Skill Steward MCP

steward_cli

Bounded stewardship CLI for Skill Steward. Product CLIs live in their own repos — see ADR 0006.

Commands

The CLI is bounded by stewardship ownership, not by command count. Use these public jobs as the mental model; they are not aliases unless a command is shown below.

Command stability tiers:

Cold-start proof starts with the claim question: what can this repo honestly claim right now? The loop assumes the steward command is installed or activated; see Portable invocation hierarchy for the supported paths.

steward doctor --jsonsteward schema check-outputs --jsonsteward schema drift --jsonsteward actions list --jsonsteward action inspect steward.contract.status.quick --jsonsteward probe --profile quick --jsonsteward benchmark \  --scenario skill_steward.contract-status-smoke \  --output .steward/benchmark-summaries/skill-steward-contract-status-smoke.json \  --strict \  --json

The schema commands sit in the loop because this dogfood route depends on machine-readable payloads. They catch route drift; they do not prove repo maturity.

Benchmark execution is durability-gated. If source.steward_contract or a file-backed scenario manifest is modified or untracked, the summary must return result: blocked with blocked_by: durability_blocked; track or commit those contract inputs, then rerun the same benchmark for executable proof. The built-in contract-status-smoke scenario proves contract discovery and durability gating only when it returns result: "pass"; it does not prove full agent navigation or diagnosis.

When a fresh command returns blocked JSON, route it directly:

steward benchmark --scenario <scenario> --strict --json | steward blocked explain --stdin --json

Use steward blocked explain --input .steward/benchmark-summaries/<scenario>.json --json only when routing an existing persisted summary. The explanation chooses the next artifact route; it does not repair the repo or turn blocked evidence into proof.

Repository ecology review inventory:

steward ecology snapshot --jsonsteward ecology route --json

The snapshot gathers Steward config status, declared action/probe inventory, benchmark declarations and summaries, schema output status, git state, active plan candidates, and evidence pointers. ecology route composes those facts into North Star dispositions: orient, compress, validate, tutor pain, promote tools, leave native, or stop. It may also emit advisory dispatch_lane_candidates for repo-wide pain, but those hints are not parent lane contracts and never authorize writes. Both commands are read-only; neither executes actions, applies repairs, awards maturity, or proves adoption.

Snapshot benchmark summaries are persisted history from .steward/benchmark-summaries/. The JSON labels them as persisted_history and includes commit/freshness hints; treat may_be_stale: true as a routing signal, not proof failure. Pipe fresh blocked JSON to steward blocked explain --stdin --json; use --output only when a fresh benchmark result should replace persisted history or feed future snapshots.

Minimal evidence bootstrap:

steward evidence init --minimal

This creates one current ledger at docs/evidence/current-status.mdx. It does not create a proof archive, benchmark, or readiness claim. Use it when the repo needs a durable current-status pointer before it needs typed actions, probes, or benchmark machinery.

Evidence growth loop:

steward observe --profile quick --jsonsteward unknown-case create --from .steward/observations/<observation>.json --jsonsteward action-candidate create \  --from .steward/unknown-cases/<case>.json \  --id example.safe.check \  --desc "Run the reviewed bounded check" \  --argv-json '["bash","tool/contracts/example_check.sh"]' \  --fs-read tool/contracts/example_check.sh \  --git false \  --benchmark example.safe-check-smoke \  --jsonsteward action-candidate inspect .steward/action-candidates/<candidate>.json --jsonsteward action-candidate review --from .steward/action-candidates/<candidate>.json --json

An action-candidate review is evidence, not a contract rewrite. Generated candidates include can_promote_in_this_run: false; promote only after repeated evidence, owner review, narrow effects, native validation, and a strict benchmark.

Development

cd packages/steward_clidart pub getdart analyze --fatal-infos   # xsoulspace_lints (library.yaml)dart run :steward validatedart run :steward doctor --jsondart run :steward list

From repo root: pnpm run steward:analyze

Portable invocation hierarchy

Use the first option that matches the job:

1. Released adopter / CI: install the released binary, then run steward <command>.

   bash

   curl -fsSL https://raw.githubusercontent.com/Arenukvern/skill_steward/main/install.sh | bashsteward doctor --json

2. Dart maintainer from this checkout: run from source without hard-coded SDK or package-config paths.

   bash

   cd packages/steward_clidart run :steward doctor --json

3. Local clone global activation: activate the checkout package, then use the normal steward command.

   bash

   dart pub global activate --source path packages/steward_clisteward doctor --json

Raw dart --packages=... bin/steward.dart commands are local provenance only. Do not publish them as adopter instructions; they encode one machine's SDK and package-config layout.

Global install from a source checkout (optional):

dart pub global activate --source path packages/steward_clisteward validate

Distribution

Maintainers: run from a repo clone (commands above). Consumers install public skills with npx skills add arenukvern/skill_steward. Use steward install/update only when a repo intentionally owns a pinned skills.json skill layer for agents and maintainers.

Skill Steward may ship zero-dependency steward binaries for consumer/bootstrap use; see ADR 0014. Product repositories still own their branded harness binaries and release contracts — see skill release-changelog-harness -> references/binary-release-contract.md.

MCP Registry (experimental)

• Registry server: io.github.Arenukvern/steward-mcp
• Manifest: server.json; OCI image: packages/steward_cli/Dockerfile.registry
• Boundary: read-only v1 tools (steward_list_actions, steward_inspect_action, steward_doctor, steward_explain_result); no action execution over MCP
• Local stdio after binary install: steward mcp (see DX_FAQ)
• OCI smoke: docker run --rm -i -v "$PWD:/workspace:ro" -w /workspace ghcr.io/arenukvern/steward-mcp:<version>
• Version drift check: pnpm run mcp:registry:check (also runs in scripts/tag_release.sh)

Decisions

• ADR 0006 — meta vs product CLIs
• ADR 0007 — Dart for meta CLI
• ADR 0008 — Skill Steward product name
• ADR 0010 — binary release scope