pi-goal-x 0.30.5 → 0.31.1

package/README.md

@@ -35,6 +35,12 @@ Sisyphus goals work well for migrations, staged refactors, release procedures, d
 
 The `/goal` and `/sisyphus` commands start a guided drafting process. The agent can ask focused questions, clarify the objective, and propose a task plan for confirmation.
 
+RPC and browser hosts use native select/input dialogs when the terminal overlay is unavailable. Questions retain their context, and proposal confirmation includes a choice of completion auditor. Cancelling or losing a dialog never confirms an unseen proposal; the draft stays available.
+
+When used with `pi-subagents`, goal mode stays in the parent session. Fresh, forked, resumed, and nested children inherit their assignment context without goal tools, goal state writes, or automatic goal continuations. Ordinary interactive forks retain goal mode.
+
+Audit progress remains visible in the dashboard. Audit transcript messages appear after the agent settles and are excluded from model input; existing sessions containing interleaved audit events can resume after loading the updated extension, without rewriting their history.
+
 The proposal is written to the conversation as a durable summary (objective, plan, verification, automatic continuation, auditor state); confirming it creates and focuses the goal and starts working automatically.
 
 ### Direct goal creation
@@ -251,6 +257,18 @@ Pressing `Esc` during active work pauses the goal.
 
 The agent can divide a goal into tasks and subtasks and update them as work progresses. The current task is tracked explicitly (persisted as the goal's execution focus) and highlighted in the dashboard; starting a task with `update_goal_task(status="start")` sets it, and completing or skipping it clears it.
 
+The model receives the objective excerpt, current task, next three pending tasks, progress, budget, and applicable rules. Full requirements, contracts, and evidence remain saved. `get_goal()` returns a compact summary; use `get_goal({section:"objective"})`, `section:"tasks"` (optionally `task_id`), or `section:"history"` for lossless pages of at most 4,000 content characters. Repeat the section and task with the returned `cursor` to continue. A changed source rejects stale cursors and asks for a restart. Existing `verbose` and `include_history` arguments remain supported.
+
+Task progress can be recorded atomically in order:
+
+```json
+{"updates":[{"task_id":"child","status":"complete","evidence":"Tests passed"},{"task_id":"parent","status":"complete"},{"task_id":"next","status":"start"}]}
+```
+
+Pass this to `update_goal_task`. A batch accepts 1–100 updates and validates every transition before applying any. Complete children before their parent; contracted tasks still require evidence, skips require reasons, and completed tasks cannot reopen. Single-task calls remain supported; do not mix the two forms. Each update retains its own ledger event. Turn-level writes check the original persisted revision, and completion audits wait for a successful flush.
+
+Goal tools appear when applicable. Outside drafting, creation and inspection remain available; outcomes appear for active, paused, or budget-limited goals, task planning for active or paused goals, and task progress for active goals with a task tree. Disabling tasks hides both task tools. Drafting keeps its question/proposal tools, and ordinary work tools remain available.
+
 Verification contracts describe the evidence required for completion. They can apply to the entire goal or to an individual task.
 
 Examples include:

package/docs/architecture.md

@@ -35,7 +35,10 @@ handlers from their dedicated modules:
 | `goal-draft.ts` | Drafting prompt/confirmation text helpers (goalDraftingPrompt, buildDraftConfirmationText, renderConfirmationTasks, GoalDraftingFocus) |
 | `goal-drafting.ts` | Guided drafting orchestration: durable `pi-goal-draft` session entries (survive compaction/tree navigation), resume/replace/cancel protection, transient drafting profile, `goal_question`/`goal_questionnaire`/`propose_goal_draft` tools, per-draft auditor selection |
 | `goal-questionnaire.ts` | Structured question/answer UI (`runGoalQuestionnaire`, `showProposalDialog`) used by the drafting tools and confirmations |
-| `goal-tool-names.ts` | The five published tool-name constants, fixed three/five profiles, work/progress classification, post-stop allowlist |
+| `goal-tool-names.ts` | The five published tool-name constants, lifecycle-dependent profiles, work/progress classification, post-stop allowlist |
+| `goal-detail.ts` | Lossless objective/task/history paging with content-bound cursors |
+| `goal-ledger-index.ts` | Incremental per-goal activity, audit, lifecycle, and Oracle projections |
+| `goal-task-index.ts` | Content-keyed task snapshots shared by prompts, tools, and dashboard models |
 | `prompts/goal-prompts.ts` | Bounded five-tool steering prompts (active-goal, continuation, stale-checkpoint, unfocused, budget-limited) |
 | `storage/goal-files.ts` | Goal path safety, serialization/parsing, active-file scanning, active-file writes, archive writes, prompt-body merge from disk |
 | `widgets/goal-widget.ts` | Above-editor Goal Beacon component |
@@ -197,19 +200,43 @@ The extension registers five normal-execution tools and three drafting-only tool
 | Tool | Purpose |
 |---|---|
 | `create_goal` | Create and focus a new goal after an explicit user request (objective 1–4000 chars, optional `mode` regular/sisyphus and `token_budget`). |
-| `get_goal` | Read-only complete focused goal snapshot. |
+| `get_goal` | Compact summary by default; lossless objective/tasks/history pages (up to 4,000 content characters), optional task selection, legacy verbose/history forms. |
 | `update_goal` | Run outcomes: `complete` (audited from actual evidence; optional `completion_summary` is an untrusted claim), `blocked` (after three consecutive identical blockers), or `paused` (immediate agent pause with required `reason`). |
 | `set_goal_tasks` | Create or structurally replace the task tree (flat parent-linked input, confirmation dialog, id-stable merge). |
-| `update_goal_task` | Update one task without stopping the turn: complete (evidence for contracted tasks), skipped (reason), pending (reopens skipped). |
+| `update_goal_task` | Update one task or an ordered atomic batch: start, complete (evidence for contracted tasks), skipped (reason), pending (reopens skipped). |
 | `goal_question` | Drafting-only structured clarification question. |
 | `goal_questionnaire` | Drafting-only multi-question clarification UI. |
 | `propose_goal_draft` | Drafting-only objective/task proposal with Confirm or Continue Chatting. |
 
-The normal execution profile is fixed: exactly five goal tools when tasks are
-enabled, exactly three when disabled. A user-started guided draft is the sole
-exception: it replaces those goal tools with question/questionnaire/proposal
-tools until confirmation or cancellation. Ordinary pi work tools are never
-touched. Invalid lifecycle calls return concise state-aware tool results.
+Outside drafting, creation and inspection are always advertised. Outcomes are
+advertised for active/paused/budget-limited goals; task planning for active or
+paused goals; task progress for active goals with tasks. Disabled tasks hide both
+task tools. A guided draft replaces goal tools with question/questionnaire/proposal
+tools until confirmation or cancellation. Ordinary pi work tools remain available.
+The SDK active profile changes only when membership changes. Executors still
+validate lifecycle state, including stale calls made after a transition.
+
+Normal prompt/dashboard reads use per-goal ledger indexes: 12 recent events,
+64 activity candidates in stable timestamp order, pinned audit/lifecycle state,
+and blocker-fingerprint Oracle state. Appends extend these indexes without
+copying history. Version 3 derived checkpoints use UTF-8 byte offsets and retain
+reconstructed Oracle advice; old or
+corrupt checkpoints rebuild from the authoritative JSONL ledger. Full history
+is loaded for explicit history/diagnostic requests. Saved goal/session formats
+remain compatible. Task presentation caches use content snapshots; usage changes
+do not rebuild task trees. Stable prompt rules precede changing goal data and
+usage counters, with retrieval instructions beside bounded excerpts.
+
+Warm settings resolution is cached by layer identity and applicable environment
+values; public resolved values remain independently editable. Task and prompt
+caches compare content fields without repeatedly serializing long contracts.
+Compiled detail pages reuse their source and cursor hash; history sources are
+invalidated by an opaque ledger generation on append or refresh. Cache size
+limits bound retained data; larger requirements remain available losslessly.
+Pure ANSI text wrapping/truncation is cached by text and layout arguments across
+widgets and dialogs. Live auditor previews scan only the report tail, while the
+complete final report remains available. Measurements and the module audit are
+in `specs/2026-09-07-comprehensive-optimization/`.
 
 The `tool_call` interceptor blocks work tools after a stop tool has fired in
 the same turn, and blocks work tools when the checkpoint that triggered the
@@ -283,7 +310,7 @@ entries and run them in one Node process with small test-only adapters for the
 SDK values used by handlers. This avoids loading unrelated model-provider and
 TUI media modules. The fast path requires Node 22.15+; `test:serial` remains
 the slow, real-SDK, process-isolated
-compatibility path. The suites cover: surface baselines (exactly the fixed five/three tool
+compatibility path. The suites cover: surface baselines (lifecycle-dependent goal tool
 profile and fourteen commands), golden file/ledger fixtures, stale-continuation behavior,
 GoalService mutation boundary, runtime/accounting, token-budget transitions,
 task-tool consolidation, verification contracts, the independent auditor,
@@ -298,7 +325,7 @@ The 2026-08-04 hardening plan
 is implemented: paused-status normalization (status authoritative, legacy
 `autoContinue: true` records stay paused), disk-fresh task transactions with
 structural-field clearing, token-budget integer validation, `task_reopened`
-ledger semantics with observable diagnostics, the three/five fixed tool
+ledger semantics with observable diagnostics, the lifecycle-dependent tool
 profile, and the supported integration/experiment coverage described above.
 (the interim drafting-surface removal was later reversed by the product
 correction in the runtime follow-up, which restores guided drafting as a

package/extensions/goal-activity.ts

@@ -136,6 +136,33 @@ function mapEvent(event: GoalLedgerEvent, taskTitles: ReadonlyMap<string, string
 	}
 }
 
+const indexedActivity = new WeakSet<readonly GoalLedgerEvent[]>();
+export function indexedActivityEvents(events: GoalLedgerEvent[]): GoalLedgerEvent[] {
+ indexedActivity.add(events);
+ return events;
+}
+
+/** Stable event identity for append-order deduplication, independent of title edits. */
+export function activityEventKey(event: GoalLedgerEvent): string | undefined {
+ const item = mapEvent(event, undefined);
+ return item ? `${item.kind}:${item.text}` : undefined;
+}
+
+const activityTypes = new Set([
+ "goal_created", "goal_tweaked", "auditor_toggled", "goal_paused", "goal_resumed", "goal_blocked",
+ "goal_budget_limited", "goal_completed", "goal_aborted", "task_started", "task_complete", "task_skipped",
+ "task_reopened", "completion_requested", "audit_started", "audit_result", "audit_skipped", "goal_archived",
+]);
+export function isActivityEvent(event: GoalLedgerEvent): boolean { return activityTypes.has(event.type); }
+
+/** Avoid formatting distinct task events during a cold ledger rebuild. */
+export function sameActivityEvent(left: GoalLedgerEvent, right: GoalLedgerEvent): boolean {
+ if (left.type !== right.type) return false;
+ if ("taskId" in left && "taskId" in right && left.taskId !== right.taskId) return false;
+ // Reasons/evidence use the exact display normalization for true duplicates.
+ return activityEventKey(left) === activityEventKey(right);
+}
+
 /**
  * Derive the readable activity feed for one goal from its durable ledger
  * events. Returns items in chronological order (oldest first), capped to the
@@ -148,6 +175,17 @@ export function deriveGoalActivity(
 ): GoalActivityItem[] {
 	const { taskTitles, limit = 5 } = options;
 	const items: GoalActivityItem[] = [];
+	if (indexedActivity.has(events)) {
+		// Runtime indexes already deduplicate and sort. Format only the requested tail.
+		for (let i = events.length - 1; i >= 0; i--) {
+			const event = events[i]!;
+			if (!("goalId" in event) || event.goalId !== goalId) continue;
+			const item = mapEvent(event, taskTitles);
+			if (item) items.push(item);
+			if (limit > 0 && items.length >= limit) break;
+		}
+		return items.reverse().slice(-limit);
+	}
 	for (const event of events) {
 		if (!("goalId" in event) || event.goalId !== goalId) continue;
 		const item = mapEvent(event, taskTitles);

package/extensions/goal-auditor.ts

@@ -109,6 +109,20 @@ function escapePromptPayload(value: string): string {
 	return value.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
 }
 
+/** Stream previews need only the tail, not a split/copy of the full growing report. */
+export function recentNonEmptyLines(text: string, limit: number): string[] {
+	const lines: string[] = [];
+	let end = text.length;
+	while (end >= 0 && lines.length < limit) {
+		const newline = end > 0 ? text.lastIndexOf("\n", end - 1) : -1;
+		const line = text.slice(newline + 1, end);
+		if (line.trim()) lines.push(line);
+		if (newline < 0) break;
+		end = newline;
+	}
+	return lines.reverse();
+}
+
 /** §60: human-readable labels for the auditor's read-only tool set. */
 export function labelForReadOnlyTool(toolName: string): string {
 	switch (toolName) {
@@ -160,15 +174,14 @@ export function buildGoalAuditorPrompt(args: {
 	warmContext?: string | null;
 }): string {
 	return [
-		"You are the independent completion auditor for pi-goal.",
-		"The executor claims the goal is complete. Your job is to decide whether the user's objective is actually satisfied.",
-		"Be skeptical and semantic. Do not approve from paperwork, intent, file count, word count, build success, or a plausible summary alone.",
-		"Use read/grep/find/ls/bash as needed to inspect real artifacts. Do not mutate files or run destructive commands.",
-		"If the work is only an alpha scaffold, generated template, shallow draft, proxy milestone, or lacks the user-facing value requested, disapprove.",
-		"If any explicit requirement is missing, weakly verified, contradicted, or not inspectable with the available evidence, disapprove.",
-		"Return a concise audit report. The final line MUST be exactly one of:",
-		"<approved/>",
-		"<disapproved/>",
+		"You are the independent completion auditor for pi-goal. Decide whether the user's objective is actually satisfied.",
+		"Audit checklist:",
+		"1. Extract the real success criteria, including every explicit requirement and quality/reader outcome. Disapprove missing, contradicted, weakly verified or uninspectable requirements.",
+		"2. Inspect real artifacts with read/grep/find/ls/bash as needed. Do not mutate files or run destructive commands. Paperwork, intent, file/word counts, build success and plausible summaries alone are not proof.",
+		...(!args.settings?.disableContracts && args.goal.verificationContract?.trim()
+			? ["3. Verify that the executor has satisfied every item in the <verification_contract>. If any item is missing or weakly addressed, disapprove."] : []),
+		"4. Explain missing or weak evidence concisely. Disapprove alpha scaffold, generated template, shallow draft or proxy milestones lacking the user-facing value requested.",
+		"5. End with exactly <approved/> only if the objective is truly complete; otherwise end with exactly <disapproved/>.",
 		"",
 		"Goal objective:",
 		"<objective>",
@@ -209,32 +222,22 @@ export function buildGoalAuditorPrompt(args: {
 			escapePromptPayload(args.warmContext.trim()),
 			"</warm_context>",
 		] : []),
-		"",
-		"Audit checklist:",
-		...[
-			"1. Extract the real success criteria from the objective, including quality/reader outcomes.",
-			"2. Inspect artifacts or command output that can prove or disprove those criteria. Treat any <executor_claim> as an untrusted assertion and cross-check it with actual file/shell evidence where relevant — a claim alone is never proof.",
-			...(!args.settings?.disableContracts && args.goal.verificationContract?.trim()
-				? ["3. Verify that the executor has satisfied every item in the <verification_contract>. If any item is missing or weakly addressed, disapprove."]
-				: []),
-			"4. Explain missing or weak evidence, especially scaffold-vs-final quality gaps.",
-			"5. End with exactly <approved/> only if the objective is truly complete; otherwise end with exactly <disapproved/>.",
-		],
+
 	].join("\n");
 }
 
-function makeAuditorResourceLoader(): ResourceLoader {
+export function makeAuditorResourceLoader(systemPrompt = [
+	"You are a read-only completion auditor running in an isolated pi agent session.",
+	"Inspect the repository and decide whether the claimed goal completion is genuinely satisfied.",
+	"Never modify files. Never approve unless the actual user objective is complete.",
+].join("\n")): ResourceLoader {
 	return {
 		getExtensions: () => ({ extensions: [], errors: [], runtime: createExtensionRuntime() }),
 		getSkills: () => ({ skills: [], diagnostics: [] }),
 		getPrompts: () => ({ prompts: [], diagnostics: [] }),
 		getThemes: () => ({ themes: [], diagnostics: [] }),
 		getAgentsFiles: () => ({ agentsFiles: [] }),
-		getSystemPrompt: () => [
-			"You are a read-only completion auditor running in an isolated pi agent session.",
-			"Inspect the repository and decide whether the claimed goal completion is genuinely satisfied.",
-			"Never modify files. Never approve unless the actual user objective is complete.",
-		].join("\n"),
+		getSystemPrompt: () => systemPrompt,
 		getSystemPromptSource: () => undefined,
 		getAppendSystemPrompt: () => [],
 		getAppendSystemPromptSources: () => [],
@@ -322,6 +325,7 @@ export async function runGoalCompletionAuditor(args: {
 	const model = resolved.model;
 	const thinkingLevel = config.thinkingLevel;
 	const outputParts: string[] = [];
+	let outputTail: string[] = [];
 	if (resolved.error) {
 		return { approved: false, disapproved: true, output: "", model: modelLabel(model), thinkingLevel, error: resolved.error };
 	}
@@ -404,10 +408,10 @@ export async function runGoalCompletionAuditor(args: {
 				const message = event.message as { role?: string; content?: Array<{ type?: string; text?: string }> };
 				if (message?.role === "assistant") {
 					for (const part of message.content ?? []) {
-						if (part.type === "text" && typeof part.text === "string" && part.text.trim()) {
+						if (part.type === "text" && typeof part.text === "string") {
 							// Keep the last 5 non-empty text lines for live display
-							const lines = part.text.split("\n").filter((l: string) => l.trim());
-							progress.recentOutput = [...lines.slice(-5)];
+							const lines = recentNonEmptyLines(part.text, 5);
+							if (lines.length) progress.recentOutput = lines;
 						}
 					}
 				}
@@ -418,18 +422,19 @@ export async function runGoalCompletionAuditor(args: {
 			const message = event.message as { role?: string; content?: Array<{ type?: string; text?: string }> };
 			if (message.role !== "assistant") return;
 			for (const part of message.content ?? []) {
-				if (part.type === "text" && typeof part.text === "string") outputParts.push(part.text);
+				if (part.type === "text" && typeof part.text === "string") {
+					outputParts.push(part.text);
+					outputTail = [...outputTail, ...recentNonEmptyLines(part.text, 8)].slice(-8);
+				}
 			}
 			// Final report production: derived label for the widget.
-			if (outputParts.some((t) => t.trim())) {
+			if (outputTail.length > 0) {
 				progress.label = "Producing report...";
 				progress.percentage = Math.max(progress.percentage ?? 0, 90);
 				progress.phase = "producing_report";
 			}
 			// Show the accumulated output in progress
-			const fullText = outputParts.join("\n\n");
-			const lines = fullText.split("\n").filter((l: string) => l.trim());
-			progress.recentOutput = lines.slice(-8);
+			progress.recentOutput = outputTail;
 			emitProgress();
 		});
 		// Wire the external AbortSignal to abort the running session when fired
@@ -451,6 +456,7 @@ export async function runGoalCompletionAuditor(args: {
 			progress.percentage = 100;
 			emitProgress();
 			unsubscribe();
+   session.dispose?.();
 		}
 		// session.abort() does NOT throw — the agent loop returns normally with
 		// whatever output was captured before the abort. Check the signal after

package/extensions/goal-commands.ts

@@ -170,7 +170,7 @@ export function registerGoalCommands(core: GoalCore): void {
 			} catch {}
 		}
 		if (!current) {
-			const openCount = core.openGoals().length;
+			const openCount = otherOpenGoalCount(core.goalsById, null);
 			ctx.ui.notify(openCount > 0 ? buildUnfocusedOpenGoalsSummary(openCount) : detailedSummary(null), "info");
 			return;
 		}
@@ -306,7 +306,7 @@ export function registerGoalCommands(core: GoalCore): void {
 	async function handleGoalPause(ctx: ExtensionContext): Promise<void> {
 		core.reconcileFocusedGoalFromDisk(ctx);
 		if (!core.state.goal) {
-			if (core.openGoals().length > 0) {
+			if (otherOpenGoalCount(core.goalsById, null) > 0) {
 				const selected = await chooseOpenGoal(ctx, "Pause which open goal?");
 				if (!selected) return;
 			} else {
@@ -329,7 +329,7 @@ export function registerGoalCommands(core: GoalCore): void {
 
 	async function handleGoalResume(ctx: ExtensionContext): Promise<void> {
 		core.reconcileFocusedGoalFromDisk(ctx);
-		if (!core.state.goal && core.openGoals().length > 0) {
+		if (!core.state.goal && otherOpenGoalCount(core.goalsById, null) > 0) {
 			const selected = await chooseOpenGoal(ctx, "Resume or focus open goal");
 			if (!selected) return;
 			if (selected.status === "active") {
@@ -660,7 +660,7 @@ export function registerGoalCommands(core: GoalCore): void {
 	}
 	async function handleGoalClear(ctx: ExtensionContext): Promise<void> {
 		core.reconcileFocusedGoalFromDisk(ctx);
-		if (!core.state.goal && core.openGoals().length > 0) {
+		if (!core.state.goal && otherOpenGoalCount(core.goalsById, null) > 0) {
 			const selected = await chooseOpenGoal(ctx, "Clear which open goal?");
 			if (!selected) return;
 		}
@@ -700,7 +700,7 @@ export function registerGoalCommands(core: GoalCore): void {
 	async function runGoalTweak(replacement: string, ctx: ExtensionContext): Promise<void> {
 		core.reconcileFocusedGoalFromDisk(ctx);
 		if (!core.state.goal) {
-			if (core.openGoals().length > 0) {
+			if (otherOpenGoalCount(core.goalsById, null) > 0) {
 				const selected = await chooseOpenGoal(ctx, "Tweak which open goal?");
 				if (!selected) return;
 			} else {

package/extensions/goal-compaction.ts

@@ -11,7 +11,8 @@ import {
   type GoalLedgerEvent,
   type LedgerStateReadResult,
 } from "./goal-ledger.ts";
-import { type GoalRecord, type GoalTask } from "./goal-record.ts";
+import { type GoalRecord } from "./goal-record.ts";
+import { taskIndex } from "./goal-task-index.ts";
 
 export function buildGoalCompactSummary(
   goal: GoalRecord,
@@ -169,17 +170,10 @@ export function buildPostCompactionGoalDelta(args: {
   lines.push(`[POST-COMPACTION RESYNC goalId=${goal.id}]`);
   // Execution focus (with contract) — the one thing memory must not lose.
   if (goal.currentTaskId && goal.taskList) {
-    const find = (tasks: GoalTask[]): GoalTask | undefined => {
-      for (const t of tasks) {
-        if (t.id === goal.currentTaskId) return t;
-        const nested = t.subtasks ? find(t.subtasks) : undefined;
-        if (nested) return nested;
-      }
-      return undefined;
-    };
-    const current = find(goal.taskList.tasks);
+    const current = taskIndex(goal.taskList.tasks).byId.get(goal.currentTaskId);
     if (current) {
-      lines.push(`Current task: ${current.id} — ${current.title}${current.verificationContract ? ` (contract: ${current.verificationContract})` : ""}`);
+      lines.push(`Current task: ${truncateText(current.id, 80)} — ${truncateText(current.title, 180)}${current.verificationContract ? ` (contract: ${truncateText(current.verificationContract, 600)})` : ""}`);
+      if (current.id.length > 80 || current.title.length > 180 || (current.verificationContract?.length ?? 0) > 600) lines.push('Retrieve omitted task requirements with get_goal(section="tasks") before continuing.');
     }
   }
   // Bounded recent-event tail.

package/extensions/goal-completion.ts

@@ -1,5 +1,5 @@
 import { type AgentToolResult, type ExtensionContext } from "@earendil-works/pi-coding-agent";
-import { GOAL_AUDIT_ENTRY, detailedSummary, goalDetails, type GoalAuditEventDetails } from "./goal-format.ts";
+import { GOAL_AUDIT_ENTRY, detailedSummary, goalDetails } from "./goal-format.ts";
 import {
 	buildCompletionReport,
 	buildTaskSummary,
@@ -9,7 +9,7 @@ import {
 import { loadGoalSettings, loadGoalSettingsFileConfig } from "./goal-settings.ts";
 import { runGoalCompletionAuditor } from "./goal-auditor.ts";
 import { nowIso, type GoalRecord } from "./goal-record.ts";
-import { latestEventsForGoal, readGoalLedger } from "./goal-ledger.ts";
+import { latestEventsForGoal, goalRuntimeEvents } from "./goal-ledger.ts";
 import { mergeGoalPromptFromDisk } from "./storage/goal-files.ts";
 import { showEscapeDialog, type EscapeDialogResult } from "./widgets/goal-escape-dialog.ts";
 import type { GoalCore } from "./goal-state.ts";
@@ -22,7 +22,8 @@ import type { GoalMutationOutcome } from "./goal-service.ts";
 // actual workspace evidence. An optional completion_summary is forwarded as an
 // UNTRUSTED executor claim — never evidence and never an approval bypass.
 export async function runGoalCompletionFlow(core: GoalCore, ctx: ExtensionContext, completionSummary?: string): Promise<AgentToolResult<unknown>> {
-	const { pi } = core;
+ const flushError = core.goalService.flushForAudit(ctx);
+ if (flushError) return {content: [{type: "text", text: flushError}], details: goalDetails(core.state.goal)};
 	core.reconcileFocusedGoalFromDisk(ctx);
 
 	// -- Completion --
@@ -129,7 +130,7 @@ function commitGoalCompletion(core: GoalCore, ctx: ExtensionContext, opts: {
 // records remain readable and honored for compatibility; no model tool or
 // task dialog creates new per-goal bypass state).
 if (auditTarget.skipAuditor) {
-	pi.sendMessage<GoalAuditEventDetails>({
+	core.auditMessages.enqueue(ctx, {
 		customType: GOAL_AUDIT_ENTRY,
 		content: `Goal completed — per-goal auditor disabled.`,
 		display: true,
@@ -159,7 +160,7 @@ if (auditTarget.skipAuditor) {
 // the auditor, records audit_skipped, and proceeds through the normal
 // deferred-completion path. No model-side bypass flag is required.
 if (settings.disabled === true) {
-	pi.sendMessage<GoalAuditEventDetails>({
+	core.auditMessages.enqueue(ctx, {
 		customType: GOAL_AUDIT_ENTRY,
 		content: `Goal completed — auditor disabled in settings.`,
 		display: true,
@@ -186,7 +187,7 @@ if (settings.disabled === true) {
 }
 
 	// Auditor is enabled — run the normal audit flow
-	await pi.sendMessage<GoalAuditEventDetails>({
+	core.auditMessages.enqueue(ctx, {
 		customType: GOAL_AUDIT_ENTRY,
 		content: [
 			"Auditor: I am starting the independent completion audit.",
@@ -195,7 +196,7 @@ if (settings.disabled === true) {
 		].filter((line): line is string => line !== undefined).join("\n"),
 		display: true,
 		details: { phase: "started", goalId: auditTarget.id, auditor: auditorLabel },
-	}, { triggerTurn: false });
+	});
 	if (!core.isFocusedOperationCurrent(completionFocus)) {
 		return core.focusedOperationCancelledResult("Goal completion", completionFocus);
 	}
@@ -239,7 +240,7 @@ if (settings.disabled === true) {
 
 	// P1-6: warm start — seed the auditor with the parent-rendered ledger tail
 	// (recent lifecycle + task evidence) so it does not re-derive session facts.
-	const ledger = readGoalLedger(ctx).events;
+	const ledger = goalRuntimeEvents(ctx, auditTarget.id);
 	const warmTail = latestEventsForGoal(ledger, auditTarget.id, 8);
 	const warmContext = warmTail.length > 0
 		? `Recent goal events (from the shared ledger):\n${warmTail.map((e) => `- ${e.at} ${e.type}${"taskId" in e ? ` (task ${e.taskId})` : ""}${"evidence" in e && e.evidence ? ` evidence: ${e.evidence}` : ""}`).join("\n")}`
@@ -296,7 +297,7 @@ if (settings.disabled === true) {
 
 		if (userChoice === "complete_without_audit") {
 			// ── Mark complete without audit ────────────────────────────
-			pi.sendMessage<GoalAuditEventDetails>({
+			core.auditMessages.enqueue(ctx, {
 				customType: GOAL_AUDIT_ENTRY,
 				content: `Goal completed — user bypassed audit via Escape.`,
 				display: true,
@@ -377,7 +378,7 @@ if (settings.disabled === true) {
 			"",
 			auditor.output || "Auditor produced no approval marker.",
 		].filter((line): line is string => line !== undefined).join("\n");
-		pi.sendMessage<GoalAuditEventDetails>({
+		core.auditMessages.enqueue(ctx, {
 			customType: GOAL_AUDIT_ENTRY,
 			content: rejectionText,
 			display: true,
@@ -394,7 +395,7 @@ if (settings.disabled === true) {
 		"",
 		auditor.output || "Auditor approved completion.",
 	].filter((line): line is string => line !== undefined).join("\n");
-	pi.sendMessage<GoalAuditEventDetails>({
+	core.auditMessages.enqueue(ctx, {
 		customType: GOAL_AUDIT_ENTRY,
 		content: approvalText,
 		display: true,