opencode-acp 1.14.25-pr.349.71 → 1.14.25-pr.350.73

package/dist/index.js

@@ -888,7 +888,6 @@ var VALID_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "compress.minContextLimit",
   "compress.modelMaxLimits",
   "compress.modelMinLimits",
-  "compress.contextLimitFallback",
   "compress.nudgeFrequency",
   "compress.minNudgeContextPercent",
   "compress.nudgeGrowthTokens",
@@ -1312,20 +1311,6 @@ function validateConfigTypes(config) {
       }
       validateModelLimits("compress.modelMaxLimits", compress.modelMaxLimits);
       validateModelLimits("compress.modelMinLimits", compress.modelMinLimits);
-      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({
@@ -1503,7 +1488,6 @@ var defaultConfig = {
     summaryBuffer: true,
     maxContextLimit: "80%",
     minContextLimit: "80%",
-    contextLimitFallback: 128e3,
     nudgeFrequency: 5,
     minNudgeContextPercent: 15,
     iterationNudgeThreshold: 15,
@@ -1639,7 +1623,6 @@ function mergeCompress(base, override) {
     minContextLimit: override.minContextLimit ?? base.minContextLimit,
     modelMaxLimits: override.modelMaxLimits ?? base.modelMaxLimits,
     modelMinLimits: override.modelMinLimits ?? base.modelMinLimits,
-    contextLimitFallback: override.contextLimitFallback ?? base.contextLimitFallback,
     nudgeFrequency: override.nudgeFrequency ?? base.nudgeFrequency,
     minNudgeContextPercent: override.minNudgeContextPercent ?? base.minNudgeContextPercent,
     nudgeGrowthTokens: override.nudgeGrowthTokens,
@@ -1659,7 +1642,8 @@ function mergeCompress(base, override) {
     lastSegmentSoftBlock: override.lastSegmentSoftBlock ?? base.lastSegmentSoftBlock,
     preserveRecentMessages: override.preserveRecentMessages ?? base.preserveRecentMessages,
     preserveRecentTokens: override.preserveRecentTokens ?? base.preserveRecentTokens,
-    preserveLastUserMessage: override.preserveLastUserMessage ?? base.preserveLastUserMessage
+    preserveLastUserMessage: override.preserveLastUserMessage ?? base.preserveLastUserMessage,
+    completionReserveTokens: override.completionReserveTokens ?? base.completionReserveTokens
   };
 }
 function mergeCommands(base, override) {
@@ -2382,16 +2366,6 @@ 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 getStorageDir() {
@@ -4237,24 +4211,6 @@ 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).
-  lazyHydrated = false;
-  async hydrateAndResolve(client, providerId, modelId) {
-    const existing = this.catalog.resolve(providerId, modelId);
-    if (existing !== void 0) {
-      return existing;
-    }
-    if (!this.lazyHydrated) {
-      this.lazyHydrated = true;
-      await this.catalog.hydrateFromClient(client);
-    }
-    return this.catalog.resolve(providerId, modelId);
-  }
   get(sessionId) {
     return this.states.get(sessionId);
   }
@@ -4346,7 +4302,8 @@ function createSessionState() {
     modelProviderID: void 0,
     modelID: void 0,
     systemPromptTokens: void 0,
-    qualityGateRetryPending: false
+    qualityGateRetryPending: false,
+    noContextLimitWarned: false
   };
 }
 function resetSessionState(state) {
@@ -4388,6 +4345,7 @@ function resetSessionState(state) {
   state.modelID = void 0;
   state.systemPromptTokens = void 0;
   state.qualityGateRetryPending = false;
+  state.noContextLimitWarned = false;
 }
 async function ensureSessionInitialized(client, state, sessionId, logger, messages, config) {
   if (state.sessionId === sessionId) {
@@ -6925,7 +6883,6 @@ 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;
@@ -6933,7 +6890,7 @@ 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));
@@ -6942,7 +6899,7 @@ 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);
   };
   const modelLimits = threshold === "max" ? config.compress.modelMaxLimits : config.compress.modelMinLimits;
   if (modelLimits && providerId !== void 0 && modelId !== void 0) {
@@ -6987,12 +6944,11 @@ 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();
@@ -8434,9 +8390,8 @@ 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) {
@@ -8500,9 +8455,8 @@ 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(
@@ -9159,7 +9113,7 @@ import { writeFile as writeFile2, mkdir as mkdir2 } from "fs/promises";
 import { join as join3 } from "path";
 import { existsSync as existsSync3 } from "fs";
 import { homedir as homedir3 } from "os";
-var LOG_VERSION = true ? "1.14.25-pr.349.71" : "dev";
+var LOG_VERSION = true ? "1.14.25-pr.350.73" : "dev";
 var LEVEL_RANK = {
   debug: 10,
   info: 20,
@@ -9983,7 +9937,6 @@ 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;
 function parseGcThreshold(threshold, modelContextLimit) {
   if (typeof threshold === "number") return threshold;
   const str = threshold ?? "100%";
@@ -9992,22 +9945,10 @@ 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) {
-    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 = [];
@@ -10052,13 +9993,153 @@ 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 lastAssistant = -1;
+    for (let i = messages.length - 1; i >= 0; i--) {
+      if (messages[i].info.role === "assistant") {
+        lastAssistant = i;
+        break;
+      }
+    }
+    let additions = 0;
+    for (let i = lastAssistant + 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 reserve = config.compress?.completionReserveTokens ?? 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;
+    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 = {
@@ -11015,12 +11096,11 @@ 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;
@@ -11037,8 +11117,7 @@ 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,
@@ -11072,6 +11151,11 @@ 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;
     }
@@ -11080,19 +11164,6 @@ 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 || state.modelProviderID !== providerID || state.modelID !== modelID;
-      state.modelContextLimit = limit;
-      state.modelProviderID = providerID;
-      state.modelID = modelID;
-      if (changed) {
-        saveSessionState(state, logger).catch(() => {
-        });
-      }
-    }
     const effectivePermission = compressPermission(state, config);
     if (effectivePermission === "deny") {
       return;
@@ -11137,17 +11208,10 @@ 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;
@@ -11177,6 +11241,16 @@ 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}) or set an absolute compress.maxContextLimit in acp.jsonc.',
+          {
+            session: state.sessionId,
+            model: `${requestModel.providerID}/${requestModel.modelID}`
+          }
+        );
+      }
     }
     syncCompressPermissionState(state, config, hostPermissions, output.messages);
     if (state.isSubAgent && !config.allowSubAgents) {
@@ -11184,11 +11258,10 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
     }
     stripHallucinations(output.messages);
     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);
@@ -11208,6 +11281,7 @@ 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);
@@ -11239,31 +11313,14 @@ ${text}`);
     stripStaleMetadata(output.messages);
     dropEmptyMessages(output.messages);
     const postTokens = getCurrentTokenUsage(state, output.messages);
-    if (postTokens !== void 0 && effectiveLimit) {
-      const budget = effectiveLimit.limit - (state.systemPromptTokens ?? 0) - OUTPUT_RESERVE_TOKENS;
-      if (postTokens > budget) {
-        logger.error(
-          "ACP hard guard: context exceeds model budget after in-flight reduction",
-          {
-            session: state.sessionId,
-            postTokens,
-            budget,
-            contextLimit: effectiveLimit.limit,
-            contextLimitSource: effectiveLimit.source,
-            hint: "request will likely be rejected; run /compact or start a new session"
-          }
-        );
-      }
-    }
     logger.info("Chat transform complete", {
       session: state.sessionId,
       model: state.modelID,
       messages: output.messages.length,
       prePruneTokens,
       postTokens,
-      contextLimit: effectiveLimit?.limit,
-      contextLimitSource: effectiveLimit?.source,
-      usagePct: postTokens !== void 0 && effectiveLimit ? `${(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) {
@@ -11650,7 +11707,7 @@ var server = (async (ctx) => {
   }
   const logger = new Logger(config.debug, config.debug ? "debug" : config.logLevel);
   logger.info("ACP plugin initialized", {
-    version: true ? "1.14.25-pr.349.71" : "dev",
+    version: true ? "1.14.25-pr.350.73" : "dev",
     workspace: ctx.directory,
     logLevel: logger.level,
     debug: config.debug,