| name | faffter-noon-architecture |
| description | Default `architecture` slot occupant — the generative architecture/infra PROPOSER. Reads the team's infra profile + the brief/spec and proposes one best-fit, build-biased, production-grade architecture, emitting a founded `faff-contract:architecture-proposal` block + ADR promotion intent. Runs as a configured slot, not the user `/` menu. |
| user-invocable | false |
| judgement_seam | architecture |
faffter-noon-architecture
The default occupant of the architecture slot — the generative half of faff's architecture story. Given a brief or spec plus the team's acquired infra profile, it proposes one best-fit, production-grade architecture + infra, reasoned. It is the missing PROPOSE box: faff already records ADRs (the adr slot), critiques a spec's design (the spec-review architectural lens), and acquires the infra profile (the profile slot) — this generates the proposal those steps then judge.
When standalone, Read the sibling faff/SKILL.md (the gateway) first — it holds the shared rules and the fixed contracts. This recap is non-normative; the gateway wins.
What it does
One LLM pass turns the brief/spec + infra profile into one proposal envelope. It proposes, never commits: it emits the envelope + ADR candidates as intent and stops. It writes nothing under docs/adr/ and makes no faff adr new call — graft Step 4b materialises any candidate. It runs no review or verdict logic; the only contact with the downstream spec-review architectural lens (the critic) is the spec artifact the proposal lands in — the proposer/critic boundary, where the proposer generates and the critic judges, sharing no logic.
The contract (faff contract architecture-proposal) validates the envelope's shape only. This producer owns the proposing strategy — how the design is fitted. Do not re-validate shape here; emit a conformant block and let the consumer pipe it.
Inputs
- The brief or spec to propose an architecture for.
- The team's infra profile, read via
faff profile show --json (the shipped read path). Never call faff profile mine — acquisition is the profile slot's job. On faff profile show exit 3 (no profile), record an explicit "no infra profile" assumption and propose against the brief alone.
How it proposes
- Fit the profile. Derive the architecture from the infra evidence (runtimes, CI, deploy targets, PaaS availability, stated prefs) and the brief — not a generic best-practice template.
- Build-biased. Default
recommendation: "build" (local-first tenet). Use "buy"/"hybrid" only when the best fit is not locally buildable; a buy/hybrid is surfaced for a human, never actioned here (procurement is a separate, out-of-scope concern).
- Born-verifiable. State concrete decisions and
assumptions a human (or a later check) can verify. "Is this production-grade?" is the human gate in v1 — this producer does not self-judge proposal quality.
- ADR candidates. Name the decisions worth an ADR as candidates only (
{title, decision, rationale}); echo each into the ## ADR promotion intent section.
Output (the contract artifact)
Emit exactly one fenced block as the producer's output — the consumer locates it, JSON.parses it, and pipes it to faff contract architecture-proposal (the sole source of contract data):
{ "chosen_architecture": "modular monolith on the existing Node runtime",
"rationale": "single Node runtime + CI already present; no ops capacity for a service mesh",
"adr_candidates": [ { "title": "Start as a modular monolith", "decision": "one deployable, module boundaries by domain", "rationale": "defers distributed-systems cost until scale demands it" } ],
"assumptions": [ "traffic stays single-region for v1" ],
"recommendation": "build" }
Then emit a matching ## ADR promotion intent section (one entry per candidate) and a closing confidence: line (high / medium / low) — an advisory self-rating of the proposal, never a quality verdict.
recommendation is one of build / buy / hybrid (enforced by the contract via violations); empty chosen_architecture/rationale or a malformed candidate is a contract violation. A swapped-in proposer (e.g. a cost-first or cloud-native strategy) conforms by emitting the same block.
