alknet/docs/research/alknet-crate-extraction/findings.md

alknet-crate-extraction — Migration from the welded core to the extracted crates

Status: Findings in progress — mapping the existing code to the target shape, phase by phase, so the migration can be ordered to keep the tree compilable at each step. Date: 2026-07-18 Scope: The three new crates (alknet-tls, alknet-endpoint, alknet-client) + the prune of alknet-core and alknet-call + the core stream unification (BiStream as the handler leaf) + the TTY control-channel fix + the channels spec cleanup + the alknet-http residual fix. The specs are confirmed tight (reviewed + amended); this doc is the how — what code moves where, in what order, with what intermediate states.

---

TL;DR

The migration is additive-then-subtractive: build all three new crates first (no breakage), then prune the old code from core and call (breakage confined to a single phase), then fix the core stream model (BiStream as the handler leaf), then fix the TTY control-channel bidirectionality, then clean up the channels spec, then fix the http residual. Ten phases (0-9), each leaving the workspace compilable. The heaviest single file (endpoint.rs, 1606 lines) is not a monolith — it's three concerns welded together, each going to a different destination.

---

What exists now (the source map)

crates/alknet-core/src/endpoint.rs — 1606 lines, three concerns

The file is not 1600 lines of one thing. It's three concerns that the extraction separates:

Lines	Concern	Destination	LOC
1-492	AlknetEndpoint struct, HandlerRegistry, dispatch_quinn/dispatch_iroh, accept loops, build_auth_context, has_iroh_identity	alknet-endpoint	~492
493-934	TlsSetup, build_rustls_server_config, build_quinn_server_config_from_rustls, RawKeyCertResolver, Ed25519SigningKey, AcceptAnyCertVerifier, SelfSignedCert, generate_self_signed_cert, load_cert_chain, load_private_key, build_iroh_endpoint	alknet-tls (TLS setup) / assembly layer (iroh builder)	~442
935-1606	42 test functions	split by subject	~671

The endpoint struct + dispatch + accept loops (the part that goes to alknet-endpoint) is ~492 lines of implementation — not 1600. The TLS setup code (~442 lines) goes to alknet-tls. The tests (~671 lines) split by what they test. The file feels monolithic because the #[cfg(feature = "quinn")] gates weave the three concerns together, but the extraction unwinds that.

crates/alknet-call/src/client/call_client.rs — 930 lines

Lines	Concern	Destination	LOC
40-88	RemoteIdentity, CallCredentials (struct + builder)	split: RemoteIdentity + new ConnectionCredentials → alknet-core (credentials.rs); CallCredentials removed (its auth_token field had no reader — see Phase 5)	~48
90-100	ClientError enum	removed (only produced by connect)	~10
102-187	CallClient struct + new + spawn_dispatch	stays (the pure protocol take-over)	~85
189-320	build_quinn_client_config, build_client_auth, select_server_verifier, load_platform_root_cert_store, load_cert_chain, load_private_key, Ed25519SigningKey, RawKeyClientCertResolver, NoClientCertResolver, FingerprintPinVerifier	alknet-tls	~130
321-640	CallConnection, Dispatcher wiring, wire-protocol helpers	stays (protocol)	~320
640-930	Tests (16 total — see Phase 5 audit)	split: 10 TLS/verifier tests → alknet-tls (Phase 1); 4 protocol-level tests stay in alknet-call unchanged; 2 CallCredentials-field tests → alknet-core (testing ConnectionCredentials); 0 lib tests call connect()	~290

The call crate's prune is ~140 lines of implementation + CallCredentials removal + from_call's credentials_auth_token dead-path removal. The test work: 10 tests move to alknet-tls (Phase 1), 2 CallCredentials tests move to alknet-core (testing ConnectionCredentials), 4 protocol-level tests stay unchanged. The integration test (two_node_call.rs, 2 tests) splits: the dial+takeover composition test moves to alknet-client/tests/ (Phase 3, rewritten with a minimal echo ProtocolHandler); the from_call test stays in alknet-call (Phase 5, rewritten to use spawn_dispatch + loopback Connection). What remains is the pure protocol: CallClient + CallConnection + Dispatcher + the wire protocol.

crates/alknet-http/src/server/adapter.rs — the residual

The HttpAdapter::handle method does connection.accept_bi() → wraps the send/recv pair in a hand-rolled QuicStream (lines 271-300, an AsyncRead+AsyncWrite adapter) → feeds it to serve_io(). But serve_io already accepts any AsyncRead+AsyncWrite (line 244). The QuicStream wrapper is a hand-rolled version of what Connection::from_bidi/from_stream (ADR-065, already landed) provides natively.

The residual: handle assumes a quinn-style multi-stream connection (accept_bi returns a fresh bidi stream). For a from_bidi connection (TCP+TLS), accept_bi yields the single bidi stream once (ADR-070's yield-once contract). The QuicStream wrapper works but is unnecessary — the streams from accept_bi are already AsyncRead+AsyncWrite. The fix is to drop QuicStream and use the streams directly (or via TokioIo::new). This is a small change (~30 lines removed) but needs verification against ADR-070's accept_bi semantics for from_bidi.

The QuicStreamDuplex test helper (lines 456-471) is the test-side equivalent — it can be replaced with Connection::from_stream test helpers (the MockConnection/ConnectionKind::Mock variants were already removed per ADR-065; tests use from_stream with tokio::io::sink/empty).

---

The seven phases

Phase 0: Move ConnectionCredentials/RemoteIdentity to alknet-core (additive, no breakage)

What: Add ConnectionCredentials + RemoteIdentity to a new crates/alknet-core/src/credentials.rs (~40 lines). ConnectionCredentials is the transport-level credential bundle (ADR-091) — it carries local_identity + remote_identity (the two dimensions the dial consumes). CallCredentials is removed (its auth_token field had no reader — connect() read only tls_identity + remote_identity; spawn_dispatch takes no credentials; the from_call forwarding path's auth_token source was a different, always-None field never connected to CallCredentials). auth_token is a per-request payload field, not a call-protocol credential. Update alknet-core/src/lib.rs to pub mod credentials + re-export. Update alknet-call to import ConnectionCredentials + RemoteIdentity from core and re-export them; remove CallCredentials and its builder methods. No other changes.

Why first: It's independent of the three new crates, purely additive (core gains types, nothing breaks), and means alknet-client (Phase 3) never has a temporary dep on alknet-call. The dep graph is clean from the start. ConnectionCredentials (not CallCredentials) is what moves — the dial consumes transport-level dimensions, not call-protocol dimensions (ADR-091).

Compilable state: cargo test passes across the workspace. alknet-call imports the types from core; its own code + tests continue to work via the re-export.

Done when: cargo test passes, ConnectionCredentials + RemoteIdentity are defined in alknet-core, alknet-call imports them from core, CallCredentials is removed from alknet-call.

Phase 1: Create alknet-tls (greenfield, additive)

What: New crate crates/alknet-tls/. Extract the TLS setup code from alknet-core/endpoint.rs (lines 493-934) + the client-side TLS helpers from alknet-call/client/call_client.rs (lines 189-320).

Types: TlsServerConfig, TlsClientConfig, TlsError, FingerprintPinVerifier, RawKeyCertResolver, RawKeyClientCertResolver, NoClientCertResolver, Ed25519SigningKey (consolidated — one copy, used by both server + client), AcceptAnyCertVerifier, SelfSignedCert, generate_self_signed_cert, load_cert_chain, load_private_key, load_platform_root_cert_store (+ the webpki-roots fallback — new, not extracted).

Deps: alknet-core (TlsIdentity, Ed25519SecretKey, fingerprint), rustls, rustls-pemfile, rustls-native-certs, webpki-roots, rcgen, tokio, optional quinn/tokio-rustls/rustls-acme. rustls-native-certs and webpki-roots are always-present (not feature-gated) — the unknown-X.509-remote CA-verification path in TlsClientConfig::new is transport-agnostic; the webpki-roots fallback merges built-in roots when the platform store is empty so NoRootAnchors is unreachable in practice (ADR-088 §5). Do not gate them under quinn/tcp.

Compilable state: alknet-tls builds and tests standalone. Core and call are unchanged — the old code still exists (duplicated). No breakage. The new crate's tests are the moved tests from endpoint.rs (the TLS-setup tests) + the moved tests from call_client.rs (the verifier/client-config tests).

Tests that move here (from endpoint.rs):

• raw_key_cert_resolver_only_raw_public_keys
• self_signed_cert_generation_produces_cert_and_key
• acme_directory_production_url / staging_url / custom_url
• tls_setup_x509_returns_no_acme_state
• build_rustls_server_config_raw_key_succeeds
• build_rustls_server_config_self_signed_succeeds
• build_quinn_server_config_from_rustls_succeeds
• load_private_key_returns_error_when_no_key_present / _file_missing
• load_cert_chain_returns_error_when_file_missing
• accept_any_cert_verifier_* (4 tests)
• ed25519_signing_key_* (6 tests)
• raw_key_cert_resolver_debug_is_implemented

Tests that move here (from call_client.rs): the FingerprintPinVerifier + build_quinn_client_config tests (need adaptation — they currently test via connect, should test via TlsClientConfig::new directly).

Done when: cargo test -p alknet-tls passes, the crate is self-contained, no other crate changed.

Phase 2: Create alknet-endpoint (greenfield, additive)

What: New crate crates/alknet-endpoint/. Extract the endpoint struct + dispatch + accept loops from alknet-core/endpoint.rs (lines 1-492), built fresh against the ADR-083 shape (new(handlers, dynamic, identity_provider, drain_timeout) + with_quinn/with_iroh/with_tcp_tls

• public dispatch + run/shutdown). No EndpointError (removed per the spec review). shutdown() is infallible.

Types: AlknetEndpoint, HandlerRegistry, TcpTlsListener, private dispatch_quinn/dispatch_iroh/dispatch_tcp_tls, build_auth_context, the extract_* helpers.

Deps: alknet-core (Connection, ProtocolHandler, AuthContext, IdentityProvider, DynamicConfig), optional quinn/iroh/tokio-rustls, tokio, arc-swap, tracing.

Module breakdown (the 492 lines split):

• registry.rs — HandlerRegistry (~40 lines + tests)
• endpoint.rs — AlknetEndpoint struct, new, builder methods, run, shutdown (~120 lines)
• dispatch.rs — dispatch (public), build_auth_context, the acme-tls/1 guard (~80 lines)
• accept/quinn.rs — dispatch_quinn, run_quinn_accept_loop, extract_quinn_alpn, extract_quinn_client_fingerprint (~80 lines)
• accept/iroh.rs — dispatch_iroh, run_iroh_accept_loop, extract_iroh_client_fingerprint (~50 lines)
• accept/tcp_tls.rs — dispatch_tcp_tls, run_tcp_tls_accept_loop, extract_tcp_tls_alpn, extract_tcp_tls_client_fingerprint (new, ~60 lines — not in the current code; written fresh per ADR-083)

This is a fresh build, not a move — the old endpoint.rs stays in core until Phase 4. The new crate is written against the target shape, not the old shape.

Compilable state: alknet-endpoint builds and tests standalone. Core's endpoint.rs still exists (duplicate). No breakage.

Tests that move here (from endpoint.rs):

• handler_registry_* (5 tests)
• build_auth_context_* (3 tests)
• dispatch_decision_logic_lookup_and_auth
• has_iroh_identity_* (3 tests)
• endpoint_constructs_with_iroh_raw_key_identity
• iroh_endpoint_runs_accept_loop_and_shutdown
• debug_for_alknet_endpoint_is_implemented_without_panicking

Done when: cargo test -p alknet-endpoint passes, the crate is self-contained, no other crate changed.

Phase 3: Create alknet-client (greenfield, additive)

What: New crate crates/alknet-client/. The AlknetClient dial seam — three dial methods (dial_quic/dial_tcp_tls/dial_iroh), all unified on &ConnectionCredentials (ADR-091), consuming TlsClientConfig from alknet-tls + ConnectionCredentials from alknet-core. The SOCKS5 proxy path (ADR-090) is feature-gated.

Types: AlknetClient, ClientDialError, Socks5ProxyConfig, Socks5Credentials (behind socks5 feature).

Deps: alknet-core (Connection, ConnectionCredentials, RemoteIdentity, Ed25519SecretKey), alknet-tls (TlsClientConfig), optional quinn/tokio-rustls/iroh/fast-socks5.

Feature gates: quinn = ["dep:quinn", "alknet-tls/quinn", "alknet-core/quinn"], tcp = ["dep:tokio-rustls", "alknet-tls/tcp"], iroh = ["dep:iroh", "alknet-core/iroh"], socks5 = ["dep:fast-socks5"].

Compilable state: alknet-client builds and tests standalone. No breakage — the old CallClient::connect still exists in alknet-call (duplicate dial). The new crate's tests use AlknetClient::dial_* + spawn_dispatch/from_connection or test the dial in isolation with mock transports.

Done when: cargo test -p alknet-client passes, the crate is self-contained, no other crate changed.

Phase 4: Prune alknet-core (subtractive, breakage confined)

What: Delete endpoint.rs from core. Remove pub mod endpoint from lib.rs. Remove the heavy deps (quinn, iroh, rcgen, rustls-pemfile, rustls-acme) from Cargo.toml — but keep the quinn/iroh features (they gate Connection::from_quinn/ from_iroh in types.rs). Add ConnectionCredentials + RemoteIdentity to core (from alknet-call — done in Phase 0; CallCredentials stays in alknet-call per ADR-091).

The lib.rs change:

// Before:
pub mod endpoint;  // ← removed
pub mod auth;
pub mod config;
// ... rest unchanged

// After:
pub mod auth;
pub mod config;
pub mod credentials;  // ← new (ConnectionCredentials, RemoteIdentity)
// ... rest unchanged

The Cargo.toml change: quinn/iroh/rcgen/rustls-pemfile/ rustls-acme leave [dependencies]. The quinn/iroh features stay (gating types.rs constructors). The acme feature is removed (vestigial — the ACME state machine is on TlsServerConfig in alknet-tls now).

The fingerprint.rs doc comment (line 13) references alknet_core::endpoint — update to reference alknet-endpoint / alknet-tls.

Breakage: anything that imported alknet_core::endpoint breaks. But per the source map, nothing does — no handler crate or call imports the endpoint module. The only consumer is the future assembly layer (hub/worker), which doesn't exist yet. So the prune is clean: delete the file, update lib.rs, update Cargo.toml, fix the fingerprint.rs comment. Core's own tests for endpoint.rs are gone (moved to alknet-tls in Phase 1 and alknet-endpoint in Phase 2).

Compilable state: cargo test -p alknet-core passes (minus the endpoint tests, which moved). The quinn/iroh features still work (Connection::from_quinn/from_iroh in types.rs).

Done when: cargo test -p alknet-core passes, the crate is lightweight (~3200 LOC, no heavy transport deps).

Phase 5: Prune alknet-call (subtractive, breakage confined)

What: Delete connect() + all TLS helpers + ClientError + CallCredentials from call_client.rs. The transport dimensions (ConnectionCredentials/RemoteIdentity) already moved to core in Phase 0; here we remove the old definitions from call_client.rs and update imports. CallCredentials is removed (its auth_token field had no reader — connect() read only tls_identity + remote_identity; spawn_dispatch takes no credentials; the from_call forwarding path's auth_token source was OpSummary.credentials_auth_token: Option<String>, always None, never connected to CallCredentials.auth_token). auth_token is a per-request payload field, not a call-protocol credential. Update Cargo.toml to drop quinn/rustls/rustls-native-certs/ rustls-pemfile. Also remove from_call's credentials_auth_token dead path: the credentials_auth_token field on OpSummary, the credentials_auth_token parameters on make_forwarding_handler / make_streaming_forwarding_handler, and the auth_token parameter on build_forwarded_payload are all removed (always None, different type than CallCredentials.auth_token, never connected). The two from_call tests asserting the Some path (build_forwarded_payload_sets_auth_token_when_provided, streaming_forwarding_handler_sets_auth_token_when_provided) are removed — they test a code path never exercised in production.

The call_client.rs after prune:

• CallClient struct + new + registry + identity_provider + spawn_dispatch (~85 lines — unchanged)
• CallConnection + Dispatcher wiring (stays — protocol)
• RemoteIdentity — removed from call_client.rs (moved to alknet-core in Phase 0; re-imported from there)
• CallCredentials — removed (its auth_token field had no reader; connect() was its only consumer and is removed in this phase; auth_token is a per-request payload field, not a credential — ADR-091, amended 2026-07-17)
• ClientError — removed
• connect + all build_*/select_*/load_*/Ed25519SigningKey/ RawKeyClientCertResolver/NoClientCertResolver/ FingerprintPinVerifier — removed (now in alknet-tls)

The Cargo.toml after prune: quinn, rustls, rustls-native-certs, rustls-pemfile all leave. The quinn feature is removed (it only gated connect + the TLS helpers, both removed — keeping it as a no-op would mislead a user enabling quinn on alknet-call expecting QUIC support; removing it surfaces any stray #[cfg(feature = "quinn")] the prune missed). alknet-call becomes a pure protocol crate.

Test impact (per the test audit below): the lib tests in call_client.rs (16 tests) split into 4 that stay unchanged (protocol-level, use spawn_dispatch(stub_connection())), 10 that move to alknet-tls in Phase 1 (TLS/verifier tests), and 2 that move to alknet-core (testing ConnectionCredentials — the call_credentials_builder_methods and remote_identity_none_is_load_bearing_not_defaulted tests, which access remote_identity/tls_identity fields that moved to ConnectionCredentials). The from_call.rs tests: 25 stay unchanged (protocol-level, use CallConnection directly), 2 are removed (the credentials_auth_token Some-path tests — see above). The integration test file (tests/two_node_call.rs, 2 tests) splits: two_node_call_round_trip (dial + take-over composition) moves to alknet-client/tests/ (Phase 3, rewritten with a minimal echo ProtocolHandler on a test ALPN — no alknet-call dependency); from_call_discovers_and_forwards_over_quic_loopback (call-protocol- specific, uses from_call) stays in alknet-call (Phase 5, rewritten to use spawn_dispatch + loopback Connection — not AlknetClient::dial_quic, which would re-create the dep the prune removes).

The implementation prune: delete ClientError, delete connect, delete CallCredentials and its builder methods, delete the TLS helpers (~140 lines), delete from_call's credentials_auth_token dead path (~30 lines). The test work: 10 TLS tests already moved in Phase 1, 2 CallCredentials tests move to alknet-core, 4 protocol tests stay, 2 from_call dead-path tests removed, the integration test splits as above.

Compilable state: cargo test -p alknet-call passes with the rewritten tests. The crate has no TLS/transport deps, no CallCredentials, no from_call dead path.

Done when: cargo test -p alknet-call passes, the crate is a pure protocol crate.

Phase 6: Core stream unification — BiStream as the handler leaf

Status: Done (commit b60a584, 2026-07-18). All "Done when" criteria met: BiStream is the only type returned by accept_bi / open_bi; Connection::from_stream is removed; from_bidi is the only public stream constructor; cargo test passes workspace-wide (one pre-existing alknet-http failure remains — see Phase 9's "Pre-existing test failure to fix in a follow-up" note; unrelated to Phase 6, verified by stashing). The QuicStream wrapper and QuicStreamDuplex test helper removals that Phase 9 called for were done as part of this phase's call-site updates (Phase 9 is now marked done — subsumed).

What was planned: Implement ADR-092 (drafted, pushed) and the stream-unification resolution from docs/research/stream-unification/findings.md. The changes to alknet-core:

• accept_bi() returns BiStream (a concrete AsyncRead + AsyncWrite newtype wrapping the inner transport), not a split (SendStream, RecvStream) pair.
• Connection::from_stream is removed. from_bidi is the only public stream constructor.
• SendStream/RecvStream collapse to thin newtypes used only internally by BiStream (they never cross a crate boundary as part of a constructor).
• BiStream implements AsyncRead + AsyncWrite directly — no hand-rolled adapter needed at call sites.

Impact on call sites: Every handler that currently calls accept_bi() and gets a (SendStream, RecvStream) pair gets a BiStream instead. The QuicStream wrapper in alknet-http (server/adapter.rs:271-314) becomes unnecessary — BiStream is already AsyncRead + AsyncWrite. The QuicStreamDuplex test helper (adapter.rs:456-493) is similarly replaceable.

What was done (cross-crate, +408/-399 lines net -57):

• alknet-core (types.rs): added concrete BiStream struct boxing Box<dyn AsyncReadWrite + Unpin> with from_joined (pub — downstream crates that produce split halves naturally, e.g. the future channels reassembly path and tests, call this) and from_bidi (pub(crate)) constructors; changed BidiStreamSource::accept_bi/open_bi return types to BiStream; the join happens once in the QuinnBidiStreamSource / IrohBidiStreamSource / StreamBidiStreamSource impls via tokio::io::join; collapsed SendStream/RecvStream to thin newtypes over Box<dyn Async* + Send + Unpin> (removed SendStreamKind / RecvStreamKind enums and the quinn/iroh per-call dispatch — the join is in the BidiStreamSource impl now); removed Connection::from_stream; promoted Connection::from_bidi to the only public stream constructor; updated tests (from_source_tests, tests) to use from_bidi + BiStream, added a SinkEmpty test helper for Connection-level-only test connections.
• alknet-http (server/adapter.rs): dropped the 44-line QuicStream wrapper and the 38-line QuicStreamDuplex test helper; tests use a single tokio::io::duplex whose ends are each AsyncRead + AsyncWrite natively.
• alknet-tty (adapter.rs): TtyAdapter::handle splits the BiStream from accept_bi via tokio::io::split for drive_session's separate AsyncWrite/AsyncRead args (the stdlib idiom for TcpStream-style duplex streams).
• alknet-call (protocol/*, client/*): Dispatcher::handle_stream takes BiStream and splits internally; CallConnection::call_with_payload/ subscribe_with_payload/write_envelope split the open_bi result via tokio::io::split at the call site; write_request / read_stream_until_closed are now generic over AsyncWrite / AsyncRead (were: concrete SendStream/RecvStream); added protocol/test_support.rs with a shared sink_empty_connection() helper replacing the 5 duplicated stub_connection() fns; updated all test handle_stream call sites to build BiStream::from_joined(recv, send) from the existing BufReader<Cursor> + duplex pair.

Compilable state: cargo test -p alknet-core passes. Handler crates (alknet-tty, alknet-call, alknet-http) may need import updates but no logic changes — they already treat the result of accept_bi() as an AsyncRead + AsyncWrite pair.

Done when: cargo test passes workspace-wide, BiStream is the only type returned by accept_bi(), from_stream is removed, from_bidi is the only public constructor.

Phase 7: TTY control-channel bidirectionality fix

What: Fix the "control isn't actually bidirectional" flaw in alknet-tty. The current STREAM_CONTROL = 3 is documented as "bidirectional" but the adapter ignores Exit from the client (adapter.rs:462-463). The fix:

• Split STREAM_CONTROL = 3 into STREAM_CTRL_IN = 3 (client→server, write half) and STREAM_CTRL_OUT = 4 (server→client, read half).
• Update InvalidStreamType bound from > 3 to > 4.
• Update ChunkReader/ChunkWriter to handle the new stream types.
• Update the adapter to properly route control messages on both halves.
• Update control.rs to reflect the split (resize/signal/eof on ctrl_in, exit on ctrl_out).

This is a TTY-layer fix — the channels layer has no stream_type concept and is unaffected.

Compilable state: cargo test -p alknet-tty passes. The control channel is properly bidirectional.

Done when: cargo test -p alknet-tty passes, STREAM_CONTROL is replaced with STREAM_CTRL_IN/STREAM_CTRL_OUT, the adapter handles both directions correctly.

Phase 8: Channels spec cleanup — 8-byte wire format, no stream_type

What: Update the channels crate specs (docs/architecture/crates/ channels/) and ADRs to reflect the stream-unification resolution:

• Wire format is 8 bytes: [channel_id:u32 BE][length:u32 BE] followed by an opaque payload. The channels layer owns channel_id and length; the payload is the handler's framing, carried transparently.
• The channels layer has no stream_type concept — not in its header, not in its code, not in its mental model.
• into_sub_streams() is removed. accept_bi is the only accessor; it yields one BiStream per channel.
• Every channel is a BiStream. Handlers sub-multiplex their BiStream however they want (TTY's 5-byte format, call's length-prefixed JSON, tunnel's raw bytes, SSH's channel protocol).
• The add/strip utility: add_channel_id(channel_id, payload_bytes) -> chunk on write; strip_channel_id(chunk) -> (channel_id, payload_bytes) on read.

ADRs amended:

• ADR-071: wire format is 8 bytes, not 9; no stream_type in the channels header.
• ADR-074: into_sub_streams removed; accept_bi is the only accessor.
• ADR-077: reversed — TTY always uses its 5-byte format; the channels layer carries it transparently in the payload.

This is a spec/docs phase — the channels crate doesn't exist yet (per ADR-081, it's planned as alknet-channels-core + alknet-channels-call). The POC at /workspace/alknet-channels-poc/ validated the 9-byte format; the spec update changes it to 8 bytes before implementation begins.

Compilable state: No code changes — spec/docs only. cargo test passes workspace-wide (unchanged).

Done when: ADRs 071/074/077 are amended, channels spec docs are updated, the POC's wire format notes are updated.

Phase 9: Fix alknet-http — drop QuicStream wrapper

Status: Done — subsumed by Phase 6 (commit b60a584, 2026-07-18). The QuicStream wrapper (44 lines) and the QuicStreamDuplex test helper (38 lines) were removed as part of Phase 6's call-site updates (per ADR-092 migration step 2 — the wrapper removal is the alknet-http call-site update). grep -r QuicStream crates/ returns no matches; HttpAdapter::handle is 4 lines; the test helpers use a single tokio::io::duplex whose ends are each AsyncRead + AsyncWrite natively.

What was planned: Remove the QuicStream wrapper from server/adapter.rs (lines 271-314) and the QuicStreamDuplex test helper (lines 456-493). After Phase 6, accept_bi() returns BiStream which is already AsyncRead + AsyncWrite — the hand-rolled adapter is unnecessary.

The change: HttpAdapter::handle calls connection.accept_bi(), gets a BiStream, and passes it directly to serve_io() (or via TokioIo::new if the hyper adapter needs it). The 44-line QuicStream wrapper is deleted. The test helper is replaced with BiStream-based test utilities.

Why this was a light pruning: The original Phase 6 was deferred because SendStream is AsyncWrite-only and RecvStream is AsyncRead-only — the QuicStream wrapper was necessary to combine them. After Phase 6, BiStream bundles both halves natively. The wrapper becomes dead code.

Pre-existing test failure to fix in a follow-up: Phase 9's "Done when" criterion was "cargo test -p alknet-http passes." One alknet-http test fails on the develop baseline (before, during, and after Phase 6 — verified by stashing Phase 6 and re-running): adapters::to_mcp::tests::search_returns_access_control_filtered_ops_excluding_subscriptions panics with "handler kind mismatch: Subscription requires HandlerKind::Stream (got HandlerKind::Once)". The bug is in the test helper full_registry_with_ops (to_mcp.rs:501-516), which always registers with HandlerKind::Once(make_echo_handler()) regardless of op_type — when the test passes OperationType::Subscription for "events/stream", the registry's kind validation (tightened in commit 9c81129 feat(call): introduce StreamingHandler, HandlerKind, ResponseStream + INVALID_OPERATION_TYPE (ADR-049)) rejects it. The fix is in the test helper: branch on op_type and use HandlerKind::Stream(make_streaming_handler(...)) for Subscription ops (mirror the pattern in dispatch.rs's registry_with_subscription). This is unrelated to Phase 6/9 — it's a test-helper bug that predates the stream-unification work — but it blocks Phase 9's "Done when" criterion and should be fixed in a small follow-up commit before Phase 9 is considered fully closed.

---

Intermediate states (compilable after each phase)

After phase	State
0 (credentials)	ConnectionCredentials/RemoteIdentity in core; call imports from core; CallCredentials removed; no breakage
1 (tls)	alknet-tls builds standalone; core/call/http unchanged (old code duplicated)
2 (endpoint)	alknet-endpoint builds standalone; core still has old endpoint.rs (duplicate)
3 (client)	alknet-client builds standalone; call still has old connect (duplicate)
4 (core prune)	core is lightweight; endpoint.rs gone; ConnectionCredentials in core
5 (call prune)	call is pure protocol; connect + TLS helpers + CallCredentials + from_call dead path gone; Category B tests already moved; 2 CallCredentials tests moved to core
6 (stream unification)	Done (b60a584). BiStream is the handler leaf; accept_bi returns BiStream; from_stream removed; from_bidi is the only public constructor; SendStream/RecvStream are thin internal newtypes. Subsumed Phase 9's QuicStream/QuicStreamDuplex removal.
7 (TTY control fix)	TTY control channel is properly bidirectional (STREAM_CTRL_IN = 3, STREAM_CTRL_OUT = 4); InvalidStreamType bound updated. Unchanged by Phase 6 — Phase 7's work is in wire.rs and control.rs, neither of which Phase 6 touched.
8 (channels spec)	Channels spec updated to 8-byte wire format; no stream_type concept; into_sub_streams removed; ADRs 071/074/077 amended. Unchanged by Phase 6 — docs-only, no code.
9 (http fix)	Done — subsumed by Phase 6 (b60a584). QuicStream wrapper removed from alknet-http; BiStream used directly; QuicStreamDuplex test helper removed. One pre-existing alknet-http test failure remains (test-helper bug in to_mcp.rs::full_registry_with_ops, unrelated to Phase 6/9) — see Phase 9's "Pre-existing test failure to fix in a follow-up" note.

Phases 0-3 are purely additive — no existing code breaks, no tests break. Phases 4-5 are subtractive — the pruned code's callers don't exist yet (no assembly layer), so the breakage is confined to the crate's own tests (and per the test audit, the call prune removes CallCredentials and the from_call dead path; the TLS tests moved in Phase 1, the CallCredentials tests moved to core, the protocol tests use spawn_dispatch directly). Phase 6 (stream unification) is a core refactor that touches all handler crates but is mechanical — handlers already treat accept_bi() results as AsyncRead + AsyncWrite. Phase 7 (TTY control fix) is scoped to alknet-tty. Phase 8 (channels spec) is docs-only. Phase 9 (http fix) is a light pruning enabled by Phase 6.

Ordering rationale

The ordering is deps before dependents, additive before subtractive:

• Phase 0 (ConnectionCredentials to core) first because it's independent, additive, and makes alknet-client (Phase 3) never depend on alknet-call. ~40 lines moved, zero breakage.
• alknet-tls (Phase 1) because both alknet-endpoint (indirectly — the assembly layer builds TlsServerConfig) and alknet-client (directly — TlsClientConfig) depend on it. It has no dep on the other new crates.
• alknet-endpoint (Phase 2) depends only on alknet-core (already exists) — it doesn't need alknet-tls (the endpoint takes pre-built transports). It could go before alknet-tls, but putting tls first means the assembly layer's transport-building code has a home from the start.
• alknet-client (Phase 3) depends on alknet-tls (TlsClientConfig) + alknet-core (ConnectionCredentials — moved in Phase 0, so the dep is clean from the start).
• Phases 4-5 (the prunes) go after the new crates because they're subtractive. The new crates (0-3) must exist first so the pruned code's functionality has a home.
• Phase 6 (stream unification) goes after the prunes because it touches all handler crates — the prunes reduce the surface area first, making the BiStream refactor simpler.
• Phase 7 (TTY control fix) goes after stream unification because TTY's wire.rs already needs updating for the BiStream change (the ChunkReader reads from an AsyncRead; after Phase 6 it reads from a BiStream which is the same trait). The control channel split is a small additional change on top.
• Phase 8 (channels spec) is docs-only and can happen anytime after the stream-unification research settles. Placed after Phase 7 because the TTY fix validates the "handler owns its sub-streams" model before the channels spec is finalized.
• Phase 9 (http fix) goes last because it depends on Phase 6 (BiStream makes the QuicStream wrapper unnecessary).

Resolved decisions

ConnectionCredentials/RemoteIdentity move — Phase 0 (before Phase 1)

Decision: Move ConnectionCredentials/RemoteIdentity to alknet-core as a standalone additive step before any new crate is created (ADR-091). It's independent of everything else, purely additive (core gains two small types, nothing breaks), and alknet-call imports them from core so its own code + tests don't change yet. This means alknet-client (Phase 3) never depends on alknet-call — the dep graph is clean from the start, no temporary dep to clean up later.

ConnectionCredentials (not CallCredentials) is what moves — it is the transport-level credential bundle (local_identity + remote_identity), carrying only the dimensions the dial consumes. CallCredentials is removed (its auth_token field had no reader — connect() read only tls_identity + remote_identity; spawn_dispatch takes no credentials; the from_call forwarding path's auth_token source was OpSummary.credentials_auth_token: Option<String>, always None, never connected to CallCredentials.auth_token). auth_token is a per-request payload field, not a call-protocol credential. See ADR-091 (amended 2026-07-17) for the full rationale and trace.

The move is ~40 lines (struct definitions + builder impls) into a new crates/alknet-core/src/credentials.rs. alknet-call's client/mod.rs imports ConnectionCredentials + RemoteIdentity from core and re-exports them; CallCredentials is removed from alknet-call. Test changes: the two CallCredentials-field tests (call_credentials_builder_methods, remote_identity_none_is_load_bearing_not_defaulted) move to alknet-core testing ConnectionCredentials; call_client_is_send_sync drops the CallCredentials/RemoteIdentity assertions (or they move with the types).

Phase 5 test audit — call_client.rs (16 tests)

The 16 tests in call_client.rs split into three categories:

Category A — protocol-level, stay in alknet-call, no rewrite needed (4 tests):

These tests use spawn_dispatch(stub_connection()) and don't touch connect, CallCredentials fields, or any TLS helper. stub_connection() (line 582) uses Connection::from_stream(tokio::io::channel(...)) — already transport-agnostic. These survive the prune unchanged.

Test	Line	What it tests
external_op_dispatches_and_populates_capabilities	665	dispatch + capabilities
unknown_op_returns_not_found	679	dispatch error path
spawn_dispatch_returns_live_call_connection	691	spawn_dispatch + ALPN
call_client_is_send_sync	705	trait bounds (import update: RemoteIdentity moved to core)

Category A2 — CallCredentials-field tests, move to alknet-core (2 tests):

These test CallCredentials fields (remote_identity, tls_identity) that moved to ConnectionCredentials in alknet-core (ADR-091). They move to alknet-core testing ConnectionCredentials::new() + with_remote_identity().

Test	Line	What it tests	Move target
call_credentials_builder_methods	652	CallCredentials builder (now ConnectionCredentials)	alknet-core
remote_identity_none_is_load_bearing_not_defaulted	921	CallCredentials::new() (now ConnectionCredentials)	alknet-core

Category B — TLS/verifier tests, move to alknet-tls (10 tests):

These test FingerprintPinVerifier, build_client_auth, select_server_verifier, and build_quinn_client_config directly. They're #[cfg(feature = "quinn")]-gated and test the TLS helpers, not the call protocol. They move to alknet-tls in Phase 1 (adapted to test TlsClientConfig::new instead of the free functions). The two build_quinn_client_config tests test the full config build (verifier + client-auth + provider wired together) and are adapted to test TlsClientConfig::new + for_quinn() instead of the free function.

Test	Line	What it tests	Move target
fingerprint_pin_verifier_matches_correct_ed25519_fingerprint	750	verifier accept	alknet-tls
fingerprint_pin_verifier_rejects_wrong_ed25519_fingerprint	769	verifier reject	alknet-tls
fingerprint_pin_verifier_matches_correct_sha256_fingerprint	789	verifier X.509 accept	alknet-tls
fingerprint_pin_verifier_rejects_wrong_sha256_fingerprint	806	verifier X.509 reject	alknet-tls
select_server_verifier_returns_ca_verifier_for_none	822	CA path	alknet-tls
select_server_verifier_returns_fingerprint_pin_for_some	839	pin path	alknet-tls
build_client_auth_presents_ed25519_raw_key_without_error	857	client cert resolver	alknet-tls
build_client_auth_none_resolves_to_no_client_cert	879	no-cert resolver	alknet-tls
build_quinn_client_config_with_raw_key_identity_builds_without_error	893	full config build	alknet-tls
build_quinn_client_config_with_no_remote_identity_builds_without_error	909	CA-verify config	alknet-tls

Category C — connect integration test, remove (0 tests):

No test in call_client.rs actually calls connect(). The tests that exercise the full QUIC dial path are in from_call.rs (which has 27 tests) and in the integration tests. call_client.rs's tests are all either protocol-level (Category A) or TLS-helper-level (Category B). This means the connect removal doesn't break any test in call_client.rs itself — the tests that need a real connection already use spawn_dispatch(stub_connection()).

The from_call.rs tests (27 tests): these use CallConnection directly (constructed from stub_connection() or a mock), not connect. 25 are protocol-level and stay in alknet-call unchanged. 2 are removed: build_forwarded_payload_sets_auth_token_when_provided and streaming_forwarding_handler_sets_auth_token_when_provided — they test the credentials_auth_token Some path, which is removed (the field was always None, never connected to CallCredentials). The one reference to connect() is in a doc comment (line 76: "the assembly layer calls from_call immediately after connect()") — update the comment to say "after AlknetClient::dial_* + spawn_dispatch".

Net Phase 5 test impact: 4 tests stay unchanged (Category A), 2 tests move to alknet-core (Category A2), 10 tests move to alknet-tls in Phase 1 (Category B), 2 from_call dead-path tests removed. The from_call.rs tests: 25 stay unchanged, 2 removed. The prune of call_client.rs is mechanical: delete ClientError, connect, CallCredentials and its builder methods, and all the TLS helpers (build_*, select_*, load_*, Ed25519SigningKey, RawKeyClientCertResolver, NoClientCertResolver, FingerprintPinVerifier); keep CallClient + new + spawn_dispatch unchanged; update imports. Also remove from_call's credentials_auth_token dead path: the field on OpSummary, the parameters on make_forwarding_handler / make_streaming_forwarding_handler, and the auth_token parameter on build_forwarded_payload. The test suite keeps the Category A tests, removes the Category A2 tests (moved to core), removes the Category B tests (moved in Phase 1), removes the 2 from_call dead-path tests, and updates the one doc comment.

This is a larger prune than the initial estimate of "~140 lines of implementation + ~290 lines of test restructuring." The actual work: delete connect + TLS helpers (~140 lines), delete CallCredentials

• builder methods (~50 lines), delete from_call's credentials_auth_token dead path (~30 lines), move 10 tests to alknet-tls (Phase 1), move 2 tests to alknet-core, keep 4 tests unchanged, remove 2 from_call dead-path tests. The connect removal breaks zero lib tests because no lib test calls connect.

Resolved questions

• Phase 9 accept_bi semantics: Resolved 2026-07-18. After Phase 6 (stream unification), accept_bi() returns BiStream which is already AsyncRead + AsyncWrite. The QuicStream wrapper (44 lines) becomes unnecessary — BiStream bundles both halves natively. Phase 9 is a light pruning: delete the wrapper, pass BiStream directly to serve_io(). The original deferral (2026-07-17) was correct at the time (SendStream was AsyncWrite-only, RecvStream was AsyncRead-only) but is obsoleted by the BiStream refactor.
• Phase 6-8 ordering: Resolved 2026-07-18. Stream unification (Phase 6) goes before TTY control fix (Phase 7) because TTY's wire.rs already needs updating for BiStream. Channels spec cleanup (Phase 8) is docs-only and goes after the TTY fix validates the "handler owns its sub-streams" model. HTTP fix (Phase 9) depends on Phase 6.
• Integration test home: Resolved. The dial + take-over composition test moves to alknet-client/tests/ with a minimal echo ProtocolHandler on a test ALPN — no alknet-call dependency, no circular path. The call-protocol-specific tests stay in alknet-call (rewritten to use spawn_dispatch + loopback Connection). See Phase 5 / open questions above.
• Workspace Cargo.toml: the three new crates need to be added to the workspace member list. Trivial but worth noting.