Portable skills for the fable-as-supervisor methodology: Claude acts as orchestrator/reviewer over an implementer agent (Codex CLI/app-server today), steering it with bounded, audited turns and judging its output with adversarial multi-model review passes.
Extracted and generalized from the bigname remediation pipeline. The skills carry the methodology; each consuming repo supplies its own bindings in a small profile file, so nothing repo-specific lives here.
.claude-plugin/ plugin + marketplace manifests
skills/
supervise-codex/ Detached Codex doctor/send/resume/review jobs plus
the stdlib-only Codex app-server bridge
supervise-kimi/ Kimi K3 High control channel: autonomous full-agent
turns plus read-only reviews; bundles a detached
stdlib-only job launcher
adversarial-review/ review standard: trust model, risk tiering, fable
subagent panel + optional Kimi K3 High external lens,
GitHub verdict lifecycle, merge gate
profiles/
PROFILE-TEMPLATE.md per-repo bindings template
As a plugin (recommended — skills stay updatable in one place):
claude plugin marketplace add <path-or-git-url-of-this-repo>
# then install/enable the fable-supervisor plugin, e.g. via /pluginOr copy the skills directly into a repo (<repo>/.claude/skills/) or
your user scope (~/.claude/skills/):
cp -r skills/supervise-codex skills/supervise-kimi \
skills/adversarial-review <dest>/Copy profiles/PROFILE-TEMPLATE.md to .claude/fable-supervisor.md in the
target repo and fill it in: workspace path, implementer identity and verify
loop, CI-mirroring gate commands, ledger/status doc paths, review-panel
config (supervise-kimi review external lens, review bot), merge gate and
style, known landmines.
The supervision and review skills read this profile first and refuse to guess missing bindings — an unfilled profile is a setup conversation, not a default.
- Implementer self-reports are evidence, not proof; audit the event log.
- One bounded instruction per turn: explicit deliverable, CI-exact gates, verify-loop-until-clean, scope tripwire.
- New semantic behavior (including in fix commits) → adversarial model panel, hand-verified, before any verdict.
- Reviewer never authors and approves the same change; the user holds merge and semantic authority (delegation per-PR only).
- A broken review lens is fixed or escalated, never silently skipped.
- Verdicts land on GitHub and in the ledger docs — durable artifacts, not chat.