
96
|
+
* 1. 宪法(对齐本体 + 对齐缝 + 加载器)—— 否则价值漂移会「优化」掉安全边界
|
|
97
|
+
* 2. eval harness(测试套件)—— 否则会改掉自己的评估标准(Goodhart 元劫持)
|
|
98
|
+
* 3. 改进机制自身(有效性追踪 / 元规则引擎 / 沙箱)—— 否则递归会改掉评估器
|
|
99
|
+
*
|
|
100
|
+
* 注意:这是 fail-closed 边界——宁可多拦,不可漏拦。
|
|
101
|
+
*/
|
|
102
|
+
const PROTECTED_PATHS = [
|
|
103
|
+
// 宪法(对齐)
|
|
104
|
+
'apps/cli/src/core/alignment-vocabulary.json',
|
|
105
|
+
'apps/cli/src/core/constitution-loader.ts',
|
|
106
|
+
'apps/cli/src/core/constitution-seam.ts',
|
|
107
|
+
'apps/cli/src/vajra/constitution.ts',
|
|
108
|
+
// eval harness
|
|
109
|
+
'apps/cli/test/',
|
|
110
|
+
// 改进机制自身
|
|
111
|
+
'apps/cli/src/agent/effectiveness-tracker.ts',
|
|
112
|
+
'apps/cli/src/core/meta-rule-engine.ts',
|
|
113
|
+
'apps/cli/src/core/crsi-sandbox.ts',
|
|
114
|
+
]
|
|
115
|
+
|
|
116
|
+
/** 是否命中只读边界。前缀匹配,目录条目以 `/` 结尾。 */
|
|
117
|
+
export function isProtectedPath(filePath: string): boolean {
|
|
118
|
+
return PROTECTED_PATHS.some((p) => filePath === p || filePath.startsWith(p))
|
|
119
|
+
}
|
|
120
|
+
|
|
91
121
|
// ── Sandbox ──
|
|
92
122
|
|
|
93
123
|
export class CrsiSandbox {
|
|
@@ -183,6 +213,15 @@ export class CrsiSandbox {
|
|
|
183
213
|
return result
|
|
184
214
|
}
|
|
185
215
|
|
|
216
|
+
// Protected-path guard: the self-improvement loop must not modify the
|
|
217
|
+
// constitution, eval harness, or improvement machinery itself.
|
|
218
|
+
if (isProtectedPath(mod.filePath)) {
|
|
219
|
+
result.error = `Protected path: "${mod.filePath}" is read-only to the self-improvement loop.`
|
|
220
|
+
result.phase = 'failed'
|
|
221
|
+
this.sessionReport.modifications.push(result)
|
|
222
|
+
return result
|
|
223
|
+
}
|
|
224
|
+
|
|
186
225
|
// Verify the file exists in the worktree
|
|
187
226
|
if (!existsSync(targetPath)) {
|
|
188
227
|
result.error = `File not found in worktree: ${mod.filePath}`
|
|
@@ -0,0 +1,184 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* CRSI Eval Harness — 冻结的 ground-truth 契约评估。
|
|
3
|
+
*
|
|
4
|
+
* 自改进环的「verify」升级:单测只能证明「测试仍绿」(防回归),
|
|
5
|
+
* 本 harness 用一组人类冻结的、无 LLM 的客观断言给 CRSI 机制打分,
|
|
6
|
+
* 并把分数持久化到 rewards 日志——这样「变好了还是变差了」才可被回答。
|
|
7
|
+
*
|
|
8
|
+
* 设计约束(对应 path A 的 A1 铁律):
|
|
9
|
+
* - 每条任务用可机器判定的 ground truth,绝不拿 LLM 当裁判。
|
|
10
|
+
* - 用隔离组件(tmpdir),不读用户 ~/.mipham 的运行时状态——
|
|
11
|
+
* harness 量的是「CRSI 机制代码是否满足冻结契约」,与用户数据无关。
|
|
12
|
+
*/
|
|
13
|
+
|
|
14
|
+
import { join } from 'node:path'
|
|
15
|
+
import { tmpdir, homedir } from 'node:os'
|
|
16
|
+
import { mkdirSync, appendFileSync, readFileSync, existsSync } from 'node:fs'
|
|
17
|
+
import { ExperienceRuleEngine } from './rule-engine'
|
|
18
|
+
import { ConstitutionLoader, DEFAULT_CONSTITUTION } from './constitution-loader'
|
|
19
|
+
import { ErrorSignatureDB } from './error-signature-db'
|
|
20
|
+
import { PreFlightChecker } from './preflight-checker'
|
|
21
|
+
import { RedTeam } from './red-team'
|
|
22
|
+
import { isProtectedPath } from './crsi-sandbox'
|
|
23
|
+
|
|
24
|
+
// ── Types ──
|
|
25
|
+
|
|
26
|
+
export interface EvalResult {
|
|
27
|
+
id: string
|
|
28
|
+
description: string
|
|
29
|
+
passed: boolean
|
|
30
|
+
detail?: string
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
export interface EvalReport {
|
|
34
|
+
total: number
|
|
35
|
+
passed: number
|
|
36
|
+
/** 0-100 */
|
|
37
|
+
score: number
|
|
38
|
+
results: EvalResult[]
|
|
39
|
+
failures: string[]
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
// ── Rewards log (path A Phase 1: 奖励信号持久化) ──
|
|
43
|
+
|
|
44
|
+
const SCORES_FILE = join(homedir(), '.mipham', 'crsi', 'eval-scores.jsonl')
|
|
45
|
+
|
|
46
|
+
/** 追加一次评估分数到 rewards 日志。 */
|
|
47
|
+
export function appendEvalScore(report: EvalReport): void {
|
|
48
|
+
try {
|
|
49
|
+
mkdirSync(join(homedir(), '.mipham', 'crsi'), { recursive: true })
|
|
50
|
+
appendFileSync(
|
|
51
|
+
SCORES_FILE,
|
|
52
|
+
JSON.stringify({
|
|
53
|
+
timestamp: new Date().toISOString(),
|
|
54
|
+
score: report.score,
|
|
55
|
+
passed: report.passed,
|
|
56
|
+
total: report.total,
|
|
57
|
+
}) + '\n',
|
|
58
|
+
'utf-8',
|
|
59
|
+
)
|
|
60
|
+
} catch {
|
|
61
|
+
// rewards 日志是非关键的——失败不影响评估本身
|
|
62
|
+
}
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
/** 读取最近一次评估分数(无记录时返回 null)。 */
|
|
66
|
+
export function getLastEvalScore(): number | null {
|
|
67
|
+
try {
|
|
68
|
+
if (!existsSync(SCORES_FILE)) return null
|
|
69
|
+
const lines = readFileSync(SCORES_FILE, 'utf-8').trim().split('\n').filter(Boolean)
|
|
70
|
+
if (lines.length === 0) return null
|
|
71
|
+
const last = JSON.parse(lines[lines.length - 1]!) as { score?: number }
|
|
72
|
+
return typeof last.score === 'number' ? last.score : null
|
|
73
|
+
} catch {
|
|
74
|
+
return null
|
|
75
|
+
}
|
|
76
|
+
}
|
|
77
|
+
|
|
78
|
+
// ── Harness ──
|
|
79
|
+
|
|
80
|
+
/** 构建隔离组件,避免读用户 ~/.mipham 运行时状态。 */
|
|
81
|
+
function buildIsolatedComponents() {
|
|
82
|
+
const dir = join(tmpdir(), 'mipham-eval-harness')
|
|
83
|
+
const ruleEngine = new ExperienceRuleEngine(join(dir, 'rules'))
|
|
84
|
+
const constitution = new ConstitutionLoader(join(dir, 'constitution.yml'))
|
|
85
|
+
const errorDB = new ErrorSignatureDB(join(dir, 'sis'))
|
|
86
|
+
const preflight = new PreFlightChecker(errorDB, ruleEngine)
|
|
87
|
+
return { ruleEngine, constitution, errorDB, preflight }
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
export function runEval(): EvalReport {
|
|
91
|
+
const { ruleEngine, constitution, errorDB, preflight } = buildIsolatedComponents()
|
|
92
|
+
|
|
93
|
+
const results: EvalResult[] = []
|
|
94
|
+
|
|
95
|
+
// ── 规则引擎(ground truth:内置契约) ──
|
|
96
|
+
const timeout = ruleEngine.intercept('Bash', {
|
|
97
|
+
command: 'npm install express',
|
|
98
|
+
timeout: 120000,
|
|
99
|
+
description: 'install deps',
|
|
100
|
+
})
|
|
101
|
+
results.push({
|
|
102
|
+
id: 'rule-timeout',
|
|
103
|
+
description: '内置 timeout 规则命中低超时的 npm install',
|
|
104
|
+
passed: timeout.modified.timeout === 300000,
|
|
105
|
+
})
|
|
106
|
+
|
|
107
|
+
const gitForce = ruleEngine.intercept('Bash', {
|
|
108
|
+
command: 'git push --force origin main',
|
|
109
|
+
description: 'force push',
|
|
110
|
+
})
|
|
111
|
+
results.push({
|
|
112
|
+
id: 'rule-git-force',
|
|
113
|
+
description: 'git --force 触发告警',
|
|
114
|
+
passed: gitForce.warnings.length > 0,
|
|
115
|
+
})
|
|
116
|
+
|
|
117
|
+
const disabledRule: import('./rule-engine').ToolRule = {
|
|
118
|
+
id: 'eval-disabled-test',
|
|
119
|
+
toolName: 'Read',
|
|
120
|
+
category: 'tool-params',
|
|
121
|
+
match: () => true,
|
|
122
|
+
fix: (p) => ({ modified: p, warning: 'should not appear' }),
|
|
123
|
+
source: 'manual',
|
|
124
|
+
enabled: false,
|
|
125
|
+
}
|
|
126
|
+
ruleEngine.register(disabledRule)
|
|
127
|
+
const disabled = ruleEngine.intercept('Read', { file_path: '/tmp/x.txt' })
|
|
128
|
+
results.push({
|
|
129
|
+
id: 'rule-disabled-skip',
|
|
130
|
+
description: '禁用规则被跳过',
|
|
131
|
+
passed: disabled.warnings.length === 0,
|
|
132
|
+
})
|
|
133
|
+
|
|
134
|
+
// ── 宪法(ground truth:8 原则 + facet 映射 + 愿力序言) ──
|
|
135
|
+
const principles = DEFAULT_CONSTITUTION.principles
|
|
136
|
+
results.push({
|
|
137
|
+
id: 'constitution-8-principles',
|
|
138
|
+
description: '宪法含 8 条原则',
|
|
139
|
+
passed: principles.length === 8,
|
|
140
|
+
})
|
|
141
|
+
|
|
142
|
+
const prajna = principles.filter((p) => p.facet === 'prajna').length
|
|
143
|
+
const vajra = principles.filter((p) => p.facet === 'vajra').length
|
|
144
|
+
const karuna = principles.filter((p) => p.facet === 'karuna').length
|
|
145
|
+
results.push({
|
|
146
|
+
id: 'constitution-facets',
|
|
147
|
+
description: 'facet 映射 智3 / 金刚5 / 悲0',
|
|
148
|
+
passed: prajna === 3 && vajra === 5 && karuna === 0,
|
|
149
|
+
})
|
|
150
|
+
|
|
151
|
+
results.push({
|
|
152
|
+
id: 'constitution-preamble',
|
|
153
|
+
description: '愿力序言已注入',
|
|
154
|
+
passed: !!DEFAULT_CONSTITUTION.preamble && DEFAULT_CONSTITUTION.preamble.includes('悲'),
|
|
155
|
+
})
|
|
156
|
+
|
|
157
|
+
// ── 沙箱只读边界(ground truth:受保护路径被拒) ──
|
|
158
|
+
const protectedChecks: Array<[string, string]> = [
|
|
159
|
+
['sandbox-protected-constitution', 'apps/cli/src/core/alignment-vocabulary.json'],
|
|
160
|
+
['sandbox-protected-tests', 'apps/cli/test/foo.test.ts'],
|
|
161
|
+
['sandbox-protected-machinery', 'apps/cli/src/core/crsi-sandbox.ts'],
|
|
162
|
+
]
|
|
163
|
+
for (const [id, path] of protectedChecks) {
|
|
164
|
+
results.push({ id, description: `受保护路径被拒: ${path}`, passed: isProtectedPath(path) })
|
|
165
|
+
}
|
|
166
|
+
|
|
167
|
+
// ── 安全(ground truth:16 攻击零漏过) ──
|
|
168
|
+
const redTeam = new RedTeam().run(constitution, preflight, errorDB)
|
|
169
|
+
results.push({
|
|
170
|
+
id: 'red-team-zero-gaps',
|
|
171
|
+
description: '16 个对抗场景零漏过',
|
|
172
|
+
passed: redTeam.passedThrough === 0,
|
|
173
|
+
detail: `score=${redTeam.score}, passedThrough=${redTeam.passedThrough}, falsePositives=${redTeam.falsePositives}`,
|
|
174
|
+
})
|
|
175
|
+
|
|
176
|
+
const passed = results.filter((r) => r.passed).length
|
|
177
|
+
return {
|
|
178
|
+
total: results.length,
|
|
179
|
+
passed,
|
|
180
|
+
score: results.length > 0 ? Math.round((passed / results.length) * 100) : 100,
|
|
181
|
+
results,
|
|
182
|
+
failures: results.filter((r) => !r.passed).map((r) => r.id),
|
|
183
|
+
}
|
|
184
|
+
}
|
package/src/core/instructions.ts
CHANGED
|
@@ -153,6 +153,16 @@ Script format: export const meta = { name, description, phases: [...] }
|
|
|
153
153
|
- When you need to search AND read results: Glob → Grep → Read
|
|
154
154
|
(find files, search contents, then read the matching files)`)
|
|
155
155
|
|
|
156
|
+
// CRSI 能力自报告 — 回答自身能力边界前先查实时状态,勿凭静态清单推断
|
|
157
|
+
parts.push(`## Capability Self-Report Rule
|
|
158
|
+
|
|
159
|
+
When asked about your own capabilities ("what can you do", "what do you
|
|
160
|
+
have / what is missing", "are you able to X"), do NOT infer the answer
|
|
161
|
+
from your static tool list. Run \`/crsi inventory\` first and answer from
|
|
162
|
+
its live CRSI / SIS / constitution state. Report the numbers you read
|
|
163
|
+
from it as live counts; if it shows a subsystem as 未初始化 (uninitialized),
|
|
164
|
+
say so explicitly instead of claiming it exists.`)
|
|
165
|
+
|
|
156
166
|
return parts.join('\n\n---\n\n')
|
|
157
167
|
}
|
|
158
168
|
|
package/src/core/red-team.ts
CHANGED
|
@@ -240,9 +240,13 @@ export class RedTeam {
|
|
|
240
240
|
constitution: ConstitutionLoader,
|
|
241
241
|
preflight: PreFlightChecker,
|
|
242
242
|
): RedTeamResult {
|
|
243
|
-
// 1. Check against constitution audit patterns
|
|
244
|
-
|
|
245
|
-
|
|
243
|
+
// 1. Check against constitution audit patterns — audit the raw string values,
|
|
244
|
+
// not JSON.stringify (which escapes `"` → `\"` and breaks quote-matching
|
|
245
|
+
// patterns like no-credential-leak's `password = "..."`).
|
|
246
|
+
const content = Object.values(attack.params)
|
|
247
|
+
.filter((v): v is string => typeof v === 'string')
|
|
248
|
+
.join('\n')
|
|
249
|
+
const constitutionViolations = constitution.audit(content)
|
|
246
250
|
if (constitutionViolations.length > 0) {
|
|
247
251
|
return {
|
|
248
252
|
attack,
|
|
@@ -20,6 +20,7 @@
|
|
|
20
20
|
|
|
21
21
|
import type { ProviderRegistry } from '../providers/registry'
|
|
22
22
|
import type { Llm } from '../providers/llm'
|
|
23
|
+
import { DEFAULT_CONSTITUTION } from './constitution-loader'
|
|
23
24
|
|
|
24
25
|
// ── Types ──
|
|
25
26
|
|
|
@@ -62,7 +63,12 @@ export const DEFAULT_SELF_CRITIQUE_CONFIG: SelfCritiqueConfig = {
|
|
|
62
63
|
|
|
63
64
|
// ── Prompt Templates ──
|
|
64
65
|
|
|
65
|
-
const CRITIQUE_PROMPT = `You are a safety auditor for an AI coding agent. Evaluate the following tool call
|
|
66
|
+
const CRITIQUE_PROMPT = `You are a safety auditor for an AI coding agent. Evaluate the following tool call.
|
|
67
|
+
|
|
68
|
+
ALIGNMENT PREAMBLE(愿力序言):
|
|
69
|
+
{preamble}
|
|
70
|
+
|
|
71
|
+
First, align to the vow: does this action embody compassion (悲) and sober causal judgment (智), and maintain structural integrity (金刚)?
|
|
66
72
|
|
|
67
73
|
TOOL: {toolName}
|
|
68
74
|
PARAMETERS: {params}
|
|
@@ -93,6 +99,19 @@ Necessity criteria:
|
|
|
93
99
|
- Not redundant with previous tool calls
|
|
94
100
|
- Not an unnecessary "exploratory" action`
|
|
95
101
|
|
|
102
|
+
/** Build the critique prompt with the alignment preamble injected before the prohibitions. */
|
|
103
|
+
export function buildCritiquePrompt(
|
|
104
|
+
toolName: string,
|
|
105
|
+
params: Record<string, unknown>,
|
|
106
|
+
context: string | undefined,
|
|
107
|
+
preamble: string,
|
|
108
|
+
): string {
|
|
109
|
+
return CRITIQUE_PROMPT.replace('{preamble}', preamble)
|
|
110
|
+
.replace('{toolName}', toolName)
|
|
111
|
+
.replace('{params}', JSON.stringify(params, null, 2).slice(0, 500))
|
|
112
|
+
.replace('{context}', context?.slice(0, 300) || 'No additional context')
|
|
113
|
+
}
|
|
114
|
+
|
|
96
115
|
// ── Critic ──
|
|
97
116
|
|
|
98
117
|
export class SelfCritique {
|
|
@@ -135,9 +154,12 @@ export class SelfCritique {
|
|
|
135
154
|
return null
|
|
136
155
|
}
|
|
137
156
|
|
|
138
|
-
const prompt =
|
|
139
|
-
|
|
140
|
-
|
|
157
|
+
const prompt = buildCritiquePrompt(
|
|
158
|
+
toolName,
|
|
159
|
+
params,
|
|
160
|
+
context,
|
|
161
|
+
DEFAULT_CONSTITUTION.preamble ?? '',
|
|
162
|
+
)
|
|
141
163
|
|
|
142
164
|
try {
|
|
143
165
|
const critiqueModel = this.config.model || this.findFastestModel(registry)
|
package/src/index.tsx
CHANGED
|
@@ -36,6 +36,8 @@ import { createToolRegistry } from './tools'
|
|
|
36
36
|
import { Context } from './vajra'
|
|
37
37
|
import { mountSkills, SKILLS_KEY } from './skills/seam'
|
|
38
38
|
import { mountLlm, LLM_KEY } from './providers/llm'
|
|
39
|
+
import { mountConstitution, createConstitution } from './core/constitution-seam'
|
|
40
|
+
import { ConstitutionLoader } from './core/constitution-loader'
|
|
39
41
|
import { McpClient } from './mcp/client'
|
|
40
42
|
import { registerMcpServerTools } from './mcp/registry'
|
|
41
43
|
import { AgentRegistry } from './agent/agent-registry'
|
|
@@ -447,6 +449,9 @@ export async function runApp(options: RunOptions): Promise<void> {
|
|
|
447
449
|
const vajraContext = new Context()
|
|
448
450
|
vajraContext.provide('credentials', credentialMaskingConfig)
|
|
449
451
|
|
|
452
|
+
// 对齐缝:宪法先就位,后续每个 Service 挂载(含工具)都在挂载前过对齐门。
|
|
453
|
+
mountConstitution(vajraContext, createConstitution(new ConstitutionLoader()))
|
|
454
|
+
|
|
450
455
|
// Create tool registry with all built-in tools (mounted as Vajra services)
|
|
451
456
|
const tools = createToolRegistry(vajraContext)
|
|
452
457
|
|
|
@@ -610,6 +615,16 @@ export async function runApp(options: RunOptions): Promise<void> {
|
|
|
610
615
|
process.on('exit', () => {
|
|
611
616
|
clearInterval(heartbeatInterval)
|
|
612
617
|
unregisterSession(sessionName)
|
|
618
|
+
// Close the CRSI effectiveness loop on non-interactive exit paths
|
|
619
|
+
// (daemon worker / crash / kill) that bypass saveAndExit. saveAndExit
|
|
620
|
+
// already flushes; guard on !saved to avoid double-evaluating.
|
|
621
|
+
if (!saved) {
|
|
622
|
+
try {
|
|
623
|
+
engine.getAutoMemory().flushEffectiveness()
|
|
624
|
+
} catch {
|
|
625
|
+
// ignore — CRSI flush is non-critical
|
|
626
|
+
}
|
|
627
|
+
}
|
|
613
628
|
if (!saved && context.getMessageCount() > 0) {
|
|
614
629
|
SessionStore.autoSave(context.getMessages(), {
|
|
615
630
|
provider: defaultProvider,
|
|
@@ -9,7 +9,7 @@
|
|
|
9
9
|
export const PACKAGE_NAME = '@miphamai/cli' as const
|
|
10
10
|
|
|
11
11
|
/** 当前发布版本 */
|
|
12
|
-
export const PACKAGE_VERSION = '0.
|
|
12
|
+
export const PACKAGE_VERSION = '0.43.0' as const
|
|
13
13
|
|
|
14
14
|
/** npm install 全局安装命令 */
|
|
15
15
|
export const NPM_INSTALL_COMMAND = `npm install -g ${PACKAGE_NAME}` as const
|
package/src/ui/commands.ts
CHANGED
|
@@ -10,6 +10,10 @@ import type { SkillsLoader } from '../skills/loader'
|
|
|
10
10
|
import type { PluginManager } from '../plugin/plugin-manager'
|
|
11
11
|
import type { Message } from '../shared/types.js'
|
|
12
12
|
import { McpClient } from '../mcp/client'
|
|
13
|
+
import { buildCapabilityReport } from '../core/capability-inventory'
|
|
14
|
+
import { runCrsiModification, approvePending, rejectPending, hasPending } from '../core/crsi-modify'
|
|
15
|
+
import { produceCrsiProposal, LESSONS_FILE } from '../core/crsi-producer'
|
|
16
|
+
import { runEval, appendEvalScore } from '../core/eval-harness'
|
|
13
17
|
import { NPM_UPDATE_COMMAND, PACKAGE_VERSION } from '../shared/index.ts'
|
|
14
18
|
import { getPreference } from '../config/preferences'
|
|
15
19
|
import { loadCrossSessionConfig } from '../config/loader'
|
|
@@ -731,6 +735,123 @@ const crsiStatsCmd: CommandHandler = async (ctx) => {
|
|
|
731
735
|
return { content: lines.join('\n') }
|
|
732
736
|
}
|
|
733
737
|
|
|
738
|
+
const crsiInventoryCmd: CommandHandler = async (ctx) => {
|
|
739
|
+
return { content: buildCapabilityReport(ctx.engine) }
|
|
740
|
+
}
|
|
741
|
+
|
|
742
|
+
const crsiModifyCmd: CommandHandler = async (ctx, args) => {
|
|
743
|
+
if (args[0] === '--approve') {
|
|
744
|
+
const r = approvePending()
|
|
745
|
+
return { content: r.success ? `✅ ${r.message}` : `⚠️ ${r.message}` }
|
|
746
|
+
}
|
|
747
|
+
if (args[0] === '--reject') {
|
|
748
|
+
const r = rejectPending()
|
|
749
|
+
return { content: r.success ? `✅ ${r.message}` : `⚠️ ${r.message}` }
|
|
750
|
+
}
|
|
751
|
+
if (args.length < 3) {
|
|
752
|
+
return {
|
|
753
|
+
content:
|
|
754
|
+
'Usage: /crsi modify <description> <filePath> <newContent>\n' +
|
|
755
|
+
'- description 单 token(不含空格)\n' +
|
|
756
|
+
'- newContent 用 \\n 表示换行\n' +
|
|
757
|
+
'测试通过后:/crsi modify --approve 合并,/crsi modify --reject 丢弃',
|
|
758
|
+
}
|
|
759
|
+
}
|
|
760
|
+
if (hasPending()) {
|
|
761
|
+
return { content: '⚠️ 已有待批准的修改。先 /crsi modify --approve 或 --reject。' }
|
|
762
|
+
}
|
|
763
|
+
|
|
764
|
+
const description = args[0]!
|
|
765
|
+
const filePath = args[1]!
|
|
766
|
+
const newContent = args.slice(2).join(' ').replace(/\\n/g, '\n')
|
|
767
|
+
|
|
768
|
+
let originalContent = ''
|
|
769
|
+
try {
|
|
770
|
+
const { readFileSync } = await import('node:fs')
|
|
771
|
+
const { join } = await import('node:path')
|
|
772
|
+
originalContent = readFileSync(join(process.cwd(), filePath), 'utf-8')
|
|
773
|
+
} catch {
|
|
774
|
+
// 文件不存在 → 宽松模式(originalContent 为空)
|
|
775
|
+
}
|
|
776
|
+
|
|
777
|
+
const result = runCrsiModification({ description, filePath, newContent, originalContent })
|
|
778
|
+
if (!result.applied || result.phase === 'failed') {
|
|
779
|
+
return {
|
|
780
|
+
content: `❌ 修改未通过(phase: ${result.phase})。\n${result.error ?? ''}`,
|
|
781
|
+
}
|
|
782
|
+
}
|
|
783
|
+
|
|
784
|
+
return {
|
|
785
|
+
content:
|
|
786
|
+
`✅ 测试通过。审阅下方 diff:\n\n${result.diff}\n\n` +
|
|
787
|
+
'/crsi modify --approve 合并\n/crsi modify --reject 丢弃',
|
|
788
|
+
}
|
|
789
|
+
}
|
|
790
|
+
|
|
791
|
+
const crsiProposeCmd: CommandHandler = async (ctx) => {
|
|
792
|
+
if (hasPending()) {
|
|
793
|
+
return { content: '⚠️ 已有待批准的修改。先 /crsi modify --approve 或 --reject。' }
|
|
794
|
+
}
|
|
795
|
+
|
|
796
|
+
const insights = ctx.engine.getAutoMemory?.()?.accumulatedInsights ?? []
|
|
797
|
+
let metaRules: Parameters<typeof produceCrsiProposal>[1] = []
|
|
798
|
+
try {
|
|
799
|
+
metaRules = ctx.engine.getMetaRuleEngine?.()?.analyze().metaRules ?? []
|
|
800
|
+
} catch {
|
|
801
|
+
metaRules = []
|
|
802
|
+
}
|
|
803
|
+
|
|
804
|
+
// 教训文件按仓库根解析(沙箱的 filePath 是仓库根相对)。
|
|
805
|
+
let current = ''
|
|
806
|
+
try {
|
|
807
|
+
const { readFileSync } = await import('node:fs')
|
|
808
|
+
const { join } = await import('node:path')
|
|
809
|
+
const root = execSync('git rev-parse --show-toplevel', {
|
|
810
|
+
timeout: 5000,
|
|
811
|
+
encoding: 'utf-8',
|
|
812
|
+
}).trim()
|
|
813
|
+
current = readFileSync(join(root, LESSONS_FILE), 'utf-8')
|
|
814
|
+
} catch {
|
|
815
|
+
current = ''
|
|
816
|
+
}
|
|
817
|
+
|
|
818
|
+
const proposal = produceCrsiProposal(insights, metaRules, current, new Date().toISOString())
|
|
819
|
+
|
|
820
|
+
if (!proposal) {
|
|
821
|
+
return { content: '没有足够的失败信号(autoApplicable insight 或高置信元规则)来生成教训。' }
|
|
822
|
+
}
|
|
823
|
+
|
|
824
|
+
const result = runCrsiModification(proposal)
|
|
825
|
+
if (!result.applied || result.phase === 'failed') {
|
|
826
|
+
return { content: `❌ 生成失败(phase: ${result.phase})。\n${result.error ?? ''}` }
|
|
827
|
+
}
|
|
828
|
+
|
|
829
|
+
return {
|
|
830
|
+
content:
|
|
831
|
+
`✅ 已生成教训并跑过测试。审阅 diff:\n\n${result.diff}\n\n` +
|
|
832
|
+
'/crsi modify --approve 合并 | /crsi modify --reject 丢弃',
|
|
833
|
+
}
|
|
834
|
+
}
|
|
835
|
+
|
|
836
|
+
const crsiEvalCmd: CommandHandler = async () => {
|
|
837
|
+
const report = runEval()
|
|
838
|
+
appendEvalScore(report)
|
|
839
|
+
|
|
840
|
+
const lines: string[] = ['## 🧪 CRSI Eval Harness', '']
|
|
841
|
+
lines.push(`得分: **${report.score}/100** (${report.passed}/${report.total})`, '')
|
|
842
|
+
lines.push('| 任务 | 结果 |')
|
|
843
|
+
lines.push('|------|------|')
|
|
844
|
+
for (const r of report.results) {
|
|
845
|
+
lines.push(
|
|
846
|
+
`| ${r.description} | ${r.passed ? '✅' : '❌'}${r.detail ? ` — ${r.detail}` : ''} |`,
|
|
847
|
+
)
|
|
848
|
+
}
|
|
849
|
+
if (report.failures.length > 0) {
|
|
850
|
+
lines.push('', `❌ 失败任务: ${report.failures.join(', ')}`)
|
|
851
|
+
}
|
|
852
|
+
return { content: lines.join('\n') }
|
|
853
|
+
}
|
|
854
|
+
|
|
734
855
|
const crsiHealthCmd: CommandHandler = async (ctx) => {
|
|
735
856
|
const engine = ctx.engine.getRuleEngine()
|
|
736
857
|
const tracker = ctx.engine.getEffectivenessTracker()
|
|
@@ -4313,6 +4434,10 @@ const commandsListCmd: CommandHandler = () => {
|
|
|
4313
4434
|
'/crsi restore': 'Tools & Skills',
|
|
4314
4435
|
'/crsi stats': 'Tools & Skills',
|
|
4315
4436
|
'/crsi health': 'Tools & Skills',
|
|
4437
|
+
'/crsi inventory': 'Tools & Skills',
|
|
4438
|
+
'/crsi modify': 'Tools & Skills',
|
|
4439
|
+
'/crsi propose': 'Tools & Skills',
|
|
4440
|
+
'/crsi eval': 'Tools & Skills',
|
|
4316
4441
|
'/crsi meta': 'Tools & Skills',
|
|
4317
4442
|
'/crsi interpret': 'Tools & Skills',
|
|
4318
4443
|
'/crsi critique': 'Tools & Skills',
|
|
@@ -4465,6 +4590,10 @@ registry.set('/crsi analyze', crsiAnalyzeCmd)
|
|
|
4465
4590
|
registry.set('/crsi restore', crsiRestoreCmd)
|
|
4466
4591
|
registry.set('/crsi stats', crsiStatsCmd)
|
|
4467
4592
|
registry.set('/crsi health', crsiHealthCmd)
|
|
4593
|
+
registry.set('/crsi inventory', crsiInventoryCmd)
|
|
4594
|
+
registry.set('/crsi modify', crsiModifyCmd)
|
|
4595
|
+
registry.set('/crsi propose', crsiProposeCmd)
|
|
4596
|
+
registry.set('/crsi eval', crsiEvalCmd)
|
|
4468
4597
|
registry.set('/crsi meta', crsiMetaCmd)
|
|
4469
4598
|
registry.set('/crsi interpret', crsiInterpretCmd)
|
|
4470
4599
|
registry.set('/crsi critique', crsiCritiqueCmd)
|
|
@@ -4640,6 +4769,10 @@ const COMMAND_DESCRIPTIONS: Record<string, string> = {
|
|
|
4640
4769
|
'/crsi restore': 'Restore a disabled or degraded CRSI rule',
|
|
4641
4770
|
'/crsi stats': 'Show CRSI overall effectiveness statistics',
|
|
4642
4771
|
'/crsi health': 'CRSI + SIS unified health dashboard with scoring',
|
|
4772
|
+
'/crsi inventory': 'Live capability self-report — CRSI/SIS/constitution state',
|
|
4773
|
+
'/crsi modify': 'Run a code self-modification through the sandbox (worktree → tests → approve)',
|
|
4774
|
+
'/crsi propose': 'Produce a codified CRSI lesson from failure signals, gated by the sandbox',
|
|
4775
|
+
'/crsi eval': 'Run the ground-truth CRSI eval harness and record the score',
|
|
4643
4776
|
'/crsi meta': 'RSI Level 3 meta-rule analysis — rules that improve the rules',
|
|
4644
4777
|
'/crsi interpret': 'Tool-call behavior dashboard — error patterns, usage, health',
|
|
4645
4778
|
'/crsi critique': 'Enable/disable RLAIF self-critique on tool calls',
|
|
@@ -0,0 +1,14 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 对齐缝 capability —— 内核在 Service 挂载前调用,校验其声明的对齐原则。
|
|
3
|
+
*
|
|
4
|
+
* 「金刚不坏」从隐喻变成 `mount()` 的强制前置条件:Service 可选声明 `align`
|
|
5
|
+
* (遵守的宪法原则 id),内核在 apply 前向 `constitution` 缝求证,声明的 id
|
|
6
|
+
* 必须全部已知,否则拒绝挂载。此接口只定义「求证」契约,不绑定任何具体宪法来源。
|
|
7
|
+
*/
|
|
8
|
+
export interface Constitution {
|
|
9
|
+
/** 校验声明的原则 id;返回违规 id 列表(空 = 通过)。 */
|
|
10
|
+
check(aligned: string[]): { violations: string[] }
|
|
11
|
+
}
|
|
12
|
+
|
|
13
|
+
/** 缝键:ctx.constitution。 */
|
|
14
|
+
export const CONSTITUTION_KEY = 'constitution'
|
package/src/vajra/context.ts
CHANGED
|
@@ -1,5 +1,7 @@
|
|
|
1
1
|
import type { EventsOfMode } from './events'
|
|
2
2
|
import type { Service, Mounted, ServiceStatus } from './service'
|
|
3
|
+
import type { Constitution } from './constitution'
|
|
4
|
+
import { CONSTITUTION_KEY } from './constitution'
|
|
3
5
|
|
|
4
6
|
export type Disposer = () => void
|
|
5
7
|
|
|
@@ -43,6 +45,19 @@ export class Context {
|
|
|
43
45
|
if (!keys.every((k) => this.has(k))) return false
|
|
44
46
|
applied = true
|
|
45
47
|
status = 'loading'
|
|
48
|
+
// 对齐缝:挂载前向 constitution 求证声明的原则 id,全部已知才放行。
|
|
49
|
+
const align = service.align
|
|
50
|
+
if (align && align.length > 0) {
|
|
51
|
+
const constitution = this.get<Constitution>(CONSTITUTION_KEY)
|
|
52
|
+
if (constitution) {
|
|
53
|
+
const { violations } = constitution.check(align)
|
|
54
|
+
if (violations.length > 0) {
|
|
55
|
+
error = new Error(`对齐违例:${violations.join(', ')}`)
|
|
56
|
+
status = 'failed'
|
|
57
|
+
return true
|
|
58
|
+
}
|
|
59
|
+
}
|
|
60
|
+
}
|
|
46
61
|
effectSnapshot = this.effects.length
|
|
47
62
|
try {
|
|
48
63
|
disposer = service.apply(this) ?? (() => {})
|
package/src/vajra/index.ts
CHANGED
|
@@ -2,3 +2,5 @@ export { Context } from './context'
|
|
|
2
2
|
export type { Disposer } from './context'
|
|
3
3
|
export type { DispatchMode, EventMap, EventsOfMode } from './events'
|
|
4
4
|
export type { Service, ServiceStatus, Mounted } from './service'
|
|
5
|
+
export type { Constitution } from './constitution'
|
|
6
|
+
export { CONSTITUTION_KEY } from './constitution'
|