@buoy-gg/agent-core 7.0.35 → 7.0.36

package/lib/commonjs/engine/historyBudget.js

@@ -0,0 +1,260 @@
+"use strict";
+
+Object.defineProperty(exports, "__esModule", {
+  value: true
+});
+exports.MAX_REQUEST_CHARS = exports.MAX_HISTORY_CHARS = void 0;
+exports.budgetForRequest = budgetForRequest;
+exports.compressConsumed = compressConsumed;
+exports.messageSize = messageSize;
+exports.trimHistory = trimHistory;
+exports.trimHistoryReport = trimHistoryReport;
+/**
+ * How much conversation the model is allowed to be sent, and what gets cut.
+ *
+ * Lives here, not in `session.ts`, because it is needed in TWO places and used
+ * to run in only one. The session trims once per turn, before the loop starts
+ * — and a turn is not one request. A twelve-step investigation appends an
+ * assistant message and a tool-results message per step, each result capped at
+ * 24,000 characters, so a long read chain could add well over a hundred
+ * thousand characters AFTER the only check had already happened and walk
+ * straight into the provider's context limit. The engine now applies the same
+ * budget before every request; the session still applies it when a turn opens.
+ */
+
+/**
+ * A long QA session must not die on the provider's context limit with an
+ * opaque error, so the history is kept under a budget.
+ *
+ * TWO STAGES, AND THE ORDER IS THE POINT. Dropping the oldest rounds was the
+ * only stage, and in a QA session the oldest round is usually the SETUP — the
+ * override that was armed, the store that was written, the user someone is
+ * impersonating. Drop it and the model can no longer explain the screen it is
+ * looking at, or that a "wrong" response is its own mock. The reply gets worse
+ * in exactly the sessions long enough to need trimming.
+ *
+ * So: COMPRESS first, DROP only if that was not enough.
+ *
+ * A tool result is "consumed" once the model has produced an assistant message
+ * after seeing it — it has already been reasoned over, and what remains is
+ * bulk. Replacing it with a marker keeps the round, its decision and its
+ * ordering while giving back nearly all the bytes. Three exemptions carry
+ * their weight:
+ *
+ * - Errors are never compressed. A failure is the diagnostic value.
+ * - The newest tool-results message is never compressed. The model is about
+ *   to answer from it.
+ * - The last read before a WRITE is never compressed: that is the shape the
+ *   model matched its edit to, and it is the thing it re-reads when the write
+ *   is refused and it has to correct itself.
+ *
+ * Only if the history is still over budget do whole rounds get dropped, as
+ * before — a round starts at a user message, so the assistant/tool-result
+ * pairing every provider requires stays intact.
+ */
+const MAX_HISTORY_CHARS = exports.MAX_HISTORY_CHARS = 150_000;
+/**
+ * The ceiling on a WHOLE request: tools, system prompt, messages and the room
+ * reserved for the reply. ~100k tokens, which fits the smallest context Ask
+ * Buoy is pointed at in practice (a 128k-token gateway) with headroom, and
+ * leaves Anthropic's 200k far from the limit.
+ */
+const MAX_REQUEST_CHARS = exports.MAX_REQUEST_CHARS = 400_000;
+const MIN_KEEP_MESSAGES = 4;
+/** Below this a marker costs more than the payload it replaces. */
+const MIN_COMPRESS_CHARS = 400;
+function messageSize(m) {
+  try {
+    return JSON.stringify(m).length;
+  } catch {
+    return 0;
+  }
+}
+
+/**
+ * Replace consumed tool results with a marker. See the note above
+ * MAX_HISTORY_CHARS for what "consumed" means and why the exemptions exist.
+ */
+function compressConsumed(messages) {
+  // The newest tool-results message: the model is about to answer from it.
+  let newest = -1;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "tool-results") {
+      newest = i;
+      break;
+    }
+  }
+  if (newest <= 0) return messages;
+
+  // The reads a later write depended on. An assistant turn that called a
+  // write was matching a shape it read in the round before, and that is the
+  // read it goes back to when the write is refused.
+  const beforeWrite = new Set();
+  for (let i = 0; i < messages.length; i++) {
+    const m = messages[i];
+    if (m.role !== "assistant" || !m.toolCalls?.length) continue;
+    if (!m.toolCalls.some(c => WRITE_ISH.test(JSON.stringify(c.input ?? {}) + c.name))) continue;
+    for (let j = i - 1; j >= 0 && j >= i - 3; j--) {
+      if (messages[j].role === "tool-results") {
+        beforeWrite.add(j);
+        break;
+      }
+    }
+  }
+  let changed = false;
+  const out = messages.map((m, i) => {
+    if (m.role !== "tool-results" || i >= newest || beforeWrite.has(i)) return m;
+    let touched = false;
+    const results = m.results.map(r => {
+      if (r.isError || r.content.length < MIN_COMPRESS_CHARS) return r;
+      if (r.content.startsWith("[earlier result")) return r;
+      touched = true;
+      return {
+        ...r,
+        content: `[earlier result — ${r.content.length} characters, already read. Ask again if you need it.]`
+      };
+    });
+    if (!touched) return m;
+    changed = true;
+    return {
+      ...m,
+      results
+    };
+  });
+  return changed ? out : messages;
+}
+
+/**
+ * Heuristic for "this call changed something". Deliberately loose: the cost of
+ * a false positive is one uncompressed read, the cost of a false negative is
+ * the model losing the shape it was editing.
+ */
+const WRITE_ISH = /set|write|dispatch|navigate|tap|override|restore|delete|remove|clear|save|run|impersonat|reload/i;
+
+/**
+ * Trim, and say how many whole rounds were dropped — the number the UI turns
+ * into "older messages were trimmed from the agent's memory".
+ *
+ * A round is a user message and everything up to the next one. Once a drop
+ * starts it ALWAYS runs to the next user boundary: the inner loop used to stop
+ * at `MIN_KEEP_MESSAGES` as well, so a history could come out starting with an
+ * assistant or tool-results message — which every provider rejects. The
+ * minimum-keep guard belongs on STARTING a drop, not on finishing one.
+ */
+function trimHistoryReport(messages) {
+  let total = 0;
+  for (const m of messages) total += messageSize(m);
+  if (total <= MAX_HISTORY_CHARS) return {
+    messages,
+    droppedRounds: 0
+  };
+
+  // Compress before dropping — a compressed round still explains the screen.
+  const compressed = compressConsumed(messages);
+  let after = 0;
+  for (const m of compressed) after += messageSize(m);
+  if (after <= MAX_HISTORY_CHARS) return {
+    messages: compressed,
+    droppedRounds: 0
+  };
+  total = after;
+  const out = [...compressed];
+  let droppedRounds = 0;
+  while (out.length > MIN_KEEP_MESSAGES && total > MAX_HISTORY_CHARS) {
+    total -= messageSize(out[0]);
+    out.shift();
+    while (out.length > 0 && out[0].role !== "user") {
+      total -= messageSize(out[0]);
+      out.shift();
+    }
+    droppedRounds += 1;
+  }
+  return {
+    messages: out,
+    droppedRounds
+  };
+}
+
+/** Exported for the history tests; not part of the package surface. */
+function trimHistory(messages) {
+  return trimHistoryReport(messages).messages;
+}
+
+/**
+ * Trim the working history of a turn that is still running.
+ *
+ * Differs from `trimHistoryReport` in one way that matters: the CURRENT round
+ * is never touched. A turn's own user message and the reads it has taken since
+ * are the thing being answered, so compressing or dropping them would make the
+ * request cheaper and the answer wrong. Everything before the last user
+ * message is fair game, in the same order as always — compress, then drop.
+ *
+ * `reserve` is what the request costs BEFORE any messages: the tool
+ * definitions (~8k tokens for a typical install, ~23k with everything
+ * installed), the system prompt, and the reply the model is allowed to write.
+ * Budgeting messages alone would leave the biggest fixed cost out of the sum.
+ */
+function budgetForRequest(messages, reserve) {
+  /**
+   * `MAX_HISTORY_CHARS` is a MESSAGES budget and always was — subtracting the
+   * reserve from it is wrong by a wide margin. The tool block alone is ~94,000
+   * characters with every tool installed, and the system prompt another
+   * ~16,000, so `150,000 - reserve` left about 24,000 characters for the
+   * conversation and dropped every earlier round of a two-turn chat. Caught by
+   * the bank, which is exactly what it is for.
+   *
+   * So the messages budget is the smaller of the two ceilings: the one that
+   * has always applied, and whatever is left of the whole-request ceiling once
+   * the fixed cost is counted. In an ordinary install the first wins and
+   * nothing changes; the second only bites when the tools are huge.
+   */
+  const budget = Math.min(MAX_HISTORY_CHARS, MAX_REQUEST_CHARS - reserve);
+  let total = 0;
+  for (const m of messages) total += messageSize(m);
+  if (total <= budget) return {
+    messages,
+    droppedRounds: 0
+  };
+
+  // Where the current round begins. Nothing from here on is eligible.
+  let currentRound = -1;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "user") {
+      currentRound = i;
+      break;
+    }
+  }
+  if (currentRound <= 0) return {
+    messages,
+    droppedRounds: 0
+  };
+  const earlier = messages.slice(0, currentRound);
+  const current = messages.slice(currentRound);
+  let currentSize = 0;
+  for (const m of current) currentSize += messageSize(m);
+  const compressed = compressConsumed(earlier);
+  let earlierSize = 0;
+  for (const m of compressed) earlierSize += messageSize(m);
+  if (earlierSize + currentSize <= budget) {
+    return {
+      messages: [...compressed, ...current],
+      droppedRounds: 0
+    };
+  }
+  const out = [...compressed];
+  let droppedRounds = 0;
+  while (out.length > 0 && earlierSize + currentSize > budget) {
+    earlierSize -= messageSize(out[0]);
+    out.shift();
+    while (out.length > 0 && out[0].role !== "user") {
+      earlierSize -= messageSize(out[0]);
+      out.shift();
+    }
+    droppedRounds += 1;
+  }
+  return {
+    messages: [...out, ...current],
+    droppedRounds
+  };
+}
+//# sourceMappingURL=historyBudget.js.map

package/lib/commonjs/engine/historyBudget.js.map

@@ -0,0 +1 @@
+{"version":3,"names":["MAX_HISTORY_CHARS","exports","MAX_REQUEST_CHARS","MIN_KEEP_MESSAGES","MIN_COMPRESS_CHARS","messageSize","m","JSON","stringify","length","compressConsumed","messages","newest","i","role","beforeWrite","Set","toolCalls","some","c","WRITE_ISH","test","input","name","j","add","changed","out","map","has","touched","results","r","isError","content","startsWith","trimHistoryReport","total","droppedRounds","compressed","after","shift","trimHistory","budgetForRequest","reserve","budget","Math","min","currentRound","earlier","slice","current","currentSize","earlierSize"],"sourceRoot":"../../../src","sources":["engine/historyBudget.ts"],"mappings":";;;;;;;;;;;AAAA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;;AAGA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACO,MAAMA,iBAAiB,GAAAC,OAAA,CAAAD,iBAAA,GAAG,OAAO;AACxC;AACA;AACA;AACA;AACA;AACA;AACO,MAAME,iBAAiB,GAAAD,OAAA,CAAAC,iBAAA,GAAG,OAAO;AACxC,MAAMC,iBAAiB,GAAG,CAAC;AAC3B;AACA,MAAMC,kBAAkB,GAAG,GAAG;AAEvB,SAASC,WAAWA,CAACC,CAAe,EAAU;EACnD,IAAI;IACF,OAAOC,IAAI,CAACC,SAAS,CAACF,CAAC,CAAC,CAACG,MAAM;EACjC,CAAC,CAAC,MAAM;IACN,OAAO,CAAC;EACV;AACF;;AAEA;AACA;AACA;AACA;AACO,SAASC,gBAAgBA,CAACC,QAAwB,EAAkB;EACzE;EACA,IAAIC,MAAM,GAAG,CAAC,CAAC;EACf,KAAK,IAAIC,CAAC,GAAGF,QAAQ,CAACF,MAAM,GAAG,CAAC,EAAEI,CAAC,IAAI,CAAC,EAAEA,CAAC,EAAE,EAAE;IAC7C,IAAIF,QAAQ,CAACE,CAAC,CAAC,CAAEC,IAAI,KAAK,cAAc,EAAE;MACxCF,MAAM,GAAGC,CAAC;MACV;IACF;EACF;EACA,IAAID,MAAM,IAAI,CAAC,EAAE,OAAOD,QAAQ;;EAEhC;EACA;EACA;EACA,MAAMI,WAAW,GAAG,IAAIC,GAAG,CAAS,CAAC;EACrC,KAAK,IAAIH,CAAC,GAAG,CAAC,EAAEA,CAAC,GAAGF,QAAQ,CAACF,MAAM,EAAEI,CAAC,EAAE,EAAE;IACxC,MAAMP,CAAC,GAAGK,QAAQ,CAACE,CAAC,CAAE;IACtB,IAAIP,CAAC,CAACQ,IAAI,KAAK,WAAW,IAAI,CAACR,CAAC,CAACW,SAAS,EAAER,MAAM,EAAE;IACpD,IAAI,CAACH,CAAC,CAACW,SAAS,CAACC,IAAI,CAAEC,CAAC,IAAKC,SAAS,CAACC,IAAI,CAACd,IAAI,CAACC,SAAS,CAACW,CAAC,CAACG,KAAK,IAAI,CAAC,CAAC,CAAC,GAAGH,CAAC,CAACI,IAAI,CAAC,CAAC,EAAE;IACtF,KAAK,IAAIC,CAAC,GAAGX,CAAC,GAAG,CAAC,EAAEW,CAAC,IAAI,CAAC,IAAIA,CAAC,IAAIX,CAAC,GAAG,CAAC,EAAEW,CAAC,EAAE,EAAE;MAC7C,IAAIb,QAAQ,CAACa,CAAC,CAAC,CAAEV,IAAI,KAAK,cAAc,EAAE;QACxCC,WAAW,CAACU,GAAG,CAACD,CAAC,CAAC;QAClB;MACF;IACF;EACF;EAEA,IAAIE,OAAO,GAAG,KAAK;EACnB,MAAMC,GAAG,GAAGhB,QAAQ,CAACiB,GAAG,CAAC,CAACtB,CAAC,EAAEO,CAAC,KAAK;IACjC,IAAIP,CAAC,CAACQ,IAAI,KAAK,cAAc,IAAID,CAAC,IAAID,MAAM,IAAIG,WAAW,CAACc,GAAG,CAAChB,CAAC,CAAC,EAAE,OAAOP,CAAC;IAC5E,IAAIwB,OAAO,GAAG,KAAK;IACnB,MAAMC,OAAO,GAAGzB,CAAC,CAACyB,OAAO,CAACH,GAAG,CAAEI,CAAC,IAAK;MACnC,IAAIA,CAAC,CAACC,OAAO,IAAID,CAAC,CAACE,OAAO,CAACzB,MAAM,GAAGL,kBAAkB,EAAE,OAAO4B,CAAC;MAChE,IAAIA,CAAC,CAACE,OAAO,CAACC,UAAU,CAAC,iBAAiB,CAAC,EAAE,OAAOH,CAAC;MACrDF,OAAO,GAAG,IAAI;MACd,OAAO;QACL,GAAGE,CAAC;QACJE,OAAO,EAAE,qBAAqBF,CAAC,CAACE,OAAO,CAACzB,MAAM;MAChD,CAAC;IACH,CAAC,CAAC;IACF,IAAI,CAACqB,OAAO,EAAE,OAAOxB,CAAC;IACtBoB,OAAO,GAAG,IAAI;IACd,OAAO;MAAE,GAAGpB,CAAC;MAAEyB;IAAQ,CAAC;EAC1B,CAAC,CAAC;EACF,OAAOL,OAAO,GAAGC,GAAG,GAAGhB,QAAQ;AACjC;;AAEA;AACA;AACA;AACA;AACA;AACA,MAAMS,SAAS,GAAG,kGAAkG;;AAEpH;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACO,SAASgB,iBAAiBA,CAACzB,QAAwB,EAGxD;EACA,IAAI0B,KAAK,GAAG,CAAC;EACb,KAAK,MAAM/B,CAAC,IAAIK,QAAQ,EAAE0B,KAAK,IAAIhC,WAAW,CAACC,CAAC,CAAC;EACjD,IAAI+B,KAAK,IAAIrC,iBAAiB,EAAE,OAAO;IAAEW,QAAQ;IAAE2B,aAAa,EAAE;EAAE,CAAC;;EAErE;EACA,MAAMC,UAAU,GAAG7B,gBAAgB,CAACC,QAAQ,CAAC;EAC7C,IAAI6B,KAAK,GAAG,CAAC;EACb,KAAK,MAAMlC,CAAC,IAAIiC,UAAU,EAAEC,KAAK,IAAInC,WAAW,CAACC,CAAC,CAAC;EACnD,IAAIkC,KAAK,IAAIxC,iBAAiB,EAAE,OAAO;IAAEW,QAAQ,EAAE4B,UAAU;IAAED,aAAa,EAAE;EAAE,CAAC;EAEjFD,KAAK,GAAGG,KAAK;EACb,MAAMb,GAAG,GAAG,CAAC,GAAGY,UAAU,CAAC;EAC3B,IAAID,aAAa,GAAG,CAAC;EACrB,OAAOX,GAAG,CAAClB,MAAM,GAAGN,iBAAiB,IAAIkC,KAAK,GAAGrC,iBAAiB,EAAE;IAClEqC,KAAK,IAAIhC,WAAW,CAACsB,GAAG,CAAC,CAAC,CAAE,CAAC;IAC7BA,GAAG,CAACc,KAAK,CAAC,CAAC;IACX,OAAOd,GAAG,CAAClB,MAAM,GAAG,CAAC,IAAIkB,GAAG,CAAC,CAAC,CAAC,CAAEb,IAAI,KAAK,MAAM,EAAE;MAChDuB,KAAK,IAAIhC,WAAW,CAACsB,GAAG,CAAC,CAAC,CAAE,CAAC;MAC7BA,GAAG,CAACc,KAAK,CAAC,CAAC;IACb;IACAH,aAAa,IAAI,CAAC;EACpB;EACA,OAAO;IAAE3B,QAAQ,EAAEgB,GAAG;IAAEW;EAAc,CAAC;AACzC;;AAEA;AACO,SAASI,WAAWA,CAAC/B,QAAwB,EAAkB;EACpE,OAAOyB,iBAAiB,CAACzB,QAAQ,CAAC,CAACA,QAAQ;AAC7C;;AAEA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACO,SAASgC,gBAAgBA,CAC9BhC,QAAwB,EACxBiC,OAAe,EACsC;EACrD;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACE,MAAMC,MAAM,GAAGC,IAAI,CAACC,GAAG,CAAC/C,iBAAiB,EAAEE,iBAAiB,GAAG0C,OAAO,CAAC;EACvE,IAAIP,KAAK,GAAG,CAAC;EACb,KAAK,MAAM/B,CAAC,IAAIK,QAAQ,EAAE0B,KAAK,IAAIhC,WAAW,CAACC,CAAC,CAAC;EACjD,IAAI+B,KAAK,IAAIQ,MAAM,EAAE,OAAO;IAAElC,QAAQ;IAAE2B,aAAa,EAAE;EAAE,CAAC;;EAE1D;EACA,IAAIU,YAAY,GAAG,CAAC,CAAC;EACrB,KAAK,IAAInC,CAAC,GAAGF,QAAQ,CAACF,MAAM,GAAG,CAAC,EAAEI,CAAC,IAAI,CAAC,EAAEA,CAAC,EAAE,EAAE;IAC7C,IAAIF,QAAQ,CAACE,CAAC,CAAC,CAAEC,IAAI,KAAK,MAAM,EAAE;MAChCkC,YAAY,GAAGnC,CAAC;MAChB;IACF;EACF;EACA,IAAImC,YAAY,IAAI,CAAC,EAAE,OAAO;IAAErC,QAAQ;IAAE2B,aAAa,EAAE;EAAE,CAAC;EAE5D,MAAMW,OAAO,GAAGtC,QAAQ,CAACuC,KAAK,CAAC,CAAC,EAAEF,YAAY,CAAC;EAC/C,MAAMG,OAAO,GAAGxC,QAAQ,CAACuC,KAAK,CAACF,YAAY,CAAC;EAC5C,IAAII,WAAW,GAAG,CAAC;EACnB,KAAK,MAAM9C,CAAC,IAAI6C,OAAO,EAAEC,WAAW,IAAI/C,WAAW,CAACC,CAAC,CAAC;EAEtD,MAAMiC,UAAU,GAAG7B,gBAAgB,CAACuC,OAAO,CAAC;EAC5C,IAAII,WAAW,GAAG,CAAC;EACnB,KAAK,MAAM/C,CAAC,IAAIiC,UAAU,EAAEc,WAAW,IAAIhD,WAAW,CAACC,CAAC,CAAC;EACzD,IAAI+C,WAAW,GAAGD,WAAW,IAAIP,MAAM,EAAE;IACvC,OAAO;MAAElC,QAAQ,EAAE,CAAC,GAAG4B,UAAU,EAAE,GAAGY,OAAO,CAAC;MAAEb,aAAa,EAAE;IAAE,CAAC;EACpE;EAEA,MAAMX,GAAG,GAAG,CAAC,GAAGY,UAAU,CAAC;EAC3B,IAAID,aAAa,GAAG,CAAC;EACrB,OAAOX,GAAG,CAAClB,MAAM,GAAG,CAAC,IAAI4C,WAAW,GAAGD,WAAW,GAAGP,MAAM,EAAE;IAC3DQ,WAAW,IAAIhD,WAAW,CAACsB,GAAG,CAAC,CAAC,CAAE,CAAC;IACnCA,GAAG,CAACc,KAAK,CAAC,CAAC;IACX,OAAOd,GAAG,CAAClB,MAAM,GAAG,CAAC,IAAIkB,GAAG,CAAC,CAAC,CAAC,CAAEb,IAAI,KAAK,MAAM,EAAE;MAChDuC,WAAW,IAAIhD,WAAW,CAACsB,GAAG,CAAC,CAAC,CAAE,CAAC;MACnCA,GAAG,CAACc,KAAK,CAAC,CAAC;IACb;IACAH,aAAa,IAAI,CAAC;EACpB;EACA,OAAO;IAAE3B,QAAQ,EAAE,CAAC,GAAGgB,GAAG,EAAE,GAAGwB,OAAO,CAAC;IAAEb;EAAc,CAAC;AAC1D","ignoreList":[]}