billion-context 0.1.75-pr.490.310 → 0.1.75-pr.491.314

package/dist/index.js

@@ -48479,11 +48479,14 @@ ${summary}` }]
   }
   return { items, replaced, dropped };
 }
-function codexCompactGate(session, effectiveLimit, transformOk) {
-  if (!transformOk || effectiveLimit <= 0) return false;
+function codexCompactGatePre(session, effectiveLimit) {
+  if (effectiveLimit <= 0) return false;
   if (session.stats.lastInputTokens >= effectiveLimit * 0.9) return false;
   return session.state.blocks.some((b2) => b2.active);
 }
+function codexCompactGate(session, effectiveLimit, transformOk) {
+  return transformOk && codexCompactGatePre(session, effectiveLimit);
+}
 function buildTriggerForgeBody(summaryText, usage, stream2) {
   const compactionItem = {
     type: "compaction",
@@ -51306,6 +51309,63 @@ function reapOrphanBlocks(session, visible, deactivate) {
   return { reaped };
 }
 
+// src/ref-gc.ts
+var CITED_REF = /\bm(\d{5})\b/g;
+var REF_CAPACITY_WARN = 9e4;
+function gcMessageRefs(session, incoming, outgoing, log2) {
+  const { byRaw, byRef } = session.state.messageRefs;
+  const tokenSnapshot = session.state.tokenSnapshot;
+  const live = /* @__PURE__ */ new Set();
+  for (const m2 of incoming) if (m2.id) live.add(m2.id);
+  for (const m2 of outgoing) if (m2.id) live.add(m2.id);
+  for (const match of citedRefs(outgoing)) {
+    const rawId = byRef[`m${match}`];
+    if (rawId) live.add(rawId);
+  }
+  for (const b2 of session.state.blocks) {
+    if (!b2.active) continue;
+    for (const id of b2.effectiveMessageIds) live.add(id);
+  }
+  const nextByRaw = {};
+  let pruned = 0;
+  for (const [rawId, ref] of Object.entries(byRaw)) {
+    if (live.has(rawId)) nextByRaw[rawId] = ref;
+    else pruned++;
+  }
+  const nextByRef = {};
+  for (const [ref, rawId] of Object.entries(byRef)) {
+    if (live.has(rawId)) nextByRef[ref] = rawId;
+  }
+  if (pruned === 0) return { pruned };
+  const nextSnapshot = {};
+  for (const [ref, tokens] of Object.entries(tokenSnapshot ?? {})) {
+    if (nextByRef[ref] !== void 0) nextSnapshot[ref] = tokens;
+  }
+  session.state = {
+    ...session.state,
+    messageRefs: {
+      ...session.state.messageRefs,
+      byRaw: nextByRaw,
+      byRef: nextByRef
+    },
+    tokenSnapshot: nextSnapshot
+  };
+  log2("info", `[${session.id}] ref gc: pruned ${pruned} dead ref mapping(s), ${Object.keys(nextByRaw).length} live`);
+  const highest = highestUsedIndex(session.state.messageRefs);
+  if (highest >= REF_CAPACITY_WARN) {
+    log2("warn", `[${session.id}] message refs at m${String(highest).padStart(5, "0")} approaching capacity 99999 \u2014 compression may stall if exhausted`);
+  }
+  return { pruned };
+}
+function citedRefs(messages) {
+  const refs = /* @__PURE__ */ new Set();
+  for (const m2 of messages) {
+    if (!m2.text) continue;
+    for (const match of m2.text.matchAll(CITED_REF)) refs.add(match[1]);
+  }
+  return refs;
+}
+
 // src/instance.ts
 import { randomUUID as randomUUID3 } from "crypto";
 import fs3 from "fs";
@@ -57074,6 +57134,21 @@ ${bodyBuffer.toString("utf8")}`);
     try {
       await withSessionLock(session, async () => {
         const runPrepare = () => countTokens ? prepareCountTokens(parsed, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, nativeWindow) : responsesCompact ? prepareResponsesCompact(bodyBuffer, parsed, session, req, core, reqConfig, log2) : prepareResponses(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin, nativeWindow);
+        const isCodexCompactTrigger = protocol === "responses" && !responsesCompact && isCodexClient(req.headers) && hasCompactionTrigger(parsed.input);
+        if (isCodexCompactTrigger) {
+          const mode = codexCompactMode();
+          const gatePre = codexCompactGatePre(session, reqConfig.modelContextLimit);
+          if (mode === "intercept" && gatePre) prepared = runPrepare();
+          if (prepared?.codexForge) {
+            await forward(req, res, opts, prepared.body, prepared, core, reqConfig, log2, route, instanceId, affinity);
+            rememberPluginMessages(sessionId, prepared.processedMessages, prepared.originalMessages, prepared.nudge);
+            return;
+          }
+          const why = mode !== "intercept" ? "BILI_CODEX_COMPACT=pass" : !gatePre ? "gate preconditions not met" : "transform/forge failed";
+          log2("info", `[${session.id}] codex compaction_trigger request not intercepted (${why}) \u2014 forwarding verbatim (no preflight, no rebuild, no window clamp)`);
+          await forward(req, res, opts, bodyBuffer, null, core, reqConfig, log2, route, instanceId, affinity);
+          return;
+        }
         prepared = runPrepare();
         if (!countTokens && !responsesCompact) {
           const outcome = await preflightCompressIfNeeded(
@@ -57198,6 +57273,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
     processedMessages = stripKernelSummaries(turn.messages, turn.state);
     applyCompactionArchive(session, activeBefore, new Set(msgs.map((m2) => m2.id)), log2);
     reapOrphanBlocks(session, msgs, deactivateBlock);
+    gcMessageRefs(session, msgs, turn.messages, log2);
     rebuiltMessages = coreToAnthropic(processedMessages, cacheControls);
     systemOut = injectSystem(parsed, opts, prompts);
     if (injectTools) {
@@ -57287,6 +57363,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
     processedMessages = stripKernelSummaries(turn.messages, turn.state);
     applyCompactionArchive(session, activeBefore, new Set(msgs.map((m2) => m2.id)), log2);
     reapOrphanBlocks(session, msgs, deactivateBlock);
+    gcMessageRefs(session, msgs, turn.messages, log2);
     rebuiltMessages = systemToUser(coreToOpenai(processedMessages));
     const sysParts = [];
     if (systemText) sysParts.push(systemText);
@@ -57379,6 +57456,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
     processedMessages = stripKernelSummaries(turn.messages, turn.state);
     reapOrphanBlocks(session, msgs, deactivateBlock);
+    gcMessageRefs(session, msgs, turn.messages, log2);
     rebuiltInput = patchResponsesInput(projection, processedMessages);
     const forgedSummaries = echoReplaced ? [] : session.metadata.codexForgedSummaries ?? [];
     if (shouldInject && !isCompactionTrigger && !process.env.ACP_NO_COMPRESS_PROMPT) {
@@ -57519,7 +57597,7 @@ function prepareResponsesCompact(body, parsed, session, req, core, config, log2)
     compressInjected: false,
     resetAfterSuccess: true
   };
-  if (codexCompactMode() !== "intercept" || !isCodexClient(req.headers) || !Array.isArray(parsed.input)) {
+  if (codexCompactMode() !== "intercept" || !isCodexClient(req.headers) || !Array.isArray(parsed.input) || !codexCompactGatePre(session, config.modelContextLimit)) {
     return base;
   }
   const prevState = session.state;