#Results and shell history

marsh results and marsh jobs read durable receipts for registered-command invocations, with different list views. Brush's history builtin continues to show ordinary shell command history.

marsh results
marsh results show 17
marsh results show 2f41f39a
marsh results --json
marsh results show 17 --json

The list is newest first and displays the full job UUID, placement, lifecycle status, cleanup state, WALL time and exit code. WALL is the receipt's timing.wall_ms: the whole request from the invocation to verified cleanup, including Kit VM preparation (vm_prepare) and image work. marsh jobs shows RUN instead: from the job record's creation (the Kit VM is ready) to its end. A cold first call therefore shows a much larger WALL than RUN; marsh results show ID --json breaks WALL down by phase. Read cleanup separately from status: a command can finish while deletion remains uncertain. show accepts the displayed cursor, a full job UUID, or a unique UUID prefix. The existing marsh jobs --json and marsh jobs show JOB --json interfaces remain available for automation.

The same-user daemon stores only structural receipt fields: command registry name, opaque execution identities, public mount targets, state, exit cause, cleanup result, timings, and bounded arguments for display. It stores the first 16 arguments, each truncated to 200 characters with lossy UTF-8 decoding (DaemonStore::set_job_args). A prompt passed as an argument can be retained. The journal captures no stdin, stdout, or stderr. Treat paths and arguments as potentially sensitive metadata. JSON receipts expose execution as a typed worker outcome, independently of the public exit code/cause. Legacy receipts without that field report unknown; do not reconstruct it by parsing diagnostic prose.

The versioned checksum journal lives under the host-only control directory at $HOME/Library/Application Support/marsh/control/<selected-home-hash>/state/results.journal. When MARSH_CONTROL_HOME is set, the journal uses $MARSH_CONTROL_HOME/<selected-home-hash>/state/results.journal. selected-home-hash means SHA256 of canonical MARSH_HOME, not its home/ child. Both status forms report the absolute control_home path. The control and state directories are mode 0700, and the journal is mode 0600; the daemon never imports a journal from guest-writable $MARSH_HOME/state.

Terminal transitions are synchronized before completion is reported. During replay, the daemon truncates an incomplete final frame, but any checksum or structural failure in a complete frame prevents startup. An owner-only lock in the state directory excludes another journal writer for the selected home, even with a different TMPDIR. Stop daemons from builds predating that lock before upgrading, since they do not hold it.

Retention keeps the newest 1,000 terminal receipts and all active receipts. After a daemon exit, replay marks queued or running jobs unknown with cause daemon_restarted; it never resumes the old containers. Quarantine blocks worker reuse within a daemon generation, but the adapter's in-memory quarantine set does not survive restart. Process admission separately retains uncertain jobs' held slots from receipts until a successful marsh workers reset KIT is recorded. Use that reset to retire the affected owned VM; do not delete the journal to recover.

#Scope configuration and job limits

Use marsh status to find the host-control directory and inspect the daemon's per-job defaults. marsh status --json exposes control_home and job_defaults; job_defaults.resources contains the effective values and job_defaults.environment_overrides lists the fixed variable names supplied at daemon startup. Fields not listed use built-in defaults. An inspection-only daemon can report no job defaults; absence is not an unlimited policy.

The writable-layer limit is enforced by detect-and-kill, not by the storage driver: the Kit VM's containerd overlayfs snapshotter ignores Docker's --storage-opt size (still passed, and honored where supported). While a job runs, its worker samples the allocated size of the container's overlay upper directory every 20 ms (less often for very large writable trees, so the measurement stays under ~10% of one CPU) and kills the container's cgroup once it exceeds the limit; the receipt's cause is limit:writable. A fast writer can exceed the limit by up to what it writes in one sampling window before the kill lands (tens of MiB on a Kit VM disk), plus whatever it writes in the first moments before the worker has found the started container. Files in mounted directories (the project, the guest home) are not part of the writable layer.

The first launcher starts the same-user daemon with its environment. That snapshot applies to jobs in the scope, not just that shell. Values must be positive decimal integers within their supported range; zero does not mean unlimited. Later shells cannot change a running daemon's defaults. A conflicting explicit MARSH_JOB_* setting is rejected rather than silently ignored. Unset conflicting variables to inspect status. To change the snapshot, finish active work, close shells, stop that scope, then launch with the desired settings; or use a separate MARSH_HOME. These are job resource limits, not hard provider billing limits or per-shell policy flags.

Host-control commands.json merges over the packaged command registry by name. Both declarations are validated, including shadowed references; the combined registry is limited to 251 commands. Names are at most 128 ASCII bytes using letters, digits, _ and -, cannot start with -, and cannot use reserved product names. Duplicate JSON keys are rejected. See the shared name rules and registry guide.

Host-control agents.json instead replaces the complete packaged ACP agent list. A missing file uses packaged defaults; [] intentionally disables ACP. It does not merge individual agents or revive omitted defaults. Malformed or unsafe declarations and invalid ACP-to-Kit bindings fail daemon startup, rather than silently disabling ACP. Fix the reported host-control or packaged file; do not delete state to bypass a configuration failure.