v0.22.0

Changed

• blueprint skill — adaptive next-stage routing clarity (skill manifest
  0.6.0 → 0.7.0). A focused follow-up to the §19 router and §9 Product
  Experience Direction, acting on an external review of a generated blueprint.
  No redesign — the changes are columns, a couple of small tables, and a short
  rationale, all tech/domain-neutral and folded into the existing
  prompts/templates/reference/example:
  • §19.2 Depends On column. The Stage Recommendations table now states
    each stage's prerequisite explicitly (e.g. security-review and ux-design
    depend on architecture-design), distinct from the revisit trigger, so the
    pipeline no longer looks more rigid — or more parallel — than intended.
  • §19.4 Recommended Pipeline split. The single flat pipeline list is now a
    Recommended Linear Path (core ordered sequence) plus a Conditional
    Follow-up Gates table (Gate | Run When | Typical Input | Output), so
    deferred conditional stages — chiefly architecture-update /
    architecture-reconciliation — are never silently dropped from the routing
    picture.
  • §19.3 ASK_USER Decision Rationale. When no stage is ASK_USER, §19 now
    justifies why (each high-impact unknown is already answered by the source
    report / §9, deferred to architecture-design, or delegated to
    security-review, with a named owner) instead of looking over-confident;
    FAIL if ASK_USER is absent despite an unresolved high-impact unknown with
    no downstream owner.
  • §19.1 complexity score labelled a routing heuristic. The output now
    states that the / 21 score is a routing heuristic, not a formal project
    estimate, to be revisited after architecture-design — preventing false
    precision.
  • §9.4 / §9.5 interaction-mode Classification. Every interaction mode now
    carries a controlled classification — primary surface / secondary
    surface / wrapper / integration surface / future surface — with an
    explicit rule that "AI Skill" must be disambiguated (usually a
    wrapper/integration surface around the CLI/core, not a separate runtime) and
    never conflated with MCP (a tool surface for external AI agents).
  • Quality gates. The Product Experience Gate gains an "Interaction modes
    classified" row; the Adaptive Stage-Gate Recommendation Gate gains
    "Stage table has Depends On", "Linear-path vs conditional-gates split",
    "ASK_USER absence explained", and "Complexity score labelled heuristic" rows,
    plus matching WARNING conditions. The §19 output-discipline budget was
    updated so the additions (columns + small tables) do not contradict the
    "keep §19 compact" rule.