opencode-acp 1.16.0-pr.360.128 → 1.16.0-pr.365.125

package/dist/index.js

@@ -890,7 +890,6 @@ var VALID_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "compress.modelMaxLimits",
   "compress.modelMinLimits",
   "compress.providers",
-  "compress.contextLimitFallback",
   "compress.nudgeFrequency",
   "compress.minNudgeContextPercent",
   "compress.nudgeGrowthTokens",
@@ -914,7 +913,6 @@ var VALID_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "compress.reasoning",
   "compress.reasoning.drop",
   "compress.reasoning.threshold",
-  "compress.completionReserveTokens",
   "gc",
   "gc.algorithm",
   "gc.promotionThreshold",
@@ -966,11 +964,7 @@ function validateConfigTypes(config) {
     errors.push({ key: "storagePath", expected: "string", actual: typeof config.storagePath });
   }
   if (config.allowSubAgents !== void 0 && typeof config.allowSubAgents !== "boolean") {
-    errors.push({
-      key: "allowSubAgents",
-      expected: "boolean",
-      actual: typeof config.allowSubAgents
-    });
+    errors.push({ key: "allowSubAgents", expected: "boolean", actual: typeof config.allowSubAgents });
   }
   if (config.pruneNotification !== void 0) {
     const validValues = ["off", "minimal", "detailed"];
@@ -1299,20 +1293,6 @@ function validateConfigTypes(config) {
           }
         }
       }
-      if (compress.completionReserveTokens !== void 0 && typeof compress.completionReserveTokens !== "number") {
-        errors.push({
-          key: "compress.completionReserveTokens",
-          expected: "number",
-          actual: typeof compress.completionReserveTokens
-        });
-      }
-      if (typeof compress.completionReserveTokens === "number" && compress.completionReserveTokens < 0) {
-        errors.push({
-          key: "compress.completionReserveTokens",
-          expected: "non-negative number (>= 0)",
-          actual: `${compress.completionReserveTokens}`
-        });
-      }
       if (typeof compress.iterationNudgeThreshold === "number" && compress.iterationNudgeThreshold < 1) {
         errors.push({
           key: "compress.iterationNudgeThreshold",
@@ -1431,20 +1411,12 @@ function validateConfigTypes(config) {
             break;
           case "nudgeForce":
             if (value !== "strong" && value !== "soft") {
-              errors.push({
-                key,
-                expected: "'strong' | 'soft'",
-                actual: JSON.stringify(value)
-              });
+              errors.push({ key, expected: "'strong' | 'soft'", actual: JSON.stringify(value) });
             }
             break;
           case "stringArray":
             if (!Array.isArray(value) || !value.every((entry) => typeof entry === "string")) {
-              errors.push({
-                key,
-                expected: "string[]",
-                actual: JSON.stringify(value)
-              });
+              errors.push({ key, expected: "string[]", actual: JSON.stringify(value) });
             }
             break;
           case "reasoningConfig":
@@ -1479,11 +1451,7 @@ function validateConfigTypes(config) {
           return;
         }
         if (typeof overrides !== "object" || overrides === null || Array.isArray(overrides)) {
-          errors.push({
-            key: prefix,
-            expected: "CompressModelOverrides",
-            actual: typeof overrides
-          });
+          errors.push({ key: prefix, expected: "CompressModelOverrides", actual: typeof overrides });
           return;
         }
         const model = overrides;
@@ -1516,11 +1484,7 @@ function validateConfigTypes(config) {
         for (const [providerId, providerValue] of Object.entries(providers)) {
           const prefix = `compress.providers.${providerId}`;
           if (typeof providerValue !== "object" || providerValue === null || Array.isArray(providerValue)) {
-            errors.push({
-              key: prefix,
-              expected: "ProviderOverrides",
-              actual: typeof providerValue
-            });
+            errors.push({ key: prefix, expected: "ProviderOverrides", actual: typeof providerValue });
             continue;
           }
           const provider = providerValue;
@@ -1557,20 +1521,6 @@ function validateConfigTypes(config) {
         }
       };
       validateProviderOverrides(compress.providers);
-      if (compress.contextLimitFallback !== void 0 && typeof compress.contextLimitFallback !== "number") {
-        errors.push({
-          key: "compress.contextLimitFallback",
-          expected: "number",
-          actual: typeof compress.contextLimitFallback
-        });
-      }
-      if (typeof compress.contextLimitFallback === "number" && compress.contextLimitFallback < 0) {
-        errors.push({
-          key: "compress.contextLimitFallback",
-          expected: "non-negative number (0 disables the fallback)",
-          actual: `${compress.contextLimitFallback}`
-        });
-      }
       const validValues = ["ask", "allow", "deny"];
       if (compress.permission !== void 0 && !validValues.includes(compress.permission)) {
         errors.push({
@@ -1656,22 +1606,13 @@ function validateConfigTypes(config) {
           });
         } else {
           if (gc.batchCleanup.lowThreshold !== void 0) {
-            validateBatchThreshold(
-              "gc.batchCleanup.lowThreshold",
-              gc.batchCleanup.lowThreshold
-            );
+            validateBatchThreshold("gc.batchCleanup.lowThreshold", gc.batchCleanup.lowThreshold);
           }
           if (gc.batchCleanup.highThreshold !== void 0) {
-            validateBatchThreshold(
-              "gc.batchCleanup.highThreshold",
-              gc.batchCleanup.highThreshold
-            );
+            validateBatchThreshold("gc.batchCleanup.highThreshold", gc.batchCleanup.highThreshold);
           }
           if (gc.batchCleanup.forceThreshold !== void 0) {
-            validateBatchThreshold(
-              "gc.batchCleanup.forceThreshold",
-              gc.batchCleanup.forceThreshold
-            );
+            validateBatchThreshold("gc.batchCleanup.forceThreshold", gc.batchCleanup.forceThreshold);
           }
         }
       }
@@ -1761,7 +1702,6 @@ var defaultConfig = {
     summaryBuffer: true,
     maxContextLimit: "80%",
     minContextLimit: "80%",
-    contextLimitFallback: 128e3,
     nudgeFrequency: 5,
     minNudgeContextPercent: 5,
     iterationNudgeThreshold: 15,
@@ -1932,7 +1872,6 @@ function mergeCompress(base, override) {
     modelMaxLimits: override.modelMaxLimits ?? base.modelMaxLimits,
     modelMinLimits: override.modelMinLimits ?? base.modelMinLimits,
     providers: mergeProviderOverrides(base.providers, override.providers),
-    contextLimitFallback: override.contextLimitFallback ?? base.contextLimitFallback,
     nudgeFrequency: override.nudgeFrequency ?? base.nudgeFrequency,
     minNudgeContextPercent: override.minNudgeContextPercent ?? base.minNudgeContextPercent,
     nudgeGrowthTokens: override.nudgeGrowthTokens,
@@ -1956,8 +1895,7 @@ function mergeCompress(base, override) {
     reasoning: {
       drop: override.reasoning?.drop ?? base.reasoning?.drop ?? DEFAULT_COMPRESS_REASONING.drop,
       threshold: override.reasoning?.threshold ?? base.reasoning?.threshold ?? DEFAULT_COMPRESS_REASONING.threshold
-    },
-    completionReserveTokens: override.completionReserveTokens ?? base.completionReserveTokens
+    }
   };
 }
 function mergeCommands(base, override) {
@@ -1995,9 +1933,10 @@ function deepCloneConfig(config) {
             ...provider,
             ...provider.models ? {
               models: Object.fromEntries(
-                Object.entries(provider.models).map(
-                  ([modelId, model]) => [modelId, { ...model }]
-                )
+                Object.entries(provider.models).map(([modelId, model]) => [
+                  modelId,
+                  { ...model }
+                ])
               )
             } : {}
           }
@@ -2071,14 +2010,8 @@ function mergeLayer(config, data) {
     ],
     compress: mergeCompress(config.compress, data.compress),
     gc: mergeGC(config.gc, data.gc),
-    qualityGate: mergeQualityGate(
-      config.qualityGate,
-      data.qualityGate
-    ),
-    messageFilters: mergeMessageFilters(
-      config.messageFilters,
-      data.messageFilters
-    )
+    qualityGate: mergeQualityGate(config.qualityGate, data.qualityGate),
+    messageFilters: mergeMessageFilters(config.messageFilters, data.messageFilters)
   };
 }
 function scheduleParseWarning(ctx, title, message) {
@@ -2752,16 +2685,6 @@ function resetOnCompaction(state) {
     nextRef: 1
   };
 }
-function resolveEffectiveContextLimit(state, config) {
-  if (typeof state.modelContextLimit === "number" && state.modelContextLimit > 0) {
-    return { limit: state.modelContextLimit, source: "model" };
-  }
-  const fallback = config.compress.contextLimitFallback;
-  if (typeof fallback === "number" && fallback > 0) {
-    return { limit: fallback, source: "fallback" };
-  }
-  return void 0;
-}
 
 // lib/state/persistence.ts
 function getDefaultStorageDir() {
@@ -4633,24 +4556,6 @@ var SessionStateRegistry = class {
   hydrateModelLimitsFromClient(client) {
     return this.catalog.hydrateFromClient(client);
   }
-  // [FIX #346] The init-time seed (above) is fire-and-forget and races
-  // server readiness: in headless spawn+resume mode the provider-config
-  // call can fail before the server is up, leaving the catalog empty for
-  // the process's lifetime. During a request the server is guaranteed up
-  // (we are inside its pipeline), so on a catalog miss we retry hydration
-  // once per process before giving up (the fallback limit then applies).
-  // The in-flight promise (not a boolean) lets concurrent callers await the
-  // same hydration instead of skipping it.
-  lazyHydration;
-  async hydrateAndResolve(client, providerId, modelId) {
-    const existing = this.catalog.resolve(providerId, modelId);
-    if (existing !== void 0) {
-      return existing;
-    }
-    this.lazyHydration ??= this.catalog.hydrateFromClient(client);
-    await this.lazyHydration;
-    return this.catalog.resolve(providerId, modelId);
-  }
   get(sessionId) {
     return this.states.get(sessionId);
   }
@@ -4744,8 +4649,7 @@ function createSessionState() {
     modelID: void 0,
     systemPromptTokens: void 0,
     storageDir: void 0,
-    qualityGateRetryPending: false,
-    noContextLimitWarned: false
+    qualityGateRetryPending: false
   };
 }
 function resetSessionState(state) {
@@ -4788,7 +4692,6 @@ function resetSessionState(state) {
   state.systemPromptTokens = void 0;
   state.storageDir = void 0;
   state.qualityGateRetryPending = false;
-  state.noContextLimitWarned = false;
 }
 async function ensureSessionInitialized(client, state, sessionId, logger, messages, config, projectDir) {
   if (state.sessionId === sessionId) {
@@ -6804,7 +6707,6 @@ function getModelInfo(messages) {
   };
 }
 function resolveContextTokenLimit(config, state, providerId, modelId, threshold) {
-  const effectiveLimit = resolveEffectiveContextLimit(state, config);
   const parseLimitValue = (limit) => {
     if (limit === void 0) {
       return void 0;
@@ -6812,7 +6714,7 @@ function resolveContextTokenLimit(config, state, providerId, modelId, threshold)
     if (typeof limit === "number") {
       return limit;
     }
-    if (!limit.endsWith("%") || effectiveLimit === void 0) {
+    if (!limit.endsWith("%") || state.modelContextLimit === void 0) {
       return void 0;
     }
     const parsedPercent = parseFloat(limit.slice(0, -1));
@@ -6821,7 +6723,7 @@ function resolveContextTokenLimit(config, state, providerId, modelId, threshold)
     }
     const roundedPercent = Math.round(parsedPercent);
     const clampedPercent = Math.max(0, Math.min(100, roundedPercent));
-    return Math.round(clampedPercent / 100 * effectiveLimit.limit);
+    return Math.round(clampedPercent / 100 * state.modelContextLimit);
   };
   if (threshold === "max") {
     const nestedLimit = resolveCompressOverrides(config, providerId, modelId).maxContextLimit;
@@ -6872,12 +6774,11 @@ function isContextOverLimits(config, state, providerId, modelId, messages) {
       if (!overMaxLimit) break;
     }
   }
-  const effectiveLimit = resolveEffectiveContextLimit(state, config);
   return {
     overMaxLimit,
     overMinLimit,
     currentTokens,
-    modelContextLimit: effectiveLimit?.limit
+    modelContextLimit: state.modelContextLimit
   };
 }
 ensureBuiltinTriggerPolicyRegistered();
@@ -7183,10 +7084,13 @@ function buildCompressibleRanges(messages, state, protectedTools = [], protected
     if (!ref) continue;
     const rn = parseInt(ref.slice(1), 10);
     if ((protectedTools.length > 0 || protectedFilePatterns.length > 0) && messageContainsProtectedTool(msg, protectedTools, protectedFilePatterns)) {
-      const tokens2 = Math.round(countMessageCharacters(msg) / 4);
+      let tokens2 = 0;
       const tools = /* @__PURE__ */ new Set();
       for (const part of msg.parts || []) {
-        if (part.type !== "text" && part.type !== "reasoning") {
+        if (part.type === "text" && typeof part.text === "string") {
+          tokens2 += Math.round(part.text.length / 4);
+        } else if (part.type !== "text" && part.type !== "reasoning") {
+          tokens2 += Math.round(JSON.stringify(part).length / 4);
           const toolName = part?.tool;
           const callID = part?.callID;
           if (toolName && callID) {
@@ -7207,13 +7111,15 @@ function buildCompressibleRanges(messages, state, protectedTools = [], protected
       protectedMsgInfo.push({ ref, refNum: rn, tokens: tokens2, tools: [...tools] });
       continue;
     }
-    const tokens = Math.round(countMessageCharacters(msg) / 4);
+    let tokens = 0;
     let isTool = false;
     let hasMeaningfulPart = false;
     for (const part of msg.parts || []) {
       if (part.type === "text" && typeof part.text === "string") {
+        tokens += Math.round(part.text.length / 4);
         if (part.text.trim().length > 0) hasMeaningfulPart = true;
       } else if (part.type !== "text" && part.type !== "reasoning") {
+        tokens += Math.round(JSON.stringify(part).length / 4);
         isTool = true;
         hasMeaningfulPart = true;
       }
@@ -8954,9 +8860,8 @@ function createDecompressTool(factoryCtx) {
     async execute(args, toolCtx) {
       const ctx = resolveToolContext(factoryCtx, toolCtx.sessionID);
       const { rawMessages } = await prepareDecompressSession(ctx, toolCtx);
-      const effectiveLimitBefore = resolveEffectiveContextLimit(ctx.state, ctx.config);
-      const contextUsageBefore = effectiveLimitBefore ? Math.round(
-        getCurrentTokenUsage(ctx.state, rawMessages) / effectiveLimitBefore.limit * 100
+      const contextUsageBefore = ctx.state.modelContextLimit ? Math.round(
+        getCurrentTokenUsage(ctx.state, rawMessages) / ctx.state.modelContextLimit * 100
       ) : void 0;
       const resolved = resolveTargets(args, ctx.state, rawMessages, ctx.logger);
       if (!resolved.ok) {
@@ -9020,9 +8925,8 @@ function createDecompressTool(factoryCtx) {
         0,
         ctx.state.stats.totalPruneTokens - restoredTokens
       );
-      const effectiveLimitAfter = resolveEffectiveContextLimit(ctx.state, ctx.config);
-      const contextUsageAfter = effectiveLimitAfter ? Math.round(
-        getCurrentTokenUsage(ctx.state, rawMessages) / effectiveLimitAfter.limit * 100
+      const contextUsageAfter = ctx.state.modelContextLimit ? Math.round(
+        getCurrentTokenUsage(ctx.state, rawMessages) / ctx.state.modelContextLimit * 100
       ) : void 0;
       await finalizeDecompressSession(ctx);
       const restoredContentPreview = buildRestoredContentPreview(
@@ -9679,7 +9583,7 @@ import { writeFile as writeFile2, mkdir as mkdir2 } from "fs/promises";
 import { join as join4 } from "path";
 import { existsSync as existsSync4 } from "fs";
 import { homedir as homedir3 } from "os";
-var LOG_VERSION = true ? "1.16.0-pr.360.128" : "dev";
+var LOG_VERSION = true ? "1.16.0-pr.365.125" : "dev";
 var LEVEL_RANK = {
   debug: 10,
   info: 20,
@@ -10503,8 +10407,6 @@ var MIN_OUTPUT_TOKENS = 1e3;
 var KEEP_PREFIX_CHARS = 2e3;
 var KEEP_SUFFIX_CHARS = 2e3;
 var PROTECT_RECENT_MESSAGES = 3;
-var OUTPUT_RESERVE_TOKENS = 16384;
-var overheadErrorLogged = /* @__PURE__ */ new Set();
 function parseGcThreshold(threshold, modelContextLimit) {
   if (typeof threshold === "number") return threshold;
   const str = threshold ?? "100%";
@@ -10513,26 +10415,10 @@ function parseGcThreshold(threshold, modelContextLimit) {
   return modelContextLimit;
 }
 function truncateLargeToolOutputs(state, config, logger, messages) {
-  const effective = resolveEffectiveContextLimit(state, config);
-  if (!effective) return;
+  if (!state.modelContextLimit) return;
   const currentTokens = getCurrentTokenUsage(state, messages);
   if (currentTokens === 0) return;
-  const configuredThreshold = parseGcThreshold(config.gc?.majorGcThresholdPercent, effective.limit);
-  const overhead = (state.systemPromptTokens ?? 0) + OUTPUT_RESERVE_TOKENS;
-  const threshold = Math.min(configuredThreshold, effective.limit - overhead);
-  if (threshold <= 0) {
-    const sessionKey = state.sessionId ?? "unknown";
-    if (!overheadErrorLogged.has(sessionKey)) {
-      overheadErrorLogged.add(sessionKey);
-      logger.error("ACP: model context window too small to fit overhead", {
-        session: state.sessionId,
-        limit: effective.limit,
-        contextLimitSource: effective.source,
-        overhead
-      });
-    }
-    return;
-  }
+  const threshold = parseGcThreshold(config.gc?.majorGcThresholdPercent, state.modelContextLimit);
   if (currentTokens < threshold) return;
   const protectedIndex = messages.length - PROTECT_RECENT_MESSAGES;
   const candidates = [];
@@ -10577,160 +10463,11 @@ function truncateLargeToolOutputs(state, config, logger, messages) {
       truncatedCount,
       estimatedSavedTokens: Math.round(savedTokens),
       currentTokens,
-      threshold,
-      contextLimit: effective.limit,
-      contextLimitSource: effective.source
+      threshold
     });
   }
 }
 
-// lib/messages/enforce-budget.ts
-var DEFAULT_COMPLETION_RESERVE_TOKENS = 32768;
-var TRUNCATION_MARKER2 = "[truncated for context space";
-var KEEP_PREFIX_CHARS2 = 2e3;
-var KEEP_SUFFIX_CHARS2 = 2e3;
-var PROTECT_RECENT_MESSAGES2 = 3;
-var MIN_CLEAR_TOKENS = 200;
-function resolveContextWindow(state) {
-  if (typeof state.modelContextLimit === "number" && state.modelContextLimit > 0) {
-    return state.modelContextLimit;
-  }
-  return void 0;
-}
-function estimateWireTokens(state, messages) {
-  const base = getCurrentTokenUsage(state, messages);
-  if (base > 0) {
-    let baseAssistant = -1;
-    for (let i = messages.length - 1; i >= 0; i--) {
-      if (messages[i].info.role !== "assistant") continue;
-      const tokens = messages[i].info.tokens;
-      if ((tokens?.input || 0) <= 0 && (tokens?.output || 0) <= 0) continue;
-      baseAssistant = i;
-      break;
-    }
-    if (baseAssistant >= 0) {
-      let additions = 0;
-      for (let i = baseAssistant + 1; i < messages.length; i++) {
-        additions += countAllMessageTokens(messages[i]);
-      }
-      return base + additions;
-    }
-  }
-  let total = 0;
-  for (const m of messages) total += countAllMessageTokens(m);
-  return total + (state.systemPromptTokens ?? 0);
-}
-function enforceContextBudget(state, config, logger, messages) {
-  const window = resolveContextWindow(state);
-  if (window === void 0) return void 0;
-  const configuredReserve = config.compress?.completionReserveTokens;
-  const reserve = typeof configuredReserve === "number" && Number.isFinite(configuredReserve) && configuredReserve >= 0 ? configuredReserve : DEFAULT_COMPLETION_RESERVE_TOKENS;
-  const budget = window - reserve;
-  if (budget <= 0) return void 0;
-  const estimatedTokens = estimateWireTokens(state, messages);
-  if (estimatedTokens <= budget) {
-    return {
-      applied: false,
-      window,
-      reserve,
-      budget,
-      estimatedTokens,
-      finalEstimate: estimatedTokens,
-      truncatedCount: 0,
-      clearedCount: 0
-    };
-  }
-  const protectedIndex = messages.length - PROTECT_RECENT_MESSAGES2;
-  const protectedTools = new Set(config.compress?.protectedTools ?? []);
-  const candidates = [];
-  for (let mi = 0; mi < protectedIndex; mi++) {
-    if (mi === 0 && messages[mi].info.role === "user") continue;
-    const msg = messages[mi];
-    const parts = Array.isArray(msg.parts) ? msg.parts : [];
-    for (const part of parts) {
-      if (part?.type !== "tool") continue;
-      if (part.state?.status !== "completed") continue;
-      if (part.tool === "compress") continue;
-      if (protectedTools.has(part.tool)) continue;
-      const content = extractCompletedToolOutput(part);
-      if (content === void 0) continue;
-      if (content === COMPACTED_TOOL_OUTPUT_PLACEHOLDER) continue;
-      const tokens = countTokens2(content);
-      if (tokens <= 0) continue;
-      candidates.push({ part, content, tokens, index: mi });
-    }
-  }
-  let saved = 0;
-  let truncatedCount = 0;
-  let clearedCount = 0;
-  const truncatable = candidates.filter(
-    (c) => !c.content.includes(TRUNCATION_MARKER2) && c.content.length > KEEP_PREFIX_CHARS2 + KEEP_SUFFIX_CHARS2
-  ).sort((a, b) => b.tokens - a.tokens);
-  for (const c of truncatable) {
-    if (estimatedTokens - saved <= budget) break;
-    const prefix = c.content.slice(0, KEEP_PREFIX_CHARS2);
-    const suffix = c.content.slice(-KEEP_SUFFIX_CHARS2);
-    const truncated = prefix + `
-
-...${TRUNCATION_MARKER2} \u2014 original ~${c.tokens} tokens]...
-
-` + suffix;
-    if (truncated.length >= c.content.length) continue;
-    c.part.state.output = truncated;
-    saved += c.tokens - countTokens2(truncated);
-    truncatedCount++;
-  }
-  if (estimatedTokens - saved > budget) {
-    const clearable = candidates.filter((c) => {
-      const out = extractCompletedToolOutput(c.part);
-      return out !== void 0 && out !== COMPACTED_TOOL_OUTPUT_PLACEHOLDER && countTokens2(out) > MIN_CLEAR_TOKENS;
-    }).sort((a, b) => a.index - b.index);
-    for (const c of clearable) {
-      if (estimatedTokens - saved <= budget) break;
-      const current = extractCompletedToolOutput(c.part);
-      if (current === void 0 || current === COMPACTED_TOOL_OUTPUT_PLACEHOLDER) continue;
-      c.part.state.output = COMPACTED_TOOL_OUTPUT_PLACEHOLDER;
-      saved += countTokens2(current) - countTokens2(COMPACTED_TOOL_OUTPUT_PLACEHOLDER);
-      clearedCount++;
-    }
-  }
-  const finalEstimate = Math.max(0, estimatedTokens - saved);
-  if (truncatedCount > 0 || clearedCount > 0) {
-    logger.warn("Context budget guard: pruned tool outputs to fit the request", {
-      session: state.sessionId,
-      estimatedTokens: Math.round(estimatedTokens),
-      budget,
-      window,
-      reserve,
-      truncatedCount,
-      clearedCount,
-      estimatedSavedTokens: Math.round(saved),
-      finalEstimate: Math.round(finalEstimate)
-    });
-  }
-  if (finalEstimate > budget) {
-    logger.warn(
-      "Context budget guard: still over budget after pruning all compressible tool outputs; the request may be rejected by the model",
-      {
-        session: state.sessionId,
-        finalEstimate: Math.round(finalEstimate),
-        budget,
-        window
-      }
-    );
-  }
-  return {
-    applied: truncatedCount > 0 || clearedCount > 0,
-    window,
-    reserve,
-    budget,
-    estimatedTokens,
-    finalEstimate,
-    truncatedCount,
-    clearedCount
-  };
-}
-
 // lib/commands/context.ts
 function analyzeTokens(state, messages) {
   const breakdown = {
@@ -11687,12 +11424,11 @@ function runBatchCleanup(state, config, logger, messages) {
     mergedCount: 0,
     savedTokens: 0
   };
-  const effective = resolveEffectiveContextLimit(state, config);
-  if (!effective) {
+  if (!state.modelContextLimit || state.modelContextLimit <= 0) {
     return noop;
   }
   const currentTokens = getCurrentTokenUsage(state, messages);
-  if (currentTokens < effective.limit) {
+  if (currentTokens < state.modelContextLimit) {
     return noop;
   }
   const maxMergedLength = config.gc.maxOldGenSummaryLength;
@@ -11709,8 +11445,7 @@ function runBatchCleanup(state, config, logger, messages) {
     mergedCount: result.mergedCount,
     savedTokens: result.savedTokens,
     currentTokens,
-    contextLimit: effective.limit,
-    contextLimitSource: effective.source
+    contextLimit: state.modelContextLimit
   });
   return {
     tier: 3,
@@ -11744,6 +11479,11 @@ function createSystemPromptHandler(registry4, logger, config, prompts) {
       input.model?.limit?.context
     );
     const state = input.sessionID ? registry4.get(input.sessionID) : void 0;
+    if (state && input.model?.limit?.context) {
+      state.modelContextLimit = input.model.limit.context;
+      state.modelProviderID = input.model?.providerID;
+      state.modelID = input.model?.id;
+    }
     if (!state || state.isSubAgent && !config.allowSubAgents) {
       return;
     }
@@ -11752,23 +11492,6 @@ function createSystemPromptHandler(registry4, logger, config, prompts) {
       logger.info("Skipping DCP system prompt injection for internal agent");
       return;
     }
-    if (input.model?.limit?.context) {
-      const limit = input.model.limit.context;
-      const providerID = input.model?.providerID;
-      const modelID = input.model?.id;
-      const changed = state.modelContextLimit !== limit || providerID !== void 0 && state.modelProviderID !== providerID || modelID !== void 0 && state.modelID !== modelID;
-      state.modelContextLimit = limit;
-      if (providerID !== void 0) {
-        state.modelProviderID = providerID;
-      }
-      if (modelID !== void 0) {
-        state.modelID = modelID;
-      }
-      if (changed) {
-        saveSessionState(state, logger).catch(() => {
-        });
-      }
-    }
     const effectivePermission = compressPermission(state, config);
     if (effectivePermission === "deny") {
       return;
@@ -11813,17 +11536,10 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
         config
       );
       const requestModel = lastUserMessage.info.model;
-      let requestModelLimit = registry4.resolveModelLimit(
+      const requestModelLimit = registry4.resolveModelLimit(
         requestModel?.providerID,
         requestModel?.modelID
       );
-      if (requestModelLimit === void 0 && requestModel?.providerID && requestModel?.modelID) {
-        requestModelLimit = await registry4.hydrateAndResolve(
-          client,
-          requestModel.providerID,
-          requestModel.modelID
-        );
-      }
       const prevModelID = state.modelID;
       if (requestModelLimit !== void 0) {
         state.modelContextLimit = requestModelLimit;
@@ -11853,16 +11569,6 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
         });
       }
       await updatePerTurnState(state, logger, messages);
-      if (state.modelContextLimit === void 0 && !state.noContextLimitWarned && requestModel?.providerID && requestModel?.modelID && registry4.resolveModelLimit(requestModel.providerID, requestModel.modelID) === void 0) {
-        state.noContextLimitWarned = true;
-        logger.warn(
-          'Model reports no context window and the catalog has no entry for it; all percentage thresholds (min/max/emergency, GC) and the context-budget guard are disabled. Set the model limit in opencode.json (e.g. "limit": {"context": 262144, "output": 16384}) to enable them (also fixes the 32000 max_tokens fallback); an absolute compress.maxContextLimit in acp.jsonc only enables proactive nudges, not the guard.',
-          {
-            session: state.sessionId,
-            model: `${requestModel.providerID}/${requestModel.modelID}`
-          }
-        );
-      }
     }
     syncCompressPermissionState(state, config, hostPermissions, output.messages);
     if (state.isSubAgent && !config.allowSubAgents) {
@@ -11888,11 +11594,10 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
       }
     }
     ensureBuiltinFiltersRegistered();
-    const effectiveLimit = resolveEffectiveContextLimit(state, config);
     applyMessageFilters(output.messages, config.messageFilters, logger, {
       sessionId: state.sessionId ?? "",
       isSubAgent: state.isSubAgent,
-      modelContextLimit: effectiveLimit?.limit
+      modelContextLimit: state.modelContextLimit
     });
     cacheSystemPromptTokens(state, output.messages);
     assignMessageRefs(state, output.messages);
@@ -11912,7 +11617,6 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
     const prePruneTokens = getCurrentTokenUsage(state, output.messages);
     prune(state, logger, config, output.messages);
     truncateLargeToolOutputs(state, config, logger, output.messages);
-    enforceContextBudget(state, config, logger, output.messages);
     hideConsumedCompressCalls(state, output.messages);
     assignMessageRefs(state, output.messages);
     const compressionPriorities = buildPriorityMap(config, state, output.messages);
@@ -11944,31 +11648,14 @@ ${text}`);
     stripStaleMetadata(output.messages);
     dropEmptyMessages(output.messages);
     const postTokens = getCurrentTokenUsage(state, output.messages);
-    if (postTokens !== void 0 && effectiveLimit) {
-      const budget = effectiveLimit.limit - (state.systemPromptTokens ?? 0) - OUTPUT_RESERVE_TOKENS;
-      if (postTokens > budget) {
-        logger.error(
-          "ACP hard guard: context exceeds model budget after in-flight reduction",
-          {
-            session: state.sessionId,
-            postTokens,
-            budget,
-            contextLimit: effectiveLimit.limit,
-            contextLimitSource: effectiveLimit.source,
-            hint: "request will likely be rejected; run /compact or start a new session"
-          }
-        );
-      }
-    }
     logger.info("Chat transform complete", {
       session: state.sessionId,
       model: state.modelID,
       messages: output.messages.length,
       prePruneTokens,
       postTokens,
-      contextLimit: effectiveLimit?.limit,
-      contextLimitSource: effectiveLimit?.source,
-      usagePct: postTokens !== void 0 && effectiveLimit ? `${(postTokens / effectiveLimit.limit * 100).toFixed(1)}%` : void 0,
+      contextLimit: state.modelContextLimit,
+      usagePct: postTokens !== void 0 && state.modelContextLimit ? `${(postTokens / state.modelContextLimit * 100).toFixed(1)}%` : void 0,
       nudged: state.nudges.shouldInjectThisTurn
     });
     if (state.sessionId) {
@@ -12000,7 +11687,12 @@ function createCommandExecuteHandler(client, registry4, logger, config, workingD
         path: { id: input.sessionID }
       });
       const messages = filterMessages(messagesResponse.data || messagesResponse);
-      const state = await registry4.getOrCreate(client, input.sessionID, messages, config);
+      const state = await registry4.getOrCreate(
+        client,
+        input.sessionID,
+        messages,
+        config
+      );
       syncCompressPermissionState(state, config, hostPermissions, messages);
       const commandCtx = {
         client,
@@ -12014,7 +11706,7 @@ function createCommandExecuteHandler(client, registry4, logger, config, workingD
       const sub = input.arguments?.trim().toLowerCase();
       if (sub === "stats" || sub === "status" || sub === "") {
         await handleStatsCommand(commandCtx);
-        return;
+        throw new Error("__DCP_CONTEXT_HANDLED__");
       }
       if (sub === "export" || sub.startsWith("export ")) {
         const exportArgs = input.arguments?.trim().slice("export".length).trim() || "";
@@ -12022,10 +11714,17 @@ function createCommandExecuteHandler(client, registry4, logger, config, workingD
         throw new Error("__DCP_CONTEXT_HANDLED__");
       }
       if (sub === "help") {
-        await sendIgnoredMessage(client, input.sessionID, buildHelpText(), {}, logger);
+        await sendIgnoredMessage(
+          client,
+          input.sessionID,
+          buildHelpText(),
+          {},
+          logger
+        );
         throw new Error("__DCP_CONTEXT_HANDLED__");
       }
       await handleContextCommand(commandCtx);
+      throw new Error("__DCP_CONTEXT_HANDLED__");
     }
   };
 }
@@ -12096,7 +11795,9 @@ function createEventHandler(registry4, logger) {
       return;
     }
     if (typeof part.callID === "string" && typeof part.messageID === "string") {
-      timing.startsByCallId.delete(buildCompressionTimingKey(part.messageID, part.callID));
+      timing.startsByCallId.delete(
+        buildCompressionTimingKey(part.messageID, part.callID)
+      );
     }
   };
 }
@@ -12373,7 +12074,7 @@ var server = (async (ctx) => {
   }
   const logger = new Logger(config.debug, config.debug ? "debug" : config.logLevel);
   logger.info("ACP plugin initialized", {
-    version: true ? "1.16.0-pr.360.128" : "dev",
+    version: true ? "1.16.0-pr.365.125" : "dev",
     workspace: ctx.directory,
     logLevel: logger.level,
     debug: config.debug,