opencode-acp 1.16.0-pr.297.127 → 1.16.0-pr.349.122

package/dist/index.js

@@ -890,6 +890,7 @@ var VALID_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "compress.modelMaxLimits",
   "compress.modelMinLimits",
   "compress.providers",
+  "compress.contextLimitFallback",
   "compress.nudgeFrequency",
   "compress.minNudgeContextPercent",
   "compress.nudgeGrowthTokens",
@@ -1521,6 +1522,20 @@ 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({
@@ -1702,6 +1717,7 @@ var defaultConfig = {
     summaryBuffer: true,
     maxContextLimit: "80%",
     minContextLimit: "80%",
+    contextLimitFallback: 128e3,
     nudgeFrequency: 5,
     minNudgeContextPercent: 5,
     iterationNudgeThreshold: 15,
@@ -1872,6 +1888,7 @@ 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,
@@ -2635,6 +2652,16 @@ 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() {
@@ -4506,6 +4533,24 @@ 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);
   }
@@ -6657,6 +6702,7 @@ 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;
@@ -6664,7 +6710,7 @@ function resolveContextTokenLimit(config, state, providerId, modelId, threshold)
     if (typeof limit === "number") {
       return limit;
     }
-    if (!limit.endsWith("%") || state.modelContextLimit === void 0) {
+    if (!limit.endsWith("%") || effectiveLimit === void 0) {
       return void 0;
     }
     const parsedPercent = parseFloat(limit.slice(0, -1));
@@ -6673,7 +6719,7 @@ 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 * state.modelContextLimit);
+    return Math.round(clampedPercent / 100 * effectiveLimit.limit);
   };
   if (threshold === "max") {
     const nestedLimit = resolveCompressOverrides(config, providerId, modelId).maxContextLimit;
@@ -6724,11 +6770,12 @@ function isContextOverLimits(config, state, providerId, modelId, messages) {
       if (!overMaxLimit) break;
     }
   }
+  const effectiveLimit = resolveEffectiveContextLimit(state, config);
   return {
     overMaxLimit,
     overMinLimit,
     currentTokens,
-    modelContextLimit: state.modelContextLimit
+    modelContextLimit: effectiveLimit?.limit
   };
 }
 ensureBuiltinTriggerPolicyRegistered();
@@ -8807,8 +8854,9 @@ function createDecompressTool(factoryCtx) {
     async execute(args, toolCtx) {
       const ctx = resolveToolContext(factoryCtx, toolCtx.sessionID);
       const { rawMessages } = await prepareDecompressSession(ctx, toolCtx);
-      const contextUsageBefore = ctx.state.modelContextLimit ? Math.round(
-        getCurrentTokenUsage(ctx.state, rawMessages) / ctx.state.modelContextLimit * 100
+      const effectiveLimitBefore = resolveEffectiveContextLimit(ctx.state, ctx.config);
+      const contextUsageBefore = effectiveLimitBefore ? Math.round(
+        getCurrentTokenUsage(ctx.state, rawMessages) / effectiveLimitBefore.limit * 100
       ) : void 0;
       const resolved = resolveTargets(args, ctx.state, rawMessages, ctx.logger);
       if (!resolved.ok) {
@@ -8872,8 +8920,9 @@ function createDecompressTool(factoryCtx) {
         0,
         ctx.state.stats.totalPruneTokens - restoredTokens
       );
-      const contextUsageAfter = ctx.state.modelContextLimit ? Math.round(
-        getCurrentTokenUsage(ctx.state, rawMessages) / ctx.state.modelContextLimit * 100
+      const effectiveLimitAfter = resolveEffectiveContextLimit(ctx.state, ctx.config);
+      const contextUsageAfter = effectiveLimitAfter ? Math.round(
+        getCurrentTokenUsage(ctx.state, rawMessages) / effectiveLimitAfter.limit * 100
       ) : void 0;
       await finalizeDecompressSession(ctx);
       const restoredContentPreview = buildRestoredContentPreview(
@@ -9530,7 +9579,7 @@ 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.297.127" : "dev";
+var LOG_VERSION = true ? "1.16.0-pr.349.122" : "dev";
 var LEVEL_RANK = {
   debug: 10,
   info: 20,
@@ -10354,6 +10403,8 @@ 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%";
@@ -10362,10 +10413,26 @@ function parseGcThreshold(threshold, modelContextLimit) {
   return modelContextLimit;
 }
 function truncateLargeToolOutputs(state, config, logger, messages) {
-  if (!state.modelContextLimit) return;
+  const effective = resolveEffectiveContextLimit(state, config);
+  if (!effective) return;
   const currentTokens = getCurrentTokenUsage(state, messages);
   if (currentTokens === 0) return;
-  const threshold = parseGcThreshold(config.gc?.majorGcThresholdPercent, state.modelContextLimit);
+  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;
+  }
   if (currentTokens < threshold) return;
   const protectedIndex = messages.length - PROTECT_RECENT_MESSAGES;
   const candidates = [];
@@ -10410,7 +10477,9 @@ function truncateLargeToolOutputs(state, config, logger, messages) {
       truncatedCount,
       estimatedSavedTokens: Math.round(savedTokens),
       currentTokens,
-      threshold
+      threshold,
+      contextLimit: effective.limit,
+      contextLimitSource: effective.source
     });
   }
 }
@@ -11371,11 +11440,12 @@ function runBatchCleanup(state, config, logger, messages) {
     mergedCount: 0,
     savedTokens: 0
   };
-  if (!state.modelContextLimit || state.modelContextLimit <= 0) {
+  const effective = resolveEffectiveContextLimit(state, config);
+  if (!effective) {
     return noop;
   }
   const currentTokens = getCurrentTokenUsage(state, messages);
-  if (currentTokens < state.modelContextLimit) {
+  if (currentTokens < effective.limit) {
     return noop;
   }
   const maxMergedLength = config.gc.maxOldGenSummaryLength;
@@ -11392,7 +11462,8 @@ function runBatchCleanup(state, config, logger, messages) {
     mergedCount: result.mergedCount,
     savedTokens: result.savedTokens,
     currentTokens,
-    contextLimit: state.modelContextLimit
+    contextLimit: effective.limit,
+    contextLimitSource: effective.source
   });
   return {
     tier: 3,
@@ -11426,11 +11497,6 @@ 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;
     }
@@ -11439,6 +11505,23 @@ 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;
@@ -11483,10 +11566,17 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
         config
       );
       const requestModel = lastUserMessage.info.model;
-      const requestModelLimit = registry4.resolveModelLimit(
+      let 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;
@@ -11541,10 +11631,11 @@ 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: state.modelContextLimit
+      modelContextLimit: effectiveLimit?.limit
     });
     cacheSystemPromptTokens(state, output.messages);
     assignMessageRefs(state, output.messages);
@@ -11595,14 +11686,31 @@ ${text}`);
     stripStaleMetadata(output.messages);
     dropEmptyMessages(output.messages);
     const postTokens = getCurrentTokenUsage(state, output.messages);
+    if (postTokens !== void 0 && effectiveLimit) {
+      const budget = effectiveLimit.limit - (state.systemPromptTokens ?? 0) - OUTPUT_RESERVE_TOKENS;
+      if (postTokens > budget) {
+        logger.error(
+          "ACP hard guard: context exceeds model budget after in-flight reduction",
+          {
+            session: state.sessionId,
+            postTokens,
+            budget,
+            contextLimit: effectiveLimit.limit,
+            contextLimitSource: effectiveLimit.source,
+            hint: "request will likely be rejected; run /compact or start a new session"
+          }
+        );
+      }
+    }
     logger.info("Chat transform complete", {
       session: state.sessionId,
       model: state.modelID,
       messages: output.messages.length,
       prePruneTokens,
       postTokens,
-      contextLimit: state.modelContextLimit,
-      usagePct: postTokens !== void 0 && state.modelContextLimit ? `${(postTokens / state.modelContextLimit * 100).toFixed(1)}%` : void 0,
+      contextLimit: effectiveLimit?.limit,
+      contextLimitSource: effectiveLimit?.source,
+      usagePct: postTokens !== void 0 && effectiveLimit ? `${(postTokens / effectiveLimit.limit * 100).toFixed(1)}%` : void 0,
       nudged: state.nudges.shouldInjectThisTurn
     });
     if (state.sessionId) {
@@ -11653,7 +11761,7 @@ 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() || "";
@@ -11671,6 +11779,7 @@ function createCommandExecuteHandler(client, registry4, logger, config, workingD
         throw new Error("__DCP_CONTEXT_HANDLED__");
       }
       await handleContextCommand(commandCtx);
+      throw new Error("__DCP_CONTEXT_HANDLED__");
     }
   };
 }
@@ -12020,7 +12129,7 @@ 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.297.127" : "dev",
+    version: true ? "1.16.0-pr.349.122" : "dev",
     workspace: ctx.directory,
     logLevel: logger.level,
     debug: config.debug,