ai-dlc product direction¶

Understand what ai-dlc promises to do, what it leaves to its neighbours, and the user outcomes it is measured against.

Status: feature-frozen as of 2026-10 by ADR 0001, which owns the in-scope list. Changes outside it are closed with a pointer; bug fixes in frozen areas are welcome. The roadmap distinguishes implemented, proposed, cancelled and unverified scope.

Product promise¶

ai-dlc is a local workflow, machine-provisioning and work-lifecycle tool. It prepares a developer machine and authors one optional workflow plugin of guidance, hooks and team skills. It finishes reviewed work only against the merged revision and its CI receipts. It does not evaluate harnesses (flightdeck does) or define what runs in cloud sandboxes (hangar's claim-time renderer does).

Its durable value is the relationship between tools, artifacts, decisions, and workflows. OpenSpec currently provides formal specification; personal projects use GitHub Issues and Projects, while Jira Cloud is the selected work target. The workflow plugin and managed sections target Claude Code and Codex; client rendering is frozen (ADR 0001) and installed clients are unqualified. These choices are replaceable.

The harness performs development and can use installed tools directly. ai-dlc provides setup, guidance, integration, validation, and evidence services where useful. Existing project policies still determine required checks and completion.

Teams should be able to scaffold their own development flow. They choose supported tools and clients, keep repository policy and acceptance commands in the project, and share only the guidance they need. Start with setup, selected guidance and useful local checks; add tracked delivery, formal specifications and other stages when the team's work needs them. Profiles and team sources carry selected practices without making this repository's own process mandatory downstream.

Immediate adoption outcome¶

Install ai-dlc on the work computer and adopt work repositories with their own Jira configuration and local credentials. Claude Code and Codex get the same guidance from the plugin or managed sections. An engine clone must not enroll work projects into ai-dlc's personal tracker. Keep Obsidian private; Confluence publication is cut.

User outcomes¶

1. Enroll another supported machine/environment and recover the same selected tools and guidance with independent local credentials, paths, and accounts.
2. Open a new or existing repository and let the harness discover selected providers, installed tools, authoritative instructions, artifacts, and next action.
3. Shape a vague idea or observed problem into a worthwhile bounded increment, with evidence, assumptions, alternatives, and explicit success criteria.
4. Translate outcomes into behavioral specifications, deliverable tickets, and verification without repeatedly copying the same information.
5. Add or replace components with clear compatibility, ownership, and update rules.
6. Hand the workflow plugin to a team harness that can switch it on or off; flightdeck measures whether it helps.

Responsibilities and replaceability¶

Evaluation of harnesses and guidance is not an ai-dlc responsibility: flightdeck owns it. Required checks and finish gates still establish delivery evidence.

Product development workflow¶

Greenfield starts with audience, problem, evidence, alternatives, and the smallest useful outcome. Brownfield starts with observed behavior, users, integrations, compatibility constraints, and an incremental change. Both converge on a shaped outcome, requirements, applicable design, behavioral specification, delivery slices, implementation, and verification. Findings can send work back to discovery.

Product requirements explain intent and scope. Specifications define observable behavior. Tickets organize deliverable slices, dependencies, and status. Tasks describe implementation steps. Links and stable IDs preserve these relationships.

UI/UX is one optional workflow in this sequence. Its generation/evaluation loop serves changes requiring interaction or visual judgment. Other work uses suitable methods: compatibility tests, architecture review, migration rehearsals, API contracts, or operational evidence. Small work uses small artifacts.

Principles and limits¶

• Keep recurring context small and relevant. Prefer short shared rules and links to task-specific guidance; measure cost and failures caught before adding more process. Required commands should test project behavior; product choices, test adequacy and design quality remain review judgments.
• Prefer existing tools; build integration where it removes coordination work. Tencent teamai repositories are a supported team-source format, not a runtime dependency; ai-dlc reads their safe, pinned subset through profile subscriptions. Prove a small component contract with real adapters before expanding it.
• Make installation, configuration, instructions, and readiness understandable together while preserving their separate data ownership.
• Carry portable state through versioned repositories; keep secrets and machine bindings local. Never infer an account from an unrelated session.
• Preserve authored content, active bindings, and existing contracts on updates.
• Make product choices explicit before coding; an executor must not invent scope, approval, evidence, or a missing cross-ticket interface.
• Measure observed outcomes. Green tests alone do not prove product value, user preference, or successful live platform qualification.
• Consistency means the same intended workflow with honest capability differences, not an assertion that every environment supports every tool.

Delivery boundaries¶

Still to qualify: machine bootstrap on macOS and Linux, the workflow plugin in Claude Code (#53), and the local lifecycle against Jira Cloud (#85). Each stays unverified until its walkthrough succeeds. Cloud execution is hangar's and evaluation flightdeck's; the scope freeze governs the rest.

The last release is v0.4.0 (September 14, 2026); the changelog lists later changes. Release-tracking issue #17 was cancelled as not planned.

Navigation¶

• ADR 0001: the scope freeze and in-scope list
• Roadmap
• Current architecture
• Migration notes
• Actual release evidence

Earlier plans are in git history and openspec/changes/archive/.