@pentoshi/clai 4.1.0 → 4.2.1

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.
Files changed (266) hide show
  1. package/README.md +14 -0
  2. package/dist/agent/compaction-executor.d.ts +68 -0
  3. package/dist/agent/compaction-executor.js +343 -0
  4. package/dist/agent/compaction-executor.js.map +1 -0
  5. package/dist/agent/context-breakdown.js +2 -0
  6. package/dist/agent/context-breakdown.js.map +1 -1
  7. package/dist/agent/context-manager.d.ts +23 -9
  8. package/dist/agent/context-manager.js +72 -43
  9. package/dist/agent/context-manager.js.map +1 -1
  10. package/dist/agent/events.d.ts +3 -0
  11. package/dist/agent/loop-guard.d.ts +2 -0
  12. package/dist/agent/loop-guard.js +48 -3
  13. package/dist/agent/loop-guard.js.map +1 -1
  14. package/dist/agent/prompt-composer.js +6 -6
  15. package/dist/agent/prompt-composer.js.map +1 -1
  16. package/dist/agent/reliability-policy.d.ts +1 -0
  17. package/dist/agent/reliability-policy.js +6 -3
  18. package/dist/agent/reliability-policy.js.map +1 -1
  19. package/dist/agent/request-accounting.d.ts +97 -0
  20. package/dist/agent/request-accounting.js +207 -0
  21. package/dist/agent/request-accounting.js.map +1 -0
  22. package/dist/agent/request-budget.d.ts +2 -2
  23. package/dist/agent/request-budget.js +2 -4
  24. package/dist/agent/request-budget.js.map +1 -1
  25. package/dist/agent/runner.d.ts +8 -2
  26. package/dist/agent/runner.js +409 -178
  27. package/dist/agent/runner.js.map +1 -1
  28. package/dist/agent/session-title.d.ts +0 -26
  29. package/dist/agent/session-title.js +0 -80
  30. package/dist/agent/session-title.js.map +1 -1
  31. package/dist/agent/tool-history.d.ts +2 -2
  32. package/dist/agent/tool-history.js +19 -6
  33. package/dist/agent/tool-history.js.map +1 -1
  34. package/dist/app/adapters/agent-event-adapter.js +19 -0
  35. package/dist/app/adapters/agent-event-adapter.js.map +1 -1
  36. package/dist/app/adapters/current-agent-adapter.js +4 -0
  37. package/dist/app/adapters/current-agent-adapter.js.map +1 -1
  38. package/dist/app/commands/catalog.js +24 -0
  39. package/dist/app/commands/catalog.js.map +1 -1
  40. package/dist/app/controllers/cancel-coordinator.d.ts +51 -0
  41. package/dist/app/controllers/cancel-coordinator.js +48 -0
  42. package/dist/app/controllers/cancel-coordinator.js.map +1 -0
  43. package/dist/app/controllers/session-compact-helper.d.ts +19 -2
  44. package/dist/app/controllers/session-compact-helper.js +100 -127
  45. package/dist/app/controllers/session-compact-helper.js.map +1 -1
  46. package/dist/app/controllers/session-context-usage.d.ts +32 -15
  47. package/dist/app/controllers/session-context-usage.js +186 -51
  48. package/dist/app/controllers/session-context-usage.js.map +1 -1
  49. package/dist/app/controllers/session-controller.d.ts +20 -11
  50. package/dist/app/controllers/session-controller.js +111 -90
  51. package/dist/app/controllers/session-controller.js.map +1 -1
  52. package/dist/app/controllers/session-naming.d.ts +31 -0
  53. package/dist/app/controllers/session-naming.js +173 -0
  54. package/dist/app/controllers/session-naming.js.map +1 -0
  55. package/dist/app/controllers/session-persistence.d.ts +2 -2
  56. package/dist/app/controllers/session-persistence.js +3 -6
  57. package/dist/app/controllers/session-persistence.js.map +1 -1
  58. package/dist/app/controllers/session-turn-request.d.ts +2 -1
  59. package/dist/app/controllers/session-turn-request.js +3 -0
  60. package/dist/app/controllers/session-turn-request.js.map +1 -1
  61. package/dist/app/controllers/turn-controller.d.ts +2 -1
  62. package/dist/app/controllers/turn-controller.js +1 -0
  63. package/dist/app/controllers/turn-controller.js.map +1 -1
  64. package/dist/app/events/app-event.d.ts +9 -0
  65. package/dist/app/events/app-event.js.map +1 -1
  66. package/dist/app/ports/agent-port.d.ts +8 -1
  67. package/dist/classic/app/ClassicApp.js +2 -1
  68. package/dist/classic/app/ClassicApp.js.map +1 -1
  69. package/dist/classic/app/app-wiring.js +1 -2
  70. package/dist/classic/app/app-wiring.js.map +1 -1
  71. package/dist/classic/app/wiring-interactions.js +2 -7
  72. package/dist/classic/app/wiring-interactions.js.map +1 -1
  73. package/dist/classic/blocks/tool-lines.js +2 -1
  74. package/dist/classic/blocks/tool-lines.js.map +1 -1
  75. package/dist/classic/input/cancel-ladder.d.ts +2 -19
  76. package/dist/classic/input/cancel-ladder.js +13 -20
  77. package/dist/classic/input/cancel-ladder.js.map +1 -1
  78. package/dist/commands/providers.js +2 -0
  79. package/dist/commands/providers.js.map +1 -1
  80. package/dist/llm/adapters/anthropic-tools.d.ts +31 -14
  81. package/dist/llm/adapters/anthropic-tools.js +124 -29
  82. package/dist/llm/adapters/anthropic-tools.js.map +1 -1
  83. package/dist/llm/adapters/gemini-tools.d.ts +16 -4
  84. package/dist/llm/adapters/gemini-tools.js +117 -16
  85. package/dist/llm/adapters/gemini-tools.js.map +1 -1
  86. package/dist/llm/adapters/openai-tools.d.ts +13 -2
  87. package/dist/llm/adapters/openai-tools.js +58 -11
  88. package/dist/llm/adapters/openai-tools.js.map +1 -1
  89. package/dist/llm/agentrouter.js +68 -45
  90. package/dist/llm/agentrouter.js.map +1 -1
  91. package/dist/llm/anthropic.js +69 -36
  92. package/dist/llm/anthropic.js.map +1 -1
  93. package/dist/llm/aws-mantle.js +70 -32
  94. package/dist/llm/aws-mantle.js.map +1 -1
  95. package/dist/llm/bynara.js +26 -69
  96. package/dist/llm/bynara.js.map +1 -1
  97. package/dist/llm/capabilities.d.ts +2 -1
  98. package/dist/llm/capabilities.js +47 -0
  99. package/dist/llm/capabilities.js.map +1 -1
  100. package/dist/llm/context-snapshot.d.ts +100 -0
  101. package/dist/llm/context-snapshot.js +277 -0
  102. package/dist/llm/context-snapshot.js.map +1 -0
  103. package/dist/llm/context-windows.d.ts +10 -0
  104. package/dist/llm/context-windows.js +129 -0
  105. package/dist/llm/context-windows.js.map +1 -0
  106. package/dist/llm/custom-provider-profile.d.ts +75 -0
  107. package/dist/llm/custom-provider-profile.js +558 -0
  108. package/dist/llm/custom-provider-profile.js.map +1 -0
  109. package/dist/llm/custom-providers.d.ts +13 -1
  110. package/dist/llm/custom-providers.js +77 -25
  111. package/dist/llm/custom-providers.js.map +1 -1
  112. package/dist/llm/effort-fallback.d.ts +44 -0
  113. package/dist/llm/effort-fallback.js +101 -0
  114. package/dist/llm/effort-fallback.js.map +1 -0
  115. package/dist/llm/fireworks.js +4 -0
  116. package/dist/llm/fireworks.js.map +1 -1
  117. package/dist/llm/free.js +4 -0
  118. package/dist/llm/free.js.map +1 -1
  119. package/dist/llm/gemini.d.ts +1 -1
  120. package/dist/llm/gemini.js +89 -51
  121. package/dist/llm/gemini.js.map +1 -1
  122. package/dist/llm/groq.js +4 -0
  123. package/dist/llm/groq.js.map +1 -1
  124. package/dist/llm/hetzner.js +4 -0
  125. package/dist/llm/hetzner.js.map +1 -1
  126. package/dist/llm/http.d.ts +49 -10
  127. package/dist/llm/http.js +417 -153
  128. package/dist/llm/http.js.map +1 -1
  129. package/dist/llm/kimchi.js +4 -0
  130. package/dist/llm/kimchi.js.map +1 -1
  131. package/dist/llm/lightning.js +4 -0
  132. package/dist/llm/lightning.js.map +1 -1
  133. package/dist/llm/meta.d.ts +2 -0
  134. package/dist/llm/meta.js +811 -655
  135. package/dist/llm/meta.js.map +1 -1
  136. package/dist/llm/modal.js +4 -0
  137. package/dist/llm/modal.js.map +1 -1
  138. package/dist/llm/model-families.d.ts +9 -0
  139. package/dist/llm/model-families.js +39 -0
  140. package/dist/llm/model-families.js.map +1 -0
  141. package/dist/llm/nvidia.js +4 -0
  142. package/dist/llm/nvidia.js.map +1 -1
  143. package/dist/llm/ollama.js +44 -40
  144. package/dist/llm/ollama.js.map +1 -1
  145. package/dist/llm/openai.js +4 -0
  146. package/dist/llm/openai.js.map +1 -1
  147. package/dist/llm/openrouter.js +4 -0
  148. package/dist/llm/openrouter.js.map +1 -1
  149. package/dist/llm/operation-ledger.d.ts +60 -0
  150. package/dist/llm/operation-ledger.js +136 -0
  151. package/dist/llm/operation-ledger.js.map +1 -0
  152. package/dist/llm/operation-usage.d.ts +32 -0
  153. package/dist/llm/operation-usage.js +172 -0
  154. package/dist/llm/operation-usage.js.map +1 -0
  155. package/dist/llm/orcarouter.d.ts +7 -0
  156. package/dist/llm/orcarouter.js +166 -0
  157. package/dist/llm/orcarouter.js.map +1 -0
  158. package/dist/llm/provider-profile-layers.d.ts +7 -0
  159. package/dist/llm/provider-profile-layers.js +961 -0
  160. package/dist/llm/provider-profile-layers.js.map +1 -0
  161. package/dist/llm/provider-profile.d.ts +181 -0
  162. package/dist/llm/provider-profile.js +329 -0
  163. package/dist/llm/provider-profile.js.map +1 -0
  164. package/dist/llm/provider-profiles.d.ts +13 -0
  165. package/dist/llm/provider-profiles.js +86 -0
  166. package/dist/llm/provider-profiles.js.map +1 -0
  167. package/dist/llm/provider.d.ts +2 -0
  168. package/dist/llm/provider.js +67 -0
  169. package/dist/llm/provider.js.map +1 -1
  170. package/dist/llm/qwen-cloud.js +4 -0
  171. package/dist/llm/qwen-cloud.js.map +1 -1
  172. package/dist/llm/reasoning-artifacts.d.ts +101 -0
  173. package/dist/llm/reasoning-artifacts.js +440 -0
  174. package/dist/llm/reasoning-artifacts.js.map +1 -0
  175. package/dist/llm/request-fingerprint.d.ts +10 -0
  176. package/dist/llm/request-fingerprint.js +313 -0
  177. package/dist/llm/request-fingerprint.js.map +1 -0
  178. package/dist/llm/request-plan.d.ts +119 -0
  179. package/dist/llm/request-plan.js +279 -0
  180. package/dist/llm/request-plan.js.map +1 -0
  181. package/dist/llm/router.d.ts +19 -1
  182. package/dist/llm/router.js +447 -75
  183. package/dist/llm/router.js.map +1 -1
  184. package/dist/llm/stream-events.d.ts +47 -0
  185. package/dist/llm/stream-events.js +55 -0
  186. package/dist/llm/stream-events.js.map +1 -0
  187. package/dist/llm/stream-terminal.d.ts +30 -0
  188. package/dist/llm/stream-terminal.js +51 -0
  189. package/dist/llm/stream-terminal.js.map +1 -0
  190. package/dist/llm/token-estimate-calibration.d.ts +23 -0
  191. package/dist/llm/token-estimate-calibration.js +140 -0
  192. package/dist/llm/token-estimate-calibration.js.map +1 -0
  193. package/dist/llm/token-usage.d.ts +35 -5
  194. package/dist/llm/token-usage.js +212 -162
  195. package/dist/llm/token-usage.js.map +1 -1
  196. package/dist/llm/tokenrouter.js +8 -2
  197. package/dist/llm/tokenrouter.js.map +1 -1
  198. package/dist/modes/ask.js +11 -5
  199. package/dist/modes/ask.js.map +1 -1
  200. package/dist/noninteractive/start-noninteractive.js +31 -2
  201. package/dist/noninteractive/start-noninteractive.js.map +1 -1
  202. package/dist/noninteractive/stream-blocks.js +15 -4
  203. package/dist/noninteractive/stream-blocks.js.map +1 -1
  204. package/dist/prompts/embedded.js +1 -1
  205. package/dist/prompts/embedded.js.map +1 -1
  206. package/dist/prompts/system.agent.md +6 -0
  207. package/dist/store/config.js +5 -1
  208. package/dist/store/config.js.map +1 -1
  209. package/dist/store/history.d.ts +2 -0
  210. package/dist/store/history.js +19 -10
  211. package/dist/store/history.js.map +1 -1
  212. package/dist/store/keys.js +11 -1
  213. package/dist/store/keys.js.map +1 -1
  214. package/dist/store/plan.js +20 -14
  215. package/dist/store/plan.js.map +1 -1
  216. package/dist/tools/elevated-shell.d.ts +1 -1
  217. package/dist/tools/elevated-shell.js +24 -72
  218. package/dist/tools/elevated-shell.js.map +1 -1
  219. package/dist/tools/nmap-runner.js +30 -36
  220. package/dist/tools/nmap-runner.js.map +1 -1
  221. package/dist/tools/sudo-session.d.ts +83 -0
  222. package/dist/tools/sudo-session.js +124 -0
  223. package/dist/tools/sudo-session.js.map +1 -0
  224. package/dist/tui-v2/app/App.js +9 -24
  225. package/dist/tui-v2/app/App.js.map +1 -1
  226. package/dist/tui-v2/components/status/context-limit-chip.js +2 -2
  227. package/dist/tui-v2/components/status/context-limit-chip.js.map +1 -1
  228. package/dist/tui-v2/components/transcript/intro-card.js +2 -1
  229. package/dist/tui-v2/components/transcript/intro-card.js.map +1 -1
  230. package/dist/tui-v2/composer/composer-editor.js +2 -1
  231. package/dist/tui-v2/composer/composer-editor.js.map +1 -1
  232. package/dist/types.d.ts +156 -1
  233. package/dist/types.js +1 -0
  234. package/dist/types.js.map +1 -1
  235. package/dist/ui-core/bootstrap/composition-root.d.ts +2 -0
  236. package/dist/ui-core/bootstrap/composition-root.js +26 -3
  237. package/dist/ui-core/bootstrap/composition-root.js.map +1 -1
  238. package/dist/ui-core/commands/picker-commands.js +11 -5
  239. package/dist/ui-core/commands/picker-commands.js.map +1 -1
  240. package/dist/ui-core/commands/session-commands.js +41 -5
  241. package/dist/ui-core/commands/session-commands.js.map +1 -1
  242. package/dist/ui-core/plan/plan-lifecycle.js +3 -2
  243. package/dist/ui-core/plan/plan-lifecycle.js.map +1 -1
  244. package/dist/ui-core/rendering/context-limit.js +4 -1
  245. package/dist/ui-core/rendering/context-limit.js.map +1 -1
  246. package/dist/ui-core/rendering/duration.d.ts +2 -1
  247. package/dist/ui-core/rendering/duration.js +13 -5
  248. package/dist/ui-core/rendering/duration.js.map +1 -1
  249. package/dist/ui-core/rendering/intro-header.js +2 -2
  250. package/dist/ui-core/rendering/intro-header.js.map +1 -1
  251. package/dist/ui-core/rendering/theme.d.ts +1 -0
  252. package/dist/ui-core/rendering/theme.js +2 -0
  253. package/dist/ui-core/rendering/theme.js.map +1 -1
  254. package/dist/ui-core/state/transcript-hydrate.js +3 -3
  255. package/dist/ui-core/state/transcript-hydrate.js.map +1 -1
  256. package/dist/ui-core/state/transcript-reducer.d.ts +5 -0
  257. package/dist/ui-core/state/transcript-reducer.js +76 -3
  258. package/dist/ui-core/state/transcript-reducer.js.map +1 -1
  259. package/dist/ui-core/state/transcript-store.d.ts +10 -1
  260. package/dist/ui-core/state/transcript-store.js +22 -1
  261. package/dist/ui-core/state/transcript-store.js.map +1 -1
  262. package/dist/ui-core/state/transcript-types.d.ts +2 -0
  263. package/dist/ui-core/state/transcript-types.js.map +1 -1
  264. package/dist/version.generated.d.ts +2 -2
  265. package/dist/version.generated.js +2 -2
  266. package/package.json +6 -1
package/README.md CHANGED
@@ -155,6 +155,7 @@ This is the core of clai's design: assemble capacity from free tiers, then survi
155
155
  | Meta (Muse Spark) | `muse-spark-1.2` | Paid (per token) | `MODEL_API_KEY` |
156
156
  | Fireworks | `accounts/fireworks/models/kimi-k2p6` | Paid (per token) | `FIREWORKS_API_KEY` |
157
157
  | Hetzner | `Qwen/Qwen3.6-35B-A3B-FP8` | Free (experiment — no billing yet) | `HETZNER_API_KEY` |
158
+ | OrcaRouter | `openai/gpt-4o-mini` | Paid (per token, zero markup) | `ORCAROUTER_API_KEY` |
158
159
 
159
160
  Several "paid" providers also expose limited free allowances — the tier label reflects what the default keys usually buy you. Flip `freeOnly` off to include paid providers in fallback.
160
161
 
@@ -251,6 +252,18 @@ clai use hetzner # model defaults to Qwen/Qwen3.6-35B-A3B
251
252
 
252
253
  Both fetch the live model catalog for `/model` (cached 1h) and rotate keys like every other provider.
253
254
 
255
+ #### OrcaRouter (one key, eleven upstreams, zero markup)
256
+
257
+ [OrcaRouter](https://docs.orcarouter.ai) is an OpenAI-compatible gateway that routes OpenAI, Anthropic, Google Gemini, DeepSeek, xAI Grok, Qwen, Kimi, MiniMax, Z.ai GLM and more behind one key — billed at each provider's published price with no token markup. Model ids are vendor-prefixed (`openai/gpt-4o-mini`, `anthropic/claude-sonnet-4.6`, `google/gemini-2.5-flash`, `deepseek/deepseek-reasoner`, …), and `orcarouter/auto` picks the cheapest model that fits the request.
258
+
259
+ ```sh
260
+ clai set orcarouter sk-your-key # key: orcarouter.ai/console (starts with sk-)
261
+ clai use orcarouter # model defaults to openai/gpt-4o-mini
262
+ /model anthropic/claude-sonnet-4.6 # or any id from the live /models catalog
263
+ ```
264
+
265
+ Streaming, native tool calling, structured outputs (`response_format`), vision via `image_url` and prompt caching all work. Reasoning uses one unified `reasoning_effort` knob (low/medium/high, plus minimal/max on some models) that the gateway translates to each upstream's native shape — `/think` and `/effort` map onto it, and thinking arrives as `reasoning_content` in the usual thinking block. `/model` reads the live catalog (cached 1h), filtered to Chat-Completions-reachable models so image/video/tts ids stay out of the picker. Keys are multi-key with rotation like everywhere else; env var is `ORCAROUTER_API_KEY`.
266
+
254
267
  ### Manage keys
255
268
 
256
269
  ```sh
@@ -265,6 +278,7 @@ clai set lightning <key> # Lightning AI Model APIs
265
278
  clai set tokenrouter sk-your-key # TokenRouter
266
279
  clai set fireworks fw_your_key # Fireworks
267
280
  clai set hetzner your-token # Hetzner Inference (experiments.hetzner.com)
281
+ clai set orcarouter sk-your-key # OrcaRouter (orcarouter.ai/console)
268
282
  clai set free <key> # optional: unlock premium models (free is keyless by default)
269
283
  clai unset modal --url # drop stored endpoint URLs, keep the keys
270
284
  clai keys # providers + masked keys (★ active) + endpoint URLs
@@ -0,0 +1,68 @@
1
+ import type { ChatMessage, CompletionRequest, ProviderId, SuccessfulRequestSnapshot } from "../types.js";
2
+ import type { OperationLedger } from "../llm/operation-ledger.js";
3
+ import { type RequestAccounting } from "./request-accounting.js";
4
+ /** The assembled snapshot-replay request cannot fit the effective safe limit. */
5
+ export declare class CompactionOverLimitError extends Error {
6
+ readonly requestTokens: number;
7
+ readonly effectiveSafeTokens: number | undefined;
8
+ constructor(message: string, requestTokens: number, effectiveSafeTokens: number | undefined);
9
+ }
10
+ export declare function isCompactionOverLimitError(error: unknown): error is CompactionOverLimitError;
11
+ export interface CompactionSummaryExecution {
12
+ readonly provider: ProviderId | undefined;
13
+ readonly model: string | undefined;
14
+ readonly systemContent: string;
15
+ readonly prompt: string;
16
+ readonly maxTokens: number;
17
+ readonly signal?: AbortSignal | undefined;
18
+ readonly sourceMessages?: readonly ChatMessage[] | undefined;
19
+ readonly baseRequest?: SuccessfulRequestSnapshot | undefined;
20
+ readonly history?: readonly ChatMessage[] | undefined;
21
+ readonly contextLimitTokens?: number | undefined;
22
+ readonly tools?: CompletionRequest["tools"] | undefined;
23
+ readonly allowModelFallback?: boolean | undefined;
24
+ readonly stream: boolean;
25
+ readonly retryOnServerError?: boolean | undefined;
26
+ /** Backoff before the error retry; defaults to COMPACTION_ERROR_RETRY_DELAY_MS. */
27
+ readonly retryDelayMs?: number | undefined;
28
+ readonly qualityRetry?: boolean | undefined;
29
+ readonly operation?: OperationLedger | undefined;
30
+ readonly onToken?: ((text: string, replace?: boolean) => void) | undefined;
31
+ }
32
+ /**
33
+ * The cache-preserving compaction request: the exact messages of the last
34
+ * successful turn request, any history messages appended since, then the
35
+ * compaction instruction as the final user turn. The previous prompt is a
36
+ * strict prefix of this request, so APC providers serve it entirely from
37
+ * cache and only the tail + instruction bill as fresh input.
38
+ */
39
+ export declare function buildCompactionReplayMessages(baseRequest: SuccessfulRequestSnapshot, history: readonly ChatMessage[], userPrompt: string): ChatMessage[];
40
+ export interface CompactionReplayPlan {
41
+ readonly messages: ChatMessage[];
42
+ readonly accounting: RequestAccounting;
43
+ }
44
+ /**
45
+ * Pre-flight the snapshot-replay request without dispatching it. Returns
46
+ * undefined when the snapshot is not a usable prefix base for the live
47
+ * history; otherwise the replay messages and their serialized-request
48
+ * accounting, so the caller can pick the cache-preserving strategy only when
49
+ * the request actually fits (`!plan.accounting.overLimit`). The accounting
50
+ * reserves the summary completion budget and adds planner slack for the
51
+ * durable-state detail the final instruction prompt gains later.
52
+ *
53
+ * Replay compatibility: identical heads are the common case (the snapshot was
54
+ * captured from this session's live message array). History restored from
55
+ * disk drops the composed system head, so a headless history is accepted when
56
+ * its messages still match into the snapshot's body. Anything else means the
57
+ * snapshot belongs to a different lineage (provider/model switch, rewound or
58
+ * replaced history) and replaying it would resurrect dropped messages.
59
+ */
60
+ export declare function planCompactionReplay(input: {
61
+ readonly baseRequest: SuccessfulRequestSnapshot;
62
+ readonly history: readonly ChatMessage[];
63
+ readonly prompt: string;
64
+ readonly maxTokens: number;
65
+ readonly contextLimitTokens?: number | undefined;
66
+ readonly stream?: boolean | undefined;
67
+ }): CompactionReplayPlan | undefined;
68
+ export declare function executeCompactionSummary(execution: CompactionSummaryExecution): Promise<string>;
@@ -0,0 +1,343 @@
1
+ import { buildCompactionRetryPrompt, isCompactionCompletionTruncated, looksLikeIncompleteCompactionSummary, looksLikeTranscriptReplay, normalizeCompactionSummary, } from "./compaction-summary.js";
2
+ import { completeWithProvider, streamWithProvider } from "../llm/router.js";
3
+ import { streamAlreadyEmitted } from "../llm/stream-progress.js";
4
+ import { createThinkingStreamParser, stripThinking } from "../ui/thinking.js";
5
+ import { accountAssembledRequest, RequestOverLimitError, SAFETY_MARGIN_TOKENS, } from "./request-accounting.js";
6
+ import { isAbortError } from "./session-policy.js";
7
+ const RETRY_SYSTEM_SUFFIX = "\nReturn only a complete continuation-memory summary. Do not include analysis, reasoning, or <think> tags.";
8
+ /**
9
+ * Pause between a failed compaction admission and its retry: the failure is
10
+ * almost always transient upstream capacity (5xx, 429, a 200 whose stream
11
+ * errored), and an instant replay just re-hits the same hot route.
12
+ */
13
+ const COMPACTION_ERROR_RETRY_DELAY_MS = 1_500;
14
+ /**
15
+ * Extra headroom the replay planner adds on top of the standard safety margin.
16
+ * The planner sizes the request with a representative instruction prompt; the
17
+ * real prompt additionally embeds durable-state snippets extracted while the
18
+ * summary runs, so the plan must leave room for them.
19
+ */
20
+ const REPLAY_PLAN_SLACK_TOKENS = 4_096;
21
+ /**
22
+ * A compaction failure worth one more admission. Retryable by default: the
23
+ * only failures excluded are ones a retry cannot fix — aborts, partial streams
24
+ * (re-sending would double-print the summary card), fit-check rejections (the
25
+ * request is deterministically too large), and definitive request errors (4xx
26
+ * other than 408/429). Everything else is treated as transient: 5xx, 429/408,
27
+ * network resets, idle timeouts, and gateway responses that returned HTTP 200
28
+ * but failed upstream mid-stream.
29
+ */
30
+ function isCompactionRetryableError(error, signal) {
31
+ if (streamAlreadyEmitted(error))
32
+ return false;
33
+ if (isAbortError(error, signal))
34
+ return false;
35
+ if (isCompactionOverLimitError(error))
36
+ return false;
37
+ if (error instanceof RequestOverLimitError)
38
+ return false;
39
+ const status = error && typeof error === "object" && "status" in error
40
+ ? Number(error.status)
41
+ : 0;
42
+ if (status === 408 || status === 429)
43
+ return true;
44
+ if (status >= 400 && status < 500)
45
+ return false;
46
+ return true;
47
+ }
48
+ /** The assembled snapshot-replay request cannot fit the effective safe limit. */
49
+ export class CompactionOverLimitError extends Error {
50
+ requestTokens;
51
+ effectiveSafeTokens;
52
+ constructor(message, requestTokens, effectiveSafeTokens) {
53
+ super(message);
54
+ this.requestTokens = requestTokens;
55
+ this.effectiveSafeTokens = effectiveSafeTokens;
56
+ this.name = "CompactionOverLimitError";
57
+ }
58
+ }
59
+ export function isCompactionOverLimitError(error) {
60
+ return error instanceof CompactionOverLimitError;
61
+ }
62
+ const FAIL_CLOSED_BY_REASON = {
63
+ truncated: "compaction failed: model hit the summary output limit — original context retained",
64
+ "reasoning-only": "compaction failed: model returned no visible summary — original context retained",
65
+ replayed: "compaction failed: model replayed the transcript — original context retained",
66
+ incomplete: "compaction failed: model returned an incomplete summary — original context retained",
67
+ };
68
+ function comparableMessage(message) {
69
+ const { images: _images, ...rest } = message;
70
+ return JSON.stringify(rest);
71
+ }
72
+ function missingHistoryTail(requestMessages, history) {
73
+ let requestIndex = 0;
74
+ let matchedHistory = 0;
75
+ while (matchedHistory < history.length) {
76
+ const expected = comparableMessage(history[matchedHistory]);
77
+ let found = -1;
78
+ for (let index = requestIndex; index < requestMessages.length; index += 1) {
79
+ if (comparableMessage(requestMessages[index]) === expected) {
80
+ found = index;
81
+ break;
82
+ }
83
+ }
84
+ if (found < 0)
85
+ break;
86
+ requestIndex = found + 1;
87
+ matchedHistory += 1;
88
+ }
89
+ return history.slice(matchedHistory).map((message) => structuredClone(message));
90
+ }
91
+ /**
92
+ * The cache-preserving compaction request: the exact messages of the last
93
+ * successful turn request, any history messages appended since, then the
94
+ * compaction instruction as the final user turn. The previous prompt is a
95
+ * strict prefix of this request, so APC providers serve it entirely from
96
+ * cache and only the tail + instruction bill as fresh input.
97
+ */
98
+ export function buildCompactionReplayMessages(baseRequest, history, userPrompt) {
99
+ return [
100
+ ...baseRequest.messages.map((message) => structuredClone(message)),
101
+ ...missingHistoryTail(baseRequest.messages, history),
102
+ { role: "user", content: userPrompt },
103
+ ];
104
+ }
105
+ /**
106
+ * Pre-flight the snapshot-replay request without dispatching it. Returns
107
+ * undefined when the snapshot is not a usable prefix base for the live
108
+ * history; otherwise the replay messages and their serialized-request
109
+ * accounting, so the caller can pick the cache-preserving strategy only when
110
+ * the request actually fits (`!plan.accounting.overLimit`). The accounting
111
+ * reserves the summary completion budget and adds planner slack for the
112
+ * durable-state detail the final instruction prompt gains later.
113
+ *
114
+ * Replay compatibility: identical heads are the common case (the snapshot was
115
+ * captured from this session's live message array). History restored from
116
+ * disk drops the composed system head, so a headless history is accepted when
117
+ * its messages still match into the snapshot's body. Anything else means the
118
+ * snapshot belongs to a different lineage (provider/model switch, rewound or
119
+ * replaced history) and replaying it would resurrect dropped messages.
120
+ */
121
+ export function planCompactionReplay(input) {
122
+ const baseRequest = input.baseRequest;
123
+ const snapshotHead = baseRequest.messages[0];
124
+ const historyHead = input.history[0];
125
+ if (!snapshotHead || !historyHead)
126
+ return undefined;
127
+ const tail = missingHistoryTail(baseRequest.messages, input.history);
128
+ if (comparableMessage(snapshotHead) !== comparableMessage(historyHead)) {
129
+ if (historyHead.role === "system")
130
+ return undefined;
131
+ if (tail.length >= input.history.length)
132
+ return undefined;
133
+ }
134
+ const messages = [
135
+ ...baseRequest.messages.map((message) => structuredClone(message)),
136
+ ...tail,
137
+ { role: "user", content: input.prompt },
138
+ ];
139
+ const { accounting } = accountAssembledRequest({
140
+ provider: baseRequest.provider,
141
+ model: baseRequest.model,
142
+ messages,
143
+ stream: input.stream ?? true,
144
+ ...(baseRequest.tools?.length ? { tools: baseRequest.tools } : {}),
145
+ ...(baseRequest.toolChoice !== undefined
146
+ ? { toolChoice: baseRequest.toolChoice }
147
+ : {}),
148
+ ...(baseRequest.parallelToolCalls !== undefined
149
+ ? { parallelToolCalls: baseRequest.parallelToolCalls }
150
+ : {}),
151
+ ...(baseRequest.thinking ? { reasoning: baseRequest.thinking } : {}),
152
+ ...(input.contextLimitTokens !== undefined
153
+ ? { contextLimitTokens: input.contextLimitTokens }
154
+ : {}),
155
+ reservedOutputTokens: input.maxTokens,
156
+ safetyMarginTokens: SAFETY_MARGIN_TOKENS + REPLAY_PLAN_SLACK_TOKENS,
157
+ });
158
+ return { messages, accounting };
159
+ }
160
+ export async function executeCompactionSummary(execution) {
161
+ const attemptMessages = (userPrompt, systemContent) => {
162
+ if (execution.baseRequest) {
163
+ return buildCompactionReplayMessages(execution.baseRequest, execution.history ?? [], userPrompt);
164
+ }
165
+ return execution.sourceMessages
166
+ ? [
167
+ ...execution.sourceMessages,
168
+ { role: "user", content: userPrompt },
169
+ ]
170
+ : [
171
+ { role: "system", content: systemContent },
172
+ { role: "user", content: userPrompt },
173
+ ];
174
+ };
175
+ const baseRequest = execution.baseRequest;
176
+ const request = {
177
+ provider: baseRequest?.provider ?? execution.provider,
178
+ model: baseRequest?.model ?? execution.model,
179
+ purpose: "compaction",
180
+ messages: attemptMessages(execution.prompt, execution.systemContent),
181
+ maxTokens: execution.maxTokens,
182
+ ...(baseRequest
183
+ ? {
184
+ ...(baseRequest.temperature !== undefined
185
+ ? { temperature: baseRequest.temperature }
186
+ : {}),
187
+ ...(baseRequest.thinking
188
+ ? { thinking: structuredClone(baseRequest.thinking) }
189
+ : {}),
190
+ ...(baseRequest.tools
191
+ ? { tools: baseRequest.tools.map((tool) => structuredClone(tool)) }
192
+ : {}),
193
+ ...(baseRequest.toolChoice !== undefined
194
+ ? { toolChoice: structuredClone(baseRequest.toolChoice) }
195
+ : {}),
196
+ ...(baseRequest.parallelToolCalls !== undefined
197
+ ? { parallelToolCalls: baseRequest.parallelToolCalls }
198
+ : {}),
199
+ }
200
+ : {
201
+ temperature: 0.1,
202
+ // Disable reasoning for the compression pass. Use "low" (not "none"):
203
+ // for openai-style gateways (TokenRouter etc.) `effort:"none"` maps to
204
+ // the wire value `reasoning_effort:"none"`, which many models reject
205
+ // ("reasoning option not supported"). "low" with enabled:false maps to
206
+ // no reasoning knob on those gateways while still disabling default
207
+ // thinking on deepseek-style models via their own mapping.
208
+ thinking: { enabled: false, effort: "low" },
209
+ ...(execution.allowModelFallback ? { allowModelFallback: true } : {}),
210
+ ...(execution.tools?.length
211
+ ? { tools: execution.tools, toolChoice: "none" }
212
+ : {}),
213
+ }),
214
+ ...(execution.signal ? { signal: execution.signal } : {}),
215
+ };
216
+ if (baseRequest) {
217
+ const accounting = accountAssembledRequest({
218
+ provider: baseRequest.provider,
219
+ model: baseRequest.model,
220
+ messages: request.messages,
221
+ stream: execution.stream,
222
+ ...(request.tools?.length ? { tools: request.tools } : {}),
223
+ ...(request.toolChoice !== undefined
224
+ ? { toolChoice: request.toolChoice }
225
+ : {}),
226
+ ...(request.parallelToolCalls !== undefined
227
+ ? { parallelToolCalls: request.parallelToolCalls }
228
+ : {}),
229
+ ...(request.thinking ? { reasoning: request.thinking } : {}),
230
+ ...(execution.contextLimitTokens !== undefined
231
+ ? { contextLimitTokens: execution.contextLimitTokens }
232
+ : {}),
233
+ reservedOutputTokens: execution.maxTokens,
234
+ }).accounting;
235
+ if (accounting.overLimit) {
236
+ throw new CompactionOverLimitError(`compaction failed: captured request plus compaction instruction needs about ${accounting.requestTokens.toLocaleString()} input tokens but only ${accounting.limit.effectiveSafeTokens?.toLocaleString()} fit after reserving summary output — original context retained`, accounting.requestTokens, accounting.limit.effectiveSafeTokens);
237
+ }
238
+ }
239
+ const runProviderAttempt = async (attemptRequest, replace = false) => {
240
+ // A pinned single dispatch is what makes the one-admission contract real:
241
+ // without it the router can still rotate keys, fall back to another
242
+ // provider, or retry a capability downgrade, each resending this prompt.
243
+ const routerOptions = {
244
+ maxRetries: 0,
245
+ singleDispatch: true,
246
+ ...(execution.operation ? { operation: execution.operation } : {}),
247
+ };
248
+ if (!execution.stream) {
249
+ return completeWithProvider(attemptRequest, routerOptions);
250
+ }
251
+ if (replace)
252
+ execution.onToken?.("", true);
253
+ const parser = createThinkingStreamParser((text) => execution.onToken?.(text), undefined, { remember: false });
254
+ const result = await streamWithProvider(attemptRequest, (token) => parser.push(token), { onStatus: () => undefined, ...routerOptions });
255
+ parser.finish();
256
+ return result;
257
+ };
258
+ const sleepBeforeRetry = async () => {
259
+ const delayMs = execution.retryDelayMs ?? COMPACTION_ERROR_RETRY_DELAY_MS;
260
+ if (delayMs <= 0)
261
+ return;
262
+ if (execution.signal?.aborted) {
263
+ throw execution.signal.reason ?? new Error("Aborted");
264
+ }
265
+ await new Promise((resolve, reject) => {
266
+ const signal = execution.signal;
267
+ const timer = setTimeout(() => {
268
+ cleanup();
269
+ resolve();
270
+ }, delayMs);
271
+ const onAbort = () => {
272
+ clearTimeout(timer);
273
+ cleanup();
274
+ reject(signal?.reason ?? new Error("Aborted"));
275
+ };
276
+ const cleanup = () => {
277
+ signal?.removeEventListener("abort", onAbort);
278
+ };
279
+ signal?.addEventListener("abort", onAbort, { once: true });
280
+ });
281
+ };
282
+ // One automatic retry after a transient failure: the replayed prefix is
283
+ // cache-hot, so the second admission only re-bills the small fresh tail.
284
+ const runAttempt = async (attemptRequest, replace = false) => {
285
+ try {
286
+ return await runProviderAttempt(attemptRequest, replace);
287
+ }
288
+ catch (error) {
289
+ if (!execution.retryOnServerError ||
290
+ !isCompactionRetryableError(error, execution.signal)) {
291
+ throw error;
292
+ }
293
+ await sleepBeforeRetry();
294
+ return await runProviderAttempt(attemptRequest, replace);
295
+ }
296
+ };
297
+ const first = await runAttempt(request);
298
+ if (first.toolCalls?.length) {
299
+ throw new Error("compaction failed: model returned tool calls instead of a summary — original context retained");
300
+ }
301
+ let visible = normalizeCompactionSummary(stripThinking(first.text).visible);
302
+ let retryReason;
303
+ if (isCompactionCompletionTruncated(first, execution.maxTokens)) {
304
+ retryReason = "truncated";
305
+ }
306
+ else if (!visible) {
307
+ retryReason = "reasoning-only";
308
+ }
309
+ else if (looksLikeTranscriptReplay(visible)) {
310
+ retryReason = "replayed";
311
+ }
312
+ else if (looksLikeIncompleteCompactionSummary(visible)) {
313
+ retryReason = "incomplete";
314
+ }
315
+ if (retryReason) {
316
+ if (execution.qualityRetry === false) {
317
+ throw new Error(FAIL_CLOSED_BY_REASON[retryReason]);
318
+ }
319
+ const retry = await runAttempt({
320
+ ...request,
321
+ messages: attemptMessages(buildCompactionRetryPrompt(execution.prompt, retryReason), `${execution.systemContent}${RETRY_SYSTEM_SUFFIX}`),
322
+ temperature: 0,
323
+ }, true);
324
+ if (retry.toolCalls?.length) {
325
+ throw new Error("compaction failed: model returned tool calls instead of a summary — original context retained");
326
+ }
327
+ if (isCompactionCompletionTruncated(retry, execution.maxTokens)) {
328
+ throw new Error("compaction failed: model hit the summary output limit twice — original context retained");
329
+ }
330
+ visible = normalizeCompactionSummary(stripThinking(retry.text).visible);
331
+ if (!visible) {
332
+ throw new Error("compaction failed: model returned an empty summary");
333
+ }
334
+ if (looksLikeTranscriptReplay(visible)) {
335
+ throw new Error("compaction failed: model replayed the transcript twice — original context retained");
336
+ }
337
+ if (looksLikeIncompleteCompactionSummary(visible)) {
338
+ throw new Error("compaction failed: model returned an incomplete summary twice — original context retained");
339
+ }
340
+ }
341
+ return visible;
342
+ }
343
+ //# sourceMappingURL=compaction-executor.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"compaction-executor.js","sourceRoot":"","sources":["../../src/agent/compaction-executor.ts"],"names":[],"mappings":"AAMA,OAAO,EACL,0BAA0B,EAC1B,+BAA+B,EAC/B,oCAAoC,EACpC,yBAAyB,EACzB,0BAA0B,GAC3B,MAAM,yBAAyB,CAAC;AACjC,OAAO,EAAE,oBAAoB,EAAE,kBAAkB,EAAE,MAAM,kBAAkB,CAAC;AAE5E,OAAO,EAAE,oBAAoB,EAAE,MAAM,2BAA2B,CAAC;AACjE,OAAO,EAAE,0BAA0B,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAC9E,OAAO,EACL,uBAAuB,EACvB,qBAAqB,EACrB,oBAAoB,GAErB,MAAM,yBAAyB,CAAC;AACjC,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AAEnD,MAAM,mBAAmB,GACvB,4GAA4G,CAAC;AAE/G;;;;GAIG;AACH,MAAM,+BAA+B,GAAG,KAAK,CAAC;AAE9C;;;;;GAKG;AACH,MAAM,wBAAwB,GAAG,KAAK,CAAC;AAEvC;;;;;;;;GAQG;AACH,SAAS,0BAA0B,CAAC,KAAc,EAAE,MAAoB;IACtE,IAAI,oBAAoB,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC9C,IAAI,YAAY,CAAC,KAAK,EAAE,MAAM,CAAC;QAAE,OAAO,KAAK,CAAC;IAC9C,IAAI,0BAA0B,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IACpD,IAAI,KAAK,YAAY,qBAAqB;QAAE,OAAO,KAAK,CAAC;IACzD,MAAM,MAAM,GACV,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,QAAQ,IAAI,KAAK;QACrD,CAAC,CAAC,MAAM,CAAE,KAA6B,CAAC,MAAM,CAAC;QAC/C,CAAC,CAAC,CAAC,CAAC;IACR,IAAI,MAAM,KAAK,GAAG,IAAI,MAAM,KAAK,GAAG;QAAE,OAAO,IAAI,CAAC;IAClD,IAAI,MAAM,IAAI,GAAG,IAAI,MAAM,GAAG,GAAG;QAAE,OAAO,KAAK,CAAC;IAChD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,iFAAiF;AACjF,MAAM,OAAO,wBAAyB,SAAQ,KAAK;IAGtC;IACA;IAHX,YACE,OAAe,EACN,aAAqB,EACrB,mBAAuC;QAEhD,KAAK,CAAC,OAAO,CAAC,CAAC;QAHN,kBAAa,GAAb,aAAa,CAAQ;QACrB,wBAAmB,GAAnB,mBAAmB,CAAoB;QAGhD,IAAI,CAAC,IAAI,GAAG,0BAA0B,CAAC;IACzC,CAAC;CACF;AAED,MAAM,UAAU,0BAA0B,CACxC,KAAc;IAEd,OAAO,KAAK,YAAY,wBAAwB,CAAC;AACnD,CAAC;AA0BD,MAAM,qBAAqB,GAAG;IAC5B,SAAS,EAAE,mFAAmF;IAC9F,gBAAgB,EACd,kFAAkF;IACpF,QAAQ,EACN,8EAA8E;IAChF,UAAU,EACR,qFAAqF;CAC/E,CAAC;AAEX,SAAS,iBAAiB,CAAC,OAAoB;IAC7C,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,GAAG,IAAI,EAAE,GAAG,OAAO,CAAC;IAC7C,OAAO,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC;AAC9B,CAAC;AAED,SAAS,kBAAkB,CACzB,eAAuC,EACvC,OAA+B;IAE/B,IAAI,YAAY,GAAG,CAAC,CAAC;IACrB,IAAI,cAAc,GAAG,CAAC,CAAC;IACvB,OAAO,cAAc,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;QACvC,MAAM,QAAQ,GAAG,iBAAiB,CAAC,OAAO,CAAC,cAAc,CAAE,CAAC,CAAC;QAC7D,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC;QACf,KAAK,IAAI,KAAK,GAAG,YAAY,EAAE,KAAK,GAAG,eAAe,CAAC,MAAM,EAAE,KAAK,IAAI,CAAC,EAAE,CAAC;YAC1E,IAAI,iBAAiB,CAAC,eAAe,CAAC,KAAK,CAAE,CAAC,KAAK,QAAQ,EAAE,CAAC;gBAC5D,KAAK,GAAG,KAAK,CAAC;gBACd,MAAM;YACR,CAAC;QACH,CAAC;QACD,IAAI,KAAK,GAAG,CAAC;YAAE,MAAM;QACrB,YAAY,GAAG,KAAK,GAAG,CAAC,CAAC;QACzB,cAAc,IAAI,CAAC,CAAC;IACtB,CAAC;IACD,OAAO,OAAO,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC;AAClF,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,6BAA6B,CAC3C,WAAsC,EACtC,OAA+B,EAC/B,UAAkB;IAElB,OAAO;QACL,GAAG,WAAW,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC;QAClE,GAAG,kBAAkB,CAAC,WAAW,CAAC,QAAQ,EAAE,OAAO,CAAC;QACpD,EAAE,IAAI,EAAE,MAAe,EAAE,OAAO,EAAE,UAAU,EAAE;KAC/C,CAAC;AACJ,CAAC;AAOD;;;;;;;;;;;;;;;GAeG;AACH,MAAM,UAAU,oBAAoB,CAAC,KAOpC;IACC,MAAM,WAAW,GAAG,KAAK,CAAC,WAAW,CAAC;IACtC,MAAM,YAAY,GAAG,WAAW,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;IAC7C,MAAM,WAAW,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;IACrC,IAAI,CAAC,YAAY,IAAI,CAAC,WAAW;QAAE,OAAO,SAAS,CAAC;IACpD,MAAM,IAAI,GAAG,kBAAkB,CAAC,WAAW,CAAC,QAAQ,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IACrE,IAAI,iBAAiB,CAAC,YAAY,CAAC,KAAK,iBAAiB,CAAC,WAAW,CAAC,EAAE,CAAC;QACvE,IAAI,WAAW,CAAC,IAAI,KAAK,QAAQ;YAAE,OAAO,SAAS,CAAC;QACpD,IAAI,IAAI,CAAC,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM;YAAE,OAAO,SAAS,CAAC;IAC5D,CAAC;IACD,MAAM,QAAQ,GAAkB;QAC9B,GAAG,WAAW,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC;QAClE,GAAG,IAAI;QACP,EAAE,IAAI,EAAE,MAAe,EAAE,OAAO,EAAE,KAAK,CAAC,MAAM,EAAE;KACjD,CAAC;IACF,MAAM,EAAE,UAAU,EAAE,GAAG,uBAAuB,CAAC;QAC7C,QAAQ,EAAE,WAAW,CAAC,QAAQ;QAC9B,KAAK,EAAE,WAAW,CAAC,KAAK;QACxB,QAAQ;QACR,MAAM,EAAE,KAAK,CAAC,MAAM,IAAI,IAAI;QAC5B,GAAG,CAAC,WAAW,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,WAAW,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAClE,GAAG,CAAC,WAAW,CAAC,UAAU,KAAK,SAAS;YACtC,CAAC,CAAC,EAAE,UAAU,EAAE,WAAW,CAAC,UAAU,EAAE;YACxC,CAAC,CAAC,EAAE,CAAC;QACP,GAAG,CAAC,WAAW,CAAC,iBAAiB,KAAK,SAAS;YAC7C,CAAC,CAAC,EAAE,iBAAiB,EAAE,WAAW,CAAC,iBAAiB,EAAE;YACtD,CAAC,CAAC,EAAE,CAAC;QACP,GAAG,CAAC,WAAW,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,WAAW,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACpE,GAAG,CAAC,KAAK,CAAC,kBAAkB,KAAK,SAAS;YACxC,CAAC,CAAC,EAAE,kBAAkB,EAAE,KAAK,CAAC,kBAAkB,EAAE;YAClD,CAAC,CAAC,EAAE,CAAC;QACP,oBAAoB,EAAE,KAAK,CAAC,SAAS;QACrC,kBAAkB,EAAE,oBAAoB,GAAG,wBAAwB;KACpE,CAAC,CAAC;IACH,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE,CAAC;AAClC,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,SAAqC;IAErC,MAAM,eAAe,GAAG,CACtB,UAAkB,EAClB,aAAqB,EACN,EAAE;QACjB,IAAI,SAAS,CAAC,WAAW,EAAE,CAAC;YAC1B,OAAO,6BAA6B,CAClC,SAAS,CAAC,WAAW,EACrB,SAAS,CAAC,OAAO,IAAI,EAAE,EACvB,UAAU,CACX,CAAC;QACJ,CAAC;QACD,OAAO,SAAS,CAAC,cAAc;YAC7B,CAAC,CAAC;gBACE,GAAG,SAAS,CAAC,cAAc;gBAC3B,EAAE,IAAI,EAAE,MAAe,EAAE,OAAO,EAAE,UAAU,EAAE;aAC/C;YACH,CAAC,CAAC;gBACE,EAAE,IAAI,EAAE,QAAiB,EAAE,OAAO,EAAE,aAAa,EAAE;gBACnD,EAAE,IAAI,EAAE,MAAe,EAAE,OAAO,EAAE,UAAU,EAAE;aAC/C,CAAC;IACR,CAAC,CAAC;IAEF,MAAM,WAAW,GAAG,SAAS,CAAC,WAAW,CAAC;IAC1C,MAAM,OAAO,GAAsB;QACjC,QAAQ,EAAE,WAAW,EAAE,QAAQ,IAAI,SAAS,CAAC,QAAQ;QACrD,KAAK,EAAE,WAAW,EAAE,KAAK,IAAI,SAAS,CAAC,KAAK;QAC5C,OAAO,EAAE,YAAY;QACrB,QAAQ,EAAE,eAAe,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,aAAa,CAAC;QACpE,SAAS,EAAE,SAAS,CAAC,SAAS;QAC9B,GAAG,CAAC,WAAW;YACb,CAAC,CAAC;gBACE,GAAG,CAAC,WAAW,CAAC,WAAW,KAAK,SAAS;oBACvC,CAAC,CAAC,EAAE,WAAW,EAAE,WAAW,CAAC,WAAW,EAAE;oBAC1C,CAAC,CAAC,EAAE,CAAC;gBACP,GAAG,CAAC,WAAW,CAAC,QAAQ;oBACtB,CAAC,CAAC,EAAE,QAAQ,EAAE,eAAe,CAAC,WAAW,CAAC,QAAQ,CAAC,EAAE;oBACrD,CAAC,CAAC,EAAE,CAAC;gBACP,GAAG,CAAC,WAAW,CAAC,KAAK;oBACnB,CAAC,CAAC,EAAE,KAAK,EAAE,WAAW,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,eAAe,CAAC,IAAI,CAAC,CAAC,EAAE;oBACnE,CAAC,CAAC,EAAE,CAAC;gBACP,GAAG,CAAC,WAAW,CAAC,UAAU,KAAK,SAAS;oBACtC,CAAC,CAAC,EAAE,UAAU,EAAE,eAAe,CAAC,WAAW,CAAC,UAAU,CAAC,EAAE;oBACzD,CAAC,CAAC,EAAE,CAAC;gBACP,GAAG,CAAC,WAAW,CAAC,iBAAiB,KAAK,SAAS;oBAC7C,CAAC,CAAC,EAAE,iBAAiB,EAAE,WAAW,CAAC,iBAAiB,EAAE;oBACtD,CAAC,CAAC,EAAE,CAAC;aACR;YACH,CAAC,CAAC;gBACE,WAAW,EAAE,GAAG;gBAChB,sEAAsE;gBACtE,uEAAuE;gBACvE,qEAAqE;gBACrE,uEAAuE;gBACvE,oEAAoE;gBACpE,2DAA2D;gBAC3D,QAAQ,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE,MAAM,EAAE,KAAc,EAAE;gBACpD,GAAG,CAAC,SAAS,CAAC,kBAAkB,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACrE,GAAG,CAAC,SAAS,CAAC,KAAK,EAAE,MAAM;oBACzB,CAAC,CAAC,EAAE,KAAK,EAAE,SAAS,CAAC,KAAK,EAAE,UAAU,EAAE,MAAe,EAAE;oBACzD,CAAC,CAAC,EAAE,CAAC;aACR,CAAC;QACN,GAAG,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,SAAS,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC1D,CAAC;IAEF,IAAI,WAAW,EAAE,CAAC;QAChB,MAAM,UAAU,GAAG,uBAAuB,CAAC;YACzC,QAAQ,EAAE,WAAW,CAAC,QAAQ;YAC9B,KAAK,EAAE,WAAW,CAAC,KAAK;YACxB,QAAQ,EAAE,OAAO,CAAC,QAAQ;YAC1B,MAAM,EAAE,SAAS,CAAC,MAAM;YACxB,GAAG,CAAC,OAAO,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC1D,GAAG,CAAC,OAAO,CAAC,UAAU,KAAK,SAAS;gBAClC,CAAC,CAAC,EAAE,UAAU,EAAE,OAAO,CAAC,UAAU,EAAE;gBACpC,CAAC,CAAC,EAAE,CAAC;YACP,GAAG,CAAC,OAAO,CAAC,iBAAiB,KAAK,SAAS;gBACzC,CAAC,CAAC,EAAE,iBAAiB,EAAE,OAAO,CAAC,iBAAiB,EAAE;gBAClD,CAAC,CAAC,EAAE,CAAC;YACP,GAAG,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,OAAO,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC5D,GAAG,CAAC,SAAS,CAAC,kBAAkB,KAAK,SAAS;gBAC5C,CAAC,CAAC,EAAE,kBAAkB,EAAE,SAAS,CAAC,kBAAkB,EAAE;gBACtD,CAAC,CAAC,EAAE,CAAC;YACP,oBAAoB,EAAE,SAAS,CAAC,SAAS;SAC1C,CAAC,CAAC,UAAU,CAAC;QACd,IAAI,UAAU,CAAC,SAAS,EAAE,CAAC;YACzB,MAAM,IAAI,wBAAwB,CAChC,+EAA+E,UAAU,CAAC,aAAa,CAAC,cAAc,EAAE,0BAA0B,UAAU,CAAC,KAAK,CAAC,mBAAmB,EAAE,cAAc,EAAE,iEAAiE,EACzQ,UAAU,CAAC,aAAa,EACxB,UAAU,CAAC,KAAK,CAAC,mBAAmB,CACrC,CAAC;QACJ,CAAC;IACH,CAAC;IAED,MAAM,kBAAkB,GAAG,KAAK,EAC9B,cAAiC,EACjC,OAAO,GAAG,KAAK,EACf,EAAE;QACF,0EAA0E;QAC1E,oEAAoE;QACpE,yEAAyE;QACzE,MAAM,aAAa,GAAG;YACpB,UAAU,EAAE,CAAC;YACb,cAAc,EAAE,IAAI;YACpB,GAAG,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,SAAS,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACnE,CAAC;QACF,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE,CAAC;YACtB,OAAO,oBAAoB,CAAC,cAAc,EAAE,aAAa,CAAC,CAAC;QAC7D,CAAC;QACD,IAAI,OAAO;YAAE,SAAS,CAAC,OAAO,EAAE,CAAC,EAAE,EAAE,IAAI,CAAC,CAAC;QAC3C,MAAM,MAAM,GAAG,0BAA0B,CACvC,CAAC,IAAI,EAAE,EAAE,CAAC,SAAS,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,EACnC,SAAS,EACT,EAAE,QAAQ,EAAE,KAAK,EAAE,CACpB,CAAC;QACF,MAAM,MAAM,GAAG,MAAM,kBAAkB,CACrC,cAAc,EACd,CAAC,KAAK,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,EAC7B,EAAE,QAAQ,EAAE,GAAG,EAAE,CAAC,SAAS,EAAE,GAAG,aAAa,EAAE,CAChD,CAAC;QACF,MAAM,CAAC,MAAM,EAAE,CAAC;QAChB,OAAO,MAAM,CAAC;IAChB,CAAC,CAAC;IAEF,MAAM,gBAAgB,GAAG,KAAK,IAAmB,EAAE;QACjD,MAAM,OAAO,GAAG,SAAS,CAAC,YAAY,IAAI,+BAA+B,CAAC;QAC1E,IAAI,OAAO,IAAI,CAAC;YAAE,OAAO;QACzB,IAAI,SAAS,CAAC,MAAM,EAAE,OAAO,EAAE,CAAC;YAC9B,MAAM,SAAS,CAAC,MAAM,CAAC,MAAM,IAAI,IAAI,KAAK,CAAC,SAAS,CAAC,CAAC;QACxD,CAAC;QACD,MAAM,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;YAC1C,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC;YAChC,MAAM,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE;gBAC5B,OAAO,EAAE,CAAC;gBACV,OAAO,EAAE,CAAC;YACZ,CAAC,EAAE,OAAO,CAAC,CAAC;YACZ,MAAM,OAAO,GAAG,GAAS,EAAE;gBACzB,YAAY,CAAC,KAAK,CAAC,CAAC;gBACpB,OAAO,EAAE,CAAC;gBACV,MAAM,CAAC,MAAM,EAAE,MAAM,IAAI,IAAI,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC;YACjD,CAAC,CAAC;YACF,MAAM,OAAO,GAAG,GAAS,EAAE;gBACzB,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAChD,CAAC,CAAC;YACF,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;QAC7D,CAAC,CAAC,CAAC;IACL,CAAC,CAAC;IAEF,wEAAwE;IACxE,yEAAyE;IACzE,MAAM,UAAU,GAAG,KAAK,EACtB,cAAiC,EACjC,OAAO,GAAG,KAAK,EACf,EAAE;QACF,IAAI,CAAC;YACH,OAAO,MAAM,kBAAkB,CAAC,cAAc,EAAE,OAAO,CAAC,CAAC;QAC3D,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,IACE,CAAC,SAAS,CAAC,kBAAkB;gBAC7B,CAAC,0BAA0B,CAAC,KAAK,EAAE,SAAS,CAAC,MAAM,CAAC,EACpD,CAAC;gBACD,MAAM,KAAK,CAAC;YACd,CAAC;YACD,MAAM,gBAAgB,EAAE,CAAC;YACzB,OAAO,MAAM,kBAAkB,CAAC,cAAc,EAAE,OAAO,CAAC,CAAC;QAC3D,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,KAAK,GAAG,MAAM,UAAU,CAAC,OAAO,CAAC,CAAC;IACxC,IAAI,KAAK,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC;QAC5B,MAAM,IAAI,KAAK,CACb,+FAA+F,CAChG,CAAC;IACJ,CAAC;IACD,IAAI,OAAO,GAAG,0BAA0B,CACtC,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,OAAO,CAClC,CAAC;IACF,IAAI,WAKS,CAAC;IACd,IAAI,+BAA+B,CAAC,KAAK,EAAE,SAAS,CAAC,SAAS,CAAC,EAAE,CAAC;QAChE,WAAW,GAAG,WAAW,CAAC;IAC5B,CAAC;SAAM,IAAI,CAAC,OAAO,EAAE,CAAC;QACpB,WAAW,GAAG,gBAAgB,CAAC;IACjC,CAAC;SAAM,IAAI,yBAAyB,CAAC,OAAO,CAAC,EAAE,CAAC;QAC9C,WAAW,GAAG,UAAU,CAAC;IAC3B,CAAC;SAAM,IAAI,oCAAoC,CAAC,OAAO,CAAC,EAAE,CAAC;QACzD,WAAW,GAAG,YAAY,CAAC;IAC7B,CAAC;IAED,IAAI,WAAW,EAAE,CAAC;QAChB,IAAI,SAAS,CAAC,YAAY,KAAK,KAAK,EAAE,CAAC;YACrC,MAAM,IAAI,KAAK,CAAC,qBAAqB,CAAC,WAAW,CAAC,CAAC,CAAC;QACtD,CAAC;QACD,MAAM,KAAK,GAAG,MAAM,UAAU,CAC5B;YACE,GAAG,OAAO;YACV,QAAQ,EAAE,eAAe,CACvB,0BAA0B,CAAC,SAAS,CAAC,MAAM,EAAE,WAAW,CAAC,EACzD,GAAG,SAAS,CAAC,aAAa,GAAG,mBAAmB,EAAE,CACnD;YACD,WAAW,EAAE,CAAC;SACf,EACD,IAAI,CACL,CAAC;QACF,IAAI,KAAK,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC;YAC5B,MAAM,IAAI,KAAK,CACb,+FAA+F,CAChG,CAAC;QACJ,CAAC;QACD,IAAI,+BAA+B,CAAC,KAAK,EAAE,SAAS,CAAC,SAAS,CAAC,EAAE,CAAC;YAChE,MAAM,IAAI,KAAK,CACb,yFAAyF,CAC1F,CAAC;QACJ,CAAC;QACD,OAAO,GAAG,0BAA0B,CAClC,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,OAAO,CAClC,CAAC;QACF,IAAI,CAAC,OAAO,EAAE,CAAC;YACb,MAAM,IAAI,KAAK,CAAC,oDAAoD,CAAC,CAAC;QACxE,CAAC;QACD,IAAI,yBAAyB,CAAC,OAAO,CAAC,EAAE,CAAC;YACvC,MAAM,IAAI,KAAK,CACb,oFAAoF,CACrF,CAAC;QACJ,CAAC;QACD,IAAI,oCAAoC,CAAC,OAAO,CAAC,EAAE,CAAC;YAClD,MAAM,IAAI,KAAK,CACb,2FAA2F,CAC5F,CAAC;QACJ,CAAC;IACH,CAAC;IAED,OAAO,OAAO,CAAC;AACjB,CAAC"}
@@ -5,6 +5,7 @@
5
5
  * Used for reliability ops and regression harnesses. Does not truncate context.
6
6
  */
7
7
  import { createHash } from "node:crypto";
8
+ import { reasoningArtifactTokensForMessage } from "../llm/reasoning-artifacts.js";
8
9
  import { estimateMessagesTokens, estimateTokens, isCompactionMemoryMessage, COMPACTION_MEMORY_PREFIX, PLAN_IMPLEMENT_MEMORY_PREFIX, MECHANICAL_MEMORY_PREFIX, } from "./context-manager.js";
9
10
  import { SESSION_STATE_PREFIX } from "./session-state.js";
10
11
  function isPlanSystem(content) {
@@ -81,6 +82,7 @@ export function buildContextBreakdown(messages, tools) {
81
82
  if (message.toolCalls?.length) {
82
83
  assistantTokens += estimateTokens(JSON.stringify(message.toolCalls));
83
84
  }
85
+ assistantTokens += reasoningArtifactTokensForMessage(message);
84
86
  break;
85
87
  case "tool":
86
88
  toolMessageCount += 1;
@@ -1 +1 @@
1
- {"version":3,"file":"context-breakdown.js","sourceRoot":"","sources":["../../src/agent/context-breakdown.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAEzC,OAAO,EACL,sBAAsB,EACtB,cAAc,EACd,yBAAyB,EACzB,wBAAwB,EACxB,4BAA4B,EAC5B,wBAAwB,GACzB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AA2B1D,SAAS,YAAY,CAAC,OAAe;IACnC,OAAO,OAAO,CAAC,UAAU,CAAC,aAAa,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;AAChF,CAAC;AAED,SAAS,aAAa,CAAC,OAAe;IACpC,OAAO,CACL,OAAO,CAAC,UAAU,CAAC,kBAAkB,CAAC;QACtC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CACvC,CAAC;AACJ,CAAC;AAED,SAAS,oBAAoB,CAAC,OAAe;IAC3C,OAAO,CACL,OAAO,CAAC,UAAU,CAAC,oBAAoB,CAAC;QACxC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CACvC,CAAC;AACJ,CAAC;AAED,SAAS,kBAAkB,CAAC,OAAe;IACzC,OAAO,CACL,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC;QAC5C,OAAO,CAAC,UAAU,CAAC,4BAA4B,CAAC;QAChD,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC;QAC5C,6CAA6C;QAC7C,yBAAyB,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,CAAC,CACvD,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,qBAAqB,CACnC,QAAgC,EAChC,KAA6C;IAE7C,IAAI,YAAY,GAAG,CAAC,CAAC;IACrB,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,eAAe,GAAG,CAAC,CAAC;IACxB,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,IAAI,qBAAqB,GAAG,CAAC,CAAC;IAC9B,IAAI,gBAAgB,GAAG,CAAC,CAAC;IAEzB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,WAAW,GAAG,CAAC,CAAC;IACpB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,sBAAsB,GAAG,CAAC,CAAC;IAC/B,IAAI,iBAAiB,GAAG,CAAC,CAAC;IAE1B,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,MAAM,MAAM,GAAG,cAAc,CAAC,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;QACnD,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;YACrB,KAAK,QAAQ,CAAC,CAAC,CAAC;gBACd,kBAAkB,IAAI,CAAC,CAAC;gBACxB,YAAY,IAAI,MAAM,CAAC;gBACvB,IAAI,kBAAkB,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBACxC,sBAAsB,IAAI,MAAM,CAAC;gBACnC,CAAC;qBAAM,IAAI,YAAY,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBACzC,UAAU,IAAI,MAAM,CAAC;gBACvB,CAAC;qBAAM,IAAI,aAAa,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBAC1C,WAAW,IAAI,MAAM,CAAC;gBACxB,CAAC;qBAAM,IAAI,oBAAoB,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBACjD,kBAAkB,IAAI,MAAM,CAAC;gBAC/B,CAAC;qBAAM,IACL,kBAAkB,KAAK,CAAC;oBACxB,kDAAkD,CAAC,IAAI,CACrD,OAAO,CAAC,OAAO,CAChB,EACD,CAAC;oBACD,6DAA6D;oBAC7D,kBAAkB,IAAI,MAAM,CAAC;gBAC/B,CAAC;qBAAM,CAAC;oBACN,iBAAiB,IAAI,MAAM,CAAC;gBAC9B,CAAC;gBACD,MAAM;YACR,CAAC;YACD,KAAK,MAAM;gBACT,gBAAgB,IAAI,CAAC,CAAC;gBACtB,UAAU,IAAI,MAAM,CAAC;gBACrB,MAAM;YACR,KAAK,WAAW;gBACd,qBAAqB,IAAI,CAAC,CAAC;gBAC3B,eAAe,IAAI,MAAM,CAAC;gBAC1B,IAAI,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC;oBAC9B,eAAe,IAAI,cAAc,CAC/B,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,SAAS,CAAC,CAClC,CAAC;gBACJ,CAAC;gBACD,MAAM;YACR,KAAK,MAAM;gBACT,gBAAgB,IAAI,CAAC,CAAC;gBACtB,gBAAgB,IAAI,MAAM,CAAC;gBAC3B,MAAM;YACR;gBACE,iBAAiB,IAAI,MAAM,CAAC;gBAC5B,MAAM;QACV,CAAC;IACH,CAAC;IAED,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,MAAM,mBAAmB,GAAG,KAAK,EAAE,MAAM,IAAI,CAAC,CAAC;IAC/C,IAAI,KAAK,EAAE,MAAM,EAAE,CAAC;QAClB,gBAAgB,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,CAAC;IAC3D,CAAC;IAED,MAAM,oBAAoB,GACxB,sBAAsB,CAAC,CAAC,GAAG,QAAQ,CAAC,CAAC,GAAG,gBAAgB,CAAC;IAE3D,OAAO;QACL,oBAAoB;QACpB,YAAY;QACZ,UAAU;QACV,eAAe;QACf,gBAAgB;QAChB,gBAAgB;QAChB,YAAY,EAAE,QAAQ,CAAC,MAAM;QAC7B,kBAAkB;QAClB,gBAAgB;QAChB,qBAAqB;QACrB,gBAAgB;QAChB,WAAW,EAAE;YACX,kBAAkB;YAClB,UAAU;YACV,WAAW;YACX,kBAAkB;YAClB,sBAAsB;YACtB,iBAAiB;SAClB;QACD,mBAAmB;KACpB,CAAC;AACJ,CAAC;AAED,0EAA0E;AAC1E,MAAM,UAAU,4BAA4B,CAC1C,SAA2B;IAE3B,OAAO;QACL,oBAAoB,EAAE,SAAS,CAAC,oBAAoB;QACpD,YAAY,EAAE,SAAS,CAAC,YAAY;QACpC,UAAU,EAAE,SAAS,CAAC,UAAU;QAChC,eAAe,EAAE,SAAS,CAAC,eAAe;QAC1C,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,YAAY,EAAE,SAAS,CAAC,YAAY;QACpC,kBAAkB,EAAE,SAAS,CAAC,kBAAkB;QAChD,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,qBAAqB,EAAE,SAAS,CAAC,qBAAqB;QACtD,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,kBAAkB,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;QAC5D,UAAU,EAAE,SAAS,CAAC,WAAW,CAAC,UAAU;QAC5C,WAAW,EAAE,SAAS,CAAC,WAAW,CAAC,WAAW;QAC9C,kBAAkB,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;QAC5D,sBAAsB,EAAE,SAAS,CAAC,WAAW,CAAC,sBAAsB;QACpE,iBAAiB,EAAE,SAAS,CAAC,WAAW,CAAC,iBAAiB;QAC1D,mBAAmB,EAAE,SAAS,CAAC,mBAAmB;KACnD,CAAC;AACJ,CAAC;AAGD,+EAA+E;AAC/E,+EAA+E;AAC/E,uCAAuC;AACvC,MAAM,UAAU,4BAA4B,CAC1C,QAAgC,EAChC,KAA6C;IAE7C,MAAM,SAAS,GAAG,qBAAqB,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;IACzD,MAAM,UAAU,GAA6C;QAC3D,EAAE,KAAK,EAAE,cAAc,EAAE,MAAM,EAAE,SAAS,CAAC,gBAAgB,EAAE;QAC7D,EAAE,KAAK,EAAE,cAAc,EAAE,MAAM,EAAE,SAAS,CAAC,gBAAgB,EAAE;QAC7D,EAAE,KAAK,EAAE,mBAAmB,EAAE,MAAM,EAAE,SAAS,CAAC,eAAe,EAAE;QACjE,EAAE,KAAK,EAAE,eAAe,EAAE,MAAM,EAAE,SAAS,CAAC,UAAU,EAAE;QACxD;YACE,KAAK,EAAE,qBAAqB;YAC5B,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;SACjD;QACD;YACE,KAAK,EAAE,mBAAmB;YAC1B,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,sBAAsB;SACrD;QACD,EAAE,KAAK,EAAE,YAAY,EAAE,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,UAAU,EAAE;QACjE;YACE,KAAK,EAAE,eAAe;YACtB,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;SACjD;KACF,CAAC;IACF,MAAM,QAAQ,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,SAAS,EAAE,EAAE,CACxD,SAAS,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,CACxD,CAAC;IACF,OAAO,GAAG,QAAQ,CAAC,KAAK,MAAM,QAAQ,CAAC,MAAM,CAAC,cAAc,EAAE,UAAU,CAAC;AAC3E,CAAC;AAED,kEAAkE;AAClE,MAAM,UAAU,cAAc,CAC5B,KAA4C;IAE5C,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC;IAChD,OAAO,UAAU,CAAC,QAAQ,CAAC;SACxB,MAAM,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,IAAI,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;SACxG,MAAM,CAAC,KAAK,CAAC;SACb,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;AAClB,CAAC"}
1
+ {"version":3,"file":"context-breakdown.js","sourceRoot":"","sources":["../../src/agent/context-breakdown.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAEzC,OAAO,EAAE,iCAAiC,EAAE,MAAM,+BAA+B,CAAC;AAClF,OAAO,EACL,sBAAsB,EACtB,cAAc,EACd,yBAAyB,EACzB,wBAAwB,EACxB,4BAA4B,EAC5B,wBAAwB,GACzB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AA2B1D,SAAS,YAAY,CAAC,OAAe;IACnC,OAAO,OAAO,CAAC,UAAU,CAAC,aAAa,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;AAChF,CAAC;AAED,SAAS,aAAa,CAAC,OAAe;IACpC,OAAO,CACL,OAAO,CAAC,UAAU,CAAC,kBAAkB,CAAC;QACtC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CACvC,CAAC;AACJ,CAAC;AAED,SAAS,oBAAoB,CAAC,OAAe;IAC3C,OAAO,CACL,OAAO,CAAC,UAAU,CAAC,oBAAoB,CAAC;QACxC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CACvC,CAAC;AACJ,CAAC;AAED,SAAS,kBAAkB,CAAC,OAAe;IACzC,OAAO,CACL,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC;QAC5C,OAAO,CAAC,UAAU,CAAC,4BAA4B,CAAC;QAChD,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC;QAC5C,6CAA6C;QAC7C,yBAAyB,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,CAAC,CACvD,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,qBAAqB,CACnC,QAAgC,EAChC,KAA6C;IAE7C,IAAI,YAAY,GAAG,CAAC,CAAC;IACrB,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,eAAe,GAAG,CAAC,CAAC;IACxB,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,IAAI,qBAAqB,GAAG,CAAC,CAAC;IAC9B,IAAI,gBAAgB,GAAG,CAAC,CAAC;IAEzB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,WAAW,GAAG,CAAC,CAAC;IACpB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,sBAAsB,GAAG,CAAC,CAAC;IAC/B,IAAI,iBAAiB,GAAG,CAAC,CAAC;IAE1B,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,MAAM,MAAM,GAAG,cAAc,CAAC,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;QACnD,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;YACrB,KAAK,QAAQ,CAAC,CAAC,CAAC;gBACd,kBAAkB,IAAI,CAAC,CAAC;gBACxB,YAAY,IAAI,MAAM,CAAC;gBACvB,IAAI,kBAAkB,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBACxC,sBAAsB,IAAI,MAAM,CAAC;gBACnC,CAAC;qBAAM,IAAI,YAAY,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBACzC,UAAU,IAAI,MAAM,CAAC;gBACvB,CAAC;qBAAM,IAAI,aAAa,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBAC1C,WAAW,IAAI,MAAM,CAAC;gBACxB,CAAC;qBAAM,IAAI,oBAAoB,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;oBACjD,kBAAkB,IAAI,MAAM,CAAC;gBAC/B,CAAC;qBAAM,IACL,kBAAkB,KAAK,CAAC;oBACxB,kDAAkD,CAAC,IAAI,CACrD,OAAO,CAAC,OAAO,CAChB,EACD,CAAC;oBACD,6DAA6D;oBAC7D,kBAAkB,IAAI,MAAM,CAAC;gBAC/B,CAAC;qBAAM,CAAC;oBACN,iBAAiB,IAAI,MAAM,CAAC;gBAC9B,CAAC;gBACD,MAAM;YACR,CAAC;YACD,KAAK,MAAM;gBACT,gBAAgB,IAAI,CAAC,CAAC;gBACtB,UAAU,IAAI,MAAM,CAAC;gBACrB,MAAM;YACR,KAAK,WAAW;gBACd,qBAAqB,IAAI,CAAC,CAAC;gBAC3B,eAAe,IAAI,MAAM,CAAC;gBAC1B,IAAI,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC;oBAC9B,eAAe,IAAI,cAAc,CAC/B,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,SAAS,CAAC,CAClC,CAAC;gBACJ,CAAC;gBACD,eAAe,IAAI,iCAAiC,CAAC,OAAO,CAAC,CAAC;gBAC9D,MAAM;YACR,KAAK,MAAM;gBACT,gBAAgB,IAAI,CAAC,CAAC;gBACtB,gBAAgB,IAAI,MAAM,CAAC;gBAC3B,MAAM;YACR;gBACE,iBAAiB,IAAI,MAAM,CAAC;gBAC5B,MAAM;QACV,CAAC;IACH,CAAC;IAED,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,MAAM,mBAAmB,GAAG,KAAK,EAAE,MAAM,IAAI,CAAC,CAAC;IAC/C,IAAI,KAAK,EAAE,MAAM,EAAE,CAAC;QAClB,gBAAgB,GAAG,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,CAAC;IAC3D,CAAC;IAED,MAAM,oBAAoB,GACxB,sBAAsB,CAAC,CAAC,GAAG,QAAQ,CAAC,CAAC,GAAG,gBAAgB,CAAC;IAE3D,OAAO;QACL,oBAAoB;QACpB,YAAY;QACZ,UAAU;QACV,eAAe;QACf,gBAAgB;QAChB,gBAAgB;QAChB,YAAY,EAAE,QAAQ,CAAC,MAAM;QAC7B,kBAAkB;QAClB,gBAAgB;QAChB,qBAAqB;QACrB,gBAAgB;QAChB,WAAW,EAAE;YACX,kBAAkB;YAClB,UAAU;YACV,WAAW;YACX,kBAAkB;YAClB,sBAAsB;YACtB,iBAAiB;SAClB;QACD,mBAAmB;KACpB,CAAC;AACJ,CAAC;AAED,0EAA0E;AAC1E,MAAM,UAAU,4BAA4B,CAC1C,SAA2B;IAE3B,OAAO;QACL,oBAAoB,EAAE,SAAS,CAAC,oBAAoB;QACpD,YAAY,EAAE,SAAS,CAAC,YAAY;QACpC,UAAU,EAAE,SAAS,CAAC,UAAU;QAChC,eAAe,EAAE,SAAS,CAAC,eAAe;QAC1C,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,YAAY,EAAE,SAAS,CAAC,YAAY;QACpC,kBAAkB,EAAE,SAAS,CAAC,kBAAkB;QAChD,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,qBAAqB,EAAE,SAAS,CAAC,qBAAqB;QACtD,gBAAgB,EAAE,SAAS,CAAC,gBAAgB;QAC5C,kBAAkB,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;QAC5D,UAAU,EAAE,SAAS,CAAC,WAAW,CAAC,UAAU;QAC5C,WAAW,EAAE,SAAS,CAAC,WAAW,CAAC,WAAW;QAC9C,kBAAkB,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;QAC5D,sBAAsB,EAAE,SAAS,CAAC,WAAW,CAAC,sBAAsB;QACpE,iBAAiB,EAAE,SAAS,CAAC,WAAW,CAAC,iBAAiB;QAC1D,mBAAmB,EAAE,SAAS,CAAC,mBAAmB;KACnD,CAAC;AACJ,CAAC;AAGD,+EAA+E;AAC/E,+EAA+E;AAC/E,uCAAuC;AACvC,MAAM,UAAU,4BAA4B,CAC1C,QAAgC,EAChC,KAA6C;IAE7C,MAAM,SAAS,GAAG,qBAAqB,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;IACzD,MAAM,UAAU,GAA6C;QAC3D,EAAE,KAAK,EAAE,cAAc,EAAE,MAAM,EAAE,SAAS,CAAC,gBAAgB,EAAE;QAC7D,EAAE,KAAK,EAAE,cAAc,EAAE,MAAM,EAAE,SAAS,CAAC,gBAAgB,EAAE;QAC7D,EAAE,KAAK,EAAE,mBAAmB,EAAE,MAAM,EAAE,SAAS,CAAC,eAAe,EAAE;QACjE,EAAE,KAAK,EAAE,eAAe,EAAE,MAAM,EAAE,SAAS,CAAC,UAAU,EAAE;QACxD;YACE,KAAK,EAAE,qBAAqB;YAC5B,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;SACjD;QACD;YACE,KAAK,EAAE,mBAAmB;YAC1B,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,sBAAsB;SACrD;QACD,EAAE,KAAK,EAAE,YAAY,EAAE,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,UAAU,EAAE;QACjE;YACE,KAAK,EAAE,eAAe;YACtB,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,kBAAkB;SACjD;KACF,CAAC;IACF,MAAM,QAAQ,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,SAAS,EAAE,EAAE,CACxD,SAAS,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,CACxD,CAAC;IACF,OAAO,GAAG,QAAQ,CAAC,KAAK,MAAM,QAAQ,CAAC,MAAM,CAAC,cAAc,EAAE,UAAU,CAAC;AAC3E,CAAC;AAED,kEAAkE;AAClE,MAAM,UAAU,cAAc,CAC5B,KAA4C;IAE5C,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC;IAChD,OAAO,UAAU,CAAC,QAAQ,CAAC;SACxB,MAAM,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,IAAI,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;SACxG,MAAM,CAAC,KAAK,CAAC;SACb,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;AAClB,CAAC"}
@@ -1,14 +1,11 @@
1
- import type { ChatImage, ChatMessage } from "../types.js";
1
+ import type { ChatMessage } from "../types.js";
2
+ import { estimateImageTokens, estimateMessagesTokens, estimateTextTokens as estimateTokens } from "./request-accounting.js";
3
+ export { estimateImageTokens, estimateMessagesTokens, estimateTokens };
2
4
  /**
3
- * Per-char token estimator. Real tokenization varies by provider, but for
4
- * budgeting a chars/3.3 heuristic is close enough for mixed text/code/JSON
5
- * (which tokenizes less efficiently than pure English prose). We
6
- * deliberately over-estimate — better to compact one turn too early than to
7
- * lose state to a provider context-window error.
5
+ * Token estimation lives in `request-accounting.ts` the one serialized-
6
+ * request accounting service. The exports above keep this module's historic
7
+ * public surface; nothing here owns a chars-per-token ratio anymore.
8
8
  */
9
- export declare function estimateTokens(text: string): number;
10
- export declare function estimateImageTokens(image: ChatImage): number;
11
- export declare function estimateMessagesTokens(messages: ChatMessage[]): number;
12
9
  /**
13
10
  * Agent-loop auto-compact threshold (estimated tokens). Shared with `/context`
14
11
  * so the reported % of budget matches when auto-compaction fires.
@@ -40,7 +37,23 @@ export interface CompactOptions {
40
37
  */
41
38
  purpose?: "default" | "plan-implement" | undefined;
42
39
  durableEnvelope?: string | undefined;
40
+ /**
41
+ * At most one summarizer dispatch: forbids automatic map/reduce. When the
42
+ * full settled range does not fit one pass, summarize the oldest complete
43
+ * message-aligned slice (emergency_prefix_slice) and retain the untouched
44
+ * middle/recent tail.
45
+ */
46
+ singleAdmission?: boolean | undefined;
47
+ /**
48
+ * Choose the direct single pass whenever source messages exist, skipping the
49
+ * estimate gate. Callers set this only after pre-flighting that the exact
50
+ * assembled request fits (e.g. a cache-preserving snapshot replay planned
51
+ * with `planCompactionReplay`) — the raw estimate gate is deliberately
52
+ * conservative and would otherwise reject requests that fit fine.
53
+ */
54
+ forceDirectSinglePass?: boolean | undefined;
43
55
  }
56
+ export type CompactionStrategy = "direct" | "single" | "emergency_prefix_slice" | "map_reduce";
44
57
  export interface CompactResult {
45
58
  messages: ChatMessage[];
46
59
  before: number;
@@ -48,6 +61,7 @@ export interface CompactResult {
48
61
  beforeTokens: number;
49
62
  afterTokens: number;
50
63
  summarized: boolean;
64
+ strategy?: CompactionStrategy | undefined;
51
65
  }
52
66
  /**
53
67
  * Content prefixes that mark a `role:"system"` message as compacted session