Skip to content

Commit ca77c4b

Browse files
authored
Merge pull request #503 from slow-stack/fix/dream-list-fence
fix(dream): 5 处 prompt 补围栏,解析失败进审计(#501 / #493 另一半)
2 parents 3ee88da + 9d5cd72 commit ca77c4b

18 files changed

Lines changed: 411 additions & 56 deletions

‎dsh-mneme/CHANGELOG.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2,6 +2,10 @@
22

33
## [Unreleased]
44

5+
## 🐛 修复
6+
7+
- **dream / sleep 侧那 5 处 prompt 一直没有围栏:巩固把 46K 清单读成「没人给我下指令」,约 43 万 token/天白烧(#501,社区报告 #493 追加)**:#493 修的是「模型把喂进去的会话转录当成自己的发言」,但围栏只装在 `autoSummarize` 那一处——`src/summarize.js` 之外,`dream.js` 的叙述条/巩固/摘要与 `sleep.js` 的冲突裁决/模式提炼这 5 处一直是把清单**裸放**在 user 消息里,system 之后没有任何分隔与角色约束。现场:`dream_consolidate` 9 次失败、输入 47K→55K tokens 而输出只有 212–907 tokens,日志里模型说「收到——这是一包记忆检索结果,但没有带具体指令,我先过一遍给你一份导览」,于是开始寒暄;报告者的对照实验同时说明方向是对的——只加 system 约束那一臂从 2/2 掉到 0/1,围栏 + 角色重申那一臂仍 1/1。修法:① `lang.js` 新增**记忆清单专用**的成对围栏与角色声明,不复用「会话转录」那对名字(喂的是清单,标错名字会让读审计与读日志的人以为这次喂进去的是一段对话);角色声明落在 **user 段开头**——#493 实测判断「这段文本是什么」的依据是它所在的那条 user 消息,只写 system 救不回来。② `fenceListBody()` 是纯函数(同一批记忆在下次阈值到来时会被原样再喂一次,重试收到的 prompt 必须与首次逐字节相同,否则「同输入同错」这条诊断依据就没了),并把正文里的 `<<<`/`>>>` 用零宽空格断链——**只中和标记、不删内容**:不然一条写着「<<<记忆清单结束>>>」的记忆就能提前闭合围栏,把后面的文字升格成指令,而记忆正文本来就是模型的可影响输入。③ 失败行的审计 `metadata` 不再为 null:`runAuditedLlm` / `auditedSleepLlm` 加 `spec.metadataFrom(result)`,dream 侧写 `parse_failure` / `raw_len` / `raw_head` / `raw_has_bracket`(`raw_has_bracket` 是区分「压根没按格式写」与「格式对但内容救不回来」的最低成本判据;报告者另指的 `raw_tail` 缺口属 #502)。审计完全复用 dream 既有那条缝,只往导出列表加名字,不重写第三份 chunk 适配。④ 顺带修掉一处**测试侧的脆弱**:多个 mock 靠「user 段以 `id=` / `候选冲突` 开头」认阶段,围栏让首行变成角色声明;改用清单形状(`| type=`)与 `options.purpose`。这不是迁就测试——`dream-candidate-hybrid.test.js` 早就注释说过前缀判据在语义预分组下会把巩固误判成摘要、只能绕开共用 mock,现在那个绕法不再需要。回归锁 5 条(围栏结构、确定性、提前闭合、consolidate 实测带围栏且失败行有读数、5 处不许再裸放的形状锁),变异自证三条(摘掉中和 → 只有提前闭合那条红;摘掉 `metadataFrom` → 只有审计那条红;拆开 consolidate 的围栏 → 审计与形状锁两条红)。评审补的两处:① `fenceListBody` 带输出契约参数——总览摘要要的是散文,对它说「只输出 JSON 数组」会让模型回 `[]`,而 `summaryText.trim()` 判非空通过,方括号就这么被存成库内总览(Major;散文契约用独立的角色声明,叙述条保留 JSON 那份,它的 system 确实要严格数组);② 流以错误结束、根本没有正文时不再写 `parse_failure` 与 `raw_len: 0`,否则一次限流会被读成「模型没按格式写」。两处各自做过变异自证。
8+
59
## [0.9.4] - 2026-10-21
610

711
## 🆕 新增

‎dsh-mneme/lib/dream.js‎

Lines changed: 57 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -313,6 +313,45 @@ function describeStreamFailure(reason) {
313313
* operation_type, model_id, related_memory_ids); `body(reportUsage)` performs
314314
* the actual stream consumption and is handed a usage reporter for the chunks.
315315
*/
316+
/**
317+
* 把喂给模型的清单包进围栏(#493 的另一半:dream / sleep 那 5 处 user 消息以前是裸放)。
318+
* 三个细节各自有原因,别当装饰删:
319+
* - 角色声明落在**开头**:判断「这段文本是什么」的是它所在的那条 user 消息,只写在 system
320+
* 里实测救不回来(#493 的 A/C 臂:只加 system 约束仍 0/2 出 JSON);
321+
* - 正文里的 `<<<` / `>>>` 用零宽空格断链——**只中和标记,不删内容**。不这么做时,一条写着
322+
* 「<<<记忆清单结束>>>」的记忆就能提前闭合围栏,把后面的文字升格成指令;而记忆正文本来就
323+
* 是模型的可影响输入,这是最省事的注入入口;
324+
* - 纯函数、不带 nonce:同一批记忆在下一次巩固阈值到来时会被原样再喂一次,重试收到的 prompt
325+
* 必须与首次逐字节相同,否则「同输入同错」这条诊断依据就没了(#493 的重试 20 次现场就靠它)。
326+
* @param {"json"|"text"} [shape] 该次调用的**输出契约**:巩固/叙述/冲突/模式要 JSON 数组,
327+
* 总览摘要要散文。角色声明里那句「只输出 JSON 数组」必须跟着契约走——对散文契约说它,
328+
* 模型会回 `[]`,而 `summaryText.trim()` 判非空通过,`[]` 就被当成摘要存进库(评审实测)。
329+
*/
330+
export function fenceListBody(language, text, shape = "json") {
331+
const open = STR.memoryFenceOpen[language] ?? STR.memoryFenceOpen.en;
332+
const close = STR.memoryFenceClose[language] ?? STR.memoryFenceClose.en;
333+
const table = shape === "text" ? STR.dreamTextRoleReminder : STR.dreamRoleReminder;
334+
const reminder = table[language] ?? table.en;
335+
const body = String(text ?? "").replace(/<<<|>>>/g, (run) => `${run[0]}\u200B${run.slice(1)}`);
336+
return `${reminder}\n\n${open}\n${body}\n${close}`;
337+
}
338+
339+
/**
340+
* 解析失败行的审计读数(#493 同款,报告者提的第 2 条)。dream 的 warn 日志本来就带
341+
* raw length + 前 300 字,缺的是**可查询**的那一面:日志要翻、按不了、也进不了面板的
342+
* 「LLM 消耗」。`raw_has_bracket` 是区分「模型压根没按格式写」与「格式对但内容救不回来」
343+
* 的唯一低成本判据。
344+
*/
345+
export function parseFailureMeta(text) {
346+
const raw = typeof text === "string" ? text : "";
347+
return {
348+
parse_failure: true,
349+
raw_len: raw.length,
350+
raw_head: raw.slice(0, 300).replace(/\s+/g, " ").trim(),
351+
raw_has_bracket: raw.includes("[")
352+
};
353+
}
354+
316355
async function runAuditedLlm(ctx, service, config, spec, body) {
317356
const audit = config?.llmAudit;
318357
if (audit?.enabled === false || typeof service?.saveLlmAudit !== "function") return body(() => {});
@@ -355,6 +394,17 @@ async function runAuditedLlm(ctx, service, config, spec, body) {
355394
errorMessage = String(error?.message ?? error);
356395
throw error;
357396
} finally {
397+
// spec.metadataFrom(result):调用方按**这一轮的原始输出**决定要不要带读数(解析成功就
398+
// 返回 null → 整键不写)。判据函数必须是纯的:它读的是 result,不读控制流里的缓存,
399+
// 否则「审计说了什么」会随调用方走到哪一行而变(sleep.js 同一条硬约定)。
400+
let metadata;
401+
if (result !== undefined && typeof spec.metadataFrom === "function") {
402+
try {
403+
metadata = spec.metadataFrom(result) ?? undefined;
404+
} catch {
405+
metadata = undefined; // 读数失败不许把审计行整个弄没
406+
}
407+
}
358408
try {
359409
service.saveLlmAudit({
360410
timestamp,
@@ -368,7 +418,8 @@ async function runAuditedLlm(ctx, service, config, spec, body) {
368418
duration_ms: Date.now() - startedAt,
369419
status,
370420
error_message: errorMessage,
371-
related_memory_ids: spec.relatedMemoryIds ?? []
421+
related_memory_ids: spec.relatedMemoryIds ?? [],
422+
...(metadata ? { metadata } : {})
372423
});
373424
} catch (auditError) {
374425
ctx.logger?.warn?.(`dsh-mneme: llm audit write failed: ${String(auditError)}`);
@@ -515,7 +566,7 @@ async function generateNarratives({ ctx, service, config, route, language, effor
515566
...(withEffort && effort ? { reasoningEffort: effort } : {}),
516567
messages: [
517568
{ role: "system", content: [{ type: "text", text: STR.prompts.narrative[language] }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } },
518-
{ role: "user", content: [{ type: "text", text: listing }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
569+
{ role: "user", content: [{ type: "text", text: fenceListBody(language, listing) }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
519570
]
520571
}, reportUsage, (reason) => { streamFailure = describeStreamFailure(reason); }));
521572
};
@@ -1210,7 +1261,8 @@ export function createDreamScheduler({ onRun, thresholdCount = 10, thresholdChar
12101261
auditError: (text) => {
12111262
decisions = extractJsonArray(text);
12121263
return Array.isArray(decisions) ? null : "no json array in llm output";
1213-
}
1264+
},
1265+
metadataFrom: (text) => (Array.isArray(decisions) ? null : parseFailureMeta(text))
12141266
}, (reportUsage) => streamText(ctx, {
12151267
provider: route.provider,
12161268
model: route.model,
@@ -1219,7 +1271,7 @@ export function createDreamScheduler({ onRun, thresholdCount = 10, thresholdChar
12191271
...(withEffort && effort ? { reasoningEffort: effort } : {}),
12201272
messages: [
12211273
{ role: "system", content: [{ type: "text", text: consolidationPrompt }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } },
1222-
{ role: "user", content: [{ type: "text", text: listText }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
1274+
{ role: "user", content: [{ type: "text", text: fenceListBody(language, listText) }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
12231275
]
12241276
}, reportUsage, (reason) => { streamFailure = describeStreamFailure(reason); }));
12251277
};
@@ -1462,7 +1514,7 @@ export function createDreamScheduler({ onRun, thresholdCount = 10, thresholdChar
14621514
...(withEffort && effort ? { reasoningEffort: effort } : {}),
14631515
messages: [
14641516
{ role: "system", content: [{ type: "text", text: STR.prompts.dreamSummary[language] }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } },
1465-
{ role: "user", content: [{ type: "text", text: summaryInputs.map((m) => `- ${m.title}: ${m.content}`).join("\n") }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
1517+
{ role: "user", content: [{ type: "text", text: fenceListBody(language, summaryInputs.map((m) => `- ${m.title}: ${m.content}`).join("\n"), "text") }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
14661518
]
14671519
}, reportUsage, (reason) => { summaryStreamFailure = describeStreamFailure(reason); }));
14681520
};

‎dsh-mneme/lib/dream/sleep.js‎

Lines changed: 28 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -21,7 +21,7 @@ import { randomUUID, createHash } from "node:crypto";
2121
import { validateDecisions, applyDecisions, ACTIONS } from "./decisions.js";
2222
import { findPotentialConflicts, cosineSimilarity } from "./clustering.js";
2323
import { scopeKeyOf } from "../scope.js";
24-
import { buildReceipt, describeStreamFailure, resolveDreamEffort, withEffortFallback, withMaxTokensFallback, maintainIndexAfterDream } from "../dream.js";
24+
import { buildReceipt, describeStreamFailure, resolveDreamEffort, withEffortFallback, withMaxTokensFallback, maintainIndexAfterDream, fenceListBody, parseFailureMeta } from "../dream.js";
2525
import { computeHeat } from "../heat.js";
2626
// 四个相位读的都是行池,而动行的写手不经汇流口:phaseDemotion 直接 setArchived /
2727
// demoteToSummary,phasePatterns 还会 mint 出新行,所以这条契约要在每处各挡一遍
@@ -106,8 +106,11 @@ async function auditedSleepLlm(ctx, service, config, logger, spec, body) {
106106
let outputTokens = 0;
107107
let status = "success";
108108
let errorMessage = null;
109+
// 提到 try 之外:finally 里的 metadataFrom 要读这一轮的原始输出(dream.js 的
110+
// runAuditedLlm 就是这么声明的)。留在 try 里,finally 看不见它,读数会静默变空。
111+
let result;
109112
try {
110-
const result = await body((usage) => {
113+
result = await body((usage) => {
111114
if (!usage) return;
112115
const i = usage.input_tokens ?? usage.inputTokens ?? usage.prompt_tokens ?? usage.promptTokens;
113116
const o = usage.output_tokens ?? usage.outputTokens ?? usage.completion_tokens ?? usage.completionTokens;
@@ -133,6 +136,17 @@ async function auditedSleepLlm(ctx, service, config, logger, spec, body) {
133136
errorMessage = String(error?.message ?? error);
134137
throw error;
135138
} finally {
139+
// spec.metadataFrom(result):与 dream.js 的 runAuditedLlm 同一形状(这里是一份并行
140+
// 实现,不是同一个函数——两份审计写法的字段必须一致,否则面板按 operation_type
141+
// 筛出来的读数会时有时没有)。判据函数只读 result,不读闭包,理由同 dream.js。
142+
let metadata;
143+
if (result !== undefined && typeof spec.metadataFrom === "function") {
144+
try {
145+
metadata = spec.metadataFrom(result) ?? undefined;
146+
} catch {
147+
metadata = undefined;
148+
}
149+
}
136150
try {
137151
service.saveLlmAudit({
138152
timestamp,
@@ -146,7 +160,8 @@ async function auditedSleepLlm(ctx, service, config, logger, spec, body) {
146160
duration_ms: Date.now() - startedAt,
147161
status,
148162
error_message: errorMessage,
149-
related_memory_ids: spec.relatedMemoryIds ?? []
163+
related_memory_ids: spec.relatedMemoryIds ?? [],
164+
...(metadata ? { metadata } : {})
150165
});
151166
} catch (auditError) {
152167
logger?.warn?.(`dsh-mneme sleep: llm audit write failed: ${String(auditError)}`);
@@ -346,7 +361,10 @@ async function phaseConflicts(ctx, service, config, logger, runId, semantic = nu
346361
// 解析结果存进闭包给下面的控制流用——审计关闭时 auditedSleepLlm 整段早退、
347362
// 这个钩子不会被调用,控制流会跟着失效(dream 的既有坑)。代价是审计开启时
348363
// 多解析一次,换来两条路径彻底解耦。
349-
auditError: (text) => (parseJsonArray(text) ? null : "no json array in llm output")
364+
auditError: (text) => (parseJsonArray(text) ? null : "no json array in llm output"),
365+
// 纯判据,与 auditError 同一条约定:重解析一次、不读闭包(审计关掉时那些闭包
366+
// 根本不会被赋值,读它会让「审计说了什么」依赖调用方走到了哪一行)。
367+
metadataFrom: (text) => (parseJsonArray(text) ? null : parseFailureMeta(text))
350368
}, (reportUsage) => streamText(ctx, {
351369
provider: route.provider,
352370
model: route.model,
@@ -355,7 +373,7 @@ async function phaseConflicts(ctx, service, config, logger, runId, semantic = nu
355373
...(withEffort && sleepEffort ? { reasoningEffort: sleepEffort } : {}),
356374
messages: [
357375
{ role: "system", content: [{ type: "text", text: conflictPrompt }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } },
358-
{ role: "user", content: [{ type: "text", text: listText }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
376+
{ role: "user", content: [{ type: "text", text: fenceListBody(language, listText) }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
359377
]
360378
}, reportUsage, (reason) => { conflictStreamFailure = describeStreamFailure(reason); }));
361379
};
@@ -547,7 +565,10 @@ async function phasePatterns(ctx, service, config, logger, runId, signal = null)
547565
relatedMemoryIds: memories.map((m) => m.id),
548566
streamError: () => patternStreamFailure,
549567
// 纯判据,同 conflict 阶段:不把解析结果存进闭包供控制流使用。
550-
auditError: (text) => (parseJsonArray(text) ? null : "no json array in llm output")
568+
auditError: (text) => (parseJsonArray(text) ? null : "no json array in llm output"),
569+
// 纯判据,与 auditError 同一条约定:重解析一次、不读闭包(审计关掉时那些闭包
570+
// 根本不会被赋值,读它会让「审计说了什么」依赖调用方走到了哪一行)。
571+
metadataFrom: (text) => (parseJsonArray(text) ? null : parseFailureMeta(text))
551572
}, (reportUsage) => streamText(ctx, {
552573
provider: route.provider,
553574
model: route.model,
@@ -556,7 +577,7 @@ async function phasePatterns(ctx, service, config, logger, runId, signal = null)
556577
...(withEffort && sleepEffort ? { reasoningEffort: sleepEffort } : {}),
557578
messages: [
558579
{ role: "system", content: [{ type: "text", text: STR.prompts.pattern[language].replace("N", String(maxPatterns)) }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } },
559-
{ role: "user", content: [{ type: "text", text: listText }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
580+
{ role: "user", content: [{ type: "text", text: fenceListBody(language, listText) }], source: { kind: "plugin:dsh-mneme", plugin: "dsh-mneme" } }
560581
]
561582
}, reportUsage, (reason) => { patternStreamFailure = describeStreamFailure(reason); }));
562583
};

0 commit comments

Comments
 (0)