@dexto/core 1.5.3 → 1.5.4
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/agent/DextoAgent.cjs +284 -1
- package/dist/agent/DextoAgent.d.ts +114 -0
- package/dist/agent/DextoAgent.d.ts.map +1 -1
- package/dist/agent/DextoAgent.js +275 -1
- package/dist/agent/schemas.d.ts +51 -21
- package/dist/agent/schemas.d.ts.map +1 -1
- package/dist/context/compaction/overflow.cjs +6 -10
- package/dist/context/compaction/overflow.d.ts +14 -11
- package/dist/context/compaction/overflow.d.ts.map +1 -1
- package/dist/context/compaction/overflow.js +6 -10
- package/dist/context/compaction/providers/reactive-overflow-provider.cjs +15 -0
- package/dist/context/compaction/providers/reactive-overflow-provider.d.ts +15 -0
- package/dist/context/compaction/providers/reactive-overflow-provider.d.ts.map +1 -1
- package/dist/context/compaction/providers/reactive-overflow-provider.js +15 -0
- package/dist/context/compaction/schemas.cjs +22 -2
- package/dist/context/compaction/schemas.d.ts +45 -0
- package/dist/context/compaction/schemas.d.ts.map +1 -1
- package/dist/context/compaction/schemas.js +22 -2
- package/dist/context/compaction/strategies/reactive-overflow.cjs +166 -26
- package/dist/context/compaction/strategies/reactive-overflow.d.ts +21 -0
- package/dist/context/compaction/strategies/reactive-overflow.d.ts.map +1 -1
- package/dist/context/compaction/strategies/reactive-overflow.js +166 -26
- package/dist/context/manager.cjs +278 -31
- package/dist/context/manager.d.ts +192 -5
- package/dist/context/manager.d.ts.map +1 -1
- package/dist/context/manager.js +285 -32
- package/dist/context/types.d.ts +6 -0
- package/dist/context/types.d.ts.map +1 -1
- package/dist/context/utils.cjs +77 -11
- package/dist/context/utils.d.ts +86 -8
- package/dist/context/utils.d.ts.map +1 -1
- package/dist/context/utils.js +71 -11
- package/dist/events/index.cjs +4 -0
- package/dist/events/index.d.ts +41 -7
- package/dist/events/index.d.ts.map +1 -1
- package/dist/events/index.js +4 -0
- package/dist/llm/executor/stream-processor.cjs +19 -1
- package/dist/llm/executor/stream-processor.d.ts +3 -0
- package/dist/llm/executor/stream-processor.d.ts.map +1 -1
- package/dist/llm/executor/stream-processor.js +19 -1
- package/dist/llm/executor/turn-executor.cjs +219 -30
- package/dist/llm/executor/turn-executor.d.ts +62 -10
- package/dist/llm/executor/turn-executor.d.ts.map +1 -1
- package/dist/llm/executor/turn-executor.js +219 -30
- package/dist/llm/executor/types.d.ts +28 -0
- package/dist/llm/executor/types.d.ts.map +1 -1
- package/dist/llm/formatters/vercel.cjs +36 -28
- package/dist/llm/formatters/vercel.d.ts.map +1 -1
- package/dist/llm/formatters/vercel.js +36 -28
- package/dist/llm/services/factory.cjs +3 -2
- package/dist/llm/services/factory.d.ts +3 -1
- package/dist/llm/services/factory.d.ts.map +1 -1
- package/dist/llm/services/factory.js +3 -2
- package/dist/llm/services/vercel.cjs +34 -6
- package/dist/llm/services/vercel.d.ts +23 -3
- package/dist/llm/services/vercel.d.ts.map +1 -1
- package/dist/llm/services/vercel.js +34 -6
- package/dist/session/chat-session.cjs +20 -11
- package/dist/session/chat-session.d.ts +9 -4
- package/dist/session/chat-session.d.ts.map +1 -1
- package/dist/session/chat-session.js +20 -11
- package/dist/session/compaction-service.cjs +139 -0
- package/dist/session/compaction-service.d.ts +81 -0
- package/dist/session/compaction-service.d.ts.map +1 -0
- package/dist/session/compaction-service.js +106 -0
- package/dist/session/session-manager.cjs +146 -0
- package/dist/session/session-manager.d.ts +50 -0
- package/dist/session/session-manager.d.ts.map +1 -1
- package/dist/session/session-manager.js +146 -0
- package/dist/session/title-generator.cjs +2 -2
- package/dist/session/title-generator.js +2 -2
- package/dist/systemPrompt/in-built-prompts.cjs +36 -0
- package/dist/systemPrompt/in-built-prompts.d.ts +18 -1
- package/dist/systemPrompt/in-built-prompts.d.ts.map +1 -1
- package/dist/systemPrompt/in-built-prompts.js +25 -0
- package/dist/systemPrompt/manager.cjs +22 -0
- package/dist/systemPrompt/manager.d.ts +10 -0
- package/dist/systemPrompt/manager.d.ts.map +1 -1
- package/dist/systemPrompt/manager.js +22 -0
- package/dist/systemPrompt/registry.cjs +2 -1
- package/dist/systemPrompt/registry.d.ts +1 -1
- package/dist/systemPrompt/registry.d.ts.map +1 -1
- package/dist/systemPrompt/registry.js +2 -1
- package/dist/systemPrompt/schemas.cjs +7 -0
- package/dist/systemPrompt/schemas.d.ts +13 -13
- package/dist/systemPrompt/schemas.d.ts.map +1 -1
- package/dist/systemPrompt/schemas.js +7 -0
- package/dist/utils/index.cjs +3 -1
- package/dist/utils/index.d.ts +1 -0
- package/dist/utils/index.d.ts.map +1 -1
- package/dist/utils/index.js +1 -0
- package/package.json +1 -1
|
@@ -38,6 +38,23 @@ export declare class ContextManager<TMessage = unknown> {
|
|
|
38
38
|
* Maximum number of tokens allowed in the conversation (if specified)
|
|
39
39
|
*/
|
|
40
40
|
private maxInputTokens;
|
|
41
|
+
/**
|
|
42
|
+
* Last known actual input token count from the LLM API response.
|
|
43
|
+
* Updated after each LLM call. Used by /context for accurate reporting.
|
|
44
|
+
*/
|
|
45
|
+
private lastActualInputTokens;
|
|
46
|
+
/**
|
|
47
|
+
* Last known actual output token count from the LLM API response.
|
|
48
|
+
* Updated after each LLM call. Used in the context estimation formula:
|
|
49
|
+
* estimatedNextInput = lastInputTokens + lastOutputTokens + newMessagesEstimate
|
|
50
|
+
*/
|
|
51
|
+
private lastActualOutputTokens;
|
|
52
|
+
/**
|
|
53
|
+
* Message count at the time of the last LLM call.
|
|
54
|
+
* Used to identify which messages are "new" since the last call.
|
|
55
|
+
* Messages after this index are estimated with length/4 heuristic.
|
|
56
|
+
*/
|
|
57
|
+
private lastCallMessageCount;
|
|
41
58
|
private historyProvider;
|
|
42
59
|
private readonly sessionId;
|
|
43
60
|
/**
|
|
@@ -72,6 +89,71 @@ export declare class ContextManager<TMessage = unknown> {
|
|
|
72
89
|
* Returns the configured maximum number of input tokens for the conversation.
|
|
73
90
|
*/
|
|
74
91
|
getMaxInputTokens(): number;
|
|
92
|
+
/**
|
|
93
|
+
* Returns the last known actual input token count from the LLM API.
|
|
94
|
+
* Returns null if no LLM call has been made yet.
|
|
95
|
+
*/
|
|
96
|
+
getLastActualInputTokens(): number | null;
|
|
97
|
+
/**
|
|
98
|
+
* Updates the last known actual input token count.
|
|
99
|
+
* Called after each LLM response with the actual usage from the API.
|
|
100
|
+
*/
|
|
101
|
+
setLastActualInputTokens(tokens: number): void;
|
|
102
|
+
/**
|
|
103
|
+
* Returns the last known actual output token count from the LLM API.
|
|
104
|
+
* Returns null if no LLM call has been made yet.
|
|
105
|
+
*/
|
|
106
|
+
getLastActualOutputTokens(): number | null;
|
|
107
|
+
/**
|
|
108
|
+
* Updates the last known actual output token count.
|
|
109
|
+
* Called after each LLM response with the actual usage from the API.
|
|
110
|
+
*/
|
|
111
|
+
setLastActualOutputTokens(tokens: number): void;
|
|
112
|
+
/**
|
|
113
|
+
* Returns the message count at the time of the last LLM call.
|
|
114
|
+
* Returns null if no LLM call has been made yet.
|
|
115
|
+
*/
|
|
116
|
+
getLastCallMessageCount(): number | null;
|
|
117
|
+
/**
|
|
118
|
+
* Records the current message count after an LLM call completes.
|
|
119
|
+
* This marks the boundary for "new messages" calculation.
|
|
120
|
+
*/
|
|
121
|
+
recordLastCallMessageCount(): Promise<void>;
|
|
122
|
+
/**
|
|
123
|
+
* Resets the actual token tracking state.
|
|
124
|
+
* Called after compaction since the context has fundamentally changed.
|
|
125
|
+
*/
|
|
126
|
+
resetActualTokenTracking(): void;
|
|
127
|
+
/**
|
|
128
|
+
* Placeholder text used when tool outputs are pruned.
|
|
129
|
+
* Shared constant to ensure consistency between preparation and estimation.
|
|
130
|
+
*/
|
|
131
|
+
private static readonly PRUNED_TOOL_PLACEHOLDER;
|
|
132
|
+
/**
|
|
133
|
+
* Prepares conversation history for LLM consumption.
|
|
134
|
+
* This is the single source of truth for history transformation logic.
|
|
135
|
+
*
|
|
136
|
+
* Transformations applied:
|
|
137
|
+
* 1. filterCompacted - Remove pre-summary messages (messages before the most recent summary)
|
|
138
|
+
* 2. Transform pruned tool messages - Replace compactedAt messages with placeholder text
|
|
139
|
+
*
|
|
140
|
+
* Used by both:
|
|
141
|
+
* - getFormattedMessagesForLLM() - For actual LLM calls
|
|
142
|
+
* - getContextTokenEstimate() - For /context command estimation
|
|
143
|
+
*
|
|
144
|
+
* @returns Prepared history and statistics about the transformations
|
|
145
|
+
*/
|
|
146
|
+
prepareHistory(): Promise<{
|
|
147
|
+
preparedHistory: InternalMessage[];
|
|
148
|
+
stats: {
|
|
149
|
+
/** Total messages in raw history */
|
|
150
|
+
originalCount: number;
|
|
151
|
+
/** Messages after filterCompacted (removed pre-summary) */
|
|
152
|
+
filteredCount: number;
|
|
153
|
+
/** Messages with compactedAt that were transformed to placeholders */
|
|
154
|
+
prunedToolCount: number;
|
|
155
|
+
};
|
|
156
|
+
}>;
|
|
75
157
|
/**
|
|
76
158
|
* Assembles and returns the current system prompt by invoking the SystemPromptManager.
|
|
77
159
|
*/
|
|
@@ -190,18 +272,123 @@ export declare class ContextManager<TMessage = unknown> {
|
|
|
190
272
|
/**
|
|
191
273
|
* Gets the conversation ready for LLM consumption with proper flow:
|
|
192
274
|
* 1. Get system prompt
|
|
193
|
-
* 2.
|
|
194
|
-
* 3. Format messages
|
|
195
|
-
* This method implements the correct ordering to avoid circular dependencies.
|
|
275
|
+
* 2. Prepare history (filter + transform pruned messages)
|
|
276
|
+
* 3. Format messages for LLM API
|
|
196
277
|
*
|
|
197
278
|
* @param contributorContext The DynamicContributorContext for system prompt contributors and formatting
|
|
198
279
|
* @param llmContext The llmContext for the formatter to decide which messages to include based on the model's capabilities
|
|
199
|
-
* @returns Object containing formatted messages and
|
|
280
|
+
* @returns Object containing formatted messages, system prompt, and prepared history
|
|
200
281
|
*/
|
|
201
|
-
|
|
282
|
+
getFormattedMessagesForLLM(contributorContext: DynamicContributorContext, llmContext: LLMContext): Promise<{
|
|
202
283
|
formattedMessages: TMessage[];
|
|
203
284
|
systemPrompt: string;
|
|
285
|
+
preparedHistory: InternalMessage[];
|
|
286
|
+
}>;
|
|
287
|
+
/**
|
|
288
|
+
* Estimates context token usage for the /context command and compaction decisions.
|
|
289
|
+
* Uses the same prepareHistory() logic as getFormattedMessagesForLLM() to ensure consistency.
|
|
290
|
+
*
|
|
291
|
+
* When actuals are available from previous LLM calls:
|
|
292
|
+
* estimatedNextInput = lastInputTokens + lastOutputTokens + newMessagesEstimate
|
|
293
|
+
*
|
|
294
|
+
* This formula is more accurate because:
|
|
295
|
+
* - lastInputTokens: exactly what the API processed (ground truth)
|
|
296
|
+
* - lastOutputTokens: exactly what the LLM returned (ground truth)
|
|
297
|
+
* - newMessagesEstimate: only estimate the delta (tool results, new user messages)
|
|
298
|
+
*
|
|
299
|
+
* When no LLM call has been made yet (or after compaction), falls back to pure estimation.
|
|
300
|
+
*
|
|
301
|
+
* @param contributorContext Context for building the system prompt
|
|
302
|
+
* @param tools Tool definitions to include in the estimate
|
|
303
|
+
* @returns Token estimates with breakdown and comparison to actual (if available)
|
|
304
|
+
*/
|
|
305
|
+
getContextTokenEstimate(contributorContext: DynamicContributorContext, tools: Record<string, {
|
|
306
|
+
name?: string;
|
|
307
|
+
description?: string;
|
|
308
|
+
parameters?: unknown;
|
|
309
|
+
}>): Promise<{
|
|
310
|
+
/** Total estimated tokens */
|
|
311
|
+
estimated: number;
|
|
312
|
+
/** Last actual token count from LLM API (null if no calls made yet) */
|
|
313
|
+
actual: number | null;
|
|
314
|
+
/** Breakdown by category */
|
|
315
|
+
breakdown: {
|
|
316
|
+
systemPrompt: number;
|
|
317
|
+
tools: {
|
|
318
|
+
total: number;
|
|
319
|
+
perTool: Array<{
|
|
320
|
+
name: string;
|
|
321
|
+
tokens: number;
|
|
322
|
+
}>;
|
|
323
|
+
};
|
|
324
|
+
messages: number;
|
|
325
|
+
};
|
|
326
|
+
/** Preparation stats */
|
|
327
|
+
stats: {
|
|
328
|
+
originalMessageCount: number;
|
|
329
|
+
filteredMessageCount: number;
|
|
330
|
+
prunedToolCount: number;
|
|
331
|
+
};
|
|
332
|
+
/** Calculation basis for debugging/display */
|
|
333
|
+
calculationBasis?: {
|
|
334
|
+
/** Whether we used the actual-based formula or pure estimation */
|
|
335
|
+
method: 'actuals' | 'estimate';
|
|
336
|
+
/** Last actual input tokens from API (if method is 'actuals') */
|
|
337
|
+
lastInputTokens?: number;
|
|
338
|
+
/** Last actual output tokens from API (if method is 'actuals') */
|
|
339
|
+
lastOutputTokens?: number;
|
|
340
|
+
/** Estimated tokens for new messages since last call (if method is 'actuals') */
|
|
341
|
+
newMessagesEstimate?: number;
|
|
342
|
+
};
|
|
204
343
|
}>;
|
|
344
|
+
/**
|
|
345
|
+
* Estimates the next input token count using actual token data from the previous LLM call.
|
|
346
|
+
* This is a lightweight version for compaction pre-checks that only returns the total.
|
|
347
|
+
*
|
|
348
|
+
* ## Formula (when actuals are available):
|
|
349
|
+
* estimatedNextInput = lastInputTokens + lastOutputTokens + newMessagesEstimate
|
|
350
|
+
*
|
|
351
|
+
* ## Why this formula works:
|
|
352
|
+
*
|
|
353
|
+
* Consider two consecutive LLM calls:
|
|
354
|
+
*
|
|
355
|
+
* ```
|
|
356
|
+
* Call N:
|
|
357
|
+
* Input sent: system + tools + [user1] = lastInput tokens
|
|
358
|
+
* Output received: assistant response = lastOutput tokens
|
|
359
|
+
*
|
|
360
|
+
* Call N+1:
|
|
361
|
+
* Input will be: system + tools + [user1, assistant1, user2, ...]
|
|
362
|
+
* ≈ lastInput + assistant1_as_input + new_messages
|
|
363
|
+
* ≈ lastInput + lastOutput + newMessagesEstimate
|
|
364
|
+
* ```
|
|
365
|
+
*
|
|
366
|
+
* The assistant's response (lastOutput) becomes part of the next input as conversation
|
|
367
|
+
* history. Text tokenizes similarly whether sent as input or received as output.
|
|
368
|
+
*
|
|
369
|
+
* ## No double-counting:
|
|
370
|
+
*
|
|
371
|
+
* The assistant message is added to history DURING streaming (before this method runs),
|
|
372
|
+
* and recordLastCallMessageCount() captures the count INCLUDING that message.
|
|
373
|
+
* Therefore, newMessages = history.slice(lastMsgCount) EXCLUDES the assistant message,
|
|
374
|
+
* so lastOutput and newMessages don't overlap.
|
|
375
|
+
*
|
|
376
|
+
* ## Pruning caveat:
|
|
377
|
+
*
|
|
378
|
+
* If tool output pruning occurs between calls, lastInput may be stale (higher than
|
|
379
|
+
* actual). This causes OVERESTIMATION, which is SAFE - we'd trigger compaction
|
|
380
|
+
* earlier rather than risk context overflow.
|
|
381
|
+
*
|
|
382
|
+
* @param systemPrompt The system prompt string
|
|
383
|
+
* @param preparedHistory Message history AFTER filterCompacted and pruning
|
|
384
|
+
* @param tools Tool definitions
|
|
385
|
+
* @returns Estimated total input tokens for the next LLM call
|
|
386
|
+
*/
|
|
387
|
+
getEstimatedNextInputTokens(systemPrompt: string, preparedHistory: readonly InternalMessage[], tools: Record<string, {
|
|
388
|
+
name?: string;
|
|
389
|
+
description?: string;
|
|
390
|
+
parameters?: unknown;
|
|
391
|
+
}>): Promise<number>;
|
|
205
392
|
/**
|
|
206
393
|
* Gets the system prompt formatted for the target LLM provider
|
|
207
394
|
* Some providers handle system prompts differently
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"manager.d.ts","sourceRoot":"","sources":["../../src/context/manager.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,sBAAsB,EAAE,MAAM,gCAAgC,CAAC;AACxE,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,KAAK,EAAE,eAAe,EAAE,gBAAgB,EAAE,QAAQ,EAAE,MAAM,YAAY,CAAC;AAE9E,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;
|
|
1
|
+
{"version":3,"file":"manager.d.ts","sourceRoot":"","sources":["../../src/context/manager.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,sBAAsB,EAAE,MAAM,gCAAgC,CAAC;AACxE,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,KAAK,EAAE,eAAe,EAAE,gBAAgB,EAAE,QAAQ,EAAE,MAAM,YAAY,CAAC;AAE9E,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAU1D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,YAAY,CAAC;AACtD,OAAO,EAAE,yBAAyB,EAAE,MAAM,0BAA0B,CAAC;AACrE,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACjE,OAAO,EAAE,4BAA4B,EAAE,MAAM,gCAAgC,CAAC;AAE9E,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AAEvD;;;;;;;;;;;;;GAaG;AACH,qBAAa,cAAc,CAAC,QAAQ,GAAG,OAAO;IAC1C;;OAEG;IACH,OAAO,CAAC,SAAS,CAAqB;IAEtC;;OAEG;IACH,OAAO,CAAC,mBAAmB,CAAsB;IAEjD;;OAEG;IACH,OAAO,CAAC,SAAS,CAAyB;IAE1C;;OAEG;IACH,OAAO,CAAC,cAAc,CAAS;IAE/B;;;OAGG;IACH,OAAO,CAAC,qBAAqB,CAAuB;IAEpD;;;;OAIG;IACH,OAAO,CAAC,sBAAsB,CAAuB;IAErD;;;;OAIG;IACH,OAAO,CAAC,oBAAoB,CAAuB;IAEnD,OAAO,CAAC,eAAe,CAA+B;IACtD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAS;IAEnC;;;;OAIG;IACH,OAAO,CAAC,eAAe,CAAkD;IAEzE,OAAO,CAAC,MAAM,CAAe;IAE7B;;;;;;;;;;OAUG;gBAEC,SAAS,EAAE,kBAAkB,EAC7B,SAAS,EAAE,sBAAsB,EACjC,mBAAmB,EAAE,mBAAmB,EACxC,cAAc,EAAE,MAAM,EACtB,eAAe,EAAE,4BAA4B,EAC7C,SAAS,EAAE,MAAM,EACjB,eAAe,EAAE,OAAO,uBAAuB,EAAE,eAAe,EAChE,MAAM,EAAE,YAAY;IAgBxB;;OAEG;IACI,kBAAkB,IAAI,OAAO,uBAAuB,EAAE,eAAe;IAI5E;;;OAGG;YACW,gBAAgB;IAoF9B;;OAEG;IACH,iBAAiB,IAAI,MAAM;IAI3B;;;OAGG;IACH,wBAAwB,IAAI,MAAM,GAAG,IAAI;IAIzC;;;OAGG;IACH,wBAAwB,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI;IAK9C;;;OAGG;IACH,yBAAyB,IAAI,MAAM,GAAG,IAAI;IAI1C;;;OAGG;IACH,yBAAyB,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI;IAK/C;;;OAGG;IACH,uBAAuB,IAAI,MAAM,GAAG,IAAI;IAIxC;;;OAGG;IACG,0BAA0B,IAAI,OAAO,CAAC,IAAI,CAAC;IAMjD;;;OAGG;IACH,wBAAwB,IAAI,IAAI;IAShC;;;OAGG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,uBAAuB,CAAuC;IAEtF;;;;;;;;;;;;;OAaG;IACG,cAAc,IAAI,OAAO,CAAC;QAC5B,eAAe,EAAE,eAAe,EAAE,CAAC;QACnC,KAAK,EAAE;YACH,oCAAoC;YACpC,aAAa,EAAE,MAAM,CAAC;YACtB,2DAA2D;YAC3D,aAAa,EAAE,MAAM,CAAC;YACtB,sEAAsE;YACtE,eAAe,EAAE,MAAM,CAAC;SAC3B,CAAC;KACL,CAAC;IA8CF;;OAEG;IACG,eAAe,CAAC,OAAO,EAAE,yBAAyB,GAAG,OAAO,CAAC,MAAM,CAAC;IAM1E;;;;;OAKG;IACG,UAAU,IAAI,OAAO,CAAC,QAAQ,CAAC,eAAe,EAAE,CAAC,CAAC;IAKxD;;;;OAIG;IACG,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;IAI5B;;;;;;;;OAQG;IACG,YAAY,IAAI,OAAO,CAAC,IAAI,CAAC;IAiBnC;;;OAGG;IACG,mBAAmB,CAAC,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;IAiCzE;;;OAGG;IACG,WAAW,CAAC,SAAS,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,GAAG,OAAO,CAAC,IAAI,CAAC;IAyBvE;;;OAGG;IACG,sBAAsB,CACxB,SAAS,EAAE,MAAM,EACjB,OAAO,EAAE,OAAO,CAAC,eAAe,CAAC,GAClC,OAAO,CAAC,IAAI,CAAC;IAqBhB;;;;;;;;;;;OAWG;IACG,uBAAuB,CAAC,UAAU,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC;IA2CpE;;;;;;;OAOG;IACG,UAAU,CAAC,OAAO,EAAE,eAAe,GAAG,OAAO,CAAC,IAAI,CAAC;IA8EzD;;;;;;OAMG;IACG,cAAc,CAAC,OAAO,EAAE,OAAO,YAAY,EAAE,WAAW,EAAE,GAAG,OAAO,CAAC,IAAI,CAAC;IAuEhF;;;;;;;;OAQG;IACG,mBAAmB,CACrB,OAAO,EAAE,MAAM,GAAG,IAAI,EACtB,SAAS,CAAC,EAAE,gBAAgB,CAAC,WAAW,CAAC,EACzC,QAAQ,CAAC,EAAE;QACP,UAAU,CAAC,EAAE,gBAAgB,CAAC,YAAY,CAAC,CAAC;QAC5C,SAAS,CAAC,EAAE,MAAM,CAAC;KACtB,GACF,OAAO,CAAC,IAAI,CAAC;IAmBhB;;;;;;;;;;;OAWG;IACG,aAAa,CACf,UAAU,EAAE,MAAM,EAClB,IAAI,EAAE,MAAM,EACZ,eAAe,EAAE,mBAAmB,EACpC,QAAQ,CAAC,EAAE;QACP,eAAe,CAAC,EAAE,OAAO,CAAC;QAC1B,cAAc,CAAC,EAAE,UAAU,GAAG,UAAU,CAAC;KAC5C,GACF,OAAO,CAAC,IAAI,CAAC;IAuChB;;;;;;;;;;;;OAYG;IACG,oBAAoB,CACtB,kBAAkB,EAAE,yBAAyB,EAC7C,UAAU,EAAE,UAAU,EACtB,YAAY,CAAC,EAAE,MAAM,GAAG,SAAS,EACjC,OAAO,CAAC,EAAE,eAAe,EAAE,GAC5B,OAAO,CAAC,QAAQ,EAAE,CAAC;IAyEtB;;;;;;;;;OASG;IACG,0BAA0B,CAC5B,kBAAkB,EAAE,yBAAyB,EAC7C,UAAU,EAAE,UAAU,GACvB,OAAO,CAAC;QACP,iBAAiB,EAAE,QAAQ,EAAE,CAAC;QAC9B,YAAY,EAAE,MAAM,CAAC;QACrB,eAAe,EAAE,eAAe,EAAE,CAAC;KACtC,CAAC;IAsBF;;;;;;;;;;;;;;;;;OAiBG;IACG,uBAAuB,CACzB,kBAAkB,EAAE,yBAAyB,EAC7C,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE;QAAE,IAAI,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,OAAO,CAAA;KAAE,CAAC,GACrF,OAAO,CAAC;QACP,6BAA6B;QAC7B,SAAS,EAAE,MAAM,CAAC;QAClB,uEAAuE;QACvE,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;QACtB,4BAA4B;QAC5B,SAAS,EAAE;YACP,YAAY,EAAE,MAAM,CAAC;YACrB,KAAK,EAAE;gBACH,KAAK,EAAE,MAAM,CAAC;gBACd,OAAO,EAAE,KAAK,CAAC;oBAAE,IAAI,EAAE,MAAM,CAAC;oBAAC,MAAM,EAAE,MAAM,CAAA;iBAAE,CAAC,CAAC;aACpD,CAAC;YACF,QAAQ,EAAE,MAAM,CAAC;SACpB,CAAC;QACF,wBAAwB;QACxB,KAAK,EAAE;YACH,oBAAoB,EAAE,MAAM,CAAC;YAC7B,oBAAoB,EAAE,MAAM,CAAC;YAC7B,eAAe,EAAE,MAAM,CAAC;SAC3B,CAAC;QACF,8CAA8C;QAC9C,gBAAgB,CAAC,EAAE;YACf,kEAAkE;YAClE,MAAM,EAAE,SAAS,GAAG,UAAU,CAAC;YAC/B,iEAAiE;YACjE,eAAe,CAAC,EAAE,MAAM,CAAC;YACzB,kEAAkE;YAClE,gBAAgB,CAAC,EAAE,MAAM,CAAC;YAC1B,iFAAiF;YACjF,mBAAmB,CAAC,EAAE,MAAM,CAAC;SAChC,CAAC;KACL,CAAC;IA+FF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OA0CG;IACG,2BAA2B,CAC7B,YAAY,EAAE,MAAM,EACpB,eAAe,EAAE,SAAS,eAAe,EAAE,EAC3C,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE;QAAE,IAAI,CAAC,EAAE,MAAM,CAAC;QAAC,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,OAAO,CAAA;KAAE,CAAC,GACrF,OAAO,CAAC,MAAM,CAAC;IAwBlB;;;;;;OAMG;IACG,wBAAwB,CAC1B,QAAQ,EAAE,yBAAyB,GACpC,OAAO,CAAC,MAAM,GAAG,IAAI,GAAG,SAAS,CAAC;IAKrC;;;OAGG;IACG,iBAAiB,IAAI,OAAO,CAAC,IAAI,CAAC;CAQ3C"}
|
package/dist/context/manager.js
CHANGED
|
@@ -3,7 +3,13 @@ import { randomUUID } from "crypto";
|
|
|
3
3
|
import { isSystemMessage, isUserMessage, isAssistantMessage, isToolMessage } from "./types.js";
|
|
4
4
|
import { DextoLogComponent } from "../logger/v2/types.js";
|
|
5
5
|
import { eventBus } from "../events/index.js";
|
|
6
|
-
import {
|
|
6
|
+
import {
|
|
7
|
+
expandBlobReferences,
|
|
8
|
+
isLikelyBase64String,
|
|
9
|
+
filterCompacted,
|
|
10
|
+
estimateContextTokens,
|
|
11
|
+
estimateMessagesTokens
|
|
12
|
+
} from "./utils.js";
|
|
7
13
|
import { ContextError } from "./errors.js";
|
|
8
14
|
class ContextManager {
|
|
9
15
|
/**
|
|
@@ -22,6 +28,23 @@ class ContextManager {
|
|
|
22
28
|
* Maximum number of tokens allowed in the conversation (if specified)
|
|
23
29
|
*/
|
|
24
30
|
maxInputTokens;
|
|
31
|
+
/**
|
|
32
|
+
* Last known actual input token count from the LLM API response.
|
|
33
|
+
* Updated after each LLM call. Used by /context for accurate reporting.
|
|
34
|
+
*/
|
|
35
|
+
lastActualInputTokens = null;
|
|
36
|
+
/**
|
|
37
|
+
* Last known actual output token count from the LLM API response.
|
|
38
|
+
* Updated after each LLM call. Used in the context estimation formula:
|
|
39
|
+
* estimatedNextInput = lastInputTokens + lastOutputTokens + newMessagesEstimate
|
|
40
|
+
*/
|
|
41
|
+
lastActualOutputTokens = null;
|
|
42
|
+
/**
|
|
43
|
+
* Message count at the time of the last LLM call.
|
|
44
|
+
* Used to identify which messages are "new" since the last call.
|
|
45
|
+
* Messages after this index are estimated with length/4 heuristic.
|
|
46
|
+
*/
|
|
47
|
+
lastCallMessageCount = null;
|
|
25
48
|
historyProvider;
|
|
26
49
|
sessionId;
|
|
27
50
|
/**
|
|
@@ -120,6 +143,119 @@ class ContextManager {
|
|
|
120
143
|
getMaxInputTokens() {
|
|
121
144
|
return this.maxInputTokens;
|
|
122
145
|
}
|
|
146
|
+
/**
|
|
147
|
+
* Returns the last known actual input token count from the LLM API.
|
|
148
|
+
* Returns null if no LLM call has been made yet.
|
|
149
|
+
*/
|
|
150
|
+
getLastActualInputTokens() {
|
|
151
|
+
return this.lastActualInputTokens;
|
|
152
|
+
}
|
|
153
|
+
/**
|
|
154
|
+
* Updates the last known actual input token count.
|
|
155
|
+
* Called after each LLM response with the actual usage from the API.
|
|
156
|
+
*/
|
|
157
|
+
setLastActualInputTokens(tokens) {
|
|
158
|
+
this.lastActualInputTokens = tokens;
|
|
159
|
+
this.logger.debug(`Updated lastActualInputTokens: ${tokens}`);
|
|
160
|
+
}
|
|
161
|
+
/**
|
|
162
|
+
* Returns the last known actual output token count from the LLM API.
|
|
163
|
+
* Returns null if no LLM call has been made yet.
|
|
164
|
+
*/
|
|
165
|
+
getLastActualOutputTokens() {
|
|
166
|
+
return this.lastActualOutputTokens;
|
|
167
|
+
}
|
|
168
|
+
/**
|
|
169
|
+
* Updates the last known actual output token count.
|
|
170
|
+
* Called after each LLM response with the actual usage from the API.
|
|
171
|
+
*/
|
|
172
|
+
setLastActualOutputTokens(tokens) {
|
|
173
|
+
this.lastActualOutputTokens = tokens;
|
|
174
|
+
this.logger.debug(`Updated lastActualOutputTokens: ${tokens}`);
|
|
175
|
+
}
|
|
176
|
+
/**
|
|
177
|
+
* Returns the message count at the time of the last LLM call.
|
|
178
|
+
* Returns null if no LLM call has been made yet.
|
|
179
|
+
*/
|
|
180
|
+
getLastCallMessageCount() {
|
|
181
|
+
return this.lastCallMessageCount;
|
|
182
|
+
}
|
|
183
|
+
/**
|
|
184
|
+
* Records the current message count after an LLM call completes.
|
|
185
|
+
* This marks the boundary for "new messages" calculation.
|
|
186
|
+
*/
|
|
187
|
+
async recordLastCallMessageCount() {
|
|
188
|
+
const history = await this.historyProvider.getHistory();
|
|
189
|
+
this.lastCallMessageCount = history.length;
|
|
190
|
+
this.logger.debug(`Recorded lastCallMessageCount: ${this.lastCallMessageCount}`);
|
|
191
|
+
}
|
|
192
|
+
/**
|
|
193
|
+
* Resets the actual token tracking state.
|
|
194
|
+
* Called after compaction since the context has fundamentally changed.
|
|
195
|
+
*/
|
|
196
|
+
resetActualTokenTracking() {
|
|
197
|
+
this.lastActualInputTokens = null;
|
|
198
|
+
this.lastActualOutputTokens = null;
|
|
199
|
+
this.lastCallMessageCount = null;
|
|
200
|
+
this.logger.debug("Reset actual token tracking state (after compaction)");
|
|
201
|
+
}
|
|
202
|
+
// ============= HISTORY PREPARATION =============
|
|
203
|
+
/**
|
|
204
|
+
* Placeholder text used when tool outputs are pruned.
|
|
205
|
+
* Shared constant to ensure consistency between preparation and estimation.
|
|
206
|
+
*/
|
|
207
|
+
static PRUNED_TOOL_PLACEHOLDER = "[Old tool result content cleared]";
|
|
208
|
+
/**
|
|
209
|
+
* Prepares conversation history for LLM consumption.
|
|
210
|
+
* This is the single source of truth for history transformation logic.
|
|
211
|
+
*
|
|
212
|
+
* Transformations applied:
|
|
213
|
+
* 1. filterCompacted - Remove pre-summary messages (messages before the most recent summary)
|
|
214
|
+
* 2. Transform pruned tool messages - Replace compactedAt messages with placeholder text
|
|
215
|
+
*
|
|
216
|
+
* Used by both:
|
|
217
|
+
* - getFormattedMessagesForLLM() - For actual LLM calls
|
|
218
|
+
* - getContextTokenEstimate() - For /context command estimation
|
|
219
|
+
*
|
|
220
|
+
* @returns Prepared history and statistics about the transformations
|
|
221
|
+
*/
|
|
222
|
+
async prepareHistory() {
|
|
223
|
+
const fullHistory = await this.historyProvider.getHistory();
|
|
224
|
+
const originalCount = fullHistory.length;
|
|
225
|
+
let history = filterCompacted(fullHistory);
|
|
226
|
+
const filteredCount = history.length;
|
|
227
|
+
if (filteredCount < originalCount) {
|
|
228
|
+
this.logger.debug(
|
|
229
|
+
`prepareHistory: filterCompacted reduced from ${originalCount} to ${filteredCount} messages`
|
|
230
|
+
);
|
|
231
|
+
}
|
|
232
|
+
let prunedToolCount = 0;
|
|
233
|
+
history = history.map((msg) => {
|
|
234
|
+
if (msg.role === "tool" && msg.compactedAt) {
|
|
235
|
+
prunedToolCount++;
|
|
236
|
+
return {
|
|
237
|
+
...msg,
|
|
238
|
+
content: [
|
|
239
|
+
{ type: "text", text: ContextManager.PRUNED_TOOL_PLACEHOLDER }
|
|
240
|
+
]
|
|
241
|
+
};
|
|
242
|
+
}
|
|
243
|
+
return msg;
|
|
244
|
+
});
|
|
245
|
+
if (prunedToolCount > 0) {
|
|
246
|
+
this.logger.debug(
|
|
247
|
+
`prepareHistory: Transformed ${prunedToolCount} pruned tool messages to placeholders`
|
|
248
|
+
);
|
|
249
|
+
}
|
|
250
|
+
return {
|
|
251
|
+
preparedHistory: history,
|
|
252
|
+
stats: {
|
|
253
|
+
originalCount,
|
|
254
|
+
filteredCount,
|
|
255
|
+
prunedToolCount
|
|
256
|
+
}
|
|
257
|
+
};
|
|
258
|
+
}
|
|
123
259
|
/**
|
|
124
260
|
* Assembles and returns the current system prompt by invoking the SystemPromptManager.
|
|
125
261
|
*/
|
|
@@ -168,6 +304,7 @@ ${prompt}`);
|
|
|
168
304
|
}
|
|
169
305
|
};
|
|
170
306
|
await this.addMessage(clearMarker);
|
|
307
|
+
this.resetActualTokenTracking();
|
|
171
308
|
this.logger.debug(`Context cleared for session: ${this.sessionId}`);
|
|
172
309
|
}
|
|
173
310
|
/**
|
|
@@ -539,51 +676,166 @@ ${prompt}`);
|
|
|
539
676
|
/**
|
|
540
677
|
* Gets the conversation ready for LLM consumption with proper flow:
|
|
541
678
|
* 1. Get system prompt
|
|
542
|
-
* 2.
|
|
543
|
-
* 3. Format messages
|
|
544
|
-
* This method implements the correct ordering to avoid circular dependencies.
|
|
679
|
+
* 2. Prepare history (filter + transform pruned messages)
|
|
680
|
+
* 3. Format messages for LLM API
|
|
545
681
|
*
|
|
546
682
|
* @param contributorContext The DynamicContributorContext for system prompt contributors and formatting
|
|
547
683
|
* @param llmContext The llmContext for the formatter to decide which messages to include based on the model's capabilities
|
|
548
|
-
* @returns Object containing formatted messages and
|
|
684
|
+
* @returns Object containing formatted messages, system prompt, and prepared history
|
|
549
685
|
*/
|
|
550
|
-
async
|
|
686
|
+
async getFormattedMessagesForLLM(contributorContext, llmContext) {
|
|
551
687
|
const systemPrompt = await this.getSystemPrompt(contributorContext);
|
|
552
|
-
const
|
|
553
|
-
let history = filterCompacted(fullHistory);
|
|
554
|
-
if (history.length < fullHistory.length) {
|
|
555
|
-
this.logger.debug(
|
|
556
|
-
`filterCompacted: Reduced history from ${fullHistory.length} to ${history.length} messages (summary present)`
|
|
557
|
-
);
|
|
558
|
-
}
|
|
559
|
-
const compactedCount = history.filter((m) => m.role === "tool" && m.compactedAt).length;
|
|
560
|
-
if (compactedCount > 0) {
|
|
561
|
-
history = history.map((msg) => {
|
|
562
|
-
if (msg.role === "tool" && msg.compactedAt) {
|
|
563
|
-
return {
|
|
564
|
-
...msg,
|
|
565
|
-
content: [
|
|
566
|
-
{ type: "text", text: "[Old tool result content cleared]" }
|
|
567
|
-
]
|
|
568
|
-
};
|
|
569
|
-
}
|
|
570
|
-
return msg;
|
|
571
|
-
});
|
|
572
|
-
this.logger.debug(
|
|
573
|
-
`Transformed ${compactedCount} compacted tool messages to placeholders`
|
|
574
|
-
);
|
|
575
|
-
}
|
|
688
|
+
const { preparedHistory } = await this.prepareHistory();
|
|
576
689
|
const formattedMessages = await this.getFormattedMessages(
|
|
577
690
|
contributorContext,
|
|
578
691
|
llmContext,
|
|
579
692
|
systemPrompt,
|
|
580
|
-
|
|
693
|
+
preparedHistory
|
|
581
694
|
);
|
|
582
695
|
return {
|
|
583
696
|
formattedMessages,
|
|
584
|
-
systemPrompt
|
|
697
|
+
systemPrompt,
|
|
698
|
+
preparedHistory
|
|
699
|
+
};
|
|
700
|
+
}
|
|
701
|
+
/**
|
|
702
|
+
* Estimates context token usage for the /context command and compaction decisions.
|
|
703
|
+
* Uses the same prepareHistory() logic as getFormattedMessagesForLLM() to ensure consistency.
|
|
704
|
+
*
|
|
705
|
+
* When actuals are available from previous LLM calls:
|
|
706
|
+
* estimatedNextInput = lastInputTokens + lastOutputTokens + newMessagesEstimate
|
|
707
|
+
*
|
|
708
|
+
* This formula is more accurate because:
|
|
709
|
+
* - lastInputTokens: exactly what the API processed (ground truth)
|
|
710
|
+
* - lastOutputTokens: exactly what the LLM returned (ground truth)
|
|
711
|
+
* - newMessagesEstimate: only estimate the delta (tool results, new user messages)
|
|
712
|
+
*
|
|
713
|
+
* When no LLM call has been made yet (or after compaction), falls back to pure estimation.
|
|
714
|
+
*
|
|
715
|
+
* @param contributorContext Context for building the system prompt
|
|
716
|
+
* @param tools Tool definitions to include in the estimate
|
|
717
|
+
* @returns Token estimates with breakdown and comparison to actual (if available)
|
|
718
|
+
*/
|
|
719
|
+
async getContextTokenEstimate(contributorContext, tools) {
|
|
720
|
+
const systemPrompt = await this.getSystemPrompt(contributorContext);
|
|
721
|
+
const { preparedHistory, stats } = await this.prepareHistory();
|
|
722
|
+
const lastInput = this.lastActualInputTokens;
|
|
723
|
+
const lastOutput = this.lastActualOutputTokens;
|
|
724
|
+
const lastMsgCount = this.lastCallMessageCount;
|
|
725
|
+
const currentHistory = await this.historyProvider.getHistory();
|
|
726
|
+
const pureEstimate = estimateContextTokens(systemPrompt, preparedHistory, tools);
|
|
727
|
+
let total;
|
|
728
|
+
let calculationBasis;
|
|
729
|
+
if (lastInput !== null && lastOutput !== null && lastMsgCount !== null) {
|
|
730
|
+
const newMessages = currentHistory.slice(lastMsgCount);
|
|
731
|
+
const newMessagesEstimate = estimateMessagesTokens(newMessages);
|
|
732
|
+
total = lastInput + lastOutput + newMessagesEstimate;
|
|
733
|
+
calculationBasis = {
|
|
734
|
+
method: "actuals",
|
|
735
|
+
lastInputTokens: lastInput,
|
|
736
|
+
lastOutputTokens: lastOutput,
|
|
737
|
+
newMessagesEstimate
|
|
738
|
+
};
|
|
739
|
+
this.logger.info(
|
|
740
|
+
`Context estimate (actuals-based): lastInput=${lastInput}, lastOutput=${lastOutput}, newMsgs=${newMessagesEstimate} (${newMessages.length} messages), total=${total}`
|
|
741
|
+
);
|
|
742
|
+
} else {
|
|
743
|
+
total = pureEstimate.total;
|
|
744
|
+
calculationBasis = {
|
|
745
|
+
method: "estimate"
|
|
746
|
+
};
|
|
747
|
+
this.logger.debug(
|
|
748
|
+
`Context estimate (pure estimate): total=${total} (no actuals available yet)`
|
|
749
|
+
);
|
|
750
|
+
}
|
|
751
|
+
const systemPromptTokens = pureEstimate.breakdown.systemPrompt;
|
|
752
|
+
const toolsTokens = pureEstimate.breakdown.tools;
|
|
753
|
+
const messagesDisplay = Math.max(0, total - systemPromptTokens - toolsTokens.total);
|
|
754
|
+
if (lastInput !== null) {
|
|
755
|
+
const pureTotal = pureEstimate.total;
|
|
756
|
+
const diff = pureTotal - lastInput;
|
|
757
|
+
const diffPercent = lastInput > 0 ? (diff / lastInput * 100).toFixed(1) : "0.0";
|
|
758
|
+
this.logger.info(
|
|
759
|
+
`Context token calibration: pureEstimate=${pureTotal}, lastActual=${lastInput}, diff=${diff} (${diffPercent}%)`
|
|
760
|
+
);
|
|
761
|
+
}
|
|
762
|
+
return {
|
|
763
|
+
estimated: total,
|
|
764
|
+
actual: lastInput,
|
|
765
|
+
breakdown: {
|
|
766
|
+
systemPrompt: systemPromptTokens,
|
|
767
|
+
tools: toolsTokens,
|
|
768
|
+
messages: messagesDisplay
|
|
769
|
+
},
|
|
770
|
+
stats: {
|
|
771
|
+
originalMessageCount: stats.originalCount,
|
|
772
|
+
filteredMessageCount: stats.filteredCount,
|
|
773
|
+
prunedToolCount: stats.prunedToolCount
|
|
774
|
+
},
|
|
775
|
+
calculationBasis
|
|
585
776
|
};
|
|
586
777
|
}
|
|
778
|
+
/**
|
|
779
|
+
* Estimates the next input token count using actual token data from the previous LLM call.
|
|
780
|
+
* This is a lightweight version for compaction pre-checks that only returns the total.
|
|
781
|
+
*
|
|
782
|
+
* ## Formula (when actuals are available):
|
|
783
|
+
* estimatedNextInput = lastInputTokens + lastOutputTokens + newMessagesEstimate
|
|
784
|
+
*
|
|
785
|
+
* ## Why this formula works:
|
|
786
|
+
*
|
|
787
|
+
* Consider two consecutive LLM calls:
|
|
788
|
+
*
|
|
789
|
+
* ```
|
|
790
|
+
* Call N:
|
|
791
|
+
* Input sent: system + tools + [user1] = lastInput tokens
|
|
792
|
+
* Output received: assistant response = lastOutput tokens
|
|
793
|
+
*
|
|
794
|
+
* Call N+1:
|
|
795
|
+
* Input will be: system + tools + [user1, assistant1, user2, ...]
|
|
796
|
+
* ≈ lastInput + assistant1_as_input + new_messages
|
|
797
|
+
* ≈ lastInput + lastOutput + newMessagesEstimate
|
|
798
|
+
* ```
|
|
799
|
+
*
|
|
800
|
+
* The assistant's response (lastOutput) becomes part of the next input as conversation
|
|
801
|
+
* history. Text tokenizes similarly whether sent as input or received as output.
|
|
802
|
+
*
|
|
803
|
+
* ## No double-counting:
|
|
804
|
+
*
|
|
805
|
+
* The assistant message is added to history DURING streaming (before this method runs),
|
|
806
|
+
* and recordLastCallMessageCount() captures the count INCLUDING that message.
|
|
807
|
+
* Therefore, newMessages = history.slice(lastMsgCount) EXCLUDES the assistant message,
|
|
808
|
+
* so lastOutput and newMessages don't overlap.
|
|
809
|
+
*
|
|
810
|
+
* ## Pruning caveat:
|
|
811
|
+
*
|
|
812
|
+
* If tool output pruning occurs between calls, lastInput may be stale (higher than
|
|
813
|
+
* actual). This causes OVERESTIMATION, which is SAFE - we'd trigger compaction
|
|
814
|
+
* earlier rather than risk context overflow.
|
|
815
|
+
*
|
|
816
|
+
* @param systemPrompt The system prompt string
|
|
817
|
+
* @param preparedHistory Message history AFTER filterCompacted and pruning
|
|
818
|
+
* @param tools Tool definitions
|
|
819
|
+
* @returns Estimated total input tokens for the next LLM call
|
|
820
|
+
*/
|
|
821
|
+
async getEstimatedNextInputTokens(systemPrompt, preparedHistory, tools) {
|
|
822
|
+
const lastInput = this.lastActualInputTokens;
|
|
823
|
+
const lastOutput = this.lastActualOutputTokens;
|
|
824
|
+
const lastMsgCount = this.lastCallMessageCount;
|
|
825
|
+
const currentHistory = await this.historyProvider.getHistory();
|
|
826
|
+
if (lastInput !== null && lastOutput !== null && lastMsgCount !== null) {
|
|
827
|
+
const newMessages = currentHistory.slice(lastMsgCount);
|
|
828
|
+
const newMessagesEstimate = estimateMessagesTokens(newMessages);
|
|
829
|
+
const total = lastInput + lastOutput + newMessagesEstimate;
|
|
830
|
+
this.logger.debug(
|
|
831
|
+
`Estimated next input (actuals-based): ${lastInput} + ${lastOutput} + ${newMessagesEstimate} = ${total}`
|
|
832
|
+
);
|
|
833
|
+
return total;
|
|
834
|
+
}
|
|
835
|
+
const pureEstimate = estimateContextTokens(systemPrompt, preparedHistory, tools);
|
|
836
|
+
this.logger.debug(`Estimated next input (pure estimate): ${pureEstimate.total}`);
|
|
837
|
+
return pureEstimate.total;
|
|
838
|
+
}
|
|
587
839
|
/**
|
|
588
840
|
* Gets the system prompt formatted for the target LLM provider
|
|
589
841
|
* Some providers handle system prompts differently
|
|
@@ -600,6 +852,7 @@ ${prompt}`);
|
|
|
600
852
|
*/
|
|
601
853
|
async resetConversation() {
|
|
602
854
|
await this.historyProvider.clearHistory();
|
|
855
|
+
this.resetActualTokenTracking();
|
|
603
856
|
this.logger.debug(
|
|
604
857
|
`ContextManager: Conversation history cleared for session ${this.sessionId}`
|
|
605
858
|
);
|
package/dist/context/types.d.ts
CHANGED
|
@@ -186,6 +186,12 @@ export interface AssistantMessage extends MessageBase {
|
|
|
186
186
|
* Present when the provider supports reasoning and returns a final reasoning trace.
|
|
187
187
|
*/
|
|
188
188
|
reasoning?: string;
|
|
189
|
+
/**
|
|
190
|
+
* Provider-specific metadata for reasoning, used for round-tripping.
|
|
191
|
+
* Contains opaque tokens (e.g., OpenAI itemId, Gemini thought signatures)
|
|
192
|
+
* that must be passed back to the provider on subsequent requests.
|
|
193
|
+
*/
|
|
194
|
+
reasoningMetadata?: Record<string, unknown>;
|
|
189
195
|
/** Token usage accounting for this response */
|
|
190
196
|
tokenUsage?: TokenUsage;
|
|
191
197
|
/** Model identifier that generated this response */
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/context/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC/D,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,2BAA2B,CAAC;AAMjE;;;GAGG;AACH,MAAM,WAAW,SAAS;IACtB,KAAK,EAAE,MAAM,GAAG,UAAU,GAAG,MAAM,GAAG,WAAW,GAAG,GAAG,CAAC;IACxD,QAAQ,CAAC,EAAE,MAAM,CAAC;CACrB;AAED;;;GAGG;AACH,MAAM,WAAW,QAAQ;IACrB,IAAI,EAAE,MAAM,GAAG,UAAU,GAAG,MAAM,GAAG,WAAW,GAAG,GAAG,CAAC;IACvD,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;CACrB;AAED;;GAEG;AACH,MAAM,WAAW,QAAQ;IACrB,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,SAAU,SAAQ,SAAS;IACxC,IAAI,EAAE,OAAO,CAAC;CACjB;AAED;;GAEG;AACH,MAAM,WAAW,QAAS,SAAQ,QAAQ;IACtC,IAAI,EAAE,MAAM,CAAC;CAChB;AAED;;;;GAIG;AACH,MAAM,WAAW,cAAc;IAC3B,IAAI,EAAE,aAAa,CAAC;IACpB,6DAA6D;IAC7D,GAAG,EAAE,MAAM,CAAC;IACZ,gFAAgF;IAChF,QAAQ,EAAE,MAAM,CAAC;IACjB,mEAAmE;IACnE,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,4DAA4D;IAC5D,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,4CAA4C;IAC5C,QAAQ,CAAC,EAAE;QACP,wCAAwC;QACxC,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,yCAAyC;QACzC,aAAa,CAAC,EAAE;YAAE,KAAK,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAA;SAAE,CAAC;KACrD,CAAC;CACL;AAED;;;GAGG;AACH,MAAM,MAAM,WAAW,GAAG,QAAQ,GAAG,SAAS,GAAG,QAAQ,GAAG,cAAc,CAAC;AAM3E;;GAEG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,QAAQ,CAE9D;AAED;;GAEG;AACH,wBAAgB,WAAW,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,SAAS,CAEhE;AAED;;GAEG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,QAAQ,CAE9D;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,cAAc,CAE1E;AAMD;;GAEG;AACH,MAAM,WAAW,mBAAmB;IAChC,uEAAuE;IACvE,OAAO,EAAE,WAAW,EAAE,CAAC;IACvB;;;OAGG;IACH,SAAS,CAAC,EAAE,KAAK,CAAC;QACd,GAAG,EAAE,MAAM,CAAC;QACZ,IAAI,EAAE,OAAO,GAAG,OAAO,GAAG,OAAO,GAAG,QAAQ,CAAC;QAC7C,QAAQ,EAAE,MAAM,CAAC;QACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;KACrB,CAAC,CAAC;IACH,IAAI,EAAE;QACF,QAAQ,EAAE,MAAM,CAAC;QACjB,UAAU,EAAE,MAAM,CAAC;QACnB,gFAAgF;QAChF,OAAO,EAAE,OAAO,CAAC;QACjB,sFAAsF;QACtF,OAAO,CAAC,EAAE,eAAe,CAAC;KAC7B,CAAC;CACL;AAQD;;GAEG;AACH,MAAM,WAAW,QAAQ;IACrB,2CAA2C;IAC3C,EAAE,EAAE,MAAM,CAAC;IACX,qEAAqE;IACrE,IAAI,EAAE,UAAU,CAAC;IACjB,4BAA4B;IAC5B,QAAQ,EAAE;QACN,mCAAmC;QACnC,IAAI,EAAE,MAAM,CAAC;QACb,uDAAuD;QACvD,SAAS,EAAE,MAAM,CAAC;KACrB,CAAC;IACF;;;;OAIG;IACH,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAC7C;AAED;;;GAGG;AACH,MAAM,MAAM,kBAAkB,GAAG,SAAS,GAAG,UAAU,GAAG,UAAU,CAAC;AAMrE;;;GAGG;AACH,UAAU,WAAW;IACjB;;;OAGG;IACH,EAAE,CAAC,EAAE,MAAM,CAAC;IAEZ;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IAEnB;;;OAGG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACtC;AAED;;GAEG;AACH,MAAM,WAAW,aAAc,SAAQ,WAAW;IAC9C,IAAI,EAAE,QAAQ,CAAC;IACf,sDAAsD;IACtD,OAAO,EAAE,WAAW,EAAE,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,WAAY,SAAQ,WAAW;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,mDAAmD;IACnD,OAAO,EAAE,WAAW,EAAE,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,gBAAiB,SAAQ,WAAW;IACjD,IAAI,EAAE,WAAW,CAAC;IAClB,kEAAkE;IAClE,OAAO,EAAE,WAAW,EAAE,GAAG,IAAI,CAAC;IAE9B;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IAEnB,+CAA+C;IAC/C,UAAU,CAAC,EAAE,UAAU,CAAC;IAExB,oDAAoD;IACpD,KAAK,CAAC,EAAE,MAAM,CAAC;IAEf,4CAA4C;IAC5C,QAAQ,CAAC,EAAE,WAAW,CAAC;IAEvB;;;OAGG;IACH,SAAS,CAAC,EAAE,QAAQ,EAAE,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,WAAY,SAAQ,WAAW;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,sDAAsD;IACtD,OAAO,EAAE,WAAW,EAAE,CAAC;IAEvB,mEAAmE;IACnE,UAAU,EAAE,MAAM,CAAC;IAEnB,4DAA4D;IAC5D,IAAI,EAAE,MAAM,CAAC;IAEb,gDAAgD;IAChD,OAAO,CAAC,EAAE,OAAO,CAAC;IAElB,qEAAqE;IACrE,eAAe,CAAC,EAAE,OAAO,CAAC;IAE1B,6CAA6C;IAC7C,cAAc,CAAC,EAAE,kBAAkB,CAAC;IAEpC;;;OAGG;IACH,WAAW,CAAC,EAAE,MAAM,CAAC;IAErB;;;OAGG;IACH,WAAW,CAAC,EAAE,eAAe,CAAC;CACjC;AAED;;;;;GAKG;AACH,MAAM,MAAM,eAAe,GAAG,aAAa,GAAG,WAAW,GAAG,gBAAgB,GAAG,WAAW,CAAC;AAM3F;;GAEG;AACH,wBAAgB,eAAe,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,aAAa,CAE1E;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,WAAW,CAEtE;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,gBAAgB,CAEhF;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,WAAW,CAEtE"}
|
|
1
|
+
{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/context/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC/D,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,2BAA2B,CAAC;AAMjE;;;GAGG;AACH,MAAM,WAAW,SAAS;IACtB,KAAK,EAAE,MAAM,GAAG,UAAU,GAAG,MAAM,GAAG,WAAW,GAAG,GAAG,CAAC;IACxD,QAAQ,CAAC,EAAE,MAAM,CAAC;CACrB;AAED;;;GAGG;AACH,MAAM,WAAW,QAAQ;IACrB,IAAI,EAAE,MAAM,GAAG,UAAU,GAAG,MAAM,GAAG,WAAW,GAAG,GAAG,CAAC;IACvD,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;CACrB;AAED;;GAEG;AACH,MAAM,WAAW,QAAQ;IACrB,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,SAAU,SAAQ,SAAS;IACxC,IAAI,EAAE,OAAO,CAAC;CACjB;AAED;;GAEG;AACH,MAAM,WAAW,QAAS,SAAQ,QAAQ;IACtC,IAAI,EAAE,MAAM,CAAC;CAChB;AAED;;;;GAIG;AACH,MAAM,WAAW,cAAc;IAC3B,IAAI,EAAE,aAAa,CAAC;IACpB,6DAA6D;IAC7D,GAAG,EAAE,MAAM,CAAC;IACZ,gFAAgF;IAChF,QAAQ,EAAE,MAAM,CAAC;IACjB,mEAAmE;IACnE,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,4DAA4D;IAC5D,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,4CAA4C;IAC5C,QAAQ,CAAC,EAAE;QACP,wCAAwC;QACxC,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,yCAAyC;QACzC,aAAa,CAAC,EAAE;YAAE,KAAK,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAA;SAAE,CAAC;KACrD,CAAC;CACL;AAED;;;GAGG;AACH,MAAM,MAAM,WAAW,GAAG,QAAQ,GAAG,SAAS,GAAG,QAAQ,GAAG,cAAc,CAAC;AAM3E;;GAEG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,QAAQ,CAE9D;AAED;;GAEG;AACH,wBAAgB,WAAW,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,SAAS,CAEhE;AAED;;GAEG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,QAAQ,CAE9D;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,WAAW,GAAG,IAAI,IAAI,cAAc,CAE1E;AAMD;;GAEG;AACH,MAAM,WAAW,mBAAmB;IAChC,uEAAuE;IACvE,OAAO,EAAE,WAAW,EAAE,CAAC;IACvB;;;OAGG;IACH,SAAS,CAAC,EAAE,KAAK,CAAC;QACd,GAAG,EAAE,MAAM,CAAC;QACZ,IAAI,EAAE,OAAO,GAAG,OAAO,GAAG,OAAO,GAAG,QAAQ,CAAC;QAC7C,QAAQ,EAAE,MAAM,CAAC;QACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;KACrB,CAAC,CAAC;IACH,IAAI,EAAE;QACF,QAAQ,EAAE,MAAM,CAAC;QACjB,UAAU,EAAE,MAAM,CAAC;QACnB,gFAAgF;QAChF,OAAO,EAAE,OAAO,CAAC;QACjB,sFAAsF;QACtF,OAAO,CAAC,EAAE,eAAe,CAAC;KAC7B,CAAC;CACL;AAQD;;GAEG;AACH,MAAM,WAAW,QAAQ;IACrB,2CAA2C;IAC3C,EAAE,EAAE,MAAM,CAAC;IACX,qEAAqE;IACrE,IAAI,EAAE,UAAU,CAAC;IACjB,4BAA4B;IAC5B,QAAQ,EAAE;QACN,mCAAmC;QACnC,IAAI,EAAE,MAAM,CAAC;QACb,uDAAuD;QACvD,SAAS,EAAE,MAAM,CAAC;KACrB,CAAC;IACF;;;;OAIG;IACH,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAC7C;AAED;;;GAGG;AACH,MAAM,MAAM,kBAAkB,GAAG,SAAS,GAAG,UAAU,GAAG,UAAU,CAAC;AAMrE;;;GAGG;AACH,UAAU,WAAW;IACjB;;;OAGG;IACH,EAAE,CAAC,EAAE,MAAM,CAAC;IAEZ;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IAEnB;;;OAGG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACtC;AAED;;GAEG;AACH,MAAM,WAAW,aAAc,SAAQ,WAAW;IAC9C,IAAI,EAAE,QAAQ,CAAC;IACf,sDAAsD;IACtD,OAAO,EAAE,WAAW,EAAE,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,WAAY,SAAQ,WAAW;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,mDAAmD;IACnD,OAAO,EAAE,WAAW,EAAE,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,gBAAiB,SAAQ,WAAW;IACjD,IAAI,EAAE,WAAW,CAAC;IAClB,kEAAkE;IAClE,OAAO,EAAE,WAAW,EAAE,GAAG,IAAI,CAAC;IAE9B;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IAEnB;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAE5C,+CAA+C;IAC/C,UAAU,CAAC,EAAE,UAAU,CAAC;IAExB,oDAAoD;IACpD,KAAK,CAAC,EAAE,MAAM,CAAC;IAEf,4CAA4C;IAC5C,QAAQ,CAAC,EAAE,WAAW,CAAC;IAEvB;;;OAGG;IACH,SAAS,CAAC,EAAE,QAAQ,EAAE,CAAC;CAC1B;AAED;;;GAGG;AACH,MAAM,WAAW,WAAY,SAAQ,WAAW;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,sDAAsD;IACtD,OAAO,EAAE,WAAW,EAAE,CAAC;IAEvB,mEAAmE;IACnE,UAAU,EAAE,MAAM,CAAC;IAEnB,4DAA4D;IAC5D,IAAI,EAAE,MAAM,CAAC;IAEb,gDAAgD;IAChD,OAAO,CAAC,EAAE,OAAO,CAAC;IAElB,qEAAqE;IACrE,eAAe,CAAC,EAAE,OAAO,CAAC;IAE1B,6CAA6C;IAC7C,cAAc,CAAC,EAAE,kBAAkB,CAAC;IAEpC;;;OAGG;IACH,WAAW,CAAC,EAAE,MAAM,CAAC;IAErB;;;OAGG;IACH,WAAW,CAAC,EAAE,eAAe,CAAC;CACjC;AAED;;;;;GAKG;AACH,MAAM,MAAM,eAAe,GAAG,aAAa,GAAG,WAAW,GAAG,gBAAgB,GAAG,WAAW,CAAC;AAM3F;;GAEG;AACH,wBAAgB,eAAe,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,aAAa,CAE1E;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,WAAW,CAEtE;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,gBAAgB,CAEhF;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,GAAG,EAAE,eAAe,GAAG,GAAG,IAAI,WAAW,CAEtE"}
|