@bastani/atomic 0.8.31-alpha.2 → 0.8.31-alpha.3

package/dist/core/agent-session.js

@@ -23,7 +23,7 @@ import { ATOMIC_GUIDE_COMMAND_NAME, ATOMIC_GUIDE_HELP_CHOICES, atomicGuideModeFo
 import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage, formatUnresolvedModelMessage, } from "./auth-guidance.js";
 import { executeBashWithOperations } from "./bash-executor.js";
 import { calculateContextTokens, collectEntriesForBranchSummary, contextCompact as runContextCompact, estimateContextTokens, generateBranchSummary, prepareContextCompaction, shouldCompact, validateContextDeletionRequest, } from "./compaction/index.js";
-import { getModelDefaultContextWindow, getSupportedContextWindows, selectContextWindow } from "./context-window.js";
+import { getEffectiveInputBudget, getModelDefaultContextWindow, getSupportedContextWindows, selectContextWindow } from "./context-window.js";
 import { formatCopilotProviderError, parseCopilotPromptLimitError } from "./copilot-errors.js";
 import { DEFAULT_THINKING_LEVEL } from "./defaults.js";
 import { exportSessionToHtml } from "./export-html/index.js";
@@ -2009,7 +2009,11 @@ export class AgentSession {
         else {
             contextTokens = calculateContextTokens(assistantMessage.usage);
         }
-        if (shouldCompact(contextTokens, contextWindow, settings)) {
+        // Compact against the effective input budget (the hard prompt cap for providers like Copilot
+        // that advertise a larger total window) so we compact before overrunning the server-side limit
+        // rather than relying on reactive overflow recovery near the cap.
+        const compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;
+        if (shouldCompact(contextTokens, compactionBudget, settings)) {
             await this._runAutoCompaction("threshold", false);
         }
     }
@@ -2017,7 +2021,11 @@ export class AgentSession {
         if (!this.model || this.model.provider !== "github-copilot" || !assistantMessage.errorMessage)
             return false;
         const promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);
-        return promptLimitError !== undefined && this.model.contextWindow > promptLimitError.limitTokens;
+        // Compare against the effective input budget (the model's real prompt cap), not the displayed
+        // total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;
+        // only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account
+        // to a lower server tier) keeps the friendly error visible instead of silently compacting down.
+        return promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;
     }
     /**
      * Internal: remove the trailing overflow error from retry context if it is still present.