mindweave 3.4.1 → 3.4.2

package/dist/dynamo/contextWindow.js

@@ -246,44 +246,31 @@ export const COLD_COMPACT_MIN_IDLE_MS = 15 * 60 * 1000;
 /** After a compaction for a cold cache, wait this long before doing it again. */
 export const COLD_COMPACT_COOLDOWN_MS = 30 * 60 * 1000;
 /**
- * Fraction of the prefix a clear must reclaim to pay for the cache rewrite it causes.
+ * How many further calls a clear is assumed to keep paying off for.
  *
- * 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:
+ * Clearing an old tool body or image is not free on a warm cache: the provider holds the
+ * conversation as a prefix, a change inside it invalidates everything after the change, and
+ * the next request writes that tail again at the full input price (plus a write premium, where
+ * the provider charges one) where it would have read it at the cached price. Written out, with
+ * T the tokens from the first changed message to the end, R the tokens the clear reclaims, h the
+ * cached read price and w the write price:
  *
- *     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
+ *     paid once, on the next call   = T * (w - h)
+ *     saved on every later call     = R * h
+ *     worth it when                   R * h * horizon >= T * (w - h)
  *
- * 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.
+ * The prices decide it, and they differ by an order of magnitude between providers: a cached read
+ * costs 10% of input on one, 2% on another, 20% on a third. At 2% the saving per call is so
+ * small that a clear which rewrites a 25K tail to reclaim 26K needs about 47 further calls to
+ * break even; at 10% the same clear needs about 11. The same clear is right on one provider and
+ * a loss on another, which a fixed fraction of the prefix could not tell apart.
  *
- * 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.
- *
- * MEASURED against the derivation above (61 real sessions, 2026-09-16), and it is too
- * pessimistic. Do not tighten this on the strength of the arithmetic:
- *
- *   - A clear does not rewrite the whole remaining prefix, only what follows the first
- *     cleared result. Real clears cost 6-12K extra uncached tokens on a ~45K prompt,
- *     not the ~43K that `1.25(P-R)` charges.
- *   - Sessions run far past one turn. Clears were followed by 37-82 more calls on
- *     average, so savings keep accruing long after the break-even horizon of one turn.
- *   - The 0.1 / 1.25 ratios are one provider's. DeepSeek reads cache at 2% of input
- *     with no write premium; GLM Flash at 20%.
- *
- * Net of cache rewrites AND the re-reads clearing causes (about 18% of file reads re-read
- * something that had been cleared), clearing saved 4.3% of spend on DeepSeek V4 Flash and
- * 10.3% on GLM 5.3 Flash, and 28 of 32 individual clears paid for themselves, including
- * most below this threshold. The policy is left as is because it is already net positive
- * and loosening it trades a few percent of saving for more re-reads, which is not yet
- * measured.
+ * A long run is 30 to 80 calls after a clear, so 40 is where it is still a fair bet. Past it
+ * the clear is declined and the context is allowed to fill: carrying the extra tokens is a
+ * cached read, cheap, and the clear is made anyway when the cache goes cold or the autocompact
+ * bar is close, both of which cost nothing extra.
  */
-const CLEAR_WORTH_FRACTION = 0.35;
+export const CLEAR_HORIZON_STEPS = 40;
 /**
  * Whether a proposed microcompaction should actually be committed.
  *
@@ -293,8 +280,8 @@ const CLEAR_WORTH_FRACTION = 0.35;
  *    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.
+ *  - otherwise: only when what it saves over the next calls out-earns the cache rewrite
+ *    it forces, at this provider's prices (see `CLEAR_HORIZON_STEPS`).
  *
  * Pure, because the alternative is discovering the economics were wrong from a bill.
  */
@@ -308,6 +295,11 @@ export function clearIsWorthIt(opts) {
     // 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;
+    const write = opts.price.cacheWrite ?? opts.price.cacheMiss;
+    const paidOnce = opts.rewritten * Math.max(0, write - opts.price.cacheHit);
+    // A provider that charges the same for a hit as for a miss has nothing to lose by a rewrite.
+    if (paidOnce <= 0)
+        return true;
+    return reclaimed * opts.price.cacheHit * (opts.horizon ?? CLEAR_HORIZON_STEPS) >= paidOnce;
 }
 //# sourceMappingURL=contextWindow.js.map

package/dist/dynamo/contextWindow.js.map

@@ -1 +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;;;;;;;;;;;;;;;;;;;;GAoBG;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;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,KAAa;IAC5C,MAAM,GAAG,GAAG,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC,CAAC;IACvD,IAAI,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC;QAAE,OAAO,GAAG,CAAC;IAChD,OAAO,gBAAgB,CAAC,KAAK,CAAC,CAAC,eAAe,EAAE,CAAC,KAAK,CAAC,IAAI,YAAY,CAAC;AAC1E,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAAa;IAC7C,MAAM,KAAK,GAAG,gBAAgB,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACnD,OAAO,KAAK,CAAC,SAAS,GAAG,CAAC,IAAI,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,SAAS,CAAC;AACjE,CAAC;AAiBD,MAAM,UAAU,WAAW,CAAC,CAM3B;IACC,MAAM,IAAI,GAAG,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;IACtC,IAAI,IAAI,IAAI,IAAI,GAAG,CAAC;QAAE,OAAO,CAAC,CAAC,GAAG,GAAG,IAAI,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC;IACtE,OAAO,CAAC,CAAC,WAAW,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC;AACtF,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,MAAM,CAAC,MAAmB;IACxC,OAAO,MAAM,KAAK,MAAM,IAAI,MAAM,KAAK,UAAU,CAAC;AACpD,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,eAAe,CAAC,MAAmB,EAAE,IAAY,EAAE,QAAgB;IACjF,OAAO,MAAM,CAAC,MAAM,CAAC,IAAI,IAAI,IAAI,QAAQ,GAAG,kBAAkB,CAAC;AACjE,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,oBAAoB,CAAC,CAUpC;IACC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,UAAU;QAAE,OAAO,KAAK,CAAC;IACrD,iGAAiG;IACjG,IAAI,CAAC,CAAC,MAAM,KAAK,MAAM,IAAI,CAAC,CAAC,MAAM,GAAG,wBAAwB;QAAE,OAAO,KAAK,CAAC;IAC7E,IAAI,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ;QAAE,OAAO,KAAK,CAAC;IACtC,OAAO,CAAC,CAAC,CAAC,CAAC,iBAAiB,IAAI,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,iBAAiB,GAAG,wBAAwB,CAAC,CAAC;AAC1F,CAAC;AAED,0GAA0G;AAC1G,MAAM,UAAU,kBAAkB,CAAC,MAAmB,EAAE,KAAa,EAAE,MAAc;IACnF,MAAM,GAAG,GACP,MAAM,KAAK,UAAU;QACnB,CAAC,CAAC,GAAG,KAAK,0CAA0C;QACpD,CAAC,CAAC,4DAA4D,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC,eAAe,CAAC;IAC1H,OAAO,4BAA4B,GAAG,wGAAwG,CAAC;AACjJ,CAAC;AAED,+GAA+G;AAC/G,MAAM,CAAC,MAAM,wBAAwB,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAEvD,iFAAiF;AACjF,MAAM,CAAC,MAAM,wBAAwB,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAEvD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;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"}
+{"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;;;;;;;;;;;;;;;;;;;;GAoBG;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;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,KAAa;IAC5C,MAAM,GAAG,GAAG,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC,CAAC;IACvD,IAAI,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC;QAAE,OAAO,GAAG,CAAC;IAChD,OAAO,gBAAgB,CAAC,KAAK,CAAC,CAAC,eAAe,EAAE,CAAC,KAAK,CAAC,IAAI,YAAY,CAAC;AAC1E,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAAa;IAC7C,MAAM,KAAK,GAAG,gBAAgB,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACnD,OAAO,KAAK,CAAC,SAAS,GAAG,CAAC,IAAI,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,SAAS,CAAC;AACjE,CAAC;AAiBD,MAAM,UAAU,WAAW,CAAC,CAM3B;IACC,MAAM,IAAI,GAAG,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;IACtC,IAAI,IAAI,IAAI,IAAI,GAAG,CAAC;QAAE,OAAO,CAAC,CAAC,GAAG,GAAG,IAAI,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC;IACtE,OAAO,CAAC,CAAC,WAAW,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC;AACtF,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,MAAM,CAAC,MAAmB;IACxC,OAAO,MAAM,KAAK,MAAM,IAAI,MAAM,KAAK,UAAU,CAAC;AACpD,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,eAAe,CAAC,MAAmB,EAAE,IAAY,EAAE,QAAgB;IACjF,OAAO,MAAM,CAAC,MAAM,CAAC,IAAI,IAAI,IAAI,QAAQ,GAAG,kBAAkB,CAAC;AACjE,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,oBAAoB,CAAC,CAUpC;IACC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,UAAU;QAAE,OAAO,KAAK,CAAC;IACrD,iGAAiG;IACjG,IAAI,CAAC,CAAC,MAAM,KAAK,MAAM,IAAI,CAAC,CAAC,MAAM,GAAG,wBAAwB;QAAE,OAAO,KAAK,CAAC;IAC7E,IAAI,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ;QAAE,OAAO,KAAK,CAAC;IACtC,OAAO,CAAC,CAAC,CAAC,CAAC,iBAAiB,IAAI,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,iBAAiB,GAAG,wBAAwB,CAAC,CAAC;AAC1F,CAAC;AAED,0GAA0G;AAC1G,MAAM,UAAU,kBAAkB,CAAC,MAAmB,EAAE,KAAa,EAAE,MAAc;IACnF,MAAM,GAAG,GACP,MAAM,KAAK,UAAU;QACnB,CAAC,CAAC,GAAG,KAAK,0CAA0C;QACpD,CAAC,CAAC,4DAA4D,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC,eAAe,CAAC;IAC1H,OAAO,4BAA4B,GAAG,wGAAwG,CAAC;AACjJ,CAAC;AAED,+GAA+G;AAC/G,MAAM,CAAC,MAAM,wBAAwB,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAEvD,iFAAiF;AACjF,MAAM,CAAC,MAAM,wBAAwB,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAEvD;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAG,EAAE,CAAC;AAEtC;;;;;;;;;;;;;GAaG;AACH,MAAM,UAAU,cAAc,CAAC,IAmB9B;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,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,IAAI,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC;IAC5D,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;IAC3E,6FAA6F;IAC7F,IAAI,QAAQ,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC/B,OAAO,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,GAAG,CAAC,IAAI,CAAC,OAAO,IAAI,mBAAmB,CAAC,IAAI,QAAQ,CAAC;AAC7F,CAAC"}

package/dist/dynamo/engine.js

@@ -53,10 +53,10 @@ import { directoryNotesFor } from "../memory/projectNotes.js";
 import { rippleCheck } from "../tools/editRipple.js";
 import { fullReadPaths, visibleReadRanges } from "../memory/presence.js";
 import { RESTORE_MAX_TOKENS_PER_FILE, renderRestored, restoreBudgetFor, selectForRestore, } from "../memory/restore.js";
-import { KEEP_LAST_N, KEEP_LAST_N_BOUNDARY, CLEARED_STUB, summaryRequest, SUMMARY_SYSTEM_PROMPT, dropOldestRounds, estimateEntriesTokens, estimateTokens, estimateTokensForChars, formatTranscriptForSummary, isContinuation, microcompact, spliceSummary, summaryForWire, summaryIsHistory, usableSummary, } from "../memory/compaction.js";
+import { KEEP_LAST_N, KEEP_LAST_N_BOUNDARY, CLEARED_STUB, summaryRequest, SUMMARY_SYSTEM_PROMPT, dropOldestRounds, estimateEntriesTokens, estimateTokens, estimateTokensForChars, formatTranscriptForSummary, isContinuation, isRecapStub, microcompact, RECAP_STUB_NUDGE, spliceSummary, summaryForWire, summaryWireIsHistory, hasUnsettledHistory, settleSummaries, firstChangedEntry, SUMMARY_HISTORY_NOTE, usableSummary, } from "../memory/compaction.js";
 import { MAX_PARALLEL_CALLS, partitionCalls, runLimited } from "./toolBatches.js";
 import { loadPlanArtifact, completePlanArtifact, renderPlanBlock, planDivergenceStop } from "./planArtifact.js";
-import { autoCompactThreshold, microCompactThreshold, cacheLikelyCold, cacheStatus, cacheLifetimeFor, cacheIsDiscounted, shouldCompactForCold, coldCompactionLine, clearIsWorthIt, measuredOverhead, sharpContextWindow, } from "./contextWindow.js";
+import { autoCompactThreshold, microCompactThreshold, cacheLikelyCold, isCold, cacheStatus, cacheLifetimeFor, cacheIsDiscounted, shouldCompactForCold, coldCompactionLine, clearIsWorthIt, measuredOverhead, sharpContextWindow, } from "./contextWindow.js";
 import { todoReminderDue, todoReminderText } from "./todoReminder.js";
 import { refreshSessionMemory, sessionMemoryDue, settleSessionMemory } from "../memory/sessionMemory.js";
 import { compactFromSessionMemory } from "../memory/sessionMemoryCompact.js";
@@ -187,8 +187,12 @@ directoryNotes = [],
 /** Where the previous turn left the shell, when this turn started back at the root. */
 cwdResetFrom = "", 
 /** A running Marathon's standing block (see marathonPrompt.ts), or "". */
-marathon = "") {
+marathon = "", 
+/** What stands in for rewriting a compaction summary the user has written past (see SUMMARY_HISTORY_NOTE), or "". */
+summaryNote = "") {
     const parts = [];
+    if (summaryNote)
+        parts.push(summaryNote);
     // Each turn starts at the project root, and a model that `cd`-ed into a subfolder last
     // turn does not know that unless it is told. It was not, and a real session ran
     // `cargo run` from the root expecting the subfolder: "could not find Cargo.toml".
@@ -540,7 +544,7 @@ directoryNotes = []) {
             // A summary is sent as background once the user has written after it, so its
             // hand-off "next step" cannot keep overriding what they actually asked.
             const said = e.role === "summary"
-                ? summaryForWire(e.content, summaryIsHistory(e, session.transcript.slice(i + 1)))
+                ? summaryForWire(e.content, summaryWireIsHistory(e, session.transcript.slice(i + 1)))
                 : e.arrival ? arrivalNote(e.arrival, e.content) : e.content;
             if (refs && refs.length > 0) {
                 const images = [];
@@ -631,7 +635,7 @@ directoryNotes = []) {
         // "keep working toward the goal" would stop being an independent check.
         session.marathon?.status === "running" && (session.toolContext.subagentDepth ?? 0) === 0
             ? marathonBlock(session.marathon.goal, session.marathon.turnsSpent === 0 ? "plan" : "run")
-            : ""),
+            : "", hasUnsettledHistory(session.transcript) ? SUMMARY_HISTORY_NOTE : ""),
         tools,
         model: session.modelConfig,
         sessionId: session.id,
@@ -1287,6 +1291,7 @@ async function respondTurn(session, options = {}) {
                 // shorter than it was.
                 session.lastCallAt = Date.now();
                 (session.lastCallByModel ??= {})[session.modelConfig.model] = session.lastCallAt;
+                session.lastRequest = { request, sent: session.transcript.slice(0, request.messages.length) };
                 result = await streamModel(request, options);
             }
             catch (error) {
@@ -1409,10 +1414,11 @@ async function respondTurn(session, options = {}) {
                 // what the user sees. ONE retry — a gate that can fire twice is a loop.
                 if (!replyRegated) {
                     const fault = replyFault(content, mutatedThisTurn);
-                    if (fault) {
+                    const copiedStub = isRecapStub(content);
+                    if (fault || copiedStub) {
                         replyRegated = true;
                         overlongReplyAt = session.transcript.length - 1; // the draft pushed just above
-                        session.transcript.push({ role: "user", content: replyRewrite(fault), synthetic: true });
+                        session.transcript.push({ role: "user", content: fault ? replyRewrite(fault) : RECAP_STUB_NUDGE, synthetic: true });
                         // The draft has been streaming into the UI's buffer, unrendered. Drop it, or
                         // the rewrite would append to it and the user would read both.
                         options.onEvent?.({ type: "replyReset" });
@@ -2473,7 +2479,7 @@ async function sweepSessionMemory(session, options, mode) {
         options.onEvent?.({ type: "usage", ...usage });
         countUsage(session, usage, options);
         recordAuxCall(session, usage, "notes");
-    }).then(async (wrote) => {
+    }, (unseenTokens) => sharedNotesPrefix(session, unseenTokens)).then(async (wrote) => {
         if (wrote)
             await options.persist?.(); // durable: the notes sidecar is written by the persister
         return wrote;
@@ -2483,6 +2489,35 @@ async function sweepSessionMemory(session, options, mode) {
     else
         void run.catch(() => { });
 }
+/**
+ * The main request the notes writer may build on, or null when it should send its own.
+ *
+ * Building on it means the provider reads the whole conversation from the cache the last call just
+ * filled, at the cached price, instead of being sent the unseen part at full price. That is only
+ * cheaper while the copy is there (this model was called within its cache lifetime), the request still
+ * describes the transcript (it ended on an entry that is still in the same place), and the cached read
+ * of the whole prefix costs less than the full-price read of what is new.
+ */
+async function sharedNotesPrefix(session, unseenTokens) {
+    const sent = session.lastRequest;
+    const model = session.modelConfig.model;
+    if (!sent || sent.request.model?.model !== model || session.transcript.length === 0)
+        return null;
+    // The provider holds the beginning of the conversation as it was sent. It only counts as the beginning
+    // of this one if nothing in it has been rewritten since (a clear, a compaction, a settled summary).
+    if (sent.sent.length > session.transcript.length || sent.sent.some((e, i) => e !== session.transcript[i]))
+        return null;
+    const status = cacheStatus({ now: Date.now(), model, lastByModel: session.lastCallByModel, ttlMs: cacheLifetimeFor(model) });
+    if (status !== "warm" || !cacheIsDiscounted(model))
+        return null;
+    const price = manifestForModel(model).price(model);
+    const prefixTokens = estimateEntriesTokens(sent.sent) + (session.contextOverhead?.model === model ? session.contextOverhead.tokens : 0);
+    if (prefixTokens * price.cacheHit > unseenTokens * price.cacheMiss)
+        return null;
+    // Same system prompt, tools and model as the main request; the conversation as it is now.
+    const { messages } = buildRequest(session, [], await loadImagePayloads(session));
+    return { request: { ...sent.request, messages }, entries: session.transcript.length, last: session.transcript[session.transcript.length - 1] };
+}
 /**
  * How much of the context window is in use, in tokens.
  *
@@ -2590,11 +2625,18 @@ async function maybeCompact(session, options) {
         const proposed = microcompact(session.transcript).entries;
         const before = estimateEntriesTokens(session.transcript);
         const after = estimateEntriesTokens(proposed);
-        if (clearIsWorthIt({ before, after, cold, autoBar })) {
+        // What the provider would have to write again: everything from the first message the clear changes.
+        const rewritten = estimateEntriesTokens(proposed.slice(firstChangedEntry(session.transcript, proposed)));
+        if (clearIsWorthIt({ before, after, cold, autoBar, rewritten, price: manifestForModel(model).price(model) })) {
+            // The prefix is being rewritten anyway, so a summary the user has written past is settled in the same move.
             session.transcript = keepClearedOriginals(session, proposed);
+            session.transcript = settleSummaries(session.transcript);
             // Silent by design — trimming stale context is background machinery.
         }
     }
+    // With the provider's copy gone there is nothing to lose by it either, clear or no clear.
+    if (isCold(status))
+        session.transcript = settleSummaries(session.transcript);
     // The history is about to be sent at full price whatever it contains, so this is the cheapest
     // moment to shrink it, and the session notes make that free: no model call. Only the top-level
     // agent; a sub-agent starts with no cache of its own and its work is short.