1 files changed, 99 insertions, 0 deletions

diff --git a/PLAN.md b/PLAN.md
new file mode 100644
index 0000000..ab277d9
--- /dev/null
+++ b/PLAN.md
@@ -0,0 +1,99 @@
+# Axon — Implementation Plan
+
+## Dependency Philosophy
+
+Minimize external dependencies at every layer. The protocol is designed so that the Go implementation requires nothing beyond the standard library and `golang.org/x/crypto`. The JS implementation requires only `@noble/curves` and `@msgpack/msgpack` — both small, audited, zero-dependency libraries. No frameworks, no ORMs, no generated code.
+
+---
+
+## Phase 1: Core Package
+
+**Everything depends on this. Do not move to Phase 2 until the signing spec is fully verified.**
+
+- `Event` struct, `Tag`, kind constants matching the registry in PROTOCOL.md
+- Canonical signing payload construction — byte-exact per the spec
+- Tag sorting (`canonical_tags` hash)
+- Ed25519 sign and verify (`crypto/ed25519`)
+- X25519 key conversion from Ed25519 keypair (`golang.org/x/crypto/curve25519`)
+- ChaCha20-Poly1305 encrypt/decrypt for DMs (`golang.org/x/crypto/chacha20poly1305`)
+- MessagePack encode/decode for events
+
+**Test vectors — write these before any implementation:**
+
+Publish a set of known-input → known-output pairs for:
+- `canonical_tags` hash given a specific tag list
+- `canonical_payload` bytes given a specific event
+- Event `id` (SHA256 of payload)
+- Signature verification
+
+These vectors must be committed to the repo and validated by the JS client in Phase 4. They are the ground truth for cross-implementation correctness. A disagreement on a test vector means the spec is ambiguous — fix the spec before fixing the code.
+
+**Go dependencies:** `golang.org/x/crypto` only.
+
+---
+
+## Phase 2: Relay
+
+- WebSocket server (stdlib `net/http` + `golang.org/x/net/websocket` or `nhooyr.io/websocket`)
+- Challenge/Auth handshake on connect
+- Allowlist: authorized pubkeys in a flat config file or SQLite table
+- Ingest pipeline:
+  1. Unmarshal msgpack envelope
+  2. Verify signature using Phase 1 core
+  3. Reject duplicates (id PRIMARY KEY)
+  4. Reject expired job requests (check `expires_at` tag)
+  5. Reject events exceeding 64KB content limit
+  6. Write to index + store
+  7. Fanout to matching subscribers
+- SQLite storage via `database/sql` + `mattn/go-sqlite3` or `modernc.org/sqlite` (pure Go, no CGo)
+- Subscription management: filter matching, per-connection subscription map
+- Ephemeral events (kind 3000–3999): fanout only, skip storage
+- Error responses with HTTP-borrowed codes
+- WebSocket keepalive: ping every 30s, close after two missed
+
+**Go dependencies:** sqlite driver, websocket library. Both should be pure Go where possible.
+
+---
+
+## Phase 3: Go Client
+
+- Connect to relay + complete auth handshake
+- Publish events
+- Subscribe with filter, receive event stream
+- Unsubscribe
+- Reconnect with backoff
+
+The Go client is the primary tool for integration testing the relay. All relay behavior should be exercised via the client before moving to Phase 4.
+
+**Go dependencies:** none beyond what Phase 1 and Phase 2 already use.
+
+---
+
+## Phase 4: JS Client
+
+- Connect + auth (Ed25519 sign the challenge)
+- Publish events (msgpack encode, Ed25519 sign)
+- Subscribe / unsubscribe
+- DM encrypt/decrypt (X25519 ECDH + ChaCha20-Poly1305)
+- Validate Phase 1 test vectors — JS and Go must agree on every vector before this phase is considered complete
+
+**JS dependencies:** `@noble/curves` (Ed25519, X25519), `@msgpack/msgpack`. No others.
+
+---
+
+## Phase 5: Integration
+
+- End-to-end: publish from JS client, receive in Go client (and vice versa)
+- Cross-language signature verification: Go signs, JS verifies; JS signs, Go verifies
+- DM round-trip: encrypt in JS, decrypt in Go
+- Job protocol flow: request → feedback → result across two separate processes
+
+At the end of Phase 5 you have a working protocol implementation ready to back an agent system.
+
+---
+
+## What Comes After
+
+- Agent discovery via `KindProfile` (kind 0) — capability advertisement
+- Relay clustering / replication (out of scope for v1)
+- Retention and event expiry policies