@remnic/core 10.0.60 → 10.0.62

package/src/storage/buffer-drain-claims.ts

@@ -0,0 +1,593 @@
+/**
+ * Durable drain claims for the extraction buffer (`remnic buffer drain`).
+ *
+ * A claim is the atomic handover of one session's buffered turns from the
+ * live buffer (daemon or CLI process) to a drain worker, so extraction can
+ * run out of process without a live buffer owner resurrecting or losing the
+ * turns. Files under `<stateDir>/buffer/drain/`:
+ *
+ *   s-<sha256(key)>.json        claim content: { key, entry, claimedAt }
+ *                               — the turns NOT yet extracted. Renamed into
+ *                               place from the session file (atomic on one
+ *                               filesystem), rewritten atomically as chunks
+ *                               drain.
+ *   s-<sha256(key)>.lease.json  { ownerId, pid, acquiredAt, heartbeatAt,
+ *                               leaseMs } — created with exclusive-create so
+ *                               exactly one drainer holds a key.
+ *
+ * Invariants (see the buffer-drain PR for the full protocol):
+ *   - At every instant the claim content file holds exactly the turns whose
+ *     extraction is not durably finished, so a kill at any point loses
+ *     nothing and the next run resumes.
+ *   - A live buffer owner that must persist a claimed session hands its
+ *     in-memory entry INTO the claim (merge, identity-deduped) instead of
+ *     writing a session file — it can never resurrect claimed turns.
+ *   - Claims are idempotent: re-claiming, double-draining, and two
+ *     concurrent drainers are all safe (per-key mutation lock + exclusive
+ *     lease; the lock also fences lease takeover and progress writes).
+ *
+ * All content reads/writes go through the storage's secure file I/O
+ * (`BufferFileIo`); only lease/existence checks touch raw paths, and claim
+ * rewrites serialize on a cross-process held file lock.
+ */
+
+import { hostname } from "node:os";
+import { z } from "zod";
+import { existsSync } from "node:fs";
+import { mkdir, readdir, rename, unlink, writeFile } from "node:fs/promises";
+import path from "node:path";
+
+import { bufferSessionFileName, type BufferFileIo } from "./buffer-session-files.js";
+import { bufferTurnsEqual, subtractBufferTurns } from "../buffer-turn-helpers.js";
+import { log } from "../logger.js";
+import { withHeldFileLock } from "../utils/serialize-mutations.js";
+import type { BufferEntryState, BufferTurn } from "../types.js";
+
+/** Result of persisting one session's entry. */
+export interface BufferSessionSaveOutcome {
+  /** "saved": the session file holds the entry. "handed_over": the entry moved into a drain claim. */
+  status: "saved" | "handed_over";
+  /** Turns a drain already extracted (claim tombstone): the owner subtracts them from its copy. */
+  drainedTurns?: BufferTurn[];
+}
+
+export interface DrainLease {
+  ownerId: string;
+  pid: number;
+  /** Host the owner runs on: a takeover may only judge liveness on that host. */
+  host: string;
+  acquiredAt: string;
+  heartbeatAt: string;
+  leaseMs: number;
+}
+
+export interface DrainClaimFile {
+  key: string;
+  entry: BufferEntryState;
+  claimedAt: string;
+  /**
+   * Turns the drain already extracted. Non-empty means the session was
+   * drained and the claim is a TOMBSTONE: a live buffer owner subtracts
+   * these from its (possibly stale) copy instead of writing them back, so
+   * drained turns can never be resurrected.
+   */
+  drained?: BufferTurn[];
+  /** Same tombstone for turns the drain extracted from `retainedTurns`. */
+  drainedRetained?: BufferTurn[];
+}
+
+export type ClaimFromSessionOutcome = "claimed" | "empty" | "already_claimed";
+export type AdoptOutcome = "adopted" | "busy";
+export type AckOutcome = "acked" | "not_empty" | "busy";
+
+export interface DrainClaimListing {
+  key: string;
+  turnCount: number;
+  leased: boolean;
+  leaseExpired: boolean;
+}
+
+/** Lock while a claim is read-modify-written; released on settle. */
+const CLAIM_LOCK_STALE_MS = 120_000;
+const CLAIM_LOCK_MAX_WAIT_MS = 120_000;
+
+function emptyEntry(): BufferEntryState {
+  return { turns: [], lastExtractionAt: null, extractionCount: 0 };
+}
+
+/** Identity-dedupe `incoming` turns into `existing` (they are the same turns a
+ * live owner still held after claiming; only genuinely NEW turns append). */
+function mergeTurns<T>(existing: T[], incoming: readonly T[], isEqual: (a: T, b: T) => boolean): T[] {
+  const merged = [...existing];
+  for (const turn of incoming) {
+    if (merged.some((kept) => isEqual(kept, turn))) continue;
+    merged.push(turn);
+  }
+  return merged;
+}
+
+function mergeEntries(claim: DrainClaimFile, incoming: BufferEntryState): DrainClaimFile {
+  const entry = emptyEntry();
+  // Turns already extracted stay out of the pending entry even when the
+  // owner's incoming copy is stale and still carries them (either tombstone
+  // list counts — a turn in either one left the claim for good).
+  const gone = [...(claim.drained ?? []), ...(claim.drainedRetained ?? [])];
+  entry.turns = mergeTurns(
+    subtractBufferTurns(claim.entry.turns, gone),
+    subtractBufferTurns(incoming.turns, gone),
+    bufferTurnsEqual,
+  );
+  const retained = [
+    ...subtractBufferTurns(claim.entry.retainedTurns ?? [], gone),
+    ...subtractBufferTurns(incoming.retainedTurns ?? [], gone),
+  ];
+  // SAFETY: an empty accumulator of the same element type.
+  const dedupedRetained = mergeTurns([] as typeof retained, retained, bufferTurnsEqual);
+  if (dedupedRetained.length > 0) entry.retainedTurns = dedupedRetained;
+  const stamps = [claim.entry.lastExtractionAt, incoming.lastExtractionAt].filter(
+    (value): value is string => value !== null && value !== undefined,
+  );
+  entry.lastExtractionAt = stamps.sort().at(-1) ?? null;
+  entry.extractionCount = Math.max(claim.entry.extractionCount, incoming.extractionCount);
+  return { ...claim, entry };
+}
+
+/**
+ * Is the lease owner still running? Only decidable on the owner's host; a
+ * lease from another host counts as alive (its expiry decides).
+ */
+function isOwnerAlive(lease: DrainLease): boolean {
+  if (lease.host !== hostname()) return true;
+  if (!Number.isInteger(lease.pid) || lease.pid <= 0) return true;
+  try {
+    process.kill(lease.pid, 0);
+    return true;
+  } catch (err) {
+    // EPERM: the process exists but belongs to another user.
+    // SAFETY: the value comes from this module's own write path; the shape is checked above.
+    return (err as NodeJS.ErrnoException)?.code === "EPERM";
+  }
+}
+
+function isLeaseExpired(lease: DrainLease, now = Date.now()): boolean {
+  const heartbeat = Date.parse(lease.heartbeatAt);
+  return !Number.isFinite(heartbeat) || now - heartbeat > lease.leaseMs;
+}
+
+const claimBodySchema = z.object({
+  key: z.string().min(1),
+  entry: z.object({
+    turns: z.array(z.unknown()),
+    retainedTurns: z.array(z.unknown()).optional(),
+    lastExtractionAt: z.string().nullable().optional(),
+    extractionCount: z.number().optional(),
+  }).optional(),
+  turns: z.array(z.unknown()).optional(),
+  retainedTurns: z.array(z.unknown()).optional(),
+  lastExtractionAt: z.string().nullable().optional(),
+  extractionCount: z.number().optional(),
+  claimedAt: z.string().optional(),
+  drained: z.array(z.unknown()).optional(),
+  drainedRetained: z.array(z.unknown()).optional(),
+});
+
+function parseBody(raw: string): DrainClaimFile | null {
+  let body: z.infer<typeof claimBodySchema>;
+  try {
+    body = claimBodySchema.parse(JSON.parse(raw));
+  } catch {
+    return null;
+  }
+  // Two on-disk shapes, one format: a claim born from rename keeps the flat
+  // session-file shape ({key, turns, ...}) until its first rewrite, which
+  // nests under `entry`. Accept both.
+  const source = body.entry && Array.isArray(body.entry.turns) ? body.entry : body;
+  if (!Array.isArray(source.turns)) return null;
+  const entry = emptyEntry();
+  // SAFETY: the schema parsed these as an array of unknown; the claim file
+  // is written only by this module (BufferSessionFiles.save), so the entries
+  // are BufferTurn values.
+  // SAFETY: the claim file is written only by this module, so the parsed
+  // turn arrays hold BufferTurn values.
+  // SAFETY: the value comes from this module's own write path; the shape is checked above.
+  entry.turns = source.turns as BufferTurn[];
+  if (source.retainedTurns) {
+    // SAFETY: same boundary as the turns above (this module writes the file).
+    entry.retainedTurns = source.retainedTurns as BufferTurn[];
+  }
+  entry.lastExtractionAt = source.lastExtractionAt ?? null;
+  entry.extractionCount = Number.isFinite(source.extractionCount) ? Number(source.extractionCount) : 0;
+  const claim: DrainClaimFile = {
+    key: body.key,
+    entry,
+    claimedAt: body.claimedAt ?? new Date().toISOString(),
+  };
+  if (Array.isArray(body.drained)) {
+    // SAFETY: same boundary as the turns above (this module writes the file).
+    claim.drained = body.drained as BufferTurn[];
+  }
+  if (Array.isArray(body.drainedRetained)) {
+    // SAFETY: same boundary as the turns above (this module writes the file).
+    claim.drainedRetained = body.drainedRetained as BufferTurn[];
+  }
+  return claim;
+}
+
+const leaseSchema = z.object({
+  ownerId: z.string().min(1),
+  pid: z.number().optional(),
+  host: z.string().optional(),
+  acquiredAt: z.string().optional(),
+  heartbeatAt: z.string(),
+  leaseMs: z.number(),
+});
+
+function parseLease(raw: string): DrainLease | null {
+  let lease: z.infer<typeof leaseSchema>;
+  try {
+    lease = leaseSchema.parse(JSON.parse(raw));
+  } catch {
+    return null;
+  }
+  return {
+    ownerId: lease.ownerId,
+    pid: lease.pid ?? 0,
+    host: lease.host ?? "",
+    acquiredAt: lease.acquiredAt ?? lease.heartbeatAt,
+    heartbeatAt: lease.heartbeatAt,
+    leaseMs: lease.leaseMs,
+  };
+}
+
+export class BufferDrainClaims {
+  constructor(
+    private readonly stateDir: string,
+    private readonly io: BufferFileIo,
+  ) {}
+
+  private get dir(): string {
+    return path.join(this.stateDir, "buffer", "drain");
+  }
+
+  private fileFor(key: string): string {
+    return path.join(this.dir, bufferSessionFileName(key));
+  }
+
+  private leaseFileFor(key: string): string {
+    return `${this.fileFor(key)}.lease.json`;
+  }
+
+  private lockFileFor(key: string): string {
+    return `${this.fileFor(key)}.lock`;
+  }
+
+  /** Sync existence check for hot-path gates (`SmartBuffer.isSessionClaimed`). */
+  isClaimed(key: string): boolean {
+    return existsSync(this.fileFor(key));
+  }
+
+  /** Test + CLI seam: the lease file path for direct inspection. */
+  leasePathForTest(key: string): string {
+    return this.leaseFileFor(key);
+  }
+
+  private async readClaimBody(key: string): Promise<DrainClaimFile | null> {
+    try {
+      return parseBody(await this.io.read(this.fileFor(key)));
+    } catch {
+      return null;
+    }
+  }
+
+  /** The claim content for `key`, or null when the key is not claimed. */
+  load(key: string): Promise<DrainClaimFile | null> {
+    return this.readClaimBody(key);
+  }
+
+  private async writeClaimBody(key: string, claim: DrainClaimFile): Promise<void> {
+    await mkdir(this.dir, { recursive: true });
+    await this.io.write(this.fileFor(key), JSON.stringify({ ...claim, key }));
+  }
+
+  private async readLease(key: string): Promise<DrainLease | null> {
+    try {
+      return parseLease(await this.io.read(this.leaseFileFor(key)));
+    } catch {
+      return null;
+    }
+  }
+
+  private async writeLease(key: string, lease: DrainLease): Promise<void> {
+    await mkdir(this.dir, { recursive: true });
+    await this.io.write(this.leaseFileFor(key), JSON.stringify(lease));
+  }
+
+  /**
+   * Atomically move a session's persisted file into a claim. The rename is
+   * the handover: after it settles the turns exist ONLY as a claim, so no
+   * crash window can lose them or leave them in both places. Callers that
+   * lose the race (`already_claimed`) simply skip the key.
+   *
+   * The whole handover runs under the claim's mutation lock, so a claim
+   * create can never land on top of an existing claim (POSIX rename would
+   * silently replace it) and the lease write cannot race an adopter.
+   */
+  async claimFromSessionFile(key: string, lease: DrainLease): Promise<ClaimFromSessionOutcome> {
+    await mkdir(this.dir, { recursive: true });
+    const sessionPath = path.join(this.stateDir, "buffer", bufferSessionFileName(key));
+    if (existsSync(sessionPath)) {
+      // Advisory pre-read: an empty live session belongs to its owner, not
+      // the drain. A turn landing between this read and the rename is fine —
+      // the rename moves whatever is there and the claim drains it.
+      try {
+        const seeded = parseBody(await this.io.read(sessionPath));
+        if (seeded && seeded.entry.turns.length === 0 && (seeded.entry.retainedTurns?.length ?? 0) === 0) {
+          return "empty";
+        }
+      } catch {
+        // Unreadable live file: proceed with the rename and let the claim
+        // path report it loudly, like the loader skips corrupt files.
+      }
+    }
+    return await withHeldFileLock(
+      this.lockFileFor(key),
+      { staleMs: CLAIM_LOCK_STALE_MS, maxWaitMs: CLAIM_LOCK_MAX_WAIT_MS },
+      async (acquired) => {
+        if (!acquired) return "already_claimed";
+        // Under the lock the claim check is authoritative: an existing claim
+        // keeps its content and its lease; the session file stays with its
+        // owner (their next save merges into the claim).
+        if (this.isClaimed(key)) return "already_claimed";
+        try {
+          await rename(sessionPath, this.fileFor(key));
+        } catch (err) {
+          // The session file is gone: either it never existed, or another
+          // drainer claimed it first.
+          if (!existsSync(sessionPath)) {
+            if (this.isClaimed(key)) return "already_claimed";
+            return "empty";
+          }
+          throw err;
+        }
+        const claim = await this.readClaimBody(key);
+        if (!claim || (claim.entry.turns.length === 0 && (claim.entry.retainedTurns?.length ?? 0) === 0)) {
+          // Empty content: nothing to drain. Delete the moved file so the
+          // empty entry keeps living with its owner, not as a stale claim.
+          await unlink(this.fileFor(key)).catch(() => {});
+          return "empty";
+        }
+        await this.writeLease(key, lease);
+        return "claimed";
+      },
+    );
+  }
+
+  /**
+   * Take over a claim whose lease is missing, expired, or whose owner died.
+   * The whole judgement runs under the claim's mutation lock and re-reads
+   * the lease inside it, so exactly one concurrent adopter wins: the loser
+   * re-reads the winner's fresh lease and reports "busy" instead of
+   * unlinking it. A lease whose owner died is taken over in the same call.
+   */
+  async adopt(key: string, lease: DrainLease): Promise<AdoptOutcome> {
+    if (!this.isClaimed(key)) return "busy";
+    return await withHeldFileLock(
+      this.lockFileFor(key),
+      { staleMs: CLAIM_LOCK_STALE_MS, maxWaitMs: CLAIM_LOCK_MAX_WAIT_MS },
+      async (acquired) => {
+        if (!acquired) return "busy";
+        const existing = await this.readLease(key);
+        if (existing && !isLeaseExpired(existing)) {
+          if (existing.ownerId === lease.ownerId) {
+            // The same drainer already holds it: renew the lease.
+            await this.writeLease(key, { ...existing, heartbeatAt: new Date().toISOString(), leaseMs: lease.leaseMs });
+            return "adopted";
+          }
+          if (isOwnerAlive(existing)) return "busy";
+          // The owner died without releasing the claim (killed worker):
+          // take it over now instead of waiting for the lease to expire.
+          await unlink(this.leaseFileFor(key)).catch(() => {});
+        } else if (existing) {
+          // Expired lease: clear it before the exclusive create. We hold the
+          // lock, so no concurrent adopter can interleave its own
+          // unlink/create between these steps.
+          await unlink(this.leaseFileFor(key)).catch(() => {});
+        }
+        try {
+          await writeFile(this.leaseFileFor(key), JSON.stringify(lease), { flag: "wx" });
+        } catch (err) {
+          // The lease file exists: another drainer took it first.
+          if (existsSync(this.leaseFileFor(key))) return "busy";
+          throw err;
+        }
+        return "adopted";
+      },
+    );
+  }
+
+  /**
+   * Merge a live owner's in-memory entry INTO the claim (the handover path a
+   * buffer owner takes instead of persisting a claimed session). Serialized
+   * against drainer progress writes on the same lock.
+   */
+  async mergeIntoClaim(key: string, incoming: BufferEntryState): Promise<void> {
+    await withHeldFileLock(this.lockFileFor(key), { staleMs: CLAIM_LOCK_STALE_MS, maxWaitMs: CLAIM_LOCK_MAX_WAIT_MS }, async (acquired) => {
+      if (!acquired) {
+        // Never fall back to a session-file write behind a held claim lock:
+        // that would resurrect claimed turns. Fail so the owner's save stays
+        // pending and retries (SmartBuffer keeps pending state on failure).
+        throw new Error(`buffer drain claim for ${key} is busy; retrying the save later`);
+      }
+      const claim = await this.readClaimBody(key);
+      if (!claim) {
+        // The claim vanished (acked between the isClaimed check and now):
+        // nothing to hand over; the drain already owns everything.
+        return;
+      }
+      await this.writeClaimBody(key, mergeEntries(claim, incoming));
+    });
+  }
+
+  /** Drainer progress: replace the claim content with what is left to do.
+   *
+   * Fenced: `ownerId` must still match the claim's lease (checked under the
+   * same lock), so a drainer that lost its lease cannot write progress that
+   * would restore turns the new owner already extracted.
+   *
+   * `seen` is the entry the drainer LOADED (its snapshot of the claim). Only
+   * turns that were in that snapshot and are absent from BOTH of the
+   * drainer's lists were actually extracted and move to the tombstones — a
+   * turn the drainer moved into `entry.retainedTurns` is retained for retry,
+   * not extracted. Turns a live owner merged into the claim afterwards stay
+   * pending. Without `seen` the pre-existing behavior is kept (every claim
+   * turn missing from `entry` counts as extracted).
+   */
+  async writeEntry(key: string, ownerId: string, entry: BufferEntryState, seen?: BufferEntryState): Promise<void> {
+    await withHeldFileLock(this.lockFileFor(key), { staleMs: CLAIM_LOCK_STALE_MS, maxWaitMs: CLAIM_LOCK_MAX_WAIT_MS }, async (acquired) => {
+      if (!acquired) throw new Error(`buffer drain claim for ${key} is busy; progress write skipped`);
+      const lease = await this.readLease(key);
+      if (!lease || lease.ownerId !== ownerId) {
+        throw new Error(`buffer drain claim for ${key} is no longer owned by ${ownerId}; progress write skipped`);
+      }
+      const claim = await this.readClaimBody(key);
+      if (!claim) return;
+      const seenTurns = seen?.turns ?? claim.entry.turns;
+      const seenRetained = seen?.retainedTurns ?? claim.entry.retainedTurns ?? [];
+      const entryRetained = entry.retainedTurns ?? [];
+      // Still held by the drainer: in its remaining turns OR retained for retry.
+      const kept = (turn: BufferTurn): boolean =>
+        entry.turns.some((held) => bufferTurnsEqual(held, turn)) || entryRetained.some((held) => bufferTurnsEqual(held, turn));
+      // Progress: the snapshot turns that vanished from the entry were extracted.
+      const drained = [...(claim.drained ?? [])];
+      for (const turn of claim.entry.turns) {
+        const extracted =
+          seenTurns.some((snapshot) => bufferTurnsEqual(snapshot, turn)) && !kept(turn);
+        if (!extracted) continue;
+        if (drained.some((gone) => bufferTurnsEqual(gone, turn))) continue;
+        drained.push(turn);
+      }
+      const drainedRetained = [...(claim.drainedRetained ?? [])];
+      for (const turn of claim.entry.retainedTurns ?? []) {
+        const extracted =
+          seenRetained.some((snapshot) => bufferTurnsEqual(snapshot, turn)) && !kept(turn);
+        if (!extracted) continue;
+        if (drained.some((gone) => bufferTurnsEqual(gone, turn)) || drainedRetained.some((gone) => bufferTurnsEqual(gone, turn))) continue;
+        drainedRetained.push(turn);
+      }
+      // Turns merged in after the drainer loaded stay pending, in extraction
+      // order behind the drainer's own remainder.
+      const late = claim.entry.turns.filter(
+        (turn) =>
+          !seenTurns.some((snapshot) => bufferTurnsEqual(snapshot, turn)) && !kept(turn),
+      );
+      const next: BufferEntryState = { ...entry, turns: [...entry.turns, ...late] };
+      const lateRetained = (claim.entry.retainedTurns ?? []).filter(
+        (turn) =>
+          !seenRetained.some((snapshot) => bufferTurnsEqual(snapshot, turn)) && !kept(turn),
+      );
+      if (lateRetained.length > 0) next.retainedTurns = mergeTurns(entryRetained, lateRetained, bufferTurnsEqual);
+      await this.writeClaimBody(key, { ...claim, entry: next, drained, drainedRetained });
+    });
+  }
+
+  /** Extend the lease. "lost" means another drainer took the claim over.
+   * Serialized on the claim lock so a stale heartbeat can never overwrite a
+   * new owner's lease mid-takeover; a contended lock just skips this tick
+   * (usually the drainer's own progress write holds it). */
+  async heartbeat(key: string, ownerId: string, leaseMs: number): Promise<"ok" | "lost"> {
+    return await withHeldFileLock(this.lockFileFor(key), { staleMs: CLAIM_LOCK_STALE_MS, maxWaitMs: CLAIM_LOCK_MAX_WAIT_MS }, async (acquired) => {
+      if (!acquired) return "ok";
+      const lease = await this.readLease(key);
+      if (!lease || lease.ownerId !== ownerId) return "lost";
+      await this.writeLease(key, { ...lease, heartbeatAt: new Date().toISOString(), leaseMs });
+      return "ok";
+    });
+  }
+
+  /** Give the key back (content stays; the next run resumes it). */
+  async releaseLease(key: string, ownerId: string): Promise<void> {
+    const lease = await this.readLease(key);
+    if (!lease || lease.ownerId !== ownerId) return;
+    await unlink(this.leaseFileFor(key)).catch(() => {});
+  }
+
+  /**
+   * Finish a claim: only its owner may ack, and only when nothing is left to
+   * extract. Non-empty claims stay (deferred content is never dropped here).
+   */
+  async ack(key: string, ownerId: string): Promise<AckOutcome> {
+    const lease = await this.readLease(key);
+    if (!lease || lease.ownerId !== ownerId) return "busy";
+    const claim = await this.readClaimBody(key);
+    if (!claim) {
+      await unlink(this.leaseFileFor(key)).catch(() => {});
+      return "acked";
+    }
+    if (claim.entry.turns.length > 0 || (claim.entry.retainedTurns?.length ?? 0) > 0) return "not_empty";
+    // Keep the (empty) claim file as a TOMBSTONE: a live buffer owner that
+    // still holds a stale copy subtracts the drained turns instead of
+    // writing them back. The owner consumes the tombstone on its next save.
+    await unlink(this.leaseFileFor(key)).catch(() => {});
+    return "acked";
+  }
+
+  /** The turns a drain already extracted for `key` (tombstone content). */
+  async drainedTurns(key: string): Promise<BufferTurn[]> {
+    const claim = await this.readClaimBody(key);
+    return claim?.drained ?? [];
+  }
+
+  /**
+   * Is the drain for `key` FINISHED (acked): the claim holds extracted turns
+   * and nothing pending, and the drainer dropped its lease. A claim with
+   * partial progress or a live lease is still being drained — its owner must
+   * merge into it, never release it.
+   */
+  async isDrained(key: string): Promise<boolean> {
+    const claim = await this.readClaimBody(key);
+    if (!claim) return false;
+    if ((claim.drained?.length ?? 0) === 0 && (claim.drainedRetained?.length ?? 0) === 0) return false;
+    if (claim.entry.turns.length > 0 || (claim.entry.retainedTurns?.length ?? 0) > 0) return false;
+    return (await this.readLease(key)) === null;
+  }
+
+  /** Drop the claim file entirely (its owner consumed the tombstone). */
+  async release(key: string): Promise<void> {
+    await unlink(this.fileFor(key)).catch(() => {});
+    await unlink(this.leaseFileFor(key)).catch(() => {});
+  }
+
+  /** Every claim with its lease state (CLI scan + summaries). */
+  async list(): Promise<DrainClaimListing[]> {
+    let names: string[];
+    try {
+      names = (await readdir(this.dir)).filter((name) => name.endsWith(".json") && !name.endsWith(".lease.json"));
+    } catch (err) {
+      // SAFETY: the value comes from this module's own write path; the shape is checked above.
+      if ((err as NodeJS.ErrnoException)?.code === "ENOENT") return [];
+      throw err;
+    }
+    const out: DrainClaimListing[] = [];
+    for (const name of names) {
+      const claim = await this.readClaimBodyByKey(name);
+      if (!claim) continue;
+      const lease = await this.readLease(claim.key);
+      out.push({
+        key: claim.key,
+        turnCount: claim.entry.turns.length + (claim.entry.retainedTurns?.length ?? 0),
+        leased: lease !== null,
+        leaseExpired: lease !== null ? isLeaseExpired(lease) : true,
+      });
+    }
+    return out;
+  }
+
+  private async readClaimBodyByKey(name: string): Promise<DrainClaimFile | null> {
+    try {
+      return parseBody(await this.io.read(path.join(this.dir, name)));
+    } catch (err) {
+      log.debug(`buffer drain: skipping unreadable claim file ${name}: ${err instanceof Error ? err.message : String(err)}`);
+      return null;
+    }
+  }
+}

package/src/storage/buffer-session-files.ts

@@ -1,7 +1,9 @@
+import { BufferDrainClaims, type BufferSessionSaveOutcome } from "./buffer-drain-claims.js";
 import { createHash } from "node:crypto";
 import type { Stats } from "node:fs";
 import { lstat, readdir, unlink } from "node:fs/promises";
 import path from "node:path";
+import { subtractBufferTurns } from "../buffer-turn-helpers.js";
 import { log } from "../logger.js";
 import { SecureStoreDecryptError, SecureStoreLockedError } from "../secure-store/secure-fs.js";
 import type { BufferEntryState, BufferState, BufferTurn } from "../types.js";
@@ -75,14 +77,32 @@ function entryFromBody(body: SessionFileBody): { key: string; entry: BufferEntry
  * One file per buffer session under `<stateDir>/buffer/`, so the observe hot
  * path persists only the session that changed. All writes happen under
  * SmartBuffer's serialized mutation chain (one writer per process).
+ *
+ * Sessions claimed by `remnic buffer drain` (see buffer-drain-claims.ts) are
+ * owned by the drain: `save` hands the in-memory entry INTO the claim
+ * (identity-deduped merge) instead of writing a session file, so a live
+ * buffer owner can never resurrect claimed turns.
  */
 export class BufferSessionFiles {
   private dirVerified = false;
+  private readonly claimsStore: BufferDrainClaims;
 
   constructor(
     private readonly stateDir: string,
     private readonly io: BufferFileIo,
-  ) {}
+  ) {
+    this.claimsStore = new BufferDrainClaims(stateDir, io);
+  }
+
+  /** The drain-claim store for this state dir (CLI scan + claim ops). */
+  get claims(): BufferDrainClaims {
+    return this.claimsStore;
+  }
+
+  /** Sync gate check: is this session claimed by a buffer drain? */
+  isClaimed(key: string): boolean {
+    return this.claimsStore.isClaimed(key);
+  }
 
   private get dir(): string {
     return path.join(this.stateDir, "buffer");
@@ -113,10 +133,37 @@ export class BufferSessionFiles {
   }
 
   /** Persist ONE session atomically. Compact JSON: machine-read state, and
-   * pretty-printing a large buffer costs CPU and bytes on every observe. */
-  async save(key: string, entry: BufferEntryState): Promise<void> {
+   * pretty-printing a large buffer costs CPU and bytes on every observe.
+   * A drain-claimed session is handed over instead (see class doc). */
+  async save(key: string, entry: BufferEntryState): Promise<BufferSessionSaveOutcome> {
+    if (this.isClaimed(key)) {
+      // Consume the tombstone ONLY when the drain finished (acked: extracted
+      // turns recorded, nothing pending, no lease). Partial `drained` just
+      // means chunks completed — the drainer may still be running, and
+      // releasing here would unlink its claim and turn its later writes into
+      // no-ops.
+      if (await this.claimsStore.isDrained(key)) {
+        const drained = await this.claimsStore.drainedTurns(key);
+        // The drain finished this session: the claim is a tombstone. Write
+        // the owner's fresh remainder back as a session file (the drained
+        // turns are subtracted here too, so a caller that did not subtract
+        // them cannot resurrect them) and release the tombstone.
+        const fresh: BufferEntryState = { ...entry, turns: subtractBufferTurns(entry.turns, drained) };
+        if (entry.retainedTurns && entry.retainedTurns.length > 0) {
+          fresh.retainedTurns = subtractBufferTurns(entry.retainedTurns, drained);
+          if (fresh.retainedTurns.length === 0) delete fresh.retainedTurns;
+        }
+        await this.assertDirSafe();
+        await this.io.write(this.fileFor(key), JSON.stringify({ key, ...fresh }));
+        await this.claimsStore.release(key);
+        return { status: "saved", drainedTurns: drained };
+      }
+      await this.claimsStore.mergeIntoClaim(key, entry);
+      return { status: "handed_over" };
+    }
     await this.assertDirSafe();
     await this.io.write(this.fileFor(key), JSON.stringify({ key, ...entry }));
+    return { status: "saved" };
   }
 
   /** Remove one session's file. Idempotent: an absent file is fine. */