MCO CLI Skill

Use this Skill when an upstream Agent needs to dispatch one task to one or more coding Agents and preserve their raw operational answers.

Before execution

Confirm the provider/model team with the user in natural language. Do not infer consent from installed binaries. Use either:

• --agent [alias=]provider:model once per invocation; or
• --providers provider,... for one default-model invocation per provider.

--agent is the model-qualified form. It supports repeated models from one provider when each repeated invocation has a distinct alias:

mco run --repo . --prompt "Inspect the parser." \
  --agent fast=pi:fast-model \
  --agent careful=pi:careful-model \
  --result-mode stdout

--providers pi,codex is an invocation-native shorthand. MCO resolves each provider's configured/default model and runs the resulting invocations through the same runtime; it is not a legacy execution path.

If MCO returns provider_selection_required, ask the user which Agents to use and retry with an explicit --providers or --agent selection.

Execution defaults

• mco run defaults to --execution-mode write.
• mco review is a thin read-only preset and defaults to --execution-mode read_only.
• Review passes the explicit prompt unchanged. With no prompt, it uses a short natural-language review prompt; it does not inject a findings schema.
• Use --execution-mode yolo only after the user explicitly requests unrestricted execution.

Output and status

MCO returns raw Agent answer text plus operational metadata. It does not infer findings, severity, confidence, consensus, or decisions from answer prose.

• Default text mode streams answer deltas to stdout. With multiple invocations it adds source labels between answers but does not rewrite answer content.
• --json prints one final JSON envelope.
• --stream jsonl prints JSONL events; each invocation's output_delta values concatenate to its formal answer.
• Progress, warnings, and provider diagnostics go to stderr. stdout remains answer text or the selected machine protocol.
• status is complete, partial, or failed; exit codes are 0, 1, and 2 respectively. A cancelled or timed-out invocation is recorded inside the corresponding task status.

For example, if fast succeeds and slow times out, the final envelope is status: partial, exit_code: 1, and still contains fast's complete answer plus slow's explicit timeout record.

Use --result-mode stdout|artifact|both. Temporary runs clean up their task directory and return artifact_root: null. Persistent modes write deterministic artifacts:

<artifact-base>/<task-id>/
  result.md
  run.json
  stages/<stage>/invocations/<invocation-id>.md
  stages/<stage>/context/manifest.json   # chain/debate/synthesis inputs
  stages/<stage>/result.md               # staged result
  stages/<stage>/run.json                 # staged operational record

--save-artifacts is shorthand for upgrading the default stdout mode to both. Answer Markdown preserves the Agent's answer body as returned by the transport decoder. Root result.md groups actual stages, places the synthesis group first when present, and retains all raw answers and explicit failures.

Multi-stage workflows

• --chain runs invocations sequentially. Each next Agent receives a manifest and paths to the complete prior Markdown answer; MCO does not summarize, sample, truncate, or paste the previous answer into the prompt.
• --debate adds a read-only stage over the prior raw answer files.
• --synthesize adds a read-only stage using --synth-provider when specified, otherwise the first selected invocation. Its manifest keeps run and (when present) debate records, including failed or missing invocations, but never reads synthesis output itself.
• Debate and synthesis prompts mark earlier files as untrusted reference material. If a prior stage partially fails, later stages continue when a valid answer exists, including a run answer after debate failed; otherwise the dependent stage records an explicit failure.
• --perspectives-json prepends a named Provider prompt focus. --divide files excludes ignored/local/build directories and round-robins the remaining sorted repository files without overlap; --divide dimensions rotates fixed review lenses in invocation declaration order without changing target paths. Dry-run shows the full resolved prompts and target paths. These are coordination only: they do not parse, rank, or rewrite invocation answers. --divide is mutually exclusive with --chain and --debate.

Parallel writing safety

Parallel writing is an explicit user choice. MCO does not create or manage worktrees and does not silently prohibit multiple write invocations.

Before dispatching parallel writers, ask the upstream Agent to partition ownership into non-overlapping --target-paths and state the edit-conflict risk. Use distinct aliases and task IDs, and warn the user if two writers may touch the same file. Prefer a read-only comparison before choosing one implementation when ownership cannot be partitioned safely.

Progressive references

• Installation and Skill sync
• Provider selection
• Execution modes
• Multi-model and write ownership
• Troubleshooting and recovery

Breaking migration

The findings command, findings schema, semantic decision/consensus pipeline, Markdown-PR and SARIF renderers, content-based INCONCLUSIVE, and findings-driven memory surfaces were removed. Old flags such as --format, --strict-contract, --memory, --space, and diff-only review flags return migration guidance instead of being silently ignored.

Migrate callers to raw prompts plus --result-mode, --json, --stream jsonl, and file-backed chain/debate/synthesis stages.