Steward H5 hosted cutover promotion — 2026-06-10
Status: bounded H5 promoted capability · Audience: maintainers and agents
Claim
fmt.check.intentcall-hosted-deps is the first bounded H5 Skill Steward capability in this repository. It promotes an existing deterministic product gate into a discoverable, inspectable, benchmarkable Steward action.
The useful outcome is not "another green benchmark." The useful outcome is that future agents can prevent local-only IntentCall path dependencies from hiding hosted dependency regressions.
Tool improvement packet
| Field | Answer |
|---|---|
| Observation | Hosted cutover can regress when a consumer pubspec regains agentkit/packages, intentcall/packages, or path: ...intentcall. |
| User value | Agents can keep mcp_flutter aligned to hosted intentcall_* packages instead of silently succeeding through local sibling paths. |
| Native owner | tool/intentcall/check_no_path_deps.sh and tool/intentcall/check_no_path_deps_test.sh. |
| Steward surface | Action fmt.check.intentcall-hosted-deps; scenario mcp_flutter.intentcall-hosted-cutover. |
| Safety and effects | bounded_local, auto policy, explicit fs_read, fs_write: [], git: false, network: false, secrets: false, destructive: false. |
| Redaction | steward/redaction/v1; benchmark summaries keep result, digests, owner, proof, durability, warnings, and subject commit. |
| Falsifier | bash tool/intentcall/check_no_path_deps_test.sh rejects synthetic stale path dependencies. |
| Positive proof | steward probe --profile quick --json runs the action on current state; strict benchmark result is pass. |
| Promotion criteria | Owner, risk, redaction, validation, falsifier, quick probe, strict benchmark, and non-claims are all recorded. |
| Educational note | Promote deterministic native gates that prevent real repo maintenance failures; keep runtime-heavy dogfood in a separate scenario. |
Tracked summary: docs/evidence/generated/mcp_flutter.intentcall-hosted-cutover.strict.redacted.json records the passing strict benchmark for proof snapshot commit bf21f6f1b0648f0fdcb0d97c84ab3174a0c23701.
Fresh-agent surface check
A separate fresh-agent exercise used declared surfaces first and ran:
steward doctor --json
steward actions list --json
steward action inspect fmt.check.intentcall-hosted-deps --json
steward probe --profile quick --json
bash tool/intentcall/check_no_path_deps_test.sh
git status --short
Verdict: the action was discoverable, inspectable, quick-probe eligible, bounded-local, and live-passed. The test rejected stale hosted-cutover path dependencies. The worktree remained clean during that exercise.
The exercise found one improvement: owner/risk/redaction/validation should be visible from the fresh-agent surface alone. The action evidence now exposes:
owner: flutter-mcp-toolkit
risk: hosted_dependency_cutover_regression
redaction: steward/redaction/v1
validation:
native_gate: make check-contracts
falsifier: bash tool/intentcall/check_no_path_deps_test.sh
benchmark: mcp_flutter.intentcall-hosted-cutover
Verification commands
Run from the repo root with a current released steward command:
steward action inspect fmt.check.intentcall-hosted-deps --json
bash tool/intentcall/check_no_path_deps_test.sh
make check-contracts
steward benchmark --scenario mcp_flutter.intentcall-hosted-cutover --strict --output .steward/benchmark-summaries/mcp_flutter.intentcall-hosted-cutover.strict.json --json