pi-pier 0.3.0 → 0.4.1

package/src/compact-coordinator.ts

@@ -1,40 +1,34 @@
 /**
- * D100 Online Context Compact Coordinator.
- *
- * Orchestrates:
- *  - Tracking todo completion boundaries (filtered to source === 'tool').
- *  - Evaluating KV-cache economics & window protection on turn_end.
- *  - Guarding against ejecting queued messages (hasPendingMessages check).
- *  - Aborting turn and executing Pi native compaction with custom instructions.
- *  - Feeding remaining tasks forward and injecting continuation turn.
- *  - Persisting state to session entries across branches and restores.
+ * D100 Online Context Compact Coordinator: tracks tool-sourced todo completion boundaries, evaluates
+ * KV-cache economics and window protection on turn_end, aborts the in-flight turn, runs pi's native
+ * compaction, then continues the task. State persists to session entries (branch rebuilds restore it).
  */
 
 import type { ExtensionAPI, ExtensionContext } from '@earendil-works/pi-coding-agent';
-import {
-  decideCompaction,
-  DEFAULT_COMPACTION_ECONOMICS,
-  estimateRemainingRequests,
-  MAX_BOUNDARY_SAMPLES,
-  nativeCompactionFeasible,
-  resolveCacheRatioFromCost,
-  type CompactionDecision,
-} from './compact-economics-core.ts';
+import { decideCompaction, DEFAULT_COMPACTION_ECONOMICS, estimateRemainingRequests, MAX_BOUNDARY_SAMPLES, nativeCompactionFeasible, resolveCacheRatioFromCost, type CompactionDecision } from './compact-economics-core.ts';
 import type { TodoItem } from './todo-core.ts';
 import type { TodoCompletionSource } from './todos-service.ts';
 import type { OnlineContextCompactConfig } from './efficiency-config-core.ts';
-import {
-  appendEfficiencyLog,
-  efficiencyLogPath,
-  resolveSessionRoot,
-} from './efficiency-store.ts';
+import { appendEfficiencyLog, efficiencyLogPath, resolveSessionRoot } from './efficiency-store.ts';
+import { swallow } from './swallow.ts';
 
 export const COMPACT_STATE_CUSTOM_TYPE = 'pi-herdr.efficiency-state';
 export const COMPACTION_CONTINUE_TYPE = 'pi-herdr.compaction-continue';
-export const DEFAULT_KEEP_RECENT_TOKENS = 20_000;
-export const DEFAULT_MEMO_TOKENS = 1_000;
+/** Shown to the model after a successful OCC. Names the abort so it is not retried as a failure. */
+export const COMPACTION_CONTINUE_TEXT =
+  'Online context compaction finished. The preceding abort ("This operation was aborted" / "Operation aborted") was this compaction aborting the in-flight turn on purpose — not a model failure, not a user cancel, and not a reason to retry that call. Active tasks are preserved. Continue the remaining work.';
+/**
+ * Cross-session compaction markers: OCC's abort lands in the child transcript as an assistant
+ * message with stopReason 'error', which a master's poller would read as a settled turn. The
+ * inflight/settled pair gives it a "do not settle" window pane state cannot provide (pane shows idle).
+ */
+export const COMPACTION_INFLIGHT_TYPE = 'pi-herdr.compaction-inflight';
+export const COMPACTION_SETTLED_TYPE = 'pi-herdr.compaction-settled';
+const DEFAULT_KEEP_RECENT_TOKENS = 20_000;
+const DEFAULT_MEMO_TOKENS = 1_000;
 
-export interface CoordinatorState {
+/** Fields persisted in the state marker; `version: 1` gates restores. */
+interface CoordinatorState {
   version: 1;
   epoch: number;
   completedBoundaryRequestCounts: number[];
@@ -45,9 +39,24 @@ export interface CoordinatorState {
   positiveContextDeltaCount: number;
   lastContextTokens: number | null;
   currentBoundaryRequestCount: number;
+  consecutiveCompactionFailures: number;
+  /** P1-3: turn-end compact decisions to skip after a failure (each abort costs an in-flight turn). */
+  compactBackoffTurnEnds: number;
 }
 
-export function initialCoordinatorState(): CoordinatorState {
+const NUMERIC_STATE_FIELDS = [
+  'epoch',
+  'carriedDebtTokens',
+  'cacheDebtRepaymentTokens',
+  'priorCompactionCount',
+  'positiveContextDeltaTotal',
+  'positiveContextDeltaCount',
+  'currentBoundaryRequestCount',
+  'consecutiveCompactionFailures',
+  'compactBackoffTurnEnds',
+] as const satisfies readonly (keyof CoordinatorState)[];
+
+function initialCoordinatorState(): CoordinatorState {
   return {
     version: 1,
     epoch: 0,
@@ -59,32 +68,52 @@ export function initialCoordinatorState(): CoordinatorState {
     positiveContextDeltaCount: 0,
     lastContextTokens: null,
     currentBoundaryRequestCount: 0,
+    consecutiveCompactionFailures: 0,
+    compactBackoffTurnEnds: 0,
   };
 }
 
 export function restoreCoordinatorState(entries: readonly unknown[]): CoordinatorState {
-  if (!Array.isArray(entries)) return initialCoordinatorState();
+  const base = initialCoordinatorState();
+  if (!Array.isArray(entries)) return base;
   for (let i = entries.length - 1; i >= 0; i--) {
     const entry = entries[i] as { type?: string; customType?: string; data?: unknown };
-    if (entry && entry.type === 'custom' && entry.customType === COMPACT_STATE_CUSTOM_TYPE) {
-      const d = entry.data as Partial<CoordinatorState>;
-      if (d && d.version === 1 && Array.isArray(d.completedBoundaryRequestCounts)) {
-        return {
-          version: 1,
-          epoch: typeof d.epoch === 'number' ? d.epoch : 0,
-          completedBoundaryRequestCounts: [...d.completedBoundaryRequestCounts],
-          carriedDebtTokens: typeof d.carriedDebtTokens === 'number' ? d.carriedDebtTokens : 0,
-          cacheDebtRepaymentTokens: typeof d.cacheDebtRepaymentTokens === 'number' ? d.cacheDebtRepaymentTokens : 0,
-          priorCompactionCount: typeof d.priorCompactionCount === 'number' ? d.priorCompactionCount : 0,
-          positiveContextDeltaTotal: typeof d.positiveContextDeltaTotal === 'number' ? d.positiveContextDeltaTotal : 0,
-          positiveContextDeltaCount: typeof d.positiveContextDeltaCount === 'number' ? d.positiveContextDeltaCount : 0,
-          lastContextTokens: typeof d.lastContextTokens === 'number' ? d.lastContextTokens : null,
-          currentBoundaryRequestCount: typeof d.currentBoundaryRequestCount === 'number' ? d.currentBoundaryRequestCount : 0,
-        };
-      }
+    if (entry?.type !== 'custom' || entry.customType !== COMPACT_STATE_CUSTOM_TYPE) continue;
+    const d = entry.data as Partial<CoordinatorState> | null | undefined;
+    if (!d || d.version !== 1 || !Array.isArray(d.completedBoundaryRequestCounts)) continue;
+    for (const field of NUMERIC_STATE_FIELDS) {
+      const value = d[field];
+      if (typeof value === 'number') base[field] = value;
     }
+    if (typeof d.lastContextTokens === 'number') base.lastContextTokens = d.lastContextTokens;
+    base.completedBoundaryRequestCounts = [...d.completedBoundaryRequestCounts];
+    return base;
   }
-  return initialCoordinatorState();
+  return base;
+}
+
+function averageContextIncrement(state: CoordinatorState): number | null {
+  return state.positiveContextDeltaCount > 0
+    ? state.positiveContextDeltaTotal / state.positiveContextDeltaCount
+    : null;
+}
+
+function remainingTodos(todos: readonly TodoItem[]): readonly TodoItem[] {
+  return todos.filter((it) => it.status !== 'completed' && it.status !== 'abandoned');
+}
+
+/** P1-3: bound the payload — a huge open-task list inflates the summary request itself. */
+function compactionInstructions(todos: readonly TodoItem[]): string {
+  const remaining = remainingTodos(todos);
+  const taskLines = remaining
+    .slice(0, 25)
+    .map((it) => `- [${it.status}] ${it.content.slice(0, 160)}${it.blocker ? ` (waiting on: ${it.blocker.slice(0, 80)})` : ''}`)
+    .join('\n');
+  const elided = remaining.length > 25 ? `\n… (+${remaining.length - 25} more, elided)` : '';
+  return [
+    'Preserve completed work, verification results, important decisions, and remaining work.',
+    remaining.length > 0 ? `Active/remaining tasks to preserve:\n${taskLines}${elided}` : 'All listed tasks are completed.',
+  ].join('\n\n');
 }
 
 export class CompactCoordinator {
@@ -95,19 +124,18 @@ export class CompactCoordinator {
   pendingBoundaryCompleted = false;
 
   recordBoundaryCompleted(count: number, source?: TodoCompletionSource): void {
-    // Only count completed transitions originating from model tool calls
+    // Only model tool calls count, and the sample is the requests spent reaching this boundary.
     if (source !== 'tool' || count <= 0) return;
     this.pendingBoundaryCompleted = true;
-    if (this.state.currentBoundaryRequestCount > 0) {
-      this.state.completedBoundaryRequestCounts.push(this.state.currentBoundaryRequestCount);
-      if (this.state.completedBoundaryRequestCounts.length > MAX_BOUNDARY_SAMPLES) {
-        this.state.completedBoundaryRequestCounts.splice(
-          0,
-          this.state.completedBoundaryRequestCounts.length - MAX_BOUNDARY_SAMPLES,
-        );
-      }
-      this.state.currentBoundaryRequestCount = 0;
+    if (this.state.currentBoundaryRequestCount <= 0) return;
+    this.state.completedBoundaryRequestCounts.push(this.state.currentBoundaryRequestCount);
+    if (this.state.completedBoundaryRequestCounts.length > MAX_BOUNDARY_SAMPLES) {
+      this.state.completedBoundaryRequestCounts.splice(
+        0,
+        this.state.completedBoundaryRequestCounts.length - MAX_BOUNDARY_SAMPLES,
+      );
     }
+    this.state.currentBoundaryRequestCount = 0;
   }
 
   onBeforeProviderRequest(contextTokens: number): void {
@@ -117,9 +145,7 @@ export class CompactCoordinator {
         0,
         this.state.carriedDebtTokens - this.state.cacheDebtRepaymentTokens,
       );
-      if (this.state.carriedDebtTokens === 0) {
-        this.state.cacheDebtRepaymentTokens = 0;
-      }
+      if (this.state.carriedDebtTokens === 0) this.state.cacheDebtRepaymentTokens = 0;
     }
     if (this.state.lastContextTokens !== null && contextTokens > this.state.lastContextTokens) {
       this.state.positiveContextDeltaTotal += contextTokens - this.state.lastContextTokens;
@@ -129,29 +155,22 @@ export class CompactCoordinator {
   }
 
   onInput(event: { text?: string; source?: string; streamingBehavior?: string }): void {
-    // Ignore internal messages dispatched by extensions (D96 reminders, pipe injections, followUps)
-    if (event.source === 'extension' || event.text?.startsWith('CORRECTION:')) {
-      return;
-    }
+    // Ignore messages dispatched by extensions (D96 reminders, pipe injections, followUps).
+    if (event.source === 'extension' || event.text?.startsWith('CORRECTION:')) return;
 
     const fromHuman =
       event.source === 'interactive' ||
       event.source === 'rpc' ||
       event.source === undefined;
+    if (!fromHuman) return;
 
-    // If human typed a new prompt or steer directive, reset expectations (correction)
-    if (fromHuman) {
-      this.selectedCompaction = null;
-      this.intentionalAbort = false;
-      this.pendingBoundaryCompleted = false;
-      // completedBoundaryRequestCounts deliberately survives: it measures
-      // requests-per-boundary pacing, which holds across human turns. Clearing it
-      // zeroed the mean and pinned expectedRemainingRequests at 1 after every
-      // prompt, so OCC never saw a horizon worth compacting under (2026-09-17
-      // review: compact.jsonl epoch kept climbing 1→9 with zero compactions).
-      this.state.carriedDebtTokens = 0;
-      this.state.epoch++;
-    }
+    this.selectedCompaction = null;
+    this.intentionalAbort = false;
+    this.pendingBoundaryCompleted = false;
+    // completedBoundaryRequestCounts deliberately survives: it measures requests-per-boundary pacing,
+    // which holds across human turns — clearing it pinned expectedRemainingRequests at 1 every prompt.
+    this.state.carriedDebtTokens = 0;
+    this.state.epoch++;
   }
 
   onTurnEnd(opts: {
@@ -165,8 +184,7 @@ export class CompactCoordinator {
 
     if (!boundary || this.selectedCompaction) return;
     if (!opts.config.enabled) return;
-
-    // Safety guard: if user typed something that is queued, don't abort!
+    // Safety guard: never abort while the user has a queued message.
     if (opts.ctx.hasPendingMessages?.()) return;
 
     const usage = opts.ctx.getContextUsage?.();
@@ -175,22 +193,10 @@ export class CompactCoordinator {
         ? usage.tokens
         : this.state.lastContextTokens ?? 0;
     const contextWindowTokens = usage?.contextWindow ?? opts.ctx.model?.contextWindow ?? null;
+    // System prompt is re-sent verbatim, so it is never part of the compressible archive; ~4 chars/token.
     const fixedTokens = Math.ceil(Buffer.byteLength(opts.ctx.getSystemPrompt?.() ?? '') / 4);
     const keepRecentTokens = opts.config.keepRecentTokens ?? DEFAULT_KEEP_RECENT_TOKENS;
     const archiveTokens = Math.max(0, contextTokens - fixedTokens - keepRecentTokens);
-    const remainingBoundaries = opts.todos.filter(
-      (it) => it.status !== 'completed' && it.status !== 'abandoned',
-    ).length;
-
-    const averageContextTokenIncrement =
-      this.state.positiveContextDeltaCount > 0
-        ? this.state.positiveContextDeltaTotal / this.state.positiveContextDeltaCount
-        : null;
-    const cacheRatio = resolveCacheRatioFromCost(
-      opts.config.cacheWriteReadRatio,
-      opts.ctx.model?.cost,
-      { provider: opts.ctx.model?.provider, modelId: opts.ctx.model?.id },
-    );
 
     const decision = decideCompaction({
       writeTokens: contextTokens,
@@ -198,13 +204,17 @@ export class CompactCoordinator {
       memoTokens: DEFAULT_MEMO_TOKENS,
       contextTokens,
       completedBoundaryRequestCounts: this.state.completedBoundaryRequestCounts,
-      remainingBoundaries,
-      averageContextTokenIncrement,
+      remainingBoundaries: remainingTodos(opts.todos).length,
+      averageContextTokenIncrement: averageContextIncrement(this.state),
       contextWindowTokens,
       priorCompactionCount: this.state.priorCompactionCount,
       carriedDebtTokens: this.state.carriedDebtTokens,
       cacheDebtRepaymentTokens: this.state.cacheDebtRepaymentTokens,
-      cacheWriteReadRatio: cacheRatio,
+      cacheWriteReadRatio: resolveCacheRatioFromCost(
+        opts.config.cacheWriteReadRatio,
+        opts.ctx.model?.cost,
+        { provider: opts.ctx.model?.provider, modelId: opts.ctx.model?.id },
+      ),
       economics: {
         ...DEFAULT_COMPACTION_ECONOMICS,
         firstCompactionRequestScale: opts.config.firstCompactionRequestScale,
@@ -212,24 +222,34 @@ export class CompactCoordinator {
       },
     });
 
-    if (decision.compact) {
-      const branch = opts.ctx.sessionManager?.getBranch?.() ?? [];
-      const keepRecentTokens = opts.config.keepRecentTokens ?? DEFAULT_KEEP_RECENT_TOKENS;
-      const feasible = nativeCompactionFeasible(branch, keepRecentTokens);
-      if (!feasible) {
-        if (opts.config.logEnabled) {
-          this.logDecision(opts.ctx, { ...decision, compact: false, reason: 'native_not_compactable' });
-        }
-        return;
+    if (!decision.compact) {
+      if (opts.config.logEnabled) this.logDecision(opts.ctx, decision);
+      return;
+    }
+
+    // P1-3 backoff: after a failed summarization, skip compact decisions for a bounded window —
+    // each attempt aborts an in-flight turn, so a doomed retry burns turns. (The token-cap failure
+    // class is gone since pi 1.4.0 uncapped summary output; transient timeouts/aborts remain.)
+    if (this.state.compactBackoffTurnEnds > 0) {
+      this.state.compactBackoffTurnEnds--;
+      if (opts.config.logEnabled) {
+        this.logDecision(opts.ctx, { ...decision, compact: false, reason: 'failure_backoff' });
       }
+      return;
+    }
 
-      this.selectedCompaction = decision;
-      this.intentionalAbort = true;
-      opts.cancelReminder?.();
-      opts.ctx.abort();
-    } else if (opts.config.logEnabled) {
-      this.logDecision(opts.ctx, decision);
+    const branch = opts.ctx.sessionManager?.getBranch?.() ?? [];
+    if (!nativeCompactionFeasible(branch, keepRecentTokens)) {
+      if (opts.config.logEnabled) {
+        this.logDecision(opts.ctx, { ...decision, compact: false, reason: 'native_not_compactable' });
+      }
+      return;
     }
+
+    this.selectedCompaction = decision;
+    this.intentionalAbort = true;
+    opts.cancelReminder?.();
+    opts.ctx.abort();
   }
 
   async onAgentSettled(opts: {
@@ -241,159 +261,157 @@ export class CompactCoordinator {
     onBeforeCompact?: (sessionRoot: string, ctx: ExtensionContext) => Promise<void>;
   }): Promise<void> {
     if (!this.selectedCompaction || !opts.ctx.isIdle()) return;
+
+    // Announce the compaction window before the summary request starts: the supervising master
+    // polls this file, and its observation window is shorter than a large-context compaction.
+    this.appendMarker(opts.pi, COMPACTION_INFLIGHT_TYPE, { at: Date.now() });
     const decision = this.selectedCompaction;
     this.selectedCompaction = null;
 
     opts.cancelReminder?.();
     this.compactionInFlight = true;
 
-    const sessionDir = opts.ctx.sessionManager?.getSessionDir?.();
-    const sessionId = opts.ctx.sessionManager?.getSessionId?.();
-    const sessionRoot = resolveSessionRoot(sessionDir, sessionId);
-
+    const sessionRoot = resolveSessionRoot(
+      opts.ctx.sessionManager?.getSessionDir?.(),
+      opts.ctx.sessionManager?.getSessionId?.(),
+    );
     if (sessionRoot && opts.onBeforeCompact) {
       try {
         await opts.onBeforeCompact(sessionRoot, opts.ctx);
       } catch {
-        /* ignore batch packing errors before compaction */
       }
     }
 
-    const remaining = opts.todos.filter(
-      (it) => it.status !== 'completed' && it.status !== 'abandoned',
-    );
-    const taskLines = remaining
-      .map((it) => `- [${it.status}] ${it.content}${it.blocker ? ` (waiting on: ${it.blocker})` : ''}`)
-      .join('\n');
-
-    const customInstructions = [
-      'Preserve completed work, verification results, important decisions, and remaining work.',
-      remaining.length > 0 ? `Active/remaining tasks to preserve:\n${taskLines}` : 'All listed tasks are completed.',
-    ].join('\n\n');
-
     const startMs = Date.now();
-    await new Promise<void>((resolve) => {
-      let settled = false;
-      const finish = () => {
-        if (!settled) {
-          settled = true;
-          resolve();
+    const settled = Promise.withResolvers<void>();
+    opts.ctx.compact({
+      customInstructions: compactionInstructions(opts.todos),
+      onComplete: (compaction) => {
+        try {
+          this.finishCompaction(opts, decision, startMs, compaction);
+        } finally {
+          settled.resolve();
         }
-      };
-
-      opts.ctx.compact({
-        customInstructions,
-        onComplete: (compaction) => {
-          try {
-            this.state.priorCompactionCount++;
-            const incrementalRatio = decision.incrementalCacheCostRatio ?? 0;
-            this.state.carriedDebtTokens = decision.writeTokens * incrementalRatio;
-            this.state.cacheDebtRepaymentTokens = Math.max(0, decision.archiveTokens - decision.memoTokens);
-            this.state.epoch++;
-
-            opts.pi.appendEntry(COMPACT_STATE_CUSTOM_TYPE, this.state);
-
-            if (opts.config.logEnabled) {
-              const sessionDir = opts.ctx.sessionManager?.getSessionDir?.();
-              const sessionId = opts.ctx.sessionManager?.getSessionId?.();
-              const root = resolveSessionRoot(sessionDir, sessionId);
-              if (root) {
-                const logPath = efficiencyLogPath(root, 'compact');
-                void appendEfficiencyLog(logPath, {
-                  schema: 'pier-efficiency/1',
-                  mechanism: 'onlineContextCompact',
-                  ts: new Date().toISOString(),
-                  sessionId: sessionId ?? 'unknown',
-                  epoch: this.state.epoch,
-                  decision: decision.reason,
-                  writeTokens: decision.writeTokens,
-                  archiveTokens: decision.archiveTokens,
-                  savedTokens: Math.max(0, decision.archiveTokens - decision.memoTokens),
-                  breakevenRequests: decision.breakevenRequests,
-                  expectedRemainingRequests: decision.expectedRemainingRequests,
-                  carriedDebtTokens: this.state.carriedDebtTokens,
-                  summaryTokens: typeof compaction?.summary === 'string' ? Math.ceil(compaction.summary.length / 4) : 0,
-                  priorCompactionCount: this.state.priorCompactionCount,
-                  durationMs: Date.now() - startMs,
-                }).catch(() => {});
-              }
-            }
-
-            try {
-              opts.pi.sendMessage(
-                {
-                  customType: COMPACTION_CONTINUE_TYPE,
-                  content: 'Online context compaction finished. Active tasks preserved. Continue working on remaining tasks.',
-                  display: false,
-                },
-                { triggerTurn: true },
-              );
-            } catch {
-              /* ignore continuation failure */
-            }
-          } finally {
-            this.compactionInFlight = false;
-            this.intentionalAbort = false;
-            finish();
-          }
-        },
-        onError: (error) => {
-          this.compactionInFlight = false;
-          this.intentionalAbort = false;
-
-          // `Compaction cancelled` / AbortError mean the user (or pi) cancelled on purpose.
-          // Telemetry must keep those apart from real failures, and a cancelled compaction
-          // must NOT be answered with a turn-triggering continuation.
-          const cancelled =
-            error instanceof Error &&
-            (error.name === 'AbortError' || error.message === 'Compaction cancelled');
-
-          if (opts.config.logEnabled) {
-            const sessionDir = opts.ctx.sessionManager?.getSessionDir?.();
-            const sessionId = opts.ctx.sessionManager?.getSessionId?.();
-            const root = resolveSessionRoot(sessionDir, sessionId);
-            if (root) {
-              const logPath = efficiencyLogPath(root, 'compact');
-              void appendEfficiencyLog(logPath, {
-                schema: 'pier-efficiency/1',
-                mechanism: 'onlineContextCompact',
-                event: 'compaction-failed',
-                ts: new Date().toISOString(),
-                sessionId: sessionId ?? 'unknown',
-                epoch: this.state.epoch,
-                decision: decision.reason,
-                cancelled,
-                error: error instanceof Error ? error.message : String(error),
-                durationMs: Date.now() - startMs,
-              }).catch(() => {});
-            }
-          }
-
-          if (cancelled) {
-            finish();
-            return;
-          }
-
-          // Resume the task: OCC aborted the turn on purpose, so a failed compaction would
-          // otherwise leave the session parked on an aborted assistant message. Pi's own
-          // threshold compaction stays available as the safety net.
-          try {
-            opts.pi.sendMessage(
-              {
-                customType: COMPACTION_CONTINUE_TYPE,
-                content: 'Context compaction failed. Continuing with the current context.',
-                display: false,
-              },
-              { triggerTurn: true },
-            );
-          } catch {
-            /* ignore continuation failure */
-          }
-
-          finish();
-        },
+      },
+      onError: (error) => {
+        try {
+          this.failCompaction(opts, decision, startMs, error);
+        } finally {
+          settled.resolve();
+        }
+      },
+    });
+    await settled.promise;
+  }
+
+  private finishCompaction(
+    opts: { ctx: ExtensionContext; pi: ExtensionAPI; config: OnlineContextCompactConfig },
+    decision: CompactionDecision,
+    startMs: number,
+    compaction: { summary?: string } | undefined,
+  ): void {
+    this.state.priorCompactionCount++;
+    this.state.carriedDebtTokens = decision.writeTokens * (decision.incrementalCacheCostRatio ?? 0);
+    this.state.cacheDebtRepaymentTokens = Math.max(0, decision.archiveTokens - decision.memoTokens);
+    this.state.epoch++;
+    this.state.consecutiveCompactionFailures = 0;
+    this.state.compactBackoffTurnEnds = 0;
+    this.appendMarker(opts.pi, COMPACT_STATE_CUSTOM_TYPE, this.state);
+
+    if (opts.config.logEnabled) {
+      this.logCompact(opts.ctx, {
+        decision: decision.reason,
+        writeTokens: decision.writeTokens,
+        archiveTokens: decision.archiveTokens,
+        savedTokens: Math.max(0, decision.archiveTokens - decision.memoTokens),
+        breakevenRequests: decision.breakevenRequests,
+        expectedRemainingRequests: decision.expectedRemainingRequests,
+        carriedDebtTokens: this.state.carriedDebtTokens,
+        summaryTokens: typeof compaction?.summary === 'string' ? Math.ceil(compaction.summary.length / 4) : 0,
+        priorCompactionCount: this.state.priorCompactionCount,
+        durationMs: Date.now() - startMs,
       });
+    }
+
+    this.sendContinuation(
+      opts.pi,
+      COMPACTION_CONTINUE_TEXT,
+      false,
+    );
+    // Settled marker last: the master's poller releases its hold as soon as it appears.
+    this.appendMarker(opts.pi, COMPACTION_SETTLED_TYPE, { outcome: 'completed', at: Date.now() });
+    this.compactionInFlight = false;
+    this.intentionalAbort = false;
+  }
+
+  private failCompaction(
+    opts: { ctx: ExtensionContext; pi: ExtensionAPI; config: OnlineContextCompactConfig },
+    decision: CompactionDecision,
+    startMs: number,
+    error: unknown,
+  ): void {
+    this.compactionInFlight = false;
+    this.intentionalAbort = false;
+    const message = error instanceof Error ? error.message : String(error);
+    // 'Compaction cancelled' / AbortError mean the user (or pi) cancelled on purpose: telemetry keeps
+    // that apart from real failures, and a cancel must NOT trigger a continuation turn.
+    const cancelled =
+      error instanceof Error &&
+      (error.name === 'AbortError' || error.message === 'Compaction cancelled');
+
+    // Release the cross-session hold in every outcome — the cancel path returns before continuing.
+    this.appendMarker(opts.pi, COMPACTION_SETTLED_TYPE, {
+      outcome: cancelled ? 'cancelled' : 'failed',
+      at: Date.now(),
     });
+    if (opts.config.logEnabled) {
+      this.logCompact(opts.ctx, {
+        event: 'compaction-failed',
+        decision: decision.reason,
+        cancelled,
+        error: message,
+        consecutiveCompactionFailures: this.state.consecutiveCompactionFailures,
+        durationMs: Date.now() - startMs,
+      });
+    }
+    if (cancelled) return;
+
+    // P1-3: record the failure and back off exponentially (cap 4 skipped decisions).
+    this.state.consecutiveCompactionFailures++;
+    this.state.compactBackoffTurnEnds = Math.min(2 ** this.state.consecutiveCompactionFailures, 4);
+    // Persist the bumped counters: state is restored from this marker, so a restart without it
+    // would reset the P1-3 backoff window and retry the doomed compaction immediately.
+    this.appendMarker(opts.pi, COMPACT_STATE_CUSTOM_TYPE, this.state);
+
+    // OCC aborted the turn on purpose, so a failed compaction would leave the session parked on an
+    // aborted message; the notice must be visible rather than a silent failure.
+    const approxTokens = this.state.lastContextTokens;
+    this.sendContinuation(
+      opts.pi,
+      `Context compaction FAILED (${message}). `
+        + `Context ≈ ${approxTokens != null ? `${Math.round(approxTokens / 1000)}K tokens` : 'near the window limit'}; `
+        + `OCC backs off for the next ${this.state.compactBackoffTurnEnds} opportunities. `
+        + 'If this repeats, run /compact manually, prune large outputs, or restart the session.',
+      true,
+    );
+  }
+
+  /** Best-effort session entry write: onAgentSettled runs detached, so a full disk must not become an
+   *  unhandled rejection (a missing marker only delays one settlement notice). */
+  private appendMarker(pi: ExtensionAPI, customType: string, data: unknown): void {
+    try {
+      pi.appendEntry(customType, data);
+    } catch (err) {
+      swallow(`compact.append-${customType}`, err);
+    }
+  }
+
+  /** Resume the task after compaction; must never throw into the detached settled handler. */
+  private sendContinuation(pi: ExtensionAPI, content: string, display: boolean): void {
+    try {
+      pi.sendMessage({ customType: COMPACTION_CONTINUE_TYPE, content, display }, { triggerTurn: true });
+    } catch {
+    }
   }
 
   rebuildFromBranch(entries: readonly unknown[]): void {
@@ -406,49 +424,49 @@ export class CompactCoordinator {
 
   getRemainingHorizon(remainingBoundaries = 3, contextWindowTokens: number | null = null): number {
     const counts = this.state.completedBoundaryRequestCounts;
-    if (counts.length === 0) return 4;
-
+    if (counts.length === 0) return 4; // No samples yet: conservative default horizon.
     const est = estimateRemainingRequests({
       completedBoundaryRequestCounts: counts,
       remainingBoundaries,
-      scale: 1.0,
-      standardDeviationK: 0.0,
+      scale: DEFAULT_COMPACTION_ECONOMICS.remainingRequestScale,
+      standardDeviationK: DEFAULT_COMPACTION_ECONOMICS.remainingRequestStddevK,
       contextTokens: this.state.lastContextTokens ?? 0,
       contextWindowTokens,
-      averageContextTokenIncrement:
-        this.state.positiveContextDeltaCount > 0
-          ? this.state.positiveContextDeltaTotal / this.state.positiveContextDeltaCount
-          : null,
+      averageContextTokenIncrement: averageContextIncrement(this.state),
     });
     return Math.max(1, est.expectedRemainingRequests);
   }
 
-  private logDecision(ctx: ExtensionContext, decision: CompactionDecision): void {
-    const sessionDir = ctx.sessionManager?.getSessionDir?.();
+  /** Append one compact.jsonl record (shared envelope); logging is always best-effort. */
+  private logCompact(ctx: ExtensionContext, fields: Record<string, unknown>): void {
     const sessionId = ctx.sessionManager?.getSessionId?.();
-    const root = resolveSessionRoot(sessionDir, sessionId);
+    const root = resolveSessionRoot(ctx.sessionManager?.getSessionDir?.(), sessionId);
     if (!root) return;
-    const logPath = efficiencyLogPath(root, 'compact');
-    void appendEfficiencyLog(logPath, {
+    void appendEfficiencyLog(efficiencyLogPath(root, 'compact'), {
       schema: 'pier-efficiency/1',
       mechanism: 'onlineContextCompact',
-      event: 'decision',
       ts: new Date().toISOString(),
       sessionId: sessionId ?? 'unknown',
       epoch: this.state.epoch,
+      ...fields,
+    }).catch(() => {});
+  }
+
+  private logDecision(ctx: ExtensionContext, decision: CompactionDecision): void {
+    this.logCompact(ctx, {
+      event: 'decision',
       decision: decision.reason,
       compact: decision.compact,
       writeTokens: decision.writeTokens,
       archiveTokens: decision.archiveTokens,
       breakevenRequests: decision.breakevenRequests,
       expectedRemainingRequests: decision.expectedRemainingRequests,
-      // 2026-09-17: without resolvedRatio/model the 61x cache_ratio_unavailable
-      // run was un-diagnosable from logs alone (which price table? which family?).
+      // Without resolvedRatio/model a cache_ratio_unavailable run is undiagnosable from logs alone.
       resolvedRatio: decision.cacheWriteReadRatio,
       incrementalCacheCostRatio: decision.incrementalCacheCostRatio,
       remainingBoundaries: decision.remainingBoundaries,
       model: ctx.model?.id ?? undefined,
       provider: ctx.model?.provider ?? undefined,
-    }).catch(() => {});
+    });
   }
 }