agentic-sdk

Review dimension for naming, idiom, comment debt, and AI tells against the codified per-language and prose standards. Invoked by reviewer agents over one module shard.

Recipe for writing C in the runtime, module placement and the ownership and control-flow discipline. Invoked by writer agents when a C unit is dispatched.

Recipe for writing Clojure, Functional Core / Imperative Shell at the function level (namespaces by domain), native wrappers, and the boundary discipline. Invoked when writing Clojure for the project.

Recipe for writing Elixir, pure functions in modules as the core and GenServer plus OTP as the shell, supervision trees, and NIF discipline for the C and Zig edge. Invoked when writing Elixir for the project.

Recipe for any English the agent produces in this project, commits, comments, docstrings, ADRs, design docs, skill bodies, changelog, error messages, project guides. Optimizes for terse, humanized prose with no em dashes and no AI tells. Invoke for any prose-writing activity.

Recipe for writing Zig, native bodies and edge wrappers, the allocator and lifetime discipline, and the hot-path rules. Invoked when writing or editing any Zig source.

Review dimension for module boundaries, dependency direction, pure/effect and native leakage, duplication, and oversized units. Invoked by reviewer agents over one module shard.

The authoritative procedure for one review round: lanes, reviewer fan-out, spine triage, level-ordered editor waves, integrate, verify, advance.

Generate a check-<dimension> recipe for a bug class the active set misses

Generate a write-<lang> recipe for a language outside the curated four

Add a non-language concern, a lane entry, a spine task, or a hook

Drive a chunk of the plan to completion unattended, phase by phase

Trace the causal chain from trigger to failure

Assess pre-deployment risk with rollback plans

Review and fix any scope, round after round until a dry round

Detect the stack, write the descriptor, materialize recipes and scaffolding

Capture a correction or interaction preference mid-session at near-zero cost, route it to the right home, and return to the work.

Create a prioritized backlog from the technical design

Define testable, scoped requirements

Clarify the user problem, constraints, and success criteria

Design architecture, components, and data flows

Establish UX guardrails and interaction patterns

Fix one bug end to end, from reproducer to one commit

Build one change end to end on the current tip, review-clean and ready to land

Promote captured guidance from the inbox into its durable home, empty the inbox, one commit.

Triage an incident to root cause and a follow-up plan

Turn a backlog item into slices, a Gherkin spec, and a plan

Turn a problem or idea into an approved backlog and feature plan

Decompose a chunk of the approved plan into a forward-only DAG of phases and tasks, write it to plan.edn, return a compact summary

Run a blameless incident review with action items

Surface risky assumptions and missing acceptance criteria

Cut a release end to end, from lane gate to tag and trunk advance

Turn noisy signals into structured analysis

The verification lane table. Runs the descriptor's three lane tiers in order and reports pass or fail per lane. Invoked by the verifier agent and the review-round-runner.

Fix the punch-list findings assigned to you, one module at one level, smallest sufficient edit, verify then land. Run by editor agents.

Design a UI surface before it is implemented, the four-layer procedure (intent, information architecture, interaction, presentation) chosen from tokens, plus the full set of states. Invoked before write-ui when a new or changed UI surface needs a design.

The project's interactive eval loop. Connect once, evaluate small forms, read the value, adjust. The discipline the write-time recipes run inside.

Produce a compact module brief for a dispatched agent, built from disk, returned as the final message, never written to disk.

Toolchain maintenance. The pin, the doctor check, the bump ritual, and the investigated-ceiling cache. Invoked by the verifier agent and maintainer dispatches.

Run a risk spike: name the load-bearing unknown, time-box it, build a throwaway harness that exercises the seam at and beyond expected load, decide go or no-go or mitigate, feed the outcome forward.