Relavium

Add a new LLM provider adapter behind the LLMProvider seam in packages/llm without leaking a single vendor SDK type across it: implement the seam interface, normalize system-prompt placement / tool schema / tool-call round-trip / streaming events / stop reasons / usage to the canonical Relavium types, map the provider into the generated model catalog for cost, wire it into the fallback runner config, and add the per-provider conformance test (recorded fixtures on PR, live nightly). USE FOR: integrating a new model provider into @relavium/llm. DO NOT USE FOR: adding a new top-level workspace package (use ../add-package/SKILL.md), changing the seam *interface* itself (that supersedes ADR-0011 — use ../write-adr/SKILL.md), or restating the seam contract (it has one canonical home in docs/reference/shared-core/llm-provider-seam.md).

Review a Relavium diff, PR, or branch against docs/standards/, the ADRs, and the CLAUDE.md non-negotiables, then report findings as file:line — severity — issue — impact — fix, sorted by severity. USE FOR: reviewing a proposed change before merge, walking a PR, or self-reviewing your own branch before requesting review. DO NOT USE FOR: implementing the change (use implement-task), a secrets/SSRF/sandbox deep-dive (use security-review), or general code questions unrelated to a diff.

The primary end-to-end flow for implementing any non-trivial change in Relavium — scope, plan, build against the standards, self-check the seam rule, run the toolchain, update docs, commit, and hand off to review. USE FOR: features, fixes, refactors, or new adapters/nodes that touch real code in packages/* or apps/*. DO NOT USE FOR: a one-line typo, a pure docs edit (use documentation-style directly), reviewing someone else's diff (use code-review), or a security-only pass (use security-review).

Run the binding Relavium security pass over a diff/branch — keys never leave the keychain or reach the frontend/logs, SSRF on custom provider base URLs, the run_command sandbox, prompt-injection posture, dependency provenance, and never hand-roll crypto. USE FOR: any change touching keys, the keychain bridge, IPC, network/base-URL handling, the command sandbox, prompt/tool-call construction, the DB encryption path, or a new dependency. DO NOT USE FOR: a general correctness review (use code-review) or implementing the change (use implement-task).

Supersede an existing Accepted ADR by writing a new one, flipping the old Status to "Superseded by", and repointing live references — never deleting or rewriting the old body. USE FOR: overriding a settled decision, reversing a prior choice, replacing one approach with another. DO NOT USE FOR: a first-time decision (use ../write-adr/SKILL.md) or a typo/clarity fix to an existing ADR (ADRs are append-only history, not living docs).

Format Conventional Commits with a per-package scope, reference the ADR/task in the body, end with the Co-Authored-By trailer, and write a PR description with a conformance checklist. USE FOR: committing finished work and opening a PR. DO NOT USE FOR: authoring a decision (use ../write-adr/SKILL.md), scoping work (use ../start-task/SKILL.md), or the pre-commit conformance screen (use ../standards-check/SKILL.md first).

Fast, mostly-grep conformance gate over a diff before deeper review — strict TS/no-any, no new dep without an ADR, no vendor type across the @relavium/llm seam, engine zero platform imports, secrets never plaintext/in logs, one-canonical-home docs, conventional commits. USE FOR: a quick pass/fail screen before requesting review or committing. DO NOT USE FOR: a full code review (use the relavium-reviewer agent / code-review.md) or a deep security audit (use security-review.md).

Author a new Architecture Decision Record in docs/decisions/ using the condensed MADR form and update the index. USE FOR: recording a settled non-trivial architectural/product/process decision, adding a new runtime dependency, justifying a stack or seam choice. DO NOT USE FOR: replacing an existing Accepted decision (use ../supersede-adr/SKILL.md), restating a spec (that belongs in docs/reference/), or scoping a unit of work (use ../start-task/SKILL.md).

Scaffold a new workspace package under packages/ or an app under apps/ in the Relavium Turborepo: package.json (@relavium/<x>, workspace:* deps), a tsconfig extending the base, shared ESLint, Vitest config, a curated src/index.ts, a test skeleton, and turbo wiring — with the import boundaries and no-circular-deps rule enforced. USE FOR: adding a new shared package or a new app surface to the monorepo. DO NOT USE FOR: adding a provider adapter inside the existing packages/llm (use ../add-llm-adapter/SKILL.md), recording the decision to add a package or dependency (use ../write-adr/SKILL.md), or authoring a spec (that has one canonical home in docs/reference/).

Interview the user relentlessly — one question at a time — to stress-test a plan or design until you reach shared understanding, resolving each branch of the decision tree and offering your own recommendation with every question. USE FOR: pressure-testing a design/plan before building, "grill me", "poke holes in this", "review my plan", interrogating a proposal for Relavium-specific risks. DO NOT USE FOR: recording the settled decision (use ../write-adr/SKILL.md), scaffolding the work (use ../add-package/SKILL.md or ../add-llm-adapter/SKILL.md), or reviewing an existing diff.

Open a scoped unit of work from a roadmap phase workstream — define in/out scope, affected packages, dependencies, and concrete acceptance criteria, then stop at the plan. USE FOR: turning a phase workstream (e.g. Phase 1 "1.K FallbackChain runner") into an actionable, bounded plan before writing code. DO NOT USE FOR: implementing the work, recording a decision (use ../write-adr/SKILL.md), or committing (use ../commit-and-pr/SKILL.md).

Add or update a docs/architecture/*.md explanation that leads with a Mermaid diagram and CITES the reference/ specs rather than restating them, preserving one-canonical-home. USE FOR: explaining how a topology/flow/model is built (engine, execution, security, multi-LLM). DO NOT USE FOR: authoring an exact spec (that lives in docs/reference/), a decision (use ../write-adr/SKILL.md), or a how-to (that is docs/runbooks/).