# porch — the writeonce web framework

> **Named `porch` on 2026-08-26.** Rename history: `writeonce-framework`
> (`use framework`) → `writeonce-serve` (`use serve`, 2026-08-25) → **`porch`**
> (`use porch`). Stories, specs, plans and the audit reports dated before each
> change still say the older name — they are dated records and were left as
> written, which is the repo's convention.
>
> Why `porch`: the structure in front of the house you actually enter through,
> and in writeonce the house *is* the database. The name appears only in `use`
> lines and the `[deps]` key — names resolve bare through `use` edges, so no
> handler body mentions it.

A web framework **written in writeonce**, consumed as a `[deps]` dependency
(iteration 15). Roadmap: [`docs/stories/porch/`](../../stories/porch/00-story.md)
— its own track, eight iterations, numbered from 1, derived from
[the Fiber parity study](../../plan/exploration/fiber/00-fiber-parity.md). Spec: `docs/superpowers/specs/2026-08-18-web-framework-design.md` §B.

```toml
[deps]
porch     = { git = "https://github.com/shoneyj/porch", rev = "v0.1.0" }
```

## What it is

- **HTTP/1.1** server core: request parsing (`Content-Length`
  bodies, %-decoded paths and query strings), response serialization, a
  blocking serve loop that answers 400 to malformed requests, 500 to
  trapping handlers (and survives), closes every fd, and honors SIGTERM.
  Connection policy: **pipelined requests are served on one connection;
  idle connections close after the response** — on a single-threaded server
  a parked keep-alive connection would block `accept` and starve every
  other client, so closing is the correct shape until fiber-per-connection
  serving lands (the arc — 8/11 — landed 2026-08-21; the serve-loop slice
  that consumes it is iteration 24's).
  A proxy in front simply reconnects.
- **Router** (`router/`): method + path table with `:param` captures into
  `req.params`; first match wins; a known path with the wrong method is
  **405 with the `Allow` header** (registration order); no matching path is
  the framework's 404. **HEAD is served free**: routed as GET, body
  suppressed, `Content-Length` still names the body a GET would carry.
- **Registration helpers**: `app.get/post/put/delete_(pattern, handler)`
  push the route for you (`delete_` because `delete` is the query
  keyword); `app.add(Route { ... })` stays for anything else. A
  request-line `Logging` middleware ships in `router/`, and
  `set_header(resp, name, value)` is the escape hatch for headers the
  builders don't set.
- **Handlers without closures**: the language has no function values by
  doctrine, so a route handler is a class satisfying the `Handler` interface
  (`fn handle(req: Req) -> Resp`), dispatched structurally — a
  non-conforming handler is a compile error (WO-E205). Middleware is its own
  interface (`fn before(req: Req) -> ?Resp`; nil = continue, a `Resp`
  short-circuits).
- **Auth mechanism in core** (`http/auth.wo`): `Authorization` header
  parsing (scheme split, case-insensitive), pure-`.wo` base64, a
  constant-time comparator (`ct_eq`, no early exit), and the blessed
  principal slot — `req.principal` is `""` until an auth middleware
  authenticates, then downstream handlers read who it is. `BearerAuth`
  and `BasicAuth` (with the `WWW-Authenticate` challenge) ship as
  middlewares; POLICY — which routes, which users, where secrets live —
  stays in the app, on top of `bearer_token`/`basic_credentials`/`ct_eq`.
- **Data layer for free**: handlers use `@table` + the query surface
  directly — durable, compiler-checked persistence in the same binary. No
  ORM, no database server.

## Honest limits (v1, all deliberate)

- **Concurrency is the APP's ten lines** (iteration 35's serving slice):
  the framework ships `serve_conn` — the keep-alive loop with read/idle
  deadlines — and the app owns accept + one spawned ConnWorker actor per
  connection (web-app's pattern; `spawn` takes a class literal, so this
  cannot live in the library). Parallel requests, stalled-client
  eviction and parked idle keep-alive are gate-proven. The plain
  `serve()` stays single-threaded for simple apps.
- **TLS: available in the runtime, not used by this sample yet.** Since
  runtime-v2 9 (2026-09-09) the runtime terminates TLS 1.3 itself —
  `net.accept_tls(listener, cert, key)` (see `docs/examples/tls-server`) — so a
  front proxy is no longer mandatory. This sample still runs plaintext
  HTTP/1.1 keep-alive behind nginx/caddy (which also supplies ALPN/HTTP/2);
  see the web-app sample's README for the nginx sketch. HTTP/2 itself is a
  separate future slice.
- `Content-Length` bodies only (no chunked encoding); **WebSockets ARE
  supported since 2026-08-27** — `ws_accept` (`http/ws.wo`) performs the RFC
  6455 handshake and hands back the hijacked `net.Conn`, and `http/wsframe.wo`
  is a pure-`.wo` frame codec; `docs/examples/chat` is the worked example and
  `just chat` its gate. **SSE is still absent**, and so is chunked encoding.
  JSON-first (no templates). Form-encoded bodies parse through
  `form_values(req)` (`+` and `%XX` decoded, nil on any other
  content-type); multipart/form-data through `multipart_parts(req)`
  (whole-body, bounded by BODY_MAX — the arc landed 2026-08-21;
  streaming uploads stay parked until their own slice) with
  `part_named` for fields; `media_type(req)` names
  the body's media type for content negotiation.

## The v1 surface — status ledger (2026-08-22)

The target surface of **framework v1**, tracked per item. The memory-rich
features (TTL cache, feature flags, durable job queue, `transaction { }`)
are **framework v2** — iteration 18, spec written, NOT part of v1.
Legend: ✅ shipped · 🔶 partial (gap named) · ⬜ candidate slice ·
⏸ parked behind a runtime iteration · 🔧 needs a runtime/compiler seam
first (pure `.wo` cannot express it yet).

### Transport

> Parity reference: [the Fiber v3.5.0 study](../../plan/exploration/fiber/00-fiber-parity.md)
> read all 32 of Fiber's middleware packages against this framework on
> 2026-08-26. **Nine already have a working counterpart here** (CORS, basic
> auth, key/bearer auth, security headers, ETag, static files, logger, host
> authorization, recover-as-500). The rows below marked ⛔/⏸ are what it found
> missing, each with an owner.

| Item | State |
| --- | --- |
| HTTP/1.1 parsing | ✅ parses + 400-and-survive; duplicate `Content-Length` rejected outright (RFC 9112 §6.3, slice 2); BODY_MAX bounds headers and body |
| Keep-alive | ✅ RETIRED close-when-idle (iteration 35's serving slice): under the app-owned fiber-per-connection pattern, idle connections PARK until the idle deadline; the sequential `serve()` keeps the old policy for simple apps |
| Read/write/idle timeouts | ✅ iteration 35: per-call deadlines (`net.read_dl`/`accept_dl`/`write_dl`, nil/false = the expected timeout); `serve_conn(read_ms, idle_ms)` bounds slow-loris AND idle keep-alive |
| Request size limits | ✅ BODY_MAX bounds headers AND body |
| Unix socket binding | ✅ `net.listen_unix(path)` (iteration 35) — stale sockets unlinked before bind, same accept/read/write after |
| Graceful SIGTERM | ✅ in-flight request completes (blocking model), listener + fds closed, storage is per-commit durable (WAL fdatasync — nothing to checkpoint) |

### Routing

| Item | State |
| --- | --- |
| Path matching | 🔶 linear scan, first-match-wins; a radix tree waits on a MEASUREMENT first — 22's harness landed (benched the DB, not the router); needs a perf-targets register entry |
| Method dispatch · path params · 404 · 405+`Allow` | ✅ |
| Wildcards | ✅ `*rest` as the LAST pattern segment captures the joined tail (empty rest matches) — slice 2 |
| Precedence rules | ✅ registration order IS the rule; wildcards capture only in last position, so order stays the whole story |
| Route groups | ✅ `Group { prefix }` + per-group before-middleware, mounted in one move — slice 2 |

### Request/response

| Item | State |
| --- | --- |
| Case-insensitive headers · query parsing | ✅ (names lowercased on read) |
| JSON · form-urlencoded · multipart | ✅ all three hooks (`json.decode`, `form_values`, `multipart_parts`) |
| Content negotiation | ✅ `media_type(req)` request-side; `accepts(req, mtype)` response-side (exact, type/*, */*; q-values stripped not ranked — ranking waits for an app serving alternates) — slice 2 |
| Trusted-proxy client IP | 🔶 `client_ip(req)` parses X-Forwarded-For; `net.peer(fd)` (iteration 35) exposes the peer — the verify middleware is now a pure-`.wo` candidate slice |
| Status/header setting · redirects | ✅ builders + `set_header` |
| WebSockets · pub/sub | ✅ **2026-08-27 (iteration 24)** — `ws_accept` does the RFC 6455 handshake and hands back the hijacked `net.Conn`; `http/wsframe.wo` is a pure-`.wo` frame codec. Rooms/presence/broadcast are actors in `docs/examples/chat`, gated by `just chat` (11 checks, 1000-client soak, both `WO_IO` backends, ASan clean). No SSE |
| Lazy body streaming + backpressure · streaming responses · explicit commit point | ⏸ UNBLOCKED by the arc (8/11 landed 2026-08-21) — stays parked until its own slice |
| ETag + conditional requests | ✅ `etag_for` (quoted base64 SHA-256) + `with_etag` (If-None-Match → 304) over iteration 34's digest builtins — slice 2 |

### Context & middleware

| Item | State |
| --- | --- |
| Ordered middleware chain | ✅ registration order, `?Resp` short-circuits |
| Request-scoped context | ✅ `req.ctx` map (slice 2): middleware writes, handlers read; identity stays in `principal` |
| Guaranteed teardown | 🔶 every fd closes on every path (gate-proven); no user teardown hooks yet |
| Cancellation into pending storage ops | ⏸ **unblocked, not built.** The arc landed 2026-08-21 and iteration 24 (2026-08-27) added the lifecycle a cancellation would ride — `call` with a catchable trap when the callee dies, bounded mailboxes, `monitor`, and `time.after` for a deadline. Nothing here consumes them yet; it stays parked until its own slice |
| Panic recovery | 🔶 trap = 500 and the server survives ✅; "rolls back the transaction" is framework v2 (needs `transaction { }`, iteration 18) |

### Storage integration (the differentiator — framework v2 territory)

| Item | State |
| --- | --- |
| Rate limiting (fixed window, durable) | ✅ counting serializes through a per-key actor pool (`middleware/keypool.wo`, `limiter.wo`) — no handler-fiber read-modify-write left to lose an increment. Gate-proven: exact count under genuine concurrency (30 parallel requests, no lost increments), WAL-durable restart still limiting, and `trust_proxy`'s peer fallback — [porch 1](../../stories/porch/01-store-backed-middleware.md). Saturation failing closed (503) shares `limiter.wo`'s own `try/catch` code path with idempotency's, but is gate-proven only via `Idempotent`/`pool_begin`'s own saturation leg in `scripts/web-app-accept.sh` — no leg drives `Limiter`'s own 503 arm directly |
| Idempotent replay of unsafe requests | ✅ the pool actor runs the route's `Handler` itself (`middleware/idempotent.wo`), so a duplicate blocks in the actor's mailbox until the owner's row commits — no in-flight heuristic, no window where a duplicate can see "nothing yet". Gate-proven: byte-identical replay, digest-mismatch refusal (422), concurrent duplicates never double-executing, a transient 5xx never replayed (solo or concurrent), ephemeral rows not leaking, and pool saturation failing closed (503) — [porch 1](../../stories/porch/01-store-backed-middleware.md) |
| Transaction-per-request middleware (commit on 2xx, roll back otherwise) | ⏸ **v2** — needs iteration 18's `transaction { }` |
| Cancellation → rollback | ⏸ arc landed; still needs v2's `transaction { }` (iteration 18) |
| Migration generation + review workflow | ⬜ recorded future story (script-based destructive migrations) |
| Eager-loading API (N+1) | ⬜ query-surface work (9-series), not framework code |
| Tenant-scoped query roots | ⬜ future; wants the query surface to grow scoped roots first |

Four things anyone wiring the rate limiter or idempotency into a real app
needs to know, found in the course of building them ([porch 1](../../stories/porch/01-store-backed-middleware.md)):

- **`Idempotent` is a `Handler` decorator, not a `Middleware`.** It holds
  `pool` + `inner` and implements `handle`, registered in place of the route's
  own handler (`app.post("/x", Idempotent { ..., inner: RealHandler {} })`),
  not via `app.use_mw`. This was forced, not stylistic: the actor has to be
  handed the route's `Handler` so it can run it inside `receive`, and only the
  handler slot exposes it.
- **`Pool` cannot live in actor state or in a message — `multi PoolSlot` can,
  and that's how a real app shards across MORE than one connection.** `Pool`
  is demand-promoted to "traced" the moment an app aliases it (a
  `Limiter`/`Idempotent`'s own `pool: Pool` field, read on every request) and
  WO-E222 refuses a traced value in actor state or a message. `PoolSlot` (and
  `multi PoolSlot`) never gets pulled into that traced set on its own — an
  actor holding `slots: multi PoolSlot` directly is the same shape
  `docs/examples/chat/main.wo`'s `Room { members: multi Mem }` already uses
  for a multi of actor handles, and it compiles and runs. Call `make_pool(n)`
  **exactly ONCE, at process start** — never per connection, which would give
  every connection its own actors and silently restore the lost-increment
  race this whole design exists to prevent — then hand `pool_slots(pool)`
  (`middleware/keypool.wo`) to every connection actor's spawn. Each
  connection rebuilds a transient `Pool` via `pool_of(self.slots)` wherever a
  `Limiter` or `Idempotent` needs one. Disclosure: the accept gate's own
  `ConnWorker` fixtures (`scripts/web-app-accept.sh`) still build a
  deliberately ONE-slot `Pool { actors: [PoolSlot { a: slot }] }` per leg —
  the limiter/idempotent legs are testing other properties, and the
  saturation leg wants exactly one actor to force mailbox overflow — so no
  gate leg yet exercises `pool_slots`/`pool_of` sharding N actors across
  connections.
- **A `call` reply is a copyable scalar only (WO-E226), and every `receive` in
  the program must agree on one return type.** That is why the stored response
  travels through the `@table` rather than the mailbox, and why outcome codes
  are packed into an `Int` (`pool_pack`/`pool_count`/`pool_begin` in
  `middleware/keypool.wo`).
- **Pool size is a capacity decision made ONCE, not a default to ignore or a
  knob to re-tune per connection.** `make_pool(n)` — called once, per the
  bullet above — spawns `n` actors, sharded by hash of the key; a hot key's
  actor has a bounded mailbox (`WO_MAILBOX`, default 1024), and once it
  saturates under load every further request for that key answers 503 rather
  than being served uncounted or queued indefinitely. Undersizing `n`
  produces more 503s under load — it does not silently let requests through
  uncounted, and it does not silently overshoot the limiter's or idempotency
  store's guarantees.

### Security

| Item | State |
| --- | --- |
| Constant-time comparison · Authorization parsing · Basic auth · principal | ✅ `http/auth.wo`, `req.principal` |
| CORS | ✅ `Cors { allow_origin }` — preflight 204 (before) + origin stamp on every response (after) — slice 2 |
| Security headers | ✅ `SecurityHeaders` after-middleware (nosniff, DENY, referrer-policy); HSTS is set wherever TLS terminates — the front proxy today, or the runtime itself once a sample adopts `net.accept_tls` (runtime-v2 9) — slice 2 |
| Host validation | ✅ `HostAllow { host }` answers 421 before any route — slice 2 |
| Strict parsing | ✅ same item as Transport's row: duplicate Content-Length is a 400 |

### Crypto (self-written, hard-stop after JWT HS256)

| Item | State |
| --- | --- |
| base64 | ✅ pure `.wo` (`http/auth.wo`) |
| SHA-1 · SHA-256 · HMAC-SHA256 | ✅ C runtime builtins (iteration 34, ids 85–87, RFC-vector gated); SHA-512/CRC32 wait for a consumer |
| Unlocks — signed cookies, webhook verification, JWT HS256 **verification** | ⬜ genuinely unblocked (integrity only needs iteration 34's HMAC); each its own slice; **hard stop at JWT HS256** — no RS256, no JOSE zoo |
| Unlocks — CSRF, session integrity, JWT **issuing** | ⛔ **BLOCKED, corrected 2026-08-26.** This row previously read "UNBLOCKED (the primitives exist since iteration 34)" and that was wrong: HMAC lets you *authenticate* a token, not *mint* one, and **writeonce has no source of randomness at all** (no `getrandom`, no CSPRNG builtin — grep the runtime). An HMAC over a guessable session id is a signed guess. A random-bytes builtin is [iteration 39](../../stories/language-runtime-database/39-web-framework-parity.md)'s first goal |
| Cookies (read + `Set-Cookie`) | ⛔ absent in BOTH directions, and `Resp.headers` is a `map<Text,Text>` so it structurally cannot carry two `Set-Cookie` lines — [iteration 39](../../stories/language-runtime-database/39-web-framework-parity.md) |
| Sessions · CSRF · rate limiting · idempotency | ⬜ [iteration 39](../../stories/language-runtime-database/39-web-framework-parity.md). Limiter and idempotency need only a `@table` + `time.ticks` and are the cheapest wins available; sessions and CSRF wait on randomness + cookies. `@table` gives all four a **durable** store, where Fiber ships in-memory and expects Redis |
| Compression · SSE · byte ranges · chunked bodies | ⏸ all four sit on the parked streaming seam (`serialize()` always emits `Content-Length`). Chunked REQUEST bodies are deliberately refused today (`internal/parse.wo:153-157`, request-smuggling note) — that refusal must survive whoever implements them |
| Typed binding of query/params/form/headers | ⏸ Fiber's `Bind` reflects over struct tags; principle 13 forbids reflection, so the answer is [iteration 29's `@derive`](../../stories/language-runtime-database/29-compile-time-metaprogramming.md). JSON bodies already work via `json.decode(t) as T` |
| PATCH/OPTIONS/HEAD/ALL helpers · named routes · per-route body limit · request id · `Location`/`Vary`/`Attachment` | ⬜ [iteration 39](../../stories/language-runtime-database/39-web-framework-parity.md) — registration and response sugar; `BODY_MAX = 1048576` is currently one compile-time number for the whole server |
| `proxy` middleware | ⛔ impossible today — no `net.connect` anywhere in the runtime ([iteration 38](../../stories/language-runtime-database/38-content-platform-capabilities.md)) |

## Layout and privacy (iteration 17)

This project declares `kind = "library"` in `wo.toml`, so `woc <dir>` runs the
FULL pipeline over it — parse, typecheck, interface satisfaction, ownership, GC
inference — with no `fn main` required, and writes nothing. That retired
iteration 16's `woc --emit` verification workaround. `woc build` on it fails
naming the kind, unless a demo `main` is added (lib+bin is allowed).

- `http/` — the public surface: `Req`/`Resp` and the response builders
  (`types.wo`), auth (`auth.wo`), multipart (`multipart.wo`), and the
  body-inspection pair `media_type`/`form_values` (`form.wo`).
- `router/` — the route table and `Logging`.
- `app.wo` — `App`, the registration helpers, the dispatch loop.
- **`internal/` — not importable by a consumer.** The connection-level request
  parser and carry-state record (`parse.wo`) and the serve loop, status text,
  and response serializer (`serve.wo`) live here. A consuming app that writes
  `use porch/internal` gets **WO-E108** at that `use`. The rule
  is Go's: a path segment named `internal` is refused across the `[deps]`
  boundary only — the framework's own modules import it freely.

One honest disclosure: privacy restricts NAMING, not code size. `internal/`
modules still compile into the consumer's single image (there is no dead-code
elimination); a consumer simply cannot name them.

## The consuming sample

`docs/examples/web-app` — a small storefront importing this framework
through `[deps]`. Its acceptance (`just web-app`) exercises the whole chain:
fetch → lock → build → serve → durable restart.

## Serving files

`StaticFiles { dir, max_bytes }` mounts a directory on a wildcard route:

```
app.get("/assets/*path", StaticFiles { dir: "assets", max_bytes: 2097152 })
app.get("/dl/*path",     StaticFiles { dir: "dist",   max_bytes: 16777216 })
```

Two rules, both refusals rather than repairs: a path containing `..` is a
404 and never reaches the filesystem, and `max_bytes` is a hard ceiling —
`fs.read_all` truncates above it, so set it above the largest file you
mean to serve. Content types come from the extension; archives
(`.tar.gz`, `.tgz`, `.zip`) also get `content-disposition: attachment`.
Text is binary-safe in this language, so archives and images travel
unchanged. Lifted out of the shop template 2026-08-25.