mindweave 0.0.1 → 2.1.2

package/dist/dynamo/contextWindow.js

@@ -0,0 +1,243 @@
+/**
+ * contextWindow.ts — model-anchored compaction thresholds (pure).
+ *
+ * Compaction is anchored to the model's REAL context window rather than a fixed
+ * number: reserve room for the summary output plus a safety buffer, then compact
+ * when usage crosses what's left. That keeps the thresholds correct per model, and
+ * automatically right when a stronger or longer model is wired in, instead of a
+ * single hard-coded ceiling.
+ *
+ * The window itself is the driver's answer, and it is each model's SHARP window —
+ * the span where attention stays reliable — not its storage maximum. A model that
+ * stores 1M tokens can still degrade from tool-noise dilution long before that, and
+ * on BYOK every token is the user's money, so compacting to the sharp window is
+ * both more accurate and cheaper.
+ *
+ * THE DIVISION OF LABOUR. A driver reports facts it can measure — its window, and
+ * the ceiling it puts on a buffered call. This module owns every decision made on
+ * top of them: how much headroom to leave, where the bars land, and the invariant
+ * that ties them together. A driver returning its own bars would be a decision in
+ * the driver, which is the line this project doesn't cross; see BOUNDARY.md.
+ */
+import { manifestForModel } from "../drivers/registry.js";
+/**
+ * Reserve used when a driver declares no buffered ceiling of its own — it sends no
+ * `max_tokens` and lets the provider's default apply, as the DeepSeek driver does.
+ * With no number to reserve, core keeps the conservative one it has always used.
+ */
+const DEFAULT_SUMMARY_RESERVE = 20_000;
+/** Headroom below the window so one turn's growth can't blow past it. */
+const TURN_HEADROOM = 13_000;
+/** Floor for the auto bar, so a tiny window still leaves a usable transcript. */
+const MIN_AUTO_BAR = 20_000;
+/** Share of the window micro-compaction aims at on an ordinary model. */
+const MICRO_SHARE = 0.3;
+/**
+ * Absolute cap on the micro bar.
+ *
+ * A flat share of the window stops making sense once windows get large: at 30% a
+ * 500K model would carry 150K of stale tool output on every turn, which is not a
+ * lean working set on any model and is the user's money on BYOK. The purpose of
+ * micro-compaction is a lean working set, and that purpose is absolute rather than
+ * proportional, so the share gets a ceiling.
+ *
+ * 96K is a judgment call under absent data, chosen so it sits ABOVE the share for
+ * every window we ship today (30% of 256K is 76.8K) and only engages past ~320K.
+ * That is deliberate: no shipped model changes behaviour, and the cap exists for
+ * the longer models it was written for. Revisit it if anyone measures where a
+ * working set actually stops paying for itself.
+ */
+const MICRO_CEILING = 96_000;
+/**
+ * Hard guard: micro never rises above this share of the auto bar.
+ *
+ * The two bars are computed from different inputs, so a large enough reserve on a
+ * small enough window could otherwise push micro past auto and invert them —
+ * clearing tool bodies at the same moment we summarize. This keeps the ordering a
+ * property of the arithmetic rather than something the numbers happen to satisfy.
+ */
+const MICRO_MAX_SHARE_OF_AUTO = 0.6;
+/**
+ * The part of the prompt that is NOT the transcript, measured rather than assumed:
+ * what the provider said the whole prompt was, minus what we measured the transcript
+ * to be on the way out. That remainder is the system prompt, every tool schema, the
+ * working-set block, the relevance map, todos and the governor — none of which the
+ * bars could see when they counted only the transcript.
+ *
+ * Clamped at zero. The transcript estimator is deliberately conservative and can read
+ * HIGHER than the provider's real count; a negative overhead would then push the bars
+ * later, which is precisely the failure this removes.
+ */
+export function measuredOverhead(promptTokens, transcriptTokens) {
+    return Math.max(0, promptTokens - transcriptTokens);
+}
+export function sharpContextWindow(model) {
+    return manifestForModel(model).contextWindow(model);
+}
+/**
+ * How much room to keep free for a compaction summary, as the model's driver
+ * reports it. A driver that declares no buffered ceiling falls back to core's
+ * conservative default rather than reserving nothing.
+ */
+export function summaryReserveFor(model) {
+    const declared = manifestForModel(model).bufferedOutputTokens?.(model);
+    return declared !== undefined && declared > 0 ? declared : DEFAULT_SUMMARY_RESERVE;
+}
+/**
+ * Autocompact bar: summarize once the transcript crosses
+ * (window − summary reserve − turn headroom). For a 256K window reserving 20K
+ * that's 223K.
+ */
+export function autoBarFor(window, summaryReserve = DEFAULT_SUMMARY_RESERVE) {
+    return Math.max(MIN_AUTO_BAR, window - summaryReserve - TURN_HEADROOM);
+}
+/**
+ * Microcompact bar: clear old tool-result bodies well before the autocompact bar,
+ * so the working set stays lean continuously rather than only at the summary point.
+ * A share of the window, capped in absolute terms, and always kept clear of the
+ * auto bar.
+ */
+export function microBarFor(window, summaryReserve = DEFAULT_SUMMARY_RESERVE) {
+    return Math.min(Math.round(window * MICRO_SHARE), MICRO_CEILING, Math.round(autoBarFor(window, summaryReserve) * MICRO_MAX_SHARE_OF_AUTO));
+}
+/**
+ * How far below the auto bar to start telling the user context is filling up.
+ *
+ * A share rather than a fixed number, so it scales with the model the way every other
+ * bar here does: 20K of warning is most of a small window's usable transcript and a
+ * rounding error on a large one. The point is to give roughly a turn or two of notice,
+ * and a turn is proportional to the window.
+ */
+const WARN_SHARE_OF_AUTO = 0.9;
+/**
+ * The point at which the user should be told, once, that a compaction is coming.
+ *
+ * Compaction rewrites the conversation. It is not a failure and it does not need
+ * permission, but arriving with no notice is how a user ends up wondering why the
+ * model forgot the middle of their session.
+ */
+export function warnBarFor(autoBar) {
+    return Math.round(autoBar * WARN_SHARE_OF_AUTO);
+}
+/**
+ * Turn a measured context size into the proactive "a compaction is coming" notice.
+ *
+ * Nothing shows until context crosses the warn bar (`warnBarFor`, 90% of the auto bar), so
+ * an ordinary session never sees it — it appears only for the turn or two of notice before
+ * the summarizing pass rewrites the conversation, which is the one thing a user should not
+ * meet by surprise. The arithmetic is the SAME auto bar the compaction decision runs on, so
+ * what the notice says and what actually happens cannot disagree.
+ */
+export function contextPressure(used, model) {
+    const auto = autoCompactThreshold(model);
+    const percentLeft = auto > 0 ? Math.max(0, Math.round(((auto - used) / auto) * 100)) : 0;
+    return { warn: used >= warnBarFor(auto), percentLeft };
+}
+/** Autocompact bar for a model, anchored to its driver's numbers. */
+export function autoCompactThreshold(model) {
+    return autoBarFor(sharpContextWindow(model), summaryReserveFor(model));
+}
+/** Microcompact bar for a model, anchored to its driver's numbers. */
+export function microCompactThreshold(model) {
+    return microBarFor(sharpContextWindow(model), summaryReserveFor(model));
+}
+/**
+ * How long a provider's prompt cache is assumed to survive between requests.
+ *
+ * 60 minutes, and the size matters more than it looks. This was 5 minutes — Anthropic's
+ * default ephemeral TTL — on the reasoning that anything past it is certainly gone. That
+ * is wrong in the DANGEROUS direction: the number decides when we clear tool bodies, and
+ * clearing rewrites the cached prefix. Guess too SHORT and we destroy an entry that was
+ * still alive, forcing a full-price rewrite that would never have happened. Guess too
+ * long and we merely miss a saving.
+ *
+ * Five minutes is only Anthropic's default tier. Its 1h tier exists, DeepSeek's context
+ * cache persists for hours, and several providers publish nothing at all — so a 5-minute
+ * assumption would have been actively wrong on most of the lineup.
+ *
+ * 60 minutes is the smallest value that is safe everywhere: past an hour no provider in
+ * the lineup is documented to still be holding the prefix, so clearing can only be
+ * reclaiming something already lost. Being wrong is then free rather than expensive.
+ */
+export const CACHE_TTL_MS = (() => {
+    const raw = Number(process.env.MINDWEAVE_CACHE_TTL_MS);
+    return Number.isFinite(raw) && raw > 0 ? raw : 60 * 60 * 1000;
+})();
+/**
+ * Below this share of the micro bar, a cold cache is not worth compacting for. A
+ * session holding almost nothing has almost nothing to save, and clearing tool bodies
+ * early costs real detail.
+ */
+const COLD_COMPACT_FLOOR = 0.5;
+/**
+ * Whether to microcompact because the provider's cache has gone cold, rather than
+ * because context is filling up.
+ *
+ * The normal gate exists for a specific reason: clearing an old tool body REWRITES the
+ * transcript, and the transcript is the cached half of the request, so on a warm cache
+ * it trades a cheap cached read for a full prefix rewrite at 1.25x. That is a loss, and
+ * it is why microcompaction waits for real pressure.
+ *
+ * None of that argument survives the cache expiring. Once the gap since the last call
+ * exceeds the TTL there is no entry left to invalidate — the next request pays to write
+ * the whole prefix no matter what it contains. Clearing stale tool bodies first is then
+ * strictly free, and the tokens never have to be paid for again.
+ *
+ * `lastCallAt` of 0 means no call has been made yet (a fresh session), which is not a
+ * cold cache — it is no cache, with nothing yet worth clearing.
+ */
+export function cacheLikelyCold(lastCallAt, now, used, microBar, ttlMs = CACHE_TTL_MS) {
+    if (lastCallAt <= 0)
+        return false;
+    if (now - lastCallAt <= ttlMs)
+        return false;
+    return used >= microBar * COLD_COMPACT_FLOOR;
+}
+/**
+ * Fraction of the prefix a clear must reclaim to pay for the cache rewrite it causes.
+ *
+ * Clearing an old tool body is not free on a warm cache: it changes the cached prefix,
+ * so the next request rewrites what remains at 1.25x instead of reading it at 0.1x.
+ * Writing out the break-even, with P the prefix and R what the clear reclaims:
+ *
+ *     keep for N steps   = N * 0.1P
+ *     clear, then N-1    = 1.25(P-R) + (N-1) * 0.1(P-R)
+ *     break-even at       N = 11.5 * (P - R) / R
+ *
+ * which is brutal at small R: reclaiming 10% of a 40K prefix needs ~103 further steps
+ * to pay off, 50% needs ~12, and 75% needs ~4. A turn does not have 103 steps. So a
+ * clear that frees a little is not a small win, it is a real loss, and the only clears
+ * worth making on a warm cache are the big ones.
+ *
+ * 0.35 sits where the break-even (~21 steps) is still optimistic but no longer absurd,
+ * and it is deliberately a FRACTION rather than a token count: the thing being paid for
+ * is proportional to the prefix, so the threshold has to be too.
+ */
+const CLEAR_WORTH_FRACTION = 0.35;
+/**
+ * Whether a proposed microcompaction should actually be committed.
+ *
+ * Three ways to say yes, and they are different kinds of reason:
+ *
+ *  - `cold`: the provider's cache has expired, so there is no entry to invalidate and
+ *    the rewrite is free. Any reclaim at all is profit.
+ *  - `urgent`: context is close enough to the autocompact bar that FITTING matters more
+ *    than cost. A request that does not fit cannot be sent at any price.
+ *  - otherwise: only when the clear reclaims enough of the prefix to out-earn the cache
+ *    write it forces, per the arithmetic above.
+ *
+ * Pure, because the alternative is discovering the economics were wrong from a bill.
+ */
+export function clearIsWorthIt(opts) {
+    const reclaimed = opts.before - opts.after;
+    if (reclaimed <= 0)
+        return false;
+    if (opts.cold)
+        return true;
+    // Close to the bar, fitting beats saving: clearing here is what defers a far more
+    // expensive summarization, and a prompt that overflows is not a cost question.
+    if (opts.before >= opts.autoBar * 0.9)
+        return true;
+    return reclaimed >= opts.before * CLEAR_WORTH_FRACTION;
+}
+//# sourceMappingURL=contextWindow.js.map

package/dist/dynamo/contextWindow.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"contextWindow.js","sourceRoot":"","sources":["../../src/dynamo/contextWindow.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,OAAO,EAAE,gBAAgB,EAAE,MAAM,wBAAwB,CAAC;AAE1D;;;;GAIG;AACH,MAAM,uBAAuB,GAAG,MAAM,CAAC;AAEvC,yEAAyE;AACzE,MAAM,aAAa,GAAG,MAAM,CAAC;AAE7B,iFAAiF;AACjF,MAAM,YAAY,GAAG,MAAM,CAAC;AAE5B,yEAAyE;AACzE,MAAM,WAAW,GAAG,GAAG,CAAC;AAExB;;;;;;;;;;;;;;GAcG;AACH,MAAM,aAAa,GAAG,MAAM,CAAC;AAE7B;;;;;;;GAOG;AACH,MAAM,uBAAuB,GAAG,GAAG,CAAC;AAEpC;;;;;;;;;;GAUG;AACH,MAAM,UAAU,gBAAgB,CAAC,YAAoB,EAAE,gBAAwB;IAC7E,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,YAAY,GAAG,gBAAgB,CAAC,CAAC;AACtD,CAAC;AAmBD,MAAM,UAAU,kBAAkB,CAAC,KAAa;IAC9C,OAAO,gBAAgB,CAAC,KAAK,CAAC,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC;AACtD,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAAa;IAC7C,MAAM,QAAQ,GAAG,gBAAgB,CAAC,KAAK,CAAC,CAAC,oBAAoB,EAAE,CAAC,KAAK,CAAC,CAAC;IACvE,OAAO,QAAQ,KAAK,SAAS,IAAI,QAAQ,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,uBAAuB,CAAC;AACrF,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,UAAU,CAAC,MAAc,EAAE,cAAc,GAAG,uBAAuB;IACjF,OAAO,IAAI,CAAC,GAAG,CAAC,YAAY,EAAE,MAAM,GAAG,cAAc,GAAG,aAAa,CAAC,CAAC;AACzE,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,WAAW,CAAC,MAAc,EAAE,cAAc,GAAG,uBAAuB;IAClF,OAAO,IAAI,CAAC,GAAG,CACb,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,WAAW,CAAC,EAChC,aAAa,EACb,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,MAAM,EAAE,cAAc,CAAC,GAAG,uBAAuB,CAAC,CACzE,CAAC;AACJ,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,kBAAkB,GAAG,GAAG,CAAC;AAE/B;;;;;;GAMG;AACH,MAAM,UAAU,UAAU,CAAC,OAAe;IACxC,OAAO,IAAI,CAAC,KAAK,CAAC,OAAO,GAAG,kBAAkB,CAAC,CAAC;AAClD,CAAC;AAcD;;;;;;;;GAQG;AACH,MAAM,UAAU,eAAe,CAAC,IAAY,EAAE,KAAa;IACzD,MAAM,IAAI,GAAG,oBAAoB,CAAC,KAAK,CAAC,CAAC;IACzC,MAAM,WAAW,GAAG,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACzF,OAAO,EAAE,IAAI,EAAE,IAAI,IAAI,UAAU,CAAC,IAAI,CAAC,EAAE,WAAW,EAAE,CAAC;AACzD,CAAC;AAED,qEAAqE;AACrE,MAAM,UAAU,oBAAoB,CAAC,KAAa;IAChD,OAAO,UAAU,CAAC,kBAAkB,CAAC,KAAK,CAAC,EAAE,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC;AACzE,CAAC;AAED,sEAAsE;AACtE,MAAM,UAAU,qBAAqB,CAAC,KAAa;IACjD,OAAO,WAAW,CAAC,kBAAkB,CAAC,KAAK,CAAC,EAAE,iBAAiB,CAAC,KAAK,CAAC,CAAC,CAAC;AAC1E,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,MAAM,YAAY,GAAG,CAAC,GAAG,EAAE;IAChC,MAAM,GAAG,GAAG,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC,CAAC;IACvD,OAAO,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAChE,CAAC,CAAC,EAAE,CAAC;AAEL;;;;GAIG;AACH,MAAM,kBAAkB,GAAG,GAAG,CAAC;AAE/B;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,UAAU,eAAe,CAC7B,UAAkB,EAClB,GAAW,EACX,IAAY,EACZ,QAAgB,EAChB,QAAgB,YAAY;IAE5B,IAAI,UAAU,IAAI,CAAC;QAAE,OAAO,KAAK,CAAC;IAClC,IAAI,GAAG,GAAG,UAAU,IAAI,KAAK;QAAE,OAAO,KAAK,CAAC;IAC5C,OAAO,IAAI,IAAI,QAAQ,GAAG,kBAAkB,CAAC;AAC/C,CAAC;AAED;;;;;;;;;;;;;;;;;;;GAmBG;AACH,MAAM,oBAAoB,GAAG,IAAI,CAAC;AAElC;;;;;;;;;;;;;GAaG;AACH,MAAM,UAAU,cAAc,CAAC,IAS9B;IACC,MAAM,SAAS,GAAG,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC;IAC3C,IAAI,SAAS,IAAI,CAAC;QAAE,OAAO,KAAK,CAAC;IACjC,IAAI,IAAI,CAAC,IAAI;QAAE,OAAO,IAAI,CAAC;IAC3B,kFAAkF;IAClF,+EAA+E;IAC/E,IAAI,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,OAAO,GAAG,GAAG;QAAE,OAAO,IAAI,CAAC;IACnD,OAAO,SAAS,IAAI,IAAI,CAAC,MAAM,GAAG,oBAAoB,CAAC;AACzD,CAAC"}