▸ 18 skills · Claude Code + Codex · skills.sh

Spec to ship, unattended.
Then it learns.

Eight autonomous phases behind one approval, bounded by deterministic evidence. Every run returns what it learned to the next one.

▶ run the whole lifecycle

One invocation. One plan approval. One bounded pass.

CClaude Code

/autonomous-pipeline "<feature>"

Run after installing the marketplace plugin. The plan is the one broad approval; surprises still stop.

XCodex plugin

$autonomous-engineering:autonomous-pipeline "<feature>"

Work in progress on the codex-native-plugin branch: build it from source, start a new session, then run. Missing capability and verification failures stay explicit.

How the full run works →Install for your platform

↻ the self-improving pipeline

▸ two native surfaces

Same contracts. Different machinery.

Use the platform you already trust. The methodology stays bounded and fail-closed; the adapters respect each provider’s actual capabilities.

C Claude Code

Available now

Plugin-native enforcement.

Install all 18 skills as slash commands, with SessionStart, PreToolUse, Stop, and SessionEnd hooks that reinforce resume, live guards, honest termination, and checkpoint cleanup.

18 canonical skills
Lifecycle and skill guards
Resume and honest-stop hooks

View Claude Code details →

X Codex

Work in progress

Provider-native control plane.

The Codex-native plugin is in development on the codex-native-plugin branch. It packages all 18 skills as isolated, plugin-qualified adapters, backed by deterministic coordination, pinned codex exec contracts, durable checkpoints, new-session discovery, and reproducible install verification — built from that branch, not yet shipped on main.

18 plugin-qualified adapters
Deterministic coordinator + pinned executor
Reproducible package + gated smoke

View Codex details →

▸ the problem

Every agent can loop. The failure modes are what kill you.

These skills separate autonomy from wishful prompting with caps, checkpoints, independent evidence, and tested guards.

✗

Fakes the green

Common skip, only, suppression, bulk-stage, and irreversible-push spellings are blocked by tested guards. Deeper cheats are caught by mutation and evidence checks.

✗

Grinds forever

Iteration, agent, time, and no-progress limits make non-convergence an honest partial result instead of an invisible runaway.

✗

Poisons the history

Disjoint task scopes, isolated writers, explicit staging, and per-task commits preserve reviewable rollback points.

✗

Lies about done

Required evidence is observed by the coordinator. Missing validation, dead verifiers, and skipped gates cannot be worded into success.

▸ the collection

Eighteen skills. Four layers.

Run one phase, compose the primitives, or take the whole lifecycle through a single governed pass.

Lifecycle phases

Eight fail-closed phases carry a change from a grounded request to release-ready work.

Explore 8 skills →05

Control primitives

The bounded loops, skeptic panels, durable state, parallel maps, and stop conditions underneath every run.

Explore 5 skills →03

Autonomy layer

Compose phases into a full run, bridge external signals, and schedule repeat-safe standing work.

Explore 3 skills →02

Context & learning

Verify the project map and feed proven lessons into context the next session loads automatically.

Explore 2 skills →

▸ deterministic enforcement

Prompt contracts bend. Evidence does not.

The guards cover specific static shortcuts; the coordinator and skill contracts cover the deeper cases they cannot safely infer from a single tool call.

guard → bounded claim → evidence
Guard	What it blocks	Verified by
`guard-test-integrity`	Common static suite-gaming forms added to test files: .only, .skip, xit, and suppressions.	scripts/test-guards.sh
`guard-git-hygiene`	Bulk staging, destructive reset or clean, commit -a, and unsafe force-push forms during a live build.	scripts/test-guards.sh
`guard-ship-irreversibles`	Unapproved forced updates and remote ref deletion while ship preparation is active.	scripts/test-guards.sh
`checkpoint.mjs`	Reinitializing a live run, demoting done work, and finalizing an incomplete checkpoint as done.	scripts/test-checkpoint.sh

See the full contract →

▸ universal skill install

Start with all eighteen.

Install individual skills through skills.sh, or add the full Claude Code plugin for a provider-native surface. The Codex-native plugin is a work in progress.

npx skills add https://github.com/ulpi-io/skills-autonomous-engineering

Compare plugin distributions →

Spec to ship, unattended.Then it learns.