AI-DLC Workflows v2 Has Been Released as a Preview, So I Tried Installing It on Kiro CLI / Claude Code / Codex CLI

Introduction

The v2 branch of AI-DLC (AI-Driven Development Life Cycle) Workflows was released as a preview on 2026-06-18.

v1 was a "rule file approach" implementation published on the AWS Blog in November 2025. It works simply by copying core-workflow.md into each harness's configuration directory, and while it supports a wide range of eight harnesses, routing and stage management depended on the LLM's judgment.

https://aws.amazon.com/blogs/devops/open-sourcing-adaptive-workflows-for-ai-driven-development-life-cycle-ai-dlc/

v2 has been completely overhauled into a native TypeScript implementation. The biggest change is that progression that relied on LLM judgment has been replaced by a deterministic engine. The main differences from v1 are summarized below.

The v2 release history began with v0.1.0 in April 2026, and reached 2.0.0 in about two months. The main turning points are excerpted from the CHANGELOG.

The v2 architecture follows a "One core, many harnesses" design. The core/ directory holds the common stage definitions, knowledge, tools, and sensors, and a build script (scripts/package.ts) generates the dist/ for each harness. Harness-specific differences (adapter hooks, configuration file formats) are separated into harness/.

In this article, I install and verify operation on three CLI-based harnesses (Kiro CLI / Claude Code / Codex CLI).

Test Environment

What I Tested

Common: Getting the Repository

Get the v2 branch with a shallow clone.

git clone --branch v2 --depth 1 https://github.com/awslabs/aidlc-workflows.git

Cloning into 'aidlc-workflows'...
remote: Enumerating objects: 1318, done.
remote: Counting objects: 100% (1318/1318), done.
remote: Compressing objects: 100% (866/866), done.
remote: Total 1318 (delta 371), reused 1146 (delta 332), pack-reused 0 (from 0)
Receiving objects: 100% (1318/1318), 3.10 MiB | 26.67 MiB/s, done.
Resolving deltas: 100% (371/371), done.

The distribution files for each harness are stored in the dist/ directory of the cloned repository. The output may change as the v2 branch is updated.

Prerequisite: bun

The v2 hooks and tools are run with bun. bun must be on the PATH for the user that runs each harness. The README also contains the following note.

which bun works in your terminal, but the harness can't find bun — add ~/.bun/bin to your non-interactive shell's PATH

Installing on Kiro CLI

Kiro CLI ≥ 2.6 is required (README "Pick your harness" table).

Copy the contents of dist/kiro/ into your project directory.

cp -r ../aidlc-workflows/dist/kiro/.kiro .

This is the directory structure after copying.

.kiro/
├── agents/
├── aidlc-common/
├── hooks/
├── knowledge/
├── scopes/
├── sensors/
├── settings/
├── skills/
├── steering/
└── tools/

settings/cli.json sets the aidlc agent as the default at startup, and the effort xhigh when using Claude Opus 4.8.

{
  "chat.defaultAgent": "aidlc",
  "chat.modelDefaults": {
    "claude-opus-4.8": {
      "output_config": {
        "effort": "xhigh"
      }
    }
  }
}

This setting does not pin the model; the actual model selection depends on the plan and settings on the Kiro side. The README explains it as follows.

AI-DLC on Kiro works best with Claude Opus 4.8, which requires a paid Kiro plan. On weaker models the conductor may skip optional stage steps (reviewer pass, learnings ritual) or rush approval gates.

Installing on Claude Code

I use Claude Code via AWS Bedrock.

cp -r ../aidlc-workflows/dist/claude/.claude .

This is the directory structure after copying.

.claude/
├── CLAUDE.md
├── agents/
├── aidlc-common/
├── hooks/
├── knowledge/
├── rules/
├── scopes/
├── sensors/
├── settings.json
├── settings.local.json.example
├── skills/
└── tools/

settings.json contains model settings and environment variables (excerpt).

{
  "env": {
    "CLAUDE_CODE_USE_BEDROCK": "1",
    "AWS_REGION": "us-east-1",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "global.anthropic.claude-opus-4-8"
  },
  "model": "opus[1m]",
  "effortLevel": "xhigh"
}

The distributed settings.json sets environment variables for using Bedrock. The project settings are configured to use Claude Opus 4.8 (1M context). If you use the Anthropic API directly, adjust env by referring to settings.local.json.example.

Installing on Codex CLI

Codex CLI requires a git repository, so if it has not been initialized, run git init first and then copy the files.

Copy three items from dist/codex/.

cp -r ../aidlc-workflows/dist/codex/.codex .
cp -r ../aidlc-workflows/dist/codex/.agents .
cp ../aidlc-workflows/dist/codex/AGENTS.md .

This is the structure after copying.

.codex/
├── agents/
├── aidlc-common/
├── aidlc-rules/
├── config.toml
├── hooks/
├── hooks.json
├── knowledge/
├── rules/
├── scopes/
├── sensors/
├── tools/
└── trust-seed.toml

.agents/
└── skills/

config.toml contains settings for the model and provider.

model = "openai.gpt-5.5"
model_provider = "amazon-bedrock"
model_context_window = 1000000
model_reasoning_effort = "high"

However, a warning appears saying that model_provider and model_providers are ignored in project-level config.

⚠ Ignored unsupported project-local config keys in .codex/config.toml:
  model_provider, model_providers.
  If you want these settings to apply, manually set them in your user-level config.toml.

You need to configure the Bedrock provider in the user-level ~/.codex/config.toml.

Doctor Check (Comparing the 3 Harnesses)

I ran /aidlc --doctor on Kiro CLI / Claude Code, and $aidlc --doctor on Codex CLI. The following is the Kiro CLI output (after init. Before init, the missing aidlc-docs/ is shown as failed).

AI-DLC Health Check
─────────────────────────────────────
✓  bun installed (required for CLI tools and hooks)
✓  aidlc-audit-logger.ts present
✓  aidlc-sync-statusline.ts present
✓  aidlc-validate-state.ts present
✓  aidlc-log-subagent.ts present
✓  aidlc-session-start.ts present
✓  aidlc-session-end.ts present
✓  aidlc-statusline.ts present
✓  aidlc-kiro-adapter.ts present
✓  agents/aidlc.json present (hook + permission wiring)
✓  settings/cli.json present (workspace default-agent activation)
✓  AWS_AIDLC_DEFAULT_SCOPE (unset — no project default)
✓  aidlc-docs/ directory exists
✓  Hooks last fired: (timestamps omitted)
✓  State Version: 7
✓  Orphan worktrees: 0 observed
✓  Stale branches: 0 observed (not a git repo)
✓  Orphan state files: 0 observed
✓  Orphan audit: 0 observed
✓  Practices staleness: never affirmed (informational)
✓  MERGE_DISPATCH: 0 orphan INVOKED (0 bracketed)
✓  Cycle detection: 0 cycles
✓  Orphan stage files: 32 graph entries all have files
✓  Scope validation: 9 scopes valid (29 advisories)
✓  Schema validation: 32/32 stages validated
✓  Graph references: 122 artifacts + edges resolved
✓  Keyword overlap: no conflicts
✓  Rule drift: no team/project rule overlaps org policy
✓  Paired sensor coverage: no sensor-bound rules (0 feedforward-only)
─────────────────────────────────────
29 passed, 0 failed

On all three harnesses, the common checks (bun, stage validation, graph consistency, etc.) all passed. The differences between harnesses are as follows.

• The Claude Code adapter is not shown as a separate item in the doctor log.

Workspace Initialization (Common to All 3 Harnesses)

I ran init on each harness. All three harnesses produced the same scaffold output.

Workspace scaffolded:
  aidlc-docs/knowledge/           (team knowledge — 11 agent dirs + aidlc-shared)
  aidlc-docs/initialization/      (3 stage artifact dirs)
  aidlc-docs/ideation/            (7 stage artifact dirs)
  aidlc-docs/inception/           (7 stage artifact dirs)
  aidlc-docs/construction/        (2 stage artifact dirs)
  aidlc-docs/operation/           (7 stage artifact dirs)
  aidlc-docs/verification/
State initialized: poc scope, 7 stages, Minimal depth
Project type: Greenfield
First post-init stage: intent-capture (IDEATION)

When AWS_AIDLC_DEFAULT_SCOPE is unset, the default scope is poc. After init completed, intent-capture (IDEATION) was indicated as the next stage, and it stopped at generating clarifying questions and selecting the interaction mode.

The utility that is executed is the same (aidlc-utility.ts init).

Notes

bun PATH Configuration

This was the first problem I encountered in this test. If bun is not installed, or is installed but cannot be found on the non-interactive shell's PATH, an error occurs when hooks run. In this test, the following error appeared in a freshly installed environment.

SessionStart:startup hook error
Failed with non-blocking status code: /bin/sh: 行 1: bun: command not found

Solutions:

• If bun is not installed: install it with curl -fsSL https://bun.sh/install | bash
• If it is installed but not on the PATH: add ~/.bun/bin in ~/.bash_profile or ~/.zshenv
• For Claude Code, it can also be specified via env in settings.json

Summary

I installed AI-DLC Workflows v2 on three harnesses (Kiro CLI, Claude Code, and Codex CLI) and verified up to doctor and init. Installation mainly consists of copying files under dist/, and it works as long as prerequisites such as the bun PATH and Codex CLI's user-level configuration are met.

Thanks to the "One core, many harnesses" design, the same workflow foundation is shared even across different harnesses. In the distributed configuration, Kiro CLI includes the effort setting for when Claude Opus 4.8 is used, and Claude Code includes the model specification for Claude Opus 4.8. Codex CLI specifies model = "openai.gpt-5.5", but the provider settings require user-level config. Since stages, agents, and sensors are also richer in functionality, please keep cost in mind when evaluating.

Reference Links

• AI-DLC Workflows 2.0 Specification (whitepaper)
• AI-DLC Method Definition Paper (a design philosophy document linked from the AWS Blog. Defines 10 principles and a 3-phase structure)