Pinned AI-DLC methodology#

The AI-DLC reference kit imports the methodology of AWS AI-DLC Workflows source at commit 2a555f7a4387cd1b8254029cccb9fd1e99974574. The imported material is licensed MIT-0; its original license is retained in upstream/LICENSE-MIT-0.

The snapshot retains the source's declared version 2.11.0. This commit is not the published v2.11.0 tag, which resolves to 6a378b53c0a4fe0641ed7d8de8dfff94264d5b6a. Comparisons with the released native binary must disclose this methodology-source difference. Historical pinned results remain bound to their original source; they do not qualify the release.

Snapshot contract#

upstream/snapshot.json contains schema_version: "1.0", an upstream record with repository, commit, and version, an ordered stages array, an agents array, and a profiles object. There are 33 stages across five phases:

The 11 profiles are bugfix, classic, enterprise, express, feature, infra, mvp, poc, refactor, security-patch, and workshop. Each profile has an ordered stages list containing exactly the compiled scope grid's EXECUTE entries and defaults containing the authored scope's policy values. name, keywords, and description are routing/display metadata and are excluded from defaults. Values such as on, off, strict, and advisory remain strings; explicit YAML true and false remain booleans. Missing properties remain absent, rather than inventing inherited runtime defaults.

Each stage retains authored execution conditions, agent roles, review settings, artifact production/consumption, stage dependencies, sensors, scope membership, unit repetition, and workspace requirements where present in the compiled graph. It also retains the original authored body in procedure, its repository-relative source_path, and source_digest (SHA-256 of the entire Markdown source including frontmatter). These records let a Kontour compiler generate skills while keeping the source and provenance inspectable.

The 14 authored agent personas are retained separately, sorted by source path. Each has slug, source_path, source_digest, the original raw source_frontmatter, and procedure (the persona body). Raw frontmatter retains folded descriptions and harness-specific tool restrictions without pretending those restrictions enforce anything in a Kontour runtime.

An EXECUTE profile entry means the stage is selected; a stage's own execution: CONDITIONAL and condition still determine applicability. Retained for_each: unit-of-work declares unit-level repetition; listing a stage once does not implement independent child-run scheduling. Dependencies on stages omitted by a profile remain in the source record so compilation must account for them explicitly instead of silently claiming the full upstream DAG is unchanged.

Reproduce the import#

Prepare the pinned upstream checkout and its already generated Claude distribution data separately. Then run from this repository:

node scripts/import-upstream.mjs /absolute/path/to/aidlc-workflows

An optional second argument selects a temporary output directory for comparison. The importer verifies the checkout HEAD, version, committed authored input bytes, and the pinned SHA-256 digests of stage-graph.json and scope-grid.json. It reads those data files and the authored Markdown; it never invokes AWS hooks, engines, packaging code, workflow commands, or commands mentioned in procedure prose. There is no timestamp or absolute checkout path in the output, so identical inputs produce byte-identical JSON and license files. Unsupported scope YAML fails loudly.

What parity means#

This snapshot establishes methodology coverage, not runtime equivalence. Kontour can compare stage selection, policy declarations, procedures, artifact contracts, and review roles against a pinned source without depending on the AWS engine. The reference kit's runtime is implemented with Kontour products; the snapshot alone does not prove its behavior.

Upstream procedure bodies include AWS-specific record paths, directives, agent identifiers, and command names. They are reference material for adapting Kontour skills. Preserving them does not make those commands available or authorize executing them. Generated instructions must bind those operations to available Kontour tools and disclose any unsupported behavior.

Runtime qualification must separately demonstrate stage advancement, conditional skip decisions, human approvals, sensor failures, review iteration limits, artifact freshness and downstream invalidation, per-unit execution and joins, source/review attribution, deployment authority, recovery, and operational feedback. Recording an upstream setting does not implement that setting, and accepting an artifact does not prove its contents were reviewed. Required evidence should fail closed when a check is unavailable or malformed, even where upstream behavior differs.

Treat this fixture as a repeatable testing ground: compare emitted Kontour flow and skill contracts to the snapshot, run representative profiles, inject failures and changed inputs, and record product-owned gaps. Update the upstream pin deliberately with new fixture digests and contract comparisons; never silently import a moving branch or substitute a second lifecycle engine.

The pinned snapshot is an independent source oracle. Output comparisons should check it directly, not only compare two outputs generated by the Kontour compiler: a shared compilation bug could make those outputs agree. A methodology comparison can verify membership, order, roles, artifacts, policies, and source digests; a behavior comparison needs AWS and Kontour run outputs for the same input fixtures and explicit acceptance criteria. Differences should be classified as matching, intentional stronger enforcement, missing capability, or observed regression.