Migrating

0.3 to 0.4

0.4 adds raw TCP and UDP servers (TCP and UDP servers), background jobs and cron (Background jobs, Cron schedules), webhooks (Webhooks), SSE topics and watch (Server-sent events and watching changes), gRPC over HTTP/3 and Unix sockets (gRPC), the S3 and SMTP clients (Object storage (S3), Email) and the Lambda deployment (AWS Lambda). They are new names and new build flags (-Dnet, -Djobs, -Dwebhooks, -Ds3, -Dsmtp, -Dlambda); an app that does not use them changes only as the table says.

0.3	0.4
An exhaustive switch on anole.db.Class	add the cases .not_found, .precondition and .invalid
An exhaustive switch on anole.db.System or anole.resilience.Kind	add the cases .s3 and .smtp
An exhaustive switch on anole.ServerError	add the cases NetNotBuilt, NetListenerNeedsPort, NetListenerHttp3, TcpAndUdp, UdpWithTls, UdpOnUnixSocket, UdpWithProxyProtocol, InvalidFraming and NetHandlerContext
A restart (process) with up to 255 HTTP/3 sockets	handoff.max_listeners is 64 TCP listeners (raw TCP ones among them) and 1024 UDP sockets (HTTP/3’s and every raw UDP listener’s); no program that fit before fails
ServerOptions.Listener built field by field	new fields tcp and udp (null); a server that names only raw listeners opens no HTTP listener, where before run() with no http or https always opened 127.0.0.1:8080 (a server that names a listener of any kind never did)
A binding built against C ABI 0.9	rebuild against the new anole.h: the ABI is 0.10 (ANOLE_ABI_COMPATIBLE asks for an exact match before 1.0); Kafka groups and batches are handle kinds 17 and 18, so a NATS handle given to a Kafka call is ANOLE_E_INVALID; new status ANOLE_E_SERVICE (-13); docs/c-abi.md, “From ABI 0.9”
The C ABI’s anole_stats of 0.9	six more u64 fields at the end (net_connections … udp_datagrams_dropped); a caller passing its size gets the fields it knows
anole.http.Client over HTTPS offered only HTTP/1.1	it offers h2 too (ClientOptions.http2 is on by default) and shares one connection per origin and executor; .http2 = false keeps 0.3’s behaviour
channel.conn (read)	channel.conn.? (the field is optional; h3 is the other connection)
A gRPC client call whose stream the server reset was UNAVAILABLE	the status follows the reset code (INTERNAL for most; only a refused stream is UNAVAILABLE, so Channel.retry no longer repeats the others)
Response.abort() in a gRPC handler answered UNKNOWN	the stream is reset; the client reads CANCELLED
error.DeploymentUnavailable for deployment = .lambda	gone: deployment = .lambda serves AWS Lambda invocations (docs/lambda.md); an exhaustive switch on anole.ServerError drops DeploymentUnavailable and adds LambdaRuntimeApiMissing, LambdaNotBuilt, LambdaRuntimeApiFailed and SnapStartUnsupported
A C consumer’s switch over every event type	add ANOLE_EV_SSE_SUBSCRIBE, ANOLE_EV_SSE_CLOSE, ANOLE_EV_WATCH_EVENT, ANOLE_EV_WATCH_STATUS and ANOLE_EV_WATCH_END (a consumer without sse routes or watches never gets them); the new handle kinds are anole_sse and anole_watch_sub
A Response held open by res.adopt for anything but a streamed 200	the same: adopt serves a streamed 200 (an event stream); anole.sse and app.sse park such a response without holding a task
A long-lived HTTP/3 response (a gRPC stream, SSE, WebSocket or WebTransport over HTTP/3) ended at the 30 s idle timeout when nothing moved	QUIC PINGs keep the connection open while a request is open; idle connections still close at the timeout
A handler on HTTP/3 ran to the end after the client’s RESET_STREAM	its task is canceled, as on HTTP/2 (the handler’s next wait returns error.Canceled)
RequestId replaced an ID the engine had already set	it keeps it (under Lambda, the invocation’s request ID)
The startup banner drew the lizard	it draws the word “anole”; the rules for when it shows, banner = false and ANOLE_NO_BANNER=1 are unchanged
addApp(b, .{ .profile = .web }) with no way to change one feature	.features = .{ .postgres = false } turns single features on or off over the preset

0.2 to 0.3

0.3 adds settings, .env files, deployment modes, retries and circuit breakers, the NATS and Kafka clients, validation and typed handlers, JSON Lines logs, OpenAPI documents and a live console. Nothing you set is changed; the table lists what reads or behaves differently.

0.2	0.3
opts.http.?.address (read)	opts.http.?.address orelse "127.0.0.1" (null is the deployment’s default)
listener.port (read)	listener.port orelse 8080
.proxy = .{ .trusted_hops = 1 }	unchanged and still works; .deployment = .behind_proxy (or .proxy = .{ .trusted = &.{"private"} }) is the recommended form: it ignores headers from peers that are not your proxy
"proxy_protocol": false written in JSON to mean “default”	unchanged; leaving it out now means “the deployment’s default” (false unless anole_lb)
X-Forwarded-For "1.2.3.4:5678" gave 0.0.0.0	gives 1.2.3.4
A Debug build read no .env	a Debug build is development and reads .env files from the working directory; ANOLE_ENV=production or .env_files = &.{} turns that off
A Debug build’s 500s said Internal Server Error	they name the error (error_detail, development); .error_detail = false keeps 0.2’s
timeouts.keepalive 60 s	75 s (set .timeouts = .{ .keepalive = .fromSeconds(60) } for the old value)
WebSockets over HTTP/2 (RFC 8441) accepted by default: Firefox opened them on the h2 connection	off: the HTTP/2 engine no longer advertises extended CONNECT, so clients open WebSockets over HTTP/1.1 (38% more video-call users, half the CPU per message, #91); .http2 = .{ .enable_connect_protocol = true } turns it back on. RFC 9220 over HTTP/3 is unchanged
PostgreSQL, Redis and MongoDB calls tried to connect every time while the database was down	a circuit breaker is on by default: after 5 connection failures in a row calls fail at once with 503 (lastError code circuit_open); .breaker = null ("breaker": null in the C ABI) keeps 0.2’s behaviour
An exhaustive switch on anole.db.System	add the cases .nats and .kafka
req.bind(T) on a body that is JSON but does not fit: 400 {"error":"Bad Request"}	422 application/problem+json with errors naming each field; not JSON or no body stays 400. req.bindWith(T, .{ .ignore_unknown_fields = true }) is the old behaviour
app.get(path, ...) with a run-time path	the route methods take comptime path; use app.addRoute(.get, path, handler)
A pattern with a wildcard before the end, a parameter named twice, an empty parameter name, or ?/#	refused: a compile error, or a startup conflict (ANOLE_E_CONFIG through the C ABI)
anole.cached on a POST route	a compile error (the cache keeps GET and HEAD answers only)
Access log JSON: {"ts":1792333306429,"method":...}	the shared record: {"time":"2026-10-23T09:41:07.415Z","level":"info","scope":"access","message":"GET /users/7 200",...,"method":...}; AccessLog.init(.{}) follows the log format (JSON in production), .format = .text keeps text
A route conflict in a test build crashed the test binary	run, listen and anole.testing.client fail with error.RouteConflict; app.routeConflict() has the message
C ABI: the log callback’s msg was the message	with log.format "json" (the default when environment is production, which is what an unset environment gets) it is a whole JSON record; "log": {"format": "text"} keeps messages. Access log lines now reach the callback (scope access); "access_log".format defaults to "auto"

C ABI JSON config (ABI 0.9)

The C ABI moves to 0.9; rebuild against the new anole.h (ANOLE_ABI_COMPATIBLE asks for an exact match before 1.0).

0.8	0.9
listeners[].address, port, proxy_protocol defaulted to "127.0.0.1", 8080, false	default null: the deployment’s value (the same three under standalone)
timeouts.keepalive_ms 60000	75000
http2.enable_connect_protocol true	false ("http2": {"enable_connect_protocol": true} for 0.8’s behaviour)
proxy_protocol_from empty: any peer	empty: proxy.trusted when set, else any peer
the log callback’s msg was the message	a whole JSON record when log.format resolves to "json" (an unset environment is production); "log": {"format": "text"} keeps messages; access lines reach the callback (scope access)
a malformed route pattern was accepted	ANOLE_E_CONFIG naming the route
PostgreSQL, Redis and MongoDB entries tried to connect on every call while the database was down	a breaker is on by default (503, error_code circuit_open); "breaker": null keeps 0.8’s behaviour
(none)	environment, deployment, error_detail, proxy.trusted, log (format, service), banner, console, brokers (nats, kafka); databases[].breaker, retry; anole_dotenv_read, anole_nats_*, anole_kafka_*, anole_console_*, ANOLE_E_BROKER, ANOLE_DB_F_RETRY

0.1 to 0.2

0.2 batches the breaking changes of the API review. Each area has its section below; the changelog lists the rest. Before 1.0 there is no stability promise, but every change here is a rename or a move, not a redesign, and the compiler names each one.

Server options, TLS and errors

ServerOptions

0.1	0.2
.server = .{ .timeout = .{ .request = d } }	.timeouts = .{ .request = d }
.server.timeout.read_header	.timeouts.read_header
(none)	.timeouts.write (30 s, new)
.server.timeout.keepalive	.timeouts.keepalive
.server.timeout.idle_park	.timeouts.idle_park
.server.timeout.shutdown (30 s)	.timeouts.shutdown (10 s)
.server.timeout.request_count	.limits.max_requests_per_connection
.server.request.max_body_size	.limits.body_max
.server.request.max_content_length	.limits.body_max (one limit for both)
.server.request.buffer_size	.limits.max_head_size
.server.request.max_header_count	.limits.max_header_count
.server.request.max_param_count / max_query_count / max_form_count / max_multiform_count	.limits.<same>
.server.request.require_host	.request.require_host
.server.request.decompress	.request.decompress
.server.request.keep_buffers_between_requests	.request.keep_buffers
.server.response.*	.response.* (same names)
.server.max_connections (null: no cap)	.limits.max_connections (0: the open-file limit; no “no cap”)
.server.trusted_proxy_hops	.proxy.trusted_hops
.server.trust_forwarded_proto / trust_forwarded_host	.proxy.trust_forwarded_proto / trust_forwarded_host
.server.errors, .server.error_handler, .server.http2, .server.listeners (ignored before)	.errors, .error_handler, .http2, .listeners
.restart_timeout_ms = 10_000	.restart_timeout = .fromSeconds(10)
.client = http.ClientConfig{...}	.client = anole.ClientOptions{...} (same fields; tls is TlsOptions)
.https.alpn_protocols, Listener.Tls.alpn_protocols	.alpn
.client_auth = .{ .ca = .{ .file = .{ .path = "ca.pem" } } }	.client_auth = .{ .ca = .fromFile("ca.pem") }
anole.http.ServerConfig.Tls.ClientAuth	anole.ClientAuth
anole.http.TlsCa	anole.Ca (.file/.dir take a path string; no dir handle)
TlsConfig.ClientCertificate{ .cert_path, .key_path }	anole.Cert.fromFiles(cert, key) (or .fromPem)
http.ClientConfig, http.TlsConfig	anole.ClientOptions, http.TlsOptions

A file-based CA or certificate used to be resolved against a dir handle; paths are now relative to the working directory, so a test that wrote its certificates into a temporary directory passes the path of the files there. A .dev_self_signed client certificate is not supported by the client types (error.DevCertificateUnavailable): make one with anole.tls.selfSigned and pass it with .fromPem.

Requests and responses

0.1	0.2
req.setTimeout(.{ .duration = .{ .raw = .fromSeconds(60), .clock = .awake } })	req.setTimeout(.fromSeconds(60))
req.setTimeout(.none)	req.setTimeout(null) (rarely needed: streams lift the deadline themselves)
req.setTimeout(.{ .deadline = t })	req.setDeadline(.{ .deadline = t })
res.writer()	try res.writer()
res.headers_written	res.headersSent()
res.conn, req.parser, req.transport, …	internal (_-prefixed); not for handlers
Response.init, Request.reset	Response._init, Request._reset (engines only)
chunked req.reader() unbounded	bounded by limits.body_max; req.readerLimited(buf, n) for more

Middleware and helpers

0.1	0.2
fn execute(self, req, res, executor: anytype) !void { return executor.next(); }	fn execute(self, req, res, next: anytype) !void { return next.call(); }
Compress.init()	Compress.init(.{})
Logger.init(), Logger.initWith(.{ ... })	Logger.init(.{ ... })
Cors.Config, RequestId.Config, Logger.Config, session.Config, csrf.Config	...Options
session.ConfigError	session.OptionsError
MemoryStore.init(gpa, io), initOptions(gpa, io, o)	MemoryStore.init(gpa, io, .{ ... })
session ttl_seconds	ttl (Duration)
ratelimit window_ms, idle_ms	window, idle
auth *_seconds (leeway_seconds, ttl_seconds, max_age_seconds, …)	same names without _seconds (Duration)
Cors max_age = 600	max_age = .fromSeconds(600)
TracingOptions.interval_ms	interval
tls.SelfSignedOptions.valid_for_seconds	valid_for
MemoryLimiter + ratelimit.Middleware.init(...)	still works; or app.rateLimit(.{ .limit = 100 })

Errors

0.1	0.2
init/listen/run return inferred anyerror sets	anole.ServerError (named; switch on it)
reloadCertificates/setCertificate	anole.TlsLoadError
out-of-order request/response use: undefined behaviour in ReleaseFast	anole.http.MisuseError, answered 500

Behaviour you may notice

• The drain waits 10 s (was 30 s). Docker’s default stop grace is also 10 s: give containers stop_grace_period: 15s.
• Event streams, streamed responses and downloads are no longer cut at 30 s; a client that stops reading is cut after timeouts.write (30 s).
• WebSocket messages over 1 MiB are refused by default (was 16 MiB).
• listen("3000") works (every interface).
• Setting one http2 field keeps 250 streams.
• app.tracing() honours OTEL_* without .process.
• Outbound requests say User-Agent: anole/<version>.

C ABI JSON config (ABI 0.7)

0.6	0.7
trusted_proxy_hops	proxy.trusted_hops
trust_forwarded_proto, trust_forwarded_host	proxy.trust_forwarded_proto, proxy.trust_forwarded_host
timeouts.*_ms: integers	integers or null (none)
middleware[].jwt.leeway_s	leeway_ms
middleware[].cors.max_age	max_age_ms
(none)	timeouts.read_header_ms, write_ms, idle_park_ms; limits.max_head_size, max_header_count, max_param_count, max_query_count, max_form_count, max_multiform_count, max_requests_per_connection; proxy, request, response, http2, errors; http3.address, port, validate_address, prefer_v2, sockets, enable_connect_protocol, inline_no_wait; tls.client_ca_dir, client_ca_system

HTTP, WebSocket and WebTransport

Middleware and noWait

A route marked noWait now takes only middleware that says it never waits. Each startup panic names the route and the middleware:

• remove noWait from the route; or
• add pub const no_wait = true; to a struct middleware of yours that truly never waits (or pub fn noWait(self: *const T) bool when it depends on its settings); a function middleware is registered app.use(anole.noWait(f)).

The built-in middleware is marked (see docs/routing.md). Interfaces gained a flag your own implementations may set: ratelimit.Limiter.never_waits, auth.apikey.Lookup.never_waits, session.Store.never_waits (the memory implementations set it; a Redis or database one leaves it off, and a noWait route behind that middleware is refused).

WebSocket and WebTransport

0.1	0.2
ws.conn == .h2, ws.conn.h1	ws.httpVersion() != .http1 (.http1, .http2, .http3)
ws.closed.load(.acquire)	ws.isClosed()
Rooms.init(a), Rooms.initOptions(a, o)	Rooms.init(a, .{}), Rooms.init(a, o)
res.upgradeWebSocket(req)	anole.websocket.upgrade(req, res, .{}) (all three HTTP versions)
session.start(...), tryAccept, tryOpen, takeDatagram, freeDatagram, trySendDatagram, generation, waitChange, wake	Session.Driver.<same>(session, ...) (Anole’s own drivers; not for handlers)
stream.tryRead, tryWrite, sendError, Stream.Poll	Stream.Driver.<same>, Stream.Driver.Poll
req.params.get("name") for an encoded segment	still raw; try req.paramDecoded("name") decodes

New and additive: UpgradeOptions.subprotocols and .headers, websocket.isUpgrade(req), ws.subprotocol, ws.httpVersion(), ws.isClosed().

C ABI 0.7: ANOLE_EV_WS_UPGRADE and anole_ws_accept are opt-in ("accept": "consumer" on a websocket route); an existing config sees nothing new except the subprotocol span in request records (empty outside ANOLE_EV_WS_OPEN). Publishes to one topic are now ordered.

The C ABI (0.6 to 0.7)

The C ABI moves to version 0.7. Before 1.0 every minor may break, so a library serves only code built against the same version: check ANOLE_ABI_COMPATIBLE(anole_abi_version()) at load, and replace any major == / minor >= comparison with it.

Rename the calls and constants (a search and replace is enough for most):

0.6	0.7
anole_respond(r, ...)	anole_req_respond(r, ...)
anole_respond_fd(r, ..., int fd, ...)	anole_req_respond_fd(r, ..., intptr_t fd, ...)
anole_stream_start/write/write_wait/end	anole_req_stream_start/write/write_wait/end
anole_release(r)	anole_req_fail(r, 0, (anole_str){0}) (500) or anole_handle_release(r) (finalizers)
(none)	anole_req_abort(r)
anole_wt_stream_release(s)	anole_handle_release(s)
anole_server_port(s, i, &port)	anole_server_listener(s, i, &info); info.port, info.udp_port
anole_server_reload_certificates(s)	anole_server_reload_tls(s, 0, (anole_str){0}, (anole_str){0})
anole_server_reload_tls(s, f, c, cl, k, kl)	anole_server_reload_tls(s, f, (anole_str){c, cl}, (anole_str){k, kl})
anole_wt_send_datagram(w, d, n)	anole_wt_send_datagram(w, d, n, 0)
ANOLE_EV_WS_DRAIN	ANOLE_EV_WRITABLE (ev.handle the socket)
ANOLE_EV_ABORT read as anole_event	anole_abort_event (reason, status)
anole_request_spans(r) = r + 1	the same helper, now reading r->spans_off
anole_request_peer_cert(r)	anole_event_str(r, r->peer_cert)
gRPC metadata at c + 1	anole_grpc_metadata(c) (c->metadata_off)
ANOLE_DB_INT…	ANOLE_DB_TYPE_INT…; ANOLE_DB_WRITE… are ANOLE_DB_F_*
ANOLE_GRPC_OK…	ANOLE_GRPC_STATUS_OK…; ANOLE_GRPC_UNARY… are ANOLE_GRPC_KIND_*; ANOLE_GRPC_SEND_FINISH is ANOLE_GRPC_F_FINISH
ANOLE_WT_UNI, ANOLE_WT_FIN	ANOLE_WT_F_UNI, ANOLE_WT_F_FIN (values kept)
ANOLE_TLS_PEM, ANOLE_SERVER_SIGNALS, ANOLE_CONSUMER_SHARED, ANOLE_POLL_NOARM	ANOLE_TLS_F_PEM, ANOLE_SERVER_F_SIGNALS, ANOLE_CONSUMER_F_SHARED, ANOLE_POLL_F_NOARM
enum-valued parameters as int (anole_ws_send’s type, the log callback’s level)	uint32_t; anole_consumer_fd returns intptr_t; anole_has_feature returns int32_t

Behaviour that changes with the rename:

• Cleanup calls are idempotent. anole_handle_release, anole_req_abort, anole_ws_close, anole_ws_unsubscribe, anole_wt_close, anole_wt_stream_reset and anole_wt_stream_stop answer ANOLE_OK on an object that is already over, set no error and are not counted in anole_stats.gone. Code that treated ANOLE_E_GONE from a close or release as harmless keeps working (it now gets OK). Answer and data calls still answer ANOLE_E_GONE.
• A handle of the wrong kind, or 0, is ANOLE_E_INVALID (token WrongHandleKind), not ANOLE_E_GONE. Code that relied on GONE to detect a wrong-kind handle must check ANOLE_E_INVALID.
• Options structs and flags are checked. An options struct larger than the library knows is refused unless the extra bytes are 0; unknown flag bits (anole_db_options.flags included) and non-zero reserved fields (anole_db_value.reserved, anole_db_options.reserved) are ANOLE_E_INVALID.
• ANOLE_E_TOOBIG from a poll writes an ANOLE_EV_NONE header with the needed len; allocate that once instead of doubling.
• Records: read the header spans through spans_off and the gRPC metadata through metadata_off, never sizeof; the client certificate is a span (peer_cert), no longer a length-prefixed blob.
• anole_ws_publish on a server that is not running is ANOLE_E_STATE (was ANOLE_E_GONE); a failed dup in anole_req_respond_fd is ANOLE_E_SYSTEM (was ANOLE_E_INVALID).
• Generate, do not hand-write: include/anole.abi.json (and include/anole.schema.json for the configuration) describe every record, constant, event and handle kind; generate your parsers from them, pinned to the Anole tag.

New in 0.7 (see docs/c-abi.md): anole_server_open, anole_server_set_ready, anole_server_listener(_count), anole_last_error_name, anole_last_errno, anole_consumer_set_wake, the allocation table, abort reasons, and a destroy that does not wait for answers only its caller could give.

Databases and gRPC

Options (every connector)

• Every call is bounded by 30 s. Options.timeout exists on all four connectors (it was MongoDB’s alone, and off) and defaults to 30 s; TxOptions.timeout = null now means the connector’s too. A call that must run longer (a migration, a background job) passes .with(.{ .timeout = .max }), or set timeout = null on the connector. C ABI: anole_db_options.timeout_ms = 0 now means the database’s timeout_ms (30000), not “none”; use ANOLE_DB_TIMEOUT_NONE.
• One vocabulary, renamed fields:

0.1	0.2
sqlite.Options.busy_timeout_ms = 5000	busy_timeout = .fromSeconds(5)
sqlite.Options.statement_cache = 64	statement_cache_size = 64 (u16)
sqlite.Options.readers: ?usize	?u32
mongodb.Options.max_idle	idle_timeout
mongodb.Options.max_connections: usize = 100 (0: no limit)	?u32, null for 100; 0 is refused
mongodb.Options.connect_timeout: Duration	?Duration (null: none)
mongodb.Options.tls: ?net.ConnectOptions.Tls	tls: anole.db.Tls (ca, client_certificate, insecure_skip_verify)
pg.Options.tls, redis.Options.tls	the same anole.db.Tls (old paths still compile)
.tls = .{ .ca = .{ .file = .{ .path = "ca.pem" } } }	.tls = .{ .ca = .{ .file = "ca.pem" } } (anole.Ca)
name: []const u8 = "postgres"	name: ?[]const u8 = null (the same default label)
sqlite.Client.execScript(sql)	execScript(&ctx, sql); pg.Client.execScript(&ctx, sql) is new (both unbounded)
pg/sqlite Query.run(req, sql, values), Tx.run	execValues (table keeps its name)
anole.mongodb.dynamic.run(collection, database, ctx, op, docs)	still there; app.mongodb.with(opts).dynamic(ctx, op, "users", docs) takes the context first
max_connections = 0	refused at init with error.InvalidOptions on PostgreSQL, Redis and MongoDB

New fields: connect_at_start and ping_timeout on MongoDB, ping_timeout on SQLite. Conn.release takes *const Conn on every connector (so const conn = try app.redis.acquire(req); defer conn.release(); compiles).

• Leaving a transaction open is caught on more paths. iterate that leaves one, and a connection from acquire given back inside one, are rolled back and logged, as plain calls already were. Tx.table, Tx.execValues and Tx.with(options) exist on both SQL connectors; a finished SQLite Tx fails with error.Database instead of asserting.
• Role switches inside functions under .auto. PostgreSQL does not report role, so reset_session = .auto does not catch SET ROLE made inside a function; use .always for apps that switch roles for row-level security, or SET LOCAL ROLE inside a transaction.

C ABI: databases

• anole_db_options is 48 bytes (req at the end) and anole_db_result_event 104 (table span at the end); read both through size/len as before.
• New: anole_db_begin, anole_db_tx_query, _exec, _commit, _rollback, _release (handle kind db_tx), ANOLE_DB_TIMEOUT_NONE, ANOLE_DB_TX_READ_ONLY, ANOLE_DB_ISOLATION_*; anole_db_query and anole_db_exec accept a NULL consumer with anole_db_options.req.
• JSON config: databases[].timeout_ms (30000), errors, client certificates (client_cert_file, client_key_file, client_cert_pem, client_key_pem), insecure_skip_verify and ping_timeout_ms for PostgreSQL, Redis and MongoDB; MongoDB idle_timeout_ms, heartbeat_ms, retry_reads, retry_writes, connect_at_start; SQLite uri, ping_timeout_ms. sqlite.statement_cache is statement_cache_size, and mongodb.max_connections 0 is refused.
• A libanole built with Redis or MongoDB but neither SQLite nor PostgreSQL answers anole_db_query (it used to refuse every call).

gRPC

• Call.addHeader and Call.addTrailer refuse the protocols’ own keys (grpc-* but grpc-status-details-bin in trailers, connect-*, trailer-*, content-type, te…): a handler could send a second grpc-status. addHeader adds a value instead of replacing one. The C ABI also refuses grpc-status-details-bin as initial metadata.
• Call.fail, addTrailer and addHeader copy their strings; Call.failFmt(code, fmt, args) is new.
• Group.grpc and Group.grpcWith register a service behind a group’s middleware, for its gRPC calls and its transcoded routes alike (before, a transcoded /v1/... route skipped a /v1 group’s auth middleware). C ABI: grpc.services[].middleware.