@bastani/atomic 0.9.8 → 0.9.9-alpha.1

package/dist/builtin/workflows/src/runs/shared/worktree-git.ts

@@ -3,6 +3,7 @@ import * as fs from "node:fs";
 import * as path from "node:path";
 import { createGitEnvironment } from "@bastani/atomic";
 import type { GitResult, GitWorktreeSetupOptions, GitWorktreeSetupResult } from "./worktree-types.js";
+import { openGitWorktreeGenerationAnchor, type GitWorktreeGenerationAnchor } from "./worktree-generation.js";
 
 const DISABLED_GIT_HOOKS_PATH = process.platform === "win32" ? "NUL" : "/dev/null";
 const GIT_COMMAND_TIMEOUT_MS = 60_000;
@@ -133,26 +134,53 @@ function runGitReadOnlyProbe(cwd: string, args: readonly string[]): GitResult {
 
 export interface GitWorktreeSetupCache {
 	get(options: GitWorktreeSetupOptions): GitWorktreeSetupResult;
+	dispose(): void;
 }
 
 function gitWorktreeSetupCacheKey(options: GitWorktreeSetupOptions): string {
+	const repoRoot = repositoryRootForGitWorktree(options.cwd);
+	const { logicalRepoRoot } = cwdWithinGitRepository(options.cwd, repoRoot);
+	const worktreeRoot = resolveGitWorktreePath(options.gitWorktreeDir, logicalRepoRoot);
 	return JSON.stringify([
-		path.resolve(options.cwd),
-		options.gitWorktreeDir.trim(),
-		options.baseBranch?.trim() ?? "",
+		comparableRealPath(gitCommonDirForWorktree(repoRoot)),
+		comparablePathThroughExistingAncestor(worktreeRoot),
 	]);
 }
 
+interface CachedGitWorktreeSetup {
+	readonly result: GitWorktreeSetupResult;
+	readonly identity: GitWorktreeIdentity;
+}
+
 export function createGitWorktreeSetupCache(): GitWorktreeSetupCache {
-	const results = new Map<string, GitWorktreeSetupResult>();
+	const results = new Map<string, CachedGitWorktreeSetup>();
+	let disposed = false;
 	return {
 		get(options) {
+			if (disposed) throw new Error("Git worktree setup cache is already disposed");
 			const key = gitWorktreeSetupCacheKey(options);
 			const cached = results.get(key);
-			if (cached !== undefined) return cached;
-			const setup = setupGitWorktree(options);
-			results.set(key, setup);
-			return setup;
+			if (cached !== undefined) {
+				assertCachedGitWorktreeIdentity(cached, options);
+				return cached.result;
+			}
+			const result = setupGitWorktree(options);
+			results.set(key, { result, identity: captureGitWorktreeIdentity(result) });
+			return result;
+		},
+		dispose() {
+			if (disposed) return;
+			disposed = true;
+			let firstError: Error | undefined;
+			for (const cached of results.values()) {
+				try {
+					cached.identity.generation.dispose();
+				} catch (error) {
+					firstError ??= error instanceof Error ? error : new Error(String(error));
+				}
+			}
+			results.clear();
+			if (firstError !== undefined) throw firstError;
 		},
 	};
 }
@@ -223,6 +251,31 @@ function comparableRealPath(value: string): string {
 	return process.platform === "win32" ? realpath.toLowerCase() : realpath;
 }
 
+function comparablePathThroughExistingAncestor(value: string): string {
+	let existing = path.resolve(value);
+	const missing: string[] = [];
+	while (!pathExistsSync(existing)) {
+		const parent = path.dirname(existing);
+		if (parent === existing) break;
+		missing.unshift(path.basename(existing));
+		existing = parent;
+	}
+	const resolved = path.resolve(comparableRealPath(existing), ...missing).replace(/\\/g, "/");
+	return process.platform === "win32" ? resolved.toLowerCase() : resolved;
+}
+
+function validateWorktreeOutsideInvokingCheckout(worktreeRoot: string, repoRoot: string): void {
+	const repository = comparableRealPath(repoRoot);
+	const candidate = comparablePathThroughExistingAncestor(worktreeRoot);
+	const relativePath = path.relative(repository, candidate);
+	if (relativePath === "") {
+		throw new Error("gitWorktreeDir must not resolve to the invoking checkout; provide a separate same-repository worktree path.");
+	}
+	if (!relativePath.startsWith(`..${path.sep}`) && relativePath !== ".." && !path.isAbsolute(relativePath)) {
+		throw new Error("gitWorktreeDir must be outside the invoking checkout; provide a separate same-repository worktree path.");
+	}
+}
+
 function gitPathFromOutput(value: string, cwd: string): string | undefined {
 	const trimmed = value.trim();
 	if (!trimmed) return undefined;
@@ -318,10 +371,74 @@ function validateExistingGitWorktreeRoot(worktreeRoot: string, repoRoot: string)
 	}
 }
 
+interface GitWorktreeIdentitySnapshot {
+	readonly worktreeRealPath: string;
+	readonly repositoryRealPath: string;
+	readonly gitDirRealPath: string;
+	readonly commonDirRealPath: string;
+	readonly device: number;
+	readonly inode: number;
+}
+
+interface GitWorktreeIdentity extends GitWorktreeIdentitySnapshot {
+	readonly generation: GitWorktreeGenerationAnchor;
+}
+
+function gitDirectoryForWorktree(cwd: string): string {
+	const result = runGitReadOnlyProbe(cwd, ["rev-parse", "--absolute-git-dir"]);
+	if (result.status !== 0) {
+		throw new Error(`Failed to validate Git directory for gitWorktreeDir ${cwd}. Git reported: ${gitFailureMessage(result)}`);
+	}
+	const gitDir = gitPathFromOutput(result.stdout, cwd);
+	if (gitDir === undefined) throw new Error("git rev-parse --absolute-git-dir returned an empty path");
+	return gitDir;
+}
+function captureGitWorktreeIdentitySnapshot(setup: GitWorktreeSetupResult): GitWorktreeIdentitySnapshot {
+	validateWorktreeOutsideInvokingCheckout(setup.worktreeRoot, setup.repositoryRoot);
+	validateExistingGitWorktreeRoot(setup.worktreeRoot, setup.repositoryRoot);
+	const stats = fs.statSync(setup.worktreeRoot);
+	return {
+		worktreeRealPath: comparableRealPath(setup.worktreeRoot),
+		repositoryRealPath: comparableRealPath(setup.repositoryRoot),
+		gitDirRealPath: comparableRealPath(gitDirectoryForWorktree(setup.worktreeRoot)),
+		commonDirRealPath: comparableRealPath(gitCommonDirForWorktree(setup.worktreeRoot)),
+		device: stats.dev,
+		inode: stats.ino,
+	};
+}
+
+function captureGitWorktreeIdentity(setup: GitWorktreeSetupResult): GitWorktreeIdentity {
+	const snapshot = captureGitWorktreeIdentitySnapshot(setup);
+	return { ...snapshot, generation: openGitWorktreeGenerationAnchor(setup.worktreeRoot) };
+}
+
+function assertCachedGitWorktreeIdentity(cached: CachedGitWorktreeSetup, options: GitWorktreeSetupOptions): void {
+	let current: GitWorktreeIdentitySnapshot;
+	try {
+		current = captureGitWorktreeIdentitySnapshot(cached.result);
+		cached.identity.generation.assertCurrent();
+	} catch (error) {
+		const message = error instanceof Error ? error.message : String(error);
+		throw new Error(`Cached gitWorktreeDir changed before reuse: ${options.gitWorktreeDir}. Reason: ${message}`);
+	}
+	const expected = cached.identity;
+	if (
+		current.worktreeRealPath !== expected.worktreeRealPath ||
+		current.repositoryRealPath !== expected.repositoryRealPath ||
+		current.gitDirRealPath !== expected.gitDirRealPath ||
+		current.commonDirRealPath !== expected.commonDirRealPath ||
+		current.device !== expected.device ||
+		current.inode !== expected.inode
+	) {
+		throw new Error(`Cached gitWorktreeDir changed before reuse: ${options.gitWorktreeDir}. Start a new run with the intended same-repository worktree.`);
+	}
+}
+
 export function setupGitWorktree(options: GitWorktreeSetupOptions): GitWorktreeSetupResult {
 	const repoRoot = repositoryRootForGitWorktree(options.cwd);
 	const { relativeCwd, logicalRepoRoot } = cwdWithinGitRepository(options.cwd, repoRoot);
 	const worktreeRoot = resolveGitWorktreePath(options.gitWorktreeDir, logicalRepoRoot);
+	validateWorktreeOutsideInvokingCheckout(worktreeRoot, repoRoot);
 	if (pathExistsSync(worktreeRoot)) {
 		validateExistingGitWorktreeRoot(worktreeRoot, repoRoot);
 		return {
@@ -346,6 +463,15 @@ export function setupGitWorktree(options: GitWorktreeSetupOptions): GitWorktreeS
 			`If another process just created this same-repository worktree, rerun the workflow to resume it. If this is an orphaned worktree from an interrupted run, recover or remove it with: ${worktreeRecoveryCommand(repoRoot, worktreeRoot)}`,
 		].join("\n"));
 	}
+	try {
+		validateWorktreeOutsideInvokingCheckout(worktreeRoot, repoRoot);
+		validateExistingGitWorktreeRoot(worktreeRoot, repoRoot);
+	} catch (error) {
+		// Best-effort rollback of a target that changed while Git was creating it.
+		runGit(repoRoot, ["worktree", "remove", "--force", worktreeRoot]);
+		const message = error instanceof Error ? error.message : String(error);
+		throw new Error(`Created gitWorktreeDir failed post-creation validation: ${message}`);
+	}
 	return {
 		worktreeRoot,
 		cwd: workspaceCwdForGitWorktreeRoot(worktreeRoot, relativeCwd),

package/dist/builtin/workflows/src/runs/shared/worktree.ts

@@ -1,4 +1,6 @@
 export { createGitWorktreeSetupCache, setupGitWorktree, setupGitWorktreeCached } from "./worktree-git.js";
+export { createGitWorktreeSetupCacheOwner } from "./worktree-cache-lifecycle.js";
+export type { GitWorktreeSetupCacheOwner } from "./worktree-cache-lifecycle.js";
 export {
 	cleanupWorktrees,
 	createWorktrees,

package/dist/builtin/workflows/src/shared/resumable-workflow-notices.ts

@@ -0,0 +1,43 @@
+import type { WorkflowSerializableValue } from "./types.js";
+import { isDurableWorkflowResumable } from "../durable/resume-eligibility.js";
+import type { ResumableWorkflowEntry } from "../durable/types.js";
+
+export interface ResumableWorkflowNotice {
+  readonly workflowId: string;
+  readonly name: string;
+}
+
+function recordOrUndefined(value: unknown): Record<string, WorkflowSerializableValue> | undefined {
+  return typeof value === "object" && value !== null
+    ? value as Record<string, WorkflowSerializableValue>
+    : undefined;
+}
+
+/** Return authoritative resumable workflow references cached in a session. */
+export function findResumableWorkflowNotices(
+  entries: readonly unknown[],
+  authoritativeCatalog: readonly ResumableWorkflowEntry[],
+): ResumableWorkflowNotice[] {
+  const authoritativeById = new Map(authoritativeCatalog.map((entry) => [entry.workflowId, entry]));
+  const sessionWorkflowIds = new Set<string>();
+  for (const value of entries) {
+    const entry = recordOrUndefined(value);
+    if (entry === undefined) continue;
+    const payload = entry["type"] === "custom" && entry["customType"] === "workflow.durable.checkpoint"
+      ? recordOrUndefined(entry["data"])
+      : entry["type"] === "workflow.durable.checkpoint"
+        ? recordOrUndefined(entry["payload"]) ?? entry
+        : undefined;
+    const workflowId = payload?.["workflowId"];
+    if (typeof workflowId === "string") sessionWorkflowIds.add(workflowId);
+  }
+
+  const notices: ResumableWorkflowNotice[] = [];
+  for (const workflowId of sessionWorkflowIds) {
+    const authoritative = authoritativeById.get(workflowId);
+    if (authoritative !== undefined && isDurableWorkflowResumable(authoritative)) {
+      notices.push({ workflowId, name: authoritative.name });
+    }
+  }
+  return notices;
+}

package/dist/builtin/workflows/src/shared/timing.ts

@@ -18,6 +18,10 @@ function nonNegative(ms: number): number {
   return Math.max(0, ms);
 }
 
+export function rebasedStageStartedAt(accumulatedDurationMs: number | undefined, resumedAt: number): number {
+  return resumedAt - nonNegative(accumulatedDurationMs ?? 0);
+}
+
 function elapsedFromStart(
   startedAt: number,
   now: number,

package/dist/builtin/workflows/src/shared/types.ts

@@ -6,7 +6,7 @@
 import type {
   AgentSession,
   AgentSessionEvent,
-  ContextCompactionResult,
+  VerbatimCompactionResult,
   CreateAgentSessionOptions,
   ModelCycleResult,
   PromptOptions,
@@ -19,7 +19,7 @@ import type * as AuthoringContract from "./authoring-contract.js";
 
 export type { TSchema };
 
-export type { AgentSessionEvent, ContextCompactionResult, ModelCycleResult, PromptOptions };
+export type { AgentSessionEvent, VerbatimCompactionResult, ModelCycleResult, PromptOptions };
 
 export type StageUserMessageContent = Parameters<AgentSession["sendUserMessage"]>[0];
 
@@ -191,6 +191,8 @@ export interface StageOptions<TSchemaDef extends TSchema | undefined = TSchema |
   resumeFromSessionFile?: string;
   /** Internal durable replay key used to map a live LM session to durable resume state. */
   durableReplayKey?: string;
+  /** Internal durable timing baseline accumulated before a process-boundary resume. */
+  durableAccumulatedDurationMs?: number;
 }
 
 // ---------------------------------------------------------------------------
@@ -352,7 +354,7 @@ export interface StageContext<TSchemaDef extends TSchema | undefined = undefined
   ): Promise<{ editorText?: string; cancelled: boolean }>;
 
   /** Compaction. */
-  compact(): Promise<ContextCompactionResult>;
+  compact(): Promise<VerbatimCompactionResult>;
   abortCompaction(): void;
 
   /** Abort current operation. */

package/dist/builtin/workflows/src/tui/graph-view-constants.ts

@@ -42,4 +42,5 @@ export const ANIMATION_TICK_MS = 100;
  * eased lerp inside `pickBorder` traces one full breath per cycle.
  */
 export const PULSE_PERIOD_MS = 2000;
+export const GRAPH_SCROLL_STEP_COLS = 4;
 export const GRAPH_SCROLL_STEP_ROWS = 4;

package/dist/builtin/workflows/src/tui/graph-view-input.ts

@@ -1,5 +1,6 @@
 import { expandedStageTarget } from "../shared/expanded-workflow-graph.js";
 import {
+  GRAPH_SCROLL_STEP_COLS,
   GRAPH_SCROLL_STEP_ROWS,
 } from "./graph-view-constants.js";
 import { GraphViewRenderer } from "./graph-view-render.js";
@@ -18,6 +19,11 @@ interface SgrMouseEvent {
   final: "M" | "m";
 }
 
+interface MouseWheelDelta {
+  cols: number;
+  rows: number;
+}
+
 /** Keyboard, mouse, switcher, prompt, and focus navigation handling. */
 export abstract class GraphViewInputController extends GraphViewRenderer {
   /** Returns true if consumed. */
@@ -46,10 +52,7 @@ export abstract class GraphViewInputController extends GraphViewRenderer {
   }
 
   private _isNonTextGraphControlBeforePrompt(data: string): boolean {
-    return (
-      this._mouseWheelDeltaRows(data) !== 0 ||
-      matchesKey(data, Key.ctrl("d"))
-    );
+    return this._mouseWheelDelta(data) !== null || matchesKey(data, Key.ctrl("d"));
   }
 
   private _handlePromptInput(data: string): boolean {
@@ -81,9 +84,10 @@ export abstract class GraphViewInputController extends GraphViewRenderer {
 
   private _handleGraphInput(data: string): boolean {
     const stageCount = this.cachedLayout.length;
-    const wheelDeltaRows = this._mouseWheelDeltaRows(data);
-    if (wheelDeltaRows !== 0) {
-      this._scrollGraphBy(wheelDeltaRows);
+    const wheelDelta = this._mouseWheelDelta(data);
+    if (wheelDelta) {
+      if (wheelDelta.rows !== 0) this._scrollGraphBy(wheelDelta.rows);
+      if (wheelDelta.cols !== 0) this._scrollGraphHorizontallyBy(wheelDelta.cols);
       return true;
     }
 
@@ -308,6 +312,11 @@ export abstract class GraphViewInputController extends GraphViewRenderer {
     this.graphScrollOffset = Math.max(0, this.graphScrollOffset + deltaRows);
   }
 
+  private _scrollGraphHorizontallyBy(deltaCols: number): void {
+    this.pendingEnsureFocusedVisible = false;
+    this.graphScrollColOffset = Math.max(0, this.graphScrollColOffset + deltaCols);
+  }
+
   private _graphNodeIndexForClick(data: string): number | null | undefined {
     const click = this._sgrLeftMousePress(data);
     if (!click) return undefined;
@@ -353,7 +362,7 @@ export abstract class GraphViewInputController extends GraphViewRenderer {
     return { col: sgr.col, row: sgr.row };
   }
 
-  private _mouseWheelDeltaRows(data: string): number {
+  private _mouseWheelDelta(data: string): MouseWheelDelta | null {
     const sgr = this._parseSgrMouse(data);
     if (sgr && sgr.final === "M") {
       return this._wheelDeltaForButtonCode(sgr.buttonCode);
@@ -361,15 +370,16 @@ export abstract class GraphViewInputController extends GraphViewRenderer {
     if (data.startsWith("\x1b[M") && data.length >= 6) {
       return this._wheelDeltaForButtonCode(data.charCodeAt(3) - 32);
     }
-    return 0;
+    return null;
   }
 
-  private _wheelDeltaForButtonCode(code: number): number {
-    if ((code & 64) === 0) return 0;
+  private _wheelDeltaForButtonCode(code: number): MouseWheelDelta | null {
+    if ((code & 64) === 0) return null;
     const direction = code & 3;
-    if (direction === 0) return -GRAPH_SCROLL_STEP_ROWS;
-    if (direction === 1) return GRAPH_SCROLL_STEP_ROWS;
-    return 0;
+    if (direction === 0) return { cols: 0, rows: -GRAPH_SCROLL_STEP_ROWS };
+    if (direction === 1) return { cols: 0, rows: GRAPH_SCROLL_STEP_ROWS };
+    if (direction === 2) return { cols: -GRAPH_SCROLL_STEP_COLS, rows: 0 };
+    return { cols: GRAPH_SCROLL_STEP_COLS, rows: 0 };
   }
 
   // ---- test seams ----

package/dist/builtin/workflows/src/tui/workflow-resume-selector.ts

@@ -10,6 +10,7 @@ import type { RunSnapshot, StageSnapshot } from "../shared/store-types.js";
 export type WorkflowResumeSelectorResult =
   | { kind: "live"; runId: string }
   | { kind: "durable"; workflowId: string }
+  | { kind: "completed"; workflowId: string }
   | { kind: "close" };
 
 export interface WorkflowResumeSelectorUiSurface {
@@ -54,47 +55,67 @@ function liveRunSession(run: RunSnapshot): WorkflowResumeSelectorItem {
   };
 }
 
-function durableWorkflowSession(entry: ResumableWorkflowEntry): WorkflowResumeSelectorItem {
+function durableWorkflowSession(
+  entry: ResumableWorkflowEntry,
+  kind: "durable" | "completed",
+): WorkflowResumeSelectorItem {
   const checkpointText = `${entry.completedCheckpoints} checkpoints`;
   const promptText = `${entry.pendingPrompts} prompts`;
-  const firstMessage = `${entry.name}  ${entry.status}  ${checkpointText}  ${promptText}`;
+  const statusText = kind === "completed" ? "✓ completed" : entry.status;
   return {
-    result: { kind: "durable", workflowId: entry.workflowId },
+    result: { kind, workflowId: entry.workflowId },
     session: {
-      path: `workflow-durable:${entry.workflowId}`,
+      path: `workflow-${kind}:${entry.workflowId}`,
       id: entry.workflowId,
-      cwd: "Durable workflow runs",
+      cwd: kind === "completed" ? "Completed workflow runs" : "Durable workflow runs",
       created: new Date(entry.createdAt),
       modified: new Date(entry.updatedAt),
       messageCount: entry.completedCheckpoints,
-      firstMessage,
-      allMessagesText: `${entry.workflowId} ${entry.name} ${entry.status} ${checkpointText} ${promptText}`,
+      firstMessage: `${entry.name}  ${statusText}  ${checkpointText}  ${promptText}`,
+      allMessagesText: `${entry.workflowId} ${entry.name} ${statusText} ${checkpointText} ${promptText}`,
+      ...(kind === "completed" ? { messageColor: "success" as const } : {}),
     },
   };
 }
 
+function compareResumeItemsByRecency(
+  left: WorkflowResumeSelectorItem,
+  right: WorkflowResumeSelectorItem,
+): number {
+  const recencyDifference = right.session.modified.getTime() - left.session.modified.getTime();
+  if (recencyDifference !== 0) return recencyDifference;
+  const idDifference = left.session.id.localeCompare(right.session.id);
+  return idDifference !== 0 ? idDifference : left.session.path.localeCompare(right.session.path);
+}
+
 export function workflowResumeSelectorItems(
   liveRuns: readonly RunSnapshot[],
   durableEntries: readonly ResumableWorkflowEntry[],
+  completedEntries: readonly ResumableWorkflowEntry[] = [],
 ): WorkflowResumeSelectorItem[] {
   const liveIds = new Set(liveRuns.map((run) => run.id));
+  const durableIds = new Set(durableEntries.map((entry) => entry.workflowId));
   return [
     ...liveRuns.map(liveRunSession),
     ...durableEntries
       .filter((entry) => !liveIds.has(entry.workflowId))
-      .map(durableWorkflowSession),
-  ];
+      .map((entry) => durableWorkflowSession(entry, "durable")),
+    ...completedEntries
+      .filter((entry) => !liveIds.has(entry.workflowId) && !durableIds.has(entry.workflowId))
+      .map((entry) => durableWorkflowSession(entry, "completed")),
+  ].sort(compareResumeItemsByRecency);
 }
 
 export function openWorkflowResumeSelector(
   ui: WorkflowResumeSelectorUiSurface,
   liveRuns: readonly RunSnapshot[],
   durableEntries: readonly ResumableWorkflowEntry[],
+  completedEntries: readonly ResumableWorkflowEntry[] = [],
 ): Promise<WorkflowResumeSelectorResult> {
   const custom = ui.custom;
   if (typeof custom !== "function") return Promise.resolve({ kind: "close" });
 
-  const items = workflowResumeSelectorItems(liveRuns, durableEntries);
+  const items = workflowResumeSelectorItems(liveRuns, durableEntries, completedEntries);
 
   const resultByPath = new Map(items.map((item) => [item.session.path, item.result]));
   const sessions = items.map((item) => item.session);
@@ -108,8 +129,11 @@ export function openWorkflowResumeSelector(
     const settle = (result: WorkflowResumeSelectorResult, done?: (result: undefined) => void): void => {
       if (settled) return;
       settled = true;
-      resolve(result);
-      done?.(undefined);
+      try {
+        done?.(undefined);
+      } finally {
+        resolve(result);
+      }
     };
 
     const factory = (
@@ -146,6 +170,12 @@ export function openWorkflowResumeSelector(
       };
     };
 
-    void custom(factory, { overlay: false });
+    try {
+      void Promise.resolve(custom(factory, { overlay: false })).catch(() => {
+        settle({ kind: "close" });
+      });
+    } catch {
+      settle({ kind: "close" });
+    }
   });
 }

package/dist/core/agent-session-auto-compaction.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent-session-auto-compaction.d.ts","sourceRoot":"","sources":["../../src/core/agent-session-auto-compaction.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAOrE,OAAO,KAAK,EAAE,2BAA2B,IAAI,YAAY,EAAE,MAAM,4BAA4B,CAAC;AAE9F;;;;;;GAMG;AACH,eAAO,MAAM,gCAAgC,IAAI,CAAC;AAElD,eAAO,MAAM,6CAA6C,IAAI,CAAC;AAc/D,wBAAsB,gBAAgB,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,gBAAgB,UAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CA8HrI;AAGD,wBAAgB,6CAA6C,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ7H;AAED,wBAAgB,uBAAuB,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAEnF;AAmDD,wBAAgB,8BAA8B,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ1F;AAED,wBAAgB,mCAAmC,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAE/F;AAED;;GAEG;AAEH,wBAAgB,kDAAkD,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQ3F;AAED;;GAEG;AAEH,wBAAgB,4CAA4C,CAAC,IAAI,EAAE,YAAY,EAC9E,OAAO,EAAE,UAAU,GAAG,WAAW,EACjC,SAAS,EAAE,OAAO,GAChB,IAAI,CAoCN;AAED,wBAAsB,uCAAuC,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAI/F;AAED;;GAEG;AAEH,wBAAsB,0BAA0B,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAWlF;AAED;;;;;;GAMG;AAEH,wBAAgB,4BAA4B,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQrE;AAOD,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,UAAU,GAAG,WAAW,EAAE,SAAS,EAAE,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAyEhI;AAED;;GAEG;AAEH,eAAO,MAAM,iCAAiC;;;;;;;;;CAS7C,CAAC","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai/compat\";\nimport { isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { getEffectiveInputBudget } from \"./context-window.ts\";\nimport { parseCopilotPromptLimitError } from \"./copilot-errors.ts\";\nimport { calculateContextTokens, estimateContextTokens, shouldCompact } from \"./compaction/index.ts\";\nimport { getLatestCompactionBoundaryEntry } from \"./session-manager.ts\";\nimport { MIN_RESPONSES_MAX_OUTPUT_TOKENS } from \"./openai-responses-payload-sanitizer.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\n\n/**\n * Upper bound on consecutive automatic continuations of a response that was\n * truncated at the output-token cap (\"length\") while the context is still\n * below the compaction budget. Each continuation regenerates the cut-off turn,\n * so a model that insists on emitting more than its per-turn output cap can\n * still terminate instead of looping forever.\n */\nexport const MAX_LENGTH_CONTINUATION_ATTEMPTS = 3;\n\nexport const MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS = 1;\n\ntype JsonValue = string | number | boolean | null | JsonValue[] | { [key: string]: JsonValue };\ntype ProviderErrorDetails = {\n\tmessage?: string;\n\tcode?: string;\n\tparam?: string;\n};\n\nconst OUTPUT_BUDGET_PARAMETER_PATTERN = /\\bmax_output_tokens\\b/;\nconst OUTPUT_BUDGET_UNDERFLOW_PATTERN = new RegExp(\n\t`(?:integer\\\\s+below\\\\s+minimum\\\\s+value|expected\\\\s+(?:a\\\\s+)?value\\\\s*>=\\\\s*${MIN_RESPONSES_MAX_OUTPUT_TOKENS}|got\\\\s+1\\\\s+instead)`,\n);\n\nexport async function _checkCompaction(this: AgentSession, assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {\n\t// The agent_end path passes skipAbortedCheck=true; the pre-prompt path passes\n\t// false. Only the live turn-completion path may auto-continue a truncated\n\t// response — before a fresh user prompt we must not resume the old turn.\n\tconst isLiveTurnCompletion = skipAbortedCheck;\n\tconst settings = this.settingsManager.getCompactionSettings();\n\tif (!settings.enabled) return;\n\n\t// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false\n\tif (skipAbortedCheck && assistantMessage.stopReason === \"aborted\") return;\n\n\tconst contextWindow = this.model?.contextWindow ?? 0;\n\n\t// Skip overflow check if the message came from a different model.\n\t// This handles the case where user switched from a smaller-context model (e.g. opus)\n\t// to a larger-context model (e.g. codex) - the overflow error from the old model\n\t// shouldn't trigger compaction for the new model.\n\tconst sameModel =\n\t\tthis.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;\n\n\t// Skip compaction checks if this assistant message is older than the latest\n\t// compaction boundary. This prevents a stale pre-compaction usage/error\n\t// from retriggering compaction on the first prompt after compaction.\n\tconst compactionBoundaryEntry = getLatestCompactionBoundaryEntry(this.sessionManager.getBranch());\n\tconst assistantIsFromBeforeCompactionBoundary =\n\t\tcompactionBoundaryEntry !== null &&\n\t\tassistantMessage.timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime();\n\tif (assistantIsFromBeforeCompactionBoundary) {\n\t\treturn;\n\t}\n\n\t// Case 1: Overflow - LLM returned context overflow error\n\t// When Copilot rejects a 1m client-budget prompt at a lower server cap (for example\n\t// because long-context/usage-based billing entitlement is missing), leave the friendly\n\t// error visible instead of auto-compacting down to a smaller server tier silently.\n\tif (sameModel && this._isCopilotServerCapBelowSelectedContextWindow(assistantMessage)) {\n\t\treturn;\n\t}\n\tif (sameModel && isContextOverflow(assistantMessage, contextWindow)) {\n\t\tconst willRetry = assistantMessage.stopReason !== \"stop\";\n\t\tif (!willRetry) {\n\t\t\tawait this._runAutoCompaction(\"overflow\", false);\n\t\t\treturn;\n\t\t}\n\n\t\tif (this._overflowRecoveryAttempted) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason: \"overflow\",\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: true,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tthis._overflowRecoveryAttempted = true;\n\t\t// Remove the error message from agent state (it IS saved to session for history,\n\t\t// but we don't want it in context for the retry)\n\t\tconst messages = this.agent.state.messages;\n\t\tif (messages.length > 0 && messages[messages.length - 1].role === \"assistant\") {\n\t\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t\t}\n\t\tawait this._runAutoCompaction(\"overflow\", willRetry);\n\t\treturn;\n\t}\n\n\t// Case 2: Threshold - context is getting large\n\t// For error messages (no usage data), estimate from last successful response.\n\t// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.\n\tlet contextTokens: number;\n\tif (assistantMessage.stopReason === \"error\") {\n\t\tconst messages = this.agent.state.messages;\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) return; // No usage data at all\n\t\t// Verify the usage source is post-compaction. Kept pre-compaction messages\n\t\t// have stale usage reflecting the old (larger) context and would falsely\n\t\t// trigger compaction right after one just finished.\n\t\tconst usageMsg = messages[estimate.lastUsageIndex];\n\t\tif (\n\t\t\tcompactionBoundaryEntry &&\n\t\t\tusageMsg.role === \"assistant\" &&\n\t\t\t(usageMsg as AssistantMessage).timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime()\n\t\t) {\n\t\t\treturn;\n\t\t}\n\t\tcontextTokens = estimate.tokens;\n\t} else {\n\t\tcontextTokens = calculateContextTokens(assistantMessage.usage);\n\t}\n\t// Compact against the effective input budget (the hard prompt cap for providers like Copilot\n\t// that advertise a larger total window) so we compact before overrunning the server-side limit\n\t// rather than relying on reactive overflow recovery near the cap.\n\tconst compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;\n\tif (shouldCompact(contextTokens, compactionBudget, settings)) {\n\t\tconst willRetry = shouldRetryAfterThresholdCompaction(assistantMessage);\n\t\tif (willRetry && isRetryWorthyOutputBudgetError(assistantMessage)) {\n\t\t\tif (this._outputBudgetErrorContinuationAttempts >= MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS) {\n\t\t\t\tthis._emit({\n\t\t\t\t\ttype: \"compaction_end\",\n\t\t\t\t\treason: \"threshold\",\n\t\t\t\t\tresult: undefined,\n\t\t\t\t\taborted: false,\n\t\t\t\t\twillRetry: false,\n\t\t\t\t\terrorMessage:\n\t\t\t\t\t\t\"Output-budget recovery stopped after a compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t\t});\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tthis._outputBudgetErrorContinuationAttempts += 1;\n\t\t}\n\t\tawait this._runAutoCompaction(\"threshold\", willRetry);\n\t\treturn;\n\t}\n\n\t// A response truncated at the output-token cap (\"length\") with the context\n\t// still below the compaction budget is genuine work cut off mid-flight, not a\n\t// context overflow. Compaction would not free any room, so continue the\n\t// generation directly instead of dead-ending on the truncation and leaving\n\t// the task half-finished.\n\tif (isLiveTurnCompletion && isRetryWorthyLengthStop(assistantMessage)) {\n\t\tthis._resumeAfterLengthTruncation();\n\t}\n}\n\n\nexport function _isCopilotServerCapBelowSelectedContextWindow(this: AgentSession, assistantMessage: AssistantMessage): boolean {\n\tif (!this.model || this.model.provider !== \"github-copilot\" || !assistantMessage.errorMessage) return false;\n\tconst promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);\n\t// Compare against the effective input budget (the model's real prompt cap), not the displayed\n\t// total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;\n\t// only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account\n\t// to a lower server tier) keeps the friendly error visible instead of silently compacting down.\n\treturn promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;\n}\n\nexport function isRetryWorthyLengthStop(assistantMessage: AssistantMessage): boolean {\n\treturn assistantMessage.stopReason === \"length\" && assistantMessage.usage.output > 0;\n}\n\nfunction isJsonRecord(value: JsonValue): value is { [key: string]: JsonValue } {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction stringField(record: { [key: string]: JsonValue }, key: string): string | undefined {\n\tconst value = record[key];\n\treturn typeof value === \"string\" ? value : undefined;\n}\n\nfunction parseProviderErrorDetails(errorMessage: string): ProviderErrorDetails | undefined {\n\tconst start = errorMessage.indexOf(\"{\");\n\tconst end = errorMessage.lastIndexOf(\"}\");\n\tif (start === -1 || end <= start) return undefined;\n\n\ttry {\n\t\tconst parsed = JSON.parse(errorMessage.slice(start, end + 1)) as JsonValue;\n\t\tif (!isJsonRecord(parsed)) return undefined;\n\n\t\tconst nestedError = parsed.error;\n\t\tif (nestedError !== undefined && isJsonRecord(nestedError)) {\n\t\t\treturn {\n\t\t\t\tmessage: stringField(nestedError, \"message\"),\n\t\t\t\tcode: stringField(nestedError, \"code\") ?? stringField(parsed, \"code\"),\n\t\t\t\tparam: stringField(nestedError, \"param\") ?? stringField(parsed, \"param\"),\n\t\t\t};\n\t\t}\n\n\t\treturn {\n\t\t\tmessage: stringField(parsed, \"message\"),\n\t\t\tcode: stringField(parsed, \"code\"),\n\t\t\tparam: stringField(parsed, \"param\"),\n\t\t};\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction isOutputBudgetUnderflowText(text: string): boolean {\n\tconst message = text.toLowerCase();\n\treturn OUTPUT_BUDGET_PARAMETER_PATTERN.test(message) && OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message);\n}\n\nfunction isStructuredOutputBudgetUnderflow(details: ProviderErrorDetails): boolean {\n\tconst message = details.message?.toLowerCase() ?? \"\";\n\tconst param = details.param?.toLowerCase();\n\tif (!OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message)) return false;\n\treturn param !== undefined ? OUTPUT_BUDGET_PARAMETER_PATTERN.test(param) : OUTPUT_BUDGET_PARAMETER_PATTERN.test(message);\n}\n\nexport function isRetryWorthyOutputBudgetError(assistantMessage: AssistantMessage): boolean {\n\tif (assistantMessage.stopReason !== \"error\" || !assistantMessage.errorMessage) return false;\n\tif (assistantMessage.api !== \"openai-responses\") return false;\n\n\tconst structuredDetails = parseProviderErrorDetails(assistantMessage.errorMessage);\n\tif (structuredDetails && isStructuredOutputBudgetUnderflow(structuredDetails)) return true;\n\n\treturn isOutputBudgetUnderflowText(assistantMessage.errorMessage);\n}\n\nexport function shouldRetryAfterThresholdCompaction(assistantMessage: AssistantMessage): boolean {\n\treturn isRetryWorthyLengthStop(assistantMessage) || isRetryWorthyOutputBudgetError(assistantMessage);\n}\n\n/**\n * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.\n */\n\nexport function _dropTrailingAutoCompactionRetryAssistantIfPresent(this: AgentSession): void {\n\tconst messages = this.agent.state.messages;\n\tconst lastMsg = messages[messages.length - 1];\n\tif (lastMsg?.role !== \"assistant\") return;\n\tconst stopReason = (lastMsg as AssistantMessage).stopReason;\n\tif (stopReason === \"error\" || stopReason === \"length\") {\n\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t}\n}\n\n/**\n * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.\n */\n\nexport function _schedulePostAutoCompactionContinuationProbe(this: AgentSession,\n\t_reason: \"overflow\" | \"threshold\",\n\twillRetry: boolean,\n): void {\n\tif (willRetry) {\n\t\tconst token = this._postCompactionContinuationToken + 1;\n\t\tthis._postCompactionContinuationToken = token;\n\t\tlet pending: Promise<void>;\n\t\tpending = new Promise<void>((resolve) => {\n\t\t\tsetTimeout(() => {\n\t\t\t\tvoid (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tif (this._postCompactionContinuationToken !== token) return;\n\t\t\t\t\t\tif (this.isCompacting || this.isStreaming) return;\n\t\t\t\t\t\tawait this._resumeAfterAutoCompaction();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tif (this._pendingPostCompactionContinuation === pending) {\n\t\t\t\t\t\t\tthis._pendingPostCompactionContinuation = undefined;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tresolve();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t}, 100);\n\t\t});\n\t\tthis._pendingPostCompactionContinuation = pending;\n\t\treturn;\n\t}\n\n\tsetTimeout(() => {\n\t\tif (this.isCompacting || this.isStreaming) {\n\t\t\treturn;\n\t\t}\n\n\t\tif (!this.agent.hasQueuedMessages()) {\n\t\t\treturn;\n\t\t}\n\n\t\tvoid this._resumeAfterAutoCompaction();\n\t}, 100);\n}\n\nexport async function _awaitPendingPostCompactionContinuation(this: AgentSession): Promise<void> {\n\tconst pending = this._pendingPostCompactionContinuation;\n\tif (pending === undefined) return;\n\tawait pending;\n}\n\n/**\n * Internal: resume generation after successful auto-compaction only when active work remains.\n */\n\nexport async function _resumeAfterAutoCompaction(this: AgentSession): Promise<void> {\n\ttry {\n\t\tawait this._runAgentContinue();\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthis._emit({\n\t\t\ttype: \"agent_continue_error\",\n\t\t\tsource: \"post_compaction\",\n\t\t\terrorMessage: `Post-compaction continuation failed: ${message}`,\n\t\t});\n\t}\n}\n\n/**\n * Internal: resume a response that was truncated at the output-token cap\n * (\"length\") when the context does not warrant compaction. The generation is\n * continued directly so the model finishes the work it was cut off from, rather\n * than dead-ending on the truncation. Bounded by MAX_LENGTH_CONTINUATION_ATTEMPTS\n * so a turn that keeps exceeding the per-turn output cap can still terminate.\n */\n\nexport function _resumeAfterLengthTruncation(this: AgentSession): void {\n\tif (this._lengthContinuationAttempts >= MAX_LENGTH_CONTINUATION_ATTEMPTS) return;\n\tthis._lengthContinuationAttempts += 1;\n\t// agent.continue() rejects an assistant tail; drop the incomplete\n\t// length-stopped message so the preceding user/tool-result anchors the\n\t// continuation. It remains persisted in session history.\n\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\tthis._schedulePostAutoCompactionContinuationProbe(\"threshold\", true);\n}\n\n\nfunction overflowUnresolved(reason: \"overflow\" | \"threshold\", aborted = false): boolean | undefined {\n\treturn reason === \"overflow\" && !aborted ? true : undefined;\n}\n\nexport async function _runAutoCompaction(this: AgentSession, reason: \"overflow\" | \"threshold\", willRetry: boolean): Promise<void> {\n\tthis._emit({ type: \"compaction_start\", reason });\n\tthis._autoCompactionAbortController = new AbortController();\n\n\ttry {\n\t\tif (!this.model) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\t// Auth is resolved lazily: only called when the planner fallback is needed.\n\t\t// This allows extension-provided deletion requests to run before auth is checked,\n\t\t// enabling local extension compaction even when API credentials are unavailable.\n\t\t// Auto-mode resolver returns undefined (rather than throwing) when auth is missing;\n\t\t// overflow recovery then falls back to deterministic no-auth eviction, while\n\t\t// threshold compaction keeps the previous no-op behavior.\n\t\tconst model = this.model;\n\t\tconst result = await this._applyContextVerbatimCompaction({\n\t\t\tresolvePlannerAuth: async () => {\n\t\t\t\tconst authResult = await this._modelRegistry.getApiKeyAndHeaders(model);\n\t\t\t\tif (!authResult.ok || !authResult.apiKey) {\n\t\t\t\t\treturn undefined;\n\t\t\t\t}\n\t\t\t\treturn { apiKey: authResult.apiKey, headers: authResult.headers };\n\t\t\t},\n\t\t\tabortController: this._autoCompactionAbortController,\n\t\t\tbackupLabel: reason === \"overflow\" ? \"overflow-auto-compact\" : \"auto-compact\",\n\t\t\treason,\n\t\t});\n\t\tif (!result) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\t\t}\n\n\t\tthis._emit({ type: \"compaction_end\", reason, result, aborted: false, willRetry });\n\t\tthis._schedulePostAutoCompactionContinuationProbe(reason, willRetry);\n\t} catch (error) {\n\t\tconst errorMessage = error instanceof Error ? error.message : \"compaction failed\";\n\t\tconst aborted = errorMessage === \"Compaction cancelled\" || (error instanceof Error && error.name === \"AbortError\");\n\t\tthis._emit({\n\t\t\ttype: \"compaction_end\",\n\t\t\treason,\n\t\t\tresult: undefined,\n\t\t\taborted,\n\t\t\twillRetry: false,\n\t\t\tunresolvedOverflow: overflowUnresolved(reason, aborted),\n\t\t\terrorMessage: aborted\n\t\t\t\t? undefined\n\t\t\t\t: reason === \"overflow\"\n\t\t\t\t\t? `Context overflow recovery failed: ${errorMessage}`\n\t\t\t\t\t: `Auto-compaction failed: ${errorMessage}`,\n\t\t});\n\t} finally {\n\t\tthis._autoCompactionAbortController = undefined;\n\t}\n}\n\n/**\n * Toggle auto-compaction setting.\n */\n\nexport const agentSessionAutoCompactionMethods = {\n\t_awaitPendingPostCompactionContinuation,\n\t_checkCompaction,\n\t_isCopilotServerCapBelowSelectedContextWindow,\n\t_dropTrailingAutoCompactionRetryAssistantIfPresent,\n\t_schedulePostAutoCompactionContinuationProbe,\n\t_resumeAfterAutoCompaction,\n\t_resumeAfterLengthTruncation,\n\t_runAutoCompaction,\n};\n"]}
+{"version":3,"file":"agent-session-auto-compaction.d.ts","sourceRoot":"","sources":["../../src/core/agent-session-auto-compaction.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAOrE,OAAO,KAAK,EAAE,2BAA2B,IAAI,YAAY,EAAE,MAAM,4BAA4B,CAAC;AAE9F;;;;;;GAMG;AACH,eAAO,MAAM,gCAAgC,IAAI,CAAC;AAElD,eAAO,MAAM,6CAA6C,IAAI,CAAC;AAc/D,wBAAsB,gBAAgB,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,gBAAgB,UAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CA8HrI;AAGD,wBAAgB,6CAA6C,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ7H;AAED,wBAAgB,uBAAuB,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAEnF;AAmDD,wBAAgB,8BAA8B,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ1F;AAED,wBAAgB,mCAAmC,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAE/F;AAED;;GAEG;AAEH,wBAAgB,kDAAkD,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQ3F;AAED;;GAEG;AAEH,wBAAgB,4CAA4C,CAAC,IAAI,EAAE,YAAY,EAC9E,OAAO,EAAE,UAAU,GAAG,WAAW,EACjC,SAAS,EAAE,OAAO,GAChB,IAAI,CAoCN;AAED,wBAAsB,uCAAuC,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAI/F;AAED;;GAEG;AAEH,wBAAsB,0BAA0B,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAWlF;AAED;;;;;;GAMG;AAEH,wBAAgB,4BAA4B,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQrE;AAOD,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,UAAU,GAAG,WAAW,EAAE,SAAS,EAAE,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAuEhI;AAED;;GAEG;AAEH,eAAO,MAAM,iCAAiC;;;;;;;;;CAS7C,CAAC","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai/compat\";\nimport { isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { getEffectiveInputBudget } from \"./context-window.ts\";\nimport { parseCopilotPromptLimitError } from \"./copilot-errors.ts\";\nimport { calculateContextTokens, estimateContextTokens, shouldCompact } from \"./compaction/index.ts\";\nimport { getLatestCompactionBoundaryEntry } from \"./session-manager.ts\";\nimport { MIN_RESPONSES_MAX_OUTPUT_TOKENS } from \"./openai-responses-payload-sanitizer.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\n\n/**\n * Upper bound on consecutive automatic continuations of a response that was\n * truncated at the output-token cap (\"length\") while the context is still\n * below the compaction budget. Each continuation regenerates the cut-off turn,\n * so a model that insists on emitting more than its per-turn output cap can\n * still terminate instead of looping forever.\n */\nexport const MAX_LENGTH_CONTINUATION_ATTEMPTS = 3;\n\nexport const MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS = 1;\n\ntype JsonValue = string | number | boolean | null | JsonValue[] | { [key: string]: JsonValue };\ntype ProviderErrorDetails = {\n\tmessage?: string;\n\tcode?: string;\n\tparam?: string;\n};\n\nconst OUTPUT_BUDGET_PARAMETER_PATTERN = /\\bmax_output_tokens\\b/;\nconst OUTPUT_BUDGET_UNDERFLOW_PATTERN = new RegExp(\n\t`(?:integer\\\\s+below\\\\s+minimum\\\\s+value|expected\\\\s+(?:a\\\\s+)?value\\\\s*>=\\\\s*${MIN_RESPONSES_MAX_OUTPUT_TOKENS}|got\\\\s+1\\\\s+instead)`,\n);\n\nexport async function _checkCompaction(this: AgentSession, assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {\n\t// The agent_end path passes skipAbortedCheck=true; the pre-prompt path passes\n\t// false. Only the live turn-completion path may auto-continue a truncated\n\t// response — before a fresh user prompt we must not resume the old turn.\n\tconst isLiveTurnCompletion = skipAbortedCheck;\n\tconst settings = this.settingsManager.getCompactionSettings();\n\tif (!settings.enabled) return;\n\n\t// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false\n\tif (skipAbortedCheck && assistantMessage.stopReason === \"aborted\") return;\n\n\tconst contextWindow = this.model?.contextWindow ?? 0;\n\n\t// Skip overflow check if the message came from a different model.\n\t// This handles the case where user switched from a smaller-context model (e.g. opus)\n\t// to a larger-context model (e.g. codex) - the overflow error from the old model\n\t// shouldn't trigger compaction for the new model.\n\tconst sameModel =\n\t\tthis.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;\n\n\t// Skip compaction checks if this assistant message is older than the latest\n\t// compaction boundary. This prevents a stale pre-compaction usage/error\n\t// from retriggering compaction on the first prompt after compaction.\n\tconst compactionBoundaryEntry = getLatestCompactionBoundaryEntry(this.sessionManager.getBranch());\n\tconst assistantIsFromBeforeCompactionBoundary =\n\t\tcompactionBoundaryEntry !== null &&\n\t\tassistantMessage.timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime();\n\tif (assistantIsFromBeforeCompactionBoundary) {\n\t\treturn;\n\t}\n\n\t// Case 1: Overflow - LLM returned context overflow error\n\t// When Copilot rejects a 1m client-budget prompt at a lower server cap (for example\n\t// because long-context/usage-based billing entitlement is missing), leave the friendly\n\t// error visible instead of auto-compacting down to a smaller server tier silently.\n\tif (sameModel && this._isCopilotServerCapBelowSelectedContextWindow(assistantMessage)) {\n\t\treturn;\n\t}\n\tif (sameModel && isContextOverflow(assistantMessage, contextWindow)) {\n\t\tconst willRetry = assistantMessage.stopReason !== \"stop\";\n\t\tif (!willRetry) {\n\t\t\tawait this._runAutoCompaction(\"overflow\", false);\n\t\t\treturn;\n\t\t}\n\n\t\tif (this._overflowRecoveryAttempted) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason: \"overflow\",\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: true,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tthis._overflowRecoveryAttempted = true;\n\t\t// Remove the error message from agent state (it IS saved to session for history,\n\t\t// but we don't want it in context for the retry)\n\t\tconst messages = this.agent.state.messages;\n\t\tif (messages.length > 0 && messages[messages.length - 1].role === \"assistant\") {\n\t\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t\t}\n\t\tawait this._runAutoCompaction(\"overflow\", willRetry);\n\t\treturn;\n\t}\n\n\t// Case 2: Threshold - context is getting large\n\t// For error messages (no usage data), estimate from last successful response.\n\t// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.\n\tlet contextTokens: number;\n\tif (assistantMessage.stopReason === \"error\") {\n\t\tconst messages = this.agent.state.messages;\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) return; // No usage data at all\n\t\t// Verify the usage source is post-compaction. Kept pre-compaction messages\n\t\t// have stale usage reflecting the old (larger) context and would falsely\n\t\t// trigger compaction right after one just finished.\n\t\tconst usageMsg = messages[estimate.lastUsageIndex];\n\t\tif (\n\t\t\tcompactionBoundaryEntry &&\n\t\t\tusageMsg.role === \"assistant\" &&\n\t\t\t(usageMsg as AssistantMessage).timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime()\n\t\t) {\n\t\t\treturn;\n\t\t}\n\t\tcontextTokens = estimate.tokens;\n\t} else {\n\t\tcontextTokens = calculateContextTokens(assistantMessage.usage);\n\t}\n\t// Compact against the effective input budget (the hard prompt cap for providers like Copilot\n\t// that advertise a larger total window) so we compact before overrunning the server-side limit\n\t// rather than relying on reactive overflow recovery near the cap.\n\tconst compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;\n\tif (shouldCompact(contextTokens, compactionBudget, settings)) {\n\t\tconst willRetry = shouldRetryAfterThresholdCompaction(assistantMessage);\n\t\tif (willRetry && isRetryWorthyOutputBudgetError(assistantMessage)) {\n\t\t\tif (this._outputBudgetErrorContinuationAttempts >= MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS) {\n\t\t\t\tthis._emit({\n\t\t\t\t\ttype: \"compaction_end\",\n\t\t\t\t\treason: \"threshold\",\n\t\t\t\t\tresult: undefined,\n\t\t\t\t\taborted: false,\n\t\t\t\t\twillRetry: false,\n\t\t\t\t\terrorMessage:\n\t\t\t\t\t\t\"Output-budget recovery stopped after a compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t\t});\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tthis._outputBudgetErrorContinuationAttempts += 1;\n\t\t}\n\t\tawait this._runAutoCompaction(\"threshold\", willRetry);\n\t\treturn;\n\t}\n\n\t// A response truncated at the output-token cap (\"length\") with the context\n\t// still below the compaction budget is genuine work cut off mid-flight, not a\n\t// context overflow. Compaction would not free any room, so continue the\n\t// generation directly instead of dead-ending on the truncation and leaving\n\t// the task half-finished.\n\tif (isLiveTurnCompletion && isRetryWorthyLengthStop(assistantMessage)) {\n\t\tthis._resumeAfterLengthTruncation();\n\t}\n}\n\n\nexport function _isCopilotServerCapBelowSelectedContextWindow(this: AgentSession, assistantMessage: AssistantMessage): boolean {\n\tif (!this.model || this.model.provider !== \"github-copilot\" || !assistantMessage.errorMessage) return false;\n\tconst promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);\n\t// Compare against the effective input budget (the model's real prompt cap), not the displayed\n\t// total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;\n\t// only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account\n\t// to a lower server tier) keeps the friendly error visible instead of silently compacting down.\n\treturn promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;\n}\n\nexport function isRetryWorthyLengthStop(assistantMessage: AssistantMessage): boolean {\n\treturn assistantMessage.stopReason === \"length\" && assistantMessage.usage.output > 0;\n}\n\nfunction isJsonRecord(value: JsonValue): value is { [key: string]: JsonValue } {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction stringField(record: { [key: string]: JsonValue }, key: string): string | undefined {\n\tconst value = record[key];\n\treturn typeof value === \"string\" ? value : undefined;\n}\n\nfunction parseProviderErrorDetails(errorMessage: string): ProviderErrorDetails | undefined {\n\tconst start = errorMessage.indexOf(\"{\");\n\tconst end = errorMessage.lastIndexOf(\"}\");\n\tif (start === -1 || end <= start) return undefined;\n\n\ttry {\n\t\tconst parsed = JSON.parse(errorMessage.slice(start, end + 1)) as JsonValue;\n\t\tif (!isJsonRecord(parsed)) return undefined;\n\n\t\tconst nestedError = parsed.error;\n\t\tif (nestedError !== undefined && isJsonRecord(nestedError)) {\n\t\t\treturn {\n\t\t\t\tmessage: stringField(nestedError, \"message\"),\n\t\t\t\tcode: stringField(nestedError, \"code\") ?? stringField(parsed, \"code\"),\n\t\t\t\tparam: stringField(nestedError, \"param\") ?? stringField(parsed, \"param\"),\n\t\t\t};\n\t\t}\n\n\t\treturn {\n\t\t\tmessage: stringField(parsed, \"message\"),\n\t\t\tcode: stringField(parsed, \"code\"),\n\t\t\tparam: stringField(parsed, \"param\"),\n\t\t};\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction isOutputBudgetUnderflowText(text: string): boolean {\n\tconst message = text.toLowerCase();\n\treturn OUTPUT_BUDGET_PARAMETER_PATTERN.test(message) && OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message);\n}\n\nfunction isStructuredOutputBudgetUnderflow(details: ProviderErrorDetails): boolean {\n\tconst message = details.message?.toLowerCase() ?? \"\";\n\tconst param = details.param?.toLowerCase();\n\tif (!OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message)) return false;\n\treturn param !== undefined ? OUTPUT_BUDGET_PARAMETER_PATTERN.test(param) : OUTPUT_BUDGET_PARAMETER_PATTERN.test(message);\n}\n\nexport function isRetryWorthyOutputBudgetError(assistantMessage: AssistantMessage): boolean {\n\tif (assistantMessage.stopReason !== \"error\" || !assistantMessage.errorMessage) return false;\n\tif (assistantMessage.api !== \"openai-responses\") return false;\n\n\tconst structuredDetails = parseProviderErrorDetails(assistantMessage.errorMessage);\n\tif (structuredDetails && isStructuredOutputBudgetUnderflow(structuredDetails)) return true;\n\n\treturn isOutputBudgetUnderflowText(assistantMessage.errorMessage);\n}\n\nexport function shouldRetryAfterThresholdCompaction(assistantMessage: AssistantMessage): boolean {\n\treturn isRetryWorthyLengthStop(assistantMessage) || isRetryWorthyOutputBudgetError(assistantMessage);\n}\n\n/**\n * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.\n */\n\nexport function _dropTrailingAutoCompactionRetryAssistantIfPresent(this: AgentSession): void {\n\tconst messages = this.agent.state.messages;\n\tconst lastMsg = messages[messages.length - 1];\n\tif (lastMsg?.role !== \"assistant\") return;\n\tconst stopReason = (lastMsg as AssistantMessage).stopReason;\n\tif (stopReason === \"error\" || stopReason === \"length\") {\n\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t}\n}\n\n/**\n * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.\n */\n\nexport function _schedulePostAutoCompactionContinuationProbe(this: AgentSession,\n\t_reason: \"overflow\" | \"threshold\",\n\twillRetry: boolean,\n): void {\n\tif (willRetry) {\n\t\tconst token = this._postCompactionContinuationToken + 1;\n\t\tthis._postCompactionContinuationToken = token;\n\t\tlet pending: Promise<void>;\n\t\tpending = new Promise<void>((resolve) => {\n\t\t\tsetTimeout(() => {\n\t\t\t\tvoid (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tif (this._postCompactionContinuationToken !== token) return;\n\t\t\t\t\t\tif (this.isCompacting || this.isStreaming) return;\n\t\t\t\t\t\tawait this._resumeAfterAutoCompaction();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tif (this._pendingPostCompactionContinuation === pending) {\n\t\t\t\t\t\t\tthis._pendingPostCompactionContinuation = undefined;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tresolve();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t}, 100);\n\t\t});\n\t\tthis._pendingPostCompactionContinuation = pending;\n\t\treturn;\n\t}\n\n\tsetTimeout(() => {\n\t\tif (this.isCompacting || this.isStreaming) {\n\t\t\treturn;\n\t\t}\n\n\t\tif (!this.agent.hasQueuedMessages()) {\n\t\t\treturn;\n\t\t}\n\n\t\tvoid this._resumeAfterAutoCompaction();\n\t}, 100);\n}\n\nexport async function _awaitPendingPostCompactionContinuation(this: AgentSession): Promise<void> {\n\tconst pending = this._pendingPostCompactionContinuation;\n\tif (pending === undefined) return;\n\tawait pending;\n}\n\n/**\n * Internal: resume generation after successful auto-compaction only when active work remains.\n */\n\nexport async function _resumeAfterAutoCompaction(this: AgentSession): Promise<void> {\n\ttry {\n\t\tawait this._runAgentContinue();\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthis._emit({\n\t\t\ttype: \"agent_continue_error\",\n\t\t\tsource: \"post_compaction\",\n\t\t\terrorMessage: `Post-compaction continuation failed: ${message}`,\n\t\t});\n\t}\n}\n\n/**\n * Internal: resume a response that was truncated at the output-token cap\n * (\"length\") when the context does not warrant compaction. The generation is\n * continued directly so the model finishes the work it was cut off from, rather\n * than dead-ending on the truncation. Bounded by MAX_LENGTH_CONTINUATION_ATTEMPTS\n * so a turn that keeps exceeding the per-turn output cap can still terminate.\n */\n\nexport function _resumeAfterLengthTruncation(this: AgentSession): void {\n\tif (this._lengthContinuationAttempts >= MAX_LENGTH_CONTINUATION_ATTEMPTS) return;\n\tthis._lengthContinuationAttempts += 1;\n\t// agent.continue() rejects an assistant tail; drop the incomplete\n\t// length-stopped message so the preceding user/tool-result anchors the\n\t// continuation. It remains persisted in session history.\n\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\tthis._schedulePostAutoCompactionContinuationProbe(\"threshold\", true);\n}\n\n\nfunction overflowUnresolved(reason: \"overflow\" | \"threshold\", aborted = false): boolean | undefined {\n\treturn reason === \"overflow\" && !aborted ? true : undefined;\n}\n\nexport async function _runAutoCompaction(this: AgentSession, reason: \"overflow\" | \"threshold\", willRetry: boolean): Promise<void> {\n\tthis._emit({ type: \"compaction_start\", reason });\n\tthis._autoCompactionAbortController = new AbortController();\n\n\ttry {\n\t\tif (!this.model) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\t// Resolve auth only after extension hooks have had an opportunity to cancel\n\t\t// compaction or provide compacted text, so local extension compaction does not\n\t\t// require provider credentials. Missing auth then fails model-driven compaction\n\t\t// before persistence or continuation, matching other provider-call failures.\n\t\tconst model = this.model;\n\t\tconst result = await this._applyVerbatimCompaction({\n\t\t\tresolvePlannerAuth: async () => {\n\t\t\t\tconst authResult = await this._modelRegistry.getApiKeyAndHeaders(model);\n\t\t\t\tif (!authResult.ok || !authResult.apiKey) {\n\t\t\t\t\treturn undefined;\n\t\t\t\t}\n\t\t\t\treturn { apiKey: authResult.apiKey, headers: authResult.headers };\n\t\t\t},\n\t\t\tabortController: this._autoCompactionAbortController,\n\t\t\tbackupLabel: reason === \"overflow\" ? \"overflow-auto-compact\" : \"auto-compact\",\n\t\t\treason,\n\t\t});\n\t\tif (!result) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\t\t}\n\n\t\tthis._emit({ type: \"compaction_end\", reason, result, aborted: false, willRetry });\n\t\tthis._schedulePostAutoCompactionContinuationProbe(reason, willRetry);\n\t} catch (error) {\n\t\tconst errorMessage = error instanceof Error ? error.message : \"compaction failed\";\n\t\tconst aborted = errorMessage === \"Compaction cancelled\" || (error instanceof Error && error.name === \"AbortError\");\n\t\tthis._emit({\n\t\t\ttype: \"compaction_end\",\n\t\t\treason,\n\t\t\tresult: undefined,\n\t\t\taborted,\n\t\t\twillRetry: false,\n\t\t\tunresolvedOverflow: overflowUnresolved(reason, aborted),\n\t\t\terrorMessage: aborted\n\t\t\t\t? undefined\n\t\t\t\t: reason === \"overflow\"\n\t\t\t\t\t? `Context overflow recovery failed: ${errorMessage}`\n\t\t\t\t\t: `Auto-compaction failed: ${errorMessage}`,\n\t\t});\n\t} finally {\n\t\tthis._autoCompactionAbortController = undefined;\n\t}\n}\n\n/**\n * Toggle auto-compaction setting.\n */\n\nexport const agentSessionAutoCompactionMethods = {\n\t_awaitPendingPostCompactionContinuation,\n\t_checkCompaction,\n\t_isCopilotServerCapBelowSelectedContextWindow,\n\t_dropTrailingAutoCompactionRetryAssistantIfPresent,\n\t_schedulePostAutoCompactionContinuationProbe,\n\t_resumeAfterAutoCompaction,\n\t_resumeAfterLengthTruncation,\n\t_runAutoCompaction,\n};\n"]}

package/dist/core/agent-session-auto-compaction.js

@@ -312,14 +312,12 @@ export async function _runAutoCompaction(reason, willRetry) {
             });
             return;
         }
-        // Auth is resolved lazily: only called when the planner fallback is needed.
-        // This allows extension-provided deletion requests to run before auth is checked,
-        // enabling local extension compaction even when API credentials are unavailable.
-        // Auto-mode resolver returns undefined (rather than throwing) when auth is missing;
-        // overflow recovery then falls back to deterministic no-auth eviction, while
-        // threshold compaction keeps the previous no-op behavior.
+        // Resolve auth only after extension hooks have had an opportunity to cancel
+        // compaction or provide compacted text, so local extension compaction does not
+        // require provider credentials. Missing auth then fails model-driven compaction
+        // before persistence or continuation, matching other provider-call failures.
         const model = this.model;
-        const result = await this._applyContextVerbatimCompaction({
+        const result = await this._applyVerbatimCompaction({
             resolvePlannerAuth: async () => {
                 const authResult = await this._modelRegistry.getApiKeyAndHeaders(model);
                 if (!authResult.ok || !authResult.apiKey) {