docs/link/v6.md

# Protocol 6 — A guarantee, and the kind that pays for it

**Status:** In review on `edge`. Part **a** of two: see §7 for what protocol 6 gains in 11b before it cuts over.
**Date:** 2026-09-04
**Codebase:** ServUO 57.4, `<servuo>`, net48 / x64, Expansion **EJ**.
**Companion to** [`PLAN.md`](PLAN.md) (1.0 read/event plane), [`PROTOCOL_2.md`](PROTOCOL_2.md) (2.0 provisioning + world-state streams), [`v3.md`](v3.md) (3.0 shard content + the visibility framework), [`v4.md`](v4.md) (4.0 guild membership), [`v5.md`](v5.md) (5.0 decay schedule, vendor fees, login result), [`INTEGRATION.md`](INTEGRATION.md) (website API).
**Driven by** [`../website/EVENTS.md`](../website/EVENTS.md) and [`../website/EVENTS_PLAN.md`](../website/EVENTS_PLAN.md) Phase 11.

---

## 1. Why this bump is not like the other five

Every protocol version so far added **data**: a field, a frame, a board. This one adds a **promise**,
and the data it also carries is along for the ride because the ride is expensive.

The promise is one sentence:

> **A command carrying an `idempotencyKey` is executed at most once. A repeat is answered with the
> original reply.**

Nothing on the wire could say that before, and [`EVENTS.md`](../website/EVENTS.md) §A row 7 records
what that cost: *"a command whose ack is lost is indistinguishable from one that never applied."*
The website's event runner retries a step that did not come back — that is what a runner is for —
so every world-writing verb had to be declared un-retryable, and Phase 9 shipped `uo.broadcast`
answering `retry: false` to *everything*, including a 503 from a shard that was merely restarting.
The reasoning was sound and the trade was real: a lost announcement is cheaper than a doubled one.

It is not a trade that survives contact with the rest of the workstream. An event that spawns
creatures, grants an item, or holds a lease on a live config value cannot be built on a wire where
"did that apply?" has no answer. §G's capability table says so in as many words: *"Required before
any world write ships."*

**So the guarantee had to come before the verbs, and this is the version that has nothing else to
do.** Protocol 12's world verbs are the payload; protocol 6 is the floor they stand on.

### 1.1 Why `champ.boss.killed` rides along

The same argument v5 §1 made: a bump costs a sidecar release, a republished bundle and an operator
update on every shard, so a field left out costs a *second* bump rather than a follow-up commit.

`champ.boss.killed` is the one piece of new data the events workstream is known to need and can
state now. §G's capability table had a boss defeat as 🔗 — *inferable* from `champ.update` going
`bossUp` true then false, correlated against a nearby `mob.killed` — and called that
*"workable and slightly fragile"*. It is more fragile than that: `bossUp` also drops when a GM
resets a spawn, when a boss despawns, and after a sidecar reconnect clears the sweep's diff cache.
And the inference is silent about the thing an event most wants to know, which is **who fought it**.

A phase condition is written over a trigger firing, so a kind nothing declares is a kind no event
can wait on. *"Advance when the boss falls"* is the canonical UO event phase boundary, and it did
not exist.

---

## 2. The shard side

### 2.1 The idempotency key

**One new optional field on every inbound command**, and no new command kinds:

```json
{"kind":"admin.broadcast","reqId":"r-17","actor":"event:412","text":"...",
 "idempotencyKey":"5f2c…40 hex"}
```

The gate is in `BridgeBoot.OnInboundLine`, before the handler is dispatched, so it covers **every**
inbound kind including the ones a later protocol adds — which is the half that is easy to forget. A
command with no key behaves exactly as it did before protocol 6, which is what leaves the admin
screens unchanged: a human pressing a button can see whether the thing happened, and sends none.

The store is `BridgeIdempotency`. Four rules, each of which is a decision rather than an
implementation detail.

**1. Reserve on receipt, not on completion.** The key is recorded *before* the handler runs. A
handler that finishes inside its own inbound call can never see a repeat — the Core thread processes
one line at a time — but a handler that defers (a lease that arms a timer, a spawn that waits for a
save) completes long after `OnInboundLine` has returned, and that is precisely the window a lost
acknowledgement opens. Reserving late would leave it uncovered.

A repeat of a key still in flight is answered **`bridge.busy`**: nothing runs, and the caller is
told to come back. It is deliberately not spelled `bridge.busy.error` — nothing is wrong, the work
is happening.

**2. A key that has begun is never released.** Not even when the handler throws. Releasing it would
let a retry re-run a command that may have applied half of itself, which is the exact failure this
file exists to prevent. A handler that throws stores a `bridge.error` reply instead, so the retry
gets a definite answer and the step fails once rather than looping.

**3. A replay is stamped with the REPEAT's correlation id.** The sidecar's `reqId` is a fresh
per-process counter, so a retry is waiting on an id the first attempt never used. Replaying the
original id verbatim would leave the call hanging until the reply timeout — the very failure being
answered. The stored reply is re-emitted with its correlation field rewritten and `"replayed": true`
appended; everything else is byte-for-byte the original, including its original `t`.

`reqId`, `code` and `id` are all recognised, in the order `rpc.rs` tries them. A repeat that
correlates on a *different* field than the original did is refused with a `bridge.error` rather than
answered under an id nobody is waiting on.

**4. The bound is loud.** TTL one hour, cap 4096 keys, drop-oldest. The hour is not a tuned number:
core's step lease is fifteen minutes and its retry backoff is bounded well inside that, so expiry
should never be the thing that ends a key's life. The cap is the guarantee's **one hole** — an
evicted key's repeat *would* be applied a second time — so an eviction that drops a key still inside
its TTL prints a console warning naming the count. If the promise is ever actually breached, an
operator reads it here rather than discovering a doubled spawn in the world.

`[bridge status` reports `idem(keys= seen= replayed= busy= evicted= uncorrelated=)`.

#### 2.1.1 How the reply is captured

The handlers emit their answers through `BridgeLink.Emit` like everything else, so there is nothing
to return. While a keyed handler runs, `Emit` offers each line to the store, and the store keeps
**only the line the sidecar would correlate with this command** — the one whose `reqId`/`code`/`id`
equals the command's.

That filter is the point. An `admin.audit` frame emitted alongside the reply is a fact about the
world and must **not** be replayed to a retry; the reply is an answer to a caller and must be. The
capture happens before `Emit`'s `_running` check, too, because a reply the link was too dead to
deliver is exactly the one a retry will come back for.

A keyed command that produces no correlated reply at all is treated as a defect rather than as a
state to model: it is logged by kind and key, and a definite `bridge.error` is stored so the retry
terminates instead of timing out forever.

### 2.2 `champ.boss.killed`

A new kind, emitted from `EventSink.CreatureDeath` in `BridgeChamps`.

```json
{"t":1788551315,"kind":"champ.boss.killed","category":"champion",
 "bossSerial":"0xD8D","boss":"Semidar","bossType":"Semidar",
 "map":"Felucca","x":1496,"y":1628,"z":-5,
 "serial":"0x400150E8","type":"Abyss","level":0,
 "region":"Britain",
 "killer":{"serial":"0x2E0","name":"tester","acct":"wttest","webId":"1","player":true},
 "damagers":[{"serial":"0x2E0","name":"tester","acct":"wttest","webId":"1","player":true,"damage":100240},
             {"serial":"0x24C","name":"Darrow","acct":"whitlocktech","webId":"14","player":true,"damage":120}]}
```

| Field | Notes |
|---|---|
| `category` | `champion` or `sea`. There is no `mini`: a `MiniChamp` has no boss |
| `bossSerial`, `boss`, `bossType` | The creature. `boss` prefers its display name and falls back to its type |
| `serial`, `type`, `level` | **The ALTAR**, present only when the kill could be attributed to one. `serial` means the spawn here, matching `champ.update`, so a consumer joins the two without a rule about which of two serials means what |
| `map`, `x`, `y`, `z`, `region` | Where it fell. `region` is the nearest **named** region and is absent in open countryside — see §6.1 |
| `killer` | The standard actor object, and the last blow only |
| `damagers` | Every player who damaged it, **highest first**, each an actor object plus `damage` |

**Detection is by TYPE, not by the sweep.** `BaseChampion` — which `BaseSeaChampion` derives from —
so one check covers both families, and a boss that popped and died inside one sweep interval is
still reported. The sweep's `bossSerial → spawnSerial` map supplies only the altar *context*; a kill
it cannot attribute arrives without `serial`, `type` and `level` rather than not arriving.

**The damage table exists here and nowhere else.** ServUO discards a creature's damage entries with
the creature, so the shard is the only party that ever sees them. Totals are summed per damager
rather than trusted to be one entry each: ServUO folds repeat damage into an existing entry, but an
entry that expired and was re-created leaves two, and a table listing the same player twice reads as
two participants. Entries are reported **whether or not ServUO considers them expired** — expiry
governs looting rights, and someone who fought the first two thirds of a champion fight and then
died took part in it regardless of what they are owed from the corpse.

Capped at 20 entries so the frame stays one line on the wire.

---

## 3. The sidecar side

`PROTOCOL_VERSION: u32 = 5` → `6`, and **one behaviour**: `bridge.busy` maps to HTTP **425 Too
Early** in all three responders.

Everything else is free. The key rides in the command body, which every write endpoint already
passes through verbatim; `champ.boss.killed` lands in `events` and on the feed through the generic
forward path with no arm of its own. There is **no store migration** — nothing gains a column.

That is the dumb-forwarder property doing its job again (v3 §3), and it is worth naming what it
means here specifically: **the sidecar makes no idempotency promise of its own.** It does not dedupe,
does not cache, and does not know what a key means. The guarantee is the shard's, end to end, which
is the only place it can be — the shard is where the world write happens.

### 3.1 Why 425 and not 409

409 is already the protocol-version gate's answer, and the two want **opposite dispositions** from a
client: a version mismatch is a deployment fault nobody should retry, and a busy shard is a retry
that will succeed on its own. Sharing a status would make the difference readable only by inspecting
the body, which is how a retry loop ends up hiding a mismatched deployment.

425 is what that status is for — a server unwilling to risk processing a request that might be a
replay — and it is unambiguous here because nothing else on this surface uses it.

A **replayed** reply, by contrast, is an ordinary **200**. The caller must be able to treat it
exactly as it would have treated the answer it lost; `replayed: true` is for the log.

---

## 4. Visibility

One classification, made on the website in `module-uo`'s `shardVisibility.js`, never in the sidecar.

| Field / kind | Audience | Why |
|---|---|---|
| `champ.boss.killed` | **`anonymous`**, on the existing `champs` feature | A champion falling is announced in-world and is the content the public board is *for*. Mapping it is required, not optional: rule 2 fails an unmapped kind closed to admin-only |
| `champ.boss.killed` → `damagers` | **`staff`**, configurable | Nested, so one rule covers the whole table |
| `champ.boss.killed` → `killer` | **not listed** | One actor, whose blow everyone present saw, and the same disclosure `mob.killed` has published on the public activity feed since before this framework existed |
| `damagers[].acct` / `.webId` | **admin, locked** | Rule 1, automatically, by suffix — inside the array, exactly as it already works inside a guild roster |

**The split is the whole point.** A shard announces that its champion fell without publishing a
ranked roll of who was strong enough to fell it. The kill is an event in the world; the damage table
is a performance record of named players that nobody consented to publish. A shard that wants a
public "who slew the champion" board lowers **one** field rule.

**The trigger carries the count, never the names.** `uo.champ.boss_killed` exposes `damagerCount`
and a `damagerNote` sentence and no damager identity at all. A trigger variable is interpolated into
mail an operator may address to every subscriber, so a name reaching the trigger's data would undo
the field rule one layer up — a distinct hole from the SSE path the rule guards, reachable without
touching the visibility config at all.

---

## 5. Cross-repo obligations

| Repo | Change |
|---|---|
| `servuo-plugins` | `BridgeIdempotency.cs` (new) · `BridgeBoot.OnInboundLine` gate · `BridgeLink.Emit` capture hook · `BridgeJson.RewriteStringField` / `WithTrueFlag` / `Damagers` · `BridgeChamps` boss-kill emitter · **`overlay.toml` `protocol = 6`, in the same PR as the emitters** |
| `link` | `PROTOCOL_VERSION` → 6 · `bridge.busy` → 425 in all three responders |
| `module-uo` | `uoLinkClient` ×3 writes carry the key · `uoEventActions` passes it and `uo.broadcast` becomes retryable · `shardVisibility` (kind + field rule) · `shardEngagement` mapper · `shardTriggers` + `engagementSeeds` for `uo.champ.boss_killed` |
| `docs` | this file · `INTEGRATION.md` · `PLAN.md` §5/§7 · `EVENTS.md` §A/§G · `EVENTS_PLAN.md` |
| `installer` | **nothing.** See below |
| `runicgateway.com` | `platform.json.protocol` → 6 — deferred to the events cutover, because `checkFacts.mjs` fetches from `main` and setting it during the `edge` period turns that repo red immediately |

**`installer` has no work, and `EVENTS_PLAN.md` Phase 11 was wrong to list it.** The plan put it in
the phase *"because of the pairing"*, but `PLAN.md` §7.4 made that unnecessary on purpose:
**no protocol version is hardcoded anywhere in the installer.** `bundle.rs` reads the number out of
the bundle document and `bundle.yml`'s Gate 1 reads both halves out of the released artefacts and
refuses a mismatch. The check is version-agnostic, so it needs no change for 6 exactly as it needed
none for 5. And `link` and `servuo-plugins` are on `edge` for this workstream, so nothing is
released or bundled until the events cutover in any case.

**The pin still has three declaration sites** — `overlay.toml`, `PROTOCOL_VERSION`, and
`module-uo`'s `uo_link_config` default — and `module-uo`'s schema test asserts that they *agree*
rather than that they equal a literal. See v5 §5 for why that phrasing is load-bearing.

---

## 6. Verification

Unit tests: 571 in `module-uo/server` (571 pass / 0 fail, 9 new), 43 in the sidecar (4 new), 42 in
`module-uo/client`; the C# compiles against the real ServUO 57.4 reference assemblies. `cargo fmt`,
`cargo clippy -D warnings` and `check:imports` clean.

Everything below was proved on the local rig — a real ServUO with a seeded world (43k mobiles, 209k
items), the **release** Rust sidecar, and `tools/scaffolding/BridgeProtocol6Probe.cs`.

| Claim | Evidence |
|---|---|
| a repeat is not re-executed | Three `POST /admin/broadcast`, two under key `K1` and one under `K2`. The shard's own audit trail holds **two** `admin.audit` rows, not three |
| a replay is the ORIGINAL reply | Attempt 2 answered `{"kind":"admin.ok","replayed":true,"reqId":"r-2","t":1788550182074}` — the first attempt's `t`, so it is stored rather than re-run |
| a replay is stamped with the REPEAT's id | The same reply carries `reqId: r-2`, not the `r-1` the first attempt used. Without this the retry hangs until the reply timeout |
| the `id`-correlated plane works too | Two `POST /towncrier` under one key → `{"kind":"towncrier.ok","id":"evt-CRIERKEY","replayed":true}` on the second, correlated on `id` rather than `reqId` |
| a different key still executes | Attempt 3 answered a fresh `r-3` with a new `t`, and produced the second audit row |
| the shard says what it did | `[Bridge] idempotency: replaying the original reply for key K1-… (admin.broadcast)` |
| `champ.boss.killed` fires for a real champion | A real `ChampionSpawn` driven through its own `SpawnChampion()`, damaged from two seeded players and killed: `boss: Semidar`, `category: champion` |
| the altar is attributed | `serial: 0x400150E8`, `type: Abyss`, `level: 0` — from the sweep's map, so the kill names the altar and not only the creature |
| the damage table ranks and SUMS | `[("tester", 100240), ("Darrow", 120)]` — descending, and `tester`'s 240 registered damage plus the 100000 killing blow folded into one entry rather than two |
| the region resolves | `region: "Britain"` — see §6.1, this is the one that took four rig cycles |
| the website's half agrees, on the real bytes | The captured frame fed through `module-uo`: trigger `uo.champ.boss_killed` with `location: "Felucca 1496, 1628 (Britain)"`, `damagerCount: 2`; projected `damagers` **absent** for `anonymous` and `player`, present for `staff`, and `acct` stripped from every actor at every rung below admin |
| the sidecar needed no change beyond the constant and one status | `GET /health` → `"protocol":6`; the key and the new kind both arrived through the generic paths |

**One thing the rig could not prove: `bridge.busy`.** With today's synchronous handlers a repeat can
never arrive while the original is still running, because the Core thread processes one inbound line
at a time — so the state is unreachable on a live shard until a handler defers. It is implemented,
`BridgeIdempotency.Hold`/`Complete` are the door a deferring handler uses, and the sidecar's 425
mapping is unit-tested on all three responders. **Phase 11b's leases are the first thing that can
actually produce it**, and proving it belongs in that walk rather than being claimed here.

### 6.1 The rig trap: the innermost region has no name

Recorded because it is the sort of defect that ships, and because the naive check *confirms* it.

The first emitter read `boss.Region` and a champion killed in the middle of Britain produced a frame
with **no region at all**. Three things were true at once:

- An active `ChampionSpawn` registers a `ChampionSpawnRegion` over its spawn area, constructed with
  a **null name** and with the town region as its `Parent`. So the most specific region containing a
  champion boss is, by construction, the one region on the map guaranteed to have no name.
- `Mobile.Region` never returns null: with no cached region it falls back to the map's **unnamed
  default region**. So the failure looks identical to "outdoors, nowhere in particular".
- `ChampionSpawn.SpawnChampion` places its boss at `Z - 15`, which was a plausible-looking second
  explanation and a red herring. It cost a rig cycle.

And the reason it survived a first look: **region registration is deferred**, so a lookup taken
immediately after the altar is placed still answers `"Britain"` and one taken at the kill twenty
seconds later does not. The probe printed `region "Britain"` at spawn on every run, including the
ones whose frame carried nothing. A read at spawn time would have confirmed the bug into the design.

The fix is general rather than a special case for champions: walk outward to the nearest **named**
ancestor. A house region, a dungeon sub-region and a guarded-zone overlay are all anonymous children
of somewhere a player would name. The map's own default region terminates the walk with its
parentless empty name, so open countryside answers null — and the field is then **omitted**, because
a consumer reading `region: ""` cannot tell "nowhere in particular" from "the shard would not say".

### 6.2 Two smaller ones

- **`deploy.ps1` overwrites `Bridge.cfg`,** so every deploy strips the scaffolding flags —
  `Protocol6ProbeOnStart` and, here, `AdminWriteEnabled` as well. The probe then silently does
  nothing on the next boot. The scaffolding README warns about this and it still cost a cycle.
- **A running ServUO holds `Scripts.dll`,** so an offline `msbuild` fails with `MSB3027` rather than
  a compile error. Stop the shard, build, restart — and read the boot log's `0 Error(s)` to confirm
  the new code is live rather than trusting a clean boot, which is what
  `servuo-plugin-build-and-run` already records.

---

## 7. What 11b adds to protocol 6

Phase 11 is split. **This document covers 11a**; 11b adds lease deadlines and the run-scoped
participation ledger to the *same* protocol version, amended in place rather than bumped to 7 —
6 will not have landed on `main` until the events cutover, and the org lead's 2026-09-03 rule is that
a protocol owes a bump once it has shipped and is amended in place before that.

Which means: **an overlay and a sidecar both declaring `6` are only interchangeable within one side
of the 11b merge.** That is tolerable exactly because nothing is released from `edge` — the bundle
CI never sees either half until the cutover, by which time 6 means one thing. It would not be
tolerable on `main`, and this paragraph exists so nobody discovers that the hard way.

What 11b will add here:

- `lease.apply` / `lease.release` / `lease.list`, with a deadline the shard honours **without being
  asked again** and a compare-and-set restore that reports `drifted` rather than overwriting a GM's
  deliberate change.
- `participation.open` / `participation.snapshot` / `participation.close`, keyed by character serial
  to match `module-uo`'s existing `memberKey`, and **persisted in the world save** — the Bridge's
  first persisted state, so a tally survives a restart mid-event.
- The first handlers that **defer**, and therefore the first that can actually answer `bridge.busy`.