@agent-compose/sdk 0.6.0 → 0.8.0

package/src/step-invocation/invoker.ts

@@ -25,7 +25,7 @@ import { randomBytes } from "node:crypto";
 import { z } from "zod";
 import { SandboxUnavailableError } from "../sandbox-errors.js";
 import type { SandboxProvider, SandboxBackgroundProcess } from "../types/sandbox.js";
-import { RUNNER_COMMAND, STEP_ENV, stepResultLinePrefix, stepPauseLinePrefix, requestContextPath, stepInputPath, stepResultFilePath } from "./protocol.js";
+import { RUNNER_COMMAND, STEP_ENV, stepResultLinePrefix, stepPauseLinePrefix, requestContextPath, stepInputPath, stepResultFilePath, stepLogFilePath, stepPidFilePath } from "./protocol.js";
 import { StepPauseRequestSchema } from "./types.js";
 import type { StepRequest, StepResult } from "./types.js";
 import type { StepObservability } from "../workflow-steps/observability.js";
@@ -194,6 +194,18 @@ export interface InvokeStepOptions {
    *  the caller's job; the activity batches and inserts at step completion. */
   onStdout?: (line: string) => void;
   onStderr?: (line: string) => void;
+  /** Called when the LIVE output stream fails mid-run (e.g. E2B's connect-web
+   *  transport throws `received unsupported compressed output` on a large
+   *  compressed frame) and the invoker falls back to recovering the result +
+   *  full logs from the durable token-keyed files. The run is NOT failed — this
+   *  is the hook to emit a structured alert so the degradation is visible/paged.
+   *  Best-effort: keep it cheap and non-throwing. */
+  onStreamDegraded?: (info: { error: unknown; runnerPid: number; resultToken: string }) => void;
+  /** Ceiling on the durable-result recovery poll after a live-stream fault
+   *  (default 45m — the wedged-runner backstop). The activity passes its step
+   *  ceiling so a stream fault on a step with hours of legitimate work left
+   *  recovers instead of timing out; startToClose is the real wall. */
+  recoveryDeadlineMs?: number;
 }
 
 /** A step running as a background command (ADR-0028). The activity races its
@@ -219,13 +231,18 @@ export interface RunningStep<TOutput = unknown> {
 function makeStreamSplitters(
   opts: Pick<InvokeStepOptions, "onStdout" | "onStderr"> | undefined,
   resultToken: string,
-): { onStdout?: (chunk: string) => void; onStderr?: (chunk: string) => void; flush: () => void } {
+): { onStdout?: (chunk: string) => void; onStderr?: (chunk: string) => void; flush: () => void; stdoutConsumedChars: () => number } {
   const resultSentinel = stepResultLinePrefix(resultToken);
   const pauseSentinel  = stepPauseLinePrefix(resultToken);
   const isSentinel = (line: string) => line.startsWith(resultSentinel) || line.startsWith(pauseSentinel);
   const make = (sink: ((line: string) => void) | undefined, filterSentinel: boolean) => {
-    if (!sink) return { onChunk: undefined, flush: () => {} };
+    if (!sink) return { onChunk: undefined, flush: () => {}, consumed: () => 0 };
     let buf = "";
+    // Chars consumed as WHOLE lines (incl. each trailing "\n"), counting sentinel
+    // lines too — this is the offset into the tee'd log file up to which the live
+    // stream has already delivered. The recovery backfill starts exactly here, so
+    // an in-flight partial line is re-read whole from the file (no split, no dup).
+    let consumed = 0;
     return {
       onChunk: (chunk: string) => {
         buf += chunk;
@@ -233,6 +250,7 @@ function makeStreamSplitters(
         while ((nl = buf.indexOf("\n")) !== -1) {
           const line = buf.slice(0, nl);
           buf = buf.slice(nl + 1);
+          consumed += line.length + 1;
           if (filterSentinel && isSentinel(line)) continue;
           sink(line);
         }
@@ -241,9 +259,11 @@ function makeStreamSplitters(
         if (buf.length === 0) return;
         const line = buf;
         buf = "";
+        consumed += line.length;
         if (filterSentinel && isSentinel(line)) return;
         sink(line);
       },
+      consumed: () => consumed,
     };
   };
   const stdout = make(opts?.onStdout, true);
@@ -252,6 +272,7 @@ function makeStreamSplitters(
     ...(stdout.onChunk ? { onStdout: stdout.onChunk } : {}),
     ...(stderr.onChunk ? { onStderr: stderr.onChunk } : {}),
     flush: () => { stdout.flush(); stderr.flush(); },
+    stdoutConsumedChars: () => stdout.consumed(),
   };
 }
 
@@ -307,6 +328,174 @@ async function classifyRunnerOutcome<TOutput>(
   };
 }
 
+/** Poll the runner's durable result file (token-keyed, written immediately before
+ *  exit on EVERY path — success, failure, pause) as the AUTHORITATIVE completion
+ *  signal, using `/proc/<pid>` liveness to tell work-in-progress from a crash.
+ *  Used wherever the live stream cannot be trusted for completion: the native
+ *  resume path (a reconnected handle's `wait()` can resolve early after a VM
+ *  pause/resume) AND the launch path after a live-stream transport fault.
+ *  `cat`-ing the result file is itself immune to the large-frame compression bug
+ *  — it is a single small JSON line, far below any compression threshold. The
+ *  activity's `startToCloseTimeout` is the real upper bound; the deadline here
+ *  is a backstop so a wedged runner can't leak this loop in the worker forever
+ *  (default 45m; a worker-death recovery reconnecting mid-step passes the full
+ *  step ceiling instead — hours of legitimate work may remain). */
+async function pollDurableResult<TOutput>(
+  sandbox: SandboxProvider,
+  runnerPid: number,
+  resultToken: string,
+  opts?: { signal?: AbortSignal; deadlineMs?: number },
+): Promise<StepResult<TOutput>> {
+  const signal = opts?.signal;
+  const resultFile = stepResultFilePath(resultToken);
+  const POLL_MS = 1000;
+  const deadlineMs = opts?.deadlineMs ?? 45 * 60_000;
+  const DEADLINE = Date.now() + deadlineMs;
+  // Both internal commands carry an explicit budget. Required for Vercel:
+  // without `timeoutMs` there is no hard client-side deadline in the provider's
+  // `commands.run`, so a wedged VM hangs one `cat` forever and the poll
+  // deadline never advances. A timeout throws → the `.catch(() => null)`
+  // treats it as a failed read/probe and the loop continues (a hung probe
+  // never misclassifies as dead). Harmless on E2B — a 1-line cat is ms-fast.
+  const CMD_TIMEOUT_MS = 10_000;
+  // A destroyed sandbox (e.g. run cancelled → sandbox killed mid-recovery)
+  // rejects EVERY command, which a lone `.catch(() => null)` renders
+  // indistinguishable from a live runner — the loop would spin until the
+  // deadline (hours) against a VM that no longer exists. Track consecutive
+  // iterations where BOTH the result read and the liveness probe rejected and
+  // exit with an honest runner-exit once the streak is unambiguous. 20
+  // iterations ≈ 30s–7min of continuous rejections (each command may burn its
+  // 10s budget) — long enough to ride out a transient provider-API blip,
+  // short enough that a dead sandbox doesn't hold the worker slot for hours.
+  const UNREACHABLE_LIMIT = 20;
+  let unreachableStreak = 0;
+  const readResult = async (): Promise<StepResult<TOutput> | "unreachable" | null> => {
+    const r = await sandbox.commands.run(`cat ${resultFile} 2>/dev/null || true`, { timeoutMs: CMD_TIMEOUT_MS }).catch(() => "unreachable" as const);
+    if (r === "unreachable") return "unreachable";
+    return r.stdout ? (parseStepResult<TOutput>(r.stdout, resultToken) ?? null) : null;
+  };
+  for (let attempt = 1; ; attempt++) {
+    // Aborted = the activity already resolved via the pause watcher (a re-pause);
+    // this poll's result is now unobserved. Return (never throw — the promise is
+    // no longer awaited) so it settles cleanly with no unhandled rejection.
+    if (signal?.aborted) {
+      return { ok: false, error: { kind: "runner-exit", message: "reconnectStep aborted (run re-paused)", exitCode: 1 } };
+    }
+    const fromFile = await readResult();
+    if (fromFile !== null && fromFile !== "unreachable") return fromFile;
+
+    const probe = await sandbox.commands
+      .run(`test -d /proc/${runnerPid} && echo alive || echo dead`, { timeoutMs: CMD_TIMEOUT_MS })
+      .catch(() => null);
+    if (fromFile === "unreachable" && probe === null) {
+      unreachableStreak++;
+      if (unreachableStreak >= UNREACHABLE_LIMIT) {
+        return {
+          ok: false,
+          error: {
+            kind: "runner-exit",
+            message: `sandbox stopped responding while waiting for runner pid ${runnerPid} to emit a result (${UNREACHABLE_LIMIT} consecutive failed polls)`,
+            exitCode: 1,
+          },
+        };
+      }
+    } else {
+      unreachableStreak = 0;
+    }
+    const dead = (probe?.stdout ?? "").includes("dead");
+    process.stderr.write(
+      `[recover] poll ${attempt}: result-file=absent runner=${dead ? "dead" : "alive"} pid=${runnerPid}\n`,
+    );
+    if (dead) {
+      // Close the write-then-exit race with one final read, else classify a crash.
+      const finalRead = await readResult();
+      if (finalRead !== null && finalRead !== "unreachable") return finalRead;
+      return {
+        ok: false,
+        error: { kind: "runner-exit", message: `runner pid ${runnerPid} exited without writing a result file`, exitCode: 1 },
+      };
+    }
+    if (Date.now() > DEADLINE) {
+      return {
+        ok: false,
+        error: {
+          kind: "runner-exit",
+          message: `timed out after ${Math.round(deadlineMs / 60_000)}m waiting for runner pid ${runnerPid} to emit a result`,
+          exitCode: 1,
+        },
+      };
+    }
+    await new Promise((r) => setTimeout(r, POLL_MS));
+  }
+}
+
+/** Recover a step's result AND its full logs after the live output stream died
+ *  mid-run (the dominant cause being E2B's connect-web transport throwing on a
+ *  compressed large frame — a LOG-TRANSPORT fault, not a runner failure). Waits
+ *  for the runner to actually finish (`pollDurableResult`), then reads the tee'd
+ *  log file back over the HTTP file transport (compression-immune) and backfills
+ *  only the lines the dead stream never delivered — everything past the last
+ *  whole line already emitted live (`stdoutConsumedChars`), so no duplication and
+ *  no split lines. Logs BEFORE the fault are already persisted by the activity's
+ *  live flush; this restores the tail so "what happened" survives intact. */
+async function recoverLogsAndResult<TOutput>(
+  sandbox: SandboxProvider,
+  runnerPid: number,
+  resultToken: string,
+  liveSplitters: { stdoutConsumedChars: () => number },
+  opts: Pick<InvokeStepOptions, "onStdout" | "onStderr"> | undefined,
+  pollOpts?: { deadlineMs?: number },
+): Promise<StepResult<TOutput>> {
+  const result = await pollDurableResult<TOutput>(sandbox, runnerPid, resultToken, pollOpts);
+  // Runner has exited → the tee'd log file is complete. Best-effort: a failed
+  // readback just means the recovered run keeps the live logs it already had.
+  if (sandbox.files.read && opts?.onStdout) {
+    const fullLog = await sandbox.files.read(stepLogFilePath(resultToken)).catch(() => null);
+    if (fullLog != null) {
+      const already = liveSplitters.stdoutConsumedChars();
+      const tail = fullLog.length > already ? fullLog.slice(already) : "";
+      if (tail.length > 0) {
+        const backfill = makeStreamSplitters(opts, resultToken);
+        backfill.onStdout?.(tail);
+        backfill.flush();
+      }
+    }
+  }
+  return result;
+}
+
+/** Read the foreground launch's pidfile back — the /proc liveness handle for
+ *  the recovery poll. Returns `null` when recovery is impossible, for either
+ *  reason: the sandbox cannot run ANY command (genuinely dead VM), or the
+ *  pidfile never appeared. The pidfile write is the FIRST statement of the
+ *  runner pipeline, so a still-absent pid after the retries means the runner
+ *  never started and no result file will ever exist — recovery would poll
+ *  blind for the full step ceiling and then misclassify; the caller must
+ *  rethrow the original terminal error instead (fast, honest). The retries
+ *  cover both a transiently-unreachable sandbox and the launch race where the
+ *  stream fault fired before the VM executed the pidfile write. */
+async function readRunnerPid(
+  sandbox: SandboxProvider,
+  resultToken: string,
+): Promise<number | null> {
+  const ATTEMPTS = 3;
+  const RETRY_DELAY_MS = 2_000;
+  for (let attempt = 1; attempt <= ATTEMPTS; attempt++) {
+    try {
+      const r = await sandbox.commands.run(
+        `cat ${stepPidFilePath(resultToken)} 2>/dev/null || true`,
+        { timeoutMs: 10_000 },
+      );
+      const pid = Number.parseInt(r.stdout.trim(), 10);
+      if (Number.isInteger(pid) && pid > 0) return pid;
+    } catch {
+      // Sandbox unreachable this attempt — fall through to the retry delay.
+    }
+    if (attempt < ATTEMPTS) await new Promise((r) => setTimeout(r, RETRY_DELAY_MS));
+  }
+  return null;
+}
+
 /** Write the step's input + request-context files and build the runner env.
  *  Shared by the foreground and background launch paths. Returns the
  *  per-invocation `resultToken` + the env map. */
@@ -337,20 +526,60 @@ export async function invokeStep<TOutput = unknown>(
 
   let result: { stdout: string; stderr: string; exitCode: number };
   try {
-    result = await sandbox.commands.run(RUNNER_COMMAND, {
-      envs,
-      timeoutMs: 0,
-      ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
-      ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
-    });
+    // Run the agent as ROOT (ADR-0019 follow-up). The factory drive (Archil)
+    // presents its S3-synced files root-owned and exposes no uid-mapped mount,
+    // so a non-root agent EACCESes on every shared file — the reason for the
+    // expensive boot-time `chmod -R` walk. As root the agent writes them
+    // directly: the walk disappears entirely. Egress is edge-enforced with NO
+    // root exemption (see sandbox.ts), so root is confined exactly like the
+    // non-root user. `HOME=/root` so the spawned `claude` finds root's skills.
+    //
+    // Same pipeline pattern as `launchStep`: tee the runner's stdout to the
+    // durable token-keyed log file so a live-stream fault can backfill the
+    // undelivered tail, and write the wrapping shell's pid (`$$`) to a pidfile
+    // first — the shell waits on the pipeline, so `/proc/<pid>` liveness is a
+    // correct "runner may still write a result" proxy for the recovery poll.
+    // `set -o pipefail` is REQUIRED: without it the pipeline's exit code is
+    // tee's (0), masking a non-zero runner exit and breaking the runner-exit
+    // classification (the provider execs via `sh -c`, where pipefail works).
+    result = await sandbox.commands.run(
+      `set -o pipefail; echo $$ > ${stepPidFilePath(resultToken)}; ${RUNNER_COMMAND} | tee ${stepLogFilePath(resultToken)}`,
+      {
+        envs: { ...envs, HOME: "/root", IS_SANDBOX: "1" },
+        sudo: true,
+        timeoutMs: 0,
+        ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
+        ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
+      },
+    );
   } catch (e) {
     // Typed sandbox-infrastructure failure: the engine's recovery contract
-    // (re-provision when retryable, honest terminal classification otherwise)
-    // keys on this error propagating INTACT — its `[sandbox-unavailable:*]`
-    // message prefix must reach the workflow as the failure leaf, not ride an
-    // embedded stderr tail that a later slice(-2000) can drop. Rethrow before
-    // the CommandExitError downgrade below.
-    if (e instanceof SandboxUnavailableError) throw e;
+    // keys on retryability. Its `[sandbox-unavailable:*]` message prefix must
+    // reach the workflow as the failure leaf, not ride an embedded stderr tail
+    // that a later slice(-2000) can drop. Handle before the CommandExitError
+    // downgrade below.
+    if (e instanceof SandboxUnavailableError) {
+      // Pre-launch (retryable): nothing user-side ran — propagate INTACT so the
+      // workflow layer re-provisions (identity + prefix contract; pinned test).
+      if (e.retryable) throw e;
+      // Post-launch terminal: the LIVE stream/wait call died (e.g. Vercel's raw
+      // "The operation timed out.") while the runner is very likely still alive
+      // and will write its durable result file. Recover instead of failing —
+      // NEVER re-launch the runner. A null pid means recovery is impossible
+      // (sandbox can't run ANY command, or the pidfile never appeared because
+      // the runner never actually started): rethrow the original terminal
+      // classification fast rather than polling blind for the step ceiling.
+      // We do NOT flush the live splitter here — the recovery backfills from
+      // the last whole line, re-reading any in-flight partial line whole from
+      // the durable log (avoids a split/duplicated line).
+      const runnerPid = await readRunnerPid(sandbox, resultToken);
+      if (runnerPid === null) throw e;
+      opts?.onStreamDegraded?.({ error: e, runnerPid, resultToken });
+      return recoverLogsAndResult<TOutput>(
+        sandbox, runnerPid, resultToken, splitters, opts,
+        opts?.recoveryDeadlineMs !== undefined ? { deadlineMs: opts.recoveryDeadlineMs } : undefined,
+      );
+    }
     // Some providers (E2B) throw a CommandExitError on a non-zero exit instead of
     // returning it. Recover stdout/stderr/exitCode from the error so we can STILL
     // parse the runner's structured `__AC_STEP_RESULT__` payload — otherwise the
@@ -384,8 +613,25 @@ export async function launchStep<TOutput = unknown>(
   }
   const { resultToken, envs } = await prepareStepLaunch(sandbox, request, opts);
   const splitters = makeStreamSplitters(opts, resultToken);
-  const proc = await sandbox.commands.runBackground(RUNNER_COMMAND, {
-    envs,
+  // Tee the runner's stdout to a durable, token-keyed LOG file (stderr stays a
+  // separate live stream). The live output rides E2B's connect-web command
+  // stream, which THROWS on a compressed large frame (gRPC-web cannot decode
+  // message compression) and kills the feed mid-run. The tee'd file, read back
+  // over the envd HTTP transport (compression-immune), lets `wait()` recover the
+  // full logs + result instead of failing a run that actually completed. `tee`
+  // runs in the same `bash -c` pipeline E2B already wraps the command in, so the
+  // background pid (the pause / reconnect-by-pid handle) is unchanged; the result
+  // sentinel still rides stdout (tee passes it through) and the result FILE is
+  // written by the runner directly, so the happy path is untouched. `set -o
+  // pipefail` is REQUIRED: without it the pipeline's exit code is tee's (0),
+  // masking a non-zero runner exit and breaking the runner-exit classification;
+  // pipefail propagates the runner's code (E2B execs via `bash -c`).
+  //
+  // Run the agent as ROOT — see invokeStep above. `sudo:true` → user:"root";
+  // `HOME=/root` so the spawned `claude` finds root's skills.
+  const proc = await sandbox.commands.runBackground(`set -o pipefail; ${RUNNER_COMMAND} | tee ${stepLogFilePath(resultToken)}`, {
+    envs: { ...envs, HOME: "/root", IS_SANDBOX: "1" },
+    sudo: true,
     timeoutMs: 0,
     ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
     ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
@@ -394,9 +640,25 @@ export async function launchStep<TOutput = unknown>(
     runnerPid:   proc.pid,
     resultToken,
     async wait() {
-      const result = await proc.wait();
-      splitters.flush();
-      return classifyRunnerOutcome<TOutput>(sandbox, result, resultToken);
+      try {
+        const result = await proc.wait();
+        splitters.flush();
+        return classifyRunnerOutcome<TOutput>(sandbox, result, resultToken);
+      } catch (e) {
+        // A genuine infra death must propagate so the engine re-provisions.
+        if (e instanceof SandboxUnavailableError) throw e;
+        // Otherwise the runner's LIVE output stream died while the runner itself
+        // is alive and writing its durable result + tee'd log files. The dominant
+        // cause is E2B's connect-web transport throwing on a compressed large
+        // frame ("received unsupported compressed output") — a LOG-TRANSPORT
+        // fault, NEVER a reason to fail a run that completed. Degrade the live
+        // feed, alert, and recover the result + full logs over the HTTP file
+        // transport. We do NOT flush the live splitter here — the recovery
+        // backfills from the last whole line, re-reading any in-flight partial
+        // line whole from the durable log (avoids a split/duplicated line).
+        opts?.onStreamDegraded?.({ error: e, runnerPid: proc.pid, resultToken });
+        return recoverLogsAndResult<TOutput>(sandbox, proc.pid, resultToken, splitters, opts);
+      }
     },
   };
 }
@@ -414,38 +676,61 @@ export async function launchStep<TOutput = unknown>(
 export async function reconnectStep<TOutput = unknown>(
   sandbox: SandboxProvider,
   resume: { runnerPid: number; resultToken: string; stepIndex: number },
-  opts?: Pick<InvokeStepOptions, "onStdout" | "onStderr">,
+  opts?: Pick<InvokeStepOptions, "onStdout" | "onStderr"> & {
+    signal?: AbortSignal;
+    /** Ceiling on the durable-result poll (default 45m — the wedged-runner
+     *  backstop). A worker-death recovery reconnecting MID-step passes the
+     *  step ceiling instead: hours of legitimate work may remain, and the
+     *  activity's per-attempt startToClose already bounds it server-side. */
+    pollDeadlineMs?: number;
+  },
 ): Promise<RunningStep<TOutput>> {
   if (!sandbox.commands.connectProcess) {
     throw new Error("reconnectStep requires a provider with background-command support (commands.connectProcess)");
   }
   const { runnerPid, resultToken } = resume;
   const splitters = makeStreamSplitters(opts, resultToken);
-  let proc: SandboxBackgroundProcess | null = null;
+  const signal = opts?.signal;
+  const pollDeadlineMs = opts?.pollDeadlineMs;
+
+  // Re-attach to the suspended runner for LIVE stdout streaming only. This is
+  // best-effort: a reconnect that throws after the native resume just means no
+  // live feed for the dashboard — completion is read from the durable result
+  // file below, NOT from this handle.
+  let attached: SandboxBackgroundProcess | null = null;
   try {
-    proc = await sandbox.commands.connectProcess(runnerPid, {
+    attached = await sandbox.commands.connectProcess(runnerPid, {
       ...(splitters.onStdout ? { onStdout: splitters.onStdout } : {}),
       ...(splitters.onStderr ? { onStderr: splitters.onStderr } : {}),
     });
   } catch (e) {
-    // A sandbox-infrastructure failure must propagate intact (recovery
-    // contract). Anything else means the runner already exited during resume
-    // (we re-attached too late) — its durable result file is written, so the
-    // handle's wait() classifies from the file below.
     if (e instanceof SandboxUnavailableError) throw e;
   }
-  const attached = proc;
+
   return {
     runnerPid,
     resultToken,
     async wait() {
-      let output: { stdout: string; stderr: string; exitCode: number } = { stdout: "", stderr: "", exitCode: 0 };
-      if (attached) {
-        try { output = await attached.wait(); }
-        catch (e) { if (e instanceof SandboxUnavailableError) throw e; }
+      // CRITICAL: do NOT trust `connect(pid).wait()` on the resume path. After a
+      // native VM pause/resume, E2B's re-attached handle can resolve its wait()
+      // EARLY — the reconnected stdout stream closes while the runner process is
+      // still alive and working — returning a default `{exitCode:0}`. That made
+      // the worker report "exited 0 without emitting" and force-kill a runner
+      // that was mid-iteration. So ignore the handle for completion and poll the
+      // runner's durable result file + `/proc` liveness instead — the same
+      // authoritative signal the launch-path recovery uses (`pollDurableResult`).
+      try {
+        const result = await pollDurableResult<TOutput>(sandbox, runnerPid, resultToken, {
+          ...(signal ? { signal } : {}),
+          ...(pollDeadlineMs !== undefined ? { deadlineMs: pollDeadlineMs } : {}),
+        });
+        splitters.flush();
+        return result;
+      } finally {
+        // Release the streaming handle (keeps it un-GC'd for the duration of the
+        // poll above; a dangling reconnect after a re-pause is cleaned up here).
+        await attached?.kill().catch(() => {});
       }
-      splitters.flush();
-      return classifyRunnerOutcome<TOutput>(sandbox, output, resultToken);
     },
   };
 }

package/src/step-invocation/protocol.ts

@@ -80,3 +80,22 @@ export function requestContextPath(stepIndex: number): string {
 export function stepResultFilePath(token: string): string {
   return `/tmp/wf/step-result-${token}.json`;
 }
+
+/** Sandbox-side pidfile the FOREGROUND launch writes (`echo $$ > …`) before
+ *  spawning the runner pipeline — the wrapping shell's pid. The recovery path
+ *  reads it back to probe `/proc/<pid>` liveness when the live stream/wait
+ *  call dies mid-step (providers with no background-process pid — Vercel). */
+export function stepPidFilePath(token: string): string {
+  return `/tmp/wf/step-pid-${token}.pid`;
+}
+
+/** Sandbox-side path where the runner's stdout is tee'd as a durable LOG file,
+ *  keyed by the per-invocation token. The live output rides E2B's connect-web
+ *  command stream, which THROWS on a compressed large frame (gRPC-web cannot
+ *  decode message compression) and kills the feed mid-run. This file, read back
+ *  over the envd HTTP file transport (compression-immune), lets the invoker
+ *  recover the FULL logs after such a fault instead of losing the tail — the
+ *  log-side analogue of `stepResultFilePath` for the result. */
+export function stepLogFilePath(token: string): string {
+  return `/tmp/wf/step-log-${token}.log`;
+}

package/src/types/api-compliance.ts

@@ -0,0 +1,79 @@
+/**
+ * Break-glass compliance session wire types (ADR-0051).
+ *
+ * A team admin has no default access to a scoped (Private/Shared) document.
+ * To view one they don't hold a grant on, they open a READ-ONLY,
+ * owner-approved, time-boxed compliance session over a scope (one factory, or
+ * the whole team). These are the shapes the server's `/api/v1/compliance/*`
+ * routes emit — `client.ts` is a thin typed wrapper over them.
+ */
+
+export type ComplianceScopeKind = "factory" | "team";
+export type ComplianceStatus = "pending" | "active" | "expired" | "revoked";
+
+/** One break-glass compliance session. `expiresAt` is null while pending —
+ *  the TTL clock starts at approval, not request. */
+export interface ComplianceSession {
+  id: string;
+  teamId: string;
+  requestedByUserId: string | null;
+  requestedByLabel: string | null;
+  approvedByUserId: string | null;
+  approvedByLabel: string | null;
+  reason: string;
+  scopeKind: ComplianceScopeKind;
+  /** The factory this session covers (factory scope); null for team scope. */
+  scopeRef: string | null;
+  status: ComplianceStatus;
+  ttlSeconds: number;
+  expiresAt: string | null;
+  revokedByUserId: string | null;
+  revokedByLabel: string | null;
+  createdAt: string;
+  approvedAt: string | null;
+  revokedAt: string | null;
+}
+
+/** One per-document access recorded under an active session — the individual
+ *  audit trail the session's reasoned entry authorizes. */
+export interface ComplianceAccess {
+  id: string;
+  sessionId: string;
+  /** Nulled when the accessed file is later deleted (ON DELETE SET NULL);
+   *  `filePath` is the durable snapshot that survives. */
+  fileId: string | null;
+  filePath: string;
+  accessedByUserId: string | null;
+  accessedByLabel: string | null;
+  accessedAt: string;
+}
+
+export interface RequestComplianceSessionInput {
+  scopeKind: ComplianceScopeKind;
+  /** Required for factory scope; must be omitted for team scope. */
+  scopeRef?: string | null;
+  reason: string;
+  /** Session lifetime once approved — 15 min floor, 7 day ceiling. */
+  ttlSeconds: number;
+}
+
+export interface ListComplianceSessionsOptions {
+  status?: ComplianceStatus;
+  limit?: number;
+  cursor?: string;
+}
+
+export interface ComplianceSessionsPage {
+  sessions: ComplianceSession[];
+  nextCursor: string | null;
+}
+
+export interface ListComplianceAccessesOptions {
+  limit?: number;
+  cursor?: string;
+}
+
+export interface ComplianceAccessesPage {
+  accesses: ComplianceAccess[];
+  nextCursor: string | null;
+}