@superblocksteam/vite-plugin-file-sync 1.99.99-next.3 → 2.0.0-SNAPSHOT.2143e2e95d0107ab

package/dist/ai-service/llm/context-v2/token-tracker.d.ts

@@ -0,0 +1,88 @@
+import type { LanguageModelUsage, ModelMessage } from "ai";
+import type { StoredMessage } from "./types.js";
+/**
+ * Either shape the estimator accepts: a wire `ModelMessage` from
+ * `Context.getMessages()`, or a raw `StoredMessage` whose `_superblocks`
+ * metadata is dropped because `toSdkMessages` strips it before the request.
+ */
+type EstimatableMessage = ModelMessage | StoredMessage;
+type UsageLike = {
+    [K in "inputTokens" | "outputTokens" | "totalTokens" | "cachedInputTokens"]?: LanguageModelUsage[K];
+} & {
+    inputTokenDetails?: Partial<NonNullable<LanguageModelUsage["inputTokenDetails"]>>;
+};
+export declare function getTotalInputTokens(usage: UsageLike): number;
+/**
+ * Tokens consumed against the model's input context window, for display in the
+ * UI's context-usage indicator. This is the sum of fresh, cache-read, and
+ * cache-write input tokens — cache reads/writes occupy the input window just
+ * like fresh input (cache is a cost discount, not a context-window discount).
+ *
+ * Excludes outputTokens and reasoningTokens: those do not consume the input
+ * window. (Note: for Anthropic, extended-thinking tokens are bundled inside
+ * `output_tokens` in the raw API response and are not separately surfaced as
+ * `reasoningTokens` by @ai-sdk/anthropic, so this function's exclusion of
+ * reasoning tokens is defensive for other providers.)
+ */
+export declare function computeContextUsed(usage: UsageLike): number;
+/**
+ * Total billed tokens: input (fresh + cache read + cache write) plus output.
+ *
+ * Do NOT use this for context-window or compaction-threshold decisions — use
+ * `computeContextUsed` instead. Output/reasoning tokens do not occupy the
+ * model's input window, and Anthropic bundles extended-thinking tokens into
+ * `output_tokens` (often 50–100k+). Counting them here inflates the measured
+ * context so compaction keeps re-triggering on the same turn — the
+ * output-inclusive compaction loop. This helper has no context/compaction
+ * callers; it is retained as a correct "total billed tokens" measure (input +
+ * output) for future billing/telemetry use, intentionally not wired into any
+ * context/compaction path.
+ *
+ * @deprecated Not for context-window or compaction decisions — use
+ * `computeContextUsed` instead. Retained only as a billing/telemetry
+ * "total billed tokens" measure.
+ */
+export declare function getTotalTokens(usage: LanguageModelUsage): number;
+/**
+ * Estimate the input tokens a message array will consume, from its content
+ * alone.
+ *
+ * Every other measure in this file needs provider-reported usage, which does
+ * not exist before the first request of a restored session — and a context
+ * restored from storage can already be too large to send (ENG-5282). Only a
+ * content-based estimate can catch that pre-flight.
+ *
+ * `_superblocks` metadata is excluded: `toSdkMessages` strips it before the
+ * wire payload is built, so counting it would overstate the request.
+ *
+ * Pass the array that is actually going on the wire, i.e. the output of
+ * `Context.getMessages()`, which already carries the system prompt and has had
+ * `_superblocks` stripped. Passing raw `StoredMessage`s instead omits the
+ * system prompt and counts a slightly different shape than the request.
+ *
+ * KNOWN GAP. The tool-definition schemas are passed to the provider alongside
+ * the messages and are not part of this array, so a session with a large tool
+ * surface can estimate under the threshold while the real request exceeds the
+ * provider limit. That case degrades to the provider rejecting it, reported as
+ * `context_too_large` rather than a generic provider error, so it stays
+ * diagnosable.
+ *
+ * Accuracy is coarse: it gates a compaction decision and the provider remains
+ * the authority on the real limit. The direction of the error is content
+ * shaped, not fixed. JSON escaping and key names add roughly a third on top of
+ * the content, so prose is over-counted while code, which real tokenizers
+ * charge closer to 3 characters per token, comes out about right.
+ */
+export declare function estimateMessagesTokens(messages: readonly EstimatableMessage[]): number;
+/**
+ * Estimate the input tokens an arbitrary serializable value will consume, using
+ * the same ratios as {@link estimateMessagesTokens}.
+ *
+ * Exists for the parts of the request that are not messages, principally the
+ * tool-definition schemas, which the provider charges for on every request but
+ * which never appear in the message array. One estimator has to govern the
+ * whole payload or the gate measures a different thing than the provider bills.
+ */
+export declare function estimateValueTokens(value: unknown): number;
+export {};
+//# sourceMappingURL=token-tracker.d.ts.map

package/dist/ai-service/llm/context-v2/token-tracker.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"token-tracker.d.ts","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/token-tracker.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,YAAY,EAAE,MAAM,IAAI,CAAC;AAE3D,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAEhD;;;;GAIG;AACH,KAAK,kBAAkB,GAAG,YAAY,GAAG,aAAa,CAAC;AA8CvD,KAAK,SAAS,GAAG;KACd,CAAC,IACE,aAAa,GACb,cAAc,GACd,aAAa,GACb,mBAAmB,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC;CACjD,GAAG;IACF,iBAAiB,CAAC,EAAE,OAAO,CACzB,WAAW,CAAC,kBAAkB,CAAC,mBAAmB,CAAC,CAAC,CACrD,CAAC;CACH,CAAC;AAiBF,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,SAAS,GAAG,MAAM,CA0B5D;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,SAAS,GAAG,MAAM,CAE3D;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,cAAc,CAAC,KAAK,EAAE,kBAAkB,GAAG,MAAM,CAYhE;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,wBAAgB,sBAAsB,CACpC,QAAQ,EAAE,SAAS,kBAAkB,EAAE,GACtC,MAAM,CAQR;AAED;;;;;;;;GAQG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAE1D"}

package/dist/ai-service/llm/context-v2/token-tracker.js

@@ -0,0 +1,193 @@
+/**
+ * Characters per token for ASCII text. Matches the ratio the phase-1 tool
+ * summarizer already uses for its pruneable-token estimates.
+ */
+const CHARS_PER_TOKEN = 4;
+/**
+ * Non-ASCII text is charged one token per character. CJK, Cyrillic, and Arabic
+ * tokenize far denser than the ASCII ratio, and undercounting is the dangerous
+ * direction here: it skips compaction and leaves the session wedged.
+ */
+const NON_ASCII_CHARS_PER_TOKEN = 1;
+/**
+ * Flat token cost attributed to one run of inline binary data.
+ *
+ * Providers tokenize images and files by their decoded dimensions, not by the
+ * length of the base64 that carries them, so charging base64 per character
+ * overstates a multimodal context by orders of magnitude — which would trigger
+ * a paid summarization (and history loss) on a context that actually fits.
+ * Roughly the cost of a large screenshot; deliberately an over- rather than
+ * under-estimate.
+ */
+const BINARY_RUN_TOKENS = 1_500;
+/**
+ * An inline `data:` URL payload — how images, PDFs, and archives are carried in
+ * message content.
+ *
+ * Anchored on the `data:` header rather than on "looks like base64" on purpose.
+ * A bare long run of the base64 alphabet is indistinguishable from a hash, a
+ * minified bundle, or any long unspaced token, and charging those a flat binary
+ * cost would over-count real text — the very failure this constant exists to
+ * avoid. Raw-base64 fields carrying no `data:` prefix are therefore still
+ * counted per character, i.e. exactly as before this change.
+ *
+ * The header class allows `;` so multi-parameter media types match. It used to
+ * exclude it, which meant `data:image/png;base64,...` matched but
+ * `data:text/plain;charset=utf-8;base64,...` did not, and that payload was
+ * charged per base64 character. `"` and `\\` stay excluded so the match cannot
+ * run past the end of the JSON string it sits in.
+ */
+const DATA_URL_PAYLOAD = /data:[^",\\]{0,120}?,[A-Za-z0-9+/=]{256,}/g;
+function getNonCachedInputTokens(usage) {
+    if (typeof usage.inputTokenDetails?.noCacheTokens === "number") {
+        return usage.inputTokenDetails.noCacheTokens;
+    }
+    if (typeof usage.totalTokens === "number" &&
+        typeof usage.outputTokens === "number" &&
+        typeof usage.inputTokens === "number" &&
+        usage.totalTokens - usage.outputTokens === usage.inputTokens) {
+        return Math.max(0, usage.inputTokens - (usage.cachedInputTokens ?? 0));
+    }
+    return usage.inputTokens ?? 0;
+}
+export function getTotalInputTokens(usage) {
+    const nonCached = getNonCachedInputTokens(usage);
+    const cacheRead = usage.cachedInputTokens ?? usage.inputTokenDetails?.cacheReadTokens ?? 0;
+    const cacheWrite = usage.inputTokenDetails?.cacheWriteTokens ?? 0;
+    const hasDetailedBreakdown = usage.inputTokenDetails?.noCacheTokens !== undefined ||
+        usage.inputTokenDetails?.cacheReadTokens !== undefined ||
+        usage.inputTokenDetails?.cacheWriteTokens !== undefined;
+    if (hasDetailedBreakdown) {
+        return nonCached + cacheRead + cacheWrite;
+    }
+    if (typeof usage.totalTokens === "number" &&
+        typeof usage.outputTokens === "number") {
+        return Math.max(0, usage.totalTokens - usage.outputTokens);
+    }
+    if (cacheRead > 0 || cacheWrite > 0) {
+        return nonCached + cacheRead + cacheWrite;
+    }
+    return usage.inputTokens ?? nonCached;
+}
+/**
+ * Tokens consumed against the model's input context window, for display in the
+ * UI's context-usage indicator. This is the sum of fresh, cache-read, and
+ * cache-write input tokens — cache reads/writes occupy the input window just
+ * like fresh input (cache is a cost discount, not a context-window discount).
+ *
+ * Excludes outputTokens and reasoningTokens: those do not consume the input
+ * window. (Note: for Anthropic, extended-thinking tokens are bundled inside
+ * `output_tokens` in the raw API response and are not separately surfaced as
+ * `reasoningTokens` by @ai-sdk/anthropic, so this function's exclusion of
+ * reasoning tokens is defensive for other providers.)
+ */
+export function computeContextUsed(usage) {
+    return getTotalInputTokens(usage);
+}
+/**
+ * Total billed tokens: input (fresh + cache read + cache write) plus output.
+ *
+ * Do NOT use this for context-window or compaction-threshold decisions — use
+ * `computeContextUsed` instead. Output/reasoning tokens do not occupy the
+ * model's input window, and Anthropic bundles extended-thinking tokens into
+ * `output_tokens` (often 50–100k+). Counting them here inflates the measured
+ * context so compaction keeps re-triggering on the same turn — the
+ * output-inclusive compaction loop. This helper has no context/compaction
+ * callers; it is retained as a correct "total billed tokens" measure (input +
+ * output) for future billing/telemetry use, intentionally not wired into any
+ * context/compaction path.
+ *
+ * @deprecated Not for context-window or compaction decisions — use
+ * `computeContextUsed` instead. Retained only as a billing/telemetry
+ * "total billed tokens" measure.
+ */
+export function getTotalTokens(usage) {
+    const recomputedTotal = getTotalInputTokens(usage) + (usage.outputTokens ?? 0);
+    const hasDetailedBreakdown = usage.inputTokenDetails?.noCacheTokens !== undefined ||
+        usage.inputTokenDetails?.cacheReadTokens !== undefined ||
+        usage.inputTokenDetails?.cacheWriteTokens !== undefined;
+    if (hasDetailedBreakdown) {
+        return recomputedTotal;
+    }
+    return usage.totalTokens ?? recomputedTotal;
+}
+/**
+ * Estimate the input tokens a message array will consume, from its content
+ * alone.
+ *
+ * Every other measure in this file needs provider-reported usage, which does
+ * not exist before the first request of a restored session — and a context
+ * restored from storage can already be too large to send (ENG-5282). Only a
+ * content-based estimate can catch that pre-flight.
+ *
+ * `_superblocks` metadata is excluded: `toSdkMessages` strips it before the
+ * wire payload is built, so counting it would overstate the request.
+ *
+ * Pass the array that is actually going on the wire, i.e. the output of
+ * `Context.getMessages()`, which already carries the system prompt and has had
+ * `_superblocks` stripped. Passing raw `StoredMessage`s instead omits the
+ * system prompt and counts a slightly different shape than the request.
+ *
+ * KNOWN GAP. The tool-definition schemas are passed to the provider alongside
+ * the messages and are not part of this array, so a session with a large tool
+ * surface can estimate under the threshold while the real request exceeds the
+ * provider limit. That case degrades to the provider rejecting it, reported as
+ * `context_too_large` rather than a generic provider error, so it stays
+ * diagnosable.
+ *
+ * Accuracy is coarse: it gates a compaction decision and the provider remains
+ * the authority on the real limit. The direction of the error is content
+ * shaped, not fixed. JSON escaping and key names add roughly a third on top of
+ * the content, so prose is over-counted while code, which real tokenizers
+ * charge closer to 3 characters per token, comes out about right.
+ */
+export function estimateMessagesTokens(messages) {
+    let tokens = 0;
+    for (const message of messages) {
+        tokens += estimateSerializedTokens(JSON.stringify(message, omitSuperblocks) ?? "");
+    }
+    return tokens;
+}
+/**
+ * Estimate the input tokens an arbitrary serializable value will consume, using
+ * the same ratios as {@link estimateMessagesTokens}.
+ *
+ * Exists for the parts of the request that are not messages, principally the
+ * tool-definition schemas, which the provider charges for on every request but
+ * which never appear in the message array. One estimator has to govern the
+ * whole payload or the gate measures a different thing than the provider bills.
+ */
+export function estimateValueTokens(value) {
+    return estimateSerializedTokens(JSON.stringify(value, omitSuperblocks) ?? "");
+}
+/**
+ * `JSON.stringify` replacer that drops `_superblocks`. A raw `StoredMessage`
+ * carries it; the wire payload does not, because `toSdkMessages` strips it, so
+ * counting it would overstate the request.
+ */
+function omitSuperblocks(key, value) {
+    return key === "_superblocks" ? undefined : value;
+}
+function estimateSerializedTokens(serialized) {
+    let tokens = 0;
+    // Charge inline binary data a flat cost rather than its base64 length.
+    const text = serialized.replace(DATA_URL_PAYLOAD, () => {
+        tokens += BINARY_RUN_TOKENS;
+        return "";
+    });
+    // Split ASCII from the rest: the two tokenize at very different densities.
+    let asciiChars = 0;
+    let nonAsciiChars = 0;
+    for (let i = 0; i < text.length; i++) {
+        if (text.charCodeAt(i) < 128) {
+            asciiChars++;
+        }
+        else {
+            nonAsciiChars++;
+        }
+    }
+    return (tokens +
+        Math.ceil(asciiChars / CHARS_PER_TOKEN) +
+        Math.ceil(nonAsciiChars / NON_ASCII_CHARS_PER_TOKEN));
+}
+//# sourceMappingURL=token-tracker.js.map

package/dist/ai-service/llm/context-v2/token-tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"token-tracker.js","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/token-tracker.ts"],"names":[],"mappings":"AAWA;;;GAGG;AACH,MAAM,eAAe,GAAG,CAAC,CAAC;AAE1B;;;;GAIG;AACH,MAAM,yBAAyB,GAAG,CAAC,CAAC;AAEpC;;;;;;;;;GASG;AACH,MAAM,iBAAiB,GAAG,KAAK,CAAC;AAEhC;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,gBAAgB,GAAG,4CAA4C,CAAC;AActE,SAAS,uBAAuB,CAAC,KAAgB;IAC/C,IAAI,OAAO,KAAK,CAAC,iBAAiB,EAAE,aAAa,KAAK,QAAQ,EAAE,CAAC;QAC/D,OAAO,KAAK,CAAC,iBAAiB,CAAC,aAAa,CAAC;IAC/C,CAAC;IACD,IACE,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,YAAY,KAAK,QAAQ;QACtC,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,KAAK,CAAC,WAAW,GAAG,KAAK,CAAC,YAAY,KAAK,KAAK,CAAC,WAAW,EAC5D,CAAC;QACD,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,WAAW,GAAG,CAAC,KAAK,CAAC,iBAAiB,IAAI,CAAC,CAAC,CAAC,CAAC;IACzE,CAAC;IACD,OAAO,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC;AAChC,CAAC;AAED,MAAM,UAAU,mBAAmB,CAAC,KAAgB;IAClD,MAAM,SAAS,GAAG,uBAAuB,CAAC,KAAK,CAAC,CAAC;IACjD,MAAM,SAAS,GACb,KAAK,CAAC,iBAAiB,IAAI,KAAK,CAAC,iBAAiB,EAAE,eAAe,IAAI,CAAC,CAAC;IAC3E,MAAM,UAAU,GAAG,KAAK,CAAC,iBAAiB,EAAE,gBAAgB,IAAI,CAAC,CAAC;IAClE,MAAM,oBAAoB,GACxB,KAAK,CAAC,iBAAiB,EAAE,aAAa,KAAK,SAAS;QACpD,KAAK,CAAC,iBAAiB,EAAE,eAAe,KAAK,SAAS;QACtD,KAAK,CAAC,iBAAiB,EAAE,gBAAgB,KAAK,SAAS,CAAC;IAE1D,IAAI,oBAAoB,EAAE,CAAC;QACzB,OAAO,SAAS,GAAG,SAAS,GAAG,UAAU,CAAC;IAC5C,CAAC;IAED,IACE,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;QACrC,OAAO,KAAK,CAAC,YAAY,KAAK,QAAQ,EACtC,CAAC;QACD,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,WAAW,GAAG,KAAK,CAAC,YAAY,CAAC,CAAC;IAC7D,CAAC;IAED,IAAI,SAAS,GAAG,CAAC,IAAI,UAAU,GAAG,CAAC,EAAE,CAAC;QACpC,OAAO,SAAS,GAAG,SAAS,GAAG,UAAU,CAAC;IAC5C,CAAC;IAED,OAAO,KAAK,CAAC,WAAW,IAAI,SAAS,CAAC;AACxC,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,kBAAkB,CAAC,KAAgB;IACjD,OAAO,mBAAmB,CAAC,KAAK,CAAC,CAAC;AACpC,CAAC;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,UAAU,cAAc,CAAC,KAAyB;IACtD,MAAM,eAAe,GACnB,mBAAmB,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,YAAY,IAAI,CAAC,CAAC,CAAC;IACzD,MAAM,oBAAoB,GACxB,KAAK,CAAC,iBAAiB,EAAE,aAAa,KAAK,SAAS;QACpD,KAAK,CAAC,iBAAiB,EAAE,eAAe,KAAK,SAAS;QACtD,KAAK,CAAC,iBAAiB,EAAE,gBAAgB,KAAK,SAAS,CAAC;IAE1D,IAAI,oBAAoB,EAAE,CAAC;QACzB,OAAO,eAAe,CAAC;IACzB,CAAC;IACD,OAAO,KAAK,CAAC,WAAW,IAAI,eAAe,CAAC;AAC9C,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,MAAM,UAAU,sBAAsB,CACpC,QAAuC;IAEvC,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,MAAM,IAAI,wBAAwB,CAChC,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,eAAe,CAAC,IAAI,EAAE,CAC/C,CAAC;IACJ,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,mBAAmB,CAAC,KAAc;IAChD,OAAO,wBAAwB,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,EAAE,eAAe,CAAC,IAAI,EAAE,CAAC,CAAC;AAChF,CAAC;AAED;;;;GAIG;AACH,SAAS,eAAe,CAAC,GAAW,EAAE,KAAc;IAClD,OAAO,GAAG,KAAK,cAAc,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,CAAC;AACpD,CAAC;AAED,SAAS,wBAAwB,CAAC,UAAkB;IAClD,IAAI,MAAM,GAAG,CAAC,CAAC;IAEf,uEAAuE;IACvE,MAAM,IAAI,GAAG,UAAU,CAAC,OAAO,CAAC,gBAAgB,EAAE,GAAG,EAAE;QACrD,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,EAAE,CAAC;IACZ,CAAC,CAAC,CAAC;IAEH,2EAA2E;IAC3E,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,aAAa,GAAG,CAAC,CAAC;IACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACrC,IAAI,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,GAAG,EAAE,CAAC;YAC7B,UAAU,EAAE,CAAC;QACf,CAAC;aAAM,CAAC;YACN,aAAa,EAAE,CAAC;QAClB,CAAC;IACH,CAAC;IAED,OAAO,CACL,MAAM;QACN,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,eAAe,CAAC;QACvC,IAAI,CAAC,IAAI,CAAC,aAAa,GAAG,yBAAyB,CAAC,CACrD,CAAC;AACJ,CAAC"}

package/dist/ai-service/llm/context-v2/truncation-fallback.d.ts

@@ -0,0 +1,10 @@
+import { LLMObsTracer } from "../../llmobs/tracer.js";
+import type { SummarizeResult } from "./phase2-session-summarizer.js";
+import type { StoredMessage } from "./types.js";
+/**
+ * Creates a truncation-based fallback when LLM summarization fails.
+ * Generates a timeline summary of the conversation without returning
+ * the actual messages, which can cause provider errors.
+ */
+export declare function createTruncationFallback(messages: StoredMessage[], tracer?: LLMObsTracer): SummarizeResult;
+//# sourceMappingURL=truncation-fallback.d.ts.map

package/dist/ai-service/llm/context-v2/truncation-fallback.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"truncation-fallback.d.ts","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/truncation-fallback.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,wBAAwB,CAAC;AAEtD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,gCAAgC,CAAC;AAEtE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AA6FhD;;;;GAIG;AACH,wBAAgB,wBAAwB,CACtC,QAAQ,EAAE,aAAa,EAAE,EACzB,MAAM,GAAE,YAAoC,GAC3C,eAAe,CA0DjB"}

package/dist/ai-service/llm/context-v2/truncation-fallback.js

@@ -0,0 +1,124 @@
+import { LLMObsTracer } from "../../llmobs/tracer.js";
+import { extractTextFromMessage } from "../client.js";
+import { toStoredMessage } from "./types.js";
+import { truncateUtf16Safe } from "./well-formed-text.js";
+const MAX_TURNS_TO_SHOW = 10;
+const MAX_USER_MESSAGE_LENGTH = 1000;
+const MAX_ASSISTANT_MESSAGE_LENGTH = 1000;
+function truncateText(text, maxLength) {
+    const normalized = text.replace(/\s+/g, " ").trim();
+    if (normalized.length <= maxLength) {
+        return normalized;
+    }
+    return truncateUtf16Safe(normalized, maxLength) + "...";
+}
+function extractTurnSummaries(messages) {
+    const summaries = [];
+    let currentUserMessage = null;
+    let lastAssistantText = null;
+    for (const msg of messages) {
+        if (msg.role === "user") {
+            const userText = extractTextFromMessage(msg, "last");
+            if (userText) {
+                if (currentUserMessage !== null) {
+                    summaries.push({
+                        userMessage: currentUserMessage,
+                        assistantMessage: lastAssistantText,
+                    });
+                }
+                currentUserMessage = userText;
+                lastAssistantText = null;
+            }
+        }
+        else if (msg.role === "assistant" && currentUserMessage !== null) {
+            const text = extractTextFromMessage(msg);
+            if (text) {
+                lastAssistantText = text;
+            }
+        }
+    }
+    if (currentUserMessage !== null) {
+        summaries.push({
+            userMessage: currentUserMessage,
+            assistantMessage: lastAssistantText,
+        });
+    }
+    return summaries;
+}
+function formatTimeline(summaries, totalMessages) {
+    const totalTurns = summaries.length;
+    const turnsToShow = summaries.slice(-MAX_TURNS_TO_SHOW);
+    const startNumber = totalTurns - turnsToShow.length + 1;
+    const isLimited = totalTurns > MAX_TURNS_TO_SHOW;
+    let header = `[Context limit reached - ${totalMessages} messages were dropped. Falling back to session summary.]\n\n`;
+    header += isLimited
+        ? `## Session History (showing last ${turnsToShow.length} of ${totalTurns} turns)\n\n`
+        : `## Session History\n\n`;
+    const lines = turnsToShow.map((turn, i) => {
+        const turnNumber = startNumber + i;
+        const isLast = i === turnsToShow.length - 1;
+        const userLabel = isLast ? "**Current request:**" : "User:";
+        // Preserve the full text for the current/last user message so the task context is not lost.
+        const userText = isLast
+            ? truncateText(turn.userMessage, Infinity)
+            : truncateText(turn.userMessage, MAX_USER_MESSAGE_LENGTH);
+        let line = `${turnNumber}. ${userLabel} "${userText}"`;
+        if (turn.assistantMessage) {
+            line += `\n    Assistant: "${truncateText(turn.assistantMessage, MAX_ASSISTANT_MESSAGE_LENGTH)}"`;
+        }
+        else if (!isLast) {
+            line += `\n    Assistant: "(no response)"`;
+        }
+        return line;
+    });
+    return header + lines.join("\n\n");
+}
+/**
+ * Creates a truncation-based fallback when LLM summarization fails.
+ * Generates a timeline summary of the conversation without returning
+ * the actual messages, which can cause provider errors.
+ */
+export function createTruncationFallback(messages, tracer = LLMObsTracer.instance) {
+    return tracer.trace({ name: "session.summarize.fallback", kind: "task" }, (fallbackSpan) => {
+        const summaries = extractTurnSummaries(messages);
+        if (summaries.length === 0) {
+            const fallbackContent = `[Context limit reached - ${messages.length} messages were dropped. ` +
+                `No user messages found to summarize.]`;
+            tracer.annotate(fallbackSpan, {
+                metadata: {
+                    input: {
+                        messageCount: messages.length,
+                    },
+                    result: {
+                        summaryLength: fallbackContent.length,
+                        turnCount: 0,
+                    },
+                },
+            });
+            return {
+                summary: fallbackContent,
+                messages: [
+                    toStoredMessage({ role: "user", content: fallbackContent }, { isSummary: true }),
+                ],
+            };
+        }
+        const summaryText = formatTimeline(summaries, messages.length);
+        const summaryMessage = toStoredMessage({ role: "user", content: summaryText }, { isSummary: true });
+        tracer.annotate(fallbackSpan, {
+            metadata: {
+                input: {
+                    messageCount: messages.length,
+                },
+                result: {
+                    summaryLength: summaryText.length,
+                    turnCount: summaries.length,
+                },
+            },
+        });
+        return {
+            summary: summaryText,
+            messages: [summaryMessage],
+        };
+    });
+}
+//# sourceMappingURL=truncation-fallback.js.map

package/dist/ai-service/llm/context-v2/truncation-fallback.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"truncation-fallback.js","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/truncation-fallback.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,wBAAwB,CAAC;AACtD,OAAO,EAAE,sBAAsB,EAAE,MAAM,cAAc,CAAC;AAEtD,OAAO,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAE7C,OAAO,EAAE,iBAAiB,EAAE,MAAM,uBAAuB,CAAC;AAE1D,MAAM,iBAAiB,GAAG,EAAE,CAAC;AAC7B,MAAM,uBAAuB,GAAG,IAAI,CAAC;AACrC,MAAM,4BAA4B,GAAG,IAAI,CAAC;AAO1C,SAAS,YAAY,CAAC,IAAY,EAAE,SAAiB;IACnD,MAAM,UAAU,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;IACpD,IAAI,UAAU,CAAC,MAAM,IAAI,SAAS,EAAE,CAAC;QACnC,OAAO,UAAU,CAAC;IACpB,CAAC;IACD,OAAO,iBAAiB,CAAC,UAAU,EAAE,SAAS,CAAC,GAAG,KAAK,CAAC;AAC1D,CAAC;AAED,SAAS,oBAAoB,CAAC,QAAyB;IACrD,MAAM,SAAS,GAAkB,EAAE,CAAC;IACpC,IAAI,kBAAkB,GAAkB,IAAI,CAAC;IAC7C,IAAI,iBAAiB,GAAkB,IAAI,CAAC;IAE5C,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;QAC3B,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACxB,MAAM,QAAQ,GAAG,sBAAsB,CAAC,GAAG,EAAE,MAAM,CAAC,CAAC;YACrD,IAAI,QAAQ,EAAE,CAAC;gBACb,IAAI,kBAAkB,KAAK,IAAI,EAAE,CAAC;oBAChC,SAAS,CAAC,IAAI,CAAC;wBACb,WAAW,EAAE,kBAAkB;wBAC/B,gBAAgB,EAAE,iBAAiB;qBACpC,CAAC,CAAC;gBACL,CAAC;gBACD,kBAAkB,GAAG,QAAQ,CAAC;gBAC9B,iBAAiB,GAAG,IAAI,CAAC;YAC3B,CAAC;QACH,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,IAAI,kBAAkB,KAAK,IAAI,EAAE,CAAC;YACnE,MAAM,IAAI,GAAG,sBAAsB,CAAC,GAAG,CAAC,CAAC;YACzC,IAAI,IAAI,EAAE,CAAC;gBACT,iBAAiB,GAAG,IAAI,CAAC;YAC3B,CAAC;QACH,CAAC;IACH,CAAC;IAED,IAAI,kBAAkB,KAAK,IAAI,EAAE,CAAC;QAChC,SAAS,CAAC,IAAI,CAAC;YACb,WAAW,EAAE,kBAAkB;YAC/B,gBAAgB,EAAE,iBAAiB;SACpC,CAAC,CAAC;IACL,CAAC;IAED,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,SAAS,cAAc,CACrB,SAAwB,EACxB,aAAqB;IAErB,MAAM,UAAU,GAAG,SAAS,CAAC,MAAM,CAAC;IACpC,MAAM,WAAW,GAAG,SAAS,CAAC,KAAK,CAAC,CAAC,iBAAiB,CAAC,CAAC;IACxD,MAAM,WAAW,GAAG,UAAU,GAAG,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC;IACxD,MAAM,SAAS,GAAG,UAAU,GAAG,iBAAiB,CAAC;IAEjD,IAAI,MAAM,GAAG,4BAA4B,aAAa,+DAA+D,CAAC;IACtH,MAAM,IAAI,SAAS;QACjB,CAAC,CAAC,oCAAoC,WAAW,CAAC,MAAM,OAAO,UAAU,aAAa;QACtF,CAAC,CAAC,wBAAwB,CAAC;IAE7B,MAAM,KAAK,GAAG,WAAW,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC,EAAE,EAAE;QACxC,MAAM,UAAU,GAAG,WAAW,GAAG,CAAC,CAAC;QACnC,MAAM,MAAM,GAAG,CAAC,KAAK,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC;QAC5C,MAAM,SAAS,GAAG,MAAM,CAAC,CAAC,CAAC,sBAAsB,CAAC,CAAC,CAAC,OAAO,CAAC;QAC5D,4FAA4F;QAC5F,MAAM,QAAQ,GAAG,MAAM;YACrB,CAAC,CAAC,YAAY,CAAC,IAAI,CAAC,WAAW,EAAE,QAAQ,CAAC;YAC1C,CAAC,CAAC,YAAY,CAAC,IAAI,CAAC,WAAW,EAAE,uBAAuB,CAAC,CAAC;QAE5D,IAAI,IAAI,GAAG,GAAG,UAAU,KAAK,SAAS,KAAK,QAAQ,GAAG,CAAC;QAEvD,IAAI,IAAI,CAAC,gBAAgB,EAAE,CAAC;YAC1B,IAAI,IAAI,qBAAqB,YAAY,CAAC,IAAI,CAAC,gBAAgB,EAAE,4BAA4B,CAAC,GAAG,CAAC;QACpG,CAAC;aAAM,IAAI,CAAC,MAAM,EAAE,CAAC;YACnB,IAAI,IAAI,kCAAkC,CAAC;QAC7C,CAAC;QAED,OAAO,IAAI,CAAC;IACd,CAAC,CAAC,CAAC;IAEH,OAAO,MAAM,GAAG,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;AACrC,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,wBAAwB,CACtC,QAAyB,EACzB,SAAuB,YAAY,CAAC,QAAQ;IAE5C,OAAO,MAAM,CAAC,KAAK,CACjB,EAAE,IAAI,EAAE,4BAA4B,EAAE,IAAI,EAAE,MAAM,EAAE,EACpD,CAAC,YAAY,EAAE,EAAE;QACf,MAAM,SAAS,GAAG,oBAAoB,CAAC,QAAQ,CAAC,CAAC;QAEjD,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC3B,MAAM,eAAe,GACnB,4BAA4B,QAAQ,CAAC,MAAM,0BAA0B;gBACrE,uCAAuC,CAAC;YAE1C,MAAM,CAAC,QAAQ,CAAC,YAAY,EAAE;gBAC5B,QAAQ,EAAE;oBACR,KAAK,EAAE;wBACL,YAAY,EAAE,QAAQ,CAAC,MAAM;qBAC9B;oBACD,MAAM,EAAE;wBACN,aAAa,EAAE,eAAe,CAAC,MAAM;wBACrC,SAAS,EAAE,CAAC;qBACb;iBACF;aACF,CAAC,CAAC;YAEH,OAAO;gBACL,OAAO,EAAE,eAAe;gBACxB,QAAQ,EAAE;oBACR,eAAe,CACb,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,eAAe,EAAE,EAC1C,EAAE,SAAS,EAAE,IAAI,EAAE,CACpB;iBACF;aACF,CAAC;QACJ,CAAC;QAED,MAAM,WAAW,GAAG,cAAc,CAAC,SAAS,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC;QAC/D,MAAM,cAAc,GAAG,eAAe,CACpC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,WAAW,EAAE,EACtC,EAAE,SAAS,EAAE,IAAI,EAAE,CACpB,CAAC;QAEF,MAAM,CAAC,QAAQ,CAAC,YAAY,EAAE;YAC5B,QAAQ,EAAE;gBACR,KAAK,EAAE;oBACL,YAAY,EAAE,QAAQ,CAAC,MAAM;iBAC9B;gBACD,MAAM,EAAE;oBACN,aAAa,EAAE,WAAW,CAAC,MAAM;oBACjC,SAAS,EAAE,SAAS,CAAC,MAAM;iBAC5B;aACF;SACF,CAAC,CAAC;QAEH,OAAO;YACL,OAAO,EAAE,WAAW;YACpB,QAAQ,EAAE,CAAC,cAAc,CAAC;SAC3B,CAAC;IACJ,CAAC,CACF,CAAC;AACJ,CAAC"}

package/dist/ai-service/llm/context-v2/turns.d.ts

@@ -0,0 +1,24 @@
+import type { StoredMessage } from "./types.js";
+interface Turn {
+    /** Index of the first message in this turn (user message) */
+    startIndex: number;
+    /** Index of the last message in this turn (assistant message with cumulativeInputTokens) */
+    endIndex: number;
+    /** Token count for this turn (delta from previous turn) */
+    tokenCount: number;
+    /** Indices of tool messages within this turn */
+    toolMessageIndices: number[];
+}
+/**
+ * Groups messages into turns. Each turn ends with an assistant message
+ * that has cumulativeInputTokens. The token count for a turn is calculated
+ * as the delta between consecutive assistant messages' cumulativeInputTokens.
+ */
+export declare function identifyTurns(messages: StoredMessage[]): Turn[];
+/**
+ * Selects the most recent turns that fit within the token budget.
+ * Walks backward through turns, accumulating token counts.
+ */
+export declare function selectRecentTurns(turns: Turn[], tokenBudget: number): Turn[];
+export {};
+//# sourceMappingURL=turns.d.ts.map

package/dist/ai-service/llm/context-v2/turns.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"turns.d.ts","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/turns.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAIhD,UAAU,IAAI;IACZ,6DAA6D;IAC7D,UAAU,EAAE,MAAM,CAAC;IACnB,4FAA4F;IAC5F,QAAQ,EAAE,MAAM,CAAC;IACjB,2DAA2D;IAC3D,UAAU,EAAE,MAAM,CAAC;IACnB,gDAAgD;IAChD,kBAAkB,EAAE,MAAM,EAAE,CAAC;CAC9B;AAiBD;;;;GAIG;AACH,wBAAgB,aAAa,CAAC,QAAQ,EAAE,aAAa,EAAE,GAAG,IAAI,EAAE,CAsD/D;AAED;;;GAGG;AACH,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE,WAAW,EAAE,MAAM,GAAG,IAAI,EAAE,CAe5E"}

package/dist/ai-service/llm/context-v2/turns.js

@@ -0,0 +1,85 @@
+import { getPrefixedLogger } from "../../../util/logger.js";
+const logger = getPrefixedLogger("[context-v2:turns]");
+function isToolMessage(msg) {
+    return msg.role === "tool" && Array.isArray(msg.content);
+}
+function isUnsummarizedToolMessage(msg) {
+    return isToolMessage(msg) && !msg._superblocks?.summarized;
+}
+function isAssistantWithTokens(msg) {
+    return (msg.role === "assistant" &&
+        msg._superblocks?.cumulativeInputTokens !== undefined);
+}
+/**
+ * Groups messages into turns. Each turn ends with an assistant message
+ * that has cumulativeInputTokens. The token count for a turn is calculated
+ * as the delta between consecutive assistant messages' cumulativeInputTokens.
+ */
+export function identifyTurns(messages) {
+    // NOTE: Unlike `getLastKnownInputTokens` (context.ts), this function does NOT
+    // skip messages flagged `preCompactionInputTokenBaseline`. A retained
+    // pre-compaction assistant intentionally forms a turn carrying its full
+    // historical `cumulativeInputTokens`, and the `Math.max(0, …)` clamp below
+    // absorbs the negative delta on the first post-compaction turn. This keeps
+    // historical per-turn accounting intact while preventing negative phase-1
+    // protect/prune budgets. Adding a flag check here would change how
+    // `selectRecentTurns` sizes pre-compaction turns — don't do it without
+    // re-deriving that interaction.
+    const turns = [];
+    let turnStart = 0;
+    let prevTokens = 0;
+    for (let i = 0; i < messages.length; i++) {
+        const msg = messages[i];
+        if (isAssistantWithTokens(msg)) {
+            const currentTokens = msg._superblocks.cumulativeInputTokens;
+            const toolIndices = [];
+            for (let j = turnStart; j <= i; j++) {
+                if (isUnsummarizedToolMessage(messages[j])) {
+                    toolIndices.push(j);
+                }
+            }
+            const delta = currentTokens - prevTokens;
+            if (delta < 0) {
+                // Expected at a compaction boundary: the next epoch's first assistant
+                // reports a smaller cumulative than the retained pre-compaction peak.
+                // Logged at debug so a negative delta arising OUTSIDE a compaction
+                // boundary (a future accounting regression) stays visible in
+                // `.superblocks/` logs instead of being silently zeroed by the clamp.
+                logger.debug(`Clamped negative turn delta (${delta}) to 0 at message index ${i}`);
+            }
+            turns.push({
+                startIndex: turnStart,
+                endIndex: i,
+                // Clamp at 0: across a compaction boundary the next epoch's first
+                // assistant reports a smaller cumulative than the retained
+                // pre-compaction peak, which would otherwise yield a negative delta and
+                // corrupt the phase-1 protect/prune token budgets that consume this.
+                tokenCount: Math.max(0, delta),
+                toolMessageIndices: toolIndices,
+            });
+            prevTokens = currentTokens;
+            turnStart = i + 1;
+        }
+    }
+    return turns;
+}
+/**
+ * Selects the most recent turns that fit within the token budget.
+ * Walks backward through turns, accumulating token counts.
+ */
+export function selectRecentTurns(turns, tokenBudget) {
+    const selected = [];
+    let accumulated = 0;
+    for (let i = turns.length - 1; i >= 0; i--) {
+        const turn = turns[i];
+        if (accumulated + turn.tokenCount <= tokenBudget) {
+            accumulated += turn.tokenCount;
+            selected.unshift(turn);
+        }
+        else {
+            break;
+        }
+    }
+    return selected;
+}
+//# sourceMappingURL=turns.js.map

package/dist/ai-service/llm/context-v2/turns.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"turns.js","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/turns.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,iBAAiB,EAAE,MAAM,yBAAyB,CAAC;AAG5D,MAAM,MAAM,GAAG,iBAAiB,CAAC,oBAAoB,CAAC,CAAC;AAavD,SAAS,aAAa,CAAC,GAAkB;IACvC,OAAO,GAAG,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;AAC3D,CAAC;AAED,SAAS,yBAAyB,CAAC,GAAkB;IACnD,OAAO,aAAa,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,YAAY,EAAE,UAAU,CAAC;AAC7D,CAAC;AAED,SAAS,qBAAqB,CAAC,GAAkB;IAC/C,OAAO,CACL,GAAG,CAAC,IAAI,KAAK,WAAW;QACxB,GAAG,CAAC,YAAY,EAAE,qBAAqB,KAAK,SAAS,CACtD,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,aAAa,CAAC,QAAyB;IACrD,8EAA8E;IAC9E,sEAAsE;IACtE,wEAAwE;IACxE,2EAA2E;IAC3E,2EAA2E;IAC3E,0EAA0E;IAC1E,mEAAmE;IACnE,uEAAuE;IACvE,gCAAgC;IAChC,MAAM,KAAK,GAAW,EAAE,CAAC;IACzB,IAAI,SAAS,GAAG,CAAC,CAAC;IAClB,IAAI,UAAU,GAAG,CAAC,CAAC;IAEnB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACzC,MAAM,GAAG,GAAG,QAAQ,CAAC,CAAC,CAAE,CAAC;QACzB,IAAI,qBAAqB,CAAC,GAAG,CAAC,EAAE,CAAC;YAC/B,MAAM,aAAa,GAAG,GAAG,CAAC,YAAa,CAAC,qBAAsB,CAAC;YAC/D,MAAM,WAAW,GAAa,EAAE,CAAC;YAEjC,KAAK,IAAI,CAAC,GAAG,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACpC,IAAI,yBAAyB,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,EAAE,CAAC;oBAC5C,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;gBACtB,CAAC;YACH,CAAC;YAED,MAAM,KAAK,GAAG,aAAa,GAAG,UAAU,CAAC;YACzC,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;gBACd,sEAAsE;gBACtE,sEAAsE;gBACtE,mEAAmE;gBACnE,6DAA6D;gBAC7D,sEAAsE;gBACtE,MAAM,CAAC,KAAK,CACV,gCAAgC,KAAK,2BAA2B,CAAC,EAAE,CACpE,CAAC;YACJ,CAAC;YACD,KAAK,CAAC,IAAI,CAAC;gBACT,UAAU,EAAE,SAAS;gBACrB,QAAQ,EAAE,CAAC;gBACX,kEAAkE;gBAClE,2DAA2D;gBAC3D,wEAAwE;gBACxE,qEAAqE;gBACrE,UAAU,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC;gBAC9B,kBAAkB,EAAE,WAAW;aAChC,CAAC,CAAC;YAEH,UAAU,GAAG,aAAa,CAAC;YAC3B,SAAS,GAAG,CAAC,GAAG,CAAC,CAAC;QACpB,CAAC;IACH,CAAC;IAED,OAAO,KAAK,CAAC;AACf,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAAa,EAAE,WAAmB;IAClE,MAAM,QAAQ,GAAW,EAAE,CAAC;IAC5B,IAAI,WAAW,GAAG,CAAC,CAAC;IAEpB,KAAK,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3C,MAAM,IAAI,GAAG,KAAK,CAAC,CAAC,CAAE,CAAC;QACvB,IAAI,WAAW,GAAG,IAAI,CAAC,UAAU,IAAI,WAAW,EAAE,CAAC;YACjD,WAAW,IAAI,IAAI,CAAC,UAAU,CAAC;YAC/B,QAAQ,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;QACzB,CAAC;aAAM,CAAC;YACN,MAAM;QACR,CAAC;IACH,CAAC;IAED,OAAO,QAAQ,CAAC;AAClB,CAAC"}

package/dist/ai-service/llm/context-v2/types-v1.d.ts

@@ -0,0 +1,36 @@
+import type { AttachmentsRecord as AttachmentsRecordV1 } from "../../attachments/store.js";
+import type { CompactionSummary } from "./types.js";
+export interface TokenizedMessage {
+    role: "system" | "user" | "assistant" | "tool";
+    content: unknown;
+    tokenCount?: number;
+    timestamp?: number;
+    hasBreakPoint?: boolean;
+}
+export type TokenizedResponseMessage = TokenizedMessage & {
+    role: "assistant" | "tool";
+};
+export interface L3TurnRecord {
+    user: TokenizedMessage;
+    assistant?: TokenizedMessage;
+}
+export interface L2TurnRecord {
+    user: TokenizedMessage;
+    response: TokenizedResponseMessage[];
+}
+export interface L1TurnRecord {
+    user: TokenizedMessage;
+    steps: TokenizedResponseMessage[][];
+}
+/**
+ * V1 context record format (no version field).
+ */
+export interface ContextRecordV1 {
+    charsPerToken: number;
+    l1Turns: L1TurnRecord[];
+    l2Turns: L2TurnRecord[];
+    l3Turns: L3TurnRecord[];
+    toolSummariesById?: Record<string, CompactionSummary>;
+    attachments?: AttachmentsRecordV1;
+}
+//# sourceMappingURL=types-v1.d.ts.map

package/dist/ai-service/llm/context-v2/types-v1.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types-v1.d.ts","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/types-v1.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,iBAAiB,IAAI,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AAC3F,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,YAAY,CAAC;AAEpD,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,QAAQ,GAAG,MAAM,GAAG,WAAW,GAAG,MAAM,CAAC;IAC/C,OAAO,EAAE,OAAO,CAAC;IACjB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,aAAa,CAAC,EAAE,OAAO,CAAC;CACzB;AAED,MAAM,MAAM,wBAAwB,GAAG,gBAAgB,GAAG;IACxD,IAAI,EAAE,WAAW,GAAG,MAAM,CAAC;CAC5B,CAAC;AAEF,MAAM,WAAW,YAAY;IAC3B,IAAI,EAAE,gBAAgB,CAAC;IACvB,SAAS,CAAC,EAAE,gBAAgB,CAAC;CAC9B;AAED,MAAM,WAAW,YAAY;IAC3B,IAAI,EAAE,gBAAgB,CAAC;IACvB,QAAQ,EAAE,wBAAwB,EAAE,CAAC;CACtC;AAED,MAAM,WAAW,YAAY;IAC3B,IAAI,EAAE,gBAAgB,CAAC;IACvB,KAAK,EAAE,wBAAwB,EAAE,EAAE,CAAC;CACrC;AAED;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,aAAa,EAAE,MAAM,CAAC;IACtB,OAAO,EAAE,YAAY,EAAE,CAAC;IACxB,OAAO,EAAE,YAAY,EAAE,CAAC;IACxB,OAAO,EAAE,YAAY,EAAE,CAAC;IACxB,iBAAiB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,iBAAiB,CAAC,CAAC;IACtD,WAAW,CAAC,EAAE,mBAAmB,CAAC;CACnC"}

package/dist/ai-service/llm/context-v2/types-v1.js

@@ -0,0 +1,3 @@
+// types from v1 context, needed for migrations
+export {};
+//# sourceMappingURL=types-v1.js.map

package/dist/ai-service/llm/context-v2/types-v1.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types-v1.js","sourceRoot":"","sources":["../../../../src/ai-service/llm/context-v2/types-v1.ts"],"names":[],"mappings":"AAAA,+CAA+C"}