Agent CLI Contract

English · 繁體中文

Orbit is designed to be driven by coding agents through the CLI. Agents should prefer structured output over human-formatted text whenever they need to parse state, decide next actions, or recover from errors.

Rule of Thumb

Use --json for programmatic reads and writes:

orbit status --json
orbit doctor --json
orbit up --infra --json
orbit logs redis --json

Human output is optimized for terminals and may change. JSON output is the agent-facing contract.

JSON Envelope

Commands converted to the agent contract return one JSON object with this top-level shape:

{
  "schema_version": "orbit.cli.v1",
  "ok": true,
  "command": "doctor",
  "instance": "ci-123",
  "data": {},
  "error": null,
  "notices": [],
  "recommended_actions": []
}

Fields:

Field	Meaning
schema_version	Contract version. Current value is orbit.cli.v1.
ok	true on success, false on failure.
command	The Orbit command that produced the response.
instance	Named runtime targeted by --instance, omitted for the default runtime.
data	Command-specific payload on success.
error	Structured error payload on failure, otherwise null.
notices	Structured non-failing events, such as a one-time legacy source migration.
recommended_actions	Follow-up commands the agent should consider.

Every lifecycle command accepts --instance <name>. Use orbit instance list --json to discover each instance's environment, state, dashboard, and resolved resource endpoints. orbit instance clean <name> --json stops the instance and removes only resources carrying that instance's ownership. Recommended actions emitted for a named runtime retain that instance target. The full targeting and cleanup model is documented in Isolated runtime instances.

When a converted command fails with --json, Orbit prints a single JSON object to stdout and exits with code 1.

orbit up --json and orbit env apply --json share one operation-wide --timeout with a five-minute default. The deadline starts when command execution begins and bounds daemon readiness, lifecycle RPCs, status polling, and failure-evidence collection. Deadline exhaustion returns timeout; caller or catchable-signal cancellation returns canceled. If an environment reconcile request was dispatched, the canceled message states that the daemon may have accepted it and recommends orbit status --json.

Streams

stdout carries the envelope and nothing else. Parse it alone.

stderr carries diagnostics and progress for humans, and is not part of the contract — its content and wording may change between releases. orbit up --json and orbit env apply --json announce a small number of lifecycle phase changes there. If a phase stays unchanged for 30 seconds, they add elapsed time and an approximate remaining operation budget. During readiness, a resource heartbeat replaces the generic phase heartbeat for that interval:

⋯ ensuring daemon
⋯ waiting for readiness
⋯ api still starting (elapsed 30s, about 4m29s remaining)

Phase wording and timing are diagnostic, not stable values for agents to parse. Use the final envelope for decisions.

Merging stderr into stdout before parsing will break: progress lines are not JSON. Redirect them separately, or discard stderr if only the envelope matters.

Error Shape

Structured errors use this shape:

{
  "schema_version": "orbit.cli.v1",
  "ok": false,
  "command": "logs",
  "data": null,
  "error": {
    "code": "unknown_resource",
    "message": "unknown resource: redisx",
    "hint": "Run orbit status --json to list configured resources.",
    "retryable": false,
    "next_command": "orbit status --json"
  },
  "recommended_actions": [
    {
      "command": "orbit status --json",
      "reason": "List configured resources and current states."
    }
  ]
}

Agents should prefer error.next_command and recommended_actions over guessing a recovery path from the message text.

Executable recommended actions

recommended_actions[].command is an executable machine command. It includes --json when structured output is available and is not intended as terminal display text; a UI may derive a human label separately.

Some failures have no command that advances them. socket_path_too_long needs a shorter ORBIT_HOME before any Orbit command can succeed, so it carries a hint and no next_command or recommended action. Treat a missing next_command as "act on the hint", not as a malformed response.

For resource_port_conflict, error.next_command is the platform-specific read-only command that inspects the current port owner. The resource's status also includes port_conflict evidence (port, resource, optional owner, and inspect_command). Do not retry or fetch logs until the owner is stopped or the shared environment selects a different host port.

Resource status includes logs_available: true only when the running daemon has buffered output for that resource. Absence means there is no historical output to inspect yet; clients should not present Logs as a recovery action for a dependency-blocked resource unless this evidence is true. A direct orbit logs <resource> --json request for a resource that never started returns logs_unavailable, not an empty success. Its sole recommended action is based on a live port recheck: inspect the current owner while occupied, or retry only that resource after the port is released.

A degraded host process may include both state_reason (for example, exited: exit status 1) and failure_evidence, the last meaningful application log line captured for that failed generation. Treat the evidence as supporting detail, not a replacement for the lifecycle reason.

A failed orbit up --json keeps its evidence in data.failed_resources: one entry per requested resource that did not reach healthy, with name, kind (container or service), role, state, state_reason, failure_kind, and log_tail (at most 20 recent log lines). Read the failure from there instead of issuing a follow-up orbit logs call.

A container that stops while Orbit is still waiting for it to become healthy ends the wait at once, with failure_kind: runtime and Docker's exit_code (plus oom_killed: true when the kernel killed it). orbit status --json reports the same fields. Look for that entry in failed_resources rather than relying on error.code alone: a dependent resource can be reported first as dependency_blocked.

Destructive steps that need an explicit go-ahead use the stable confirmation_required error code. The refusal changes nothing; the first recommended action is the same command with --yes. Emit it only when the caller (or its operator) intends the destruction.

Stable error codes

error.code values are part of the versioned contract. The current set:

Code	Meaning
agent_plugin_failed	a coding agent's plugin command failed; the replacement actions are the commands to run by hand
checks_failed	doctor checks failed; resolve them, then run doctor again
canceled	the caller or a supported catchable signal canceled the operation
command_failed	unclassified failure; act on the hint
confirmation_required	a destructive step needs --yes
daemon_unreachable	Orbit is not running
dashboard_port_conflict	the dashboard port is owned by another process
dependency_blocked	a required dependency cannot become healthy
env_mismatch	the selected config differs from the running daemon's; next_command is orbit up -c <config> --json, which asks for --yes before stopping running resources
env_repo_access	environment repository access or authentication failed
env_repo_unavailable	environment repository unreachable
environment_changed	the env file changed; apply before continuing
environment_schema_newer	the env schema is newer than this Orbit
environment_schema_outdated	the env schema predates this Orbit
environment_selection_required	no environment is selected
init_incomplete	initialization stopped before finishing
invalid_argument	conflicting or unknown command selection
invalid_environment	the environment file fails validation
invalid_environment_schema	the environment file fails schema validation
json_unsupported_destructive_command	this destructive command refuses --json
logs_unavailable	no buffered output exists for that resource
not_configured	the active env does not enable this feature
orbit_update_pending	an installed update needs a daemon restart
package_managed_install	Orbit is owned by Homebrew or Scoop; use the reported package-manager command
package_managed_rollback	rollback must be performed through the package manager that owns Orbit
project_context_inactive	a project config was found but the daemon serves another env
resource_port_conflict	a resource's host port is owned by another process
service_start_failed	a resource failed to become healthy
service_working_directory_missing	a host service path does not resolve
setup_required	orbit init has not completed
socket_path_too_long	ORBIT_HOME produces an over-long socket path
timeout	the operation exceeded its operation-wide --timeout
unknown_group	--group names no defined group
unknown_resource	a named resource is not in the env

Converted Commands

These commands currently use the orbit.cli.v1 envelope when --json is set:

Command	JSON behavior
orbit version --json	Returns the installed Orbit version.
orbit doctor --json	Returns diagnostic checks in data.
orbit inspect --json	Returns an agent-ready state snapshot with readiness, daemon/env summaries, resource risks, release-update state, and recommended follow-up commands. Read-only inspection remains available during an update transaction.
orbit status --json	Returns setup/selection readiness, the selected and available environments, managed repository URL/ref/commit when applicable, daemon and release-update state, and configured resource state in data.resources. Read-only status remains available during an update transaction.
orbit env info --json	Returns the env's identity and, per resource, ports and URL with provenance: declared comes from the environment file, observed from the running daemon. Observed values are withheld when the daemon serves a different environment (data.daemon.config_match: false). Resource environment values appear only with --show-secrets; key names are always listed.
orbit logs <resource> --json	Returns recent log lines in one JSON object.
orbit logs <resource> -f --json	Streams NDJSON events, one JSON object per line.
orbit up --json	Returns the resources actually selected by the daemon (including dependencies and group filtering), observed final states, degraded/timed-out resources, and recommended follow-up commands. When it applies pending config edits, data.environment_changes reports running intent preserved across the handoff. An environment with no enabled resources succeeds immediately with empty arrays.
orbit down --json	Returns final lifecycle result after stopping resources. It is a successful no-op with empty arrays when Orbit is already stopped. Before setup it recommends orbit init; after setup it recommends the next normal orbit up.
orbit down <resources...> --json	Returns the final lifecycle result for the requested resources. If that stop degrades running dependents, they are included in resources and degraded_resources with the one dependency-restoration action. --group and --infra use the same mutually exclusive selection modes as orbit up; group shutdown stops members without stopping shared dependencies.
orbit restart --json	Returns final lifecycle result and verifies restart evidence.
orbit env list --json	Returns data.environment with the selection state, prior selection when unavailable, exact available environment choices, and managed repository URL/ref/commit when applicable.
orbit env use <path> --json	Returns the selected env, env name, daemon running state, and whether restart is required.
orbit source sync [<name>] --json	Returns per-source results; --all continues independent sources and reports every success or failure.
orbit env apply --json	Applies pending environment changes without interrupting unchanged resources, then returns the resources that were running, preserved or restarted, or removed from the new config. It accepts the same operation-wide --timeout as up.
orbit switch <env> --json	Returns the selected env, daemon start/restart action, final daemon state, config path, dashboard URL, and the new env's prerequisite checks/readiness.
orbit update --json	Checks/downloads when no target is staged, applies the verified target across registered default/named runtimes, and returns additive phase and per-runtime outcomes. Existing binary/resource fields remain compatible. --rollback uses the same installation-wide journal.
orbit daemon start --json	Returns daemon running state, PID, config path, and dashboard URL.
orbit daemon stop --json	Returns stopped state, previous PID, and whether service shutdown was requested.
orbit daemon restart --json	Returns previous/new daemon state, PID, config path, dashboard URL, and service shutdown effect.
orbit uninstall --json	Previews binary artifacts and whether user data is preserved; --yes is required before removal.
orbit skill install --json	Previews or, with --yes, applies the two agent-skill routes: data.claude_plugin (claude_code_detected, installed, action, the plugin commands) and data.agents_directory (path, state of current/outdated/modified/missing, action). data.applied is true when every planned action completed: false for the preview, and false in an agent_plugin_failed envelope, whose data still reports the directory route's result. An action of none carries a reason; a plugin route handed back (for example because claude is not on the shell's PATH) is none with the commands to run listed. A second run reports both actions as none.
orbit skill uninstall --json	Same shape; removes the plugin and any copy Orbit wrote (current or outdated), reports a modified copy as left in place, and leaves the marketplace registration alone.
`orbit sqlserver publish <database	project> --json`
orbit sqlserver list --json	Returns the configured SQL projects and their databases in data.
orbit sqlserver diff --json	Returns per-database drift in the envelope; --script adds the change script to data.
orbit sqlserver reset --json	Never resets: refuses with the stable json_unsupported_destructive_command error so data loss stays a human decision.
orbit init --json	Returns the setup result in the envelope, with failures carrying the incomplete step.
orbit tunnel list --json	Returns active claims in data.claims (paths, owner, target, status); --all includes other owners' claims with mine flags.
orbit tunnel release --json	Returns the released count, paths or local port, and gateway in the envelope.
orbit trace --json	Returns recent trace summaries in data.traces, newest first.
orbit trace -f --json	Streams NDJSON trace-summary events, one JSON object per line.
orbit trace <id> --json	Returns one full trace (summary fields + spans) in data.
orbit tracing status --json	Returns the receiver's health, the port in use, and ingest counters in data.

Lifecycle commands keep decorative and diagnostic progress off stdout in JSON mode so the final envelope remains parseable.

Lifecycle actions are outcome-specific. A successful up returns one primary next action, orbit open --json. A failed start recommends status, logs for the root failed resource, and a targeted restart after the reported cause is fixed; it does not send agents through unrelated setup diagnostics.

Lifecycle payloads use resource vocabulary consistently:

• requested_resources is the daemon-resolved set, including selected dependencies and excluding resources filtered out by groups.
• resources contains the observed final state for that set.
• degraded_resources and timed_out_resources identify unsuccessful outcomes without treating containers as services.
• environment_changes, when present, contains previously_running, restored_resources, started_dependencies, and unavailable_resources from the config handoff performed before startup.

This is the exact set whose terminal state the command waits for. Status and inspect likewise expose mixed host processes and containers under resources. Log payloads and NDJSON events identify their source with resource.

When a command discovers the nearest project orbit.yaml, status reports that actual active config as data.environment.source: "project", with selected_path pointing to the discovered file. This project context takes precedence over a managed environment selected elsewhere. Agents should use the reported source and path rather than infer the active config from ~/.orbit/envs.

Daemon-backed responses also expose context: kind, canonical identity and config_path, display_name, availability, running state, and (for projects) project_root. A managed selection overridden by a project remains present under managed_selection with active: false. An explicit -c <path> has kind explicit, even when its basename is orbit.yaml.

If another project-local environment is running, status remains successful and describes only the current project's configured resources. It sets data.environment.context_switch_required: true, reports running_name and running_path, and recommends exactly orbit up --json. Doctor validates the current project and treats ports owned by the running project as releasable during that switch. Operational commands that could control the other project's resources fail with project_context_inactive; agents should follow its sole orbit up --json action. A successful switch reports data.context_switch in the up result, including both project names and the resources stopped before startup. When resources are running, non-interactive and JSON switches without --yes fail with confirmation_required and the exact orbit up --yes --json action. The target is validated before that confirmation and no resource is stopped on refusal. orbit up -c <config> switches an explicit config the same way, keeping -c in the confirmed action; every other command still fails closed with env_mismatch. Inspect follows the same ownership boundary and never reports the other project's resources or dashboard as belonging to the current project.

Before first setup, status remains a successful state query with data.setup_required: true, a human-readable setup_message, and exactly one orbit init --yes --json action. A present but invalid environment file is instead an invalid_environment error, because startup cannot safely use it. If the previously selected environment was renamed or removed, status instead returns data.selection_required: true, a selection_message, and exact orbit switch <env> --json choices; this does not mean setup must be repeated.

orbit init --yes --json never invents a source workspace from the current directory. The field is omitted for self-contained environments. If a custom environment requires ${WORKSPACE_ROOT} and no proven local workspace exists, init returns service_working_directory_missing and explains that the source must be removed and added again with --workspace "$PWD". Other unresolved path variables preserve their name and lead to orbit settings set-env <NAME> "$PWD" --json; they never produce a source-workspace instruction.

When GitHub reports that an environment repository was not found, Orbit returns env_repo_unavailable without recommended actions. GitHub deliberately uses the same response for a misspelled or missing repository and a private repository hidden from the current credentials, so an agent must not automatically log in or retry unchanged. Verify the owner and repository name first; authenticate Git only after confirming that the URL is correct and private. Unambiguous authentication failures remain env_repo_access.

Stable data.operation values for converted control commands:

Command	data.operation
orbit env list --json	env_list
orbit env info --json	env_info
orbit tunnel list --json	tunnel_list
orbit tunnel release --json	tunnel_release
orbit env use <path> --json	env_use
orbit source sync --json	source_sync
orbit env apply --json	env_apply
orbit switch <env> --json	switch
orbit daemon start --json	daemon_start
orbit daemon stop --json	daemon_stop
orbit daemon restart --json	daemon_restart
orbit update --json	update
orbit update --rollback --json	rollback
orbit settings set <key> <value> --json	settings_set
orbit settings set-env <name> <value> --json	settings_set_env
orbit settings list --json	settings_list
orbit uninstall --json	uninstall
orbit skill install --json	skill_install
orbit skill uninstall --json	skill_uninstall

Settings map fields

orbit settings list --json returns data.settings.env_toggles and data.settings.user_env as JSON objects in both daemon-backed and offline reads. Empty maps serialize as {} rather than disappearing, so consumers can distinguish "configured but empty" from a response that violates the contract.

switch stops the running environment first, so with resources running and no --yes it returns confirmation_required instead of acting — switching is a machine-wide provisioning step, not something a harness should trigger implicitly. For switch, previous_environment_stopped says whether Orbit stopped a previously running environment before making the selection. A stopped Orbit is not started merely to record a selection; orbit up remains the sole startup action. prerequisites_ready is false when the newly selected env is missing a required runtime or package installation; prerequisites carries the same checks as Doctor, and recommended_actions includes exact runnable remedies when Orbit can determine one.

An unresolved or missing host-service path uses the stable service_working_directory_missing error code. switch leaves the daemon stopped, and Doctor or up returns the exact workspace-setting or config-edit action. Execute that action before retrying; up has not started dependencies.

Daemon stop/restart payloads may include stop_method when a prior daemon was running. Stable values are graceful, terminated, and killed.

Legacy JSON Commands

Some commands predate the envelope and intentionally keep their existing JSON shape for compatibility:

Command	Behavior
orbit daemon status --json	Returns the legacy daemon status object.
orbit history --json	Keeps its existing history payload.
orbit history gaps --json	Keeps its existing history gaps payload.
orbit tunnel claim --json	Streams Tunlease-shaped NDJSON events (schema_version: 1, type per event) — it is a live stream, not a request-response. The upstream event shape also stays available on list/release behind -o json.

Open

orbit open [resource] --json opens the target and returns one success envelope. The payload contains url, opened, and a target of dashboard, service, or container. A named target includes its name in resource. Service targets also retain the same value in service for backward compatibility; container targets do not use that legacy field.

Passthrough Commands

These commands wrap an interactive or foreign process, so --json (a global flag) is accepted but has no effect — output is the wrapped tool's raw stream:

orbit exec, orbit query redis|mongo|postgres, orbit topics *, orbit sqlserver query. orbit seed prints human progress on success; only its failures use the error envelope.

Do not assume every --json command has an envelope. Check the command-specific contract before parsing.

The daemon HTTP API is structured for the Orbit UI and CLI internals, but the current agent-facing compatibility contract is the CLI --json surface unless a future design explicitly says otherwise.

Logs

Non-streaming logs return one envelope:

orbit logs redis --json

The success payload includes the service name, requested line count, a non-null lines array, and whether output was truncated.

Streaming logs use newline-delimited JSON:

orbit logs redis -f --json

Each emitted line is a complete JSON object. Agents should parse it as NDJSON, not as one JSON array.

If a followed log stream fails after NDJSON output has begun, Orbit emits a final NDJSON event with type: "error" instead of appending an envelope in a different format, then exits non-zero.

Traces

Tracing is on by default; it is off only when the active env sets an explicit tracing.enabled: false (see tracing.md). Check the live receiver state with orbit tracing status --json.

orbit trace --json             # recent traces; --limit N caps the count
orbit trace <trace-id> --json  # one full trace with its span tree
orbit trace -f --json          # NDJSON stream of trace-summary updates

A trace summary carries traceId, rootService, rootName, startUnixNano, durationMs, spanCount, services (distinct, first-seen order), and status (ok | error). The full trace adds spans, each with traceId/spanId/parentId, service, name, kind, startUnixNano/ endUnixNano/durationMs, status/statusMsg, and attributes. These key names are pinned by contract tests (app/traces_json_test.go) — treat renames as breaking.

In -f --json mode the same traceId may be emitted repeatedly as its trace accumulates spans; the latest event supersedes earlier ones.

A typical agent workflow: orbit trace --json → pick the errored trace → orbit trace <id> --json to find the failing span → the trace's log lines via GET /api/traces/{id}/logs on the daemon, or orbit trace <id> --logs for human-readable output.

Inspect

orbit inspect --json is the recommended first command for agents that need to understand whether Orbit is ready for automation. It returns the normal orbit.cli.v1 envelope, with data.schema_version set to orbit.inspect.v2.

The inspect payload contains:

Field	Meaning
readiness	Stable decision state with state, blocked, and summary.
daemon	Daemon running state, PID, version, upgrade info, and dashboard URL when available.
environment	The same state, selected_name, selected_path, and environments selection object used by status and env list, plus preview/daemon details when available.
resources	Bounded summary of the running daemon's resource states. It does not enumerate declared configuration; use orbit env info --json for that.
risks	Ordered machine-readable risks such as setup_required, environment_selection_required, orbit_update_pending, config_invalid, environment_stopped, env_mismatch, status_unavailable, dependency_readiness_ambiguous, resource_degraded, resource_converging, and resource_stopped. A dependency-readiness risk is advisory config evidence and can accompany a blocking lifecycle risk.
recommended_actions	Safe next commands the agent should consider.

Stable readiness.state values:

State	Blocked	Meaning
setup_required	true	No usable environment has been selected; the only next action is orbit init --yes --json.
selection_required	true	The previous selection is unavailable. Actions contain exact orbit switch <source>/<env> --json choices, or orbit source sync --json when none are available.
config_invalid	true	The selected config cannot be loaded. An older shared schema has the sole action orbit source sync --json; a newer schema has the sole action orbit update --json. Syntax errors and unknown fields require editing the reported file, so Orbit does not return an unchanged inspect self-loop.
update_required	true	A newer Orbit binary is installed but the daemon still runs the previous version; the only action is orbit daemon restart --json.
stopped	true	The selected environment is configured but not running; configured resources are listed as stopped and the only action is orbit up --json, or orbit up -c <config> --json when the config was given with -c (each with --instance when one is set).
needs_daemon	true	A running daemon is serving a different env than the selected config.
degraded	false	At least one resource reports degraded.
converging	false	At least one resource is pending, starting, building, stopping, or restarting.
partial	false	At least one resource is stopped and no higher-priority risk exists.
ready	false	No inspect risk was detected.

For a terminal degraded resource with buffered output, inspect, status, and doctor lead only to orbit logs <resource> --json. After returning the exit output, logs rechecks live state. If a supported project dependency check is failing, its sole action installs the declared dependencies and then restarts only that resource; otherwise the sole action is orbit restart <resource> --json. Follow that sequence instead of skipping ahead: a newly observed port, dependency, or runtime cause can replace a blind restart with its safer cause-specific action.

converging may also be returned when daemon resource status is temporarily unavailable. In that case, a status_unavailable risk explains the condition.

Agents should treat blocked: true as a stop-and-recover state. Agents may run non-destructive recommended_actions directly, but must still require explicit human approval for any action marked destructive: true.

When readiness is needs_daemon, agents should run the recommended switch action before acting on service state because the running environment differs from the selected CLI config. Outside a discovered project that action is orbit up -c <config> --json: up switches only for a config it is explicitly given, selects nothing, and confirms like a project switch.

When readiness is update_required, agents must restart Orbit before sending resource mutations. Orbit returns orbit_update_pending from those mutations and recommends only orbit daemon restart --json.

Recommended Agent Workflow

This document owns the JSON contract, not agent decision policy. The Orbit skill defines the inspect, action, verification, failure-recovery, instance-targeting, and destructive-operation workflow. Agents should follow response recommended_actions before falling back to ad hoc debugging.

Exit Codes

Code	Meaning
0	Success.
1	Error. In --json mode, converted commands still emit structured JSON.

Exit status reports whether the command itself succeeded, not whether the reported environment is ready. In particular, orbit inspect --json exits 0 with ok: true when it successfully reports readiness.blocked: true. Agents must inspect data.readiness.blocked; orbit status --json may instead exit 1 when the selected state prevents status from being produced.