@leanandmean/coding-agent 0.74.1-scramjet.22 → 0.74.1-scramjet.24

package/dist/core/agent-session.js

@@ -17,7 +17,7 @@ import { basename, dirname, resolve } from "node:path";
 // SCRAMJET-DIVERGENCE: provider-safe harness tool-call id generation for the persisted-settlement
 // acknowledgement key (#341).
 import { generateHarnessToolCallId } from "@leanandmean/agent";
-import { clampThinkingLevel, cleanupSessionResources, flattenSystemPrompt, getSupportedThinkingLevels, isContextOverflow, modelsAreEqual, resetApiProviders, } from "@leanandmean/ai";
+import { clampThinkingLevel, cleanupSessionResources, flattenSystemPrompt, getContextWindowBudget, getSupportedThinkingLevels, isContextOverflow, modelsAreEqual, resetApiProviders, } from "@leanandmean/ai";
 import { theme } from "../modes/interactive/theme/theme.js";
 import { stripFrontmatter } from "../utils/frontmatter.js";
 import { sleep } from "../utils/sleep.js";
@@ -1602,7 +1602,8 @@ export class AgentSession {
         // Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false
         if (skipAbortedCheck && assistantMessage.stopReason === "aborted")
             return;
-        const contextWindow = this.model?.contextWindow ?? 0;
+        // SCRAMJET-DIVERGENCE: operational context checks use the provider budget, not advertised capacity.
+        const contextWindow = this.model ? getContextWindowBudget(this.model) : 0;
         // Skip overflow check if the message came from a different model.
         // This handles the case where user switched from a smaller-context model (e.g. opus)
         // to a larger-context model (e.g. codex) - the overflow error from the old model
@@ -2158,7 +2159,8 @@ export class AgentSession {
         if (message.stopReason !== "error" || !message.errorMessage)
             return false;
         // Context overflow is handled by compaction, not retry
-        const contextWindow = this.model?.contextWindow ?? 0;
+        // SCRAMJET-DIVERGENCE: retry classification uses the provider's operational context budget.
+        const contextWindow = this.model ? getContextWindowBudget(this.model) : 0;
         if (isContextOverflow(message, contextWindow))
             return false;
         const err = message.errorMessage;
@@ -2628,7 +2630,8 @@ export class AgentSession {
         if (!model)
             return undefined;
         const contextWindow = model.contextWindow ?? 0;
-        if (contextWindow <= 0)
+        const contextWindowBudget = getContextWindowBudget(model);
+        if (contextWindow <= 0 || contextWindowBudget <= 0)
             return undefined;
         // After compaction, the last assistant usage reflects pre-compaction context size.
         // We can only trust usage from an assistant that responded after the latest compaction.
@@ -2653,14 +2656,15 @@ export class AgentSession {
                 }
             }
             if (!hasPostCompactionUsage) {
-                return { tokens: null, contextWindow, percent: null };
+                return { tokens: null, contextWindow, contextWindowBudget, percent: null };
             }
         }
         const estimate = estimateContextTokens(this.messages);
-        const percent = (estimate.tokens / contextWindow) * 100;
+        const percent = (estimate.tokens / contextWindowBudget) * 100;
         return {
             tokens: estimate.tokens,
             contextWindow,
+            contextWindowBudget,
             percent,
         };
     }