Full-featured composed scenes — one per row so each design has room to breathe. Click for the live render, the spec, and the source.
GO · safety primitive · live drill
Step 1 / 3
The hinged cover prevents accidental fire. Lift it to expose the dome. Unarming reseals the cover automatically after 30s of inactivity.
Step 2 / 3
Two-key challenge prevents reflex strikes. Type HALT to confirm.
Step 3 / 3
Autonomy ceiling dropped to 0%. In-flight outputs are quarantined for human review. Audit-log entry written.
14:02:11 CEILING autonomy → 0% 14:02:11 FREEZE 12 agents · 4 toolchains 14:02:12 AUDIT halt · operator=alex · reason=DRILL
DO · intent radar
GO · sign-off ceremony
Every 30 days the signing key must rotate. Old verifications continue to succeed during a 24-hour grace window. No request returns 401 due to rotation alone.
Draft spec under review
A structured adversarial review where three agent personas interrogate a draft spec.
Spec is sent to each persona in parallel; each returns one sharpest unresolved question.
A ranked list of objections gated by severity, plus a verdict band the author must clear.
What happens when the upstream token service is partially degraded — only some keys served?
Failure Mode 2 mentions clock skew but never bounds the acceptable window. Define it.
Runbook entry for the 401-with-refresh-hint case is missing from on-call playbook.
What you get
What it is
A heuristic 0–100 score across four dimensions, computed on any markdown spec.
How it runs
CI calls the scorer on changed spec files. Below 60 the build fails; the diff carries the per-dimension deltas.
CI gate snippet
ics-score specs/*.md --gate 60 \
--report ics-report.json \
--fail-on-regressionProposed plan
What it is
An explicit per-step approval gate sitting between the agent's plan and any side-effect.
How it runs
Each step is tiered — read, write, sandbox, production. Auto-approve thresholds live in policy; everything above the line waits for human assent.
What you get
Spec ↔ code drift map
A continuous diff between the canonical spec and the running implementation.
Detector samples production traces and source-of-truth specs on a cadence; structural drift always pages.
A drift budget. Three structural events in a sprint and the convergence loop fires.
Each arrow is an explicit contract. The envelope carries only the payload — never the originating agent's chain-of-thought.
Context payload
What you get
A reproducible handoff — the next agent gets exactly the same context the previous one finished with, nothing more.
Workflow with overlay
Candidate pattern
withIdempotency(key, () =>
retry({ attempts: 3, jitter: true }, fn))Evidence trail
Outcome
Promoted to convention. Future PRs that reinvent retry get a soft suggestion to adopt this pattern.
The ladder
Current rung
What it is
A composed team of role-tagged agents converging on a single artefact.
How it runs
Each role has a tool tier. None of them edit the artefact in place — they propose. The Reviewer is the only role that merges.
What you get
One artefact per task. Four explicit contributions. Zero shared mutable state.
Write a failing test that pins the intent.
Make the test pass by the smallest change.
Clean the shape without changing behaviour.
// Never refactor on red. Never skip green.
Most recent first
What it is
A hash-linked append-only log of every governed action, with role attribution.
How it runs
Each entry hashes the previous entry's hash into its own. Rewriting history breaks every downstream link.
What you get
A timeline you can replay during incident review — and one you cannot quietly edit.
DoCoDeGo is in Alpha. The framework is documented, the practices are battle-tested at small scale, and the next release is being shaped in public.
If it produces anything, it should produce engineers and teams who think more clearly about what they are building and why.
Discord is where specs are debated, the framework gets sharper, and decisions land in writing. The conversation is the artefact.