feat: stream Reginald's replies token-by-token

The 26b is slow (~30s/call); the chat now shows the answer forming instead of
freezing until it's done. The final prose streams over SSE; tool-calling turns
stay structured (no partial tokens), so streaming kicks in for the narration.

core (@commitea/core):
- chat-client.complete gains an optional onToken — when set, it requests
  stream:true and parses the OpenAI SSE stream, emitting content deltas and
  assembling streamed tool-call argument fragments into the final result.
- GiteaHttpResponse exposes the optional `body` stream (real fetch has it; stubs
  don't). agent-loop threads onToken to each completion.

app:
- model:chat forwards each delta to the renderer (event.sender.send); preload
  exposes model.onToken(cb) → unsubscribe. useChat accumulates the live stream
  into a growing bubble (with a cursor), replaced by the authoritative final
  content when the turn resolves. Unconfigured → scripted reply, unchanged.

Verified: 118 core tests green (2 streaming: SSE content deltas + tool-call
fragment assembly), desktop typecheck clean, 14 fixture e2e green. Live: a real
turn against gemma-4-26b assembles the correct answer via the streaming path
(live-reginald green) — the reply now renders token-by-token.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

apps/desktop/e2e/live-persistence.spec.ts

@@ -0,0 +1,37 @@
+import { dirname, join } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+import { _electron as electron, expect, test } from '@playwright/test'
+
+const here = dirname(fileURLToPath(import.meta.url))
+const MAIN = join(here, '..', 'out', 'main', 'index.js')
+
+// Opt-in (GITEA_LIVE=1). A real launch persists the snapshot; a second launch with
+// gitea unreachable must still show the board + real scheduler output from the
+// persisted cache (offline reads). No model needed.
+test.describe('live persistence', () => {
+  test('offline: serves the persisted snapshot', async () => {
+    test.skip(!process.env.GITEA_LIVE, 'GITEA_LIVE not set — opt-in live test')
+    test.setTimeout(120_000)
+
+    // launch 1 — real reconcile writes the snapshot to disk
+    const app1 = await electron.launch({ args: [MAIN], env: { ...process.env } })
+    const w1 = await app1.firstWindow()
+    await w1.waitForLoadState('domcontentloaded')
+    // a scheduler-only phrase confirms real data reconciled (never emitted by fixtures)
+    await expect(w1.getByText(/on the critical path|unblocks #|waits on #|· ready/).first()).toBeVisible({ timeout: 30000 })
+    await app1.close()
+
+    // launch 2 — gitea unreachable; the reconcile must fall back to the persisted snapshot
+    const app2 = await electron.launch({
+      args: [MAIN],
+      env: { ...process.env, GITEA_BASE_URL: 'http://127.0.0.1:9' },
+    })
+    const w2 = await app2.firstWindow()
+    await w2.waitForLoadState('domcontentloaded')
+    // Focus still renders real scheduler output — proving it came from the cache, offline
+    await expect(w2.getByText(/on the critical path|unblocks #|waits on #|· ready/).first()).toBeVisible({ timeout: 20000 })
+    await w2.screenshot({ path: join(here, '.artifacts', 'screens', 'live-offline.png'), fullPage: true, animations: 'disabled' })
+    await app2.close()
+  })
+})

apps/desktop/src/main/gitea.ts

@@ -27,6 +27,8 @@ import {
 } from '@commitea/core'
 import { ipcMain } from 'electron'
 
+import { loadSnapshot, saveSnapshot } from './snapshot-store.js'
+
 /** Walk up from cwd looking for a .env.local with a GITEA_TOKEN (dev convenience). */
 function loadEnvLocalToken(): string | undefined {
   let dir = process.cwd()
@@ -126,16 +128,81 @@ export async function reconcileSnapshot(
   return { issues, milestones, deps, timelines }
 }
 
+type Snapshot = Awaited<ReturnType<typeof reconcileSnapshot>>
+
+/**
+ * A single in-memory reconcile cache shared across the app. A full reconcile is
+ * ~2N gitea calls (deps + timelines per issue); without this, every agent tool
+ * call refetched the whole repo. Reads within `maxAgeMs` reuse the cache;
+ * `getSnapshot({ maxAgeMs: 0 })` forces a fresh pull (the explicit UI reconcile),
+ * and any write calls `invalidateSnapshot()` so the next read sees it. The cache
+ * is rebuildable — the durable truth stays in gitea (the purity split, D4).
+ */
+let snapshotCache: { snap: Snapshot; at: number } | null = null
+
+export async function getSnapshot(client: GiteaClient, opts?: { maxAgeMs?: number }): Promise<Snapshot> {
+  const maxAgeMs = opts?.maxAgeMs ?? 0
+  if (snapshotCache && maxAgeMs > 0 && Date.now() - snapshotCache.at <= maxAgeMs) {
+    return snapshotCache.snap
+  }
+  const snap = await reconcileSnapshot(client)
+  snapshotCache = { snap, at: Date.now() }
+  saveSnapshot(snap, new Date().toISOString()) // persist for instant boot + offline
+  return snap
+}
+
+/** Drop the cache so the next read reflects a just-made write. */
+export function invalidateSnapshot(): void {
+  snapshotCache = null
+}
+
+/**
+ * The last persisted snapshot (from a previous session), for instant boot. The
+ * renderer shows it immediately, then a real reconcile supersedes it
+ * (stale-while-revalidate). Returns null when there's nothing on disk; its
+ * `savedAt` marks staleness. It does NOT seed the cache — agent tool calls
+ * always reconcile fresh so they never reason over stale data.
+ */
+export function bootSnapshot(): (Snapshot & { savedAt: string }) | null {
+  const persisted = loadSnapshot()
+  if (!persisted) return null
+  return {
+    issues: persisted.issues,
+    milestones: persisted.milestones,
+    deps: persisted.deps,
+    timelines: persisted.timelines,
+    savedAt: persisted.savedAt,
+  } as unknown as Snapshot & { savedAt: string }
+}
+
+/** Agent tool calls tolerate a slightly stale snapshot (seconds) to stay responsive. */
+export const AGENT_SNAPSHOT_TTL_MS = 30_000
+
 export function registerGiteaIpc(): void {
   const client = getGiteaClient()
   const repo = client ? `${process.env.GITEA_OWNER ?? 'christian'}/${process.env.GITEA_REPO ?? 'commitea'}` : null
 
   ipcMain.handle('gitea:status', () => ({ configured: !!client, repo }))
 
+  // Instant boot: the last persisted snapshot, shown before the fresh reconcile lands.
+  ipcMain.handle('gitea:boot', () => {
+    if (!client) return { configured: false }
+    const persisted = bootSnapshot()
+    return persisted ? { configured: true, cached: true, ...persisted } : { configured: true, cached: false }
+  })
+
   ipcMain.handle('gitea:reconcile', async () => {
     if (!client) return { configured: false, issues: [], milestones: [], deps: [], timelines: {} }
-    const snap = await reconcileSnapshot(client)
-    return { configured: true, ...snap }
+    try {
+      // explicit UI sync — force fresh, and warm the cache for agent tool calls
+      const snap = await getSnapshot(client, { maxAgeMs: 0 })
+      return { configured: true, stale: false, ...snap }
+    } catch (e) {
+      // offline / gitea down — serve the last persisted snapshot rather than error out
+      const persisted = bootSnapshot()
+      if (persisted) return { configured: true, stale: true, ...persisted }
+      throw e
+    }
   })
 
   ipcMain.handle('gitea:getIssue', async (_event, index: number) => {
@@ -167,6 +234,7 @@ export function registerGiteaIpc(): void {
     const ids = await resolveLabelIds(plan.labels)
     await client.setIssueLabels(change.issue, ids)
     const issue = await client.getIssue(change.issue)
+    invalidateSnapshot() // the board + forecast must reflect the label change
     return { ok: true as const, plan, issue }
   })
 
@@ -183,6 +251,7 @@ export function registerGiteaIpc(): void {
         const issue = await client.createIssue({ title: it.title, body: it.body, labelIds })
         created.push({ number: issue.number, title: issue.title })
       }
+      if (created.length) invalidateSnapshot() // new issues enter the board/scope
       return { ok: true as const, created }
     },
   )

apps/desktop/src/main/model.ts

@@ -26,7 +26,13 @@ import {
 } from '@commitea/core'
 import { ipcMain } from 'electron'
 
-import { appendDirectiveEntry, getGiteaClient, getPmStateClient, reconcileSnapshot } from './gitea.js'
+import {
+  AGENT_SNAPSHOT_TTL_MS,
+  appendDirectiveEntry,
+  getGiteaClient,
+  getPmStateClient,
+  getSnapshot,
+} from './gitea.js'
 
 /** Small local model for prose + the read tool; big model reserved for later decomposition. */
 function resolveModelRouter(): ModelRouter | null {
@@ -80,11 +86,15 @@ export function registerModelIpc(): void {
     return { configured: true, model }
   })
 
-  ipcMain.handle('model:chat', async (_event, messages: ChatMessage[]) => {
+  ipcMain.handle('model:chat', async (event, messages: ChatMessage[]) => {
     if (!router) return { ok: false as const, reason: 'unconfigured' as const }
     const client = getGiteaClient()
     const model = await resolveLoadedModel(router.small.baseUrl, router.small.model)
     const chat = createChatClient({ ...router.small, model }, fetch)
+    // stream the model's prose to the renderer token-by-token
+    const onToken = (delta: string) => {
+      if (!event.sender.isDestroyed()) event.sender.send('model:chat:token', delta)
+    }
 
     // Proposals the model formulates this turn; the renderer approves them (the
     // write happens through gitea:applyChange, never inside the loop).
@@ -93,7 +103,8 @@ export function registerModelIpc(): void {
     const execute = async (name: string, args: unknown) => {
       if (!client) return { error: 'gitea is not configured' }
       if (name === 'query_project') {
-        const snap = await reconcileSnapshot(client)
+        // reuse a recent reconcile — a multi-tool turn shouldn't refetch the repo each call
+        const snap = await getSnapshot(client, { maxAgeMs: AGENT_SNAPSHOT_TTL_MS })
         const a = (args ?? {}) as { view: ProjectView; filters?: QueryFilters }
         return buildProjectView(a.view, a.filters, snap, new Date())
       }
@@ -122,10 +133,11 @@ export function registerModelIpc(): void {
 
     try {
       const turn = await runAgentTurn({
-        complete: (m, t) => chat.complete(m, t),
+        complete: (m, t, ot) => chat.complete(m, t, ot),
         messages: [{ role: 'system', content: REGINALD_SYSTEM }, ...messages],
         tools: REGINALD_TOOLS,
         execute,
+        onToken,
       })
       return { ok: true as const, content: turn.content, steps: turn.steps, proposals }
     } catch (e) {

apps/desktop/src/main/snapshot-store.ts

@@ -0,0 +1,47 @@
+/**
+ * Durable snapshot store — the reconcile cache, persisted to disk. On boot the
+ * app shows the last snapshot instantly (stale-while-revalidate) instead of a
+ * blank board while ~2N gitea calls run; if gitea is unreachable, reads fall
+ * back to it (offline). It's a rebuildable mirror — the durable truth stays in
+ * gitea (the purity split, D4). A plain JSON file: the whole snapshot fits in
+ * memory at this scale, so indexed SQL buys nothing yet (see the PR).
+ */
+
+import { readFileSync, writeFileSync } from 'node:fs'
+import { join } from 'node:path'
+
+import { app } from 'electron'
+
+/** The shape we persist — kept loose so a schema drift degrades to "no cache", not a crash. */
+export interface PersistedSnapshot {
+  issues: unknown[]
+  milestones: unknown[]
+  deps: unknown[]
+  timelines: Record<number, unknown[]>
+  /** ISO time the snapshot was reconciled — shown as "cached since". */
+  savedAt: string
+}
+
+function snapshotPath(): string {
+  return join(app.getPath('userData'), 'commitea-snapshot.json')
+}
+
+/** Load the last persisted snapshot, or null if absent/corrupt. Never throws. */
+export function loadSnapshot(): PersistedSnapshot | null {
+  try {
+    const parsed = JSON.parse(readFileSync(snapshotPath(), 'utf8')) as PersistedSnapshot
+    if (parsed && Array.isArray(parsed.issues)) return parsed
+    return null
+  } catch {
+    return null // missing file, bad JSON, or drift — treat as no cache
+  }
+}
+
+/** Persist a freshly reconciled snapshot. Best-effort — a write failure never breaks a reconcile. */
+export function saveSnapshot(snap: Omit<PersistedSnapshot, 'savedAt'>, savedAt: string): void {
+  try {
+    writeFileSync(snapshotPath(), JSON.stringify({ ...snap, savedAt }), 'utf8')
+  } catch {
+    // disk full / permissions — the in-memory cache still works this session
+  }
+}

apps/desktop/src/preload/index.ts

@@ -5,6 +5,8 @@ const api = {
   gitea: {
     /** Whether the main process has a gitea token + target repo configured. */
     status: () => ipcRenderer.invoke('gitea:status'),
+    /** The last persisted snapshot, for instant boot before the fresh reconcile. */
+    boot: () => ipcRenderer.invoke('gitea:boot'),
     /** Full read of the managed repo — every issue + milestone. */
     reconcile: () => ipcRenderer.invoke('gitea:reconcile'),
     /** One issue by index, normalized (or null if unconfigured). */
@@ -23,6 +25,12 @@ const api = {
     status: () => ipcRenderer.invoke('model:status'),
     /** One agent turn: messages in, Reginald's prose + the tools it consulted out. */
     chat: (messages: unknown) => ipcRenderer.invoke('model:chat', messages),
+    /** Subscribe to streamed prose tokens for the in-flight turn; returns an unsubscribe. */
+    onToken: (cb: (delta: string) => void) => {
+      const listener = (_e: unknown, delta: string) => cb(delta)
+      ipcRenderer.on('model:chat:token', listener)
+      return () => ipcRenderer.removeListener('model:chat:token', listener)
+    },
     /** Decompose a braindump into a proposed issue set (capture_work). */
     capture: (braindump: string) => ipcRenderer.invoke('model:capture', braindump),
   },

apps/desktop/src/renderer/src/components/shell/chat-panel.tsx

@@ -19,7 +19,7 @@ export interface ChatPanelProps {
 }
 
 export function ChatPanel({ onOpenDirectives, offline, onApplyChange }: ChatPanelProps) {
-  const { msgs, thinking, live, model, steps, proposals, send: sendChat, approve, dismiss } = useChat(onApplyChange)
+  const { msgs, thinking, live, model, steps, proposals, streaming, send: sendChat, approve, dismiss } = useChat(onApplyChange)
   // shorten "google/gemma-4-26b-a4b-qat" → "gemma-4-26b" for the header chip
   const modelLabel = model ? (model.split('/').pop() ?? model).replace(/-(qat|instruct|it|gguf)$/i, '') : 'gemma-4'
   const [text, setText] = useState('')
@@ -28,7 +28,7 @@ export function ChatPanel({ onOpenDirectives, offline, onApplyChange }: ChatPane
   useEffect(() => {
     const el = scrollRef.current
     if (el) el.scrollTop = el.scrollHeight
-  }, [msgs, thinking])
+  }, [msgs, thinking, streaming])
 
   const send = () => {
     const t = text.trim()
@@ -98,7 +98,14 @@ export function ChatPanel({ onOpenDirectives, offline, onApplyChange }: ChatPane
             The model is away from its desk. Reads still work; writes will wait their turn.
           </div>
         ) : null}
-        {thinking ? <div style={{ font: 'var(--text-agent)', color: 'var(--ink-3)' }}>considering…</div> : null}
+        {streaming ? (
+          <div style={{ font: 'var(--text-agent)', color: 'var(--ink-1)', lineHeight: 1.55 }}>
+            {streaming}
+            <span style={{ opacity: 0.5 }}>▊</span>
+          </div>
+        ) : thinking ? (
+          <div style={{ font: 'var(--text-agent)', color: 'var(--ink-3)' }}>considering…</div>
+        ) : null}
         {!thinking && steps.length ? (
           <div style={{ font: 'var(--text-caption)', color: 'var(--ink-3)', display: 'flex', alignItems: 'center', gap: 5 }}>
             <Icon name="eye" size={11} /> consulted {Array.from(new Set(steps.map((s) => s.replace('query_project', 'the project').replace('propose_change', 'the labels').replace('record_directive', 'the directive ledger')))).join(', ')}

apps/desktop/src/renderer/src/global.d.ts

@@ -28,17 +28,31 @@ export type CaptureResult =
   | { ok: false; reason: 'unconfigured' | 'error'; message?: string }
   | ({ ok: true } & CaptureProposal)
 
-/** The gitea bridge exposed by the preload over IPC (main-process backed). */
-export interface GiteaBridge {
-  status(): Promise<{ configured: boolean; repo: string | null }>
-  reconcile(): Promise<{
+/** A reconciled snapshot as it crosses the bridge. */
+export interface SnapshotPayload {
   configured: boolean
   issues: GiteaIssue[]
   milestones: GiteaMilestone[]
   deps: DependencyEdge[]
   /** Normalized lifecycle events keyed by issue number. */
   timelines: Record<number, LifecycleEvent[]>
-  }>
+  /** true when served from the persisted cache (offline / instant boot). */
+  stale?: boolean
+  /** ISO time the persisted snapshot was reconciled (present on cached reads). */
+  savedAt?: string
+}
+
+/** Boot payload — the persisted snapshot, or a marker that there's none yet. */
+export type BootPayload =
+  | { configured: false }
+  | { configured: true; cached: false }
+  | ({ configured: true; cached: true } & Omit<SnapshotPayload, 'configured'>)
+
+/** The gitea bridge exposed by the preload over IPC (main-process backed). */
+export interface GiteaBridge {
+  status(): Promise<{ configured: boolean; repo: string | null }>
+  boot(): Promise<BootPayload>
+  reconcile(): Promise<SnapshotPayload>
   getIssue(index: number): Promise<GiteaIssue | null>
   applyChange(change: IssueChange): Promise<ApplyChangeResult>
   createIssues(issues: ProposedIssue[]): Promise<CreateIssuesResult>
@@ -54,6 +68,8 @@ export interface ModelBridge {
   status(): Promise<{ configured: boolean; model: string | null }>
   chat(messages: ChatMessage[]): Promise<ChatResult>
   capture(braindump: string): Promise<CaptureResult>
+  /** Subscribe to streamed prose tokens; returns an unsubscribe fn. */
+  onToken(cb: (delta: string) => void): () => void
 }
 
 /** The result of reading the directive ledger. */

apps/desktop/src/renderer/src/lib/use-backlog.ts

@@ -12,13 +12,18 @@ export type BacklogState =
       milestones: GiteaMilestone[]
       deps: DependencyEdge[]
       timelines: Record<number, LifecycleEvent[]>
+      /** true while showing the persisted snapshot (instant boot / offline). */
+      stale: boolean
+      /** ISO time the shown snapshot was reconciled, when stale. */
+      savedAt?: string
     }
 
 /**
- * Reconcile the managed repo through the main-process bridge. Runs once on
- * mount; the returned `refetch` re-reconciles after a write so the board and
- * forecast reflect the change. `unconfigured` means no token — the UI falls
- * back to demo fixtures. Errors (network, bad token) surface as `error`.
+ * Reconcile the managed repo through the main-process bridge, stale-while-
+ * revalidate: on mount it shows the persisted snapshot instantly (marked stale),
+ * then a fresh reconcile supersedes it. If gitea is unreachable, the fresh
+ * reconcile falls back to the persisted snapshot (offline). `refetch` re-syncs
+ * after a write. `unconfigured` means no token — the UI uses demo fixtures.
  */
 export function useBacklog(): [BacklogState, () => void] {
   const [state, setState] = useState<BacklogState>({ status: 'loading' })
@@ -27,6 +32,30 @@ export function useBacklog(): [BacklogState, () => void] {
 
   useEffect(() => {
     let alive = true
+
+    // instant boot from the persisted snapshot (only on first mount, not refetch)
+    if (nonce === 0) {
+      window.commitea.gitea
+        .boot()
+        .then((b) => {
+          if (!alive || !('cached' in b) || !b.cached) return
+          setState((prev) =>
+            prev.status === 'ready' && !prev.stale
+              ? prev // a fresh reconcile already won the race
+              : {
+                  status: 'ready',
+                  issues: b.issues,
+                  milestones: b.milestones,
+                  deps: b.deps,
+                  timelines: b.timelines,
+                  stale: true,
+                  savedAt: b.savedAt,
+                },
+          )
+        })
+        .catch(() => {})
+    }
+
     window.commitea.gitea
       .reconcile()
       .then((r) => {
@@ -39,12 +68,19 @@ export function useBacklog(): [BacklogState, () => void] {
                 milestones: r.milestones,
                 deps: r.deps,
                 timelines: r.timelines,
+                stale: r.stale ?? false,
+                savedAt: r.savedAt,
               }
             : { status: 'unconfigured' },
         )
       })
       .catch((e: unknown) => {
-        if (alive) setState({ status: 'error', message: e instanceof Error ? e.message : String(e) })
+        if (alive) {
+          // keep a shown boot snapshot rather than clobbering it with an error
+          setState((prev) =>
+            prev.status === 'ready' ? prev : { status: 'error', message: e instanceof Error ? e.message : String(e) },
+          )
+        }
       })
     return () => {
       alive = false

apps/desktop/src/renderer/src/lib/use-chat.ts

@@ -20,6 +20,8 @@ export interface ChatState {
   steps: string[]
   /** Changes Reginald has proposed and is awaiting approval on. */
   proposals: ChangeProposal[]
+  /** The in-flight streamed prose (grows token-by-token) before the turn finalizes. */
+  streaming: string
   send: (text: string) => void
   approve: (p: ChangeProposal) => void
   dismiss: (p: ChangeProposal) => void
@@ -41,6 +43,7 @@ export function useChat(onApplyChange?: (change: IssueChange) => Promise<{ ok: b
   const [model, setModel] = useState<string | null>(null)
   const [steps, setSteps] = useState<string[]>([])
   const [proposals, setProposals] = useState<ChangeProposal[]>([])
+  const [streaming, setStreaming] = useState('')
   const convoRef = useRef(convo)
   convoRef.current = convo
 
@@ -86,10 +89,17 @@ export function useChat(onApplyChange?: (change: IssueChange) => Promise<{ ok: b
         role: m.from === 'user' ? 'user' : 'assistant',
         content: m.text,
       }))
+      setStreaming('')
+      const unsubscribe = window.commitea.model.onToken((delta) => setStreaming((s) => s + delta))
+      const finish = () => {
+        unsubscribe()
+        setThinking(false)
+        setStreaming('')
+      }
       window.commitea.model
         .chat(wire)
         .then((res) => {
-          setThinking(false)
+          finish()
           if (res.ok) {
             setSteps(res.steps.map((s) => s.tool))
             setProposals(res.proposals)
@@ -105,7 +115,7 @@ export function useChat(onApplyChange?: (change: IssueChange) => Promise<{ ok: b
           }
         })
         .catch(() => {
-          setThinking(false)
+          finish()
           setConvo((c) => [...c, { from: 'agent', text: 'I could not reach the model.' }])
         })
     },
@@ -137,5 +147,5 @@ export function useChat(onApplyChange?: (change: IssueChange) => Promise<{ ok: b
     setConvo((c) => [...c, { from: 'agent', text: `Left #${p.change.issue} as it was.` }])
   }, [])
 
-  return { msgs: [...seed, ...convo], thinking, live, model, steps, proposals, send, approve, dismiss }
+  return { msgs: [...seed, ...convo], thinking, live, model, steps, proposals, streaming, send, approve, dismiss }
 }

packages/core/src/agent/agent-loop.ts

@@ -31,18 +31,20 @@ function stringify(result: unknown): string {
 }
 
 export async function runAgentTurn(opts: {
-  complete: (messages: ChatMessage[], tools?: ToolDecl[]) => Promise<CompletionResult>
+  complete: (messages: ChatMessage[], tools?: ToolDecl[], onToken?: (delta: string) => void) => Promise<CompletionResult>
   messages: ChatMessage[]
   tools: ToolDecl[]
   execute: ToolExecutor
   maxSteps?: number
+  /** Streams content deltas as the model produces prose (final-answer streaming). */
+  onToken?: (delta: string) => void
 }): Promise<AgentTurn> {
   const maxSteps = opts.maxSteps ?? DEFAULT_MAX_STEPS
   const convo: ChatMessage[] = [...opts.messages]
   const steps: AgentStep[] = []
 
   for (let step = 0; step < maxSteps; step++) {
-    const { content, toolCalls } = await opts.complete(convo, opts.tools)
+    const { content, toolCalls } = await opts.complete(convo, opts.tools, opts.onToken)
     if (toolCalls.length === 0) {
       convo.push({ role: 'assistant', content })
       return { content, steps, messages: convo }
@@ -63,7 +65,7 @@ export async function runAgentTurn(opts: {
   }
 
   // Out of tool budget — force a final prose answer with tools withheld.
-  const final = await opts.complete(convo, [])
+  const final = await opts.complete(convo, [], opts.onToken)
   convo.push({ role: 'assistant', content: final.content })
   return { content: final.content, steps, messages: convo }
 }

packages/core/src/agent/agent.test.ts

@@ -25,6 +25,48 @@ describe('createChatClient', () => {
     return { fetch, calls }
   }
 
+  function sseStub(chunks: string[]) {
+    const calls: { url: string; body: unknown }[] = []
+    const fetch: FetchLike = (url, init) => {
+      calls.push({ url, body: init?.body ? JSON.parse(init.body) : undefined })
+      const enc = new TextEncoder()
+      const body = new ReadableStream<Uint8Array>({
+        start(c) {
+          for (const ch of chunks) c.enqueue(enc.encode(ch))
+          c.close()
+        },
+      })
+      return Promise.resolve({ ok: true, status: 200, body, json: () => Promise.resolve({}), text: () => Promise.resolve('') })
+    }
+    return { fetch, calls }
+  }
+
+  it('streams content deltas via onToken and returns the assembled result', async () => {
+    const { fetch, calls } = sseStub([
+      'data: {"choices":[{"delta":{"content":"Right "}}]}\n\n',
+      'data: {"choices":[{"delta":{"content":"now: #2."}}]}\n\n',
+      'data: [DONE]\n\n',
+    ])
+    const client = createChatClient({ baseUrl: 'http://x/v1', model: 'm' }, fetch)
+    const tokens: string[] = []
+    const res = await client.complete([{ role: 'user', content: 'now?' }], undefined, (d) => tokens.push(d))
+
+    expect(tokens).toEqual(['Right ', 'now: #2.'])
+    expect(res.content).toBe('Right now: #2.')
+    expect((calls[0].body as { stream?: boolean }).stream).toBe(true)
+  })
+
+  it('assembles a streamed tool call from argument deltas', async () => {
+    const { fetch } = sseStub([
+      'data: {"choices":[{"delta":{"tool_calls":[{"index":0,"id":"c1","function":{"name":"query_project","arguments":"{\\"view\\""}}]}}]}\n\n',
+      'data: {"choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":":\\"focus\\"}"}}]}}]}\n\n',
+      'data: [DONE]\n\n',
+    ])
+    const client = createChatClient({ baseUrl: 'http://x/v1', model: 'm' }, fetch)
+    const res = await client.complete([{ role: 'user', content: 'x' }], [{ name: 'query_project', description: '', parameters: {} }], () => {})
+    expect(res.toolCalls).toEqual([{ id: 'c1', name: 'query_project', arguments: '{"view":"focus"}' }])
+  })
+
   it('POSTs to /chat/completions and parses content', async () => {
     const { fetch, calls } = stub({ choices: [{ message: { content: 'the focus is #2' } }] })
     const client = createChatClient({ baseUrl: 'http://localhost:1234/v1', model: 'gemma' }, fetch)

packages/core/src/agent/chat-client.ts

@@ -48,8 +48,12 @@ export interface CompletionResult {
   toolCalls: ToolCall[]
 }
 
+/** Called with each streamed content delta (final-prose streaming). */
+export type OnToken = (delta: string) => void
+
 export interface ChatClient {
-  complete(messages: ChatMessage[], tools?: ToolDecl[]): Promise<CompletionResult>
+  /** When `onToken` is given, the response streams (SSE) and each content delta is emitted. */
+  complete(messages: ChatMessage[], tools?: ToolDecl[], onToken?: OnToken): Promise<CompletionResult>
 }
 
 /** Map our message shape to the OpenAI wire shape. */
@@ -85,7 +89,7 @@ export function createChatClient(config: ModelConfig, fetchImpl: FetchLike): Cha
   const url = `${config.baseUrl.replace(/\/+$/, '')}/chat/completions`
 
   return {
-    async complete(messages, tools) {
+    async complete(messages, tools, onToken) {
       const body: Record<string, unknown> = {
         model: config.model,
         messages: messages.map(toWireMessage),
@@ -95,11 +99,13 @@ export function createChatClient(config: ModelConfig, fetchImpl: FetchLike): Cha
         body.tools = tools.map(toWireTool)
         body.tool_choice = 'auto'
       }
+      const stream = !!onToken
+      if (stream) body.stream = true
       const res = await fetchImpl(url, {
         method: 'POST',
         headers: {
           'Content-Type': 'application/json',
-          Accept: 'application/json',
+          Accept: stream ? 'text/event-stream' : 'application/json',
           ...(config.apiKey ? { Authorization: `Bearer ${config.apiKey}` } : {}),
         },
         body: JSON.stringify(body),
@@ -108,6 +114,8 @@ export function createChatClient(config: ModelConfig, fetchImpl: FetchLike): Cha
         const text = await res.text().catch(() => '')
         throw new Error(`model completion failed (${res.status}): ${text.slice(0, 200)}`)
       }
+      if (stream && res.body) return readStream(res.body, onToken!)
+
       const json = (await res.json()) as { choices?: { message: RawChoiceMessage }[] }
       const msg = json.choices?.[0]?.message
       return {
@@ -121,3 +129,56 @@ export function createChatClient(config: ModelConfig, fetchImpl: FetchLike): Cha
     },
   }
 }
+
+/** Streamed tool-call delta: name arrives first, arguments accumulate across chunks. */
+interface RawToolDelta {
+  index: number
+  id?: string
+  function?: { name?: string; arguments?: string }
+}
+
+/** Parse an OpenAI SSE stream: emit content deltas via onToken, accumulate the final result. */
+async function readStream(body: ReadableStream<Uint8Array>, onToken: OnToken): Promise<CompletionResult> {
+  const reader = body.getReader()
+  const decoder = new TextDecoder()
+  let buffer = ''
+  let content = ''
+  const toolAcc: { id: string; name: string; arguments: string }[] = []
+
+  const handle = (data: string) => {
+    if (data === '[DONE]') return
+    let chunk: { choices?: { delta?: { content?: string; tool_calls?: RawToolDelta[] } }[] }
+    try {
+      chunk = JSON.parse(data)
+    } catch {
+      return
+    }
+    const delta = chunk.choices?.[0]?.delta
+    if (!delta) return
+    if (delta.content) {
+      content += delta.content
+      onToken(delta.content)
+    }
+    for (const tc of delta.tool_calls ?? []) {
+      const slot = (toolAcc[tc.index] ??= { id: '', name: '', arguments: '' })
+      if (tc.id) slot.id = tc.id
+      if (tc.function?.name) slot.name = tc.function.name
+      if (tc.function?.arguments) slot.arguments += tc.function.arguments
+    }
+  }
+
+  for (;;) {
+    const { done, value } = await reader.read()
+    if (done) break
+    buffer += decoder.decode(value, { stream: true })
+    const lines = buffer.split('\n')
+    buffer = lines.pop() ?? ''
+    for (const line of lines) {
+      const t = line.trim()
+      if (t.startsWith('data:')) handle(t.slice(5).trim())
+    }
+  }
+  if (buffer.trim().startsWith('data:')) handle(buffer.trim().slice(5).trim())
+
+  return { content, toolCalls: toolAcc.filter((t) => t.name).map((t) => ({ id: t.id, name: t.name, arguments: t.arguments })) }
+}

packages/core/src/gitea/types.ts

@@ -31,6 +31,8 @@ export interface GiteaHttpResponse {
   status: number
   json(): Promise<unknown>
   text(): Promise<string>
+  /** Present on the real fetch Response; used for SSE streaming (chat). */
+  body?: ReadableStream<Uint8Array> | null
 }
 
 export type FetchLike = (url: string, init?: GiteaRequestInit) => Promise<GiteaHttpResponse>

packages/core/src/index.ts

@@ -80,7 +80,7 @@ export type {
 } from './calibration/calibration-v0.js'
 
 export { createChatClient } from './agent/chat-client.js'
-export type { ChatClient, ChatMessage, CompletionResult, ModelConfig, ToolCall, ToolDecl } from './agent/chat-client.js'
+export type { ChatClient, ChatMessage, CompletionResult, ModelConfig, OnToken, ToolCall, ToolDecl } from './agent/chat-client.js'
 export { pickModel } from './agent/model-router.js'
 export type { ModelRouter, TaskKind } from './agent/model-router.js'
 export { runAgentTurn } from './agent/agent-loop.js'