@pentoshi/clai 4.11.2 → 4.11.3

package/dist/agent/runner.js

@@ -986,6 +986,7 @@ export async function runAgentTurn(prompt, options = {}) {
         // provider's reported prompt size below so the estimator learns this route's
         // bias instead of permanently over-reporting it.
         let dispatchedRawRequestTokens = 0;
+        let dispatchedRequestRoute;
         let consecutiveSynthesizedRounds = 0;
         const engagementPolicy = new EngagementPolicyEngine();
         const probeStateKey = (call) => {
@@ -3011,6 +3012,11 @@ export async function runAgentTurn(prompt, options = {}) {
         }
         async function maybeAutoCompact(reason, force = false) {
             const beforeTokens = estimateNextRequestTokens(messages);
+            emit({
+                type: "context-estimate",
+                estimatedTokens: beforeTokens,
+                model,
+            });
             const contextLimitTokens = currentContextLimitTokens();
             const compactTrigger = autoCompactTriggerTokens(getReliabilityPolicy(), {
                 provider,
@@ -3351,6 +3357,7 @@ export async function runAgentTurn(prompt, options = {}) {
                             ...(contextLimitTokens !== undefined ? { contextLimitTokens } : {}),
                         }).accounting;
                         dispatchedRawRequestTokens = finalAccounting.rawRequestTokens;
+                        dispatchedRequestRoute = { provider, model };
                         // The chip reports the request that is actually about to be sent, from
                         // the same accounting the fit gate and the compaction card use, so the
                         // three can never disagree.
@@ -3665,7 +3672,11 @@ export async function runAgentTurn(prompt, options = {}) {
                 provider = completion.provider;
                 model = completion.model;
                 if (completion.usage) {
-                    if (completion.usage.exact && completion.usage.promptTokens > 0) {
+                    const requestRouteMatched = dispatchedRequestRoute?.provider === completion.provider &&
+                        dispatchedRequestRoute.model === completion.model;
+                    if (completion.usage.exact &&
+                        completion.usage.promptTokens > 0 &&
+                        requestRouteMatched) {
                         recordRequestTokenObservation({
                             provider: completion.provider,
                             model: completion.model,
@@ -3865,7 +3876,30 @@ export async function runAgentTurn(prompt, options = {}) {
                                         // next turn is not orphaned, then nudge for append.
                                         const salvageHistoryCalls = ensureUniqueToolCallIds(nativeToolCalls, toolCallIdsInHistory(messages));
                                         const salvagedCallIndex = nativeToolCalls.indexOf(writeTc);
+                                        const salvagedToolName = salvaged.operation === "append"
+                                            ? "fs.append"
+                                            : "fs.write";
+                                        const salvagedHistoryArgs = {
+                                            path: salvaged.path,
+                                            content: salvaged.content,
+                                            ...(salvaged.operation === "append"
+                                                ? {
+                                                    position: "end",
+                                                    ...(typeof salvaged.expectedPriorBytes === "number"
+                                                        ? { expectedPriorBytes: salvaged.expectedPriorBytes }
+                                                        : {}),
+                                                }
+                                                : {}),
+                                        };
                                         const salvagedCallId = salvageHistoryCalls[salvagedCallIndex]?.id ?? writeTc.id;
+                                        if (salvagedCallIndex >= 0) {
+                                            salvageHistoryCalls[salvagedCallIndex] = {
+                                                id: salvagedCallId,
+                                                name: salvagedToolName,
+                                                args: salvagedHistoryArgs,
+                                                rawArguments: JSON.stringify(salvagedHistoryArgs),
+                                            };
+                                        }
                                         appendAssistantWithTools(messages, assistantText.visible, salvageHistoryCalls, completion.reasoningBlock ??
                                             (assistantText.hasThinking && assistantText.thinkContent
                                                 ? { text: assistantText.thinkContent }
@@ -3876,17 +3910,14 @@ export async function runAgentTurn(prompt, options = {}) {
                                                 : `Tool ${tc.name} result (exit=1, ok=false):\nCancelled — sibling write was truncated and salvaged.`, tc.name, tc.id === salvagedCallId);
                                         }
                                         const priorBytes = writeResult.bytesOnDisk;
-                                        const salvagedToolName = salvaged.operation === "append"
-                                            ? "fs.append"
-                                            : "fs.write";
                                         const appendNudge = toolsAttached
                                             ? `Your ${salvagedToolName} tool call was cut off at the token limit, but the system salvaged the partial content and wrote ${lineCount} lines (file is now ${priorBytes} bytes) to ${salvaged.path}. ` +
                                                 `The file ends with: ${JSON.stringify(salvaged.lastLine)}\n\n` +
-                                                `CONTINUE by calling fs.append now with path=${JSON.stringify(salvaged.path)}, expectedPriorBytes=${priorBytes}, and content set to ONLY the remaining content not already on disk (prefer hundreds of lines per call). ` +
+                                                `CONTINUE by calling fs.append with path=${JSON.stringify(salvaged.path)}, expectedPriorBytes=${priorBytes}, and content set to ONLY the next remaining chunk not already on disk. Keep each chunk under 24,000 characters and wait for its receipt before sending another. ` +
                                                 `Do not re-read the full file; do not re-send content already saved. Use the platform tool interface — no markdown fences.`
                                             : `Your ${salvagedToolName} tool call was cut off at the token limit, but the system salvaged the partial content and wrote ${lineCount} lines (file is now ${priorBytes} bytes) to ${salvaged.path}. ` +
                                                 `The file ends with: ${JSON.stringify(salvaged.lastLine)}\n\n` +
-                                                `CONTINUE with ONE large fs.append of the remaining content:\n` +
+                                                `CONTINUE with one fs.append chunk under 24,000 characters, then wait for its receipt before sending another:\n` +
                                                 '```tool\n{"name":"fs.append","args":{"path":' +
                                                 JSON.stringify(salvaged.path) +
                                                 ',"expectedPriorBytes":' +