A Agent Spine / v0.3.0
v0.3.0 — strict verification, profiles, and pilot evidence are live
Tag v0.3.0 License MIT Updated 2026 · 04 · 28
Codex-native core / Claude Code compatible / agent-ready repos

Make agent work inspectable, repeatable, and safe.

Agent Spine gives a repository one visible operating contract: AGENTS.md, native .codex hooks and subagents, reusable skills, verification gates, optional Claude Code compatibility, and merge-first adoption for real codebases.

View on GitHub Read v0.3.0 notes See the pilot
./scripts/install-agent-contract.sh ../existing-repo auto --dry-run
v0.3.0Current release
1 doneTarget-backed pilot
3 plannedPilot runbooks
strictVerification mode
§ 01Artifacts

What it installs — three files, one contract.

A · root
AGENTS.md

The canonical, repo-level instruction contract. The first file every agent reads — task framing, scope, files of record, and house rules.

1 file · plaintextspec ↗
B · runtime
.codex/

Native Codex configuration: hooks, policies, subagent roles, and an audit log. The control surface that turns guidance into enforcement.

5 entries · TOML/JSONbrowse ↗
C · library
.agents/skills/

Reusable workflows for kickoff, research, decisions, production changes, review, and handoff — composable across repos and teams.

7 skills · Markdownread skills ↗
§ 02Adoption

Greenfield or existing repo — pick your entry.

Bootstrap a fresh repository, or use the merge-first installer to add the agent control layer around an existing README, AGENTS.md and CI — without rewriting what works.

~/work · zsh installverifylogs
1$./scripts/bootstrap-template.sh sample-fullstack "fullstack-product" drop-claude
2 ↳ scaffolds AGENTS.md, .codex/, .agents/skills/
3$./scripts/install-agent-contract.sh "../existing-repo" auto
4 ↳ merges contract; preserves docs & CI
5$git diff --stat
6 3 files changed · 184 insertions(+) · 0 deletions(−)

Preview

./scripts/install-agent-contract.sh ../repo auto preserve-agents --dry-run

Verify

./scripts/verify-template.sh --strict

Report

./scripts/verification-summary.sh .codex/logs/verification-summary.json
§ 03Thesis

Good agent work leaves evidence.

i

Context

What the agent saw

  • Task framing
  • Files inspected
  • Assumptions & unknowns
ii

Control

What it was allowed

  • Narrow implementation scope
  • Policy hooks
  • Bounded subagent roles
iii

Proof

What it produced

  • Focused verification
  • Review or risk notes
  • Handoff artifacts
§ 04Compatibility

Codex is the core; Claude Code is an adapter.

Codex-native core

AGENTS.md, .codex/, hooks, policy, subagents, verification, and release gates define the canonical control plane.

Claude Code adapter

.claude/ and CLAUDE.md mirror the same contract for Claude Code users without becoming required template infrastructure.

Prompt docs alone

Plain prompt files document intent, but Agent Spine adds install, drift, policy, and release checks around that intent.

§ 05Verification

Verification — three scripts, one contract.

verify · ci installverifylogs
1$./scripts/verify-template.sh
2 ✓ AGENTS.md present · 12 sections
3 ✓ .codex/ valid · 0 policy violations
4$./scripts/smoke-test-bootstrap.sh
5 ✓ scaffold dry-run · 3 files staged
6$./scripts/smoke-test-installer.sh
7 ✓ merge-first installer · existing CI preserved
8 done in 1.42s

Three deterministic scripts you can wire into any CI. They check the contract, never your runtime.

3scripts
Verify · smoke · install
0deps
Pure shell, no runtimes
~1.4s
Typical full run