@kenkaiiii/ggcoder 5.60.2 → 5.60.4

package/dist/core/agent-session.d.ts

@@ -255,6 +255,15 @@ export declare class AgentSession {
      *  only on a real edge. */
     private verificationArmed;
     private compactionOccurred;
+    /**
+     * Re-grounding carry-over for post-turn compaction. `resetHookState` clears
+     * `compactionOccurred` at run start — before the injection point — so a
+     * background compaction that finished after the final response parks its
+     * "re-ground next turn" signal here for pickup at the next run start.
+     */
+    private compactionArmedForNextRun;
+    /** In-flight post-turn (background) compaction, if any. */
+    private postTurnCompaction?;
     private lastCompactionCompacted;
     private compactionRetryAfter;
     /** A restored oversized checkpoint must be canonicalized before its first prompt is persisted. */
@@ -569,6 +578,26 @@ export declare class AgentSession {
     /** Canonicalize a deferred restore before a new prompt can fork stale history. */
     private adoptDeferredCheckpointBeforePrompt;
     private persistCompactionCheckpoint;
+    /**
+     * Wait out any in-flight post-turn background compaction before a new
+     * entry point mutates `this.messages` (prompt, attachments, model switch).
+     * The background compact() snapshots and then REPLACES the array, so a
+     * message pushed while it runs would be silently dropped from the live
+     * history — the pre-run await in runLoop() sits after the push and cannot
+     * protect it.
+     */
+    private settlePostTurnCompaction;
+    /**
+     * Post-turn compaction: once the final response has been delivered, compact
+     * in the background while the user reads the answer, instead of making the
+     * next prompt pay the summarizer latency up front (the pre-run path stays as
+     * the backstop, so a skipped or failed attempt costs nothing). Mirrors the
+     * Codex `model_post_turn_compact_threshold_percent` guards: skip when user
+     * input is already queued (it would race the next turn), when the run was
+     * aborted, or during the failure cooldown — and never let a compaction
+     * error surface in the completed turn.
+     */
+    private maybeCompactPostTurn;
     compact(existingCredentials?: {
         accessToken: string;
         accountId?: string;

package/dist/core/agent-session.js

@@ -37,6 +37,7 @@ import { log } from "./logger.js";
 import { setEstimatorModel, calibrateEstimatorFromUsage } from "./compaction/token-estimator.js";
 import { calculateActiveContextTokens } from "./compaction/active-context.js";
 import { resolveCompactionPolicy } from "./compaction/policy.js";
+import { clampThinkingForPlanMode } from "./thinking-level.js";
 import { pruneStaleToolResults } from "./compaction/tool-result-pruner.js";
 import { discoverAgents } from "./agents.js";
 import { enhancePrompt } from "../utils/prompt-enhancer.js";
@@ -199,6 +200,15 @@ export class AgentSession {
      *  only on a real edge. */
     verificationArmed = false;
     compactionOccurred = false;
+    /**
+     * Re-grounding carry-over for post-turn compaction. `resetHookState` clears
+     * `compactionOccurred` at run start — before the injection point — so a
+     * background compaction that finished after the final response parks its
+     * "re-ground next turn" signal here for pickup at the next run start.
+     */
+    compactionArmedForNextRun = false;
+    /** In-flight post-turn (background) compaction, if any. */
+    postTurnCompaction;
     lastCompactionCompacted = false;
     compactionRetryAfter = 0;
     /** A restored oversized checkpoint must be canonicalized before its first prompt is persisted. */
@@ -895,6 +905,7 @@ export class AgentSession {
         kind: "prompt",
         visibility: "transcript",
     }, options = {}) {
+        await this.settlePostTurnCompaction();
         await this.adoptDeferredCheckpointBeforePrompt();
         const slash = await this.resolveSlashInput(content);
         if (slash?.kind === "template") {
@@ -929,6 +940,7 @@ export class AgentSession {
      * attachments are always a direct conversational turn.
      */
     async promptWithAttachments(text, attachments) {
+        await this.settlePostTurnCompaction();
         await this.adoptDeferredCheckpointBeforePrompt();
         const parts = this.buildAttachmentParts(text, attachments);
         if (parts.length === 0)
@@ -1069,6 +1081,13 @@ export class AgentSession {
                 this.backgroundVerification.delete(id);
         }
         this.compactionOccurred = false;
+        // Post-turn compaction may have landed between runs — adopt its armed
+        // signal so the re-grounding hook fires for this run exactly as it would
+        // after a pre-run compaction.
+        if (this.compactionArmedForNextRun) {
+            this.compactionArmedForNextRun = false;
+            this.compactionOccurred = true;
+        }
         this.originalRequest = originalRequest;
     }
     /**
@@ -1936,6 +1955,11 @@ export class AgentSession {
         // of the public API model window. Failed/no-op attempts cool down across
         // prompts; provider overflow recovery still bypasses this path entirely.
         if (this.settingsManager.get("autoCompact") && Date.now() >= this.compactionRetryAfter) {
+            // A post-turn compaction may still be running in the background. Let it
+            // settle first: it usually already compacted this history, so the
+            // decision below turns into a no-op instead of duplicate work.
+            if (this.postTurnCompaction)
+                await this.postTurnCompaction;
             const contextWindow = getContextWindow(this.model, {
                 provider: this.provider,
                 accountId: creds.accountId,
@@ -2003,7 +2027,13 @@ export class AgentSession {
                 maxTokens: this.maxTokens,
                 maxTurns: this.opts.maxTurns,
                 maxTurnExtensions: this.opts.maxTurnExtensions,
-                thinking: this.thinkingLevel,
+                // Plan mode caps effort at medium (Codex `plan_mode_reasoning_effort`
+                // preset): read-only exploration doesn't need xhigh/max reasoning, and
+                // deep-reasoning models left at the ceiling burn enormous thinking
+                // budgets re-deriving context they cannot act on.
+                thinking: this.planModeRef.current
+                    ? clampThinkingForPlanMode(this.thinkingLevel)
+                    : this.thinkingLevel,
                 apiKey,
                 // Per-turn credential resolution. A run can span many minutes; if any
                 // process sharing auth.json refreshes this grant meanwhile, the token
@@ -2269,8 +2299,15 @@ export class AgentSession {
             await this.persistMessage(this.messages[i]);
         }
         this.lastPersistedIndex = this.messages.length;
+        // Final response is delivered — compact in the background while the user
+        // reads it, instead of charging the summarizer latency to their next
+        // prompt. Detached by design; the pre-run path above remains the backstop.
+        this.maybeCompactPostTurn(creds);
     }
     async switchModel(provider, model) {
+        // The model-switch note is appended to `this.messages`; settle any
+        // background compaction first so the note cannot be swapped out.
+        await this.settlePostTurnCompaction();
         const prevProvider = this.provider;
         const prevModel = this.model;
         // Diff gate: a "switch" to the model already in use is not state change.
@@ -2471,6 +2508,101 @@ export class AgentSession {
             newCount: result.newCount,
         });
     }
+    /**
+     * Wait out any in-flight post-turn background compaction before a new
+     * entry point mutates `this.messages` (prompt, attachments, model switch).
+     * The background compact() snapshots and then REPLACES the array, so a
+     * message pushed while it runs would be silently dropped from the live
+     * history — the pre-run await in runLoop() sits after the push and cannot
+     * protect it.
+     */
+    async settlePostTurnCompaction() {
+        if (this.postTurnCompaction)
+            await this.postTurnCompaction;
+    }
+    /**
+     * Post-turn compaction: once the final response has been delivered, compact
+     * in the background while the user reads the answer, instead of making the
+     * next prompt pay the summarizer latency up front (the pre-run path stays as
+     * the backstop, so a skipped or failed attempt costs nothing). Mirrors the
+     * Codex `model_post_turn_compact_threshold_percent` guards: skip when user
+     * input is already queued (it would race the next turn), when the run was
+     * aborted, or during the failure cooldown — and never let a compaction
+     * error surface in the completed turn.
+     */
+    maybeCompactPostTurn(creds) {
+        if (!this.settingsManager.get("autoCompact"))
+            return;
+        if (this.opts.signal?.aborted)
+            return;
+        if (this.userQueue.length > 0)
+            return;
+        if (this.postTurnCompaction)
+            return;
+        if (Date.now() < this.compactionRetryAfter)
+            return;
+        // One compaction per turn boundary: a pre-run or overflow-recovery
+        // compaction already shrank this run's history — re-probing right after
+        // the final response would only re-derive that decision.
+        if (this.compactionOccurred)
+            return;
+        const contextWindow = getContextWindow(this.model, {
+            provider: this.provider,
+            accountId: creds.accountId,
+        });
+        const policy = resolveCompactionPolicy({
+            provider: this.provider,
+            model: this.model,
+            contextWindow,
+            threshold: this.settingsManager.get("compactThreshold"),
+            accountId: creds.accountId,
+            approvedPlanPath: this.approvedPlanPath,
+        });
+        let activeTokens;
+        if (this.providerContext) {
+            const anchorIndex = this.messages.lastIndexOf(this.providerContext.anchor);
+            if (anchorIndex >= 0) {
+                activeTokens = calculateActiveContextTokens(this.messages, {
+                    usage: this.providerContext.usage,
+                    pendingMessages: this.messages.slice(anchorIndex + 1),
+                });
+            }
+        }
+        if (!shouldCompact(this.messages, contextWindow, policy.threshold, activeTokens))
+            return;
+        log("INFO", "compaction", "Post-turn compaction decision — compacting in background", {
+            provider: this.provider,
+            model: this.model,
+            transport: this.provider === "openai" && creds.accountId ? "codex_oauth" : "public_api",
+            contextWindow: String(contextWindow),
+            activeTokens: activeTokens === undefined ? "estimated" : String(activeTokens),
+            triggerLimit: String(policy.targetTokens),
+        });
+        this.postTurnCompaction = (async () => {
+            try {
+                await this.compact(creds, "automatic");
+                if (this.lastCompactionCompacted) {
+                    // Arm re-grounding for the next turn. resetHookState clears
+                    // compactionOccurred at run start, before the injection point, so
+                    // park the signal for pickup there.
+                    this.compactionArmedForNextRun = true;
+                    this.compactionRetryAfter = 0;
+                }
+                else {
+                    this.compactionRetryAfter = Date.now() + 30_000;
+                }
+            }
+            catch (error) {
+                this.compactionRetryAfter = Date.now() + 30_000;
+                if (isAbortError(error) || this.opts.signal?.aborted)
+                    return;
+                log("WARN", "compaction", `Post-turn compaction failed; cooling down for 30s: ${error instanceof Error ? error.message : String(error)}`);
+            }
+            finally {
+                this.postTurnCompaction = undefined;
+            }
+        })();
+    }
     async compact(existingCredentials, mode = "manual", 
     /** User-stated focus (`/compact <focus>`): what must survive verbatim. */
     focus) {
@@ -3445,6 +3577,12 @@ export class AgentSession {
         return this.getPromptCacheKey();
     }
     async dispose() {
+        // Quiesce any in-flight post-turn compaction BEFORE tearing down state:
+        // the background compact() snapshots and replaces `this.messages`, so
+        // letting it run past this point would checkpoint a near-empty history
+        // and leak a junk session file after teardown.
+        if (this.postTurnCompaction)
+            await this.postTurnCompaction;
         this.diagnosticsRecorder?.finalize();
         this.managerAbortSignal?.removeEventListener("abort", this.managerAbortHandler);
         this.processManager?.shutdownAll();