UNPKG

@tanstack/ai-sandbox

Version:

Provider-agnostic sandbox layer for TanStack AI — run harness adapters inside isolated sandboxes (defineSandbox, defineWorkspace, withSandbox) with a uniform SandboxHandle, workspace bootstrap, policy, and resumable lifecycle.

139 lines (138 loc) 7.39 kB
import { ProcessOptions, SandboxHandle } from './contracts.js'; import { RunStore } from '@tanstack/ai'; export interface SpawnNdjsonOptions extends ProcessOptions { /** * Called for each raw stdout line that is non-empty but fails JSON parsing * (e.g. a CLI banner). Defaults to ignoring it. Stderr is never parsed. */ onNonJsonLine?: (line: string) => void; /** * Written to the process stdin (then stdin is closed) right after spawn — * e.g. the agent prompt for `claude -p`. Avoids putting the prompt in argv. */ input?: string; /** * Route the agent's stdout through an in-sandbox journal rather than holding * its pipe directly. * * This is what makes a run survive host death: with nothing piped, there is * no reader whose disappearance can SIGPIPE the agent, and a later host reads * the same file from byte 0. Opt-in so every existing caller (and every * existing test) is unaffected. */ journal?: JournalOptions; } /** Journaling configuration for {@link spawnNdjson}. */ export interface JournalOptions { /** Run id the journal path is derived from. Must match across hosts. */ runId: string; /** Journal directory. Defaults to `/tmp/tanstack-runs`. */ dir?: string; /** * Read an EXISTING journal instead of starting the agent. The read still * begins at byte 0 — the alignment step, not the reader, decides what has * already been delivered. */ attach?: boolean; /** Poll interval for providers that cannot follow. */ pollIntervalMs?: number; /** * Run record store, consulted ONLY on an attach and only when the journal is * absent, to tell "not written yet" from "will never be written" (see * `attach-preflight.ts`). Optional so an attach with no store wired keeps the * bounded wait while losing the unknown/terminal classification. */ runs?: RunStore; /** * Bounded wait for a live run's journal to appear on an attach, AND — on every * path, attach or fresh — the bound on the read's first byte * (`ReadJournalOptions.firstByteTimeoutMs`). One knob for both because they * bound the same question from two sides: the preflight covers "the file does * not exist", the read covers "the file exists but nothing is writing to it", * and a caller that widens one always means to widen the other. * * Defaults to `DEFAULT_ATTACH_JOURNAL_WAIT_MS`. */ attachWaitMs?: number; } type JournaledOptions = SpawnNdjsonOptions & { journal: JournalOptions; }; /** Split a stream of arbitrary string chunks into complete lines. */ export declare function toLines(chunks: AsyncIterable<string>): AsyncIterable<string>; /** * Start the agent with its stdout (and the `{"__exit":N}` sentinel) redirected * into the journal, then return. * * Deliberately does NOT wait for the process and does NOT read its stdout: the * whole point of journaling is that the host holds no handle on the agent's * output, so a host that dies mid-run cannot take the agent down with it (no * pipe to SIGPIPE). The spawned process is left running in the sandbox; the * sentinel line the wrapper appends on exit is how anyone — this host or a * successor — learns it finished. Stdin is still written directly to the * spawned process, exactly as the unjournaled path does, since that transport * is unaffected by where stdout goes. */ export declare function startJournaledAgent(handle: SandboxHandle, command: string, options: JournaledOptions): Promise<void>; /** * Read a run's journal and yield each line parsed as JSON. * * Always reads from byte 0 — the alignment step (a later phase), not this * reader, decides what a client has already seen. Stops at the `{"__exit":N}` * sentinel, and throws for a non-zero N so the calling adapter's existing * `catch` turns it into a `RUN_ERROR`, the same observable outcome the * unjournaled path produces from a non-zero `wait()`. There is nothing to * `wait()` on here: the host holds a `tail`, not the agent process, so the * sentinel line IS the exit code. * * The sentinel is also what bounds journal growth: reaching it means the run is * terminal, and a terminal run's record is the event log, so both journal files * are deleted before this iterable finishes. The ordering below is load-bearing * and is asserted, not merely commented: * * - The sentinel is captured and the loop is `break`-ed, so the source's * `finally` kills the `tail` BEFORE the `rm` runs — the reader is stopped, then * its input is deleted, never the other way round. * - `exitCode` stays `undefined` if the journal stream ends without a sentinel. * NOTHING is deleted on that path either way — the run may be mid-flight and a * successor host may still need every byte — but the two causes are then * separated by `options.signal.aborted`: an aborted consumer returns quietly, * while a stream that died on its own (killed `tail`, destroyed sandbox, torn * pipe) THROWS. Returning for both is how a truncated read used to reach the * client as a normally-completing run. * - A non-zero sentinel deletes too. The run is terminal either way. * - The stderr sidecar is read BEFORE the deletion that destroys it, so a * non-zero exit carries up to {@link STDERR_ERROR_CHARS} chars of the agent's * own diagnostics, exactly as the unjournaled path below does. That closes the * "Known regression" this function used to document; the read is bounded and * failure-swallowing (see {@link readStderrTail}), so it cannot turn a run * failure into a cleanup failure. * * On an ATTACH (`journal.attach === true`) the read is preceded by * {@link awaitAttachableJournal}, which fails fast for a runId the store does not * know or has already terminalized and otherwise waits a BOUNDED time for a live * run's journal to appear. Without it, an attach to a runId with no journal * created an empty one (`journalFollowCommand` does that deliberately) and tailed * it forever — no sentinel, no error, no timeout. * * One case this does NOT bound: a run that reaches its sentinel while DETACHED * has no host reading it, so nothing observes the sentinel and nothing here * runs. Sweeping those is `pruneJournals`' job (`journal-sweep.ts`): it consults * the run store's status for each journal it finds and deletes only the terminal * ones, from a cron the application schedules rather than from a run. */ export declare function readJournalNdjson(handle: SandboxHandle, options: JournaledOptions): AsyncIterable<unknown>; /** * Spawn `command` in the sandbox and yield each stdout line parsed as JSON. * * Without `options.journal`, behavior is byte-identical to before: resolves * the spawn handle's exit via `wait()` after stdout closes; a non-zero exit * with no events surfaced is the adapter's concern to detect. * * With `options.journal`, the agent's stdout is redirected into an in-sandbox * journal (unless `journal.attach` is set, meaning a run already in flight) * and then read back from byte 0 — one code path for a fresh run and an * attach, both going through {@link readJournalNdjson}. */ export declare function spawnNdjson(handle: SandboxHandle, command: string, options?: SpawnNdjsonOptions): AsyncIterable<unknown>; export {};