CLI reference

The installed help output is authoritative:

mco --help
mco run --help
mco review --help

Commands

Command	Purpose
mco run	General multi-Agent invocation
mco review	Thin read-only raw-answer preset over the same invocation runtime
mco doctor	Provider presence, auth, version, risk, and Skill checks
mco agent	List, inspect, and discover provider models
mco skills	Read, inspect, and sync the bundled Skill
mco session	Persistent multi-turn provider sessions
mco serve	Run the MCP server

The old findings command and findings-oriented memory commands were removed. Calling them returns migration guidance and does not start a provider.

Invocation selection

Option	Purpose
--agent [alias=]provider:model	Repeatable explicit invocation declaration
--providers provider,...	Invocation-native shorthand: one configured/default model per provider
--provider-models-json	Model mapping used by the provider shorthand
--target-paths	Comma-separated task scope paths
--task-id	Safe stable task/artifact identifier
--prompt / --file	Inline prompt, prompt file, or stdin with --file -

Aliases must be unique. Repeating the same provider/model without distinct aliases is rejected. Configuration and provider/model validation happen before any Agent invocation starts.

Runtime and access options

Option	Default	Purpose
--execution-mode	run: write; review: read_only	Provider permission profile
--allow-paths	.	Fail-closed MCO scope boundary
--enforcement-mode	strict	Reject unsupported provider policy or use best_effort
--provider-permissions-json	unset	Provider-specific permission overrides
--provider-context-json	unset	Provider context policy
--provider-timeouts	unset	Provider-specific invocation hard-timeout overrides
--invocation-hard-timeout	180	Per-invocation wall-clock deadline; 0 disables
--stall-timeout	900	Maximum time without Provider output progress
--review-hard-timeout	1800	Global task deadline; 0 disables
--max-provider-parallelism	0	Parallelism policy for configured execution

Output options

Option	Purpose
--result-mode stdout	Stream/return answers and clean up temporary artifacts
--result-mode artifact	Persist artifacts and return the operational result
--result-mode both	Persist artifacts and stream/return answers
--save-artifacts	Upgrade the default stdout mode to both
--json	Print one final machine-readable envelope
--stream jsonl	Print machine-readable event lines as invocations progress
--stream live	Human live mode; non-TTY output falls back to JSONL
--include-token-usage	Preserve reliable provider usage metadata when available

--json, --quiet, and --stream are mutually exclusive. In JSON/JSONL modes stdout contains only the selected protocol. Provider diagnostics and progress warnings go to stderr.

Multi-stage options

Option	Behavior
--chain	Run invocations sequentially and pass complete prior Markdown through a manifest
--debate	Add a read-only stage over prior raw answers
--synthesize	Add a read-only synthesis stage over the available run/debate raw records
--synth-provider	Select the provider invocation used for synthesis
--perspectives-json	Add an explicit per-provider prompt perspective
--divide files|dimensions	Assign non-overlapping scope files, or declaration-ordered rotating review lenses, without interpreting answers

Chain, debate, and division are mutually exclusive. Perspectives and division are explicit prompt/scope coordination only: --perspectives-json prepends a Provider-specific Review Perspective; --divide files sorts repository files in the selected scope, excludes ignored/local/build directories, and assigns the remaining files round-robin in declaration order; --divide dimensions rotates the fixed review lenses in that order without changing target_paths. Dry-run shows the complete resolved invocation prompts and target paths. These options preserve raw invocation answers and never derive semantic findings or consensus. Debate and synthesis mark earlier answer files as untrusted reference material. A valid earlier answer allows later stages to continue after a partial failure; no valid input produces an explicit dependent-stage failure.

mco review uses the same runtime. Its default prompt is a short natural-language review request, and an explicit --prompt is passed unchanged. It never injects a findings schema.

Artifacts

With persistent result mode, <artifact-base>/<task-id>/ contains:

result.md
run.json
stages/<stage>/invocations/<invocation-id>.md
stages/<stage>/context/manifest.json
stages/<stage>/result.md
stages/<stage>/run.json
provider-runs/                 # internal transport/provider evidence

Each stage is deterministic in declaration order. Root result.md groups the stages that actually ran; synthesis comes first when present, while every raw answer and explicit failure record remains below it. Per-invocation Markdown preserves the decoded Agent answer body. Temporary execution removes its task directory and reports artifact_root: null.

Exit codes and task status

Code	Meaning
0	All invocations completed successfully (complete)
1	At least one invocation succeeded and at least one did not (partial)
2	No invocation completed successfully, or input/configuration failed (failed)

Invocation-level status remains explicit: success, failed, timeout, or cancelled. The task status is only complete, partial, or failed.

Removed surfaces and migration

The findings command/schema, semantic normalization, deduplication, confidence, consensus, passive lifecycle, findings-driven memory, Markdown-PR, SARIF, and content-based INCONCLUSIVE surfaces are gone. The old --format, --strict-contract, --memory, --space, --diff, --staged, --unstaged, and --diff-base flags return migration errors. Use --target-paths and a raw prompt for scope, then select text, JSON, JSONL, or file-backed artifacts.