feat: add guided recipe and operator authoring

[zimuwang-real]

What

• turn the public evolve-agent skill into a backend-neutral router for initialized workspaces, RSIHub source checkouts, partial/ambiguous contexts, and external target projects
• add guided experiment design, recipe authoring, reusable named-operator authoring, deployment, approval, recovery, and replay playbooks
• expose one canonical skill to Codex through .agents/skills/evolve-agent and Claude through .claude/skills/evolve-agent
• keep both platform wrappers identical and thin; they delegate to skills/evolve-agent/SKILL.md without copying its workflow
• add machine-readable invocation/behavior coverage, packaging checks, and frozen snapshot-protocol metadata

Why

Users should be able to invoke the same repository-local skill from Codex or Claude, collaborate with their own agent to define requirements, choose or author recipe/operator modules, qualify the evaluator first, and deploy only after explicit architecture, source, and deployment decisions.

The repository paths follow the current platform conventions: Codex discovers project skills from .agents/skills; Claude discovers them from .claude/skills. No nonfunctional .codex/skills compatibility copy is added.

Safety and behavior

• guided authoring accepts named operator: bindings only and hard-stops on script:
• inspection requires a verified pre-existing executable in a read-only, credential-free boundary
• source/deployment approvals bind immutable identities through externally anchored hash-chained records
• remote targets require a recipe-pinned revision; local targets use exact copy-semantic manifests and immutable snapshots
• prospective preflight stops unless trusted containment and allowlist-sanitization tooling is actually provisioned
• parent replay requires exact candidate, evaluator, runtime, artifact, purpose, and partition identities; children always execute fresh

Evaluation clarity

• snapshot_protocols.json stores frozen protocol/inventory metadata for historical snapshots that did not embed complete scoring semantics
• existing 16 behavior-case rubrics and historical score snapshots remain unchanged
• the current inventory is 29 behavior cases, including 13 newly unreported cases, plus 8 invocation cases
• no formal paired results were fabricated for the new cases

Impact

The feature diff is limited to skills, repository discovery wrappers, documentation, evaluation assets, and packaging/tests. It does not add feature-authored runtime code or dependency/lockfile changes. Repository-only wrappers are excluded from the sdist; the canonical skill remains present and byte-identical in the sdist and wheel.

Checks

• uv run --frozen pytest -q: 1229 passed, 3 skipped
• focused adapter/skill/eval/resource/public and rename-affected contracts: 71 passed
• final controller-focused check: 33 passed
• Ruff check and format check: passed
• ty: passed
• fresh RSIHub sdist and wheel: built successfully
• release-artifact tests against the fresh distribution: 2 passed
• independent final updated-base review: spec compliant, no Critical, Important, or Minor findings; ready to merge

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: fc6198a1-e2f4-4d0d-a2b1-d9e05bb94d22

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}