H4 fresh-agent workflow — 2026-06-10

Status: H4 bounded workflow passed · Audience: maintainers and agents

Claim

A fresh agent completed one bounded mcp_flutter workflow using declared repository surfaces before raw exploration: prove the first Skill Steward dogfood scenario status. This is an H4 proof for the bounded contract-status workflow, not a claim that all product or runtime workflows are H4.

Task

Prove the first Skill Steward dogfood scenario status for mcp_flutter using declared repository surfaces before raw exploration.

Repo state

git status
# On branch codex/steward-adoption-h2
# nothing to commit, working tree clean

git status --short --branch
# ## codex/steward-adoption-h2

Declared surfaces used first

1. AGENTS.md Governance & Skill Steward section.
2. makefile / make check-contracts native gate.
3. At proof time, tool/steward/run.sh was the temporary Skill Steward bridge named in AGENTS.md; current reusable guidance uses the released steward executable.

Transcript

bash tool/steward/run.sh doctor --json
# Passed after selecting a compatible Steward command surface.

make check-contracts
# Passed.
# Existing warnings only: skill metadata warnings for missing license and references/sources.md.

bash tool/steward/run.sh doctor --json
# Passed: config valid, no diagnostics, one quick-eligible action.

bash tool/steward/run.sh actions list --json
# Passed: listed fmt.check.tool-prefix.

bash tool/steward/run.sh action inspect fmt.check.tool-prefix --json
# Passed: bounded_local, auto, no writes, no git, no network, no secrets, no destructive effects.

bash tool/steward/run.sh probe --json --profile quick
# Passed: executed fmt.check.tool-prefix, exit_code 0.

bash tool/steward/run.sh benchmark --scenario mcp_flutter.web-dogfood-warm --strict --json
# Passed: result=pass, proof.status=ready, durability.status=ready.

Outcome

• Fresh-agent workflow: passed.
• Fallback needed at proof time: yes. The global steward binary was present but stale for the declared contract workflow, so the agent used the then-documented temporary wrapper and local Skill Steward override.
• Files edited by fresh agent: none.
• Outputs persisted by fresh agent: none.
• Review artifact added later: docs/evidence/generated/mcp_flutter.web-dogfood-warm.strict.redacted.json.
• Current adoption note: the temporary wrapper was not promoted. New adopters should install and run steward directly.

Remaining gaps and non-claims

• This proves one bounded contract-status workflow, not the WebMCP runtime dogfood path.
• The scenario source.commit names the pre-proof subject commit. The benchmark treats a local HEAD mismatch as a warning rather than remote equivalence proof.
• Existing skill metadata/source warnings remain in make check-contracts; they do not fail the native gate.
• The original raw sibling-checkout command and temporary wrapper are local provenance only and are no longer reusable adopter guidance.
• H5 remains unclaimed because no repeated evidence has promoted a reusable diagnostic, action, eval, or harness capability with owner, risk, redaction, and validation.