feat(x): reference-based model requests + wire-form composer

model_call_requested carried three duplications that dominated turn-file
size (measured on a real 647KB two-step turn): the system prompt (28KB)
and the 38-tool snapshot (40KB) repeated per call despite being byte-
identical to turn_created.agent.resolved, and messages re-inlining the
whole current-turn transcript (230KB re-stating a 207KB tool result).

ModelRequest is now a list of references into the turn's own events —
call 0: [{context}?, {input}]; call N: [{assistant: N-1}, ...that
batch's toolResults in source order]; re-issue after an interruption:
[]. Every referenced byte exists exactly once in the file; the measured
turn drops to ~285KB and each further model call costs ~200 bytes. The
reducer's ordering invariant got stricter: reference lists are matched
exactly against the transcript.

Debuggability is now byte-for-byte at the wire level: ResolvedModel
gains encodeMessages (the structural->wire conversion — user-message
context weaving, attachment rendering, tool-result enveloping, i.e.
convertFromMessages), and composeModelRequest rebuilds the exact
provider payload (resolved system prompt + wrapped tools + materialized
prefix + resolved refs, encoded). The loop transmits exactly the
composer's output, so the durable file plus the composer reproduce what
the model received — pinned by a property test asserting composed ==
sent for every call, and a 2KB size guard on request events. A bridge
test demonstrates the woven wire form (no raw userMessageContext).

Breaking for existing dev turn files (same schemaVersion, pre-release):
wipe ~/.rowboat/storage.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

apps/x/apps/renderer/src/hooks/useSessionChat.test.tsx

@@ -82,7 +82,7 @@ describe('useSessionChat', () => {
     // A new turn streams in over the feed.
     act(() => {
       emit({ kind: 'turn-event', sessionId: S1, turnId: 'turn-2', event: created('turn-2', S1, user('q2')) })
-      emit({ kind: 'turn-event', sessionId: S1, turnId: 'turn-2', event: requested('turn-2', 0, [user('q2')]) })
+      emit({ kind: 'turn-event', sessionId: S1, turnId: 'turn-2', event: requested('turn-2', 0) })
       emit({
         kind: 'turn-event',
         sessionId: S1,

apps/x/apps/renderer/src/lib/session-chat/store.test.ts

@@ -143,7 +143,7 @@ describe('SessionChatStore', () => {
     await store.setSession(S1)
 
     emit(turnEvent(S1, 'turn-1', created('turn-1', S1, user('go'))))
-    emit(turnEvent(S1, 'turn-1', requested('turn-1', 0, [user('go')])))
+    emit(turnEvent(S1, 'turn-1', requested('turn-1', 0)))
     expect(store.getSnapshot().chatState?.isThinking).toBe(true)
 
     emit(turnEvent(S1, 'turn-1', completed('turn-1', 0, assistantText('done'))))
@@ -161,7 +161,7 @@ describe('SessionChatStore', () => {
     client.sessions.set(S1, sessionState(S1, []))
     await store.setSession(S1)
     emit(turnEvent(S1, 'turn-1', created('turn-1', S1, user('go'))))
-    emit(turnEvent(S1, 'turn-1', requested('turn-1', 0, [user('go')])))
+    emit(turnEvent(S1, 'turn-1', requested('turn-1', 0)))
     emit(turnEvent(S1, 'turn-1', { type: 'text_delta', turnId: 'turn-1', modelCallIndex: 0, delta: 'he' }))
     emit(turnEvent(S1, 'turn-1', { type: 'text_delta', turnId: 'turn-1', modelCallIndex: 0, delta: 'y' }))
     expect(store.getSnapshot().chatState?.currentAssistantMessage).toBe('hey')
@@ -199,7 +199,7 @@ describe('SessionChatStore', () => {
     // The feed attached mid-turn: we never saw turn_created for turn-9.
     client.turns.set('turn-9', [
       created('turn-9', S1, user('missed')),
-      requested('turn-9', 0, [user('missed')]),
+      requested('turn-9', 0),
     ])
     emit(turnEvent(S1, 'turn-9', completed('turn-9', 0, assistantText('caught up'))))
     await Promise.resolve()
@@ -273,7 +273,7 @@ describe('SessionChatStore', () => {
 
     await store.setSession(S1)
     emit(turnEvent(S1, 'turn-1', created('turn-1', S1, user('go'))))
-    emit(turnEvent(S1, 'turn-1', requested('turn-1', 0, [user('go')])))
+    emit(turnEvent(S1, 'turn-1', requested('turn-1', 0)))
     emit(turnEvent(S1, 'turn-1', completed('turn-1', 0, assistantText('done'))))
     emit(turnEvent(S1, 'turn-1', turnCompleted('turn-1')))
     expect(store.getSnapshot().chatState?.isProcessing).toBe(false)

apps/x/apps/renderer/src/lib/session-chat/test-fixtures.ts

@@ -46,16 +46,24 @@ export function created(
   }
 }
 
-export function requested(turnId: string, index: number, messages: unknown[]): TEvent {
+type RequestRef =
+  | { kind: 'context' }
+  | { kind: 'input' }
+  | { kind: 'assistant'; modelCallIndex: number }
+  | { kind: 'toolResult'; toolCallId: string }
+
+export function requested(
+  turnId: string,
+  index: number,
+  refs: RequestRef[] = [{ kind: 'input' }],
+): TEvent {
   return {
     type: 'model_call_requested',
     turnId,
     ts: TS,
     modelCallIndex: index,
     request: {
-      systemPrompt: 'SYS',
-      messages: messages as never,
-      tools: [],
+      messages: refs,
       parameters: {},
     },
   }
@@ -127,7 +135,7 @@ export function completedTurnLog(
 ): TEvent[] {
   return [
     created(turnId, sessionId, user(question)),
-    requested(turnId, 0, [user(question)]),
+    requested(turnId, 0),
     completed(turnId, 0, assistantText(answer)),
     turnCompleted(turnId, answer),
   ]

apps/x/apps/renderer/src/lib/session-chat/turn-view.test.ts

@@ -87,14 +87,13 @@ describe('buildTurnConversation', () => {
     const call = assistantCalls(toolCallPart('tc1', 'echo', { x: 1 }))
     const state = reduceTurn([
       created(T1, S1, user('run it')),
-      requested(T1, 0, [user('run it')]),
+      requested(T1, 0),
       completed(T1, 0, call),
       invocation(T1, 'tc1', 'echo'),
       toolResult(T1, 'tc1', 'echo', { echoed: true }),
       requested(T1, 1, [
-        user('run it'),
-        call,
-        { role: 'tool', content: '{"echoed":true}', toolCallId: 'tc1', toolName: 'echo' },
+        { kind: 'assistant', modelCallIndex: 0 },
+        { kind: 'toolResult', toolCallId: 'tc1' },
       ]),
       completed(T1, 1, assistantText('all done')),
       turnCompleted(T1, 'all done'),
@@ -113,7 +112,7 @@ describe('buildTurnConversation', () => {
   it('marks permission-pending tools pending and running tools running', () => {
     const state = reduceTurn([
       created(T1, S1),
-      requested(T1, 0, [user('hello')]),
+      requested(T1, 0),
       completed(T1, 0, assistantCalls(toolCallPart('p1', 'executeCommand'), toolCallPart('r1', 'echo'))),
       {
         type: 'tool_permission_required',
@@ -139,7 +138,7 @@ describe('buildTurnConversation', () => {
     }
     const state = reduceTurn([
       created(T1, S1, input as never),
-      requested(T1, 0, [input]),
+      requested(T1, 0),
       { type: 'model_call_failed', turnId: T1, ts: TS, modelCallIndex: 0, error: 'boom' },
       { type: 'turn_failed', turnId: T1, ts: TS, error: 'boom', usage: {} },
     ])
@@ -154,7 +153,7 @@ describe('buildSessionChatState', () => {
     const prior = reduceTurn(completedTurnLog('turn-1', S1, 'first?', 'first answer'))
     const latest = reduceTurn([
       created('turn-2', S1, user('second?')),
-      requested('turn-2', 0, [user('second?')]),
+      requested('turn-2', 0),
     ])
     const state = buildSessionChatState([prior, latest], emptyOverlay())
     expect(
@@ -174,7 +173,7 @@ describe('buildSessionChatState', () => {
   it('exposes pending permissions as request events; waiting is not thinking', () => {
     const turn = reduceTurn([
       created(T1, S1),
-      requested(T1, 0, [user('hello')]),
+      requested(T1, 0),
       completed(T1, 0, assistantCalls(toolCallPart('p1', 'executeCommand', { command: 'rm -rf /' }))),
       {
         type: 'tool_permission_required',
@@ -209,7 +208,7 @@ describe('buildSessionChatState', () => {
   it('maps human decisions to responses and classifier decisions to auto-decisions', () => {
     const turn = reduceTurn([
       created(T1, S1),
-      requested(T1, 0, [user('hello')]),
+      requested(T1, 0),
       completed(T1, 0, assistantCalls(toolCallPart('h1', 'echo'), toolCallPart('c1', 'echo'))),
       { type: 'tool_permission_required', turnId: T1, ts: TS, toolCallId: 'h1', toolName: 'echo', request: { kind: 'command', commandNames: ['x'] } },
       { type: 'tool_permission_required', turnId: T1, ts: TS, toolCallId: 'c1', toolName: 'echo', request: { kind: 'command', commandNames: ['y'] } },
@@ -230,7 +229,7 @@ describe('buildSessionChatState', () => {
   it('exposes pending ask-human calls with question and options', () => {
     const turn = reduceTurn([
       created(T1, S1),
-      requested(T1, 0, [user('hello')]),
+      requested(T1, 0),
       completed(T1, 0, assistantCalls(toolCallPart('ah1', 'ask-human', { question: 'Deploy?', options: ['Yes', 'No'] }))),
       {
         type: 'tool_invocation_requested',
@@ -254,7 +253,7 @@ describe('buildSessionChatState', () => {
   it('stitches live tool output onto the matching tool item', () => {
     const turn = reduceTurn([
       created(T1, S1),
-      requested(T1, 0, [user('hello')]),
+      requested(T1, 0),
       completed(T1, 0, assistantCalls(toolCallPart('tc1', 'executeCommand'))),
       invocation(T1, 'tc1', 'executeCommand'),
     ])
@@ -265,7 +264,7 @@ describe('buildSessionChatState', () => {
   })
 
   it('surfaces streaming text as currentAssistantMessage', () => {
-    const turn = reduceTurn([created(T1, S1), requested(T1, 0, [user('hello')])])
+    const turn = reduceTurn([created(T1, S1), requested(T1, 0)])
     const state = buildSessionChatState([turn], { ...emptyOverlay(), text: 'typing…' })
     expect(state.currentAssistantMessage).toBe('typing…')
   })