|
| 1 | +#!/usr/bin/env bash |
| 2 | +# Gated inference smoke-test for baudbot. |
| 3 | +# |
| 4 | +# Verifies that the control-agent can complete at least one real LLM turn |
| 5 | +# end-to-end via session-control RPC. Default OFF for PR CI; enabled by |
| 6 | +# setting BAUDBOT_CI_INFERENCE_SMOKE=1. |
| 7 | +# |
| 8 | +# Optional fail-open mode: BAUDBOT_CI_INFERENCE_SMOKE_OPTIONAL=1 logs failure |
| 9 | +# as a warning instead of failing the build (useful for non-nightly runs where |
| 10 | +# flaky provider errors shouldn't block merges). |
| 11 | +# |
| 12 | +# Expects baudbot to be already installed and stoppable via `sudo baudbot`. |
| 13 | + |
| 14 | +set -Eeuo pipefail |
| 15 | + |
| 16 | +readonly AGENT_USER="baudbot_agent" |
| 17 | +readonly AGENT_HOME="/home/${AGENT_USER}" |
| 18 | +readonly AGENT_ENV="${AGENT_HOME}/.config/.env" |
| 19 | +readonly CONTROL_DIR="${AGENT_HOME}/.pi/session-control" |
| 20 | +readonly CONTROL_ALIAS="${CONTROL_DIR}/control-agent.alias" |
| 21 | +readonly START_TIMEOUT_SECONDS=60 |
| 22 | +readonly INFERENCE_TIMEOUT_SECONDS=120 |
| 23 | +readonly EXPECTED_TOKEN="CI_INFERENCE_OK" |
| 24 | + |
| 25 | +started=0 |
| 26 | + |
| 27 | +log() { |
| 28 | + printf '[inference-smoke] %s\n' "$*" |
| 29 | +} |
| 30 | + |
| 31 | +cleanup() { |
| 32 | + local exit_code=$? |
| 33 | + if [[ $started -eq 1 ]]; then |
| 34 | + log "cleanup: stopping baudbot" |
| 35 | + sudo baudbot stop >/dev/null 2>&1 || true |
| 36 | + fi |
| 37 | + exit "$exit_code" |
| 38 | +} |
| 39 | +trap cleanup EXIT |
| 40 | + |
| 41 | +wait_for_control_socket() { |
| 42 | + local deadline=$((SECONDS + START_TIMEOUT_SECONDS)) |
| 43 | + local target="" |
| 44 | + |
| 45 | + while (( SECONDS < deadline )); do |
| 46 | + if [[ -L "$CONTROL_ALIAS" ]]; then |
| 47 | + target="$(readlink -- "$CONTROL_ALIAS" 2>/dev/null || true)" |
| 48 | + if [[ -n "$target" ]]; then |
| 49 | + if [[ "$target" != /* ]]; then |
| 50 | + target="${CONTROL_DIR}/${target}" |
| 51 | + fi |
| 52 | + if [[ -S "$target" ]]; then |
| 53 | + printf '%s\n' "$target" |
| 54 | + return 0 |
| 55 | + fi |
| 56 | + fi |
| 57 | + fi |
| 58 | + sleep 1 |
| 59 | + done |
| 60 | + |
| 61 | + return 1 |
| 62 | +} |
| 63 | + |
| 64 | +dump_diagnostics() { |
| 65 | + log "--- diagnostics ---" |
| 66 | + sudo baudbot status 2>&1 || true |
| 67 | + log "--- end diagnostics ---" |
| 68 | +} |
| 69 | + |
| 70 | +# Send a message via session-control RPC and wait for turn_end. |
| 71 | +# Prints the assistant response content on success, exits non-zero on failure. |
| 72 | +rpc_send_wait_turn_end() { |
| 73 | + local socket_path="$1" |
| 74 | + local message="$2" |
| 75 | + local timeout_seconds="$3" |
| 76 | + |
| 77 | + sudo -u "$AGENT_USER" python3 - "$socket_path" "$message" "$timeout_seconds" <<'PY' |
| 78 | +import json |
| 79 | +import socket |
| 80 | +import sys |
| 81 | +
|
| 82 | +sock_path = sys.argv[1] |
| 83 | +message = sys.argv[2] |
| 84 | +timeout_seconds = int(sys.argv[3]) |
| 85 | +
|
| 86 | +send_cmd = {"type": "send", "message": message, "mode": "steer"} |
| 87 | +subscribe_cmd = {"type": "subscribe", "event": "turn_end"} |
| 88 | +
|
| 89 | +client = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM) |
| 90 | +try: |
| 91 | + client.settimeout(timeout_seconds) |
| 92 | + client.connect(sock_path) |
| 93 | +
|
| 94 | + # Send both commands |
| 95 | + client.sendall((json.dumps(send_cmd) + "\n").encode("utf-8")) |
| 96 | + client.sendall((json.dumps(subscribe_cmd) + "\n").encode("utf-8")) |
| 97 | +
|
| 98 | + buf = b"" |
| 99 | + send_response = None |
| 100 | +
|
| 101 | + while True: |
| 102 | + chunk = client.recv(8192) |
| 103 | + if not chunk: |
| 104 | + print("connection closed before turn_end", file=sys.stderr) |
| 105 | + sys.exit(1) |
| 106 | + buf += chunk |
| 107 | +
|
| 108 | + while b"\n" in buf: |
| 109 | + line, buf = buf.split(b"\n", 1) |
| 110 | + line = line.strip() |
| 111 | + if not line: |
| 112 | + continue |
| 113 | +
|
| 114 | + try: |
| 115 | + msg = json.loads(line.decode("utf-8", errors="replace")) |
| 116 | + except json.JSONDecodeError: |
| 117 | + continue |
| 118 | +
|
| 119 | + if msg.get("type") == "response": |
| 120 | + cmd = msg.get("command", "") |
| 121 | + if cmd == "send": |
| 122 | + if not msg.get("success", False): |
| 123 | + print(f"send failed: {msg.get('error', 'unknown')}", file=sys.stderr) |
| 124 | + sys.exit(1) |
| 125 | + send_response = msg |
| 126 | + # Ignore subscribe response |
| 127 | + continue |
| 128 | +
|
| 129 | + if msg.get("type") == "event" and msg.get("event") == "turn_end": |
| 130 | + if send_response is None: |
| 131 | + print("received turn_end before send response", file=sys.stderr) |
| 132 | + sys.exit(1) |
| 133 | + data = msg.get("data", {}) |
| 134 | + assistant_msg = data.get("message", {}) |
| 135 | + content = assistant_msg.get("content", "") |
| 136 | + if not content: |
| 137 | + print("turn completed but no assistant content", file=sys.stderr) |
| 138 | + sys.exit(1) |
| 139 | + print(content) |
| 140 | + sys.exit(0) |
| 141 | +
|
| 142 | + print("stream ended without turn_end event", file=sys.stderr) |
| 143 | + sys.exit(1) |
| 144 | +except socket.timeout: |
| 145 | + print("timeout waiting for inference response", file=sys.stderr) |
| 146 | + sys.exit(1) |
| 147 | +finally: |
| 148 | + client.close() |
| 149 | +PY |
| 150 | +} |
| 151 | + |
| 152 | +inject_api_key() { |
| 153 | + # If a real API key is available in the environment, inject it into the |
| 154 | + # agent's .env so the runtime can authenticate with the provider. |
| 155 | + if [[ -n "${CI_ANTHROPIC_API_KEY:-}" ]]; then |
| 156 | + log "injecting CI API key into agent .env" |
| 157 | + sed -i "s|^ANTHROPIC_API_KEY=.*|ANTHROPIC_API_KEY=${CI_ANTHROPIC_API_KEY}|" "$AGENT_ENV" |
| 158 | + elif grep -q "ANTHROPIC_API_KEY=sk-ant-testkey" "$AGENT_ENV" 2>/dev/null; then |
| 159 | + log "WARNING: agent .env has dummy API key; inference will likely fail" |
| 160 | + log " set CI_ANTHROPIC_API_KEY to provide a real key" |
| 161 | + fi |
| 162 | +} |
| 163 | + |
| 164 | +main() { |
| 165 | + inject_api_key |
| 166 | + |
| 167 | + log "starting baudbot" |
| 168 | + sudo baudbot start |
| 169 | + started=1 |
| 170 | + |
| 171 | + log "waiting for control-agent socket" |
| 172 | + local socket_path="" |
| 173 | + if ! socket_path="$(wait_for_control_socket)"; then |
| 174 | + log "control-agent socket did not become ready within ${START_TIMEOUT_SECONDS}s" |
| 175 | + dump_diagnostics |
| 176 | + return 1 |
| 177 | + fi |
| 178 | + log "control socket ready: ${socket_path}" |
| 179 | + |
| 180 | + log "sending inference prompt (timeout ${INFERENCE_TIMEOUT_SECONDS}s)" |
| 181 | + local response="" |
| 182 | + if ! response="$(rpc_send_wait_turn_end "$socket_path" \ |
| 183 | + "Reply with exactly: ${EXPECTED_TOKEN}" \ |
| 184 | + "$INFERENCE_TIMEOUT_SECONDS")"; then |
| 185 | + log "inference failed" |
| 186 | + dump_diagnostics |
| 187 | + return 1 |
| 188 | + fi |
| 189 | + |
| 190 | + # Validate response contains expected token |
| 191 | + if [[ "$response" == *"$EXPECTED_TOKEN"* ]]; then |
| 192 | + log "inference response contains expected token" |
| 193 | + else |
| 194 | + log "unexpected response (missing '${EXPECTED_TOKEN}'):" |
| 195 | + log " ${response:0:500}" |
| 196 | + dump_diagnostics |
| 197 | + return 1 |
| 198 | + fi |
| 199 | + |
| 200 | + log "stopping baudbot" |
| 201 | + sudo baudbot stop |
| 202 | + started=0 |
| 203 | + |
| 204 | + log "inference smoke passed" |
| 205 | +} |
| 206 | + |
| 207 | +main "$@" |
0 commit comments