Skip to content

AI-DLC Workflows 2.0 - Roadmap

Status as of 2026-07-13. Current released version: 2.3.4 (origin/v2 tip 3c768787). Full GA is declared at 2.4.0.

North star reference

The seven functional goals of the AI-DLC Workflows 2.0 North Star, verbatim in intent:

  1. Mimic what we practice in the real world - a stage executed by a configurable ensemble (Owner, Collaborator, Verifier) with consistent semantics across harnesses.
  2. Customization of behaviour - encode new behaviours/policies/constraints in no more than two targeted changes, reusable across harnesses without tool-specific rewrites.
  3. Adaptiveness of workflows - scale in (report triage to compact Fix, Test, PR) and scale out (decide next stages at boundaries); composition not hard-wired.
  4. Verifier as a true adversary - adversarial quality gate; may use a different LLM than the producer; validates against machine-checkable evidence; budgeted self-heal loop escalating to HITL.
  5. Support for cyclic, directional flows - forward progression plus governed, directional feedback loops.
  6. Preserve artefact traceability - downstream stages enrich upstream artefacts rather than spawning disconnected ones.
  7. Organizational, not project-local, artefact repository - shared org knowledge layer across projects, intents, and repos; six named scenarios.

Goal scorecard

# Goal Status Delivered by Remaining lands in
1 Real-world ensemble Partial 2.0.x roles + 2.1.8 reviewer wiring 2.5.0
2 Customization Shipped 2.0.x rules stack + 2.3.0 plugins plugin deferred surfaces, stage rules layer
3 Adaptiveness Shipped 2.2.0 composer + scale-in (boundary auto-assessment deliberately human-initiated)
4 Verifier as adversary Partial 2.0.x reviewer loop, 2.3.1 model split 2.4.0
5 Cyclic flows Partial within-stage review loop, HITL hardening 2.6.0
6 Traceability Partial artefact graph + upstream-coverage sensor 2.7.0 (+ community PRs below)
7 Org repository Shipped 2.1.0 spaces/intents/org-KB done

Delivered

Version Feature Goal Key PRs
2.0.0 - 2.0.2 GA Preview: reviewer mechanism, multi-harness core, agent roster 1, 4 (v2 baseline)
2.1.0 Per-intent workspace: spaces, intents, multi-repo, org-KB 7 #429
2.1.2 Per-unit for_each iteration 3 #444
2.1.3 - 2.1.7 Loop-integrity hardening: artifact gate, stop-hook, human-presence, dangling consumes 4, 5 #443, #466, #482
2.1.8 Reviewer wired across all harnesses 1, 4 #405
2.2.0 Adaptive workflows: composer, scale-in triage, in-flight recompose 3 #477
2.2.1 - 2.2.19 Hardening line: unit kinds, per-unit reviewer scope, Kiro IDE hooks, optional produces, path quoting - #491, #509 - #512, #520 - #522, #525, #537, #538, #471, #544
2.3.0 Extensions: plugin mechanism (plugins//, per-harness emission, compose seam) 2 #475
2.3.1 Tunable agent tiers (judgment / balanced / templated) 4 #546
2.3.2 - 2.3.3 Doctor surfaces hook drops; scope-change refused under autonomy - #541, #542
2.3.4 Deterministic PreToolUse enforcement of reviewer read scope 4 #545

In flight (ours, open)

Version claim Work PR State
2.3.5 Plugin content buckets + install-time plugin selection #550 Open, tiers green, merge gated
2.3.5 - 2.3.6 Single-command CLI dispatcher (dark launch) #560 Draft, stacked on #550; rebase + retarget after #550 merges
2.3.7 Phase Progress rows advance at phase boundaries #562 Open
2.3.8 upstream-coverage matches real citation forms #563 Open
2.3.17 opencode harness (5th harness) #571 Open, rebased onto v2, merge gated

Version slots are multi-claimed; whoever merges second rebases and re-bumps (t68 catches misses).

Outstanding (planned minors)

2.4.0 - Reviewer-as-verifier (declares Full GA)

Interpretation locked 2026-07-13: prompt-only change. Of Goal 4's four sentences, "different LLM" is delivered (2.3.1 tier split) and the budget is delivered (reviewer_max_iterations + HITL escalation, reading "iterations / tokens" as or). What remains:

  • Adversarial framing: the shared refute-not-confirm + evidence-grounding contract written once into stage-protocol section 12a, plus a short domain-voiced line in each of the two reviewer personas (aidlc-product-lead-agent, aidlc-architecture-reviewer-agent).
  • Evidence grounding folds sentence 3 in: the reviewer runs the machine checks that already exist (listed validation tools, tests, lint, typecheck, acceptance criteria, consumed contracts); opinion-only findings are suggestions, not NOT-READY.
  • Explicitly NOT in scope: token budget, bespoke domain validator tools, new stage-schema fields. Record these scope decisions in the changelog entry.
  • No new agent: the two existing reviewers cover all 11 reviewer-declared stages; the per-stage reviewer: field already supports adding more later.

2.5.0 - Three-role ensemble (Goal 1)

  • Collaborators become independent subagents rather than inline voices adopted by the conductor.
  • Collaboration pattern (pipeline / swarm / review-loop / mob) becomes a per-stage selectable knob.

2.6.0 - Governed cyclic flows (Goal 5)

  • Cross-stage backward edges: a downstream stage exposing an upstream gap triggers a governed, engine-managed return. Today the graph compile rejects all cycles and requires_stage is forward-only.

2.7.0 - Progressive enrichment (Goal 6)

  • Downstream stages enrich the same upstream artefact in place instead of emitting linked-but-separate files.
  • ADRs as a core design artefact.

Community PRs advancing goals (open, need shepherding)

PR Work Goal
#401 Per-stage traceability enforcement, deterministic JSON sensor 6
#402 / #403 / #404 Design-vs-code hygiene, unit-test dedupe, observability NFR consistency 6
#432 V2 extension mechanism (likely superseded by 2.3.0 plugins; close or rebase onto the seam) 2
#526 Product discovery in Ideation 3
#535 In-place upgrade subcommand -
#552 / #553 Cold-start scope routing; Codex structured selections as human turns -

Known gaps not on the minor ladder

  • Rules enforcement (#495): rules_in_context paths are emitted but nothing forces the conductor to read them; quietly undermines Goal 2 at runtime.
  • Stage-level rules layer (aidlc-stage-<slug>.md): reserved, unbuilt (Goal 2).
  • Plugin deferred surfaces: agents/scopes/memory/knowledge projection, when: evaluation, marketplace (Goal 2).
  • Sensor blocking severity (#431): sensors are advisory-only; a gate cannot be halted by a machine check (adjacent to Goal 4, deliberately out of 2.4.0).
  • Conductor health cluster (#547, #548, #551) and Kiro IDE 1.0 regressions (#543, #555): pre-GA quality, not goal-mapped.
  • Cross-unit discovery propagation (Goal 6): still a real v2 gap, but the only open PR (#300) targets the retired v1 tree (aidlc-rules/, base main, 2026-05); it needs a fresh v2 implementation, not a rebase.