@octopi-agent/engine 0.90.11 → 0.97.15
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/dist/config-admin/agent-presets.d.ts +70 -0
- package/dist/config-admin/agent-presets.js +243 -0
- package/dist/config-admin/discover.d.ts +30 -0
- package/dist/config-admin/discover.js +48 -0
- package/dist/config-admin/env.d.ts +22 -0
- package/dist/config-admin/env.js +41 -0
- package/dist/config-admin/index.d.ts +20 -0
- package/dist/config-admin/index.js +15 -0
- package/dist/config-admin/model-presets.d.ts +75 -0
- package/dist/config-admin/model-presets.js +249 -0
- package/dist/config-admin/name-guard.d.ts +14 -0
- package/dist/config-admin/name-guard.js +24 -0
- package/dist/config-admin/patch.d.ts +56 -0
- package/dist/config-admin/patch.js +201 -0
- package/dist/config-admin/port.d.ts +68 -0
- package/dist/config-admin/port.js +21 -0
- package/dist/config-admin/redact.d.ts +39 -0
- package/dist/config-admin/redact.js +42 -0
- package/dist/config-schema/agent.d.ts +43 -1
- package/dist/config-schema/context.d.ts +6 -1
- package/dist/config-schema/context.js +8 -0
- package/dist/config-schema/engine.d.ts +106 -4
- package/dist/config-schema/governance.d.ts +43 -1
- package/dist/config-schema/governance.js +157 -0
- package/dist/config-schema/models.d.ts +5 -2
- package/dist/config-schema/models.js +4 -1
- package/dist/config-schema/observability.d.ts +9 -0
- package/dist/config-schema/observability.js +29 -0
- package/dist/config-schema/run.d.ts +3 -0
- package/dist/config-schema/run.js +14 -0
- package/dist/config.d.ts +29 -4
- package/dist/config.js +15 -4
- package/dist/harness/activation/dispatcher.d.ts +8 -2
- package/dist/harness/activation/dispatcher.js +33 -5
- package/dist/harness/activation/inbox.d.ts +87 -0
- package/dist/harness/activation/inbox.js +169 -0
- package/dist/harness/activation/runtime.d.ts +9 -8
- package/dist/harness/activation/runtime.js +100 -8
- package/dist/harness/activation/types.d.ts +19 -0
- package/dist/harness/activation/types.js +2 -0
- package/dist/harness/agent/builder.d.ts +15 -1
- package/dist/harness/agent/builder.js +31 -10
- package/dist/harness/agent/config-bridge.d.ts +5 -0
- package/dist/harness/agent/config-bridge.js +3 -0
- package/dist/harness/agent/config-integrity.d.ts +72 -0
- package/dist/harness/agent/config-integrity.js +146 -0
- package/dist/harness/agent/wire-agent-builder.js +2 -1
- package/dist/harness/collaboration/autonomous-subsystem/think/executor.js +3 -1
- package/dist/harness/collaboration/delegation/child-run.js +21 -5
- package/dist/harness/collaboration/delegation/notifier.js +4 -3
- package/dist/harness/collaboration/delegation/service.js +2 -3
- package/dist/harness/collaboration/delegation/wiring.js +6 -5
- package/dist/harness/collaboration/orchestration/reflector/llm-reflector.d.ts +1 -1
- package/dist/harness/collaboration/orchestration/reflector/llm-reflector.js +1 -1
- package/dist/harness/context/capabilities/compact/engine.d.ts +30 -0
- package/dist/harness/context/capabilities/compact/engine.js +198 -46
- package/dist/harness/context/capabilities/compact/handover.d.ts +78 -0
- package/dist/harness/context/capabilities/compact/handover.js +107 -0
- package/dist/harness/context/capabilities/compact/index.d.ts +3 -1
- package/dist/harness/context/capabilities/compact/index.js +2 -1
- package/dist/harness/context/capabilities/compact/types.d.ts +30 -1
- package/dist/harness/context/capabilities/summary/executor.js +2 -1
- package/dist/harness/context/capabilities/summary/policies.js +8 -5
- package/dist/harness/context/constitution/default-agents.md +11 -10
- package/dist/harness/context/default-context-engine.d.ts +34 -4
- package/dist/harness/context/default-context-engine.js +194 -49
- package/dist/harness/context/hybrid-compressor.js +6 -0
- package/dist/harness/context/llm-summarizer.d.ts +2 -8
- package/dist/harness/context/llm-summarizer.js +69 -91
- package/dist/harness/context/message-selector.js +4 -3
- package/dist/harness/context/summarize.d.ts +2 -0
- package/dist/harness/context/summarize.js +7 -2
- package/dist/harness/context/truncate-compressor.js +7 -0
- package/dist/harness/context/turn-envelope.d.ts +3 -2
- package/dist/harness/context/turn-envelope.js +3 -2
- package/dist/harness/context/types.d.ts +33 -1
- package/dist/harness/extension/execution-environment/dangerous-commands.d.ts +28 -0
- package/dist/harness/extension/execution-environment/dangerous-commands.js +129 -0
- package/dist/harness/extension/execution-environment/effect-policy.d.ts +127 -8
- package/dist/harness/extension/execution-environment/effect-policy.js +244 -19
- package/dist/harness/extension/execution-environment/index.d.ts +4 -2
- package/dist/harness/extension/execution-environment/index.js +2 -1
- package/dist/harness/extension/execution-environment/kill-switch.d.ts +34 -0
- package/dist/harness/extension/execution-environment/kill-switch.js +46 -0
- package/dist/harness/extension/execution-environment/ops-audit.d.ts +200 -0
- package/dist/harness/extension/execution-environment/ops-audit.js +338 -0
- package/dist/harness/extension/execution-environment/privileged-exec.d.ts +158 -0
- package/dist/harness/extension/execution-environment/privileged-exec.js +418 -0
- package/dist/harness/extension/execution-environment/process-info.d.ts +180 -0
- package/dist/harness/extension/execution-environment/process-info.js +466 -0
- package/dist/harness/extension/execution-environment/process-isolation.d.ts +7 -0
- package/dist/harness/extension/execution-environment/process-isolation.js +9 -0
- package/dist/harness/extension/execution-environment/sandbox.d.ts +109 -0
- package/dist/harness/extension/execution-environment/sandbox.js +228 -21
- package/dist/harness/extension/execution-environment/tool-capabilities.d.ts +1 -1
- package/dist/harness/extension/execution-environment/tool-capabilities.js +6 -0
- package/dist/harness/extension/execution-environment/types.d.ts +19 -0
- package/dist/harness/extension/plugin-ecosystem/channel-types.d.ts +5 -0
- package/dist/harness/extension/plugin-ecosystem/tools/allowlisted-run.js +33 -0
- package/dist/harness/extension/plugin-ecosystem/tools/ask-user.d.ts +4 -5
- package/dist/harness/extension/plugin-ecosystem/tools/ask-user.js +5 -15
- package/dist/harness/extension/plugin-ecosystem/tools/builtin.d.ts +3 -0
- package/dist/harness/extension/plugin-ecosystem/tools/builtin.js +53 -2
- package/dist/harness/extension/plugin-ecosystem/tools/index.d.ts +2 -0
- package/dist/harness/extension/plugin-ecosystem/tools/index.js +2 -0
- package/dist/harness/extension/plugin-ecosystem/tools/privileged-tool.d.ts +43 -0
- package/dist/harness/extension/plugin-ecosystem/tools/privileged-tool.js +350 -0
- package/dist/harness/extension/plugin-ecosystem/tools/process-tools.d.ts +30 -0
- package/dist/harness/extension/plugin-ecosystem/tools/process-tools.js +137 -0
- package/dist/harness/governance/human-in-the-loop/decision-cache.d.ts +54 -4
- package/dist/harness/governance/human-in-the-loop/decision-cache.js +42 -16
- package/dist/harness/governance/security/temp-path.d.ts +9 -0
- package/dist/harness/governance/security/temp-path.js +13 -2
- package/dist/harness/governance/session-acl/seed-roles.js +14 -0
- package/dist/harness/index.d.ts +2 -0
- package/dist/harness/index.js +2 -0
- package/dist/harness/knowledge/fts.js +1 -2
- package/dist/harness/knowledge/index-store.js +59 -9
- package/dist/harness/memory/sqlite/keyword-search.d.ts +1 -0
- package/dist/harness/memory/sqlite/keyword-search.js +4 -3
- package/dist/harness/observability/observer/hub.js +8 -0
- package/dist/harness/observability/observer/run-snapshot.js +1 -0
- package/dist/harness/observability/observer/types.d.ts +4 -0
- package/dist/harness/run/agent/agent.d.ts +7 -0
- package/dist/harness/run/agent/agent.js +23 -1
- package/dist/harness/run/reliability/fallback-provider.d.ts +117 -5
- package/dist/harness/run/reliability/fallback-provider.js +262 -22
- package/dist/harness/run/reliability/harness-events.d.ts +24 -1
- package/dist/harness/run/reliability/model-error-policy.d.ts +32 -0
- package/dist/harness/run/reliability/model-error-policy.js +42 -0
- package/dist/harness/run/reliability/provider-wrapper.js +29 -2
- package/dist/harness/run/reliability/refusal-guard.d.ts +20 -0
- package/dist/harness/run/reliability/refusal-guard.js +102 -0
- package/dist/harness/run/reliability/run-agent.js +231 -33
- package/dist/harness/run/run-guard/default-run-guard.js +1 -1
- package/dist/harness/run/run-scope.d.ts +31 -0
- package/dist/harness/run/runner.d.ts +116 -2
- package/dist/harness/run/runner.js +547 -41
- package/dist/harness/session/attachments/llm-intent.js +2 -1
- package/dist/harness/session/tasks/render.d.ts +4 -1
- package/dist/harness/session/tasks/render.js +10 -7
- package/dist/harness/shared/events/agent-event-map.d.ts +66 -0
- package/dist/harness/shared/events/agent-event-map.js +5 -0
- package/dist/index.d.ts +3 -3
- package/dist/index.js +2 -2
- package/dist/integration/agent-runtime/channel-message-source.js +19 -1
- package/dist/integration/index.d.ts +2 -0
- package/dist/integration/index.js +1 -0
- package/dist/integration/observability/loop-watchdog.d.ts +56 -0
- package/dist/integration/observability/loop-watchdog.js +177 -0
- package/dist/integration/observability/metrics.d.ts +3 -0
- package/dist/integration/observability/metrics.js +18 -7
- package/dist/integration/providers/anthropic.d.ts +23 -0
- package/dist/integration/providers/anthropic.js +97 -12
- package/dist/integration/providers/fetch-timeout.d.ts +42 -0
- package/dist/integration/providers/fetch-timeout.js +51 -0
- package/dist/integration/providers/openai.d.ts +10 -0
- package/dist/integration/providers/openai.js +58 -11
- package/dist/integration/providers/provider-diagnostics.d.ts +80 -0
- package/dist/integration/providers/provider-diagnostics.js +89 -0
- package/dist/integration/providers/timeout-coherence.d.ts +52 -0
- package/dist/integration/providers/timeout-coherence.js +68 -0
- package/dist/integration/providers/wire-dump.d.ts +62 -0
- package/dist/integration/providers/wire-dump.js +111 -2
- package/dist/integration/storage/jsonl.js +4 -1
- package/dist/subsystems/cognition-steward/conceptualize/handler.js +1 -1
- package/dist/subsystems/memory-steward/backfill/handler.js +1 -1
- package/dist/subsystems/wisdom-steward/formation/handler.js +1 -1
- package/package.json +3 -2
|
@@ -7,9 +7,11 @@
|
|
|
7
7
|
* 摘要策略:
|
|
8
8
|
* 1. 将消息格式化为上下文
|
|
9
9
|
* 2. 调用 LLM 生成结构化摘要
|
|
10
|
-
* 3.
|
|
10
|
+
* 3. 摘要预算(P1c/F8):clamp(2%, 2k, 8k) 软目标写指令 + 4× 保险丝硬顶——与 compact 引擎同语义
|
|
11
11
|
*/
|
|
12
12
|
import { getTextContent } from '@octopi-agent/core/types.js';
|
|
13
|
+
import { buildHandoverPrompt, buildHandoverUserContent, clampSummaryProviderCap, detectSummaryContamination, summarySoftTargetTokens, } from './capabilities/compact/index.js';
|
|
14
|
+
import { estimateLLMMessages } from './token-estimator.js';
|
|
13
15
|
import { HeuristicTokenEstimator } from './token-estimator.js';
|
|
14
16
|
import { toolResultBody } from './tool-result-body.js';
|
|
15
17
|
/**
|
|
@@ -42,61 +44,7 @@ export function buildToolCallLedger(messages) {
|
|
|
42
44
|
}
|
|
43
45
|
return lines.join('\n');
|
|
44
46
|
}
|
|
45
|
-
/**
|
|
46
|
-
* 结构化摘要模板(F-d 硬约束:Done 逐条可核验 + 交接语防重做)
|
|
47
|
-
*/
|
|
48
|
-
const SUMMARY_TEMPLATE = `## Conversation Summary
|
|
49
|
-
|
|
50
|
-
> HANDOFF: Your working memory was compacted. First check Done/ToDo and the session task list (task_*). DO NOT redo work listed under Done.
|
|
51
|
-
|
|
52
|
-
### Goal
|
|
53
|
-
[What the user is trying to accomplish]
|
|
54
|
-
|
|
55
|
-
### Progress
|
|
56
|
-
#### Done
|
|
57
|
-
- [each item MUST be] what was done + artifact path(s) / exact identifiers
|
|
58
|
-
- ...
|
|
59
|
-
|
|
60
|
-
#### In Progress
|
|
61
|
-
[What is currently being worked on]
|
|
62
|
-
|
|
63
|
-
#### Blocked
|
|
64
|
-
[Any issues or blockers]
|
|
65
|
-
|
|
66
|
-
### Key Decisions
|
|
67
|
-
[Important technical decisions and reasons]
|
|
68
|
-
|
|
69
|
-
### Relevant Files
|
|
70
|
-
[Files read, modified, or created — absolute or workspace-relative paths]
|
|
71
|
-
|
|
72
|
-
### Next Steps
|
|
73
|
-
[What needs to be done next; align with task_* list when present]
|
|
74
|
-
|
|
75
|
-
### Key Context
|
|
76
|
-
[Specific values, error messages, configuration details]`;
|
|
77
|
-
/**
|
|
78
|
-
* 摘要指令(缓存友好:只作 **user 尾部** 追加,绝不改写 system/历史)。
|
|
79
|
-
* 规则与旧 SUMMARIZE_INSTRUCTION 对齐;模板不变。
|
|
80
|
-
*/
|
|
81
|
-
function buildSummarizeTailInstruction(params) {
|
|
82
|
-
const header = params.previousSummary
|
|
83
|
-
? `You are updating an existing conversation summary. Merge the conversation above into the summary below.\n\nPrevious Summary:\n${params.previousSummary}`
|
|
84
|
-
: `You are a conversation summarizer. Create a concise, structured summary of the conversation above.`;
|
|
85
|
-
return `${header}
|
|
86
|
-
|
|
87
|
-
Rules:
|
|
88
|
-
1. Focus on factual information, not pleasantries
|
|
89
|
-
2. Preserve specific details (file paths, error messages, code snippets)
|
|
90
|
-
3. Use the exact structure provided in the template
|
|
91
|
-
4. If updating a previous summary, merge new information rather than replacing everything
|
|
92
|
-
5. Keep the summary under ${params.maxTokens} tokens
|
|
93
|
-
6. HARD: Every Done item must state what was completed AND the artifact path or exact identifier
|
|
94
|
-
7. HARD: Keep the HANDOFF line at the top unchanged in meaning — it prevents redoing finished work
|
|
95
|
-
8. HARD: Do not invent paths or claim work that is not evidenced in the history
|
|
96
|
-
|
|
97
|
-
Template:
|
|
98
|
-
${SUMMARY_TEMPLATE}`;
|
|
99
|
-
}
|
|
47
|
+
/** 模板/指令/合并语义已单源化到 capabilities/compact/handover.ts(F8 模板统一:四段式交接书)。 */
|
|
100
48
|
export class LLMSummaryCompressor {
|
|
101
49
|
/**
|
|
102
50
|
* 压缩消息
|
|
@@ -111,13 +59,15 @@ export class LLMSummaryCompressor {
|
|
|
111
59
|
const estimator = tokenEstimator ?? new HeuristicTokenEstimator();
|
|
112
60
|
// 如果没有摘要函数,返回截断结果
|
|
113
61
|
if (!summarize) {
|
|
114
|
-
return this.fallbackTruncate(messages, targetTokens, estimator);
|
|
62
|
+
return this.fallbackTruncate(messages, targetTokens, estimator, false);
|
|
115
63
|
}
|
|
116
64
|
// 如果没有消息需要压缩
|
|
117
65
|
if (messages.length === 0) {
|
|
118
66
|
return {
|
|
119
67
|
result: [],
|
|
120
68
|
estimatedTokens: 0,
|
|
69
|
+
fresh: false,
|
|
70
|
+
summarizeAttempted: false,
|
|
121
71
|
};
|
|
122
72
|
}
|
|
123
73
|
// 计算当前 token 数
|
|
@@ -128,47 +78,79 @@ export class LLMSummaryCompressor {
|
|
|
128
78
|
return {
|
|
129
79
|
result: [...messages],
|
|
130
80
|
estimatedTokens: currentTokens,
|
|
81
|
+
fresh: false,
|
|
82
|
+
summarizeAttempted: false,
|
|
131
83
|
};
|
|
132
84
|
}
|
|
133
|
-
//
|
|
134
|
-
const
|
|
85
|
+
// F8 单一预算语义:P1c clamp(2%, 2k, 8k) 软目标写指令;provider 硬顶 = 4× 保险丝 + 窗口钳制
|
|
86
|
+
const softTarget = summarySoftTargetTokens(currentTokens, params.summaryBudget);
|
|
87
|
+
const prompt = buildHandoverPrompt({
|
|
88
|
+
maxTokens: softTarget,
|
|
89
|
+
previousSummary,
|
|
90
|
+
foldRange: params.foldRange,
|
|
91
|
+
taskSnapshot: params.taskSnapshot,
|
|
92
|
+
});
|
|
135
93
|
// 格式化消息为 LLM 可读的上下文(仅无 llmPrefix 时的旁路)
|
|
136
94
|
const formattedContext = this.formatMessagesForSummary(messages);
|
|
137
95
|
// 构建 LLM 消息
|
|
138
96
|
// 缓存优先:llmPrefix(会话 system+history 同形)+ 尾部 user 指令。
|
|
139
97
|
// 禁止改写 system / 重排历史——否则打穿 provider 前缀缓存。
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
|
|
144
|
-
|
|
145
|
-
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
|
|
151
|
-
|
|
98
|
+
const buildPrefixed = () => [
|
|
99
|
+
...params.llmPrefix,
|
|
100
|
+
{ role: 'user', content: prompt.tailUser },
|
|
101
|
+
];
|
|
102
|
+
const buildTextify = () => [
|
|
103
|
+
{ role: 'system', content: prompt.instruction },
|
|
104
|
+
{ role: 'user', content: buildHandoverUserContent(params, formattedContext) },
|
|
105
|
+
];
|
|
106
|
+
let usePrefix = Boolean(params.llmPrefix?.length);
|
|
107
|
+
let llmMessages = usePrefix ? buildPrefixed() : buildTextify();
|
|
108
|
+
// 回声比对源 = 模型实际可见文本投影(评审 F8):前缀模式只比前缀内容——formattedContext 是
|
|
109
|
+
// 文本化重渲染,模型并未看见,算进比对源只增误伤面;与 compact 引擎同判据
|
|
110
|
+
let echoSource = usePrefix
|
|
111
|
+
? (params.llmPrefix ?? []).map((m) => String(m.content ?? '')).join('\n')
|
|
112
|
+
: formattedContext;
|
|
113
|
+
const hardCap = softTarget * 4;
|
|
114
|
+
const capOf = (msgs) => clampSummaryProviderCap({
|
|
115
|
+
hardCap,
|
|
116
|
+
// estimateLLMMessages 口径含 tool_calls/content block/消息开销——摘要恰在工具密集
|
|
117
|
+
// 会话触发,String(content) 口径会把 tool_calls 全漏掉、多模态块变 [object Object]
|
|
118
|
+
inputTokens: estimateLLMMessages(msgs),
|
|
119
|
+
contextWindow: params.contextWindow,
|
|
120
|
+
maxOutputTokens: params.maxOutputTokens,
|
|
121
|
+
});
|
|
122
|
+
let providerMaxTokens = capOf(llmMessages);
|
|
123
|
+
// F22 可观测:无前缀/被迫旁路 = cache-miss,如实上报,不静默换请求形状
|
|
124
|
+
if (!usePrefix) {
|
|
125
|
+
params.emit?.({ type: 'compact.summarize_fallback', reason: 'no_prefix', cached: false });
|
|
126
|
+
}
|
|
127
|
+
if (providerMaxTokens < 256 && usePrefix) {
|
|
128
|
+
usePrefix = false;
|
|
129
|
+
llmMessages = buildTextify();
|
|
130
|
+
echoSource = formattedContext;
|
|
131
|
+
providerMaxTokens = capOf(llmMessages);
|
|
132
|
+
params.emit?.({ type: 'compact.summarize_fallback', reason: 'window_guard', cached: false });
|
|
152
133
|
}
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
content: buildSummarizeTailInstruction({ previousSummary, maxTokens: summaryTokens }),
|
|
158
|
-
},
|
|
159
|
-
{ role: 'user', content: formattedContext },
|
|
160
|
-
];
|
|
134
|
+
if (providerMaxTokens < 256) {
|
|
135
|
+
// F12 可观测:窗口穷尽是最终失败,事件如实上报(不止于换旁路时)
|
|
136
|
+
params.emit?.({ type: 'compact.summarize_fallback', reason: 'window_exhausted', cached: false });
|
|
137
|
+
return this.fallbackTruncate(messages, targetTokens, estimator, false);
|
|
161
138
|
}
|
|
162
139
|
try {
|
|
163
140
|
// 调用 LLM 生成摘要
|
|
164
141
|
const summary = await summarize(llmMessages, {
|
|
165
142
|
previousSummary,
|
|
166
|
-
maxTokens:
|
|
143
|
+
maxTokens: providerMaxTokens,
|
|
167
144
|
});
|
|
168
145
|
// 空摘要(拒答/内容过滤)按失败处理:不得用空正文替换中间段、更不得冒充语义摘要
|
|
169
146
|
if (!summary || !summary.trim()) {
|
|
170
147
|
return this.fallbackTruncate(messages, targetTokens, estimator);
|
|
171
148
|
}
|
|
149
|
+
// 输出闸门(评审 F8):proactive/overflow 与手动路径同一污染判据——续写污染
|
|
150
|
+
//(对话体/工具调用形状、成段回声)按失败处理,走截断降级(诚实降级,不落坏摘要)。
|
|
151
|
+
if (detectSummaryContamination(summary, echoSource)) {
|
|
152
|
+
return this.fallbackTruncate(messages, targetTokens, estimator);
|
|
153
|
+
}
|
|
172
154
|
// 摘要用 user 角色:中段 system 会被部分网关拒绝
|
|
173
155
|
const summaryMessage = {
|
|
174
156
|
role: 'user',
|
|
@@ -182,6 +164,8 @@ export class LLMSummaryCompressor {
|
|
|
182
164
|
result: [summaryMessage],
|
|
183
165
|
estimatedTokens: summaryMessageTokens,
|
|
184
166
|
droppedSummary: `Conversation compressed from ${currentTokens} to ~${summaryMessageTokens} tokens`,
|
|
167
|
+
fresh: true,
|
|
168
|
+
summarizeAttempted: true,
|
|
185
169
|
};
|
|
186
170
|
}
|
|
187
171
|
catch (err) {
|
|
@@ -190,16 +174,6 @@ export class LLMSummaryCompressor {
|
|
|
190
174
|
return this.fallbackTruncate(messages, targetTokens, estimator);
|
|
191
175
|
}
|
|
192
176
|
}
|
|
193
|
-
/**
|
|
194
|
-
* 计算摘要 token 预算
|
|
195
|
-
*
|
|
196
|
-
* 公式:contentTokens × 0.20
|
|
197
|
-
* 上下限:2000 - 12000
|
|
198
|
-
*/
|
|
199
|
-
calculateSummaryBudget(contentTokens) {
|
|
200
|
-
const budget = Math.floor(contentTokens * 0.20);
|
|
201
|
-
return Math.max(2000, Math.min(12000, budget));
|
|
202
|
-
}
|
|
203
177
|
/**
|
|
204
178
|
* 格式化消息为 LLM 可读的上下文
|
|
205
179
|
*
|
|
@@ -237,7 +211,7 @@ export class LLMSummaryCompressor {
|
|
|
237
211
|
* 作为 **middle 替换体**:账本说明 + 近尾若干条。
|
|
238
212
|
* 外层 head/tail 由调用方(applyProactiveCompact)拼接,此处不得重复携带。
|
|
239
213
|
*/
|
|
240
|
-
fallbackTruncate(messages, _targetTokens, estimator) {
|
|
214
|
+
fallbackTruncate(messages, _targetTokens, estimator, summarizeAttempted = true) {
|
|
241
215
|
// 近尾保留工具步上下文;更早的换成「账本」——禁止只留 4 条且无索引
|
|
242
216
|
const keepCount = Math.min(Math.max(6, Math.ceil(messages.length * 0.15)), messages.length);
|
|
243
217
|
const kept = messages.slice(-keepCount);
|
|
@@ -247,6 +221,8 @@ export class LLMSummaryCompressor {
|
|
|
247
221
|
return {
|
|
248
222
|
result: [...messages],
|
|
249
223
|
estimatedTokens: estimator.estimateMessages(messages),
|
|
224
|
+
fresh: false,
|
|
225
|
+
summarizeAttempted,
|
|
250
226
|
};
|
|
251
227
|
}
|
|
252
228
|
const ledger = buildToolCallLedger(dropped);
|
|
@@ -266,6 +242,8 @@ export class LLMSummaryCompressor {
|
|
|
266
242
|
result,
|
|
267
243
|
estimatedTokens,
|
|
268
244
|
droppedSummary: `${dropped.length} messages truncated`,
|
|
245
|
+
fresh: false,
|
|
246
|
+
summarizeAttempted,
|
|
269
247
|
};
|
|
270
248
|
}
|
|
271
249
|
}
|
|
@@ -34,14 +34,15 @@ export class DefaultMessageSelector {
|
|
|
34
34
|
};
|
|
35
35
|
}
|
|
36
36
|
// 1. 划分头部保护区域
|
|
37
|
-
const headCount = Math.min(protectFirstN, messages.length);
|
|
37
|
+
const headCount = Math.max(0, Math.min(protectFirstN, messages.length));
|
|
38
38
|
const headMessages = messages.slice(0, headCount);
|
|
39
39
|
// 2. 划分尾部保护区域(排除头部已保护的)
|
|
40
40
|
const remainingAfterHead = messages.slice(headCount);
|
|
41
41
|
const tailCount = Math.min(protectLastN, remainingAfterHead.length);
|
|
42
|
-
|
|
42
|
+
// slice(-0)===slice(0) 返回全量、slice(0,-0) 返回空:tailCount=0 时必须显式取空/全量
|
|
43
|
+
const tailMessages = tailCount > 0 ? remainingAfterHead.slice(-tailCount) : [];
|
|
43
44
|
// 3. 中间区域(可压缩/丢弃)
|
|
44
|
-
const middleMessages = remainingAfterHead.slice(0, -tailCount);
|
|
45
|
+
const middleMessages = tailCount > 0 ? remainingAfterHead.slice(0, -tailCount) : remainingAfterHead;
|
|
45
46
|
// 4. 计算各区域 token 数
|
|
46
47
|
const headTokens = estimator.estimateMessages(headMessages);
|
|
47
48
|
const tailTokens = estimator.estimateMessages(tailMessages);
|
|
@@ -11,6 +11,8 @@ export interface CreateProviderSummarizeOptions {
|
|
|
11
11
|
model?: string;
|
|
12
12
|
/** 摘要温度(默认 0.3,偏确定性) */
|
|
13
13
|
temperature?: number;
|
|
14
|
+
/** 模型实际输出上限(评审 F6):摘要硬顶的源头钳制,防大窗口下 0.1×W 顶穿 max_output_tokens */
|
|
15
|
+
maxOutputTokens?: number;
|
|
14
16
|
}
|
|
15
17
|
/**
|
|
16
18
|
* 创建基于 ModelProvider 的摘要函数
|
|
@@ -16,8 +16,13 @@ export function createProviderSummarize(provider, options) {
|
|
|
16
16
|
const response = await provider.chat({
|
|
17
17
|
messages,
|
|
18
18
|
model: options?.model,
|
|
19
|
-
|
|
20
|
-
|
|
19
|
+
// 未显式配置即不发送:由模型配置(ModelInfo.temperature)或服务端默认决定
|
|
20
|
+
temperature: options?.temperature,
|
|
21
|
+
// 源头钳制(评审 F6):三条摘要路径(manual/proactive/overflow)共用本闭包,
|
|
22
|
+
// 硬顶不得越过模型实际输出上限,否则窗口钳制在大窗口下仍会顶穿被 provider 400
|
|
23
|
+
maxTokens: options?.maxOutputTokens && options.maxOutputTokens > 0
|
|
24
|
+
? Math.min(opts?.maxTokens ?? options.maxOutputTokens, options.maxOutputTokens)
|
|
25
|
+
: opts?.maxTokens,
|
|
21
26
|
});
|
|
22
27
|
return response.content;
|
|
23
28
|
};
|
|
@@ -24,6 +24,8 @@ export class TruncateCompressor {
|
|
|
24
24
|
return {
|
|
25
25
|
result: [],
|
|
26
26
|
estimatedTokens: 0,
|
|
27
|
+
fresh: false,
|
|
28
|
+
summarizeAttempted: false,
|
|
27
29
|
};
|
|
28
30
|
}
|
|
29
31
|
// 计算当前 token 数
|
|
@@ -33,6 +35,8 @@ export class TruncateCompressor {
|
|
|
33
35
|
return {
|
|
34
36
|
result: [...messages],
|
|
35
37
|
estimatedTokens: currentTokens,
|
|
38
|
+
fresh: false,
|
|
39
|
+
summarizeAttempted: false,
|
|
36
40
|
};
|
|
37
41
|
}
|
|
38
42
|
// 需要压缩的 token 数
|
|
@@ -83,6 +87,9 @@ export class TruncateCompressor {
|
|
|
83
87
|
result,
|
|
84
88
|
estimatedTokens,
|
|
85
89
|
droppedSummary: `${removedCount} messages removed (~${removedTokens} tokens)`,
|
|
90
|
+
// 纯截断不是语义摘要(F8R-01 显式契约);调用方冷却语义:截断未发起 LLM
|
|
91
|
+
fresh: false,
|
|
92
|
+
summarizeAttempted: false,
|
|
86
93
|
};
|
|
87
94
|
}
|
|
88
95
|
/**
|
|
@@ -5,8 +5,9 @@
|
|
|
5
5
|
* - 开 turn:user 消息正文 = `<turn_start id at>…<user_message>…</user_message></turn_start>`
|
|
6
6
|
* - 闭 turn:独立消息 `<turn_end id/>`,`metadata.source='turnEnd'`
|
|
7
7
|
*
|
|
8
|
-
*
|
|
9
|
-
*
|
|
8
|
+
* 不变式:每个信封(turn_start 与其 turn_end 之间)恰一枚闭标;id 会话内单调,
|
|
9
|
+
* 但崩溃续跑的历史数据可能同 id 多信封(计数器丢失遗留),以信封区间判形为准。
|
|
10
|
+
* 定形后字节冻结。正文经 XML 转义嵌入,用户原文不得截断信封。
|
|
10
11
|
*/
|
|
11
12
|
import type { Message } from '@octopi-agent/core/types.js';
|
|
12
13
|
/** turn_start 元素的 id 形态(会话内单调 t1,t2,…) */
|
|
@@ -5,8 +5,9 @@
|
|
|
5
5
|
* - 开 turn:user 消息正文 = `<turn_start id at>…<user_message>…</user_message></turn_start>`
|
|
6
6
|
* - 闭 turn:独立消息 `<turn_end id/>`,`metadata.source='turnEnd'`
|
|
7
7
|
*
|
|
8
|
-
*
|
|
9
|
-
*
|
|
8
|
+
* 不变式:每个信封(turn_start 与其 turn_end 之间)恰一枚闭标;id 会话内单调,
|
|
9
|
+
* 但崩溃续跑的历史数据可能同 id 多信封(计数器丢失遗留),以信封区间判形为准。
|
|
10
|
+
* 定形后字节冻结。正文经 XML 转义嵌入,用户原文不得截断信封。
|
|
10
11
|
*/
|
|
11
12
|
import { getTextContent } from '@octopi-agent/core/types.js';
|
|
12
13
|
/** turn_start 元素的 id 形态(会话内单调 t1,t2,…) */
|
|
@@ -48,7 +48,7 @@ export type ContextCompactReason = 'proactive' | 'overflow';
|
|
|
48
48
|
* 避免 LLM 摘要耗时被误判为会话卡死。失败不中断本轮(引擎回退截断)。
|
|
49
49
|
*/
|
|
50
50
|
export interface ContextCompactEvent {
|
|
51
|
-
type: 'context.compact.start' | 'context.compact.end' | 'context.compact.error';
|
|
51
|
+
type: 'context.compact.start' | 'context.compact.end' | 'context.compact.error' | 'context.compact.summarize_fallback';
|
|
52
52
|
sessionId: string;
|
|
53
53
|
reason?: ContextCompactReason;
|
|
54
54
|
/** 压缩前估算 token */
|
|
@@ -61,6 +61,8 @@ export interface ContextCompactEvent {
|
|
|
61
61
|
durationMs?: number;
|
|
62
62
|
/** 走缓存摘要重建,未调 LLM */
|
|
63
63
|
cached?: boolean;
|
|
64
|
+
/** summarize_fallback:cache-miss 旁路原因(no_prefix | window_guard | window_exhausted,F22) */
|
|
65
|
+
fallback?: string;
|
|
64
66
|
error?: string;
|
|
65
67
|
}
|
|
66
68
|
/** 事件回调(Builder 注入 → EventBus) */
|
|
@@ -83,6 +85,8 @@ export interface AssembleParams {
|
|
|
83
85
|
tokenBudget?: number;
|
|
84
86
|
/** 模型上下文窗口(仅显式配置时传入) */
|
|
85
87
|
contextWindow?: number;
|
|
88
|
+
/** 模型 max_output_tokens(评审 F8R-04):摘要硬顶按输出上限钳制,防自动路径顶穿被 400 */
|
|
89
|
+
maxOutputTokens?: number;
|
|
86
90
|
/**
|
|
87
91
|
* 显式压缩目标 token(context.compactTargetTokens)。
|
|
88
92
|
* 与 contextWindow 无关;未知窗口时仍可作手动/结构压缩的目标。
|
|
@@ -274,12 +278,40 @@ export interface CompressParams {
|
|
|
274
278
|
* 摘要请求只许在其**尾部**追加指令;禁止改 system / 历史字节——否则打穿 provider 前缀缓存。
|
|
275
279
|
*/
|
|
276
280
|
llmPrefix?: LLMMessage[];
|
|
281
|
+
/** 摘要软目标预算(P1c clamp(2%, 2k, 8k) 的可配置 ratio/min/max) */
|
|
282
|
+
summaryBudget?: {
|
|
283
|
+
ratioPct?: number;
|
|
284
|
+
minTokens?: number;
|
|
285
|
+
maxTokens?: number;
|
|
286
|
+
};
|
|
287
|
+
/** 折叠区间(评审 F7):交接书替换范围与原样保留尾部——计数语义 */
|
|
288
|
+
foldRange?: {
|
|
289
|
+
replacedCount: number;
|
|
290
|
+
total: number;
|
|
291
|
+
preservedTail: number;
|
|
292
|
+
};
|
|
293
|
+
/** task 快照(B0:易变块走 content 尾部,不进 system) */
|
|
294
|
+
taskSnapshot?: string;
|
|
295
|
+
/** 会话模型窗口/输出上限(评审 F5):摘要硬顶按窗口钳制 */
|
|
296
|
+
contextWindow?: number;
|
|
297
|
+
maxOutputTokens?: number;
|
|
298
|
+
/** 事件回调(F22:cache-miss 旁路可观测) */
|
|
299
|
+
emit?: (e: {
|
|
300
|
+
type: 'compact.summarize_fallback';
|
|
301
|
+
reason?: string;
|
|
302
|
+
cached?: boolean;
|
|
303
|
+
}) => void;
|
|
277
304
|
}
|
|
278
305
|
export interface CompressResult {
|
|
279
306
|
/** 压缩后的消息/摘要 */
|
|
280
307
|
result: Message[];
|
|
281
308
|
/** 估算 token 数 */
|
|
282
309
|
estimatedTokens: number;
|
|
310
|
+
/** 本次产出**新的**语义摘要(评审 F8-F1):仅 true 时调用方才可推进覆盖点/previousSummary。
|
|
311
|
+
* 机械降级(截断/兜底/占位说明)必须显式 false——占位说明不得冒充语义摘要(fail-closed,F8R-01)。 */
|
|
312
|
+
fresh: boolean;
|
|
313
|
+
/** LLM 摘要调用是否真正发起过(成功/失败均算):只在 true 时推进冷却时间戳(F8R-01)。 */
|
|
314
|
+
summarizeAttempted: boolean;
|
|
283
315
|
/** 被丢弃的信息摘要 */
|
|
284
316
|
droppedSummary?: string;
|
|
285
317
|
}
|
|
@@ -0,0 +1,28 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 危险命令分类(§6:减速带,不是安全边界)— ops P4 pre-exec 闸 / P5 确认通道共用
|
|
3
|
+
*
|
|
4
|
+
* 诚实定位(必须与文档一致):**只拦事故,不拦攻击**。shell/argv 分类必然可绕过
|
|
5
|
+
* (`sh -c`、`python -c`、base64、变量拼接、别名、重定向)。任何「靠模式识别危险
|
|
6
|
+
* 命令」的机制都拦不住刻意构造的注入——真正的边界是权限范围(escalation.policy /
|
|
7
|
+
* L2 白名单 / OS uid)+ 审计。
|
|
8
|
+
*
|
|
9
|
+
* 分类表保守(§6 命中表):命中常见破坏性模式 → 要求确认;未命中 → 放行。
|
|
10
|
+
* **不做「未知即危险」**(否则误报淹没真信号)。
|
|
11
|
+
*
|
|
12
|
+
* 已知绕过(tests 钉死为用例,防后人误当边界):
|
|
13
|
+
* - `sh -c 'rm -rf /'` —— 嵌套 shell 的命令串本分类器不解析;
|
|
14
|
+
* - `> /etc/*` 重定向 —— argv 形态根本表达不了重定向。
|
|
15
|
+
*/
|
|
16
|
+
/** 分类结论 */
|
|
17
|
+
export interface DangerousCommandVerdict {
|
|
18
|
+
dangerous: boolean;
|
|
19
|
+
/** 命中的 §6 分类表条目(审计/诊断用);未命中为 undefined */
|
|
20
|
+
pattern?: string;
|
|
21
|
+
}
|
|
22
|
+
/**
|
|
23
|
+
* 按 §6 命中表对 argv 分类。
|
|
24
|
+
*
|
|
25
|
+
* @param argv - 完整 argv(argv[0] = 可执行文件;提权前缀 sudo/doas 会被剥掉后分类)
|
|
26
|
+
* @returns 是否危险 + 命中条目名
|
|
27
|
+
*/
|
|
28
|
+
export declare function classifyDangerousArgv(argv: readonly string[]): DangerousCommandVerdict;
|
|
@@ -0,0 +1,129 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 危险命令分类(§6:减速带,不是安全边界)— ops P4 pre-exec 闸 / P5 确认通道共用
|
|
3
|
+
*
|
|
4
|
+
* 诚实定位(必须与文档一致):**只拦事故,不拦攻击**。shell/argv 分类必然可绕过
|
|
5
|
+
* (`sh -c`、`python -c`、base64、变量拼接、别名、重定向)。任何「靠模式识别危险
|
|
6
|
+
* 命令」的机制都拦不住刻意构造的注入——真正的边界是权限范围(escalation.policy /
|
|
7
|
+
* L2 白名单 / OS uid)+ 审计。
|
|
8
|
+
*
|
|
9
|
+
* 分类表保守(§6 命中表):命中常见破坏性模式 → 要求确认;未命中 → 放行。
|
|
10
|
+
* **不做「未知即危险」**(否则误报淹没真信号)。
|
|
11
|
+
*
|
|
12
|
+
* 已知绕过(tests 钉死为用例,防后人误当边界):
|
|
13
|
+
* - `sh -c 'rm -rf /'` —— 嵌套 shell 的命令串本分类器不解析;
|
|
14
|
+
* - `> /etc/*` 重定向 —— argv 形态根本表达不了重定向。
|
|
15
|
+
*/
|
|
16
|
+
/**
|
|
17
|
+
* 剥掉提权前缀(`sudo` / `doas` 及其选项)——分类对象是真实目标命令。
|
|
18
|
+
* 保守剥离:只处理选项 token 与 `-u <user>` 形态;剥不干净时宁可少报(§6 未命中放行)。
|
|
19
|
+
*
|
|
20
|
+
* @param argv - 完整 argv
|
|
21
|
+
* @returns 剥离后的 argv(至少含一个 token)
|
|
22
|
+
*/
|
|
23
|
+
function stripPrivilegePrefix(argv) {
|
|
24
|
+
if (argv.length === 0)
|
|
25
|
+
return argv;
|
|
26
|
+
const base = basenameOf(argv[0]);
|
|
27
|
+
if (base !== 'sudo' && base !== 'doas')
|
|
28
|
+
return argv;
|
|
29
|
+
const rest = [];
|
|
30
|
+
for (let i = 1; i < argv.length; i += 1) {
|
|
31
|
+
const token = argv[i];
|
|
32
|
+
if (token === '--') {
|
|
33
|
+
rest.push(...argv.slice(i + 1));
|
|
34
|
+
break;
|
|
35
|
+
}
|
|
36
|
+
if (token.startsWith('-')) {
|
|
37
|
+
// `-u user` / `--user=user` 类选项会吞掉下一个 token
|
|
38
|
+
if (token === '-u' || token === '--user' || token === '-g' || token === '--group') {
|
|
39
|
+
i += 1;
|
|
40
|
+
}
|
|
41
|
+
continue;
|
|
42
|
+
}
|
|
43
|
+
rest.push(...argv.slice(i));
|
|
44
|
+
break;
|
|
45
|
+
}
|
|
46
|
+
return rest.length > 0 ? rest : argv.slice(1);
|
|
47
|
+
}
|
|
48
|
+
/** 可执行名(跨平台取末段;不解析 Windows 扩展名以外的语义) */
|
|
49
|
+
function basenameOf(path) {
|
|
50
|
+
const normalized = path.replace(/\\/g, '/');
|
|
51
|
+
const idx = normalized.lastIndexOf('/');
|
|
52
|
+
return (idx >= 0 ? normalized.slice(idx + 1) : normalized).toLowerCase();
|
|
53
|
+
}
|
|
54
|
+
/** 短选项聚合字母(`-rf` → ['r','f']);长选项/非选项不参与 */
|
|
55
|
+
function shortFlagLetters(tokens) {
|
|
56
|
+
const letters = new Set();
|
|
57
|
+
for (const token of tokens) {
|
|
58
|
+
if (token === '--')
|
|
59
|
+
break;
|
|
60
|
+
if (!token.startsWith('-') || token.startsWith('--') || token === '-')
|
|
61
|
+
continue;
|
|
62
|
+
for (const ch of token.slice(1))
|
|
63
|
+
letters.add(ch.toLowerCase());
|
|
64
|
+
}
|
|
65
|
+
return letters;
|
|
66
|
+
}
|
|
67
|
+
/** 是否出现递归/强制类长选项 */
|
|
68
|
+
function hasLongFlag(tokens, names) {
|
|
69
|
+
return tokens.some((t) => names.includes(t));
|
|
70
|
+
}
|
|
71
|
+
/**
|
|
72
|
+
* 按 §6 命中表对 argv 分类。
|
|
73
|
+
*
|
|
74
|
+
* @param argv - 完整 argv(argv[0] = 可执行文件;提权前缀 sudo/doas 会被剥掉后分类)
|
|
75
|
+
* @returns 是否危险 + 命中条目名
|
|
76
|
+
*/
|
|
77
|
+
export function classifyDangerousArgv(argv) {
|
|
78
|
+
const effective = stripPrivilegePrefix(argv);
|
|
79
|
+
if (effective.length === 0)
|
|
80
|
+
return { dangerous: false };
|
|
81
|
+
const cmd = basenameOf(effective[0]);
|
|
82
|
+
const rest = effective.slice(1);
|
|
83
|
+
// `rm -rf`:递归 + 强制同现(-rf / -fr / -r -f / --recursive --force)
|
|
84
|
+
if (cmd === 'rm' || cmd === 'rm.exe') {
|
|
85
|
+
const flags = shortFlagLetters(rest);
|
|
86
|
+
const recursive = flags.has('r') || hasLongFlag(rest, ['--recursive']);
|
|
87
|
+
const force = flags.has('f') || hasLongFlag(rest, ['--force']);
|
|
88
|
+
if (recursive && force)
|
|
89
|
+
return { dangerous: true, pattern: 'rm -rf' };
|
|
90
|
+
return { dangerous: false };
|
|
91
|
+
}
|
|
92
|
+
// `dd of=/dev/*`:直接写块设备
|
|
93
|
+
if (cmd === 'dd' || cmd === 'dd.exe') {
|
|
94
|
+
if (rest.some((a) => a.startsWith('of=/dev/') || a.startsWith('of=\\\\.\\'))) {
|
|
95
|
+
return { dangerous: true, pattern: 'dd of=/dev/*' };
|
|
96
|
+
}
|
|
97
|
+
return { dangerous: false };
|
|
98
|
+
}
|
|
99
|
+
// `mkfs.*`:格式化
|
|
100
|
+
if (cmd.startsWith('mkfs'))
|
|
101
|
+
return { dangerous: true, pattern: 'mkfs.*' };
|
|
102
|
+
// `iptables -F`:清空防火墙规则
|
|
103
|
+
if (cmd === 'iptables' || cmd === 'iptables.exe' || cmd === 'nft') {
|
|
104
|
+
if (rest.includes('-F') || rest.includes('--flush')) {
|
|
105
|
+
return { dangerous: true, pattern: 'iptables -F' };
|
|
106
|
+
}
|
|
107
|
+
return { dangerous: false };
|
|
108
|
+
}
|
|
109
|
+
// `systemctl disable`:关服务自启
|
|
110
|
+
if (cmd === 'systemctl' || cmd === 'systemctl.exe') {
|
|
111
|
+
if (rest.includes('disable') || rest.includes('mask')) {
|
|
112
|
+
return { dangerous: true, pattern: 'systemctl disable' };
|
|
113
|
+
}
|
|
114
|
+
return { dangerous: false };
|
|
115
|
+
}
|
|
116
|
+
// `chmod -R` / `chown -R`:递归改权限/属主
|
|
117
|
+
if (cmd === 'chmod' || cmd === 'chown' || cmd === 'chmod.exe' || cmd === 'chown.exe') {
|
|
118
|
+
const flags = shortFlagLetters(rest);
|
|
119
|
+
if (flags.has('r') || hasLongFlag(rest, ['--recursive'])) {
|
|
120
|
+
return { dangerous: true, pattern: `${cmd} -R` };
|
|
121
|
+
}
|
|
122
|
+
return { dangerous: false };
|
|
123
|
+
}
|
|
124
|
+
// `truncate`:截断文件(静默毁数据的典型手滑)
|
|
125
|
+
if (cmd === 'truncate' || cmd === 'truncate.exe') {
|
|
126
|
+
return { dangerous: true, pattern: 'truncate' };
|
|
127
|
+
}
|
|
128
|
+
return { dangerous: false };
|
|
129
|
+
}
|