Device Smoke Cycle

device-smoke-cycle

Use this skill for a fast iteration loop: build something locally, stage it on the camera, run a consumer that links/executes against the staged copy, collect logs, reboot. Each project wires the specifics; the cycle shape is the same.

When to use

• You changed a library (e.g. libimp.so, libsysutils.so) and want to see how a consumer (e.g. Raptor's S31raptor) behaves against the new build.
• You want a clean dmesg + logread (+ logcat) bundle from a freshly-started consumer.
• You want the device returned to a known state (rebooted) before the next iteration.
• You want agent-friendly output: a small, comparable log bundle per cycle.

Inputs

Provide via env vars, CLI args, or a ## device-smoke-cycle block in the project's CLAUDE.md. CLAUDE.md takes precedence.

• BUILD_CMD - build invocation. Probe order if unset: ./build-for-device.sh → ./build.sh → ./script.sh → make.
• ARTIFACTS - space-separated host paths to stage (e.g. lib/libimp.so lib/libsysutils.so).
• STAGE_DIR - device directory to scp into. Default: /opt.
• START_CMD - command run on device that consumes the staged artifacts. Typically prefixed with LD_LIBRARY_PATH=$STAGE_DIR so the staged libs win over /usr/lib.
• SETTLE_SECS - seconds to wait after start before capturing. Default: 10.
• COLLECT - commands whose output to capture. Default: dmesg logread logcat.
• POST_ACTION - action after capture. Default: reboot.
• IP - camera IP (required).
• SSH_USER - default root.

Workflow

1. Resolve the build command:

   • Use BUILD_CMD if set (env, arg, or CLAUDE.md).
   • Otherwise probe the repo root in order: build-for-device.sh, build.sh, script.sh, then make.
   • Refuse to guess if none match - ask for BUILD_CMD explicitly.

2. Build locally:

   • Run BUILD_CMD (pass platform/flags as the project expects, e.g. ./build-for-device.sh T31).

3. Stage ARTIFACTS under STAGE_DIR on the device:

   • scp $ARTIFACTS "$SSH_USER@$IP:$STAGE_DIR/"
   • If the build script itself uploads, skip this step.

4. Start the consumer against the staged copy and show initial syslog:

   • ssh "$SSH_USER@$IP" "$START_CMD & logread"
   • Typical START_CMD: LD_LIBRARY_PATH=/opt /opt/S31raptor start.

5. Wait SETTLE_SECS, then capture each COLLECT command's output to a timestamped file:

   • TS=$(date -u +%Y%m%dT%H%M%SZ)
   • For each cmd in COLLECT:
     • ssh "$SSH_USER@$IP" "sleep $SETTLE_SECS; $cmd 2>/dev/null" > "logs/$TS-$cmd.log" (only sleep on the first one)
   • Or collapse into one SSH round-trip with clear ===CMD=== banners between sections.

6. Run POST_ACTION to reset state:

   • Default: ssh "$SSH_USER@$IP" 'reboot' (use reboot -f only if hung).

7. Hand the captured log files back to the agent for higher-level analysis and to plan the next change.

Project wiring via CLAUDE.md

Preferred: add a pinned block so the cycle is explicit and greppable. Example for openimp → Raptor:

## device-smoke-cycle
BUILD_CMD=./build-for-device.sh T31
ARTIFACTS=lib/libimp.so lib/libsysutils.so
STAGE_DIR=/opt
START_CMD=LD_LIBRARY_PATH=/opt /opt/S31raptor start

The skill reads this block first, env/CLI second, probed defaults last.

Notes

• LD_LIBRARY_PATH=$STAGE_DIR is the load-bearing trick: it lets the staged libs win without overwriting /usr/lib/ or unpacking a new firmware image.
• If the consumer's init script launches daemons from /usr/bin/, LD_LIBRARY_PATH still applies (inherited through start-stop-daemon), but the daemon binaries themselves come from /usr/bin/. To test modified daemon binaries, either bind-mount $STAGE_DIR/<daemon> over /usr/bin/<daemon> first, or run the daemons directly instead of via the init script.
• Stop any already-running instance of the consumer before starting the staged copy, or the two will race.
• logcat is not present on every Thingino build - redirect stderr and do not fail the cycle if it is missing.
• Keep a cycle under ~60 seconds wall-clock so successive runs produce comparable logs.
• SSH must be non-interactive (key-based auth) for agent-driven runs.
• A reboot after each cycle guarantees /dev/shm/rss_ring_*, stale daemons, and any mount tricks are cleared - do not skip it when iterating.