billion-context 0.1.71-pr.467.282 → 0.1.71

package/dist/index.js

@@ -56684,7 +56684,6 @@ ${bodyBuffer.toString("utf8")}`);
       nativeFromFallback = false;
       log2("info", `[${session.id}] self-healed context window: ${resolved} \u2192 ${learnedLimit} (learned from an upstream overflow)`);
     }
-    const nativeWindow = reqConfig.modelContextLimit;
     if (shouldReserveOutputHeadroom(protocol)) {
       const p2 = parsed;
       const rawMax = p2.max_tokens ?? p2.max_completion_tokens ?? p2.max_output_tokens;
@@ -56699,7 +56698,7 @@ ${bodyBuffer.toString("utf8")}`);
     acquireInFlight(session);
     try {
       await withSessionLock(session, async () => {
-        const runPrepare = () => countTokens ? prepareCountTokens(parsed, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, nativeWindow) : responsesCompact ? prepareResponsesCompact(bodyBuffer, parsed, session, req, core, reqConfig, log2) : prepareResponses(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin, nativeWindow);
+        const runPrepare = () => countTokens ? prepareCountTokens(parsed, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : responsesCompact ? prepareResponsesCompact(bodyBuffer, parsed, session, req, core, reqConfig, log2) : prepareResponses(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin);
         prepared = runPrepare();
         if (!countTokens && !responsesCompact) {
           const outcome = await preflightCompressIfNeeded(
@@ -56783,7 +56782,7 @@ function diagNudge(turn, sessionId, tokenCount, limit, model, willInject) {
   const interval = b2["nudgeGrowthTokens"] ?? 0;
   const pendingT1 = b2["pendingT1"] ?? 0;
   const ref = b2["growthReference"] ?? 0;
-  const inject = willInject ? n.shouldInject ? `INJECT T${n.tier ?? "?"}` : `INJECT-ESC T${n.tier ?? "?"}` : n.shouldInject ? `ARMED-SUPPRESSED T${n.tier ?? "?"}` : "idle";
+  const inject = !n.shouldInject ? "idle" : willInject ? `INJECT T${n.tier ?? "?"}` : `ARMED-SUPPRESSED T${n.tier ?? "?"}`;
   const modelTag = model ? ` model=${model}` : "";
   return `[${sessionId}] nudge ${inject}: usage=${pct2} (${tokenCount}/${limit}), growth=${growth}/${floor} (ref=${ref}, interval=${interval}), pendingT1=${pendingT1}/${interval}${modelTag}, reason="${n.reason.slice(0, 120)}"`;
 }
@@ -56818,7 +56817,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
       if (t) session.meta.title = t;
     }
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
-    const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
+    const willInjectNudge = opts.compress.injectNudge && !!turn.nudge?.shouldInject;
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
     processedMessages = stripKernelSummaries(turn.messages, turn.state);
     reapOrphanBlocks(session, msgs, deactivateBlock);
@@ -56846,36 +56845,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
   delete rebuilt.prompt_cache_key;
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, anthropicSystem: parsed.system, protocol: "anthropic", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, renderTags: "text-only" };
 }
-var OUTPUT_CLAMP_MARGIN_PCT = 0.05;
-var OUTPUT_CLAMP_MIN_MARGIN = 2048;
-var OUTPUT_CLAMP_FLOOR = 1024;
-var EMERGENCY_NUDGE_ESCALATION_PCT = 0.7;
-function estimateInputTokens(processedMessages, systemText, tools, lastInputTokens) {
-  const est = estimateCoreMessages(processedMessages) + defaultCountTokens(systemText ?? "") + defaultCountTokens(JSON.stringify(tools ?? []));
-  return Math.max(lastInputTokens > 0 ? lastInputTokens : 0, est);
-}
-function clampOutputBudget(requested, inputEstimate, nativeWindow) {
-  const margin = Math.max(OUTPUT_CLAMP_MIN_MARGIN, Math.ceil(inputEstimate * OUTPUT_CLAMP_MARGIN_PCT));
-  const cap = nativeWindow - inputEstimate - margin;
-  if (cap < OUTPUT_CLAMP_FLOOR || cap >= requested) return void 0;
-  return cap;
-}
-function emergencyNudge(nudge, escalationPct = EMERGENCY_NUDGE_ESCALATION_PCT) {
-  if (!nudge || nudge.shouldInject) return false;
-  if (nudge.compressibleRanges.length === 0) return false;
-  return nudge.contextUsage >= escalationPct;
-}
-function clampOutgoingOutput(rebuilt, field, ctx, sessionId, log2) {
-  const raw = rebuilt[field];
-  if (typeof raw !== "number") return;
-  const inputEstimate = estimateInputTokens(ctx.processedMessages, ctx.systemText, ctx.tools, ctx.lastInputTokens);
-  const capped = clampOutputBudget(raw, inputEstimate, ctx.nativeWindow);
-  if (capped !== void 0) {
-    rebuilt[field] = capped;
-    log2("info", `[${sessionId}] output budget clamped ${raw} -> ${capped} (input~${inputEstimate}, window=${ctx.nativeWindow}); prevents input+output overflow (#453)`);
-  }
-}
-function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session, pluginMode, nativeWindow) {
+function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session, pluginMode) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
@@ -56905,7 +56875,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
       if (t) session.meta.title = t;
     }
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
-    const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
+    const willInjectNudge = opts.compress.injectNudge && !!turn.nudge?.shouldInject && shouldInject;
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
     processedMessages = stripKernelSummaries(turn.messages, turn.state);
     reapOrphanBlocks(session, msgs, deactivateBlock);
@@ -56931,7 +56901,6 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
     processedMessages = [];
   }
   const rebuilt = { ...parsed, messages: rebuiltMessages, tools: toolsOut };
-  clampOutgoingOutput(rebuilt, typeof parsed.max_completion_tokens === "number" ? "max_completion_tokens" : "max_tokens", { systemText: openaiSystemText, tools: toolsOut, processedMessages, lastInputTokens: session.stats.lastInputTokens, nativeWindow }, sessionId, log2);
   delete rebuilt.prompt_cache_retention;
   if (stream2 && rebuilt.stream_options === void 0) {
     rebuilt.stream_options = { include_usage: true };
@@ -56940,7 +56909,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
   markDirty(session);
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, protocol: "openai", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, openaiSystemText, renderTags: "text-only" };
 }
-function prepareResponses(parsed, req, opts, core, config, prompts, log2, session, identity, pluginMode, upstreamOrigin, nativeWindow) {
+function prepareResponses(parsed, req, opts, core, config, prompts, log2, session, identity, pluginMode, upstreamOrigin) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
@@ -56997,7 +56966,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
       if (t) session.meta.title = t;
     }
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
-    const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && !isCompactionTrigger && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
+    const willInjectNudge = opts.compress.injectNudge && !!turn.nudge?.shouldInject && shouldInject && !isCompactionTrigger;
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
     processedMessages = stripKernelSummaries(turn.messages, turn.state);
     reapOrphanBlocks(session, msgs, deactivateBlock);
@@ -57048,9 +57017,6 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     log2("info", `[${sessionId}] codex compact intercepted (trigger); forged SSE with ${summaries.length} block summary(s), upstream not contacted`);
   }
   const rebuilt = { ...parsed, input: rebuiltInput, tools: toolsOut };
-  if (!isCompactionTrigger) {
-    clampOutgoingOutput(rebuilt, "max_output_tokens", { systemText: (responsesProjection?.systemParts ?? []).join("\n"), tools: toolsOut, processedMessages, lastInputTokens: session.stats.lastInputTokens, nativeWindow }, sessionId, log2);
-  }
   const promptCacheKey = resolvePromptCacheKey(
     rebuilt.prompt_cache_key,
     identity,