@evident-ai/runner-cdk 3.4.1-dev.2f1b44b → 3.4.1-dev.51df4de

package/dist/microvm-image-context/hooks/common.sh

@@ -312,14 +312,21 @@ fetch_runner_secret() {
     log "runner secret is not configured; continuing without GitHub and MCP credentials"
     return 0
   fi
+  # From here, RUNNER_SECRET_ARN IS configured: every failure branch below
+  # returns 1, not 0. Process env is fixed at spawn — nothing later in this
+  # VM's life can retry a fetch that fails here — so a VM that was promised
+  # these credentials and never got them is contrary evidence it can do the
+  # job it was launched for, exactly the class check_runner_key already treats
+  # as fatal (#1997): fail /run now and let the lifecycle machinery replace
+  # this VM, rather than run the whole session silently degraded.
   remaining="$(remaining_credential_budget "${step_started_s}")"
   if [ "${remaining}" -lt 1 ]; then
     warn "CREDENTIAL-RESTORE-SKIPPED: runner-secret skipped; the ${CREDENTIAL_RESTORE_DEADLINE_SECONDS}s credential-restore budget is exhausted"
-    return 0
+    return 1
   fi
   if ! stderr_file="$(mktemp /dev/shm/runner-secret-stderr.XXXXXX)"; then
     warn "RUNNER-SECRET-STDERR-UNAVAILABLE: could not allocate diagnostic storage; continuing without runner credentials"
-    return 0
+    return 1
   fi
   started_ms="$(now_ms)"
   payload="$(timeout -k "${CREDENTIAL_RESTORE_KILL_GRACE_SECONDS}" "${remaining}" aws secretsmanager get-secret-value --secret-id "${RUNNER_SECRET_ARN}" --query SecretString --output text 2>"${stderr_file}")" || rc=$?
@@ -327,12 +334,12 @@ fetch_runner_secret() {
   if [ "${rc}" -eq 124 ] || [ "${rc}" -eq 137 ]; then
     warn "CREDENTIAL-RESTORE-TIMEOUT: runner-secret did not finish within its ${remaining}s share of the ${CREDENTIAL_RESTORE_DEADLINE_SECONDS}s credential-restore deadline; continuing without it"
     rm -f "${stderr_file}"
-    return 0
+    return 1
   fi
   if [ "${rc}" -ne 0 ]; then
     warn "RUNNER-SECRET-UNREADABLE: $(<"${stderr_file}")"
     rm -f "${stderr_file}"
-    return 0
+    return 1
   fi
   rm -f "${stderr_file}"
   if ! jq -e 'type == "object"' >/dev/null 2>&1 <<<"${payload}"; then
@@ -363,7 +370,10 @@ restore_credentials() {
 
   load_state_config || return 1
   # This shares the existing bounded window so /run's worst-case duration does not grow.
-  fetch_runner_secret "${step_started_s}"
+  # Fatal when RUNNER_SECRET_ARN is configured (see fetch_runner_secret's own
+  # comment); a no-ARN self-hosted runner already returned 0 above and never
+  # reaches this `||`.
+  fetch_runner_secret "${step_started_s}" || return 1
   bounded_restore_call restore-claude "${step_started_s}" restore claude || return 1
   bounded_restore_call restore-opencode "${step_started_s}" restore opencode || return 1
 
@@ -517,6 +527,35 @@ prewarm_litestream() {
   log "pre-warming ${binary} in the background"
 }
 
+# Reads the aws CLI v2 install tree into the page cache, in the background, so
+# fetch_runner_secret's first `aws secretsmanager get-secret-value` does not
+# pay first-touch I/O on the critical path.
+#
+# A single `cat` of the `aws` entrypoint (litestream's pattern above) is NOT
+# enough here: v2 ships as a real Python distribution (~7,500 files under the
+# resolved binary's own directory), and a cold invocation demand-pages a
+# scattered set of them (botocore's endpoints.json/partitions.json, service
+# model JSON, shared libs) — one boot measured #1997's fetch at 7,243ms cold
+# vs ~400ms warm, an ~18x gap the litestream read-ahead trick alone cannot
+# close. `find -exec cat` walks that whole tree instead of one file.
+prewarm_aws_cli() {
+  local binary tree
+  binary="$(command -v aws 2>/dev/null || true)"
+  if [ -z "${binary}" ]; then
+    warn "aws CLI is not on PATH; skipping the boot pre-warm"
+    return 0
+  fi
+  tree="$(dirname "$(readlink -f "${binary}")")"
+
+  (
+    local started_ms rc=0
+    started_ms="$(now_ms)"
+    find "${tree}" -type f -exec cat {} + >/dev/null 2>&1 || rc=$?
+    log_elapsed_since aws-cli-prewarm "${started_ms}" "${rc}"
+  ) &
+  log "pre-warming ${tree} in the background"
+}
+
 # Generates ${LITESTREAM_CONFIG_FILE} from runner-synchroniser's own renderer
 # — the SAME config module the credential restore/sync already goes through,
 # so there is no second copy of the bucket/prefix/path logic to drift

package/dist/microvm-image-context/hooks/run

@@ -54,8 +54,12 @@ printf '%s\n' "${state_prefix}" >"${STATE_PREFIX_FILE}"
 
 # 3 — start reading the litestream binary NOW so that read overlaps step 4,
 # which needs a different binary (node), instead of landing inside step 5, which
-# start_opencode has to wait for. Backgrounded and never waited on.
+# start_opencode has to wait for. Backgrounded and never waited on. Same for the
+# aws CLI: step 4's runner-secret fetch is its first-ever invocation in this
+# VM's life, and #1997 measured that cold first touch alone blowing the whole
+# credential-restore deadline.
 prewarm_litestream
+prewarm_aws_cli
 
 # 4 — credential stores, restored before anything that reads them starts. Fatal
 # only if persistence itself is unavailable (an unset bucket/prefix, or a broken

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@evident-ai/runner-cdk",
-  "version": "3.4.1-dev.2f1b44b",
+  "version": "3.4.1-dev.51df4de",
   "description": "Reusable CDK constructs for an Evident agent runner: a single scale-to-zero Fargate runner (task + service + per-agent self-stop role + waker Lambda), or a per-session AWS Lambda MicroVM that boots on demand and suspends between messages. Instantiate once per agent from your own stack.",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",