@kenkaiiii/ggcoder 5.67.1 → 5.68.1

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

@@ -1,5 +1,5 @@
 import { type AgentTool } from "@kenkaiiii/gg-agent";
-import { type Message, type MessageProvenance, type Provider, type ThinkingLevel } from "@kenkaiiii/gg-ai";
+import { type Message, type MessageProvenance, type Provider, type Usage, type ThinkingLevel } from "@kenkaiiii/gg-ai";
 import { EventBus } from "./event-bus.js";
 import { type CompletionReview } from "./completion-review.js";
 import { SlashCommandRegistry } from "./slash-commands.js";
@@ -224,6 +224,7 @@ export declare class AgentSession {
     private autopilotMarkers;
     private appMarkers;
     private turnMetrics;
+    private readonly cacheDiagnostics;
     /** Internal-only (GG_INTERNAL): live per-session cost/reliability recorder.
      * Absent entirely in public builds — see core/internal-diagnostics.ts. */
     private diagnosticsRecorder?;
@@ -299,6 +300,12 @@ export declare class AgentSession {
     private postTurnCompaction?;
     private lastCompactionCompacted;
     private compactionRetryAfter;
+    /**
+     * SoL-Pi plan-step compaction bookkeeping (see compaction/plan-step-policy.ts).
+     * Step progress comes from `[DONE:n]` markers in assistant text — the same
+     * contract the approved-plan UI tracks.
+     */
+    private planStepState;
     /** A restored oversized checkpoint must be canonicalized before its first prompt is persisted. */
     private deferredCompactionPending;
     /** Latest provider count, anchored to the assistant response it measured. */
@@ -601,8 +608,11 @@ export declare class AgentSession {
     private getHookFollowUpMessages;
     private reviewLspEvidence;
     private withReviewLspEvidence;
-    /** Auto-compact if needed, run agent loop with auth retry, and persist messages. */
+    /** Wraps the real run: cancels an in-flight cache prewarm and tracks run
+     *  activity / last real request time for {@link prewarm}. */
     private runLoop;
+    /** Auto-compact if needed, run agent loop with auth retry, and persist messages. */
+    private runLoopInner;
     switchModel(provider: string, model: string): Promise<void>;
     /**
      * Record the switch as its own trailing message at the point it happened,
@@ -629,6 +639,20 @@ export declare class AgentSession {
      * protect it.
      */
     private settlePostTurnCompaction;
+    /**
+     * Advance plan-step bookkeeping over the messages added since the last
+     * observation and, when a plan step was newly completed (`[DONE:n]`), run
+     * the SoL-Pi cost rule. Called between model steps (in-flight) and once
+     * after the run; `scanIndex` and `doneSteps` make each message and each
+     * step count once across both paths. Returns undefined when no step
+     * completed since the last observation.
+     */
+    private observePlanStepProgress;
+    /**
+     * Plan-step bookkeeping after ANY successful compaction (as in the bench:
+     * every compaction leaves a cache-write cost to be repaid by later savings).
+     */
+    private recordPlanStepCompaction;
     /**
      * Post-turn compaction: once the final response has been delivered, compact
      * in the background while the user reads the answer, instead of making the
@@ -637,7 +661,9 @@ export declare class AgentSession {
      * 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.
+     * error surface in the completed turn. Besides the size trigger, a newly
+     * completed approved-plan step may compact when the cache-cost rule in
+     * compaction/plan-step-policy.ts says the shrink pays for itself.
      */
     private maybeCompactPostTurn;
     compact(existingCredentials?: {
@@ -789,6 +815,11 @@ export declare class AgentSession {
      * the standard prompt therefore needs a rebuild. Custom and sub-agent
      * prompts never render packs, so detection is skipped for them.
      */
+    /**
+     * Bring the system prompt to the exact state the next provider request will
+     * send. Shared by real runs and {@link prewarm} so both see one prefix.
+     */
+    private prepareSystemPromptForRequest;
     private refreshActiveLanguages;
     /** Rebuild messages[0] from current plan-mode + approved-plan state. */
     private rebuildSystemPromptInPlace;
@@ -924,6 +955,22 @@ export declare class AgentSession {
     /** Replace the abort signal (e.g. after cancellation). */
     setSignal(signal: AbortSignal): void;
     private bindManagerCancellation;
+    private runLoopDepth;
+    private lastRealRequestAt;
+    private lastPrewarmAt;
+    private prewarmController;
+    /**
+     * Best-effort Anthropic prompt-cache prewarm before the user's next turn
+     * (desktop app calls this on the first keystroke after opening a chat or an
+     * idle pause). Sends the exact request prefix the next real turn will use —
+     * same system/tools/thinking/cache options — with `max_tokens: 1`, so the
+     * first real reply is a cache read instead of a cold write.
+     */
+    prewarm(signal?: AbortSignal): Promise<{
+        ok: boolean;
+        reason: string;
+        usage?: Usage;
+    }>;
     /** True when speedProfile is "optimized" (1-h cache TTL + pre-warm), or the
      *  session was constructed with `forceLongCacheRetention` (Ken sessions). */
     private isSpeedOptimized;