agentsclimarketplace

Architecture aware development

Skill CauchyIO/awow/.agents/skills/architecture-aware-development

Rivets a KB-alignment action onto each lifecycle moment — reuse prior art, align plans with ratified ADRs, re-check governed tasks. Config-gated: inert unless {PROJECT}/context/tooling/architecture.md declares an architecture plane. Read when forming a plan there.From its SKILL.md

Install
npx -y skills add CauchyIO/awow --skill architecture-aware-development

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.8 KB, as published. Nobody here has run it

Architecture-aware development

awow keeps the agent honest against the board plane (board-aware-development) — what to build. This is the second plane: how we have already decided to build — ratified ADRs, solution-design records, durable pattern notes. In repos that run a retrieval agent over that knowledge, it is queryable in one call. Bind it to the lifecycle the same way: each design/plan moment rivets onto a KB-alignment action.

This is opt-in and config-gated. It does nothing unless the repo declares an architecture plane.

Plane check first (graceful no-op)

Read {PROJECT}/context/tooling/architecture.md. It names the plane:

adr_dir:      <path to ratified decisions, may be empty>
pattern_dir:  <path to durable pattern notes, may be empty>
kb_agent:     <retrieval tool handle, e.g. an ask_brain-style MCP — or "none">
strictness:   stop-and-surface | warn-only
  • No pointer → do nothing, silently. (The hook will not even nudge.)
  • kb_agent: none → fall back to reading the named adr_dir / pattern_dir directly; nothing on disk → no-op.
  • Never invent an ADR or pattern that is not there.

The crosswalk

Engine skill (the moment)Architecture action
brainstormingPrior-art probe, folded into the existing "go to the board" beat — not a second ritual. Ask the KB agent: does an ADR / design / pattern already cover this? If yes, reuse or extend it rather than designing from scratch. Keep it light; the start edge stays whisper-thin.
writing-plansAlign + flag. Query the governing ADRs/patterns with a query built from the plan's domain nouns (specificity is the biggest lever on recall), then flag the specific plan tasks that touch governed surfaces.
executing-plans / subagent-driven-development / test-driven-developmentRe-check the flagged tasks only — carried by the plan, not a blind per-task query.

Completion-edge skills (verification-before-completion, requesting-code-review, finishing-a-development-branch) stay board-only — architecture alignment is a design/build concern.

On what you find

  • Conflict with a ratified ADR or established pattern → do not proceed silently. Surface it concretely — the ADR id + file path + the specific contradiction — and seek human reconciliation. This is a disposition, not a technically enforced gate; strictness: warn-only downgrades it to a warning the human can wave through.
  • Alignment → cite "checked against: [the ADRs/patterns you actually retrieved]". Never write "no conflicts found" — that falsely implies completeness. The citation claims scope of check, not a global guarantee. A retrieval miss is possible: this reduces architecture-blindness, it does not eliminate it.

Query tooling

Use the cheap retrieval call (e.g. ask_brain) for the prior-art probe and the per-part re-checks. Escalate to the deeper, server-synthesised call (e.g. ask_brain_ultra) for the single plan-time alignment, or whenever the cheap evidence is thin/ambiguous on a potential conflict — do not stop, or wave through, on weak evidence.

When NOT to apply this

  • No {PROJECT}/context/tooling/architecture.md pointer — there is no plane to consult; stay silent.
  • No inner-loop engine installed — there is no lifecycle seam to bind.
  • The work is not an initiative — skip it, like the board check.
  • You are inside /process-workitem — that command owns the seam; do not double-apply.

Keep looking

Skills are one crate of 608,666. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.