adlc-engineering

ADLC

Run repeatable Build, Fix, and Review loops with your coding agent, with every completion claim tied to evidence.

ADLC gives an existing coding agent one public skill and a deterministic control plane for state, permissions, verification, approvals, recovery, and evidence. It does not replace the agent or silently merge, deploy, publish, or spend money.

Install in 30 seconds

Install the scoped public beta from the adlc-engineering distribution, then use the transactional lifecycle to install and diagnose one canonical skill:

python3 -m pip install "adlc-engineering==0.9.2"
adlc-skill install --provider codex --target /path/to/your-repo
adlc-skill doctor --provider codex --target /path/to/your-repo

Use claude instead of codex for Claude Code. The package lifecycle (install, doctor, update, rollback, and uninstall) is the transactional path for those two providers. Cursor, Antigravity, and Factory are experimental compatibility targets without rollback claims. See installation for exact ownership and migration behavior.

Run a five-minute Fix

In the target repository, ask the installed coding agent:

/adlc fix the failing average calculation. Reproduce it first, make the smallest repair, run the affected tests, and stop with PR-ready evidence.

The expected sequence is red verifier → bounded repair → green verifier → independent review → pr_ready. To replay the repository’s deterministic version of that first success in an isolated temporary repository:

bash tests/acceptance/run_readme_quickstart.sh

That replay routes Fix through the installed public facade, then proves the deterministic kernel path. It does not invoke Codex or make a live-provider claim. The First Fix guide explains every stop state and artifact.

Three evidence-bound loops

Start with the Build, Fix, or Review guide. Resume interrupted work without replaying completed effects with the Resume guide.

Current provider evidence

This block is checked against support-matrix.json; edit the evidence, then run python3 scripts/render_support_matrix.py, rather than hand-editing claims. Labels apply only to the named provider, version, harness, model, loop, fixture, commit, and dimensions. See the generated support matrix for raw evidence links and limitations.

How it works

your intent
    │
    ▼
one `adlc` skill ──► Build / Fix / Review loop contract
    │                              │
    ▼                              ▼
coding-agent judgment       deterministic ADLC kernel
                             state · admission · tests
                             approvals · recovery · evidence
                                      │
                                      ▼
                         proved / blocked / awaiting human

The skill routes only the context needed for the selected command. The kernel owns deterministic truth and fails closed when approval, credentials, evidence, or compatibility proof is missing. Read skills, loops, and kernel for the boundary.

Safety and human approval

• Review, status, and doctor are read-only.
• Local mutation starts only inside an authorized Build or Fix boundary.
• External writes, publish, merge, release, deploy, destructive recovery, privileged access, and paid execution require explicit approval.
• Persisted effects carry idempotency evidence so resume does not intentionally replay completed work.
• Telemetry is off by default. Provider credentials and target-repository content stay under the operator’s provider and repository controls.
• pr_ready does not mean merged, deployed, adopted, secure, or generally available.

Read the security and privacy boundary, security policy, and compatibility/deprecation policy.

Proof, not promises

• Provider conformance evidence
• Deterministic public Fix replay
• Replayable invoice Fix demo and public benchmark
• Legacy-surface migration ledger
• Canonical local verification

The public benchmark preserves three live Codex runs plus an independent three-run replay, with exact red-before-green verification, same-session interrupt/resume, one-file scope, distinct read-only review, and completion-audit evidence. Its versioned report discloses provider/model versions, tokens, duration, marginal account cost, failures, and limitations. These proofs are scoped evidence, not proof of future behavior, market traction, compliance certification, GA readiness, or universal provider support.

Public beta candidate

The 0.9.2 beta launch packet accompanies the immutable release candidate. It includes the technical narrative, five-minute demo and recording plan, local-first metric definitions, and beta operating model. Every launch claim resolves to repository evidence, telemetry remains off, and package, GitHub Release, Pages, and launch communication actions remain human-approval-bound and must be verified at their public endpoints.

Use the beta feedback issue template for sanitized product outcomes and friction. Do not put vulnerabilities, credentials, prompts, source code, private paths, or target-repository content in a public issue; use the private Security Advisory path instead.

The beta learning funnel is intentionally local and ordered:

1. README visit. Confirm that the evidence boundary and current provider row match the intended use.
2. Install. Complete the pinned package install shown above.
3. Doctor. Run adlc-skill doctor and proceed only when required checks pass.
4. First loop. Run the five-minute Fix path, or an eligible Build, and retain its local evidence-backed terminal report.
5. Returning project. Within seven days, start another eligible Build, Fix, or Review in the same project; this is a local learning signal, not a retention or traction claim.

Documentation and community

• Documentation home
• Public command reference
• Configuration
• Artifacts and schemas
• Stop reasons
• Contributing and governance
• Code of Conduct
• Changelog
• Migration guide and historical archive

Status and limitations

ADLC is an unreleased beta candidate. The source checkout, deterministic tests, exact Codex Fix configuration shown above, documentation-site build, scoped live public benchmark, secure release preparation, and final-form launch packet have evidence; package publication, production docs deployment, and public communication remain human-gated. Only released and tested configurations may graduate support labels.

doc_honesty_section: This page describes the current source product and links to its evidence; it is not a release artifact or adoption proof.

no_overclaim: ADLC does not claim GA, autonomous delivery, universal provider behavior, benchmark superiority, compliance certification, or production support.

limitations: Current live-provider evidence covers only the exact Codex Fix row above. Claude Code invocation is credential-blocked in the recorded run, and compatibility targets are not live-provider claims.

License

MIT. See LICENSE.