@nimbus-sh/core 0.5.0 → 0.7.0

package/src/observability/oom-classify.ts

@@ -1,260 +0,0 @@
-/**
- * oom-classify.ts — discriminator for OOM-like errors at Nimbus
- * boundaries.
- *
- * Why this exists
- * ───────────────
- * Nimbus has at least five distinct "the work failed" failure modes
- * that all surface today as either a thrown JS Error or a console
- * line:
- *
- *   1. SQLITE_NOMEM at the storage layer (per-DO SQLite cap, post-
- *      STOR/SPEC: Address SQLITE_NOMEM issues).
- *   2. Isolate memory exhaustion → 'oom'.
- *   3. Isolate CPU-budget exhaustion → 'cpu_exceeded'.
- *   4. Structured-clone refusal (`Cannot deserialize cloned data`)
- *      between supervisor ↔ facet RPC (32 MiB cap).
- *   5. RPC timeout (TimeoutError from facet-pool's per-task race).
- *
- * Plus a few platform-side terminations (subrequest cap, condemnation,
- * hard eviction) that the user sees but Nimbus has no first-party
- * signal for.
- *
- * Memory and CPU are separate buckets
- * ───────────────────────────────────
- * workerd models them as distinct trace outcomes — `exceededMemory` vs
- * `exceededCpu` (EventOutcome, workerd/io/outcome.capnp) — and words the
- * two message families so neither one's signature appears in the other:
- *
- *   memory  "Worker has exceeded memory limit."
- *           "Worker exceeded memory limit."
- *           "broken.exceededMemory; jsg.Error: Durable Object's isolate
- *            exceeded its memory limit …"
- *           "Memory limit exceeded"                    (RangeError)
- *   cpu     "Worker exceeded CPU time limit."
- *           "Durable Object exceeded its CPU time limit and was reset."
- *           "Python Worker exceeded CPU time limit"
- *
- * Read a low 'cpu_exceeded' count carefully: neither condition reliably
- * produces a message. Both are uncatchable inside the isolate that died,
- * so the string is only observable by a CALLER across an RPC boundary,
- * and Nimbus has repeatedly measured isolates vanishing with no throw at
- * all. Absence of 'cpu_exceeded' entries is therefore NOT evidence that
- * CPU was not the cause — confirm against `wrangler tail` either way.
- *
- * Without a classifier, every recordFailure() call has to stringify-
- * match its own error site. With this module, we pin the regex once
- * and reuse it everywhere.
- *
- * Inputs we accept
- * ────────────────
- * The classifier accepts:
- *   - Error instances (read e.message)
- *   - plain strings (use directly — for stderr-line classification at
- *     facet-exit boundaries where we don't have an Error)
- *   - anything else (return 'unknown')
- *
- * Output values
- * ─────────────
- * Same set used by `DiagFailure.cause` in oom-discriminator.ts. Keep
- * the union narrow and additive — adding a new value is fine, but
- * never re-purpose an existing one.
- */
-
-const OOM_CAUSES = [
-  'sqlite_nomem',
-  'oom',
-  'cpu_exceeded',
-  'clone_refused',
-  'rpc_timeout',
-  'subrequest_cap',
-  'dynamic_worker_cap',
-  'condemnation',
-  'hard_evict',
-  'unknown',
-] as const;
-
-export type OomCause = typeof OOM_CAUSES[number];
-
-export function isOomCause(input: unknown): input is OomCause {
-  return OOM_CAUSES.some((cause) => cause === input);
-}
-
-/**
- * Classify an error or message string into an OomCause. Returns
- * 'unknown' when no signature matches — callers should still record
- * the message via DiagFailure.message so a human can later widen the
- * classifier rules.
- */
-export function classifyError(input: unknown): OomCause {
-  const msg = readMessage(input);
-  if (msg.length === 0) return 'unknown';
-  return classifyMessage(msg);
-}
-
-/**
- * Variant for callers that already have the message string. Prefer
- * classifyError() at boundaries; this is exposed for cases where the
- * message has already been extracted (e.g. truncated / sanitised).
- */
-export function classifyMessage(msg: string): OomCause {
-  // Use lower-case match for forgiveness; SQLITE_NOMEM is canonically
-  // upper but stderr can be either.
-  const m = msg.toLowerCase();
-
-  // SQLITE_NOMEM signals
-  if (m.includes('sqlite_nomem')) return 'sqlite_nomem';
-  if (m.includes('out of memory')) return 'sqlite_nomem';
-  if (m.includes('database or disk is full')) return 'sqlite_nomem';
-
-  // Structured-clone refusal — a 32 MiB-cap cousin
-  if (m.includes('deserialize cloned data')) return 'clone_refused';
-  if (m.includes('could not be cloned')) return 'clone_refused';
-
-  // RPC timeout (Nimbus's TimeoutError from facet-pool)
-  if (m.includes('timeouterror')) return 'rpc_timeout';
-  if (/\btimed?\s*out\b/.test(m)) return 'rpc_timeout';
-
-  // Subrequest cap (Cloudflare platform)
-  if (m.includes('too many subrequests')) return 'subrequest_cap';
-
-  // Worker Loader concurrency cap (Cloudflare platform): a Durable Object
-  // admits ~5-6 concurrent dynamic workers, and loader-cache entries are
-  // never released — every distinct loader.get(id) permanently consumes a
-  // slot, so this cap recurs until the DO itself is replaced.
-  if (m.includes('too many concurrent dynamic workers')) return 'dynamic_worker_cap';
-
-  // Memory exhaustion. workerd trace outcome `exceededMemory`; in-band it
-  // arrives as a `broken.exceededMemory` actor break, a "…exceeded [its]
-  // memory limit" kill, or the "Memory limit exceeded" RangeError.
-  if (m.includes('exceededmemory')) return 'oom';
-  if (/exceeded (?:its )?memory limit/.test(m)) return 'oom';
-  if (m.includes('memory limit exceeded')) return 'oom';
-  if (m.includes('memory limit') && m.includes('reset')) return 'oom';
-
-  // CPU-budget exhaustion. A DIFFERENT platform condition from memory —
-  // workerd reports it as trace outcome `exceededCpu`. Never fold the two:
-  // a CPU kill recurs on the same input, a memory kill recurs on the same
-  // working-set size, and the remedies are opposites.
-  //
-  // `cpu time limit` covers both scopes, which differ only by the
-  // possessive: "Worker exceeded CPU time limit." and "Durable Object
-  // exceeded its CPU time limit and was reset."
-  //
-  // Wall-clock handler timeouts are deliberately NOT matched here.
-  // "Actor exceeded event execution time and was disconnected." is the
-  // caller-configurable setHibernatableWebSocketEventTimeout() bound and
-  // "Alarm exceeded its allowed execution time" is its alarm twin. Both
-  // are elapsed-time bounds a handler blocked on I/O trips without burning
-  // CPU, so bucketing them here would repeat the merge this split undoes.
-  if (m.includes('cpu time limit')) return 'cpu_exceeded';
-  if (m.includes('exceeded cpu limit')) return 'cpu_exceeded';
-
-  // Explicit eviction (per CF research §A.1.2)
-  if (m.includes('condemned')) return 'condemnation';
-  if (m.includes('hard evict')) return 'hard_evict';
-
-  return 'unknown';
-}
-
-/**
- * Transient Durable Object reset — the object was reset by the platform
- * mid-request for a reason unrelated to the request's own resource use:
- * a code deploy rolling over, or a storage-subsystem cold-start hiccup.
- * The in-flight request/RPC rejects, but the work itself never ran to a
- * conclusion and is safe to re-attempt.
- *
- * Deliberately narrow: it must NOT match resource kills ("isolate exceeded
- * its memory limit and was reset", "Durable Object exceeded its CPU time
- * limit and was reset"), which classifyMessage() routes to 'oom' and
- * 'cpu_exceeded' — those recur on retry and must surface, not loop. Note
- * both end in "and was reset"; the checks below key on the CAUSE clause,
- * never on the reset itself.
- * Observed verbatim signatures:
- *   - "Durable Object reset because its code was updated."
- *   - "Internal error while starting up Durable Object storage caused
- *      object to be reset; reference = ..."
- *   - "Internal error in Durable Object storage caused object to be reset;
- *      reference = ..."
- *   - "Durable Object storage operation exceeded timeout which caused
- *      the object to be reset."
- *
- * The second and third are the SAME condition worded for two moments —
- * startup versus a live write — and only the first was matched here, which is
- * why an object reset mid-write still failed a whole install. Nimbus has
- * measured the live-write wording twice: a 45.7 MB single-turn write reset the
- * object once and the same write then succeeded 12/12 on retry
- * (`vfs/facet-resident-store.ts`), and it reset a session DO mid-npm-install
- * on 2026-08-10 (probe `agentic-cli/new/pi-official-installer`). Retrying is
- * bounded, so the one case that is NOT transient — crossing the object's
- * storage budget, which the platform also reports this way
- * (`loaders/process-fabric.ts`) — recurs, exhausts the budget and surfaces.
- */
-export function isTransientDoReset(input: unknown): boolean {
-  const m = readMessage(input).toLowerCase();
-  if (m.length === 0) return false;
-  if (m.includes('reset because its code was updated')) return true;
-  if (m.includes('durable object storage caused object to be reset')) return true;
-  if (m.includes('storage operation') && m.includes('reset')) return true;
-  return false;
-}
-
-/**
- * Workerd shed the call because the target Durable Object's input-gate
- * queue was too deep or too old: "Durable Object is overloaded."
- *
- * Distinct from `isTransientDoReset` on purpose. The object is alive and
- * the work never started, so the call is safe to re-attempt — but the
- * cause is queue pressure, so the only useful retry is one that backs off
- * long enough for the queue to drain. Callers pair this with a longer
- * backoff than a reset retry uses.
- */
-export function isDoOverloaded(input: unknown): boolean {
-  return readMessage(input).toLowerCase().includes('durable object is overloaded');
-}
-
-/**
- * One line about a failure, in terms someone can act on.
- *
- * npm install's catch sites each wrote `e?.remoteMessage || e?.message ||
- * String(e)`. That reduces a failure to a bare sentence, and when the platform
- * declines to describe a rejected Durable Object call the sentence is the word
- * `internal error` — which is how `resolver-fanout layer 2 failed: internal
- * error` reached users for months carrying nothing at all. The message is
- * often the least we know: the error's class, whether the description came
- * from a remote isolate, and which condition it classifies as are all still in
- * hand, and naming them costs nothing and invents nothing.
- */
-export function describeError(input: unknown): string {
-  const remote = typeof input === 'object' && input !== null
-    ? (input as { remoteMessage?: unknown }).remoteMessage
-    : undefined;
-  const message = (typeof remote === 'string' && remote.length > 0)
-    ? `${remote} (remote)`
-    : readMessage(input) || String(input);
-  const name = input instanceof Error && input.name && input.name !== 'Error'
-    ? `${input.name}: `
-    : '';
-  const cause = classifyError(input);
-  const suffix = cause !== 'unknown' ? ` [${cause}]`
-    : isTransientDoReset(input) ? ' [transient-do-reset]'
-    : isDoOverloaded(input) ? ' [do-overloaded]'
-    : '';
-  return `${name}${message}${suffix}`;
-}
-
-function readMessage(input: unknown): string {
-  if (input == null) return '';
-  if (typeof input === 'string') return input;
-  if (input instanceof Error) return input.message ?? '';
-  // Some thrown values are plain objects with .message
-  if (typeof input === 'object') {
-    const m = (input as { message?: unknown }).message;
-    if (typeof m === 'string') return m;
-  }
-  try {
-    return String(input);
-  } catch {
-    return '';
-  }
-}

package/src/observability/oom-discriminator.ts

@@ -1,405 +0,0 @@
-/**
- * oom-discriminator.ts — singleton ring buffer for OOM / failure
- * forensics + last-known-RPC-frame + last-known-facet-id +
- * session-recovery-event ring, plus snapshot/rehydrate hooks for DO
- * storage persistence.
- *
- * W5 (Lever 5 / J.1.1) — original module. Pairs with oom-classify.ts.
- * C'.2 — added the recovery_event ring.
- *
- * Why singleton-per-isolate
- * ─────────────────────────
- * Same pattern as src/observability/diag-counters.ts. The supervisor bundle
- * is the consumer; all writers (sqlite-vfs, facet-pool, facet-manager,
- * supervisor-rpc, npm-installer, nimbus-session) live in the same
- * isolate. Module scope provides one process-local diagnostic without
- * threading a handle through each writer.
- *
- * Two distinct rings
- * ──────────────────
- * - failures: things that failed (the original W5 ring).
- * - recoveryEvents: lifecycle transitions of the session. Cold isolate
- *   boot → 'cold' → 'hydrated' → 'active' → 'drained'.
- *
- * Bounded-size guarantees
- * ───────────────────────
- *   - failures ring: 50 entries (RING_SIZE).
- *   - recoveryEvents ring: 50 entries (RECOVERY_RING_SIZE).
- *   - per-message cap: 200 chars (truncated at insert).
- *   - per-RPC-frame: single slot, one object.
- *   - per-facet-id: single slot, one object.
- *
- * Snapshot size stays bounded even with both rings full.
- */
-
-import { isOomCause, type OomCause } from './oom-classify.js';
-
-const RING_SIZE = 50;
-const RECOVERY_RING_SIZE = 50;
-const MESSAGE_CAP = 200;
-
-export interface RpcFrame {
-  method: string;
-  payloadBytes: number;
-  atMs: number;
-}
-
-export interface FacetId {
-  codeId: string;
-  slotIndex: number;
-  atMs: number;
-}
-
-export interface DiagFailure {
-  /** ms epoch */
-  at: number;
-  /** Lifecycle stage (e.g. 'install', 'resolve', 'rpc', 'facet', 'ws'). */
-  phase: string;
-  /** Discriminated cause. See oom-classify.ts. */
-  cause: OomCause;
-  /** Best-effort RSS estimate in bytes (peak observed). */
-  rssEstimateBytes: number;
-  /** process.memoryUsage().heapUsed if available; 0 in DO contexts. */
-  heapUsedBytes: number;
-  /** SqliteVFS LRU hot bytes at time of failure. */
-  lruBytes: number;
-  /** Sum of in-flight RPC payload bytes (best-effort). */
-  inFlightBytes: number;
-  /** Snapshot of the most recent RPC frame, if any. */
-  lastRpcFrame: RpcFrame | null;
-  /** Snapshot of the most recent facet dispatch, if any. */
-  lastFacetId: FacetId | null;
-  /** Optional facet exit code (when failure represents a facet termination). */
-  exitCode?: number;
-  /** Optional facet pid. */
-  pid?: number;
-  /** Free-form message; truncated to MESSAGE_CAP characters. */
-  message?: string;
-}
-
-/**
- * Session lifecycle states for the C'.2 recovery_event ring.
- *
- * This module records state-machine transitions; it does not enforce them.
- *
- * State semantics:
- * - 'cold'      : fresh DO instance, no in-memory session state yet.
- * - 'hydrated'  : Phase R completed — SQL state read into in-memory cache.
- * - 'active'    : Phase B+W+O completed — kernel/shell/terminal wired.
- * - 'drained'   : webSocketError/Close fired; in-memory caches nulled,
- *                 SQL state freshly persisted; awaiting next reconnect.
- *
- * Trigger labels: human-readable event that caused the transition.
- * Examples: 'first-fetch', 'ws-upgrade', 'ws-close', 'ws-error',
- * 'isolate-evicted', 'manual-test'.
- */
-/** Session state union.
- *
- * The original four are the high-level lifecycle markers:
- *   cold     — never inited
- *   hydrated — re-init complete (warm path; Phase O skipped)
- *   active   — live shell processing input
- *   drained  — webSocketError/Close fired; SQL freshly persisted
- *
- * The four [B'.4] additions are the fine-grained phases of
- * initSession itself, recorded as the function progresses:
- *   rehydrate — Phase R: loading state from DO SQLite
- *   build     — Phase B: constructing Kernel/Shell/registry
- *   wire      — Phase W: attaching terminal, replaying scrollback
- *   online    — Phase O: cold-start MOTD + starter hint (cold only;
- *               on warm re-init this phase is skipped and the
- *               initSession ends with 'hydrated' instead)
- *
- * The phase transitions are SUPPLEMENTAL to the high-level markers,
- * not a replacement. Probes that look for the legacy `'hydrated'`
- * marker continue to work; probes that want fine-grained debugging
- * can grep for the phase states.
- */
-export type SessionState =
-  | 'cold' | 'hydrated' | 'active' | 'drained'
-  | 'rehydrate' | 'build' | 'wire' | 'online';
-
-export interface DiagRecoveryEvent {
-  /** ms epoch */
-  at: number;
-  /** Source state. */
-  fromState: SessionState;
-  /** Destination state. */
-  toState: SessionState;
-  /** Human-readable trigger ('ws-close', 'isolate-evicted', etc.). */
-  trigger: string;
-  /** _isolateGen at the time of the event. Lets a probe detect the
-   *  difference between an in-isolate transition and a cross-isolate one
-   *  (the latter implies workerd recycled the DO). */
-  isolateGen: number;
-  /** True when the transition could not preserve state that should
-   *  have survived (e.g. SQL persist threw or a snapshot was missing). */
-  dataLoss: boolean;
-  /** Number of SQL keys/rows rehydrated on a hydrated transition.
-   *  Zero when the transition is one that doesn't read SQL. */
-  snapshotKeysRehydrated: number;
-  /** Optional free-form notes (200 char cap). */
-  notes?: string;
-}
-
-interface RingState {
-  failures: DiagFailure[];
-  recoveryEvents: DiagRecoveryEvent[];
-  lastRpcFrame: RpcFrame | null;
-  lastFacetId: FacetId | null;
-}
-
-const state: RingState = {
-  failures: [],
-  recoveryEvents: [],
-  lastRpcFrame: null,
-  lastFacetId: null,
-};
-
-function getState(): RingState {
-  return state;
-}
-
-/** Append a failure to the ring. Newest first. Capped at RING_SIZE. */
-export function recordFailure(f: DiagFailure): void {
-  const s = getState();
-  // Defensive copy + message cap.
-  const entry: DiagFailure = {
-    at: Number(f.at) || Date.now(),
-    phase: String(f.phase ?? 'unknown'),
-    cause: f.cause,
-    rssEstimateBytes: Number(f.rssEstimateBytes) || 0,
-    heapUsedBytes: Number(f.heapUsedBytes) || 0,
-    lruBytes: Number(f.lruBytes) || 0,
-    inFlightBytes: Number(f.inFlightBytes) || 0,
-    lastRpcFrame: f.lastRpcFrame ?? s.lastRpcFrame,
-    lastFacetId: f.lastFacetId ?? s.lastFacetId,
-  };
-  if (typeof f.exitCode === 'number') entry.exitCode = f.exitCode;
-  if (typeof f.pid === 'number') entry.pid = f.pid;
-  if (f.message) {
-    const m = String(f.message);
-    entry.message = m.length > MESSAGE_CAP ? m.slice(0, MESSAGE_CAP) : m;
-  }
-  s.failures.unshift(entry);
-  if (s.failures.length > RING_SIZE) s.failures.length = RING_SIZE;
-}
-
-/** Read a snapshot of the ring. Newest first. Caller-side mutations
- *  do not affect the singleton. */
-export function getFailures(): DiagFailure[] {
-  return getState().failures.slice(0);
-}
-
-/** Record the current RPC frame (called at every RPC entry). Bounded
- *  to a single slot — the LATEST frame wins. */
-export function setLastRpcFrame(method: string, payloadBytes: number): void {
-  const s = getState();
-  s.lastRpcFrame = {
-    method: String(method),
-    payloadBytes: Number(payloadBytes) || 0,
-    atMs: Date.now(),
-  };
-}
-
-export function getLastRpcFrame(): RpcFrame | null {
-  return getState().lastRpcFrame;
-}
-
-/** Record the most recent facet dispatch. */
-export function setLastFacetId(codeId: string, slotIndex: number): void {
-  const s = getState();
-  s.lastFacetId = {
-    codeId: String(codeId),
-    slotIndex: Number(slotIndex) || 0,
-    atMs: Date.now(),
-  };
-}
-
-export function getLastFacetId(): FacetId | null {
-  return getState().lastFacetId;
-}
-
-// ── C'.2 recovery_event ring ────────────────────────────────────────────
-//
-// The ring is bounded at RECOVERY_RING_SIZE; the diag endpoint reads via
-// getRecoveryEvents() (newest first).
-
-/** Append a recovery event to the ring. Newest first. Capped at
- *  RECOVERY_RING_SIZE. */
-export function recordRecoveryEvent(e: DiagRecoveryEvent): void {
-  const s = getState();
-  // Defensive copy + notes cap. We intentionally do NOT validate the
-  // state-machine direction here (e.g. that 'drained' only follows
-  // 'active') — the state machine lives elsewhere; this ring is a
-  // recorder, not an enforcer. If a probe sees an impossible
-  // transition, that's a real bug in the state machine and the
-  // probe must fail.
-  const entry: DiagRecoveryEvent = {
-    at: Number(e.at) || Date.now(),
-    fromState: e.fromState,
-    toState: e.toState,
-    trigger: String(e.trigger ?? 'unknown'),
-    isolateGen: Number(e.isolateGen) || 0,
-    dataLoss: !!e.dataLoss,
-    snapshotKeysRehydrated: Number(e.snapshotKeysRehydrated) || 0,
-  };
-  if (e.notes) {
-    const n = String(e.notes);
-    entry.notes = n.length > MESSAGE_CAP ? n.slice(0, MESSAGE_CAP) : n;
-  }
-  s.recoveryEvents.unshift(entry);
-  if (s.recoveryEvents.length > RECOVERY_RING_SIZE) {
-    s.recoveryEvents.length = RECOVERY_RING_SIZE;
-  }
-}
-
-/** Read a snapshot of the recovery ring. Newest first. */
-export function getRecoveryEvents(): DiagRecoveryEvent[] {
-  return getState().recoveryEvents.slice(0);
-}
-
-/** Reset the recovery ring. Tests + manual operator use only. */
-export function resetRecoveryEvents(): void {
-  getState().recoveryEvents.length = 0;
-}
-
-/**
- * Build a JSON-serializable snapshot suitable for ctx.storage.put.
- * Bounded ≤ 40 KB even with both rings full of MESSAGE_CAP-truncated
- * messages. Schema version embedded so a future shape change can
- * cleanly reject old snapshots.
- *
- * Schema version 2 includes both failure and recovery-event rings.
- */
-export interface DiagSnapshot {
-  /** Schema version. Bump when shape changes. */
-  v: 2;
-  failures: DiagFailure[];
-  recoveryEvents: DiagRecoveryEvent[];
-  lastRpcFrame: RpcFrame | null;
-  lastFacetId: FacetId | null;
-}
-
-export function snapshotForStorage(): DiagSnapshot {
-  const s = getState();
-  return {
-    v: 2,
-    failures: s.failures.slice(0),
-    recoveryEvents: s.recoveryEvents.slice(0),
-    lastRpcFrame: s.lastRpcFrame,
-    lastFacetId: s.lastFacetId,
-  };
-}
-
-/**
- * Restore from a snapshot. Garbage / null / wrong-shape input is
- * silently ignored. Does NOT throw — constructor-time rehydration must
- * never block DO startup.
- *
- * Only the current v2 schema is accepted.
- */
-export function rehydrateFromStorage(blob: unknown): void {
-  if (!isRecord(blob) || blob.v !== 2) return;
-  const b = blob;
-  if (!Array.isArray(b.failures)) return;
-
-  const s = getState();
-  s.failures.length = 0;
-  for (const f of b.failures) {
-    const entry = parseFailure(f);
-    if (!entry) continue;
-    s.failures.push(entry);
-    if (s.failures.length >= RING_SIZE) break;
-  }
-  s.recoveryEvents.length = 0;
-  if (Array.isArray(b.recoveryEvents)) {
-    for (const e of b.recoveryEvents) {
-      const entry = parseRecoveryEvent(e);
-      if (!entry) continue;
-      s.recoveryEvents.push(entry);
-      if (s.recoveryEvents.length >= RECOVERY_RING_SIZE) break;
-    }
-  }
-  const lastRpcFrame = parseRpcFrame(b.lastRpcFrame);
-  if (lastRpcFrame) s.lastRpcFrame = lastRpcFrame;
-  const lastFacetId = parseFacetId(b.lastFacetId);
-  if (lastFacetId) s.lastFacetId = lastFacetId;
-}
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return typeof value === 'object' && value !== null;
-}
-
-function parseRpcFrame(value: unknown): RpcFrame | null {
-  if (!isRecord(value)
-      || typeof value.method !== 'string'
-      || typeof value.payloadBytes !== 'number'
-      || typeof value.atMs !== 'number') return null;
-  return { method: value.method, payloadBytes: value.payloadBytes, atMs: value.atMs };
-}
-
-function parseFacetId(value: unknown): FacetId | null {
-  if (!isRecord(value)
-      || typeof value.codeId !== 'string'
-      || typeof value.slotIndex !== 'number'
-      || typeof value.atMs !== 'number') return null;
-  return { codeId: value.codeId, slotIndex: value.slotIndex, atMs: value.atMs };
-}
-
-function parseFailure(value: unknown): DiagFailure | null {
-  if (!isRecord(value)
-      || typeof value.at !== 'number'
-      || typeof value.phase !== 'string'
-      || !isOomCause(value.cause)
-      || typeof value.rssEstimateBytes !== 'number'
-      || typeof value.heapUsedBytes !== 'number'
-      || typeof value.lruBytes !== 'number'
-      || typeof value.inFlightBytes !== 'number') return null;
-  const entry: DiagFailure = {
-    at: value.at,
-    phase: value.phase,
-    cause: value.cause,
-    rssEstimateBytes: value.rssEstimateBytes,
-    heapUsedBytes: value.heapUsedBytes,
-    lruBytes: value.lruBytes,
-    inFlightBytes: value.inFlightBytes,
-    lastRpcFrame: parseRpcFrame(value.lastRpcFrame),
-    lastFacetId: parseFacetId(value.lastFacetId),
-  };
-  if (typeof value.exitCode === 'number') entry.exitCode = value.exitCode;
-  if (typeof value.pid === 'number') entry.pid = value.pid;
-  if (typeof value.message === 'string') entry.message = value.message.slice(0, MESSAGE_CAP);
-  return entry;
-}
-
-const SESSION_STATES: readonly SessionState[] = [
-  'cold', 'hydrated', 'active', 'drained',
-  'rehydrate', 'build', 'wire', 'online',
-];
-
-function isSessionState(value: unknown): value is SessionState {
-  return SESSION_STATES.some((state) => state === value);
-}
-
-function parseRecoveryEvent(value: unknown): DiagRecoveryEvent | null {
-  if (!isRecord(value)
-      || typeof value.at !== 'number'
-      || !isSessionState(value.fromState)
-      || !isSessionState(value.toState)
-      || typeof value.trigger !== 'string'
-      || typeof value.isolateGen !== 'number'
-      || typeof value.dataLoss !== 'boolean'
-      || typeof value.snapshotKeysRehydrated !== 'number') return null;
-  const entry: DiagRecoveryEvent = {
-    at: value.at,
-    fromState: value.fromState,
-    toState: value.toState,
-    trigger: value.trigger,
-    isolateGen: value.isolateGen,
-    dataLoss: value.dataLoss,
-    snapshotKeysRehydrated: value.snapshotKeysRehydrated,
-  };
-  if (typeof value.notes === 'string') entry.notes = value.notes.slice(0, MESSAGE_CAP);
-  return entry;
-}