opencode-acp 1.16.0-pr.349.122 → 1.16.0-pr.360.124

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",
@@ -1522,20 +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({
@@ -1717,7 +1702,6 @@ var defaultConfig = {
     summaryBuffer: true,
     maxContextLimit: "80%",
     minContextLimit: "80%",
-    contextLimitFallback: 128e3,
     nudgeFrequency: 5,
     minNudgeContextPercent: 5,
     iterationNudgeThreshold: 15,
@@ -1888,7 +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,
@@ -2652,16 +2635,6 @@ 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() {
@@ -4533,24 +4506,6 @@ 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);
   }
@@ -6702,7 +6657,6 @@ 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;
@@ -6710,7 +6664,7 @@ 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));
@@ -6719,7 +6673,7 @@ 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;
@@ -6770,12 +6724,11 @@ 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();
@@ -7081,13 +7034,10 @@ 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)) {
-      let tokens2 = 0;
+      const tokens2 = Math.round(countMessageCharacters(msg) / 4);
       const tools = /* @__PURE__ */ new Set();
       for (const part of msg.parts || []) {
-        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);
+        if (part.type !== "text" && part.type !== "reasoning") {
           const toolName = part?.tool;
           const callID = part?.callID;
           if (toolName && callID) {
@@ -7108,15 +7058,13 @@ function buildCompressibleRanges(messages, state, protectedTools = [], protected
       protectedMsgInfo.push({ ref, refNum: rn, tokens: tokens2, tools: [...tools] });
       continue;
     }
-    let tokens = 0;
+    const tokens = Math.round(countMessageCharacters(msg) / 4);
     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;
       }
@@ -8854,9 +8802,8 @@ 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) {
@@ -8920,9 +8867,8 @@ 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(
@@ -9579,7 +9525,7 @@ 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.349.122" : "dev";
+var LOG_VERSION = true ? "1.16.0-pr.360.124" : "dev";
 var LEVEL_RANK = {
   debug: 10,
   info: 20,
@@ -10403,8 +10349,6 @@ 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%";
@@ -10413,26 +10357,10 @@ 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 = [];
@@ -10477,9 +10405,7 @@ function truncateLargeToolOutputs(state, config, logger, messages) {
       truncatedCount,
       estimatedSavedTokens: Math.round(savedTokens),
       currentTokens,
-      threshold,
-      contextLimit: effective.limit,
-      contextLimitSource: effective.source
+      threshold
     });
   }
 }
@@ -11440,12 +11366,11 @@ 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;
@@ -11462,8 +11387,7 @@ 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,
@@ -11497,6 +11421,11 @@ 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;
     }
@@ -11505,23 +11434,6 @@ 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;
@@ -11566,17 +11478,10 @@ 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;
@@ -11631,11 +11536,10 @@ 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);
@@ -11686,31 +11590,14 @@ ${text}`);
     stripStaleMetadata(output.messages);
     dropEmptyMessages(output.messages);
     const postTokens = getCurrentTokenUsage(state, output.messages);
-    if (postTokens !== void 0 && effectiveLimit) {
-      const budget = effectiveLimit.limit - (state.systemPromptTokens ?? 0) - OUTPUT_RESERVE_TOKENS;
-      if (postTokens > budget) {
-        logger.error(
-          "ACP hard guard: context exceeds model budget after in-flight reduction",
-          {
-            session: state.sessionId,
-            postTokens,
-            budget,
-            contextLimit: effectiveLimit.limit,
-            contextLimitSource: effectiveLimit.source,
-            hint: "request will likely be rejected; run /compact or start a new session"
-          }
-        );
-      }
-    }
     logger.info("Chat transform complete", {
       session: state.sessionId,
       model: state.modelID,
       messages: output.messages.length,
       prePruneTokens,
       postTokens,
-      contextLimit: effectiveLimit?.limit,
-      contextLimitSource: effectiveLimit?.source,
-      usagePct: postTokens !== void 0 && effectiveLimit ? `${(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) {
@@ -12129,7 +12016,7 @@ 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.349.122" : "dev",
+    version: true ? "1.16.0-pr.360.124" : "dev",
     workspace: ctx.directory,
     logLevel: logger.level,
     debug: config.debug,