Skip to content

ai-dlc roadmap

Know what ai-dlc does today, which three increments come next, and what is not planned.

The ai-dlc Project and GitHub issues own ticket priority and status. Product direction owns the durable product promise, and release verification owns what a release must prove.

Current outcome

  • Frozen to three jobs. ai-dlc is feature-frozen (ADR 0001, epic #189). It authors the workflow plugin, provisions machines from pinned profiles and finishes work only against the merged revision's CI receipts.
  • The plugin ships. plugins/ai-dlc-workflow/ is published into a team harness for a with/without comparison (#191–#193).
  • Works today. ai-dlc runs from a cloned installation on the work computer and gives Claude Code, Codex and Antigravity one consistent development workflow. Delivery runs through reviewed work records, OpenSpec changes, pull requests and a gated ai-dlc work finish; release publication from a tag is implemented.
  • Default providers. Personal projects use GitHub Issues with repository-associated Projects; work repositories use Jira Cloud; Obsidian is the private journal. hangar intake starts agent jobs from tickets, and ai-dlc keeps its tracker adapters for the local lifecycle (see the tracker boundary).
  • Unsupported. Plane is shipped but unsupported and unqualified; the outstanding qualification list is in release verification.
  • Moved or removed. Evaluation moved to flightdeck, and hangar renders and runs cloud sandboxes. The Linear adapter, ai-dlc eval and ai-dlc scaffold are removed.

Installing the engine from this repository is distinct from adopting a work repository, which has its own tracker and independent local credentials.

Three September team-flow improvements are closed: #167 (guidance for selected project features), #168 (concise team skill discovery) and #169 (focused checks and meaningful starter verification). Their effect on time, tokens and failures caught is unmeasured; flightdeck owns the with/without comparison.

Next three increments

The September delivery-path, diagnostics and release increments are complete (#71–#78, #103). The next increments test whether the workflow helps on real work before adding surface area.

  1. Adoption on real repositories. Adopt one work repository with Jira Cloud and one personal repository with GitHub Issues, deliver ordinary work through them for two weeks, and keep a friction log of every refusal, repeated step and workaround. The work-repository run supplies the live lifecycle record that #85 requires for the local lifecycle steps hangar does not perform (work publish, a blocked work finish before merge and a closing work finish after merge); on success, present one qualified work tracker. Plane already reads as unsupported. Friction-log entries become issues; nothing else enters the queue until they are triaged.
  2. A shorter delivery path. Count the manual commands between a tracker item and work finish on the adopted repositories, then remove or fold steps until the count is halved. The delivery path baseline records the starting counts, candidate reductions and the delivered per-work, content-bound documentation evidence (#146). Re-record only new or stale targets; unchanged decisions survive unrelated edits and target-branch movement. Use observed adoption friction to prioritize the remaining reductions, including the baseline's proposal for one pre-merge command.
  3. Evaluation moved to flightdeck. The runner, grader, report and skill evaluation now live in flightdeck under ADR 0001; ai-dlc's copy was removed in #195 and its last version is tag eval-handoff-2026-10. The paired comparison planned in #138–#140 happens there.

Waiting on environments or decisions, not scheduled: #53 native client, company and cross-platform qualification. The Confluence publication half of #50 is cut under ADR 0001; its team-document-publication change (no tasks started) is archived as cancelled, as is native-harness-verification (#175). The FDE scaffold half of

50 is delivered.

How status is tracked

  • The GitHub Project above holds priority and status; issue closure reasons distinguish completed from not-planned scope.
  • ai-dlc context --brief (MCP work_context) lists open work records, the required checks and the next command; ai-dlc work status <id> reads one record.
  • Work completes only through ai-dlc work finish, which requires the archived specification, the merged pull request and exact merged-revision CI receipts. A Done board value, a local check or a direct adapter call is not completion.

Not planned

Cancelled or deferred scope stays cancelled; nothing here reopens it.

  • Qualification of clean machines, hosted clients, live providers and human calibration that was closed as not planned (#14, #15, #16, #17, #20, #21).
  • Selective Confluence publication (#22, #50), cut under ADR 0001.
  • Package index publication.

The September 2026 roadmap ledger preserves the delivered-and-cancelled table and the reconciliation narrative; the historical executor handoff preserves the earlier execution checkpoints.