@evident-ai/runner-cdk 3.4.1-dev.1856549 → 3.4.1-dev.2b2679a

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

@@ -1,11 +1,9 @@
 #!/usr/bin/env bash
 #
 # Re-dial with the identity /run left behind. There is no step that could fetch
-# a fresh runner key, so the one from /run is what resumes. Also restarts the
-# session-DB replicator /suspend stopped before the snapshot (#812 WI-4):
-# litestream does not survive a suspend/resume freeze on this design (Q2) —
-# /suspend stops it and /resume starts a fresh one, the same pattern already
-# proven for the tunnel.
+# a fresh runner key, so the one from /run is what resumes. The CLI delegated by
+# start_tunnel owns the credential loop and the session-DB replicator's
+# post-resume start.
 set -euo pipefail
 
 # shellcheck source=./common.sh
@@ -39,13 +37,18 @@ fi
   read -r runner_key
 } <"${CONTEXT_FILE}"
 
+# Restore the id so the resumed CLI can self-report and acknowledge any
+# outstanding recycle request (#1906). Older images may not have this file.
+if [ -s "${MICROVM_ID_FILE}" ]; then
+  MICROVM_ID="$(cat "${MICROVM_ID_FILE}")"
+  export MICROVM_ID
+fi
+
 # Tolerant, never fatal: a resume that fails costs the user their whole
 # session, so a broken durable-state config degrades to "no session-DB
-# restart" rather than a failed resume (the same Q3 reasoning /run's
-# session-DB step follows). start_litestream's own guards (disabled/marker/
-# already-running) handle the rest — this needs no logic of its own.
+# replication" rather than a failed resume. The CLI's own guards handle the
+# rest — this needs no logic of its own.
 load_state_config || warn "could not resolve durable-state config; the session DB will not resume replicating"
-start_litestream
 
 # Diagnostic-only, unlike /run's gate: a failed resume costs the user their
 # whole session, so this never exits — it only converts a silent "resumed but

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

@@ -51,20 +51,20 @@ regenerate_machine_id
 # reads it, and left behind for the hooks that flush back to the same prefix.
 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.
+# 3 — start reading the litestream and aws CLI binaries NOW. The reads run in
+# the background and overlap the CLI's own node startup and auth round trip,
+# keeping their cold first touches off the path to the first credential fetch.
 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
-# synchroniser bundle) — a VM with no model credentials yet still boots.
-restore_credentials
+# 4 — export the durable-state config before anything that reads it. Fatal,
+# because the CLI and the teardown hooks must agree on this runner's state prefix.
+load_state_config || exit 1
 
 # 5 — does the runner key in this payload actually authenticate against
-# Evident? Checked here, before opencode or the tunnel spend any of this boot's
-# SIGTERM budget on a key that cannot work, and before anything has started
-# that cleanup would need to tear down. Fatal ONLY on contrary evidence
+# Evident? Checked here, before the tunnel spends any of this boot's time on a key
+# that cannot work, and before anything has started that cleanup would need to
+# tear down. Fatal ONLY on contrary evidence
 # (check_runner_key's own contract, common.sh): a key the API actively rejects
 # cannot work regardless, and failing here costs ~5s against the ~10 minutes a
 # runner that can never connect would otherwise burn before the lifecycle cron
@@ -72,35 +72,14 @@ restore_credentials
 # it warns and this VM still boots.
 check_runner_key "${runner_key}" "${api_url}" || exit 1
 
-# 6 — the OpenCode session DB (#812 WI-2), restored before opencode opens it —
-# the only place this can happen. Bare, like restore_credentials above:
-# restore_session_db never returns non-zero (Q3 — nothing about the session
-# DB may ever fail /run), so there is deliberately no `||` here to catch.
-restore_session_db
-log "session DB restore done ${SECONDS}s into the hook"
-
-# 7 — apply the overlay before OpenCode resolves its project configuration.
-apply_runner_opencode_config
-
-# 8 — configure git after credentials are restored and before agent shells start.
-configure_github_access
-
-# 9 — opencode. Started here, not at build time: a warm process in the shared
-# snapshot would carry its installation id and database into every VM. Not
-# waited on: a slow opencode boot is not a reason to fail /run (the tunnel CLI
-# auto-starts opencode when it finds none healthy,
-# apps/cli/src/commands/ensure-opencode.ts, and the api-worker lifecycle cron
-# reclaims a runner that never comes online).
-start_opencode
-
-# 10 — begin replicating the session DB (#812 WI-3), now that opencode has
-# opened it and before any work can arrive over the tunnel. Bare, like
-# restore_session_db above: start_litestream never returns non-zero (every
-# guard inside it is its own `return 0`), so there is nothing here for
-# `set -e` to abort on.
-start_litestream
+# 6 — preserve the VM id across suspend/resume. The runtime always supplies it
+# for a /run hook, and the subshell keeps the persisted value protected.
+(
+  umask 077
+  printf '%s\n' "${MICROVM_ID}" >"${MICROVM_ID_FILE}"
+)
 
-# 11 — the first per-VM identity on the wire. The subshell's umask makes the file
+# 7 — the first per-VM identity on the wire. The subshell's umask makes the file
 # unreadable to anyone else from the moment it exists, before the key is in it.
 (
   umask 077
@@ -112,6 +91,6 @@ start_litestream
 # attempt cap (apps/cli/src/lib/tunnel/connection.ts), and a control-plane
 # reclaim pass (the api-worker lifecycle cron) is what decides whether a
 # runner that never connects gets suspended.
-start_tunnel "${runner_key}" "${api_url}" "${tunnel_url}"
+start_tunnel "${runner_key}" "${api_url}" "${tunnel_url}" true true
 
 run_succeeded=true

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

@@ -8,8 +8,10 @@ set -euo pipefail
 # shellcheck source=./common.sh
 source "$(dirname "$0")/common.sh"
 
-sync_credentials
-stop_tunnel
+# load_state_config exports PERSISTENCE_BUCKET, which flush_session_db below
+# requires; the CLI flush completes, or is proven absent, before teardown continues.
+load_state_config || warn "could not resolve durable-state config; neither the credential flush nor the session-DB flush below can run"
+stop_runner_and_flush_credentials
 
 # opencode is deliberately NOT stopped here (#812 WI-4) — it is inside the
 # snapshot and must resume with it — so a turn still in flight may not be

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

@@ -8,8 +8,10 @@ set -uo pipefail
 # shellcheck source=./common.sh
 source "$(dirname "$0")/common.sh"
 
-sync_credentials
-stop_tunnel
+# load_state_config exports PERSISTENCE_BUCKET, which flush_session_db below
+# requires; the CLI flush completes, or is proven absent, before teardown continues.
+load_state_config || warn "could not resolve durable-state config; neither the credential flush nor the session-DB flush below can run"
+stop_runner_and_flush_credentials
 
 # Writers stopped BEFORE litestream's final sync (#812 WI-4, the
 # ordered-shutdown invariant, plan §6): otherwise litestream could snapshot
@@ -29,6 +31,6 @@ flush_session_db
 # this hook runs under `set -u` (above), so an unbound one aborts the script
 # AT THIS LINE, leaving the runner key sitting in CONTEXT_FILE on a VM that is
 # being torn down. That is the one thing this line exists to prevent.
-rm -f "${CONTEXT_FILE}" "${SESSION_DB_NO_REPLICATE_MARKER}"
+rm -f "${CONTEXT_FILE}" "${SESSION_DB_NO_REPLICATE_MARKER}" "${CREDENTIAL_FLUSH_MARKER_FILE}"
 
 exit 0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@evident-ai/runner-cdk",
-  "version": "3.4.1-dev.1856549",
+  "version": "3.4.1-dev.2b2679a",
   "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",