neohaskell/docs/archive/2026-07-ai-artifacts/claude-skills/feature-pipeline-preview/02-classify-feature/01-decide

ARCHIVAL KNOWLEDGE — DO NOT USE OR REFERENCE. Superseded by codemap/ + root AGENTS.md. Scheduled for deletion once the pipeline is verified (Phase 6). Manifest: docs/archive/2026-07-ai-artifacts/MANIFEST.md

name: 01-decide description: Judges the feature's complexity tier from the ADR, issue, and diff scope, with explicit rationale. kind: leaf executor: opus model: claude-opus-4-7

Decide complexity tier

Returns the single tier that best describes the feature's blast radius. The output drives every later grounding pass — trivial features skip cryptographic and hot-path checks entirely; security-critical features cannot skip them. A miscalled tier compounds across phases.

Inputs

• feature_name — string, the human-readable feature name from pipeline.py get feature_name.
• issue_text — string, the GitHub issue body if available (may be empty).
• module_path — string, the planned module location from pipeline.py get module_path.
• adr_path — optional, the path to the ADR if phase 3 has already produced one.
• diff_scope — optional list of file paths the feature touches, if known.

The orchestrator gathers what is available. Any missing field is "" or [] — never invent content.

The five tiers

| Tier | The feature… | Grounding intensity | | --- | --- | --- | | trivial | Adds no new trust boundary, no external IO, no new dependency, no new control flow. Examples: a Mock integration used only in tests, a pure rename, a doc fix, a test fixture, an internal-only refactor. | Skips cryptographic, hot-path, and authorisation rubrics entirely. | | simple | Adds a new pure value type or pure function with no IO and no dependency. Example: a Decimal type, a Uuid newtype. | Eyeball-only security/perf checks; rejects every cascade recommendation by default. | | moderate | Adds a new query / command / decider / serialiser / persistence schema, or a new in-process integration. Touches external IO OR adds a dependency, but not both. | Keeps specific findings; demotes cascade recommendations unless directly applicable. | | complex | Adds a new event-store layer, new networked subsystem, schema migration, replication, queue, or contention-sensitive concurrent structure. Touches both external IO AND a new dependency. | Full Top-10 + STRIDE + SSDF; profiling evidence required for any speedup claim. | | security-critical | Produces, compares, signs, verifies, or stores a value whose secrecy gates access. Touches authn / authz, multi-tenant isolation, HMAC / JWT / OAuth / session / API-key handling. | Every methodology section active; no demotions. |

Decision algorithm

State assumptions, define the verifiable goal, then execute:

1. Read every input present. If the ADR exists at adr_path, read it end-to-end and consult its "Decision drivers" and "Public API" sections. If only feature_name + issue_text are present, work from those.
2. Collect six signals. For each, record true / false / unknown:
   • touches_secrets — does the feature produce, compare, sign, verify, or store a secret value (API key, token, HMAC, password, OAuth state, session id, share-link)?
   • touches_auth — does it touch authentication, authorisation, multi-tenant isolation, or canAccess / canView semantics?
   • external_io — does it perform network, file-system, or process IO?
   • adds_dep — does it add a new cabal / npm / nix dependency?
   • is_test_only — is the feature scoped to test code (Mock, fixture, stub, harness)? Mock integrations are test-only.
   • is_pure_change — is the entire diff a pure rename, doc edit, internal refactor, comment fix, or formatting change?
3. Apply the cascade. First rule that fires wins. Do not short-circuit early; finish reading inputs first.
   • Rule 1 — touches_secrets ∨ touches_auth ⇒ security-critical. No exceptions, even for test-only code that handles real secrets.
   • Rule 2 — is_pure_change ∨ is_test_only ⇒ trivial. This rule overrides keyword matches like "integration" or "http" when the feature is a mock or fixture. A "Mock HTTP integration for tests" is trivial.
   • Rule 3 — event-store / migration / networked-subsystem / multi-writer-concurrent-structure language present, OR (external_io ∧ adds_dep) ⇒ complex.
   • Rule 4 — query / command / decider / serialiser / integration vocabulary, OR external_io, OR adds_dep ⇒ moderate.
   • Rule 5 — otherwise ⇒ simple.
4. Sanity-check. Re-read the chosen tier's row in the table. If the description does not fit the feature, name the mismatch and pick again. If two tiers fit equally, pick the lower-intensity one (the grounding pass can always promote findings; cascade demotions are harder to undo).

Output

Emit exactly one JSON document on stdout, no preamble, no trailing commentary:

{
  "tier": "trivial",
  "rationale": "<one to three sentences naming the rule that fired and the decisive signals>",
  "signals": {
    "touches_secrets": false,
    "touches_auth": false,
    "external_io": false,
    "adds_dep": false,
    "is_test_only": true,
    "is_pure_change": false
  }
}

The persist step will read this JSON verbatim. Any deviation (preamble, multiple JSON blocks, missing field) breaks the pipeline.

Refusals

• All inputs are empty or contradictory → refuse: "insufficient evidence to classify"; do not invent a tier.
• Any signal remains unknown at emit time → refuse: "signal <name> cannot be determined from inputs; ask the maintainer". The JSON output schema is boolean-only; unknown is a scratchpad state for the decision algorithm and must never appear in the emitted JSON.
• The chosen tier contradicts the explicit issue body (e.g. issue says "this stores OAuth tokens" but no auth/secret signal was set) → refuse: name the contradiction and stop.