dsh-workplan 1.12.0 → 1.12.2
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/CHANGELOG.md +51 -0
- package/README.en.md +7 -5
- package/README.md +13 -5
- package/README.zh.md +239 -0
- package/SECURITY.md +25 -2
- package/docs/SBOM.md +17 -9
- package/docs/cc/ADV-FSP-FUNCTIONAL-SPEC.md +44 -0
- package/docs/cc/ADV-TDS-TOE-DESIGN.md +54 -0
- package/docs/cc/AGD-OPS-SECURE-OPERATION.md +70 -0
- package/docs/cc/AVA-VULNERABILITY-ANALYSIS.md +16 -0
- package/docs/cc/CC-2026-ASSESSMENT.md +33 -1
- package/docs/cc/CM-PLAN.md +45 -0
- package/docs/cc/SECURITY-TARGET.md +60 -11
- package/docs/cc/T6-RECHECK-REPORT.md +83 -0
- package/docs/cc/TEST-COVERAGE-MATRIX.md +40 -0
- package/docs/decision-1.12.0-cc1-keep-20261104.md +46 -0
- package/docs/release-record-1.12.0-20261103.md +99 -0
- package/lib/departments.js +41 -1
- package/lib/host.js +46 -3
- package/lib/plan.js +1 -0
- package/lib/runtime.js +81 -6
- package/lib/understanding.js +304 -23
- package/package.json +1 -1
- package/src/departments.ts +39 -1
- package/src/host.ts +46 -4
- package/src/plan.ts +1 -0
- package/src/runtime.ts +78 -6
- package/src/understanding.ts +330 -26
package/src/runtime.ts
CHANGED
|
@@ -13,8 +13,8 @@
|
|
|
13
13
|
*/
|
|
14
14
|
|
|
15
15
|
import { randomUUID } from 'node:crypto'
|
|
16
|
-
import { existsSync, statSync } from 'node:fs'
|
|
17
|
-
import { isAbsolute, resolve } from 'node:path'
|
|
16
|
+
import { existsSync, realpathSync, statSync } from 'node:fs'
|
|
17
|
+
import { basename, dirname, isAbsolute, join, resolve } from 'node:path'
|
|
18
18
|
|
|
19
19
|
import {
|
|
20
20
|
MAX_CONCURRENT,
|
|
@@ -2443,7 +2443,12 @@ export function reconstructFromSession(session: any, targetPlanId?: string) {
|
|
|
2443
2443
|
if (snap) {
|
|
2444
2444
|
try {
|
|
2445
2445
|
snap.statuses = applyStatus(snap.statuses, update.subtaskId, {
|
|
2446
|
-
|
|
2446
|
+
/*
|
|
2447
|
+
* handoff 的 durable 记录携带的是状态快照,不是一次状态迁移。
|
|
2448
|
+
* 重放时若再走 done 分支,会把已经 reported 的行打回 done 并抛
|
|
2449
|
+
* INVALID_TRANSITION(wp-79c31abf 的 H1 就是这样被丢掉的)。
|
|
2450
|
+
*/
|
|
2451
|
+
status: update.handoffs === undefined ? update.status : undefined,
|
|
2447
2452
|
progress: typeof update.progress === 'number' ? update.progress : undefined,
|
|
2448
2453
|
summary: typeof update.note === 'string' ? update.note : undefined,
|
|
2449
2454
|
// [1.8.13] Evidence written through the update tool must survive replay.
|
|
@@ -2451,6 +2456,7 @@ export function reconstructFromSession(session: any, targetPlanId?: string) {
|
|
|
2451
2456
|
acceptance: update.acceptance,
|
|
2452
2457
|
artifacts: update.artifacts,
|
|
2453
2458
|
missingArtifacts: update.missingArtifacts,
|
|
2459
|
+
handoffs: update.handoffs,
|
|
2454
2460
|
revision: Number.isInteger(update.revision) ? update.revision : undefined,
|
|
2455
2461
|
updatedBy: 'agent',
|
|
2456
2462
|
})
|
|
@@ -5391,6 +5397,36 @@ export function createWorkplanRuntime(ctx: any, options: WorkplanRuntimeOptions
|
|
|
5391
5397
|
entries: Array<{ raw: string; canonical?: string; planIds: string[] }>
|
|
5392
5398
|
}>()
|
|
5393
5399
|
|
|
5400
|
+
/*
|
|
5401
|
+
* [1.12.2] G3 修复:符号链接必须归一到同一个文件。
|
|
5402
|
+
* 1.8.18 的 canonicalPath 只做 path.resolve,同一个文件因此有无限种写法——
|
|
5403
|
+
* 交付的是 /a/report.md,改 /link/report.md(link 指向 a)就被放过。
|
|
5404
|
+
* 这里再用 realpath.native 把路径解析成真实身份,两侧都解析才能比对。
|
|
5405
|
+
* 文件尚不存在时解析最近的存在祖先后拼回剩余段落;完全解析不了时
|
|
5406
|
+
* 返回 undefined,调用方回落到旧的 path.resolve 结果(不改变旧行为)。
|
|
5407
|
+
*/
|
|
5408
|
+
function realPathIdentity(value: string) {
|
|
5409
|
+
try {
|
|
5410
|
+
return realpathSync.native(value)
|
|
5411
|
+
} catch {
|
|
5412
|
+
// 文件可能尚未创建,继续找最近的存在祖先。
|
|
5413
|
+
}
|
|
5414
|
+
let cursor = value
|
|
5415
|
+
const tail: string[] = []
|
|
5416
|
+
for (let depth = 0; depth < 64; depth += 1) {
|
|
5417
|
+
const parent = dirname(cursor)
|
|
5418
|
+
if (!parent || parent === cursor) return undefined
|
|
5419
|
+
tail.unshift(basename(cursor))
|
|
5420
|
+
cursor = parent
|
|
5421
|
+
try {
|
|
5422
|
+
return join(realpathSync.native(cursor), ...tail)
|
|
5423
|
+
} catch {
|
|
5424
|
+
// 继续向上。
|
|
5425
|
+
}
|
|
5426
|
+
}
|
|
5427
|
+
return undefined
|
|
5428
|
+
}
|
|
5429
|
+
|
|
5394
5430
|
function canonicalPath(session: any, value: any) {
|
|
5395
5431
|
if (typeof value !== 'string' || value.trim() === '') return undefined
|
|
5396
5432
|
if (isAbsolute(value)) return resolve(value)
|
|
@@ -5398,6 +5434,12 @@ export function createWorkplanRuntime(ctx: any, options: WorkplanRuntimeOptions
|
|
|
5398
5434
|
return typeof cwd === 'string' && cwd ? resolve(cwd, value) : undefined
|
|
5399
5435
|
}
|
|
5400
5436
|
|
|
5437
|
+
function identityPath(session: any, value: any) {
|
|
5438
|
+
const canonical = canonicalPath(session, value)
|
|
5439
|
+
if (canonical === undefined) return undefined
|
|
5440
|
+
return realPathIdentity(canonical) ?? canonical
|
|
5441
|
+
}
|
|
5442
|
+
|
|
5401
5443
|
function presentedEntries(session: any) {
|
|
5402
5444
|
if (typeof session?.eventAt !== 'function' || !(Number(session.seq) >= 0)) return []
|
|
5403
5445
|
const cache = presentedPaths.get(session.id)
|
|
@@ -5422,7 +5464,7 @@ export function createWorkplanRuntime(ctx: any, options: WorkplanRuntimeOptions
|
|
|
5422
5464
|
const planIds = [...seen]
|
|
5423
5465
|
for (const file of event.data?.files ?? []) {
|
|
5424
5466
|
if (typeof file?.path !== 'string' || file.path.trim() === '') continue
|
|
5425
|
-
entries.push({ raw: file.path, canonical:
|
|
5467
|
+
entries.push({ raw: file.path, canonical: identityPath(session, file.path), planIds })
|
|
5426
5468
|
}
|
|
5427
5469
|
}
|
|
5428
5470
|
presentedPaths.set(session.id, {
|
|
@@ -5441,15 +5483,22 @@ export function createWorkplanRuntime(ctx: any, options: WorkplanRuntimeOptions
|
|
|
5441
5483
|
|
|
5442
5484
|
function presentedWriteDenialReason(agent: any, args: any) {
|
|
5443
5485
|
const session = agent?.session
|
|
5444
|
-
if (!session
|
|
5486
|
+
if (!session) return undefined
|
|
5445
5487
|
if (isPlanModeActive(ctx, session)) return undefined
|
|
5446
5488
|
const entries = presentedEntries(session)
|
|
5447
5489
|
if (entries.length === 0) return undefined
|
|
5448
5490
|
const requested = args?.file_path ?? args?.path
|
|
5449
5491
|
if (typeof requested !== 'string' || requested.trim() === '') return undefined
|
|
5450
|
-
|
|
5492
|
+
// [1.12.2] G3: compare by real file identity (symlinks resolved).
|
|
5493
|
+
const canonical = identityPath(session, requested)
|
|
5494
|
+
/*
|
|
5495
|
+
* [1.12.2] G3: compare like-for-like. The entry now holds the real identity
|
|
5496
|
+
* of the presented path, so a link delivered earlier and a write through
|
|
5497
|
+
* the real path (or vice versa) resolve to the same string.
|
|
5498
|
+
*/
|
|
5451
5499
|
const matched = entries.filter((entry: any) =>
|
|
5452
5500
|
entry.raw === requested
|
|
5501
|
+
|| (entry.raw !== undefined && identityPath(session, entry.raw) === canonical)
|
|
5453
5502
|
|| (canonical !== undefined && entry.canonical !== undefined && entry.canonical === canonical))
|
|
5454
5503
|
if (matched.length === 0) return undefined
|
|
5455
5504
|
const planIds = new Set<string>()
|
|
@@ -5469,6 +5518,28 @@ export function createWorkplanRuntime(ctx: any, options: WorkplanRuntimeOptions
|
|
|
5469
5518
|
+ '(present 之后新建的计划不受此限制。)'
|
|
5470
5519
|
}
|
|
5471
5520
|
|
|
5521
|
+
/**
|
|
5522
|
+
* [1.12.2] G2: every path a single write call can touch.
|
|
5523
|
+
* apply_patch carries many files in one string; parsing the directives here
|
|
5524
|
+
* lets the guard deny the whole call instead of letting an unrelated path
|
|
5525
|
+
* in the arguments form a blind spot.
|
|
5526
|
+
*/
|
|
5527
|
+
function writeTargets(_agent: any, args: any): string[] {
|
|
5528
|
+
const targets: string[] = []
|
|
5529
|
+
const single = args?.file_path ?? args?.path
|
|
5530
|
+
if (typeof single === 'string' && single.trim() !== '') targets.push(single)
|
|
5531
|
+
const patch = typeof args?.input === 'string' ? args.input : (typeof args?.patch === 'string' ? args.patch : '')
|
|
5532
|
+
if (patch) {
|
|
5533
|
+
const directive = /^\*\*\* (?:Update File|Add File|Delete File|Move to): (.+)$/gm
|
|
5534
|
+
let match
|
|
5535
|
+
while ((match = directive.exec(patch)) !== null) {
|
|
5536
|
+
const path = match[1].trim()
|
|
5537
|
+
if (path !== '') targets.push(path)
|
|
5538
|
+
}
|
|
5539
|
+
}
|
|
5540
|
+
return targets
|
|
5541
|
+
}
|
|
5542
|
+
|
|
5472
5543
|
function plansPayload(sessionId: any) {
|
|
5473
5544
|
/*
|
|
5474
5545
|
* [1.8.15] The feed must carry build identity too.
|
|
@@ -5561,6 +5632,7 @@ export function createWorkplanRuntime(ctx: any, options: WorkplanRuntimeOptions
|
|
|
5561
5632
|
presentDenialVerdict,
|
|
5562
5633
|
supersededDenialVerdict,
|
|
5563
5634
|
presentedWriteDenialReason,
|
|
5635
|
+
writeTargets,
|
|
5564
5636
|
updateStatus,
|
|
5565
5637
|
cancelPlan,
|
|
5566
5638
|
dispatchTool,
|
package/src/understanding.ts
CHANGED
|
@@ -76,6 +76,24 @@ export const MAX_AUDITED_SUBTASKS = 15
|
|
|
76
76
|
*/
|
|
77
77
|
export const MISMATCH_BLOCKING_SCORE = 3
|
|
78
78
|
|
|
79
|
+
/**
|
|
80
|
+
* [P0-1] Domain coverage threshold.
|
|
81
|
+
*
|
|
82
|
+
* The MS2 run (2026-11-03) measured 11/29 required departments present: plans
|
|
83
|
+
* silently dropped whole domains (design/legal/security/devops) the user's own
|
|
84
|
+
* request names. A single incidental keyword is not evidence of a domain, so a
|
|
85
|
+
* department counts as *implied by the request* only when at least two of its
|
|
86
|
+
* catalogue keywords occur in the user text. Named so tests pin it.
|
|
87
|
+
*/
|
|
88
|
+
export const REQUIRED_DEPARTMENT_MIN_HITS = 2
|
|
89
|
+
|
|
90
|
+
/**
|
|
91
|
+
* Wording that turns a missing department into a deliberate exclusion. The plan
|
|
92
|
+
* must say 「不安排<部门>:理由」 in taskUnderstanding.assumptions; silence is
|
|
93
|
+
* what the gate rejects.
|
|
94
|
+
*/
|
|
95
|
+
export const DEPARTMENT_EXCLUSION_CUES = /(?:不安排|不需要|未涉及|不涉及|排除|不含|无需|不启用)/
|
|
96
|
+
|
|
79
97
|
/** Hard anchors are the only elements that can fail a plan by themselves. */
|
|
80
98
|
export type TaskElementKind =
|
|
81
99
|
| 'version'
|
|
@@ -98,9 +116,15 @@ export interface TaskElement {
|
|
|
98
116
|
|
|
99
117
|
export interface TaskReference {
|
|
100
118
|
token: string
|
|
101
|
-
/**
|
|
119
|
+
/**
|
|
120
|
+
* `context` = an antecedent for this token was actually found in the recent
|
|
121
|
+
* conversation and quoted in `antecedent`; `inline-noun` = e.g. 「这个插件」;
|
|
122
|
+
* `none` = still ambiguous.
|
|
123
|
+
*/
|
|
102
124
|
resolvedBy: 'context' | 'inline-noun' | 'none'
|
|
103
125
|
resolved: boolean
|
|
126
|
+
/** The noun phrase taken from the conversation that this token points at. */
|
|
127
|
+
antecedent?: string
|
|
104
128
|
evidence: string
|
|
105
129
|
}
|
|
106
130
|
|
|
@@ -138,6 +162,13 @@ export interface DepartmentAssignment {
|
|
|
138
162
|
declaredRationale: string
|
|
139
163
|
}
|
|
140
164
|
|
|
165
|
+
/** One department the user's request implies, with the evidence words. */
|
|
166
|
+
export interface DepartmentCoverageEntry {
|
|
167
|
+
department: string
|
|
168
|
+
departmentLabel: string
|
|
169
|
+
hits: string[]
|
|
170
|
+
}
|
|
171
|
+
|
|
141
172
|
export interface TaskAnchor {
|
|
142
173
|
text: string
|
|
143
174
|
sha256: string
|
|
@@ -160,6 +191,12 @@ export interface UnderstandingAudit {
|
|
|
160
191
|
assignments: DepartmentAssignment[]
|
|
161
192
|
departmentsUsed: string[]
|
|
162
193
|
mismatches: string[]
|
|
194
|
+
/** Departments the user's own request implies (>= REQUIRED_DEPARTMENT_MIN_HITS keywords). */
|
|
195
|
+
requiredDepartments: DepartmentCoverageEntry[]
|
|
196
|
+
/** Implied departments with no subtask and no stated exclusion — the U3 gap. */
|
|
197
|
+
missingDepartments: DepartmentCoverageEntry[]
|
|
198
|
+
/** Implied departments that the plan deliberately excluded, with the stated reason. */
|
|
199
|
+
excludedDepartments: Array<DepartmentCoverageEntry & { reason: string }>
|
|
163
200
|
}
|
|
164
201
|
stage3: {
|
|
165
202
|
verdict: 'aligned' | 'deviation' | 'needs-clarification' | 'needs-context' | 'degraded'
|
|
@@ -431,22 +468,192 @@ function extractReferences(text: string): TaskReference[] {
|
|
|
431
468
|
})
|
|
432
469
|
}
|
|
433
470
|
}
|
|
434
|
-
|
|
435
|
-
|
|
436
|
-
|
|
437
|
-
|
|
438
|
-
|
|
439
|
-
|
|
440
|
-
|
|
441
|
-
|
|
442
|
-
|
|
443
|
-
|
|
444
|
-
|
|
445
|
-
|
|
471
|
+
/*
|
|
472
|
+
* [P0-2] Option letters are references only when a choice cue selects one.
|
|
473
|
+
*
|
|
474
|
+
* The old rule fired on any A-D letter as soon as the text said 建议/方案
|
|
475
|
+
* anywhere: MS2 case T05-04 ("A/B 探针…建议移入") was blocked as an unresolved
|
|
476
|
+
* pronoun it never contained, and Markdown list markers "A." would do the
|
|
477
|
+
* same. A letter now counts only when it directly follows a selection verb,
|
|
478
|
+
* which is the only structure that can actually mean "choose this option".
|
|
479
|
+
*/
|
|
480
|
+
const option = /(?:选|选项|方案|可选|建议|采用|按)\s*(?:用\s*)?([A-D])(?![A-Za-z\/])/.exec(text)
|
|
481
|
+
if (option && !seen.has(option[1])) {
|
|
482
|
+
seen.add(option[1])
|
|
483
|
+
references.push({
|
|
484
|
+
token: option[1],
|
|
485
|
+
resolvedBy: 'none',
|
|
486
|
+
resolved: false,
|
|
487
|
+
evidence: clauseOf(text, option.index, option[0].length),
|
|
488
|
+
})
|
|
446
489
|
}
|
|
447
490
|
return references
|
|
448
491
|
}
|
|
449
492
|
|
|
493
|
+
/**
|
|
494
|
+
* [P0-2] Pronoun antecedent extraction.
|
|
495
|
+
*
|
|
496
|
+
* Before this, any non-empty conversation turned every pronoun into
|
|
497
|
+
* `resolvedBy: 'context'` — the audit claimed "它 was resolved" even when the
|
|
498
|
+
* context was "进展到哪一步了" or talked about something else entirely, so a
|
|
499
|
+
* plan could be dispatched on a guess. Resolution now requires an actual noun
|
|
500
|
+
* phrase in the recent conversation:
|
|
501
|
+
*
|
|
502
|
+
* - a cue (`用户…:` / `使用` / `针对` / `把`…) followed by the phrase, or
|
|
503
|
+
* - a topic sentence (`我们在聊…`, `讨论了…`);
|
|
504
|
+
*
|
|
505
|
+
* and phrases made only of function words (`把` / `问题` / `这个`…) are refused,
|
|
506
|
+
* so an empty promise cannot be manufactured. When nothing qualifies the
|
|
507
|
+
* reference stays unresolved and the plan is blocked as `needs-context`
|
|
508
|
+
* (strict mode) or recorded as a warning (rollback switch).
|
|
509
|
+
*/
|
|
510
|
+
const ANTECEDENT_CUES = [
|
|
511
|
+
/(?:^|[::。;;\n])\s*(?:用户|需求|任务|要求|目标|对象)?\s*(?:上一轮|上一条|刚才|之前|前面)?\s*[说讲提]:?\s*/g,
|
|
512
|
+
/(?:使用|针对|关于|围绕|涉及|改造|修改|调整|优化|修复|处理|把|将|对)\s*/g,
|
|
513
|
+
/(?:和|与|同|跟)\s*[\u4e00-\u9fffA-Za-z0-9_\-\.]{2,12}?(?:相关|有关|有关联)/g,
|
|
514
|
+
// Options: "给出 A 方案(…)与 B 方案(只跑增量)", "我们在聊 X"/"讨论了 X".
|
|
515
|
+
/(?:在聊|聊到|讨论了|商量了|谈到了|提到|列出|提供|给出|可选)\s*/g,
|
|
516
|
+
]
|
|
517
|
+
const FUNCTION_WORD_ONLY = /^(?:那|这|些|个|位|次|样|里|边|面|中|上|下|我们|你们|他们|它|他|她|是|的|了|在|有|做|干|弄|搞|看|说|问|聊|讨论|事情|东西|问题|内容|功能|东西|部分|那块|这块)+$/
|
|
518
|
+
const ANTECEDENT_MAX_LENGTH = 12
|
|
519
|
+
|
|
520
|
+
export interface AntecedentResult {
|
|
521
|
+
/** The noun phrase the pronoun points at. */
|
|
522
|
+
antecedent: string
|
|
523
|
+
/** The conversation snippet the phrase was taken from. */
|
|
524
|
+
source: string
|
|
525
|
+
}
|
|
526
|
+
|
|
527
|
+
/** Refuse function-word-only fragments: they carry no referent. */
|
|
528
|
+
function isMeaningfulAntecedent(candidate: string): boolean {
|
|
529
|
+
const value = candidate.trim()
|
|
530
|
+
if (value.length < 2 || value.length > ANTECEDENT_MAX_LENGTH) return false
|
|
531
|
+
// An option label ("A 方案") is a referent: it points at a lettered choice.
|
|
532
|
+
if (/^[A-D](?:方案|选项)$/.test(value)) return true
|
|
533
|
+
/*
|
|
534
|
+
* A pronoun is not its own antecedent: "把它改一下" leaves "它" behind at the
|
|
535
|
+
* start of the clause, and accepting that would resolve 它 to itself.
|
|
536
|
+
*/
|
|
537
|
+
if (/^(?:它|他|她|这|那|其)(?:们)?$/.test(value) || /^(?:它|他|她|这|那|其)/.test(value)) return false
|
|
538
|
+
if (!/[\u4e00-\u9fffA-Za-z]/.test(value)) return false
|
|
539
|
+
return !FUNCTION_WORD_ONLY.test(value)
|
|
540
|
+
}
|
|
541
|
+
|
|
542
|
+
/**
|
|
543
|
+
* Find the phrase a pronoun points at in the recent conversation.
|
|
544
|
+
*
|
|
545
|
+
* Returns `null` — never a fabricated phrase — when the conversation names no
|
|
546
|
+
* candidate. Callers must treat `null` as "still ambiguous".
|
|
547
|
+
*/
|
|
548
|
+
export function resolvePronounAntecedent(token: string, conversationContext: string): AntecedentResult | null {
|
|
549
|
+
const text = String(conversationContext ?? '').trim()
|
|
550
|
+
if (!text || text.length > 20000) return null
|
|
551
|
+
for (const pattern of ANTECEDENT_CUES) {
|
|
552
|
+
pattern.lastIndex = 0
|
|
553
|
+
for (const match of text.matchAll(pattern)) {
|
|
554
|
+
const start = (match.index ?? 0) + match[0].length
|
|
555
|
+
const found = antecedentAfterCue(text, start)
|
|
556
|
+
if (found) return found
|
|
557
|
+
}
|
|
558
|
+
}
|
|
559
|
+
return null
|
|
560
|
+
}
|
|
561
|
+
|
|
562
|
+
/**
|
|
563
|
+
* Split one clause into candidate noun phrases.
|
|
564
|
+
*
|
|
565
|
+
* Whitespace, 、 and the connectors 与/和/及 separate phrases — but an option
|
|
566
|
+
* letter is a prefix, not a phrase of its own: "A 方案" must stay one candidate,
|
|
567
|
+
* otherwise the letter is lost and the option cannot be matched.
|
|
568
|
+
*/
|
|
569
|
+
export function splitAntecedentCandidates(clause: string): string[] {
|
|
570
|
+
const parts: string[] = []
|
|
571
|
+
let current = ''
|
|
572
|
+
const isWordChar = (char: string | undefined): boolean => !!char && /[\u4e00-\u9fffA-Za-z0-9_]/.test(char)
|
|
573
|
+
for (let index = 0; index < clause.length; index += 1) {
|
|
574
|
+
const char = clause[index]
|
|
575
|
+
if (',,、'.includes(char)) {
|
|
576
|
+
parts.push(current)
|
|
577
|
+
current = ''
|
|
578
|
+
continue
|
|
579
|
+
}
|
|
580
|
+
if ('与和及'.includes(char)) {
|
|
581
|
+
// "A 方案与 B 方案": a connector right after an option label is a separator.
|
|
582
|
+
const soFar = current.trim()
|
|
583
|
+
if (/[A-D](?:方案|选项)$/.test(soFar)) {
|
|
584
|
+
parts.push(current)
|
|
585
|
+
current = ''
|
|
586
|
+
continue
|
|
587
|
+
}
|
|
588
|
+
}
|
|
589
|
+
if (/s/.test(char) || char === ' ') {
|
|
590
|
+
const soFar = current.trim()
|
|
591
|
+
const next = clause.slice(index).trimStart()
|
|
592
|
+
// Glue an option letter to the word that follows it.
|
|
593
|
+
if (/^[A-D]$/.test(soFar) && /^(?:方案|选项)/.test(next)) {
|
|
594
|
+
current += char
|
|
595
|
+
continue
|
|
596
|
+
}
|
|
597
|
+
if (isWordChar(clause[index - 1]) || isWordChar(next[0])) {
|
|
598
|
+
parts.push(current)
|
|
599
|
+
current = ''
|
|
600
|
+
continue
|
|
601
|
+
}
|
|
602
|
+
continue
|
|
603
|
+
}
|
|
604
|
+
if ('()()'.includes(char)) {
|
|
605
|
+
// Parentheses are skipped entirely; their content is not the referent.
|
|
606
|
+
const close = clause.indexOf(char === '(' ? ')' : char === '(' ? ')' : char, index + 1)
|
|
607
|
+
if (close >= 0) {
|
|
608
|
+
index = close
|
|
609
|
+
parts.push(current)
|
|
610
|
+
current = ''
|
|
611
|
+
continue
|
|
612
|
+
}
|
|
613
|
+
}
|
|
614
|
+
current += char
|
|
615
|
+
}
|
|
616
|
+
parts.push(current)
|
|
617
|
+
return parts
|
|
618
|
+
.map((part) => part.replace(/[(())]/g, '').trim())
|
|
619
|
+
.filter((part) => part.length > 0)
|
|
620
|
+
}
|
|
621
|
+
|
|
622
|
+
function cleanAntecedentCandidate(part: string): string {
|
|
623
|
+
return part
|
|
624
|
+
.replace(/\u0000/g, '')
|
|
625
|
+
.replace(/^(?:把|将|对|使用|针对|关于|围绕|涉及|改造|修改|调整|优化|修复|处理)\s*/, '')
|
|
626
|
+
.replace(/(?:改成|改为|修正成|调整成|设成|变为|换成|变成|默认为|默认|保持|不用|不要|可以|需要|看看|检查一下|是什么|怎么样|如何).*$/, '')
|
|
627
|
+
// Parentheticals are not part of the referent: "B 方案(只跑增量)" -> "B 方案".
|
|
628
|
+
.replace(/[((][^))]*[))]/g, '')
|
|
629
|
+
.replace(/[,,。!?!?;;::、\s]+$/, '')
|
|
630
|
+
.trim()
|
|
631
|
+
}
|
|
632
|
+
|
|
633
|
+
/**
|
|
634
|
+
* Read the clause that starts at `start` and return the first meaningful noun
|
|
635
|
+
* phrase in it. A verb left at the start (`把…`) is dropped by
|
|
636
|
+
* `cleanAntecedentCandidate`; the clause never crosses a sentence boundary.
|
|
637
|
+
*/
|
|
638
|
+
function antecedentAfterCue(text: string, start: number): AntecedentResult | null {
|
|
639
|
+
const tail = text.slice(start)
|
|
640
|
+
const stop = tail.search(/[。!?!?\n;;]/)
|
|
641
|
+
const clause = (stop >= 0 ? tail.slice(0, stop) : tail).trim()
|
|
642
|
+
if (!clause) return null
|
|
643
|
+
/*
|
|
644
|
+
* The clause may chain several noun phrases ("A 方案…与 B 方案…"). Split at
|
|
645
|
+
* the connector and at whitespace so each candidate is one noun phrase; the
|
|
646
|
+
* caller compares them to the pronoun token and picks the matching one.
|
|
647
|
+
*/
|
|
648
|
+
const parts = splitAntecedentCandidates(clause)
|
|
649
|
+
for (const part of parts) {
|
|
650
|
+
const cleaned = cleanAntecedentCandidate(part).replace(/^[((]|[))]$/g, '')
|
|
651
|
+
if (isMeaningfulAntecedent(cleaned)) return { antecedent: cleaned, source: clause.slice(0, 80) }
|
|
652
|
+
}
|
|
653
|
+
return null
|
|
654
|
+
}
|
|
655
|
+
|
|
656
|
+
|
|
450
657
|
function extractConflicts(text: string): TaskConflict[] {
|
|
451
658
|
const conflicts: TaskConflict[] = []
|
|
452
659
|
const seen = new Set<string>()
|
|
@@ -551,6 +758,57 @@ function flattenPlanSubtasks(plan: any): any[] {
|
|
|
551
758
|
return subtasks
|
|
552
759
|
}
|
|
553
760
|
|
|
761
|
+
/**
|
|
762
|
+
* [P0-1] Which department domains the user's request itself implies, and which
|
|
763
|
+
* of them the plan never addresses.
|
|
764
|
+
*
|
|
765
|
+
* Root cause this closes: before 1.12.1 the audit only asked "does the chosen
|
|
766
|
+
* department fit?" — a plan could omit an entire domain the request names and
|
|
767
|
+
* still pass, because nothing asked "is a domain missing?". MS2 measured the
|
|
768
|
+
* consequence: 11/29 required departments covered.
|
|
769
|
+
*
|
|
770
|
+
* A missing department is forgiven only when the plan states the exclusion in
|
|
771
|
+
* `taskUnderstanding.assumptions` / `unknowns` with one of
|
|
772
|
+
* DEPARTMENT_EXCLUSION_CUES, so omission must be a written decision instead of
|
|
773
|
+
* silence. Pure and deterministic: no model, no I/O.
|
|
774
|
+
*/
|
|
775
|
+
export function auditDepartmentCoverage(taskText: string, plan: any): {
|
|
776
|
+
required: DepartmentCoverageEntry[]
|
|
777
|
+
missing: DepartmentCoverageEntry[]
|
|
778
|
+
excluded: Array<DepartmentCoverageEntry & { reason: string }>
|
|
779
|
+
} {
|
|
780
|
+
const text = typeof taskText === 'string' ? taskText.trim() : ''
|
|
781
|
+
if (text === '') return { required: [], missing: [], excluded: [] }
|
|
782
|
+
const used = new Set<string>(
|
|
783
|
+
flattenPlanSubtasks(plan)
|
|
784
|
+
.map((subtask) => (typeof subtask?.department === 'string' ? subtask.department : ''))
|
|
785
|
+
.filter((department) => department !== ''),
|
|
786
|
+
)
|
|
787
|
+
const statements: string[] = []
|
|
788
|
+
for (const key of ['assumptions', 'unknowns']) {
|
|
789
|
+
const list = plan?.taskUnderstanding?.[key]
|
|
790
|
+
if (!Array.isArray(list)) continue
|
|
791
|
+
for (const item of list) if (typeof item === 'string' && item.trim() !== '') statements.push(item.trim())
|
|
792
|
+
}
|
|
793
|
+
const required: DepartmentCoverageEntry[] = []
|
|
794
|
+
const missing: DepartmentCoverageEntry[] = []
|
|
795
|
+
const excluded: Array<DepartmentCoverageEntry & { reason: string }> = []
|
|
796
|
+
for (const capability of DEPARTMENT_CAPABILITIES) {
|
|
797
|
+
const hits = matchDepartmentKeywords(text, capability.id)
|
|
798
|
+
if (hits.length < REQUIRED_DEPARTMENT_MIN_HITS) continue
|
|
799
|
+
const entry: DepartmentCoverageEntry = { department: capability.id, departmentLabel: capability.zh, hits }
|
|
800
|
+
required.push(entry)
|
|
801
|
+
if (used.has(capability.id)) continue
|
|
802
|
+
const reason = statements.find((statement) => (
|
|
803
|
+
(statement.includes(capability.zh) || statement.includes(capability.id))
|
|
804
|
+
&& DEPARTMENT_EXCLUSION_CUES.test(statement)
|
|
805
|
+
))
|
|
806
|
+
if (reason) excluded.push({ ...entry, reason: reason.slice(0, 300) })
|
|
807
|
+
else missing.push(entry)
|
|
808
|
+
}
|
|
809
|
+
return { required, missing, excluded }
|
|
810
|
+
}
|
|
811
|
+
|
|
554
812
|
/**
|
|
555
813
|
* Explain one department choice with the catalog keywords that support it.
|
|
556
814
|
*
|
|
@@ -682,7 +940,7 @@ function emptyAudit(anchor: TaskAnchor | null, reason: string): UnderstandingAud
|
|
|
682
940
|
version: UNDERSTANDING_AUDIT_VERSION,
|
|
683
941
|
anchor,
|
|
684
942
|
stage1: { elements: [], hardCount: 0, references: [], conflicts: [], deliverableClauses: [] },
|
|
685
|
-
stage2: { assignments: [], departmentsUsed: [], mismatches: [] },
|
|
943
|
+
stage2: { assignments: [], departmentsUsed: [], mismatches: [], requiredDepartments: [], missingDepartments: [], excludedDepartments: [] },
|
|
686
944
|
stage3: {
|
|
687
945
|
verdict: 'degraded',
|
|
688
946
|
rewritten: false,
|
|
@@ -713,6 +971,7 @@ export function auditTaskUnderstanding(plan: any, context: PlanAuditContext = {}
|
|
|
713
971
|
const planText = collectPlanText(plan)
|
|
714
972
|
const assignments = flattenPlanSubtasks(plan).map((subtask) => departmentFit(subtask, taskText))
|
|
715
973
|
const departmentsUsed = [...new Set(assignments.map((entry) => entry.department))]
|
|
974
|
+
const coverage = auditDepartmentCoverage(taskText, plan)
|
|
716
975
|
const mismatches = assignments
|
|
717
976
|
.filter((entry) => entry.verdict === 'mismatch')
|
|
718
977
|
.map((entry) => `${entry.subtaskId || '(无 ID)'} → ${entry.departmentLabel}:${entry.reason}`)
|
|
@@ -741,15 +1000,25 @@ export function auditTaskUnderstanding(plan: any, context: PlanAuditContext = {}
|
|
|
741
1000
|
const hardTotal = extracted.elements.filter((element) => element.hard).length
|
|
742
1001
|
const summaryCoverage = hardTotal === 0 ? 1 : (hardTotal - droppedFromSummary.length) / hardTotal
|
|
743
1002
|
|
|
744
|
-
|
|
745
|
-
|
|
746
|
-
|
|
747
|
-
|
|
748
|
-
|
|
749
|
-
|
|
750
|
-
|
|
751
|
-
|
|
752
|
-
)
|
|
1003
|
+
/*
|
|
1004
|
+
* [P0-2] Resolution requires an antecedent, not merely a non-empty context.
|
|
1005
|
+
* The old rule marked every pronoun resolved as soon as any conversation
|
|
1006
|
+
* existed, so "把它改成默认收起" passed with a context that talked about
|
|
1007
|
+
* something else. Now the phrase is extracted and quoted; otherwise the
|
|
1008
|
+
* reference stays unresolved and blocks (or warns in rollback mode).
|
|
1009
|
+
*/
|
|
1010
|
+
const conversation = typeof context.conversationContext === 'string' ? context.conversationContext : ''
|
|
1011
|
+
const resolvedByContext = extracted.references.map((reference) => {
|
|
1012
|
+
if (reference.resolved) return reference
|
|
1013
|
+
const found = resolvePronounAntecedent(reference.token, conversation)
|
|
1014
|
+
if (!found) return reference
|
|
1015
|
+
// Option letters are resolved only when the conversation lists that option
|
|
1016
|
+
// as a listed choice ("A 方案" / "B 选项"), not a bare letter in a word.
|
|
1017
|
+
const listedOption = new RegExp('(?:^|[^A-Za-z])' + reference.token + '\\s*(?:方案|选项)')
|
|
1018
|
+
if (/^[A-D]$/.test(reference.token) && !listedOption.test(conversation)) return reference
|
|
1019
|
+
return { ...reference, resolved: true, resolvedBy: 'context' as const, antecedent: found.antecedent }
|
|
1020
|
+
})
|
|
1021
|
+
const unresolved = resolvedByContext.filter((reference) => !reference.resolved)
|
|
753
1022
|
|
|
754
1023
|
const reasons: string[] = []
|
|
755
1024
|
let verdict: UnderstandingAudit['stage3']['verdict'] = 'aligned'
|
|
@@ -757,6 +1026,10 @@ export function auditTaskUnderstanding(plan: any, context: PlanAuditContext = {}
|
|
|
757
1026
|
verdict = 'deviation'
|
|
758
1027
|
reasons.push(`用户原要求中的 ${missingHardAnchors.length} 个硬要素在计划中完全缺失:${missingHardAnchors.slice(0, 6).join('、')}`)
|
|
759
1028
|
}
|
|
1029
|
+
if (coverage.missing.length > 0) {
|
|
1030
|
+
verdict = 'deviation'
|
|
1031
|
+
reasons.push(`用户原要求涉及 ${coverage.missing.length} 个部门领域,计划既未安排也未写明排除理由:${coverage.missing.map((entry) => entry.departmentLabel).join('、')}`)
|
|
1032
|
+
}
|
|
760
1033
|
if (mismatches.length > 0) {
|
|
761
1034
|
if (verdict === 'aligned') verdict = 'deviation'
|
|
762
1035
|
reasons.push(`${mismatches.length} 个子任务的部门与用户原要求矛盾:${mismatches.slice(0, 4).join(';')}`)
|
|
@@ -767,7 +1040,7 @@ export function auditTaskUnderstanding(plan: any, context: PlanAuditContext = {}
|
|
|
767
1040
|
}
|
|
768
1041
|
if (unresolved.length > 0) {
|
|
769
1042
|
if (verdict === 'aligned') verdict = 'needs-context'
|
|
770
|
-
reasons.push(
|
|
1043
|
+
reasons.push(`含指代的任务在最近对话里找不到先行词:${unresolved.map((reference) => reference.token).join('、')},不应在未澄清前派发`)
|
|
771
1044
|
}
|
|
772
1045
|
if (verdict === 'aligned') {
|
|
773
1046
|
reasons.push(`硬要素 ${hardTotal - missingHardAnchors.length}/${hardTotal} 全部保留;${assignments.length} 个子任务的部门有目录关键词支撑`)
|
|
@@ -783,7 +1056,14 @@ export function auditTaskUnderstanding(plan: any, context: PlanAuditContext = {}
|
|
|
783
1056
|
conflicts: extracted.conflicts,
|
|
784
1057
|
deliverableClauses: extracted.deliverableClauses,
|
|
785
1058
|
},
|
|
786
|
-
stage2: {
|
|
1059
|
+
stage2: {
|
|
1060
|
+
assignments,
|
|
1061
|
+
departmentsUsed,
|
|
1062
|
+
mismatches,
|
|
1063
|
+
requiredDepartments: coverage.required,
|
|
1064
|
+
missingDepartments: coverage.missing,
|
|
1065
|
+
excludedDepartments: coverage.excluded,
|
|
1066
|
+
},
|
|
787
1067
|
stage3: {
|
|
788
1068
|
verdict,
|
|
789
1069
|
rewritten: !verbatim,
|
|
@@ -831,11 +1111,19 @@ export function validatePlanContext(plan: any, context: PlanAuditContext = {}):
|
|
|
831
1111
|
errors.push(`CLARIFY: ${audit.stage3.reasons.find((reason) => reason.includes('自相矛盾')) ?? '用户要求存在矛盾,必须先向用户确认'}`)
|
|
832
1112
|
}
|
|
833
1113
|
if (audit.stage3.verdict === 'needs-context') {
|
|
834
|
-
errors.push(`CONTEXT: ${audit.stage3.reasons.find((reason) => reason.includes('指代')) ?? '
|
|
1114
|
+
errors.push(`CONTEXT: ${audit.stage3.reasons.find((reason) => reason.includes('指代')) ?? '任务含指代但在最近对话中找不到先行词,必须先向用户确认'}`)
|
|
835
1115
|
}
|
|
836
1116
|
for (const missing of audit.stage3.missingHardAnchors.slice(0, 6)) {
|
|
837
1117
|
errors.push(`DEVIATION: 用户原要求中的硬要素「${missing}」在计划中完全缺失,请补回或说明为什么可以省略`)
|
|
838
1118
|
}
|
|
1119
|
+
/*
|
|
1120
|
+
* [P0-1] Missing domain coverage is blocking, not a warning: MS2 showed a
|
|
1121
|
+
* plan can look internally consistent while silently dropping design/legal/
|
|
1122
|
+
* security/devops, which is exactly what "必须高" cannot tolerate.
|
|
1123
|
+
*/
|
|
1124
|
+
for (const entry of audit.stage2.missingDepartments.slice(0, 6)) {
|
|
1125
|
+
errors.push(`COVERAGE: 用户原要求涉及「${entry.departmentLabel}」领域(命中 ${entry.hits.slice(0, 4).join('、')}),计划既没有该部门的子任务,也没有在 taskUnderstanding.assumptions 写明「不安排${entry.departmentLabel}:理由」`)
|
|
1126
|
+
}
|
|
839
1127
|
for (const assignment of audit.stage2.assignments) {
|
|
840
1128
|
if (assignment.verdict !== 'mismatch') continue
|
|
841
1129
|
errors.push(`DEVIATION: 子任务 ${assignment.subtaskId || '(无 ID)'} 分配给「${assignment.departmentLabel}」缺乏依据:${assignment.reason}`)
|
|
@@ -879,6 +1167,22 @@ export function compactUnderstandingAudit(audit: UnderstandingAudit): any {
|
|
|
879
1167
|
})),
|
|
880
1168
|
departmentsUsed: audit.stage2.departmentsUsed,
|
|
881
1169
|
mismatches: audit.stage2.mismatches,
|
|
1170
|
+
requiredDepartments: audit.stage2.requiredDepartments.map((entry) => ({
|
|
1171
|
+
department: entry.department,
|
|
1172
|
+
departmentLabel: entry.departmentLabel,
|
|
1173
|
+
hits: entry.hits.slice(0, 4),
|
|
1174
|
+
})),
|
|
1175
|
+
missingDepartments: audit.stage2.missingDepartments.map((entry) => ({
|
|
1176
|
+
department: entry.department,
|
|
1177
|
+
departmentLabel: entry.departmentLabel,
|
|
1178
|
+
hits: entry.hits.slice(0, 4),
|
|
1179
|
+
})),
|
|
1180
|
+
excludedDepartments: audit.stage2.excludedDepartments.map((entry) => ({
|
|
1181
|
+
department: entry.department,
|
|
1182
|
+
departmentLabel: entry.departmentLabel,
|
|
1183
|
+
hits: entry.hits.slice(0, 4),
|
|
1184
|
+
reason: entry.reason,
|
|
1185
|
+
})),
|
|
882
1186
|
},
|
|
883
1187
|
stage3: {
|
|
884
1188
|
verdict: audit.stage3.verdict,
|