Activity

You can let the agent work unattended and review what it did afterwards. A task keeps its log, model calls are metered as they go through the gateway, and each sandbox’s start and end are recorded.

droneship logs <id>          # what a task did
droneship logs -f <id>       # follow it while it works
droneship logs --raw <id>    # every event, as JSONL
droneship list -a            # how your runs and recent sandboxes ended

The task log

droneship logs shows one row per action, with its time, what the agent did (read, cmd, write, tool and others), its target and the outcome, Claude’s messages between them, and an end row that reads done or failed. Background runs and logs shows an example. --raw prints the agent’s own JSONL log, one event per line, for jq or your own tooling.

In the console, open a task to see its Logs tab: the same raw log, following live while the task runs, with a filter and a copy button.

A stopped task keeps its log for 7 days. logs reads tasks only, not interactive sandboxes. Background runs and logs has the flags and exit codes.

How each sandbox ended

droneship list -a shows your running sandboxes and your tasks. An interactive sandbox that ended in the last day is listed too, with how it ended: killed, idle, time limit, stopped or lost, and whether its workspace was kept:

workspace saved — droneship run claude picks it up

Model spend

Every model call that goes through the Highflame gateway is metered against the sandbox that made it. The console shows it:

• On the sandbox page: estimated AI spend, with tokens.
• On the Tasks list: each task’s AI spend, and the total.
• On the Usage page: each sandbox’s time running and AI spend for the project. The spend reads Subscription for a sandbox billed to your own Claude or ChatGPT plan, and not metered for one whose model calls skipped the gateway (--direct).

How events are linked to a sandbox

Each sandbox has its own identity. Droneship records an event when a sandbox launches and when it ends, tagged with that identity’s id. The gateway tags each model call it serves for the sandbox with the same id. That shared id is what ties a sandbox’s spend to the sandbox.

The launch event records the harness, egress bundle, backend and isolation level, how long the launch took, and, for a refused launch, why. The end event records how it ended (killed, idle, time limit) and how long it ran.

More: The console, droneship logs, droneship list.

Next

• Hand off a task: start a task and read its result.
• Background runs and logs: logs, exit codes and cleanup.
• Auto-stop: when a sandbox ends on its own.