the skill catalog · works in your agent, CI or a crew job · MIT

The tools your coding agent can’t build alone.

Skills for Claude Code, Codex, Antigravity, OpenCode and Pi; also in CI, or inside a wicked-crew job. The first one matters most: a proof gate that re-derives “done” from evidence instead of taking the agent’s word.

agent · buildasserted
the retry change: all acceptance tests pass
the agent says so. the gate checks the evidence itself.
  1. vault
  2. verifier
  3. hash
  4. attestation
checking the evidence…
hover to hold · pick a claim

Call a skill by name: in Claude Code, /wicked-garden-prove; in Codex, OpenCode, Pi or Antigravity, ask your agent to use wicked-garden-prove.

load out garden and its peers — one command

interactive · detects your CLIs · installs what you pick

01the gate

Play the lying agent. Watch the gate refuse.

Proof, not claims. Run the proof gate and “done” is re-checked from evidence: an unsupported claim is rejected, and by default the gate fails closed if its checks cannot run (an explicit claim-only fallback is labelled). Which entry paths block and which only advise is spelled out.

claim the retry change: all acceptance tests passgate: armed
pull the lever — the gate re-runs the verifier, re-hashes the recording and checks the vault is there
  1. vault
  2. verifier
  3. hash
  4. attestation
auto-playing — flip a switch or pull the lever to take control

It refuses where it runs: when a skill invokes it, in a compiled pre-push or CI trigger, or inside a wicked-crew job.

02the graph

Sees what grep can’t, with wicked-estate.

On estate’s code graph it finds the event and command links grep never sees, for blast radius, and its patch skill renames across files as a graph operation.

blast radius of maxRetry

grep -rn maxRetry 3 hits: retry.ts, queue.ts, worker.ts. notify.ts never shows up.

patch · one graph operation
retry.tsinterface Policy { maxRetry → maxRetries: number }
queue.tsif (n > policy.maxRetry → maxRetries) fail(job);
worker.tsconst { maxRetry → maxRetries } = policy;

the field appears in three files; the rename edits all three as one graph operation

auto-playing — pick grep or the graph, or run the rename

03the qe domain · absorbed wicked-testing

In acceptance, no agent grades its own homework.

Its acceptance pipeline never runs the reviewer inline with the test writer or the executor; those two also run as separate workers where the host can start them. The reviewer reads cold evidence files only: never the executor’s conversation or reasoning; command output reaches it only as recorded evidence.

report card · same changeself-grade vs the wall
author grades itself
PASS
always
independent reviewer
?
awaiting review…

illustrative: the same change, graded twice

inbox · reads only these
crosses: evidence filesbounced: executor context · chain-of-thought · executor’s notes · prior verdicts

Evidence lands in .wicked-qe/evidence/<run-id>/. The reviewer is a separate worker whose brief carries only evidence paths; if none can run, the verdict is FAIL inconclusive, never an inline self-review.

04the fleet

Forty specialists, five actions, one router.

40 qe-* specialist skills, grouped under five actions of the wicked-garden-qe router. Pick an action to see who reports to it.

auto-cycling

all five actions All 40, grouped by the router action they report to.

05a second opinion

A disputed decision gets a second opinion.

When a decision is disputed, a review panel weighs in using the other agents you have installed; a same-family fallback is labelled, and without a quorum it gives no verdict.

disputedship the retry change? the builder says yes, the reviewer found a race

agents you have installed

auto-playing
  1. codexhold: race in retry
  2. antigravityhold: race in retry
  3. opencode · dissentship
panel 2 of 3hold · the race in retry; 1 dissent recorded (ship)

06memory with a source

In a crew job, agents propose. You approve.

Cross-session memory and knowledge on wicked-estate: your agent stores decisions, ingests documents (PDFs included) and gets cited answers back. In a wicked-crew job an agent can only propose: each memory waits for your approval, with its source recorded.

1 · an agent in a crew job proposes
decision · proposed“cap retries at 1 until the race in retry is fixed”source: retry.ts · witnessed in this job
2 · your queue
nothing waiting
3 · three weeks later, a new session
“why do we cap retries at 1?”
the answer cites what you approved
auto-playing — approve or drop it yourself

What’s new in 12.66

  • Reports with sources: wicked-garden-learn has a report mode for wicked-crew’s six report jobs, from research to due diligence: every claim is numbered, cites its source and is re-checked there.
  • Respond and rework: draft answers an RFP, RFI or questionnaire from your materials, item by item, or reworks a document: condense, re-audience, generalize, convert or refresh.
  • Phase plugins: editor-scaffold also scaffolds a run plugin: a section in every studio job’s details, drawn from its events, reading only the topics you approve.

07the whole catalog

140 skills, by what they do.

The full surface, including the 40-specialist QE fleet, all reading the same evidence-first discipline. Everything here is real and in the repo today. Pick a family, or type a task.

type to match the families
quality engineeringthe wall ↑

The QE fleet — 40 specialists under five router actions, the 3-agent acceptance pipeline, and the prove gate. See the wall.

acceptplanauthorexecutereviewinsightprove
try it

One install brings what garden needs. The proof gate needs wicked-vault: the family installer adds it with garden, and without it the gate fails closed. wicked-estate adds the code graph and memory; wicked-bus and wicked-interactive are optional.

wicked-estatecode graph + memorywicked-vaultrequired by the gatewicked-busoptionalwicked-interactiveoptional

08build on it

Your domain. Your pack. Same gates.

The catalog is open — the built-in domains follow a naming contract anyone can follow. Ship a pack with your unique take — a fintech-risk domain, a games-QA fleet, a compliance cast — and it sits beside the built-ins, under the same evidence discipline.

acme-fintech-pack/
wicked-pack.jsonthe manifest
skills/acme-fintech/router · user-invocable
skills/acme-fintech-modeler/worker
skills/acme-fintech-auditor/worker · evaluator ≠ creator
a wicked-pack.json manifest + kebab-case SKILL.md names (≤64 chars)
one router per domain

A single user-invocable skill fronts the domain and routes its actions — the same shape as qe, mem, and product.

{vendor}-{domain}-{role} workers

Each specialist is its own skill that declares metadata.role: worker, routed by your domain’s router: the same shape as the 40 qe-* specialists.

evidence via vault

Your pack can call the same proof gate: its claims are re-derived against wicked-vault, and the gate fails closed when it cannot run.

in your pack’s directory, check it against the contract

09the bench

One command. The whole loadout.

The family installer is the fastest way in: one interactive command that detects your coding-agent CLIs and installs the products you pick (garden brings wicked-vault with it). Prefer just this plugin? It is the second tab.

the family installer · recommended
  1. detects the coding-agent CLIs you have
  2. you pick a bundle, or single products
  3. installs them; garden pulls in wicked-vault
  4. a readiness check shows what is ready

MIT · open-source · v12.66.1 · local-first: memory, evidence and the code graph stay in local files, with no server to run.
The wicked platform

One surface. One control plane.
One catalog. One record.

Your coding agents do the work. A different agent checks it against your rules. You get the evidence. Four layers make that work. Hover or tab through one to see what it does and what passes between them.

Experience Where you direct your agents The screens you use: start a job, watch it, answer the decisions that need you.
wicked-studiothe surface

The ADE (agentic development environment) for running a team of coding agents: start a job with a chat, build apps over your data, answer what needs you.

ws.wickedagile.com↗ repo
start a job · watch it · answer its questions — through one API
Control Intent in, gated work out Runs the jobs, checks every step, keeps the evidence: your coding agents as sandboxed workers.
wicked-crewthe control plane

Runs your agents in sandboxes, with one read-only gate to your databases. By default a different agent reviews the work, and every gate says what actually ran.

wc.wickedagile.com↗ repo
calls skills from sandboxed agents — a failed check sends the work back, waits for you or stops the job; every verdict lands in the record
Capability What agents can do Skills, tools, playbooks and the QE specialists: how agents touch the record.
wicked-gardenthe catalog

The skills your coding agent can’t build alone: proof from evidence, graph-aware refactors, 40 QE specialists.

you are here
reads & writes the record through its contract — never around it
Foundation The system of record Code graph · memory · knowledge · evidence · events. Local, with no servers to run.
wicked-estatethe record

What your agents know about your system: code graph, memory, knowledge and database structure, never rows.

we.wickedagile.com↗ repo
wicked-interactivethe document engine

The engine behind the docs, decks and captioned demo recordings studio produces. You use it through studio.

repo ↗