0001 Refocus AI-DLC on workflow, machines and work lifecycle¶

Context¶

AI-DLC grew to about 28,000 lines of Python in src/ai_dlc/, 46 OpenSpec capability specifications and several megabytes of documentation. Its product promise read as a harness definer and evaluator: it rendered guidance for every runtime, ran skill evaluations and described hosted follow-ons.

Three neighbouring tools now own those jobs more directly. Hangar runs coding agents in cloud sandboxes and renders their configuration at claim time. Flightdeck judges harness changes with paired experiments over recorded runs. The governed workspace owns policy decisions, approvals and receipts. Keeping parallel versions of those responsibilities in AI-DLC adds maintenance without adding evidence. The planning record is the 2026-10 harness plan.

Options considered¶

1. Keep growing AI-DLC as the harness compiler and evaluator. Rejected: hangar's claim-time renderer is what actually runs in sandboxes, and evaluation needs run data that only the sandbox fleet produces.
2. Retire AI-DLC. Rejected: developers still need a reproducible machine, a repository-owned workflow and a CI-receipt-gated finish.
3. Freeze AI-DLC to three jobs and move or cut the rest. Accepted.

Decision¶

AI-DLC is a local workflow, machine-provisioning and work-lifecycle tool. It does three jobs:

1. Workflow plugin author. It authors one optional workflow plugin (guidance, hooks and team skills) that a team harness can pin and switch on or off, and it renders managed sections into developer repositories.
2. Developer-machine provisioner. It enrolls, bootstraps and provisions developer machines from versioned profiles.
3. CI-receipt-gated work lifecycle. It binds reviewed work to a tracker item, branch and pull request and finishes it only against the merged revision and its CI receipts.

New features, documentation growth and new OpenSpec capabilities outside the in-scope list below stop. Changes outside the list are closed with a pointer to this decision; bug fixes in frozen areas are welcome.

What moved where¶

In-scope list¶

Active:

• Work lifecycle: work new/start/pr/archive/finish, the finish gates specification-current, pr-merged and ci-green, and CI receipts.
• Machine enrollment, bootstrap and provisioning, including finishing native Windows (#172).
• Ownership-digest managed sections (harness/agents.py, .ai-dlc/agent-ownership.json).
• SCOPES layering (config.py).
• The skills lockfile (agents/skills.lock.json).
• The workflow plugin (#191 to #193).
• Local MCP (ai-dlc mcp serve).
• The documentation gate.
• The local knowledge provider (Obsidian vault).

Frozen, bug fixes only:

• FDE scaffold (ai-dlc fde).
• Design capture and frontend packs.
• Team sources and teamai import.
• Effective-environment reports.
• The documentation subsystem (ai-dlc docs ...).
• GitHub Issues/Projects and Jira Cloud adapters.

Moving out:

• ai-dlc eval and skill evaluation, to flightdeck (#194).
• Per-runtime rendering growth, to hangar (#196). No new clients, client formats or per-runtime translation in AI-DLC's renderer; cloud sandboxes are rendered by hangar at claim time. The devcontainer, codex-cloud and claude-cloud execution targets are removed.

Cut:

• Linear (#199).
• Legacy scaffold (#201).
• Confluence publication (#50).
• Native harness verification (#175).

Consequences¶

• AI-DLC no longer claims to evaluate harnesses or to define what runs in cloud sandboxes. The workflow plugin is one optional component that hangar and flightdeck can carry and compare; AI-DLC does not decide whether it is enabled.
• Issues outside the in-scope list receive the scope:frozen label and a pointer here. Bug reports in frozen areas stay open.
• Code moves and deletions happen in their own issues; this decision does not move code. Moved and cut areas keep working until those issues land.
• A later decision may reopen a frozen area. Until then, this list is the scope boundary for contributors and agents.

Links¶

• Product direction
• Roadmap
• Repository guidance
• Contributing