The codegen consumes a schema from any backend and emits typed client code for any frontend — it doesn't belong inside a backend adapter. That placement was historical sediment from when there was only a Django backend; it predates the AFI generalization. New top-level slot: `protocol/` for protocol-level tooling. Tree is now: backends/ server protocol adapters frontends/ client kernel + per-framework adapters cores/ shared language-level primitives protocol/ protocol-level tooling workers/ runtime workers / bridges Codegen moves to `protocol/mizan-generate/`. Same file layout under `generator/` (cli.mjs, lib/), preserved via git mv. Package metadata cleaned up: - name: "generate" (placeholder) → "mizan-generate" - description filled in - type: module (cli.mjs is .mjs ESM, was previously declared "commonjs") - bin entry added so `npx mizan-generate --config <config.mjs>` works once the package is published, instead of `node path/to/cli.mjs`. Path-reference fixups: - backends/mizan-django/README.md: `node path/to/...` → `npx mizan-generate` - backends/mizan-fastapi/README.md: same - ISSUES.md: file paths in three issue entries - CLAUDE.md: codegen description + Package Layout section refreshed (added protocol/, mizan-fastapi entry, mizan-python entry) - docs/AFI_ARCHITECTURE.md: Package Layout refreshed identically Verified codegen runs from new location: regenerated the FastAPI example harness's api/ output, identical to pre-move. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
16 KiB
Mizan — Technical Reference
What Mizan Is
Mizan is an Application Framework Interface (AFI). One decorator on a server function. Typed client generated. Invalidation automatic. Caching protocol-driven. SSR via subprocess.
Django + React ships first. The protocol is language-agnostic (proven by mizan-ts).
Package Layout
Tree organized by role. Per-framework adapters wrap a single shared kernel; codegen targets the adapter.
backends/ server protocol adapters
mizan-django/ Django adapter
mizan-fastapi/ FastAPI adapter (RPC + context + invalidation; AFI-common scope)
mizan-ts/ TypeScript adapter (proves the protocol is language-agnostic)
frontends/ client kernel + per-framework adapters
mizan-base/ framework-agnostic kernel; owns data, status, error; adapters subscribe
mizan-react/ React contexts + hooks over the kernel
mizan-vue/ Vue composables over the kernel
mizan-svelte/ Svelte stores/runes over the kernel
cores/ shared language-level primitives
mizan-python/ @client decorator, registry, MWT, HMAC cache keys; consumed by both Python backends
protocol/ protocol-level tooling
mizan-generate/ codegen — fetches schema from any backend, emits typed React/Vue/Svelte client
workers/ runtime workers / bridges
mizan-ssr/ Bun subprocess used by the Django template backend
The Three Protocols
1. RPC Protocol (Anti-REST)
No resources. No CRUD. Functions in, results out.
Context fetch (reads):
GET /api/mizan/ctx/<context_name>/?param1=val1¶m2=val2
200 OK
Cache-Control: no-store
Content-Type: application/json
{
"user_profile": {"name": "Ryth", "email": "ryth@example.com"},
"user_orders": [{"id": 1, "total": 100}]
}
All functions sharing a context name are bundled into one response. Keys are function names. Values are return values.
Mutation call (writes):
POST /api/mizan/call/
Content-Type: application/json
{"fn": "update_profile", "args": {"user_id": 5, "name": "Ryth"}}
200 OK
Cache-Control: no-store
X-Mizan-Invalidate: user;user_id=5
{
"result": {"ok": true},
"invalidate": [{"context": "user", "params": {"user_id": 5}}]
}
2. Invalidation-on-Mutation Protocol
Two transports for the same signal. Both are first-class.
Transport 1 — JSON body (for RPC/SPA clients):
{"result": {...}, "invalidate": ["user"]}
{"result": {...}, "invalidate": [{"context": "user", "params": {"user_id": 5}}]}
Transport 2 — HTTP header (for Edge, htmx, view-path functions):
X-Mizan-Invalidate: user
X-Mizan-Invalidate: user;user_id=5
X-Mizan-Invalidate: user;user_id=5, notifications
Format: comma-separated contexts, semicolon-separated URL-encoded params per context.
Three-tier auto-scoping (no developer annotation needed):
- Argument name matching: mutation has
user_idparam, context hasuser_idparam → scoped automatically - Auth inference: Edge-side concern (reads JWT/MWT to extract user identity)
- Broad fallback: invalidate all instances of the context
Return-type branching determines which transport:
- Function returns data (dict, BaseModel) → RPC path → JSON body + header
- Function returns HttpResponse (redirect, HTML) → View path → header only
3. Frontend-Agnostic Rendering (SSR + PSR)
SSR — Django template backend integration. render(request, 'ProfilePage', props) calls a persistent Bun subprocess that runs renderToString.
PSR (Preemptive Static Rendering) — pages re-rendered on mutation, not on request. Edge caches the result. Controlled by the manifest's render_strategy field.
The Bun worker protocol — JSON-RPC over stdin/stdout:
→ {"id": 1, "method": "render", "params": {"component": "ProfilePage", "props": {"userId": 5}}}
← {"id": 1, "html": "<div>...</div>"}
Worker stays alive across requests. Django's SSRBridge manages the subprocess lifecycle with thread-safe request correlation via message IDs.
The @client Decorator — Full API
from mizan import client, ReactContext, GlobalContext
UserContext = ReactContext('user')
Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
context |
ReactContext | str | False |
False |
Named context for grouping. False = standalone function. |
affects |
ReactContext | str | list |
None |
What this mutation invalidates. Mutually exclusive with context. |
private |
bool |
False |
Not client-callable. No RPC endpoint. No codegen. Still in invalidation graph. |
route |
str | None |
None |
Mizan-owned URL pattern for view-path functions. |
methods |
list[str] | None |
None |
HTTP methods for route. Default: ['GET'] for context, ['POST'] for mutation. |
auth |
bool | str | callable | None |
None |
Auth requirement: True, 'staff', 'superuser', or callable(request) -> bool. |
websocket |
bool |
False |
Enable WebSocket RPC transport. |
rev |
int |
0 |
Cache revision. Increment to bust cached entries on deploy. |
cache |
int | False |
(default) | Cache TTL hint. False = never cache. Integer = TTL seconds. |
Usage Patterns
# Global context — auto-mounted at root, SSR-hydrated
@client(context=GlobalContext)
def current_user(request) -> UserShape:
return UserShape.query(lambda qs: qs.filter(pk=request.user.pk))[0]
# Named context — bundled GET, generates typed hooks
@client(context=UserContext)
def user_profile(request, user_id: int) -> UserShape:
return UserShape.query(lambda qs: qs.filter(pk=user_id))[0]
@client(context=UserContext)
def user_orders(request, user_id: int) -> list[OrderShape]:
return OrderShape.query(lambda qs: qs.filter(user_id=user_id))
# Mutation — auto-scoped invalidation (user_id matches)
@client(affects=UserContext)
def update_profile(request, user_id: int, name: str) -> dict:
request.user.name = name
request.user.save()
return {"ok": True}
# Function-level affects — only user_profile refetches
@client(affects='user_profile')
def update_name(request, user_id: int, name: str) -> dict:
...
# View-path context — registered in invalidation graph, no codegen
@client(context=UserContext, route='/profile/<user_id>/')
def profile_page(request, user_id: int) -> HttpResponse:
return render(request, 'profile.html', {...})
# View-path mutation — invalidation via header on the redirect
@client(affects=UserContext, route='/profile/<user_id>/update/', methods=['POST'])
def update_profile_view(request, user_id: int) -> HttpResponse:
form = ProfileForm(request.POST)
if form.is_valid():
form.save()
return redirect(f'/profile/{user_id}/')
# Private webhook — not client-callable, emits invalidation
@client(affects='subscription', private=True, route='/webhooks/stripe/', methods=['POST'])
def stripe_webhook(request) -> HttpResponse:
event = json.loads(request.body)
process_stripe_event(event)
return HttpResponse(status=200)
# Auth guards
@client(auth=True)
def secret(request) -> dict: ...
@client(auth='staff')
def admin_action(request) -> dict: ...
@client(auth=lambda req: req.user.email.endswith('@company.com'))
def internal_tool(request) -> dict: ...
_meta Dict Structure
After decoration, the function class has _meta with these possible keys:
{
"context": "user", # context name string (if context=)
"affects": [ # normalized affects targets (if affects=)
{"type": "context", "name": "user"},
{"type": "function", "name": "user_profile", "context": "user"},
],
"private": True, # if private=True
"route": "/webhooks/stripe/", # if route=
"methods": ["POST"], # if route= (defaults applied)
"view_path": True, # if return type is HttpResponse
"websocket": True, # if websocket=True
"auth": "required", # "required" | "staff" | "superuser" | callable
"rev": 3, # if rev=
"cache": 60, # if cache=
"form": True, # if form function
"form_name": "contact", # form name
"form_role": "schema", # "schema" | "validate" | "submit"
}
Cache System
Required Settings
# settings.py
MIZAN_CACHE_SECRET = "your-32-byte-hmac-signing-key" # Required for cache
MIZAN_CACHE_REDIS_URL = "redis://localhost:6379/0" # Required for cache
Both must be set. If either is missing, caching is disabled with a warning.
HMAC Key Derivation
Cache keys are derived from HMAC-SHA256 over a JSON-canonical form:
derive_cache_key(secret, context, params, user_id=None, rev=0) -> str
Canonical form (the HMAC message):
{"c":"user","p":{"user_id":"5"},"r":0}
With optional "u":"5" for user-scoped entries.
c= context namep= sorted params dict (all values stringified)r= revision numberu= user ID (for auth-scoped cache entries)
Key format: ctx:{context}:{hmac_hex}
- Example:
ctx:user:605a1ca5ad5994e9b765c8d1b330474c2a0d51a7b8fbbdc402f992da7ba902f6
Cross-language conformance: The TypeScript adapter (mizan-ts/src/cache/keys.ts) produces identical keys for identical inputs. Pin tests verify this.
Cache Operations
from mizan.cache import cache_get, cache_put, cache_purge
# Store
cache_put(secret, backend, "user", {"user_id": "5"}, b'{"name":"Ryth"}')
# Retrieve
data = cache_get(secret, backend, "user", {"user_id": "5"})
# Scoped purge (recomputes HMAC, deletes one key)
cache_purge(backend, "user", params={"user_id": "5"}, secret=secret)
# Broad purge (SCAN by prefix "ctx:user:*")
cache_purge(backend, "user")
Backends
MemoryCache — dict-based, for testing. No persistence.
RedisCache — production backend.
- Connection pooling (50 max connections)
- 24h default TTL safety net
- Key prefix:
mizan:(configurable) delete_by_prefixuses Redis SCAN (1000 keys per batch)deleteuses UNLINK (non-blocking)
Cache Integration in Dispatch
context_fetch_view checks origin-side cache before executing functions. On cache miss, executes functions and stores the result. On mutation, purges affected cache entries based on the invalidation targets.
All HTTP responses emit Cache-Control: no-store. Origin-side caching is internal — the HTTP layer never caches at the CDN. Edge caching is managed by Mizan Edge (closed-source Cloudflare Workers) which uses the manifest and MWT tokens.
MWT (Mizan Web Token) and JWT
Two Token Systems
JWT — standard user authentication tokens. Access + refresh pair. Session-tied for revocation.
# settings.py
JWT_PRIVATE_KEY = "your-secret-key" # Required
JWT_ALGORITHM = "HS256" # Default, or RS256 for asymmetric
JWT_ACCESS_TOKEN_EXPIRES_IN = 300 # 5 minutes
JWT_REFRESH_TOKEN_EXPIRES_IN = 604800 # 7 days
JWT_VALIDATE_SESSION = True # Check session exists on use
JWT claims: sub (user ID), sid (session key), staff, super, type (access/refresh), iat, exp.
Session validation: on every JWT use, checks that the session still exists. Logging out destroys the session → immediately revokes all tokens tied to it.
MWT — Mizan Web Token. Protocol-owned identity for Edge cache keying. Separate secret from JWT and cache.
# settings.py
MIZAN_MWT_SECRET = "your-mwt-signing-key" # Separate from JWT_PRIVATE_KEY
MIZAN_MWT_TTL = 300 # 5 minutes
MWT is used by Mizan Edge to derive user-scoped cache keys without exposing the cache secret to the client. The MWT carries claims that Edge needs (user identity, permissions) in a short-lived token that travels on a custom header (X-Mizan-Token).
Secret Separation
Three independent secrets, each with its own blast radius:
| Secret | Setting | Purpose | Compromise Impact |
|---|---|---|---|
| JWT secret | JWT_PRIVATE_KEY |
User auth tokens | Auth bypass |
| Cache secret | MIZAN_CACHE_SECRET |
HMAC cache keys | Cache poisoning |
| MWT secret | MIZAN_MWT_SECRET |
Edge identity tokens | Cache key spoofing |
SSR Implementation
Django Template Backend
# settings.py
TEMPLATES = [
{
'BACKEND': 'mizan.ssr.MizanTemplates',
'OPTIONS': {
'worker_path': 'frontend/ssr-worker.tsx',
'timeout': 5,
},
},
]
Usage in Views
from django.shortcuts import render
def profile_page(request, user_id):
profile = get_user_profile(user_id)
return render(request, 'ProfilePage', {'profile': profile})
render() calls MizanTemplates.get_template('ProfilePage') which returns a MizanTemplate. The template's render(context) sends JSON-RPC to the Bun worker.
SSR Bridge (bridge.py)
- Spawns
bun run <worker_path>on first render - Persistent subprocess — stays alive across requests
- JSON-RPC over stdin/stdout with message ID correlation
- Thread-safe: multiple Django workers can call
render()concurrently - Auto-restarts on crash
- Waits for
{"id": 0, "ready": true}before accepting requests
Bun Worker (worker.tsx)
- Reads newline-delimited JSON from stdin
- Component registry:
registerComponent('ProfilePage', ProfilePage) - Calls
renderToString(createElement(Component, props)) - Returns
{"id": N, "html": "..."}or{"id": N, "error": "..."} - Health check:
{"method": "ping"}→{"pong": true}
Edge Manifest
Generated by generate_edge_manifest() or python manage.py export_edge_manifest.
{
"contexts": {
"user": {
"functions": [
{"name": "user_profile", "path": "rpc"},
{"name": "profile_page", "path": "view", "route": "/profile/<user_id>/"}
],
"endpoints": ["/api/mizan/ctx/user/"],
"params": ["user_id"],
"user_scoped": true,
"render_strategy": "dynamic_cached",
"page_routes": ["/profile/<user_id>/"]
}
},
"mutations": {
"update_profile": {
"affects": ["user"],
"auto_scoped_params": ["user_id"]
},
"stripe_webhook": {
"affects": ["subscription"],
"private": true,
"route": "/webhooks/stripe/",
"methods": ["POST"]
}
}
}
render_strategy: "psr" (no user-scoped params) or "dynamic_cached" (user-scoped). Derived automatically from whether params overlap with {user_id, user, owner_id, account_id}.
URL Patterns
# mizan/urls.py
urlpatterns = [
path("session/", session_init_view), # GET — CSRF cookie
path("call/", function_call_view), # POST — RPC dispatch
path("ctx/<str:context_name>/", context_fetch_view), # GET — bundled context fetch
]
Mounted at /api/mizan/ by convention:
urlpatterns = [
path("api/mizan/", include("mizan.urls")),
]
Codegen — Current State
The codegen is protocol/mizan-generate/ — framework-agnostic, two-stage. Stage 1 emits the protocol layer (callXxx for mutations, fetchXxx for context bundles, types). Stage 2 emits per-framework hooks/composables/stores that subscribe to the mizan-base kernel.
What's in place:
- Function hooks (
useEcho,useUserProfile, etc.) in the React adapter, subscribing to kernel state viauseSyncExternalStore - Context hooks for named contexts and
global - Channel hooks for WebSocket transport
- Vue and Svelte equivalents (Stage 2 templates compile, but no live-backend example exercises them — see
ISSUES.mdA4)
What's not yet emitted (the wrapper layer):
<MizanContext>provider component for React (callsconfigure()and mounts the kernel into the component tree)useMizan()hook for accessing the kernel from React- Framework-named error class (e.g.
DjangoError) wrappingMizanErrorfrom the kernel - Vue and Svelte equivalents
The legacy MizanProvider in mizan-react/src/context.tsx (~750 lines) and the forms.ts consumer (~1163 lines) still depend on the pre-kernel API. They're tracked as ISSUES.md A1 and A3. Removing them is gated on the wrapper layer above being emitted.
The SSR pipeline is independent of the codegen — it renders whatever components are registered in the Bun worker.