🧭 Let your coding agent off the leash — not off the rails. Guardrails, a hard budget cap & a self-fixing PR loop for Claude Code / Codex / Gemini. Eval-gated 100/100, you always merge.
This repository also configures its own agents. See what compass tells them →
Agent Claude Code
Designs API contracts — REST, gRPC/protobuf, and proto-first service interfaces. Use when shaping endpoints or messages, versioning a public surface, or reviewing an API for backward compatibility and devex. A wrong contract is expensive to unship, so it reasons about compatibility first.
Agent Claude Code
Designs the implementation approach for a non-trivial feature or change before any code is written. Use for multi-file, multi-service, or ambiguous work. Produces a concrete plan, identifies risks and the files to touch, and weighs tradeoffs. Read-only — never edits.
Agent Claude Code
Reviews a diff or set of changes for correctness, security, and convention adherence. Use proactively after writing a non-trivial chunk of code, before committing, or when the user asks for a review. Returns prioritized, actionable findings — not a rewrite.
Agent Claude Code
Designs schemas, writes safe migrations, and diagnoses slow queries — SQL (Postgres/MySQL) and common stores (Redis, Mongo, key-value). Use for schema/index design, migration safety, and query-plan tuning. Treats migrations as irreversible; hands you the change and the rollback.
Agent Claude Code
Roots out the cause of a specific failing test, stack trace, panic, or wrong behavior. Use when something is broken and the cause isn't obvious. Forms a hypothesis, proves it, then proposes the minimal fix.
Agent Claude Code
Writes and updates documentation — READMEs, ADRs, API docs, runbooks — matching the repo's existing voice. Use to document a feature, write an ADR, or refresh stale docs. Cheap and prose-focused.
Agent Claude Code
Implements Go changes idiomatically — services, gRPC, concurrency, tests. Use for focused Go feature/bugfix work, especially control-plane / backend services. Writes code and tests, runs vet and tests before handing back.
Agent Claude Code
Diagnoses Kubernetes/cluster state and drafts manifest changes (Kustomize/Helm/ArgoCD). Use to investigate failing pods, read cluster state, or prepare deployment changes. Reads freely; never applies or deletes — it hands you the change to run.
Agent Claude Code
Profiles slow or resource-heavy code and fixes the real bottleneck — latency, throughput, CPU, memory, allocations. Use when something is slow (not broken), before/after a suspected regression, or to validate an optimization with numbers. Measures first, changes one thing, measures again.
Agent Claude Code
Implements Python changes idiomatically — 3.11+, full type hints, pydantic/dataclasses, tests. Use for focused Python feature/bugfix work, especially services, data pipelines, and ML/inference glue. Writes code and tests, runs ruff and tests before handing back.
Agent Claude Code
Implements Rust changes idiomatically — hot-path services, async, error handling, tests. Use for focused Rust feature/bugfix work, especially latency-sensitive gateway/router/runtime code. Writes code and tests, runs clippy and tests before handing back.
Agent Claude Code
Deep security review of changes or a component — authz, secrets, injection, trust boundaries, crypto, multi-tenancy. Use when touching auth, encryption, tenant isolation, external inputs, or before shipping security-sensitive code. Read-only.
Agent Claude Code
Generates and hardens the test suite — unit AND end-to-end — for a change, runs it, and validates the behavior actually works (not just that assertions pass). Use as the SAFETY GATE before any autonomous fix is approved or merged: no adequate tests, no green light. The thing that makes the autonomous loops safe.
Agent Claude Code
Runs the test suite (or a subset), parses failures, and reports a tight summary with root-cause hypotheses. Use to execute tests and triage failures without spending driver-model tokens. Does not fix code unless asked.
Agent Claude Code
Implements TypeScript/Node changes idiomatically — strict types, discriminated unions, zod at boundaries, tests. Use for focused TS/JS feature/bugfix work, especially typed backends, SDKs, and server/client boundaries. Writes code and tests, runs typecheck and lint before handing back.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: