@coreplane/switchboard 0.0.0 → 1.18.0

package/dist/assets/src/deploy/restart.ts

@@ -0,0 +1,393 @@
+import { parseIngressTokenMap, type IngressIdentity } from "../core/ingressTokens.js";
+import { hasAction } from "../core/authz/authorize.js";
+import type { GrantsLookup } from "../core/authz/actor.js";
+import { LIVE_GATE_DEADLINE_MS, parseHealthz, type HealthzBody } from "./liveGate.js";
+import { profileUrls, type DeploymentProfile } from "./profile.js";
+
+// `deploy restart` — restart the bot container WITHOUT an image build, so a
+// rotated bot secret goes live in seconds instead of a full `deploy all --only
+// bot` (docs/reference/specs/slack-channel.md item 8). Cloudflare's model: `wrangler secret
+// put` updates the Worker's env, but a running container keeps the env it
+// started with, and a rollout only happens on an image/config change. The
+// documented restart is the Container DO calling `stop()` (SIGTERM → the bot's
+// graceful drain finishes in-flight runs and exits) and the NEXT request
+// starting it again — with envVars computed at start time from the DO's
+// current env (deploy/cloudflare/worker.ts). The Worker exposes that as
+// `POST /admin/restart`; this module is the pure half shared by the Worker
+// (authorization, the refusal decision, the wire shapes) and the CLI runner
+// (src/deploy/run.ts `runBotRestart`), so both sides are unit-tested here and
+// nothing in this file imports node:*.
+//
+// Authorization: the bearer must be an entry of the bot's own
+// `SWITCHBOARD_INGRESS_TOKENS` map whose `http:<subject>` actor holds
+// `deploy:write` in the bot's config (`grants`, authorization.md item 9) — the
+// very action the `deploy.restart` command declares, so the Worker route is
+// authorized exactly as `/api/deploy.restart` would be if the bot served it.
+// Two halves, because the two sides hold different things: the Worker holds the
+// token map (it fires scheduled runs with the `cron` entry) and can tell WHO a
+// bearer is — `authenticateRestart`, 401/503 without touching the container —
+// but the grants live in the container's config, so it asks the bot
+// (`POST /admin/restart/authorize`, src/channels/adminRestartAuthorize.ts) whether
+// that subject holds the action and relays the answer (`parseRestartAuthorization`).
+// The Worker names the subject it authenticated in `RESTART_SUBJECT_HEADER` and
+// the bot decides the grant for THAT subject without re-authenticating the
+// bearer: the two sides can hold different generations of the token map (a
+// `wrangler secret put` reaches the Worker's env at once and the container's
+// only after a restart), so re-authenticating in the container would refuse the
+// very rotation the restart exists to finish. The header is trustworthy because
+// the container is reachable only through the Worker, which strips it from every
+// proxied request (`stripRestartSubject`) and sets it only on its own internal
+// call. The bot's own `/admin/crash` runs the whole check in one place
+// (`authorizeRestart`).
+//
+// The route's URL is the installation's: the deployment profile names the
+// bot's hostname, `planRestart` derives `https://<bot>/admin/restart` from it.
+
+/** The operator's env var holding a `SWITCHBOARD_INGRESS_TOKENS` bearer with `deploy:write`. */
+export const RESTART_TOKEN_ENV = "SWITCHBOARD_DEPLOY_TOKEN";
+/** The scope the bearer's identity must carry — the `deploy.restart` command's own. */
+export const RESTART_SCOPE = "deploy:write";
+
+// ---- refusal decision (Worker side; the CLI relies on the 409) -----------------------------
+
+export interface RestartVerdict {
+  allow: boolean;
+  forced: boolean;
+  /** What refuses (fail-closed: no JSON body, an impossible count). */
+  problems: string[];
+  /** What is said but does not refuse: runs in flight (they hand off), a drain under way. */
+  warnings: string[];
+  message: string;
+}
+/**
+ * Whether the container may be stopped now — the deploy preflight's rules
+ * (deploy/cloudflare/preflight.mjs `decide`) minus the rollout-state check (a
+ * restart is not a rollout). Since the handoff (docs/reference/specs/run-history.md item
+ * 39) runs in flight and a drain under way are WARNINGS, not refusals: SIGTERM
+ * hands every resumable run to the next generation. Fail closed on a body
+ * that is not JSON or an impossible count; `force` allows those anyway, with
+ * the warning.
+ */
+export function decideRestart(body: HealthzBody | undefined, opts: { force: boolean }): RestartVerdict {
+  const problems: string[] = [];
+  const warnings: string[] = [];
+  if (!body) {
+    problems.push(
+      "bot not answering with JSON on /healthz (container restarting, unreachable, or a Worker that predates the preflight)",
+    );
+  } else {
+    if (!Number.isInteger(body.inFlight) || (body.inFlight as number) < 0) {
+      problems.push(`bot reports an impossible inFlight=${JSON.stringify(body.inFlight)} (counter bug or old Worker)`);
+    } else if ((body.inFlight as number) > 0) {
+      warnings.push(
+        `${body.inFlight} run(s) in flight — handed to the next generation on SIGTERM (run-history item 39); they continue there`,
+      );
+    }
+    if (body.draining === true)
+      warnings.push(
+        "bot is already draining (a deploy or an earlier restart is in progress) — it restarts on its own when the drain ends; a second stop is harmless",
+      );
+  }
+  const said =
+    warnings.length > 0 ? ` —\n${warnings.map((w) => `  - ${w}`).join("\n")}` : ": no runs in flight, not draining";
+  if (problems.length === 0) return { allow: true, forced: false, problems, warnings, message: `restart ok${said}` };
+  const detail = problems.map((p) => `  - ${p}`).join("\n");
+  if (opts.force)
+    return {
+      allow: true,
+      forced: true,
+      problems,
+      warnings,
+      message: `restart WARNING: stopping by force despite —\n${detail}`,
+    };
+  return {
+    allow: false,
+    forced: false,
+    problems,
+    warnings,
+    message: `restart REFUSED —\n${detail}\n  wait and retry, or pass --force to stop blind`,
+  };
+}
+
+// ---- authorization (Worker side) ---------------------------------------------------------------
+
+/** Byte-wise equality whose running time depends only on the lengths, never on
+ *  where the first difference is. */
+export function constantTimeEqual(a: string, b: string): boolean {
+  const enc = new TextEncoder();
+  const ab = enc.encode(a);
+  const bb = enc.encode(b);
+  let diff = ab.length ^ bb.length;
+  for (let i = 0; i < Math.max(ab.length, bb.length); i++) diff |= (ab[i] ?? 0) ^ (bb[i] ?? 0);
+  return diff === 0;
+}
+
+/** Find the presented bearer among the configured tokens by comparing against
+ *  EVERY entry (fixed work; a plain object-key lookup would let a probe learn
+ *  which prefixes exist from timing). Returns the matched identity, if any. */
+export function lookupConstantTime<T>(tokens: Record<string, T>, presented: string): T | undefined {
+  let found: T | undefined;
+  for (const [token, identity] of Object.entries(tokens)) if (constantTimeEqual(token, presented)) found = identity;
+  return found;
+}
+
+export type RestartAuth = { ok: true; subject: string } | { ok: false; status: 401 | 403 | 503; reason: string };
+export type RestartAuthn = { ok: true; identity: IngressIdentity } | { ok: false; status: 401 | 503; reason: string };
+
+/** The bot route the Worker asks before stopping the container: 200 `{ ok, subject }`
+ *  when the bearer's actor holds `deploy:write`, else `authorizeRestart`'s 401/403/503. */
+export const RESTART_AUTHORIZE_PATH = "/admin/restart/authorize";
+/** The header the Worker sets on its internal authorize call, naming the subject it
+ *  authenticated from its own token map; stripped from every proxied request. */
+export const RESTART_SUBJECT_HEADER = "x-switchboard-restart-subject";
+
+/** WHETHER, for a subject the Worker already authenticated: the bot's grants alone. */
+export function authorizeRestartSubject(subject: string, grantsFor: GrantsLookup): RestartAuth {
+  if (subject.trim() === "") return { ok: false, status: 401, reason: "unauthorized: an empty subject" };
+  if (!hasAction(grantsFor(`http:${subject}`).actions, RESTART_SCOPE))
+    return {
+      ok: false,
+      status: 403,
+      reason: `forbidden: identity "${subject}" holds no ${RESTART_SCOPE} grant (grants["http:${subject}"] in config.yaml)`,
+    };
+  return { ok: true, subject };
+}
+
+/** The request without the subject header — what the Worker forwards to the container for
+ *  every route it does not answer itself, so a caller can never assert a subject. */
+export function stripRestartSubject(request: Request): Request {
+  if (!request.headers.has(RESTART_SUBJECT_HEADER)) return request;
+  const headers = new Headers(request.headers);
+  headers.delete(RESTART_SUBJECT_HEADER);
+  return new Request(request, { headers });
+}
+
+/** WHO the bearer is — the Worker's half. Check `Authorization: Bearer <token>`
+ *  against the raw `SWITCHBOARD_INGRESS_TOKENS` value: no usable map → 503 (the
+ *  route is disabled, never open); no/unknown bearer → 401. Says nothing about
+ *  what the identity may do. Never echoes token material. */
+export function authenticateRestart(authorization: string | undefined, tokensRaw: string | undefined): RestartAuthn {
+  return authenticateIngressBearer(authorization, tokensRaw, "restart");
+}
+
+/** Who the bearer is, from the ingress token map — the same map every
+ *  `/admin/*` route on the bot checks (`feature` names the route in the 503,
+ *  which is what an operator sees when the map is missing). */
+export function authenticateIngressBearer(
+  authorization: string | undefined,
+  tokensRaw: string | undefined,
+  feature: string,
+): RestartAuthn {
+  const parsed = parseIngressTokenMap(tokensRaw);
+  if (!parsed.ok || Object.keys(parsed.tokens).length === 0) {
+    return {
+      ok: false,
+      status: 503,
+      reason: `${feature} disabled: SWITCHBOARD_INGRESS_TOKENS is ${parsed.ok ? "not set" : parsed.reason}`,
+    };
+  }
+  const m = /^Bearer\s+(\S+)$/i.exec(authorization ?? "");
+  const identity = m ? lookupConstantTime(parsed.tokens, m[1]) : undefined;
+  if (!identity)
+    return {
+      ok: false,
+      status: 401,
+      reason: "unauthorized: a Bearer token from SWITCHBOARD_INGRESS_TOKENS is required",
+    };
+  return { ok: true, identity };
+}
+
+/** WHO and WHETHER — the whole check, where the grants are (the bot). The
+ *  bearer's `http:<subject>` actor must hold `deploy:write` (`grantsFor` —
+ *  config's entry for the token's subject); a known identity without it → 403. */
+export function authorizeRestart(
+  authorization: string | undefined,
+  tokensRaw: string | undefined,
+  grantsFor: GrantsLookup,
+): RestartAuth {
+  const authn = authenticateRestart(authorization, tokensRaw);
+  if (!authn.ok) return authn;
+  return authorizeRestartSubject(authn.identity.subject, grantsFor);
+}
+
+/** Any other `/admin/*` route's whole check: an ingress bearer (401/503) whose
+ *  actor `http:<subject>` holds `scope` (403) — the restart's rule with the
+ *  scope as a parameter, so no route invents its own. */
+export function authorizeIngressBearer(
+  authorization: string | undefined,
+  tokensRaw: string | undefined,
+  grantsFor: GrantsLookup,
+  scope: string,
+  feature: string,
+): RestartAuth {
+  const authn = authenticateIngressBearer(authorization, tokensRaw, feature);
+  if (!authn.ok) return authn;
+  const { identity } = authn;
+  if (!hasAction(grantsFor(`http:${identity.subject}`).actions, scope))
+    return {
+      ok: false,
+      status: 403,
+      reason: `forbidden: identity "${identity.subject}" holds no ${scope} grant (grants["http:${identity.subject}"] in config.yaml)`,
+    };
+  return { ok: true, subject: identity.subject };
+}
+
+/** The bot's `POST /admin/restart/authorize` answer, as the Worker reads it: 200
+ *  `{ ok: true, subject }` → allowed; 401 / 403 / 503 `{ ok: false, error }` →
+ *  relayed as they are; anything else (a bot without the route, a non-JSON body,
+ *  an unexpected status) → 503, fail-closed — the Worker never restarts on an
+ *  answer it cannot read. */
+export function parseRestartAuthorization(status: number, text: string): RestartAuth {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(text);
+  } catch {
+    parsed = undefined;
+  }
+  const body = typeof parsed === "object" && parsed !== null ? (parsed as Record<string, unknown>) : undefined;
+  if (status === 200 && body?.ok === true && typeof body.subject === "string" && body.subject !== "")
+    return { ok: true, subject: body.subject };
+  if ((status === 401 || status === 403 || status === 503) && body?.ok === false && typeof body.error === "string")
+    return { ok: false, status, reason: body.error };
+  return {
+    ok: false,
+    status: 503,
+    reason: `restart disabled: the bot did not answer the authorization check (HTTP ${status}) — is it running this version?`,
+  };
+}
+
+// ---- wire shapes ----------------------------------------------------------------------------------
+
+export type ParsedRestartRequest = { ok: true; force: boolean } | { ok: false; reason: string };
+
+/** The request body: empty, or a JSON object with an optional boolean `force`. */
+export function parseRestartRequest(text: string): ParsedRestartRequest {
+  if (text.trim() === "") return { ok: true, force: false };
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(text);
+  } catch {
+    return { ok: false, reason: "body is not valid JSON" };
+  }
+  if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed))
+    return { ok: false, reason: "body must be a JSON object" };
+  const force = (parsed as { force?: unknown }).force;
+  if (force !== undefined && typeof force !== "boolean") return { ok: false, reason: "`force` must be a boolean" };
+  return { ok: true, force: force === true };
+}
+
+/** What the Container DO reports back to the Worker route. */
+export type RestartOutcome =
+  /** The container is not running: nothing to stop; the next request starts it with the current env. */
+  | { kind: "not-running" }
+  | { kind: "refused"; problems: string[] }
+  /** SIGTERM sent; `previousStartedAt` is what the OLD container reported (the CLI's baseline). */
+  | { kind: "stopping"; forced: boolean; inFlight: number; previousStartedAt: string | undefined };
+
+export interface RestartHttpResponse {
+  status: number;
+  body: Record<string, unknown>;
+}
+
+export function restartResponse(outcome: RestartOutcome): RestartHttpResponse {
+  switch (outcome.kind) {
+    case "stopping":
+      return {
+        status: 202,
+        body: {
+          ok: true,
+          stopping: true,
+          forced: outcome.forced,
+          inFlight: outcome.inFlight,
+          previousStartedAt: outcome.previousStartedAt,
+        },
+      };
+    case "refused":
+      return { status: 409, body: { ok: false, refused: true, problems: outcome.problems } };
+    case "not-running":
+      return {
+        status: 200,
+        body: {
+          ok: true,
+          stopping: false,
+          note: "container not running — nothing to stop; the next request starts it with the current env",
+        },
+      };
+  }
+}
+
+/** The CLI's reading of the route's answer. */
+export type RestartResponseClass =
+  | { kind: "stopping"; previousStartedAt: string | undefined }
+  | { kind: "not-running" }
+  /** 409 — retryable while runs finish. */
+  | { kind: "refused"; reason: string }
+  | { kind: "unauthorized"; reason: string }
+  | { kind: "failed"; reason: string };
+
+export function classifyRestartResponse(status: number, text: string): RestartResponseClass {
+  const body = parseHealthz(text) as Record<string, unknown> | undefined;
+  if (status === 202)
+    return {
+      kind: "stopping",
+      previousStartedAt: typeof body?.previousStartedAt === "string" ? body.previousStartedAt : undefined,
+    };
+  if (status === 200 && body?.stopping === false) return { kind: "not-running" };
+  if (status === 409) {
+    const problems = Array.isArray(body?.problems)
+      ? body.problems.filter((p): p is string => typeof p === "string")
+      : [];
+    return { kind: "refused", reason: problems[0] ?? "restart refused" };
+  }
+  if (status === 401 || status === 403)
+    return { kind: "unauthorized", reason: `HTTP ${status}: ${text.slice(0, 200)}` };
+  return { kind: "failed", reason: `HTTP ${status}: ${text.slice(0, 200)}` };
+}
+
+// ---- the plan (what `deploy restart` does, as data) --------------------------------------------
+
+export interface RestartOptions {
+  /** The only supported target for now: the bot is the one Worker with a long-lived container. */
+  only: "bot";
+  force: boolean;
+  waitMaxMinutes: number;
+  pollSeconds: number;
+}
+
+export interface RestartPlan {
+  target: "bot";
+  adminUrl: string;
+  healthUrl: string;
+  tokenEnv: string;
+  force: boolean;
+  /** Budget for waiting out a 409 (runs in flight) before giving up. */
+  waitMaxMs: number;
+  pollMs: number;
+  /** Budget for the new container to answer with a later `startedAt` (drain + cold start). */
+  liveDeadlineMs: number;
+}
+
+export function planRestart(opts: RestartOptions, profile: DeploymentProfile): RestartPlan {
+  const urls = profileUrls(profile);
+  return {
+    target: opts.only,
+    adminUrl: urls.botAdminRestartUrl,
+    healthUrl: `${urls.publicBaseUrl}/healthz`,
+    tokenEnv: RESTART_TOKEN_ENV,
+    force: opts.force,
+    waitMaxMs: opts.waitMaxMinutes * 60_000,
+    pollMs: opts.pollSeconds * 1000,
+    liveDeadlineMs: LIVE_GATE_DEADLINE_MS,
+  };
+}
+
+export function formatRestartPlan(plan: RestartPlan): string {
+  const gate = plan.force
+    ? `preflight FORCED — in-flight runs are drained (SIGTERM), killed only at the drain deadline`
+    : `refused while runs are in flight or draining (409) — retry every ${plan.pollMs / 1000}s up to ${plan.waitMaxMs / 60_000} min`;
+  return [
+    `Restart ${plan.target}: POST ${plan.adminUrl} (bearer from $${plan.tokenEnv}, needs ${RESTART_SCOPE}) — ${gate}`,
+    `  then wait until ${plan.healthUrl} answers not draining with a later startedAt (up to ${Math.round(plan.liveDeadlineMs / 60_000)} min: drain + cold start)`,
+    `  no image build: the container restarts on the same build with the Worker's CURRENT secrets`,
+  ].join("\n");
+}

package/dist/assets/src/effort.ts

@@ -0,0 +1,17 @@
+// Model effort: how hard the model thinks per turn (Anthropic
+// `output_config.effort`; skipped for models without support). A first-class
+// config dimension resolved through the SAME layers as the model ref
+// (docs/reference/specs/routing-and-config.md item 2): request directive > thread-sticky
+// > user scope > channel scope > defaults > the agent definition > the
+// provider's own default. Lower effort = much faster turns; the wall clock is
+// the real budget, so effort is what decides how much of it goes to thinking.
+
+export const EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"] as const;
+export type Effort = (typeof EFFORT_LEVELS)[number];
+
+export function isEffort(value: unknown): value is Effort {
+  return typeof value === "string" && (EFFORT_LEVELS as readonly string[]).includes(value);
+}
+
+/** The valid levels, for error messages: `low, medium, high, xhigh, max`. */
+export const EFFORT_LEVELS_HINT = EFFORT_LEVELS.join(", ");

package/dist/assets/src/execution/bashTimeout.ts

@@ -0,0 +1,78 @@
+// Per-call bash timeout policy (docs/reference/specs/execution.md item 11): ONE clamp and
+// one set of bounds, shared by the tool layer, every executor, and both deploy
+// Workers (which import from src/execution like shellQuote/residentDetach).
+// Deliberately free of node: imports so wrangler can bundle it into Workers.
+
+/** Default per-command budget when the caller passes no timeoutMs. */
+export const BASH_TIMEOUT_MS = 5 * 60_000;
+
+/** Floor: anything lower is a typo or an attack, not a budget. */
+export const BASH_TIMEOUT_MIN_MS = 1_000;
+
+/** Hard ceiling a caller can raise the budget to. Not by itself a guarantee
+ *  against one command eating a run — 20 minutes is 80% of a 25-minute review
+ *  — which is what `bashBudgetWithinRun` below is for. */
+export const BASH_TIMEOUT_MAX_MS = 20 * 60_000;
+
+/** Wall clock a run keeps back from its last command for the write-up: the
+ *  runner's deadline forces a final answer, and a command still running at
+ *  that moment would have been wasted anyway. */
+export const RUN_DEADLINE_RESERVE_MS = 60_000;
+
+/** Margin the remote exec clients add to their HTTP wait over the command
+ *  budget, so the server's own timeout answer (a streamed exit 124) wins the
+ *  race against the client's transport deadline instead of both firing at the
+ *  same instant. */
+export const EXEC_CALL_MARGIN_MS = 30_000;
+
+/** The documented clamp rule, applied identically bot-side and server-side
+ *  (a server never trusts the client's number): a finite number is truncated
+ *  to an integer and clamped into [BASH_TIMEOUT_MIN_MS, BASH_TIMEOUT_MAX_MS]
+ *  — so 0/negative run at the 1s floor, a 25-minute ask runs at 20 minutes —
+ *  and anything else (absent, NaN, Infinity, a string) falls back to the
+ *  5-minute default. */
+export function clampBashTimeout(requested: unknown): number {
+  if (typeof requested !== "number" || !Number.isFinite(requested)) return BASH_TIMEOUT_MS;
+  return Math.min(Math.max(Math.trunc(requested), BASH_TIMEOUT_MIN_MS), BASH_TIMEOUT_MAX_MS);
+}
+
+/** What a command may actually get when the run's wall clock ends in
+ *  `remainingMs`: `unchanged` when the wanted budget fits before the reserve,
+ *  `clipped` to what fits (with the line the model sees so it knows why the
+ *  command ended early), or `exhausted` when even the 1s floor does not fit —
+ *  the tool then refuses to start a command that cannot finish, and the model
+ *  writes up what it has. Without the reserve one first command at the
+ *  20-minute ceiling can consume most of a 25-minute run budget and leave
+ *  nothing for the write-up. */
+export type RunBudget =
+  { kind: "unchanged" } | { kind: "clipped"; timeoutMs: number; note: string } | { kind: "exhausted"; note: string };
+
+export function bashBudgetWithinRun(wantedMs: number, remainingMs: number): RunBudget {
+  const secs = (ms: number) => Math.max(0, Math.round(ms / 1000));
+  const cap = Math.trunc(remainingMs - RUN_DEADLINE_RESERVE_MS);
+  if (cap < BASH_TIMEOUT_MIN_MS) {
+    return {
+      kind: "exhausted",
+      note:
+        `run budget exhausted — ${secs(remainingMs)}s of wall clock left, inside the ` +
+        `${secs(RUN_DEADLINE_RESERVE_MS)}s write-up reserve, so the command was not run; write up what you have now`,
+    };
+  }
+  if (cap >= wantedMs) return { kind: "unchanged" };
+  return {
+    kind: "clipped",
+    timeoutMs: cap,
+    note: `[timeout clipped to ${secs(cap)}s — the run's wall clock ends in ${secs(remainingMs)}s]`,
+  };
+}
+
+/** The line a timed-out command shows the model: names the limit that fired
+ *  and the knob that raises it, so the model can self-correct (re-run with a
+ *  larger timeoutMs, split the command, or background it) instead of guessing
+ *  at a generic abort. */
+export function bashTimeoutNote(timeoutMs: number): string {
+  return (
+    `aborted at the ${Math.round(timeoutMs / 1000)}s command timeout ` +
+    `(pass the bash tool's timeoutMs for longer commands, max ${BASH_TIMEOUT_MAX_MS} ms)`
+  );
+}

package/dist/assets/src/execution/bindingPurge.ts

@@ -0,0 +1,43 @@
+// The resident keeps a thread binding after eviction by design
+// (docs/reference/specs/resident-repos.md item 23): the ref stays sticky for the next
+// attach. A load run that attaches fifty synthetic threads therefore leaves
+// fifty evicted rows on the resident's detail page forever. `POST /debug
+// {op: "purge-bindings", prefix}` (item 60) deletes exactly the bindings a
+// harness created and no longer needs. This is its decision, pure and
+// imported by the resident Worker like `residentDiskBudget.ts`.
+
+export interface PurgeableBinding {
+  threadKey: string;
+  evicted?: boolean;
+}
+
+export type PurgeDecision = { ok: true; purge: string[]; keptLive: string[] } | { ok: false; error: string };
+
+/** Thread-key namespaces real channels mint (docs/reference/specs/http-ingress.md item 1,
+ *  slack, mcp, the CLI). A purge is for synthetic keys only. */
+const PRODUCTION_NAMESPACES = new Set(["slack", "http", "mcp", "cli"]);
+
+/** A prefix must be a whole namespace (`load:`) or longer (`load:r1:`), never
+ *  empty, never a bare partial namespace, never a production namespace. */
+export function selectBindingsToPurge(bindings: readonly PurgeableBinding[], prefix: string): PurgeDecision {
+  const m = /^([a-z][a-z0-9-]{0,31}):/.exec(prefix);
+  if (!m)
+    return {
+      ok: false,
+      error: `prefix must name a whole thread-key namespace, like "load:" (got ${JSON.stringify(prefix)})`,
+    };
+  if (PRODUCTION_NAMESPACES.has(m[1])) {
+    return {
+      ok: false,
+      error: `prefix ${JSON.stringify(prefix)} is a production namespace; a purge is for synthetic thread keys only`,
+    };
+  }
+  const purge: string[] = [];
+  const keptLive: string[] = [];
+  for (const b of bindings) {
+    if (!b.threadKey.startsWith(prefix)) continue;
+    if (b.evicted === true) purge.push(b.threadKey);
+    else keptLive.push(b.threadKey);
+  }
+  return { ok: true, purge, keptLive };
+}

package/dist/assets/src/execution/residentBackupTransfer.ts

@@ -0,0 +1,50 @@
+/** Which way a resident's snapshot bytes travel (docs/reference/specs/resident-repos.md
+ *  item 61), kept pure so the decision is a unit test and the Worker only
+ *  reads it.
+ *
+ *  The Sandbox SDK has two transfer modes for `createBackup` / `restoreBackup`:
+ *  - `localBucket: true` — the Durable Object reads the archive from the R2
+ *    binding and pumps it to the container over the control RPC (and the
+ *    reverse on upload). The SDK documents this as the LOCAL-DEVELOPMENT mode
+ *    ("required for local development where presigned URLs and FUSE are
+ *    unavailable"). It puts a 128 MB isolate in the data path of every
+ *    transfer: a checkout restore larger than the isolate's memory (a full
+ *    checkout runs to gigabytes) resets the isolate mid-transfer, and a
+ *    resident whose restore was reset sits in `restoring` with no way out but
+ *    a manual rebuild.
+ *  - presigned — the DO signs GET/PUT URLs; the container downloads
+ *    (`downloadBackupParallel`, resumable) and uploads the bytes itself. The
+ *    DO orchestrates and judges; its memory no longer scales with the archive.
+ *    This is the mode the SDK's `requirePresignedURLSupport` gates on four env
+ *    values, named below exactly as it reads them.
+ *
+ *  Fail-closed: any input absent → local mode, with the gap named, so a
+ *  misprovisioned secret degrades to today's behavior loudly instead of
+ *  breaking every snapshot. The mode is reported on /healthz for the receipt. */
+
+export const PRESIGNED_BACKUP_ENV = [
+  "CLOUDFLARE_ACCOUNT_ID",
+  "BACKUP_BUCKET_NAME",
+  "R2_ACCESS_KEY_ID",
+  "R2_SECRET_ACCESS_KEY",
+] as const;
+
+export type BackupTransferMode = "presigned" | "local";
+
+export interface BackupTransferDecision {
+  mode: BackupTransferMode;
+  /** What to pass the SDK as `localBucket`. */
+  localBucket: boolean;
+  /** The presigned inputs that are absent or blank, in PRESIGNED_BACKUP_ENV order; empty in presigned mode. */
+  missing: string[];
+}
+
+export function backupTransferMode(env: Record<string, unknown>): BackupTransferDecision {
+  const missing = PRESIGNED_BACKUP_ENV.filter((name) => {
+    const v = env[name];
+    return typeof v !== "string" || v.trim() === "";
+  });
+  return missing.length === 0
+    ? { mode: "presigned", localBucket: false, missing: [] }
+    : { mode: "local", localBucket: true, missing: [...missing] };
+}