fix: resolve provider-qualified context window via direct config scan

linwebs · linwebs · commit 80b8160bcabb · 2026-03-14T00:18:10.000+08:00
applyConfiguredContextWindows now writes only bare model-id keys to
MODEL_CACHE, reserving the provider-qualified key space exclusively for
raw discovery entries (e.g. OpenRouter slash-paths).

Call sites that need per-provider disambiguation (get-reply-directives,
memory-flush, agent-runner-memory) now call resolveContextTokensForModel
with the full cfg so they scan models.providers directly, rather than
relying on a qualified cache key that was never reliably populated.

resolveContextTokens in model-selection.ts reverted to remove the now-
unused provider param; resolveMemoryFlushContextWindowTokens accepts cfg
and delegates to resolveContextTokensForModel.
diff --git a/src/agents/context.ts b/src/agents/context.ts
@@ -62,7 +62,7 @@ export function applyConfiguredContextWindows(params: {
   if (!providers || typeof providers !== "object") {
     return;
   }
-  for (const [providerId, provider] of Object.entries(providers)) {
+  for (const provider of Object.values(providers)) {
     if (!Array.isArray(provider?.models)) {
       continue;
     }
@@ -73,18 +73,13 @@ export function applyConfiguredContextWindows(params: {
       if (!modelId || !contextWindow || contextWindow <= 0) {
         continue;
       }
-      // Store with provider-qualified key (e.g. "rdsec/claude-4.6-sonnet") so
-      // models that share the same id across providers but differ in context
-      // window are resolved correctly via lookupContextTokens.
-      // Use normalizeProviderId so the key format matches resolveContextTokensForModel
-      // which also normalizes before constructing the qualified lookup key.
-      if (providerId) {
-        params.cache.set(`${normalizeProviderId(providerId)}/${modelId}`, contextWindow);
-      }
-      // Also store by bare model id as a fallback for callers that only have
-      // the model name. When multiple providers define the same id the last
-      // one wins, but a provider-qualified lookup above will always take
-      // precedence when provider info is available.
+      // Only write the bare model id. The provider-qualified key space is
+      // reserved for raw discovery entries (e.g. "google-gemini-cli/gemini-3.1-pro-preview")
+      // and must not be polluted by synthetic config writes that could corrupt
+      // slash-containing model IDs (e.g. OpenRouter's "anthropic/claude-opus-4-6").
+      // Callers with provider info use resolveContextTokensForModel which scans
+      // cfg.models.providers directly and never relies on this cache for the
+      // qualified lookup.
       params.cache.set(modelId, contextWindow);
     }
   }
diff --git a/src/auto-reply/reply/agent-runner-memory.ts b/src/auto-reply/reply/agent-runner-memory.ts
@@ -289,6 +289,7 @@ export async function runMemoryFlushIfNeeded(params: {
     params.sessionEntry ??
     (params.sessionKey ? params.sessionStore?.[params.sessionKey] : undefined);
   const contextWindowTokens = resolveMemoryFlushContextWindowTokens({
+    cfg: params.cfg,
     modelId: params.followupRun.run.model ?? params.defaultModel,
     providerId: params.followupRun.run.provider,
     agentCfgContextTokens: params.agentCfgContextTokens,
diff --git a/src/auto-reply/reply/get-reply-directives.ts b/src/auto-reply/reply/get-reply-directives.ts
@@ -1,4 +1,6 @@
 import type { ExecToolDefaults } from "../../agents/bash-tools.js";
+import { resolveContextTokensForModel } from "../../agents/context.js";
+import { DEFAULT_CONTEXT_TOKENS } from "../../agents/defaults.js";
 import { resolveFastModeState } from "../../agents/fast-mode.js";
 import type { ModelAliasIndex } from "../../agents/model-selection.js";
 import { resolveSandboxRuntimeStatus } from "../../agents/sandbox.js";
@@ -17,7 +19,7 @@ import { applyInlineDirectiveOverrides } from "./get-reply-directives-apply.js";
 import { clearExecInlineDirectives, clearInlineDirectives } from "./get-reply-directives-utils.js";
 import { defaultGroupActivation, resolveGroupRequireMention } from "./groups.js";
 import { CURRENT_MESSAGE_MARKER, stripMentions, stripStructuralPrefixes } from "./mentions.js";
-import { createModelSelectionState, resolveContextTokens } from "./model-selection.js";
+import { createModelSelectionState } from "./model-selection.js";
 import { formatElevatedUnavailableMessage, resolveElevatedPermissions } from "./reply-elevated.js";
 import { stripInlineStatus } from "./reply-inline.js";
 import type { TypingController } from "./typing.js";
@@ -418,11 +420,14 @@ export async function resolveReplyDirectives(params: {
     resolvedReasoningLevel = await modelState.resolveDefaultReasoningLevel();
   }
 
-  let contextTokens = resolveContextTokens({
-    agentCfg,
-    model,
-    provider,
-  });
+  let contextTokens =
+    resolveContextTokensForModel({
+      cfg,
+      provider,
+      model,
+      contextTokensOverride: agentCfg?.contextTokens,
+      fallbackContextTokens: DEFAULT_CONTEXT_TOKENS,
+    }) ?? DEFAULT_CONTEXT_TOKENS;
 
   const initialModelLabel = `${provider}/${model}`;
   const formatModelSwitchEvent = (label: string, alias?: string) =>
diff --git a/src/auto-reply/reply/memory-flush.ts b/src/auto-reply/reply/memory-flush.ts
@@ -1,7 +1,6 @@
-import { lookupContextTokens } from "../../agents/context.js";
+import { resolveContextTokensForModel } from "../../agents/context.js";
 import { resolveCronStyleNow } from "../../agents/current-time.js";
 import { DEFAULT_CONTEXT_TOKENS } from "../../agents/defaults.js";
-import { normalizeProviderId } from "../../agents/model-selection.js";
 import { DEFAULT_PI_COMPACTION_RESERVE_TOKENS_FLOOR } from "../../agents/pi-settings.js";
 import { parseNonNegativeByteSize } from "../../config/byte-size.js";
 import type { OpenClawConfig } from "../../config/config.js";
@@ -160,19 +159,19 @@ function ensureMemoryFlushSafetyHints(text: string): string {
 }
 
 export function resolveMemoryFlushContextWindowTokens(params: {
+  cfg?: OpenClawConfig;
   modelId?: string;
   providerId?: string;
   agentCfgContextTokens?: number;
 }): number {
-  const qualified =
-    params.providerId && params.modelId
-      ? lookupContextTokens(`${normalizeProviderId(params.providerId)}/${params.modelId}`)
-      : undefined;
   return (
-    qualified ??
-    lookupContextTokens(params.modelId) ??
-    params.agentCfgContextTokens ??
-    DEFAULT_CONTEXT_TOKENS
+    resolveContextTokensForModel({
+      cfg: params.cfg,
+      provider: params.providerId,
+      model: params.modelId,
+      contextTokensOverride: params.agentCfgContextTokens,
+      fallbackContextTokens: DEFAULT_CONTEXT_TOKENS,
+    }) ?? DEFAULT_CONTEXT_TOKENS
   );
 }
 
diff --git a/src/auto-reply/reply/model-selection.ts b/src/auto-reply/reply/model-selection.ts
@@ -603,16 +603,8 @@ export function resolveModelDirectiveSelection(params: {
 export function resolveContextTokens(params: {
   agentCfg: NonNullable<NonNullable<OpenClawConfig["agents"]>["defaults"]> | undefined;
   model: string;
-  provider?: string;
 }): number {
-  const qualified =
-    params.provider
-      ? lookupContextTokens(`${normalizeProviderId(params.provider)}/${params.model}`)
-      : undefined;
   return (
-    params.agentCfg?.contextTokens ??
-    qualified ??
-    lookupContextTokens(params.model) ??
-    DEFAULT_CONTEXT_TOKENS
+    params.agentCfg?.contextTokens ?? lookupContextTokens(params.model) ?? DEFAULT_CONTEXT_TOKENS
   );
 }
