nextcraft/.ciagent/ARCHITECTURE.md

Nextcraft — ARCHITECTURE.md

Overview

Nextcraft is a TypeScript monorepo (pnpm workspaces + turborepo) with a Next.js application hosting four surfaces (Learner, Marketplace, Employer Dashboard, Admin), a shared component library, typed mock data layer, and shared types package. As of v0.2, a Python FastAPI application (apps/ai-service) hosts six AI tutor agents backed by a provider-agnostic LLM layer.

v0.3 additions (Credential Engines): real credential engines replace v0.2 mock inputs — a sandbox fabric (isolated per-learner coding environments via Linux user/mount/pid/net namespaces), a live build-telemetry pipeline (WebSocket ingest + SQLite-ordered event log), a process-trace grading engine, seeded per-learner variant task generation, and a voice-based oral defense (STT/TTS via a new provider-agnostic voice layer). First real persistence introduced: SQLite (ai_service/telemetry/, grading, variant, defense stores). Lab/Assessor/Proctor agents are re-grounded onto real telemetry/traces. Identity/age-gating (KYC) deferred per founder directive — no security engineer persona; secrets-hygiene checklist only.

Confirmed Technology Stack (v0.2)

Technology	Version	Purpose
Node.js	v24.15.0	Runtime (web)
pnpm	12.3.4	Package manager + workspaces
turborepo	2.3.3	Build orchestration
Next.js	15 (App Router)	Web application framework
React	19+	UI library
TypeScript	5.x	Type system
Tailwind CSS	v4	Utility-first CSS
lucide-react	latest	Icons
recharts	latest	Charts
@xyflow/react	latest	Competency graph viewer
Python	3.11.2	Runtime (ai-service)
FastAPI	0.141.x	AI service framework
uvicorn	0.52.x	ASGI server
pydantic	2.13.x	Request/response models, structured outputs
pydantic-settings	2.15.x	Settings + env-file loading (replaces python-dotenv)
httpx	0.28.x	Async LLM HTTP client (ollama-cloud + local providers)
sse-starlette	3.4.x	SSE framing, ping keep-alive
pytest	9.x	Test runner
pytest-asyncio	1.4.x	Async tests (auto mode)
ruff	latest	Python lint (check-only, no formatter) — pnpm ai:lint
ollama-cloud	https://ollama.com/v1	Default LLM provider (OpenAI-compatible, Bearer auth)

Deliberately not used: openai-python SDK (the LLMProvider protocol is the port; raw httpx keeps delta passthrough and ollama-cloud quirk tolerance), python-dotenv (pydantic-settings reads .env natively), respx (httpx MockTransport is built in).

v0.2 Architecture Decisions (from Research)

1. D-016 SSE envelope — meta event (agent/session/model, flushed before first token) → raw OpenAI-compatible chunk passthrough → optional done → error event before [DONE] on mid-stream failure. Pre-first-byte failures use proper HTTP status. Headers: Cache-Control: no-cache, X-Accel-Buffering: no.
2. D-017 httpx direct client — lifespan-managed httpx.AsyncClient (10s connect / 300s read), shared by ollama-cloud and local providers; no SDK.
3. D-018 Agent framework — BaseAgent ABC (system_prompt/build_messages/stream_reply/structured_reply) + explicit registry; prompts are versioned code in prompts/.
4. D-019 Session store — SessionStore protocol + InMemorySessionStore (asyncio.Lock, 20-message window, 500-cap LRU, agent-scoped sessions). DB-migration-ready.
5. D-020 Structured outputs — 4-layer defense: response_format (auto-degrade) → prompt-embedded schema → fence-strip/first-balanced-object parse → single bounded retry.
6. D-021 Mock corpus in Python — ai_service/corpus/ pydantic-typed, convention-aligned with TS packages/mock-data (shared IDs, cross-referencing headers); no codegen in v0.2.
7. D-022 Monorepo integration — zero-dependency shim package.json in apps/ai-service + ai#* turbo passthrough tasks (cache:false, outputs:[]) + root ai:dev/ai:test scripts + idempotent venv bootstrap.
8. D-023 Testing — pytest-asyncio auto mode; TestClient client.stream() for SSE; httpx MockTransport for byte-exact provider parser tests; scripted mock provider incl. failure modes. Tests never call the cloud.

v0.3 Architecture Decisions (from Research — Credential Engines)

9. D-024 Sandbox isolation = Linux namespaces via unshare — per-learner sandbox runs as a subprocess entered into fresh user+mount+pid+network namespaces (unshare --user --map-root-user --mount --pid --fork --net). Probe-verified on this box: in-namespace uid=0, network fully isolated (0 interfaces), learner writes land in a per-sandbox directory; proc-remount not permitted here but not required. Chosen because no container runtime (docker/podman/bwrap/firejail) exists on the box and there is no sudo. A SandboxBackend protocol abstracts the spawner so a future containerd/runc backend can replace namespace-spawning without touching callers.
10. D-025 Sandbox scope = coding IDE only (v0.3) — the sandbox fabric provisions a single build environment (shell + filesystem + run/test). REQ-F-021's design-tool and simulation environments are deferred to v0.4; one real build path proves the full credential pipeline (telemetry → trace → grade → defense).
11. D-026 Telemetry = WebSocket ingest + SQLite ordered event log — in-sandbox capture agent streams structured events over WebSocket to ai_service (/v1/telemetry/ingest); events persisted to SQLite with a per-(learner,task) monotonic seq for gap detection, giving durability + at-least-once delivery + replay without a message broker.
12. D-027 First persistence = SQLite, protocol-wrapped — introduces a real DB (ai_service/data/*.db) for telemetry traces, grades, variants, and defenses. Access via SQLModel. Every store is a protocol (TraceStore, VariantStore, DefenseStore, GradeStore) with a SQLite implementation — Postgres-migration-ready, mirroring D-019's SessionStore pattern.
13. D-028 Process-trace grading = hybrid deterministic + LLM — deterministic features (test pass/fail, edit count, error/fix cycles, idle gaps, command categories) computed in code into a compact trace digest; the digest feeds an Assessor-style rubric prompt and returns structured scores via D-020 JSON defense. The LLM never sees the raw trace — only the digest.
14. D-029 Variant generation = seeded template instantiation — task templates with typed parameter slots; an LLM instantiates a unique variant per learner from a seed; seed+parameters persisted (D-027) for grading fairness and proctoring cross-check.
15. D-030 Voice = provider-agnostic, mock-first, browser-fallback — a VoiceProvider protocol (mirror of LLMProvider) with STT (OpenAI-compatible /audio/transcriptions) + TTS (/audio/speech) against a configurable endpoint, a deterministic mock (canned transcript/audio) for tests, and browser-native SpeechRecognition/speechSynthesis as a no-key fallback. Voice defense reuses BaseAgent + the existing SSE pipeline (new Examiner agent).
16. D-031 No new apps — extend ai-service — telemetry, grading, variant, voice, and sandbox orchestration are new modules inside apps/ai-service (sharing the LLM pool, config, and session infra). Only the in-sandbox capture agent is a separate tiny Python process shipped into the namespace. No new top-level apps/ entry.
17. D-032 Capacity = single-box, 1–5 concurrent sandboxes — concurrency guard returns 503 when the sandbox pool is full. No queueing, no horizontal scaling in v0.3 (solo-founder/pilot scale).

---

Components

apps/ai-service — AI Tutor Service (v0.2 NEW)

Component	Description	Boundaries	Depends On
ai_service/main.py	FastAPI app factory, lifespan (httpx client pool, provider factory), CORS (localhost only), /health	App entry	config, llm, agents, api
ai_service/config.py	pydantic-settings Settings (env_prefix="AI_", env_file, SecretStr key)	Configuration only	None
ai_service/api/	Endpoints: chat.py (POST /v1/chat/stream, SSE), lab.py, assessment.py, proctor.py, mentor.py; deps.py (DI)	Composes agents + sessions; never imported by llm/ or agents/	agents, llm
ai_service/llm/	types.py (Message; ChatDelta/ChoiceDelta removed in P3 — no consumers), base.py (LLMProvider protocol), openai_compat.py (ollama-cloud + local), mock.py (deterministic), factory.py	Never imports agents/ or api/	config
ai_service/agents/	base.py (BaseAgent ABC), registry.py, session.py (SessionStore), structured.py (JSON defense), coach/tutor/lab/assessor/proctor/mentor.py	Never imports api/	llm, prompts, corpus
ai_service/prompts/	Per-agent system prompt constants + render_context functions (str.format_map)	Data only	None
ai_service/corpus/	Mock engine inputs: learner_context.py, telemetry.py (Lab/Proctor scenarios), artifacts.py (pre-baked artifacts, rubrics, transcripts)	Pydantic-typed; aligned with TS packages/mock-data by convention	None
scripts/	bootstrap.sh (venv + pip install idempotent), dev.sh (exports keys from .ciagent/.env.secrets → uvicorn), test.sh (pytest), lint.sh (ruff check, G-3)	Dev entry points	pyproject.toml
tests/	conftest.py (mock provider, settings override, TestClient), health, llm (MockTransport parser), agents (framework + per-agent), api (SSE stream tests)	Mock provider only — no cloud	all

Module boundary rules: llm/ never imports agents/ or api/; agents/ never imports api/; api/ composes both via DI. corpus/ is the only home of mock engine data. Prompts are code — versioned and reviewed in git.

apps/ai-service — v0.3 Credential Engine modules (NEW)

Component	Description	Boundaries	Depends On
ai_service/sandbox/	backend.py (SandboxBackend protocol), unshare_backend.py (userns/mount/pid/net spawner, D-024), manager.py (lifecycle: create/list/snapshot/destroy + concurrency guard D-032), workdir.py (per-sandbox fs layout)	Never imports api/ or agents/; spawns subprocesses only	config
ai_service/telemetry/	models.py (TelemetryEvent, TraceSpan), store.py (TraceStore protocol + SQLite impl D-027), ingest.py (WebSocket /v1/telemetry/ingest, seq gap detection D-026)	Persistence; never imports agents/	config
ai_service/grading/	features.py (deterministic trace digest D-028), engine.py (rubric scoring orchestration), store.py (GradeStore)	LLM only via digest; never sees raw trace	llm, telemetry, prompts
ai_service/variants/	templates.py (task template library), generator.py (seeded LLM instantiation D-029), store.py (VariantStore)	LLM via structured output	llm, grading
ai_service/voice/	base.py (VoiceProvider protocol D-030), browser.py (native SR/TTS fallback descriptor), mock.py (deterministic; the real server STT/TTS provider is the v0.4 seam — GRILL CUT-1/G-7)	Never imports agents/ or api/	config
ai_service/agents/examiner.py	Seventh agent: oral defense examiner; streams over existing SSE, consumes process traces + emits integrity signals	reuses BaseAgent (D-018)	llm, prompts, telemetry
ai_service/data/*.db	SQLite databases (telemetry/grades/variants/defenses)	gitignored	—
scripts/sandbox-agent.py	Tiny in-namespace capture process shipped into the sandbox; streams telemetry to ingest	standalone	stdlib only

Boundary additions: sandbox/, telemetry/, grading/, variants/, voice/ are engine modules — they never import api/ (which composes them via DI) and never import agents/ (agents call engines through narrow interfaces, not vice versa).

apps/web — Next.js Application

Component	Description	Boundaries	Depends On
app/(learner)/	Learner surface route group: landing, catalog, competency stack, dashboard, byte viewer, sandbox mockup, assessment mockup	Learner-only routes and layouts	packages/ui, packages/mock-data, packages/types
app/(marketplace)/	Marketplace surface route group: job board, job detail, employer profile, search/filter, pricing	Marketplace-only routes and layouts	packages/ui, packages/mock-data, packages/types
app/(employer)/	Employer dashboard route group: overview, talent search, candidate profile, posting management	Employer-only routes and layouts	packages/ui, packages/mock-data, packages/types
app/(admin)/	Admin surface route group: overview, learner management, competency graph viewer, moderation	Admin-only routes and layouts	packages/ui, packages/mock-data, packages/types
app/layout.tsx	Root layout: theme provider, navigation shell, responsive container	All routes	packages/ui
components/	Surface-specific components (learner/, marketplace/, employer/, admin/) plus shared chrome (navigation-shell, header/footer, role-switcher, theme-provider, breadcrumbs, dark-mode-toggle)	App-level components	packages/ui
hooks/	use-chat-stream.ts — SSE client hook: fetch + ReadableStream, byte buffering + frame reassembly, idempotent AbortController cleanup	Client components only	ai-service SSE
lib/	sse.ts (shared SSE frame parser — CRLF normalization + : ping immunity, G-1), breadcrumbs.ts, format.ts	Pure utilities	None

packages/ui — Shared Component Library

Component	Description	Boundaries	Depends On
tokens/	Design tokens as TS constants: colors, spacing, radii, shadows, breakpoints (mirrored as Tailwind v4 @theme tokens in apps/web globals.css)	Foundation layer — no dependencies	None
primitives/	Button, Input, Card, Badge, Avatar — each with a Storybook story	Atomic UI components	tokens, packages/types

Composite/layout/theme components (navigation shell, tables, chat panels, graph viewer, theme provider) live in apps/web/components/ as app-level components, not in packages/ui.

packages/mock-data — Mock Data Layer

Component	Description	Boundaries	Depends On
competency-stacks.ts	5 competency stacks (AI Orchestration Engineer, AI Safety & Governance, Human-AI Product Designer, AI-Augmented Field Operator, Computational Sciences), each with 12-18 competencies	Typed mock data	packages/types
jobs.ts	20+ mock AI-era job listings with skills, seniority, salary, match scores	Typed mock data	packages/types
candidates.ts	15+ mock candidate profiles with artifacts, process traces, defense scores, microcredentials	Typed mock data	packages/types
employers.ts	10+ mock employer profiles with logos, descriptions, open positions	Typed mock data	packages/types
learner-progress.ts	Mock learner progress data: active competencies, completion percentages, recent artifacts	Typed mock data	packages/types
admin.ts	Admin surface mock data: platform metrics, activity feed, system health, learner roster (admin view), moderation queues	Typed mock data	packages/types
ai-scenarios.ts	AI engine-input scenario IDs + display metadata for the learner agent panels; IDs string-identical to ai_service/corpus/ (D-021)	Typed mock data	packages/types

packages/types — Shared Types

Component	Description	Boundaries	Depends On
domain.ts	Competency, CompetencyStack, Microcredential, Artifact, ProcessTrace, OralDefense, AssessmentRubric	Domain types	None
marketplace.ts	Job, Employer, Candidate, JobPosting, TalentMatch, SearchFilter	Marketplace types	None
user.ts	Learner, Admin, EmployerUser, AgeGroup, Role	User types	None
ui.ts	Component props, theme config, breakpoint definitions	UI types	None

---

Data Flow

[packages/mock-data + packages/types]      [ai_service/corpus]
        │ (TS, web surfaces)                      │ (Python, agent inputs)
        ▼                                         ▼
[Next.js Route Groups]                    [ai-service agents]
  (learner)/ (marketplace)/                 coach tutor lab assessor proctor mentor
  (employer)/ (admin)/                                │
        │                                             ▼
        │      SSE (fetch + ReadableStream)     [LLMProvider]
        └────── client components ◄───────────────────┤
               http://localhost:8420          ollama-cloud / local / mock
                                               (https://ollama.com/v1)

• Web surfaces remain server-component-first; client components (chat, filters, graph viewer, dark mode toggle) fetch directly from ai-service over SSE (A-002: no Next.js API-route proxy in v0.2).
• The LLM provider layer is a dumb pipe — OpenAI-compatible chunks pass through byte-identical; envelope logic (meta/done/error) lives only in the API layer (D-016).
• Lab/Assessor/Proctor read mock scenarios from ai_service/corpus/ — real engines are v0.3+.
• All automated tests use the deterministic mock provider; the cloud is for manual probes only.

---

Build Order (v0.3)

1. Sandbox fabric — SandboxBackend protocol + unshare namespace spawner + lifecycle manager (create/list/snapshot/destroy) + concurrency guard + per-sandbox workdir; isolation + resource-limit probes
2. Live build telemetry — TelemetryEvent models + SQLite TraceStore + WebSocket ingest endpoint + seq gap detection + in-sandbox capture agent
3. Process-trace grading engine — deterministic feature/digest computation + rubric scoring via LLM structured output + GradeStore; calibrated against v0.2 mock corpora
4. Variant task generation — template library + seeded LLM instantiation + VariantStore + difficulty normalization anchors
5. Oral / voice defense — VoiceProvider protocol + STT/TTS + mock + browser fallback + Examiner agent + transcript/integrity-signal capture
6. Agent re-grounding + learner surface integration — Lab/Assessor/Proctor consume real telemetry/grades/defense signals; learner sandbox mockup → real in-browser xterm.js build/run; assessment mockup → live defense + live grading

---

Build Order (v0.2 — complete)

1. AI service scaffolding — apps/ai-service: FastAPI app, config, provider layer (ollama-cloud/local/mock), SSE chat endpoint, pytest harness, turbo integration
2. Agent framework — BaseAgent, registry, session store, structured output, prompts scaffolding, learner-context corpus
3. Coach + Tutor agents — full implementations, chat endpoint agent routing
4. Lab + Assessor agents — telemetry scenarios + pre-baked artifacts corpus, /v1/lab/feedback + /v1/assessment/evaluate
5. Proctor + Mentor agents — proctor scenarios, /v1/proctor/signals + /v1/mentor/narrative
6. Learner surface integration — useChatStream hook, agent switcher, streaming/error/loading states, agent output panels across the four learner surfaces

The v0.1 build order (monorepo → types → mock data → tokens → primitives → layout → composites → surfaces → polish) is complete and preserved in git history (tags v0.0.1–v0.1.0).

---

Future Architecture (Post-v0.3, for reference)

v0.3 delivers the real credential engines; later milestones fill in the remaining platform:

• In-memory sessions → PostgreSQL + Drizzle/SQLModel — SessionStore + v0.3 TraceStore/GradeStore/VariantStore/DefenseStore protocols swap SQLite→Postgres with no API changes
• userns subprocess sandboxes → containerd/runc backend — D-024 SandboxBackend protocol swap; same lifecycle API
• Coding-IDE sandbox → design tool + simulation environments — REQ-F-021 full scope (v0.4)
• Mock provider → per-agent model routing — provider factory already selects by config; per-agent AI_<AGENT>_MODEL overrides
• No auth → real KYC + sessions — deferred per founder directive; REQ-F-017 identity/age-gating lands post-v0.3 (v0.4+). Age-gating remains the v0.1 visual flow mockup
• No search → Semantic vector search (pgvector) — Filter UI replaced with vector similarity search
• No payments → Payment processing — Pricing page replaced with real subscription/payment flows

The monorepo structure (apps/web + apps/ai-service + packages/*) accommodates further apps without restructuring.