@infituit/opencode-acp 2.0.0-alpha.8 → 2.0.0-alpha.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -109,7 +109,13 @@ ACP 使用自己的配置文件,按以下顺序加载:
109
109
  }
110
110
  ```
111
111
 
112
- 上下文阈值接受 token 数或模型上下文窗口的百分比。`maxContextLimit` 是触发压缩提示的软阈值,不是硬性截断上限;达到阈值不代表模型一定立即执行压缩。根据模型的可用上下文调整阈值。
112
+ 上下文阈值接受 token 数或模型上下文窗口的百分比。`maxContextLimit` 是触发压缩提示的软阈值,不是模型的真实上下文窗口,也不是硬性截断上限;达到阈值不代表模型一定立即执行压缩。根据模型的可用上下文调整阈值。
113
+
114
+ `compress.contextLimitFallback` 是模型上下文窗口未知时使用的备用 token 数,默认 `128000`,设为 `0` 可禁用此备用值。从 OpenCode 模型目录(catalog)等来源获知的有效模型窗口优先于 fallback;fallback 不会扩大模型的真实容量。
115
+
116
+ - fallback 偏小,可能使压缩提醒(nudge)、GC 等依赖上下文窗口的机制提前触发;偏大则可能高估可用预算,来不及在真实窗口耗尽前提醒或回收。
117
+ - 优先在 OpenCode 的 provider/model 配置中补齐或修正实际的 `limit.context`(参见 [OpenCode 模型配置](https://opencode.ai/v2/docs/models)),不要盲目调大 fallback 来延后提醒。
118
+ - 禁用 fallback 后,若仍无法获知模型窗口,依赖该窗口的机制可能不可用;这不代表模型拥有无限上下文。
113
119
 
114
120
  ### 按 Provider / 按模型覆盖
115
121
 
package/dist/index.js CHANGED
@@ -3568,7 +3568,7 @@ function applyPendingCompressionDurations(state) {
3568
3568
  }
3569
3569
 
3570
3570
  // src/state/model-limits.ts
3571
- function createModelLimitCatalog() {
3571
+ function createModelLimitCatalog(logger) {
3572
3572
  const modelLimits = /* @__PURE__ */ new Map();
3573
3573
  return {
3574
3574
  record(providerId, modelId, limit) {
@@ -3580,14 +3580,27 @@ function createModelLimitCatalog() {
3580
3580
  return modelLimits.get(`${providerId}/${modelId}`);
3581
3581
  },
3582
3582
  /**
3583
- * Best-effort one-time seed from the V2 model catalog. Never throws;
3583
+ * Best-effort load from the V2 model catalog. Never throws;
3584
3584
  * returns the number of model-limit entries recorded.
3585
3585
  */
3586
3586
  async hydrateFromClient(client) {
3587
3587
  try {
3588
3588
  const v2Client = client;
3589
+ if (typeof v2Client?.model?.list !== "function") {
3590
+ logger?.warn("Model limit catalog load failed", {
3591
+ reason: "api_unavailable",
3592
+ catalogSize: modelLimits.size
3593
+ });
3594
+ return 0;
3595
+ }
3589
3596
  const response = await v2Client.model.list();
3590
- if (!Array.isArray(response.data)) return 0;
3597
+ if (!Array.isArray(response?.data)) {
3598
+ logger?.warn("Model limit catalog load failed", {
3599
+ reason: "invalid_response",
3600
+ catalogSize: modelLimits.size
3601
+ });
3602
+ return 0;
3603
+ }
3591
3604
  let recorded = 0;
3592
3605
  for (const value of response.data) {
3593
3606
  if (!value || typeof value !== "object") continue;
@@ -3598,8 +3611,19 @@ function createModelLimitCatalog() {
3598
3611
  modelLimits.set(`${model.providerID}/${model.id}`, model.limit.context);
3599
3612
  recorded++;
3600
3613
  }
3614
+ logger?.info("Model limit catalog loaded", {
3615
+ reason: response.data.length === 0 ? "empty_catalog" : recorded === 0 ? "no_valid_limits" : "loaded",
3616
+ entries: response.data.length,
3617
+ recorded,
3618
+ skipped: response.data.length - recorded,
3619
+ catalogSize: modelLimits.size
3620
+ });
3601
3621
  return recorded;
3602
3622
  } catch {
3623
+ logger?.warn("Model limit catalog load failed", {
3624
+ reason: "request_failed",
3625
+ catalogSize: modelLimits.size
3626
+ });
3603
3627
  return 0;
3604
3628
  }
3605
3629
  }
@@ -5185,10 +5209,12 @@ async function updatePerTurnState(state, logger, messages) {
5185
5209
  state.currentTurn = countTurns(state, messages);
5186
5210
  }
5187
5211
  var REGISTRY_SOFT_CAP = 32;
5212
+ var MODEL_LIMIT_RETRY_INTERVAL_MS = 3e4;
5188
5213
  var SessionStateRegistry = class {
5189
5214
  constructor(logger, projectDir) {
5190
5215
  this.logger = logger;
5191
5216
  this.projectDir = projectDir;
5217
+ this.catalog = createModelLimitCatalog(logger);
5192
5218
  }
5193
5219
  logger;
5194
5220
  projectDir;
@@ -5202,38 +5228,41 @@ var SessionStateRegistry = class {
5202
5228
  // model named on the request's user message instead of waiting one turn
5203
5229
  // for the system hook. Shared implementation — the test registry stub
5204
5230
  // composes the same factory.
5205
- catalog = createModelLimitCatalog();
5231
+ catalog;
5206
5232
  recordModelLimit(providerId, modelId, limit) {
5207
5233
  this.catalog.record(providerId, modelId, limit);
5208
5234
  }
5209
5235
  resolveModelLimit(providerId, modelId) {
5210
5236
  return this.catalog.resolve(providerId, modelId);
5211
5237
  }
5212
- /**
5213
- * Best-effort one-time seed from the host's provider catalog
5214
- * V2 model catalog. Never throws; returns the number of model-limit entries recorded.
5215
- * returns the number of model-limit entries recorded.
5216
- */
5238
+ /** Best-effort seed from the V2 model catalog; returns the number of recorded entries. */
5217
5239
  hydrateModelLimitsFromClient(client) {
5218
5240
  return this.catalog.hydrateFromClient(client);
5219
5241
  }
5220
- // [FIX #346] The init-time seed (above) is fire-and-forget and races
5221
- // server readiness: in headless spawn+resume mode the provider-config
5222
- // call can fail before the server is up, leaving the catalog empty for
5223
- // the process's lifetime. During a request the server is guaranteed up
5224
- // (we are inside its pipeline), so on a catalog miss we retry hydration
5225
- // once per process before giving up (the fallback limit then applies).
5226
- // The in-flight promise (not a boolean) lets concurrent callers await the
5227
- // same hydration instead of skipping it.
5242
+ // Initialization can precede provider readiness. Retry request-time misses,
5243
+ // sharing in-flight work and cooling down completed attempts so an unknown
5244
+ // model cannot cause a catalog request on every turn. Cache hits never wait.
5228
5245
  lazyHydration;
5246
+ nextModelLimitRetryAt = 0;
5229
5247
  async hydrateAndResolve(client, providerId, modelId) {
5230
5248
  const existing = this.catalog.resolve(providerId, modelId);
5231
5249
  if (existing !== void 0) {
5232
5250
  return existing;
5233
5251
  }
5234
- this.lazyHydration ??= this.catalog.hydrateFromClient(client);
5252
+ if (!this.lazyHydration && Date.now() >= this.nextModelLimitRetryAt) {
5253
+ this.lazyHydration = this.catalog.hydrateFromClient(client).finally(() => {
5254
+ this.nextModelLimitRetryAt = Date.now() + MODEL_LIMIT_RETRY_INTERVAL_MS;
5255
+ this.lazyHydration = void 0;
5256
+ });
5257
+ }
5235
5258
  await this.lazyHydration;
5236
- return this.catalog.resolve(providerId, modelId);
5259
+ const limit = this.catalog.resolve(providerId, modelId);
5260
+ this.logger.debug("Model limit catalog lookup", {
5261
+ model: `${providerId}/${modelId}`,
5262
+ reason: limit === void 0 ? "model_not_found" : "model",
5263
+ retryAfterMs: limit === void 0 ? Math.max(0, this.nextModelLimitRetryAt - Date.now()) : 0
5264
+ });
5265
+ return limit;
5237
5266
  }
5238
5267
  get(sessionId) {
5239
5268
  return this.states.get(sessionId);
@@ -11517,7 +11546,7 @@ import { writeFile as writeFile2, mkdir as mkdir2 } from "fs/promises";
11517
11546
  import { join as join5 } from "path";
11518
11547
  import { existsSync as existsSync5 } from "fs";
11519
11548
  import { homedir as homedir3 } from "os";
11520
- var LOG_VERSION = true ? "2.0.0-alpha.8" : "dev";
11549
+ var LOG_VERSION = true ? "2.0.0-alpha.9" : "dev";
11521
11550
  var LEVEL_RANK = {
11522
11551
  debug: 10,
11523
11552
  info: 20,
@@ -13399,11 +13428,14 @@ function createChatMessageTransformHandler(client, registry4, logger, config, pr
13399
13428
  checkpoint("permissions");
13400
13429
  if (state.modelContextLimit === void 0 && !state.noContextLimitWarned && requestModel?.providerID && requestModel?.modelID && registry4.resolveModelLimit(requestModel.providerID, requestModel.modelID) === void 0) {
13401
13430
  state.noContextLimitWarned = true;
13431
+ const effectiveLimit2 = resolveEffectiveContextLimit(state, config);
13402
13432
  logger.warn(
13403
- '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.',
13433
+ effectiveLimit2 ? "Model context window is unavailable in the catalog; using compress.contextLimitFallback. A smaller fallback may trigger compression early; a larger one may overestimate the available budget. Check the provider/model limit in OpenCode and the model catalog load logs. Catalog misses are retried on later requests with a cooldown." : "Model context window is unavailable in the catalog and compress.contextLimitFallback is disabled; checks that require a context window cannot run. Absolute compress.maxContextLimit can still trigger proactive nudges. Check the provider/model limit in OpenCode and the model catalog load logs. Catalog misses are retried on later requests with a cooldown.",
13404
13434
  {
13405
13435
  session: state.sessionId,
13406
- model: `${requestModel.providerID}/${requestModel.modelID}`
13436
+ model: `${requestModel.providerID}/${requestModel.modelID}`,
13437
+ contextLimit: effectiveLimit2?.limit ?? 0,
13438
+ contextLimitSource: effectiveLimit2?.source ?? "unknown"
13407
13439
  }
13408
13440
  );
13409
13441
  }