cadora 0.2.0

An AI-DLC workflow conductor — drive the AWS AI-DLC method on a headless coding agent (Claude Code), with deterministic gates and run capture.

Cadora

An AI-DLC workflow conductor — drive the AWS AI-DLC method on a coding agent, headlessly, and capture every run.

Cadora implements the AWS AI-DLC method (AI-Driven Development Life Cycle) as a thin, backend-agnostic conductor. Point it at a product vision and it drives a coding agent through the full AI-DLC lifecycle — 🔵 Inception (requirements → planning → design) then 🟢 Construction (code generation → build & test) — autonomously and headlessly, applies a deterministic gate, and captures the whole run (artifacts, events, cost) for inspection and comparison.

Cadora is the conductor, not the agent: it doesn't implement the agent loop (the backend CLI does), and it isn't an in-session/in-IDE assistant. It installs the AI-DLC rules into a workspace, drives an external headless coding-agent CLI, gates the result, and archives it — so runs are reproducible, scriptable, gate-able, and comparable.

Backends: Cadora supports both Claude Code (claude -p, subscription-funded by default) and OpenAI Codex (codex exec --json, using your Codex login). Select either with --executor claude or --executor codex; both use the same AI-DLC topology, deterministic gates, toolchain-integrity evaluation, and run archive, so their results can be A/B-compared directly.

Both backends ship in v0.2.0 — Claude Code since v0.1.0, and OpenAI Codex promoted to a supported backend after live verification.

Status

v0.2.0 — released (PyPI). Adds the pluggable HITL review surface over MCP (cadora mcp — Claude Code/Desktop, Codex, remote HTTP), the OpenAI Codex backend, gate-prerequisite provisioning, and toolchain integrity on top of the v0.1.0 conductor. Live-proven end-to-end — from a tiny CLI to a multi-stack AWS app (CDK + Lambda + Next.js, cdk synth + tests passing). pytest green, ruff clean, CI on Python 3.10–3.12.

How it works

vision.md ─▶ install AI-DLC rules (agent memory + .aidlc-rule-details/) into a workspace
          ─▶ drive the selected backend through the AI-DLC lifecycle
          ─▶ deterministic build-test gate (blocks on failure)
          ─▶ run archive: aidlc-docs/ + events + cost   →   cadora archive show

Install

Requires Python 3.10+ and at least one authenticated backend CLI:

• claude for Claude Code.
• codex for OpenAI Codex.

pip install cadora

From source (for development):

git clone https://github.com/yeychenne/cadora.git && cd cadora
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Quickstart

# 1. Set up an AI-DLC workspace from your product vision.
cadora aidlc-init ./my-project --vision vision.md

# 2. Drive the full AI-DLC lifecycle on Claude Code (autonomous, subscription-funded).
cadora run examples/aidlc.topology.yaml --vision vision.md --cwd ./my-project

# 3. Inspect what happened.
cadora archive ls
cadora archive show <run-id>

Codex development instance:

cadora run examples/aidlc.topology.yaml \
  --executor codex \
  --model gpt-5.4 \
  --integrity-mode repair \
  --vision vision.md \
  --cwd ./my-project

Cadora installs the workflow into the backend-native project memory file: CLAUDE.md for Claude Code or AGENTS.md for Codex. Existing project instructions are preserved outside a managed block. Toolchain integrity defaults to non-blocking audit. Use --integrity-mode enforce to block local packages/scripts that impersonate declared tools, or repair to allow one fresh, constrained agent session to remove the substitution and rerun the external gate.

The deterministic gate also distinguishes a real test/build failure from an unavailable prerequisite. By default, Python workspaces that declare requirements-dev.txt, dev-requirements.txt, or requirements/dev.txt get a cached isolated environment under .cadora/gate-venv; the original --gate-cmd is then executed unchanged inside it. If dependency provisioning is unavailable, the archive records blocked_prerequisite and the missing packages instead of misreporting the application as failing. For disconnected environments, provide a local wheel cache with --gate-wheelhouse /path/to/wheels; use --gate-setup off to manage the gate environment yourself.

Human review is explicit and fail-closed. In a multi-stage topology, mark selected nodes with review: true and run with --hitl. At each declared point the operator must approve, request changes, or abort. Requested changes rerun the same stage—with a maximum of three revision cycles—before any downstream node starts. Closed/non-interactive stdin aborts rather than silently approving. Decisions, comments, individual attempt outputs, and aggregate revision cost are stored in the archive. See examples/aidlc-hitl.topology.yaml. The review surface is pluggable: besides the terminal, cadora mcp serves the review gate over the Model Context Protocol to any MCP client — Claude Code, Claude Desktop, or the Codex CLI (local stdio), or a networked client over streamable HTTP. See docs/hitl-mcp.md.

Scan any existing workspace without running an agent:

cadora integrity ./my-project
cadora integrity ./my-project --json

The run lands in runs/<run-id>/ — manifest.json (ok, cost, model, structured gate status, funding) + a per-node aidlc-docs/ snapshot + the event stream. The generated application code lands in your workspace.

Funding (Claude Code)

Cadora defaults to your Claude Code subscription — it removes any ambient ANTHROPIC_API_KEY from the run so a stray key can't silently meter you. Metered API is explicit opt-in (--funding api). Set the subscription token once with claude setup-token, or just be logged in to Claude Code.

The AI-DLC method

Cadora vendors the AWS AI-DLC rule-set (awslabs/aidlc-workflows, MIT-0) under cadora/aidlc_rules/ and installs it per run as backend-native project memory + .aidlc-rule-details/ (per-stage rules). v0.1.0 drives the lifecycle as a single autonomous session (examples/aidlc.topology.yaml); the per-stage DAG (examples/aidlc-stages.topology.yaml, one node per stage) is the roadmap.

Architecture

• cadora/topology.py — the workflow DAG (schema + loader + dependency-ordered waves).
• cadora/executors/ — the NodeExecutor seam + structured Claude Code and Codex CLI backends.
• cadora/mcp/ — the MCP interface seam: serves the HITL review gate + run control to any MCP client (Claude Code/Desktop, Codex, or remote HTTP).
• cadora/workspace.py — install the AI-DLC rules + inputs into a run workspace.
• cadora/gates.py — deterministic-first gates (shell checks that block).
• cadora/archive.py — run capture (runs/<id>/manifest.json + artifacts) + cadora archive ls/show.
• cadora/runner.py — wires it together.

See docs/ARCHITECTURE.md.

Roadmap

cadora eval / compare (deterministic AI-DLC compliance + LLM-judge over captured runs), per-stage wave concurrency, hardening for Kiro/Antigravity, and a consulting deliverable pack.

License

MIT — see LICENSE. The vendored AI-DLC rules are MIT-0 (cadora/aidlc_rules/LICENSE).