feat(P04): variants API + TS types (Wave 3)

Task 4-3-01: POST /v1/variants (template_id or competency_id resolution; cache-first D-029; 404 unknown template/competency; 422 neither), GET /v1/variants/{task_id}, GET /v1/variants?learner_id=. SQLiteVariantStore + VariantGenerator wired into lifespan. Task 4-3-02: packages/types/variants.ts (TaskVariant, VariantParams) + grading.ts (RubricScore + GradeRecord, criteria-typed) — field-for-field Python parity, scores typed as the documented rubric-or-gate union. 11 endpoint tests green (distinct learners -> distinct statements at API level; cache hit -> zero LLM calls); suite 324 green; typecheck 7/7; ruff clean. ---ci--- phase: 4 milestone: v0.3 status: execute requirements: {covered: [REQ-3-005], partial: []} ---/ci---
2026-09-12 03:21:11 +00:00
parent 82b9de382a
commit 4acffac71e
8 changed files with 639 additions and 1 deletions
@@ -10,6 +10,7 @@ from .mentor import router as mentor_router
 from .proctor import router as proctor_router
 from .sandboxes import router as sandboxes_router
 from .telemetry import router as telemetry_router
+from .variants import router as variants_router

 __all__ = [
    "assessment_router",
@@ -19,4 +20,5 @@ __all__ = [
    "proctor_router",
    "sandboxes_router",
    "telemetry_router",
+    "variants_router",
 ]
@@ -12,6 +12,8 @@ from ..sandbox.manager import SandboxManager
 from ..sandbox.workdir import SandboxDir
 from ..telemetry.ingest import TraceIntegrityMap
 from ..telemetry.store import TraceStore
+from ..variants.generator import VariantGenerator
+from ..variants.store import VariantStore


 def get_settings(request: Request) -> Settings:
@@ -53,3 +55,11 @@ def get_grade_store(request: Request) -> GradeStore:

 def get_grading_engine(request: Request) -> GradingEngine:
    return request.app.state.grading_engine
+
+
+def get_variant_generator(request: Request) -> VariantGenerator:
+    return request.app.state.variant_generator
+
+
+def get_variant_store(request: Request) -> VariantStore:
+    return request.app.state.variant_store
@@ -0,0 +1,187 @@
+"""/v1/variants — seeded per-learner task variant endpoints (REQ-3-005, D-029).
+
+Three faces over the variant engine (the generator + store stay FastAPI-free;
+this module owns all HTTP wiring — the D-027/D-032 house pattern):
+
+    POST /v1/variants {learner_id, template_id | competency_id}
+        → 200 the learner's variant — GENERATED on the first request,
+          CACHED (store read, zero LLM calls) on every repeat: D-029
+          reproducibility means one (learner_id, template_id) is ONE
+          variant forever, so a regenerate is always a 200 of the SAME
+          variant, never a second render.
+        → 404 unknown template_id, or competency_id with no bound template.
+        → 422 neither template_id nor competency_id given.
+    GET  /v1/variants/{task_id}
+        → 200 the stored variant owning the task key (the grading and
+          telemetry join path); 404 when no variant was ever generated
+          for the task.
+    GET  /v1/variants?learner_id=...
+        → 200 the learner's variants, chronological; [] when none.
+
+Template resolution: an explicit `template_id` wins; without it the FIRST
+template bound to `competency_id` is used (`template_for_competency`,
+D-021 corpus alignment). The response carries `competency_id` resolved
+from the template library at read time — an enrichment, not a persisted
+column (the seed re-derives the whole variant, D-029) — so the learner
+surface can bind a variant to its competency without a library round-trip.
+
+Distinctness (REQ-3-005): different learners on the same template draw
+different seeded params and receive distinct statements and task_ids;
+tests/api/test_variants.py asserts this end-to-end through the API.
+"""
+
+from datetime import datetime
+from typing import Self
+
+from fastapi import APIRouter, Depends, HTTPException
+from pydantic import BaseModel, Field, model_validator
+
+from ..variants.generator import VariantGenerator
+from ..variants.store import VariantRecord, VariantStore
+from ..variants.templates import TaskTemplate, get_template, template_for_competency
+from .deps import get_variant_generator, get_variant_store
+
+router = APIRouter(prefix="/v1/variants", tags=["variants"])
+
+
+# -- contracts ------------------------------------------------------------------
+
+
+class VariantGenerateRequest(BaseModel):
+    """One variant identity: an explicit `template_id`, or the first
+    template bound to a `competency_id` (D-021). `template_id` wins when
+    both are given (explicit identity beats derived); at least one is
+    required — 422 otherwise.
+    """
+
+    learner_id: str = Field(min_length=1)
+    template_id: str | None = None
+    competency_id: str | None = None
+
+    @model_validator(mode="after")
+    def _require_template_or_competency(self) -> Self:
+        if self.template_id is None and self.competency_id is None:
+            raise ValueError("template_id or competency_id is required")
+        return self
+
+
+class VariantResponse(BaseModel):
+    """VariantRecord over HTTP, plus the `competency_id` enrichment.
+
+    Every field except `competency_id` mirrors `VariantRecord` exactly
+    (snake_case; `created_at` is an ISO 8601 UTC datetime) — the wire shape
+    typed as `TaskVariant` in packages/types/variants.ts.
+    """
+
+    learner_id: str
+    task_id: str
+    template_id: str
+    competency_id: str
+    seed: str
+    params: dict[str, str | int]
+    statement: str
+    starter_files: dict[str, str]
+    created_at: datetime
+
+
+class VariantListResponse(BaseModel):
+    variants: list[VariantResponse]
+
+
+# -- resolution + rendering -----------------------------------------------------
+
+
+def _resolve_template(body: VariantGenerateRequest) -> TaskTemplate:
+    """Template for the request: the explicit id, else the first template
+    bound to the competency; 404 when neither resolves."""
+    if body.template_id is not None:
+        template = get_template(body.template_id)
+        if template is None:
+            raise HTTPException(
+                status_code=404,
+                detail=f"no task template with id {body.template_id!r}",
+            )
+        return template
+    # The request validator guarantees the disjunction, so reaching here
+    # means a competency_id was given (never None).
+    assert body.competency_id is not None
+    templates = template_for_competency(body.competency_id)
+    if not templates:
+        raise HTTPException(
+            status_code=404,
+            detail=f"no task template for competency {body.competency_id!r}",
+        )
+    return templates[0]
+
+
+def _competency_for(template_id: str) -> str:
+    """competency_id enrichment for stored records (read paths)."""
+    template = get_template(template_id)
+    if template is None:
+        # Integrity guard: a stored variant referencing a template that is
+        # no longer in the library cannot be enriched; fail loudly rather
+        # than fabricate a competency binding.
+        raise HTTPException(
+            status_code=500,
+            detail=f"stored variant references unknown template {template_id!r}",
+        )
+    return template.competency_id
+
+
+def _to_response(record: VariantRecord, competency_id: str) -> VariantResponse:
+    return VariantResponse(
+        learner_id=record.learner_id,
+        task_id=record.task_id,
+        template_id=record.template_id,
+        competency_id=competency_id,
+        seed=record.seed,
+        params=dict(record.params),
+        statement=record.statement,
+        starter_files=dict(record.starter_files),
+        created_at=record.created_at,
+    )
+
+
+# -- endpoints ------------------------------------------------------------------
+
+
+@router.post("", response_model=VariantResponse)
+async def generate_variant(
+    body: VariantGenerateRequest,
+    generator: VariantGenerator = Depends(get_variant_generator),
+) -> VariantResponse:
+    """The learner's variant for the resolved template — generated on the
+    first request, cached (no LLM call) on every repeat: D-029 makes a
+    regenerate a 200 of the SAME stored variant.
+    """
+    template = _resolve_template(body)
+    record = await generator.generate(body.learner_id, template.id)
+    return _to_response(record, competency_id=template.competency_id)
+
+
+@router.get("", response_model=VariantListResponse)
+async def list_variants(
+    learner_id: str,
+    store: VariantStore = Depends(get_variant_store),
+) -> VariantListResponse:
+    """All stored variants for the learner, chronological; [] when none."""
+    variants = [
+        _to_response(record, competency_id=_competency_for(record.template_id))
+        for record in store.list_for_learner(learner_id)
+    ]
+    return VariantListResponse(variants=variants)
+
+
+@router.get("/{task_id}", response_model=VariantResponse)
+async def get_variant(
+    task_id: str,
+    store: VariantStore = Depends(get_variant_store),
+) -> VariantResponse:
+    """The stored variant owning the task key — the grading and telemetry
+    join path; 404 when no variant was ever generated for the task."""
+    record = store.get_by_task(task_id)
+    if record is None:
+        raise HTTPException(
+            status_code=404, detail=f"no stored variant for task {task_id!r}"
+        )
+    return _to_response(record, competency_id=_competency_for(record.template_id))
@@ -19,6 +19,7 @@ from .api import (
    proctor_router,
    sandboxes_router,
    telemetry_router,
+    variants_router,
 )
 from .config import Settings
 from .grading.engine import GradingEngine
@@ -27,6 +28,8 @@ from .llm import create_provider
 from .sandbox import SandboxManager, UnshareBackend
 from .telemetry.ingest import TraceIntegrityMap
 from .telemetry.store import SQLiteTraceStore
+from .variants.generator import VariantGenerator
+from .variants.store import SQLiteVariantStore

 logger = logging.getLogger(__name__)

@@ -94,6 +97,27 @@ def create_app(settings: Settings | None = None) -> FastAPI:
                model=settings.model,
            )

+        # Variant generation (REQ-3-005, Wave 3): VariantStore from the same
+        # SQLite file as traces/grades (D-027), one VariantGenerator singleton
+        # wired through app.state — the generator receives store + provider
+        # via constructor DI and knows nothing of FastAPI (api/ composes it,
+        # same pattern as GradingEngine). Tests may pre-set
+        # app.state.variant_store / app.state.variant_generator (the same
+        # state-injection override); the lifespan adopts a pre-set store but
+        # NEVER rebuilds a pre-set generator (its provider binding is part
+        # of the test fixture).
+        variant_store = getattr(app.state, "variant_store", None)
+        if variant_store is None:
+            settings.db_path.parent.mkdir(parents=True, exist_ok=True)
+            variant_store = SQLiteVariantStore(db_path=settings.db_path)
+            app.state.variant_store = variant_store
+        if getattr(app.state, "variant_generator", None) is None:
+            app.state.variant_generator = VariantGenerator(
+                variant_store,
+                app.state.provider,
+                model=settings.model,
+            )
+
        async def _reaper_loop() -> None:
            # Wall-clock timeout + G-2 workdir-size sweep, one pass per tick.
            while True:
@@ -115,6 +139,7 @@ def create_app(settings: Settings | None = None) -> FastAPI:
            await manager.destroy_all()
            trace_store.close()
            grade_store.close()
+            variant_store.close()
            await app.state.http_client.aclose()

    app = FastAPI(title="Nextcraft AI Service", version="0.3.0", lifespan=lifespan)
@@ -143,6 +168,7 @@ def create_app(settings: Settings | None = None) -> FastAPI:
    app.include_router(proctor_router)
    app.include_router(sandboxes_router)
    app.include_router(telemetry_router)
+    app.include_router(variants_router)
    return app


@@ -0,0 +1,277 @@
+"""Variant API tests — generation, cache, distinctness over HTTP (Task 4-3-01).
+
+Contract under test (api/variants.py, REQ-3-005):
+
+    POST /v1/variants {learner_id, template_id}
+        first request       → 200 generated variant (task_id, seed, params,
+                              statement, starter_files, competency_id)
+        repeat request      → 200 the SAME cached variant with ZERO LLM
+                              calls (D-029 reproducibility through the
+                              whole HTTP stack)
+        unknown template    → 404
+    POST /v1/variants {learner_id, competency_id}
+        bound competency    → 200 the first template for that competency
+        unknown competency  → 404
+        neither id given    → 422
+    GET  /v1/variants/{task_id}
+        generated earlier  → 200 the stored variant (generate→get roundtrip)
+        unknown task       → 404
+    GET  /v1/variants?learner_id=...
+        → 200 that learner's variants only (scoping); [] for a learner
+          with none.
+
+Distinctness (REQ-3-005, API level): two learners POSTing the same
+template receive distinct statements, seeds and task_ids, and GET by
+task_id hands each back their own variant.
+
+Wiring: per-test tmp-path SQLiteVariantStore + a pre-set VariantGenerator
+(state-injection override — the lifespan adopts variant_store /
+variant_generator from app.state instead of constructing them; same
+pattern as test_grading.py / test_telemetry_ingest.py). The generator
+binds a ScriptedRenderProvider so each test controls the render LLM
+exactly, and counts calls so the cache test can assert the LLM was never
+reached on the second POST.
+
+Zero network: providers are MockProvider family members only (conftest
+rule, enforced in _make_client).
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from pathlib import Path
+
+import pytest
+from fastapi.testclient import TestClient
+
+from ai_service.config import Settings
+from ai_service.llm.mock import MockProvider
+from ai_service.main import create_app
+from ai_service.variants.generator import VariantGenerator
+from ai_service.variants.store import SQLiteVariantStore
+from ai_service.variants.templates import get_template
+
+LEARNER_A = "variant-learner-a"
+LEARNER_B = "variant-learner-b"
+TEMPLATE = "tpl-llm-judge"
+COMPETENCY = "stack-orchestration-c007"  # tpl-llm-judge's D-021 binding
+
+
+class ScriptedRenderProvider(MockProvider):
+    """Deterministic render whose statement embeds the params (distinct per
+    draw); counts calls so the cache test asserts zero LLM calls on the
+    second POST. Mirrors tests/variants/test_generator.py."""
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.calls = 0
+
+    async def chat(self, messages, *, model, temperature=0.7, response_format=None):  # noqa: ANN001
+        self.calls += 1
+        import json
+
+        user = next(m.content for m in reversed(messages) if m.role == "user")
+        # Distinct per distinct params: hash the seeded slot lines.
+        fingerprint = abs(hash(user)) % 10_000
+        return json.dumps({"statement": f"Scripted variant #{fingerprint} — build it."})
+
+
+@pytest.fixture()
+def store(tmp_path: Path) -> Iterator[SQLiteVariantStore]:
+    s = SQLiteVariantStore(db_path=tmp_path / "variants.db")
+    yield s
+    s.close()
+
+
+@pytest.fixture()
+def provider() -> ScriptedRenderProvider:
+    return ScriptedRenderProvider()
+
+
+def _make_client(
+    tmp_path: Path, store: SQLiteVariantStore, provider: MockProvider
+) -> TestClient:
+    """App + TestClient with a pre-set store + generator.
+
+    The lifespan adopts both (state-injection override); the generator
+    binds OUR provider, so the fixture — not Settings — scripts the LLM.
+    Cloud-free guard: the provider must be a MockProvider family member
+    (conftest rule, enforced here because this module builds its own
+    client rather than consuming the conftest one).
+    """
+    assert isinstance(provider, MockProvider)
+    settings = Settings(
+        provider="mock",
+        db_path=tmp_path / "variant-test.db",
+        sandbox_dir=tmp_path / "sandboxes",
+    )
+    app = create_app(settings)
+    app.state.variant_store = store
+    if getattr(app.state, "variant_generator", None) is None:
+        app.state.variant_generator = VariantGenerator(
+            store, provider, model="gemma4:31b"
+        )
+    return TestClient(app)
+
+
+def _post(
+    client: TestClient,
+    learner_id: str,
+    template_id: str | None = TEMPLATE,
+    competency_id: str | None = None,
+):
+    return client.post(
+        "/v1/variants",
+        json={
+            "learner_id": learner_id,
+            **({"template_id": template_id} if template_id is not None else {}),
+            **({"competency_id": competency_id} if competency_id is not None else {}),
+        },
+    )
+
+
+# -- POST: generation + response shape --------------------------------------------
+
+
+class TestGenerate:
+    def test_post_returns_full_variant_payload(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = _post(client, LEARNER_A)
+
+        assert response.status_code == 200
+        body = response.json()
+        assert body["learner_id"] == LEARNER_A
+        assert body["template_id"] == TEMPLATE
+        assert body["competency_id"] == COMPETENCY  # enrichment from get_template
+        assert body["task_id"].startswith("task-") and len(body["task_id"]) == len("task-") + 16
+        assert body["seed"]  # non-empty D-029 seed
+        assert body["statement"]  # rendered, non-empty
+        assert body["starter_files"] == get_template(TEMPLATE).starter_files
+        assert body["params"]  # seeded slot values
+        assert body["created_at"]  # ISO timestamp travels
+        assert provider.calls == 1  # first POST renders exactly once
+
+    def test_post_then_get_roundtrip(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            posted = _post(client, LEARNER_A)
+            assert posted.status_code == 200
+
+            fetched = client.get(f"/v1/variants/{posted.json()['task_id']}")
+
+        assert fetched.status_code == 200
+        assert fetched.json() == posted.json()  # identical stored variant
+
+    def test_post_regenerate_is_cache_hit_no_llm_call(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            first = _post(client, LEARNER_A)
+            assert first.status_code == 200
+            assert provider.calls == 1
+
+            second = _post(client, LEARNER_A)  # same (learner, template) → cache
+            assert second.status_code == 200
+            assert second.json() == first.json()  # D-029: the SAME variant
+            assert provider.calls == 1, "cached regenerate must not render again"
+
+
+# -- POST: distinctness (REQ-3-005, API level) ------------------------------------
+
+
+class TestDistinctLearners:
+    def test_two_learners_distinct_statements_and_task_ids(
+        self, tmp_path, store, provider
+    ):
+        with _make_client(tmp_path, store, provider) as client:
+            a = _post(client, LEARNER_A)
+            b = _post(client, LEARNER_B)
+            assert a.status_code == 200 and b.status_code == 200
+
+            a_body, b_body = a.json(), b.json()
+
+            # ...and each learner GETs back exactly their own variant
+            a_fetched = client.get(f"/v1/variants/{a_body['task_id']}")
+            b_fetched = client.get(f"/v1/variants/{b_body['task_id']}")
+
+        assert a_body["statement"] != b_body["statement"]
+        assert a_body["seed"] != b_body["seed"]
+        assert a_body["task_id"] != b_body["task_id"]
+        assert a_body["competency_id"] == b_body["competency_id"]  # same bar (a-5)
+        assert a_fetched.status_code == 200 and a_fetched.json() == a_body
+        assert b_fetched.status_code == 200 and b_fetched.json() == b_body
+
+
+# -- POST: template / competency resolution ---------------------------------------
+
+
+class TestTemplateResolution:
+    def test_unknown_template_404(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = _post(client, LEARNER_A, template_id="tpl-does-not-exist")
+
+        assert response.status_code == 404
+        assert "no task template" in response.json()["detail"]
+
+    def test_competency_lookup_uses_first_bound_template(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = _post(
+                client, LEARNER_A, template_id=None, competency_id=COMPETENCY
+            )
+
+        assert response.status_code == 200
+        body = response.json()
+        assert body["template_id"] == TEMPLATE  # the competency's first template
+        assert body["competency_id"] == COMPETENCY
+
+    def test_unknown_competency_404(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = _post(
+                client, LEARNER_A, template_id=None, competency_id="stack-none-c999"
+            )
+
+        assert response.status_code == 404
+        assert "no task template for competency" in response.json()["detail"]
+
+    def test_neither_id_given_422(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = _post(client, LEARNER_A, template_id=None, competency_id=None)
+
+        assert response.status_code == 422
+
+
+# -- GET: stored reads ------------------------------------------------------------
+
+
+class TestStoredReads:
+    def test_get_unknown_task_404(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = client.get("/v1/variants/task-0000000000000000")
+
+        assert response.status_code == 404
+        assert "no stored variant" in response.json()["detail"]
+
+    def test_list_scoped_to_one_learner(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            assert _post(client, LEARNER_A).status_code == 200
+            assert _post(
+                client, LEARNER_A, template_id="tpl-guardrail-schema"
+            ).status_code == 200
+            assert _post(client, LEARNER_B).status_code == 200
+
+            listed = client.get("/v1/variants", params={"learner_id": LEARNER_A})
+            empty = client.get("/v1/variants", params={"learner_id": "nobody"})
+
+        assert listed.status_code == 200
+        variants = listed.json()["variants"]
+        assert len(variants) == 2  # LEARNER_A's two, LEARNER_B's excluded
+        assert {v["learner_id"] for v in variants} == {LEARNER_A}
+        assert {v["template_id"] for v in variants} == {TEMPLATE, "tpl-guardrail-schema"}
+        # chronological ordering; both carry the competency enrichment
+        assert all(v["competency_id"] for v in variants)
+
+        assert empty.status_code == 200
+        assert empty.json()["variants"] == []
+
+    def test_list_missing_learner_param_422(self, tmp_path, store, provider):
+        with _make_client(tmp_path, store, provider) as client:
+            response = client.get("/v1/variants")
+
+        assert response.status_code == 422