feat: Reginald follows the model you load (auto-detect) + shows it in the header

Instead of a hardcoded model name (which forces LM Studio to JIT-swap your loaded
model out — and fails when a big model already fills memory), resolve the model
at request time: an explicit env override wins, else ask the server which model
is *loaded* (LM Studio's native /api/v0/models), else the first non-embedding
model, else a default. Reginald now uses whatever you load, no config churn.

- main/model.ts: resolveLoadedModel() drives both model:status and model:chat;
  COMMITEA_MODEL_SMALL still overrides.
- useChat exposes the resolved model id; the panel header shows it
  (google/gemma-4-26b-a4b-qat → "gemma-4-26b-a4b · local").
- live-reginald e2e: header assertion relaxed to the loaded model; timeouts
  raised for a slow big local model (~2 calls/turn + a reconcile).

Verified: 14 fixture e2e green; live e2e drives the app against the loaded
gemma-4-26b — "What now?" → "You should work on #2 … on the critical path,
unblocks #33 and #4" (the real scheduler pick), header shows the live model.

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

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

@@ -0,0 +1,33 @@
+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 + COMMITEA_MODEL_LIVE=1 + a local model on :1234). Launches
+// WITHOUT COMMITEA_E2E so Reginald runs the real agent loop against the real repo.
+test.describe('live Reginald', () => {
+  test('answers a question by consulting the real project', async () => {
+    test.skip(!process.env.GITEA_LIVE || !process.env.COMMITEA_MODEL_LIVE, 'live model test — opt-in')
+    test.setTimeout(300_000) // a big local model is slow: ~2 calls/turn + a reconcile
+    const app = await electron.launch({ args: [MAIN], env: { ...process.env } })
+    const win = await app.firstWindow()
+    await win.waitForLoadState('domcontentloaded')
+
+    // model configured → the live greeting + header (the loaded model, not the scripted demo)
+    await expect(win.getByText(/· local$/)).toBeVisible({ timeout: 20000 })
+    await expect(win.getByText(/I check the real board before I answer/)).toBeVisible()
+
+    const composer = win.getByPlaceholder(/Tell me what to do/)
+    await composer.fill('What should I work on right now?')
+    await composer.press('Enter')
+
+    // the agent loop ran end-to-end: it consulted the project, then answered
+    await expect(win.getByText(/consulted the project/)).toBeVisible({ timeout: 240_000 })
+    await win.screenshot({ path: join(here, '.artifacts', 'screens', 'live-reginald.png'), fullPage: true, animations: 'disabled' })
+
+    await app.close()
+  })
+})

apps/desktop/src/main/gitea.ts

@@ -9,7 +9,13 @@
 import { readFileSync } from 'node:fs'
 import { dirname, join } from 'node:path'
 
-import { createGiteaClient, type GiteaConfig, type LifecycleEvent } from '@commitea/core'
+import {
+  createGiteaClient,
+  type GiteaClient,
+  type GiteaConfig,
+  type LifecycleEvent,
+  type ProjectSnapshot,
+} from '@commitea/core'
 import { ipcMain } from 'electron'
 
 /** Walk up from cwd looking for a .env.local with a GITEA_TOKEN (dev convenience). */
@@ -43,15 +49,20 @@ function resolveConfig(): GiteaConfig | null {
   }
 }
 
-export function registerGiteaIpc(): void {
+// Memoized client so both the gitea and model bridges share one instance.
+let sharedClient: GiteaClient | null | undefined
+export function getGiteaClient(): GiteaClient | null {
+  if (sharedClient === undefined) {
     const config = resolveConfig()
-  const client = config ? createGiteaClient(config, fetch) : null
+    sharedClient = config ? createGiteaClient(config, fetch) : null
-  const repo = config ? `${config.owner}/${config.repo}` : null
+  }
+  return sharedClient
+}
 
-  ipcMain.handle('gitea:status', () => ({ configured: !!config, repo }))
+/** Full reconcile: issues + milestones + native deps + lifecycle timelines. */
-
+export async function reconcileSnapshot(
-  ipcMain.handle('gitea:reconcile', async () => {
+  client: GiteaClient,
-    if (!client) return { configured: false, issues: [], milestones: [], deps: [], timelines: {} }
+): Promise<ProjectSnapshot & { milestones: Awaited<ReturnType<GiteaClient['listMilestones']>> }> {
   const [issues, milestones] = await Promise.all([client.listIssues(), client.listMilestones()])
   // dependency edges among the open scope (the scheduler only plans what's left)
   const open = issues.filter((i) => i.state === 'open')
@@ -64,7 +75,19 @@ export function registerGiteaIpc(): void {
     issues.map(async (i) => [i.number, await client.getIssueTimeline(i.number)] as const),
   )
   const timelines: Record<number, LifecycleEvent[]> = Object.fromEntries(timelineEntries)
-    return { configured: true, issues, milestones, deps, timelines }
+  return { issues, milestones, deps, timelines }
+}
+
+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 }))
+
+  ipcMain.handle('gitea:reconcile', async () => {
+    if (!client) return { configured: false, issues: [], milestones: [], deps: [], timelines: {} }
+    const snap = await reconcileSnapshot(client)
+    return { configured: true, ...snap }
   })
 
   ipcMain.handle('gitea:getIssue', async (_event, index: number) => {

apps/desktop/src/main/index.ts

@@ -3,6 +3,7 @@ import { join } from 'node:path'
 import { BrowserWindow, app, shell } from 'electron'
 
 import { registerGiteaIpc } from './gitea.js'
+import { registerModelIpc } from './model.js'
 
 function createWindow(): void {
   const win = new BrowserWindow({
@@ -35,6 +36,7 @@ function createWindow(): void {
 
 void app.whenReady().then(() => {
   registerGiteaIpc()
+  registerModelIpc()
   createWindow()
 
   app.on('activate', () => {

apps/desktop/src/main/model.ts

@@ -0,0 +1,103 @@
+/**
+ * Main-process model bridge — Reginald's brain runs here. Model traffic (like
+ * gitea's) stays in main: the renderer is CSP-locked and never talks to the
+ * LLM directly. On `model:chat` it drives the agent loop against the configured
+ * OpenAI-compatible endpoint, executing `query_project` by reconciling the repo
+ * and building the requested view. v0 is read-only — writes still go through the
+ * propose-approve controls.
+ */
+
+import {
+  buildProjectView,
+  type ChatMessage,
+  createChatClient,
+  type ModelRouter,
+  type ProjectView,
+  type QueryFilters,
+  REGINALD_SYSTEM,
+  REGINALD_TOOLS,
+  runAgentTurn,
+} from '@commitea/core'
+import { ipcMain } from 'electron'
+
+import { getGiteaClient, reconcileSnapshot } from './gitea.js'
+
+/** Small local model for prose + the read tool; big model reserved for later decomposition. */
+function resolveModelRouter(): ModelRouter | null {
+  if (process.env.COMMITEA_E2E === '1') return null // e2e uses the scripted fixture Reginald
+  const baseUrl = process.env.COMMITEA_MODEL_URL ?? 'http://localhost:1234/v1'
+  if (!baseUrl) return null
+  return {
+    small: { baseUrl, model: process.env.COMMITEA_MODEL_SMALL ?? '' },
+    big: { baseUrl, model: process.env.COMMITEA_MODEL_BIG ?? '' },
+  }
+}
+
+/**
+ * Resolve which model to actually ask for. An explicit env override wins;
+ * otherwise ask the server which model is *loaded* (LM Studio's native
+ * `/api/v0/models`) so Reginald follows whatever you load — no config churn on a
+ * model switch. Falls back to the first non-embedding model, then a sane default.
+ */
+async function resolveLoadedModel(baseUrl: string, override: string): Promise<string> {
+  if (override) return override
+  const root = baseUrl.replace(/\/v1\/?$/, '')
+  try {
+    const res = await fetch(`${root}/api/v0/models`)
+    if (res.ok) {
+      const data = (await res.json()) as { data?: { id: string; state?: string; type?: string }[] }
+      const loaded = (data.data ?? []).find((m) => m.state === 'loaded' && m.type !== 'embeddings')
+      if (loaded) return loaded.id
+    }
+  } catch {
+    // native API unavailable — fall through to the OpenAI listing
+  }
+  try {
+    const res = await fetch(`${baseUrl.replace(/\/+$/, '')}/models`)
+    if (res.ok) {
+      const data = (await res.json()) as { data?: { id: string }[] }
+      const first = (data.data ?? []).find((m) => !/embed/i.test(m.id))
+      if (first) return first.id
+    }
+  } catch {
+    // ignore — use the default
+  }
+  return 'google/gemma-4-e4b'
+}
+
+export function registerModelIpc(): void {
+  const router = resolveModelRouter()
+
+  ipcMain.handle('model:status', async () => {
+    if (!router) return { configured: false, model: null }
+    const model = await resolveLoadedModel(router.small.baseUrl, router.small.model)
+    return { configured: true, model }
+  })
+
+  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)
+
+    const execute = async (name: string, args: unknown) => {
+      if (name !== 'query_project') return { error: `unknown tool: ${name}` }
+      if (!client) return { error: 'gitea is not configured' }
+      const snap = await reconcileSnapshot(client)
+      const a = (args ?? {}) as { view: ProjectView; filters?: QueryFilters }
+      return buildProjectView(a.view, a.filters, snap, new Date())
+    }
+
+    try {
+      const turn = await runAgentTurn({
+        complete: (m, t) => chat.complete(m, t),
+        messages: [{ role: 'system', content: REGINALD_SYSTEM }, ...messages],
+        tools: REGINALD_TOOLS,
+        execute,
+      })
+      return { ok: true as const, content: turn.content, steps: turn.steps }
+    } catch (e) {
+      return { ok: false as const, reason: 'error' as const, message: e instanceof Error ? e.message : String(e) }
+    }
+  })
+}

apps/desktop/src/preload/index.ts

@@ -10,6 +10,12 @@ const api = {
     /** One issue by index, normalized (or null if unconfigured). */
     getIssue: (index: number) => ipcRenderer.invoke('gitea:getIssue', index),
   },
+  model: {
+    /** Whether a model endpoint is configured (else the UI keeps the scripted Reginald). */
+    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),
+  },
 }
 
 export type CommiteaApi = typeof api

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

@@ -1,12 +1,13 @@
 import React, { useEffect, useRef, useState } from 'react'
 
-import { CANNED_REPLY, CHAT, type ChatMessage } from '../../data/fixtures.js'
+import { useChat } from '../../lib/use-chat.js'
 import { Icon, IconButton } from '../ui/index.js'
 
 /**
- * Reginald's panel — chat is the write-path (decisions.md D1). This is the P3-2
+ * Reginald's panel — chat is the write-path (decisions.md D1). Wired to the
- * fixture shell: it echoes a canned reply so the layout + interactions are real,
+ * model bridge via `useChat`: when a model is configured, sending drives a real
- * but no model is wired. P4 replaces `send` with the model router + tools.
+ * agent turn (query_project + prose); otherwise it echoes the scripted fixture
+ * reply so the layout stays real. Writes still go through propose-approve.
  */
 export interface ChatPanelProps {
   onOpenDirectives?: () => void
@@ -14,9 +15,10 @@ export interface ChatPanelProps {
 }
 
 export function ChatPanel({ onOpenDirectives, offline }: ChatPanelProps) {
-  const [msgs, setMsgs] = useState<ChatMessage[]>(CHAT)
+  const { msgs, thinking, live, model, steps, send: sendChat } = useChat()
+  // 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('')
-  const [thinking, setThinking] = useState(false)
   const scrollRef = useRef<HTMLDivElement>(null)
 
   useEffect(() => {
@@ -27,13 +29,8 @@ export function ChatPanel({ onOpenDirectives, offline }: ChatPanelProps) {
   const send = () => {
     const t = text.trim()
     if (!t) return
-    setMsgs((m) => [...m, { from: 'user', text: t }])
     setText('')
-    setThinking(true)
+    sendChat(t)
-    setTimeout(() => {
-      setThinking(false)
-      setMsgs((m) => [...m, { from: 'agent', text: CANNED_REPLY }])
-    }, 900)
   }
 
   return (
@@ -61,7 +58,7 @@ export function ChatPanel({ onOpenDirectives, offline }: ChatPanelProps) {
         <Icon name="sparkles" size={16} style={{ color: offline ? 'var(--ink-3)' : 'var(--jade)' }} />
         <span style={{ font: 'var(--text-body-strong)', color: 'var(--ink-1)' }}>Reginald</span>
         <span style={{ font: '400 11px var(--font-mono)', color: 'var(--ink-3)', marginLeft: 'auto' }}>
-          {offline ? 'offline · queueing' : 'gemma-4b · local'}
+          {offline ? 'offline · queueing' : live ? `${modelLabel} · local` : 'demo · scripted'}
         </span>
         <IconButton icon="history" label="Directive log" size="sm" onClick={onOpenDirectives} />
       </header>
@@ -98,6 +95,11 @@ export function ChatPanel({ onOpenDirectives, offline }: ChatPanelProps) {
           </div>
         ) : null}
         {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')))).join(', ')}
+          </div>
+        ) : null}
       </div>
 
       <div style={{ padding: 14, borderTop: '1px solid var(--line-1)', flexShrink: 0 }}>

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

@@ -1,4 +1,11 @@
-import type { DependencyEdge, GiteaIssue, GiteaMilestone, LifecycleEvent } from '@commitea/core'
+import type {
+  AgentStep,
+  ChatMessage,
+  DependencyEdge,
+  GiteaIssue,
+  GiteaMilestone,
+  LifecycleEvent,
+} from '@commitea/core'
 
 /** The gitea bridge exposed by the preload over IPC (main-process backed). */
 export interface GiteaBridge {
@@ -14,11 +21,23 @@ export interface GiteaBridge {
   getIssue(index: number): Promise<GiteaIssue | null>
 }
 
+/** One agent turn's result. */
+export type ChatResult =
+  | { ok: false; reason: 'unconfigured' | 'error'; message?: string }
+  | { ok: true; content: string; steps: AgentStep[] }
+
+/** The model bridge (Reginald) exposed by the preload over IPC. */
+export interface ModelBridge {
+  status(): Promise<{ configured: boolean; model: string | null }>
+  chat(messages: ChatMessage[]): Promise<ChatResult>
+}
+
 declare global {
   interface Window {
     commitea: {
       platform: string
       gitea: GiteaBridge
+      model: ModelBridge
     }
   }
 }

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

@@ -0,0 +1,105 @@
+import { useCallback, useEffect, useRef, useState } from 'react'
+
+import type { ChatMessage as WireMessage } from '@commitea/core'
+
+import { CANNED_REPLY, CHAT, type ChatMessage } from '../data/fixtures.js'
+
+const LIVE_GREETING: ChatMessage = {
+  from: 'agent',
+  text: 'Morning. Ask me anything about the project — I check the real board before I answer.',
+}
+
+export interface ChatState {
+  msgs: ChatMessage[]
+  thinking: boolean
+  /** true once a model endpoint is confirmed; otherwise the panel echoes the demo reply. */
+  live: boolean
+  /** The loaded model's id when live (for the header). */
+  model: string | null
+  /** Tools Reginald consulted on the last turn (for a subtle activity line). */
+  steps: string[]
+  send: (text: string) => void
+}
+
+/**
+ * Reginald's conversation. When a model is configured, `send` drives one agent
+ * turn through the main-process bridge (which runs the tool loop). Otherwise it
+ * echoes the scripted fixture reply, so the layout stays real with no model and
+ * fixture e2e is unaffected. The fixture greeting is display-only — only real
+ * turns (`convo`) are sent to the model as history.
+ */
+export function useChat(): ChatState {
+  const [seed, setSeed] = useState<ChatMessage[]>(CHAT)
+  const [convo, setConvo] = useState<ChatMessage[]>([])
+  const [thinking, setThinking] = useState(false)
+  const [live, setLive] = useState(false)
+  const [model, setModel] = useState<string | null>(null)
+  const [steps, setSteps] = useState<string[]>([])
+  const convoRef = useRef(convo)
+  convoRef.current = convo
+
+  useEffect(() => {
+    let alive = true
+    window.commitea.model
+      .status()
+      .then((s) => {
+        if (alive && s.configured) {
+          setLive(true)
+          setModel(s.model)
+          setSeed([LIVE_GREETING])
+        }
+      })
+      .catch(() => {})
+    return () => {
+      alive = false
+    }
+  }, [])
+
+  const send = useCallback(
+    (raw: string) => {
+      const text = raw.trim()
+      if (!text) return
+      const nextConvo: ChatMessage[] = [...convoRef.current, { from: 'user', text }]
+      setConvo(nextConvo)
+      setThinking(true)
+      setSteps([])
+
+      if (!live) {
+        window.setTimeout(() => {
+          setThinking(false)
+          setConvo((c) => [...c, { from: 'agent', text: CANNED_REPLY }])
+        }, 900)
+        return
+      }
+
+      const wire: WireMessage[] = nextConvo.map((m) => ({
+        role: m.from === 'user' ? 'user' : 'assistant',
+        content: m.text,
+      }))
+      window.commitea.model
+        .chat(wire)
+        .then((res) => {
+          setThinking(false)
+          if (res.ok) {
+            setSteps(res.steps.map((s) => s.tool))
+            setConvo((c) => [...c, { from: 'agent', text: res.content || '…' }])
+          } else {
+            setConvo((c) => [
+              ...c,
+              {
+                from: 'agent',
+                text: res.reason === 'error' ? `I hit a snag: ${res.message ?? 'unknown error'}` : 'No model is configured.',
+              },
+            ])
+          }
+        })
+        .catch(() => {
+          setThinking(false)
+          setConvo((c) => [...c, { from: 'agent', text: 'I could not reach the model.' }])
+        })
+    },
+    [live],
+  )
+
+  return { msgs: [...seed, ...convo], thinking, live, model, steps, send }
+}

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

@@ -0,0 +1,62 @@
+import { describe, expect, it } from 'vitest'
+
+import { extractLabelFacts } from '../labels/label-schema.js'
+import type { FetchLike, GiteaIssue } from '../gitea/types.js'
+import { runAgentTurn } from './agent-loop.js'
+import { REGINALD_SYSTEM, REGINALD_TOOLS } from './agent-tools.js'
+import { createChatClient } from './chat-client.js'
+import { buildProjectView, type ProjectSnapshot } from './query-project.js'
+
+/**
+ * Opt-in (COMMITEA_MODEL_LIVE=1). Drives the real chat client + agent loop
+ * against a local OpenAI-compatible server (LM Studio on :1234 by default),
+ * proving the model calls query_project and narrates the real result.
+ */
+const LIVE = !!process.env.COMMITEA_MODEL_LIVE
+const BASE = process.env.COMMITEA_MODEL_URL ?? 'http://localhost:1234/v1'
+const MODEL = process.env.COMMITEA_MODEL_SMALL ?? 'google/gemma-4-e4b'
+
+function issue(over: Partial<GiteaIssue>): GiteaIssue {
+  const labels = over.labels ?? []
+  return {
+    number: 1, title: '#1', body: '', state: 'open', labels, facts: extractLabelFacts(labels),
+    milestone: null, assignee: null, assignees: [], createdAt: '2026-01-05T09:00:00Z',
+    updatedAt: '2026-01-05T09:00:00Z', closedAt: null, url: '', ...over,
+  }
+}
+
+const SNAP: ProjectSnapshot = {
+  issues: [
+    issue({ number: 2, title: 'ChangeSource interface + polling source', labels: ['est/3d', 'p/1'] }),
+    issue({ number: 3, title: 'SQLite cache bootstrap', labels: ['est/2d', 'p/1'] }),
+  ],
+  timelines: {},
+  deps: [{ issue: 3, dependsOn: 2 }],
+}
+
+describe('agent loop (live model)', () => {
+  it.skipIf(!LIVE)(
+    'calls query_project and narrates the real focus',
+    async () => {
+      const client = createChatClient({ baseUrl: BASE, model: MODEL }, globalThis.fetch as unknown as FetchLike)
+      const turn = await runAgentTurn({
+        complete: (m, t) => client.complete(m, t),
+        messages: [
+          { role: 'system', content: REGINALD_SYSTEM },
+          { role: 'user', content: 'What should I work on right now?' },
+        ],
+        tools: REGINALD_TOOLS,
+        execute: async (name, args) =>
+          name === 'query_project'
+            ? buildProjectView((args as { view: any }).view, (args as any).filters, SNAP, new Date())
+            : { error: `unknown tool ${name}` },
+      })
+
+      // it consulted the project, then answered in prose about the real top item (#2)
+      expect(turn.steps.some((s) => s.tool === 'query_project')).toBe(true)
+      expect(turn.content.trim().length).toBeGreaterThan(0)
+      expect(turn.content).toMatch(/#?2\b|ChangeSource/i)
+    },
+    60_000,
+  )
+})

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

@@ -0,0 +1,69 @@
+/**
+ * The agent turn loop. Given a model `complete` fn, the conversation, the tool
+ * declarations, and an `execute` that actually runs a tool, it drives the
+ * call→tool→result→call cycle until the model answers in prose (or a step
+ * budget is hit). Pure orchestration with injected I/O — the model and the tool
+ * executor are both stubbable, so the loop is fully unit-testable offline.
+ */
+
+import type { ChatMessage, CompletionResult, ToolDecl } from './chat-client.js'
+
+/** A tool the loop ran, with the raw args and its stringified result — for the UI's activity trail. */
+export interface AgentStep {
+  tool: string
+  arguments: string
+  result: string
+}
+
+export type ToolExecutor = (name: string, args: unknown) => Promise<unknown>
+
+export interface AgentTurn {
+  content: string
+  steps: AgentStep[]
+  /** The full conversation including this turn's assistant/tool messages. */
+  messages: ChatMessage[]
+}
+
+const DEFAULT_MAX_STEPS = 4
+
+function stringify(result: unknown): string {
+  return typeof result === 'string' ? result : JSON.stringify(result)
+}
+
+export async function runAgentTurn(opts: {
+  complete: (messages: ChatMessage[], tools?: ToolDecl[]) => Promise<CompletionResult>
+  messages: ChatMessage[]
+  tools: ToolDecl[]
+  execute: ToolExecutor
+  maxSteps?: number
+}): 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)
+    if (toolCalls.length === 0) {
+      convo.push({ role: 'assistant', content })
+      return { content, steps, messages: convo }
+    }
+    convo.push({ role: 'assistant', content, toolCalls })
+    for (const tc of toolCalls) {
+      let result: unknown
+      try {
+        const args = tc.arguments ? JSON.parse(tc.arguments) : {}
+        result = await opts.execute(tc.name, args)
+      } catch (e) {
+        result = { error: e instanceof Error ? e.message : String(e) }
+      }
+      const resultStr = stringify(result)
+      steps.push({ tool: tc.name, arguments: tc.arguments, result: resultStr })
+      convo.push({ role: 'tool', toolCallId: tc.id, name: tc.name, content: resultStr })
+    }
+  }
+
+  // Out of tool budget — force a final prose answer with tools withheld.
+  const final = await opts.complete(convo, [])
+  convo.push({ role: 'assistant', content: final.content })
+  return { content: final.content, steps, messages: convo }
+}

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

@@ -0,0 +1,46 @@
+/**
+ * Reginald's tool surface. v0 ships the one read tool (`query_project`); the
+ * three write tools (capture_work, apply_changes, record_directive) layer on
+ * later against the same loop. Few, fat tools so a small local model survives
+ * with one thing to reach for (docs/agent-tools.md).
+ */
+
+import type { ToolDecl } from './chat-client.js'
+
+export const QUERY_PROJECT_TOOL: ToolDecl = {
+  name: 'query_project',
+  description:
+    'Read the current project state. A `view` selects the shape; deterministic code (scheduler, ' +
+    'lifecycle inference, calibration) backs every number — you report it, you never compute it.',
+  parameters: {
+    type: 'object',
+    properties: {
+      view: {
+        type: 'string',
+        enum: ['focus', 'board', 'calibration', 'issue', 'search'],
+        description:
+          'focus = Now/Next/Later; board = issues by lifecycle column; calibration = estimate-vs-actual; ' +
+          'issue = one issue (needs filters.issueId); search = issues matching filters.query.',
+      },
+      filters: {
+        type: 'object',
+        properties: {
+          issueId: { type: 'number' },
+          query: { type: 'string' },
+          limit: { type: 'number' },
+        },
+      },
+    },
+    required: ['view'],
+  },
+}
+
+export const REGINALD_TOOLS: ToolDecl[] = [QUERY_PROJECT_TOOL]
+
+export const REGINALD_SYSTEM = [
+  'You are Reginald, the calm, dry project manager inside CommiTea — a tool that runs projects on Gitea.',
+  'Call query_project to ground every answer in the real project; never invent issues, numbers, or dates.',
+  'The scheduler and forecasts are deterministic code — report their output, do not recompute it.',
+  'Forecasts are ranges, never single dates. Refer to issues as #<number>.',
+  'Be brief and plain. A sentence or two is usually enough. No preamble, no bullet-point dumps.',
+].join(' ')

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

@@ -0,0 +1,239 @@
+import { describe, expect, it } from 'vitest'
+
+import { extractLabelFacts } from '../labels/label-schema.js'
+import type { GiteaIssue } from '../gitea/types.js'
+import type { FetchLike } from '../gitea/types.js'
+import { type ChatMessage, type CompletionResult, createChatClient, type ToolDecl } from './chat-client.js'
+import { runAgentTurn } from './agent-loop.js'
+import { pickModel } from './model-router.js'
+import { buildProjectView, type ProjectSnapshot } from './query-project.js'
+
+// ---- chat client ----
+
+describe('createChatClient', () => {
+  function stub(body: unknown) {
+    const calls: { url: string; body: unknown }[] = []
+    const fetch: FetchLike = (url, init) => {
+      calls.push({ url, body: init?.body ? JSON.parse(init.body) : undefined })
+      return Promise.resolve({
+        ok: true,
+        status: 200,
+        json: () => Promise.resolve(body),
+        text: () => Promise.resolve(''),
+      })
+    }
+    return { fetch, calls }
+  }
+
+  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)
+    const res = await client.complete([{ role: 'user', content: 'what now?' }])
+
+    expect(res.content).toBe('the focus is #2')
+    expect(res.toolCalls).toEqual([])
+    expect(calls[0].url).toBe('http://localhost:1234/v1/chat/completions')
+    expect((calls[0].body as { model: string }).model).toBe('gemma')
+  })
+
+  it('parses tool calls from the wire shape', async () => {
+    const { fetch } = stub({
+      choices: [
+        {
+          message: {
+            content: null,
+            tool_calls: [{ id: 'c1', function: { name: 'query_project', arguments: '{"view":"focus"}' } }],
+          },
+        },
+      ],
+    })
+    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('serializes assistant tool_calls + tool results in the request', async () => {
+    const { fetch, calls } = stub({ choices: [{ message: { content: 'ok' } }] })
+    const client = createChatClient({ baseUrl: 'http://x/v1', model: 'm' }, fetch)
+    const convo: ChatMessage[] = [
+      { role: 'user', content: 'hi' },
+      { role: 'assistant', content: '', toolCalls: [{ id: 'c1', name: 'query_project', arguments: '{}' }] },
+      { role: 'tool', toolCallId: 'c1', name: 'query_project', content: '{"now":null}' },
+    ]
+    await client.complete(convo)
+    const sent = (calls[0].body as { messages: Record<string, unknown>[] }).messages
+    expect(sent[1].tool_calls).toBeDefined()
+    expect(sent[2]).toEqual({ role: 'tool', tool_call_id: 'c1', content: '{"now":null}' })
+  })
+})
+
+// ---- model router ----
+
+describe('pickModel', () => {
+  const router = { small: { baseUrl: 'u', model: 'small' }, big: { baseUrl: 'u', model: 'big' } }
+  it('routes plan → big, everything else → small', () => {
+    expect(pickModel(router, 'plan').model).toBe('big')
+    expect(pickModel(router, 'ritual').model).toBe('small')
+  })
+})
+
+// ---- agent loop ----
+
+describe('runAgentTurn', () => {
+  const tools: ToolDecl[] = [{ name: 'query_project', description: '', parameters: {} }]
+
+  it('runs a tool then returns the model prose, recording the step', async () => {
+    // first completion asks for a tool; second answers in prose
+    const scripted: CompletionResult[] = [
+      { content: '', toolCalls: [{ id: 'c1', name: 'query_project', arguments: '{"view":"focus"}' }] },
+      { content: 'Right now: #2.', toolCalls: [] },
+    ]
+    let i = 0
+    const executed: { name: string; args: unknown }[] = []
+    const turn = await runAgentTurn({
+      complete: async () => scripted[i++],
+      messages: [{ role: 'user', content: 'what now?' }],
+      tools,
+      execute: async (name, args) => {
+        executed.push({ name, args })
+        return { now: { issue: 2 } }
+      },
+    })
+
+    expect(turn.content).toBe('Right now: #2.')
+    expect(turn.steps).toEqual([
+      { tool: 'query_project', arguments: '{"view":"focus"}', result: '{"now":{"issue":2}}' },
+    ])
+    expect(executed).toEqual([{ name: 'query_project', args: { view: 'focus' } }])
+    // conversation carries user → assistant(tool) → tool → assistant(prose)
+    expect(turn.messages.map((m) => m.role)).toEqual(['user', 'assistant', 'tool', 'assistant'])
+  })
+
+  it('answers directly when the model needs no tool', async () => {
+    const turn = await runAgentTurn({
+      complete: async () => ({ content: 'Hello.', toolCalls: [] }),
+      messages: [{ role: 'user', content: 'hi' }],
+      tools,
+      execute: async () => ({}),
+    })
+    expect(turn.content).toBe('Hello.')
+    expect(turn.steps).toEqual([])
+  })
+
+  it('captures a tool executor error as the tool result instead of throwing', async () => {
+    const scripted: CompletionResult[] = [
+      { content: '', toolCalls: [{ id: 'c1', name: 'query_project', arguments: '{}' }] },
+      { content: 'Something went wrong reading that.', toolCalls: [] },
+    ]
+    let i = 0
+    const turn = await runAgentTurn({
+      complete: async () => scripted[i++],
+      messages: [{ role: 'user', content: 'x' }],
+      tools,
+      execute: async () => {
+        throw new Error('boom')
+      },
+    })
+    expect(turn.steps[0].result).toContain('boom')
+    expect(turn.content).toBe('Something went wrong reading that.')
+  })
+
+  it('forces a final prose answer when the step budget is exhausted', async () => {
+    // always asks for a tool; the loop must still terminate with prose
+    let toolRounds = 0
+    const turn = await runAgentTurn({
+      complete: async (_m, tls) => {
+        if (tls && tls.length) {
+          toolRounds++
+          return { content: '', toolCalls: [{ id: `c${toolRounds}`, name: 'query_project', arguments: '{}' }] }
+        }
+        return { content: 'final answer', toolCalls: [] } // called with tools withheld
+      },
+      messages: [{ role: 'user', content: 'x' }],
+      tools,
+      execute: async () => ({}),
+      maxSteps: 2,
+    })
+    expect(toolRounds).toBe(2)
+    expect(turn.content).toBe('final answer')
+  })
+})
+
+// ---- query_project view builder ----
+
+describe('buildProjectView', () => {
+  const asOf = new Date('2026-02-01T00:00:00Z')
+
+  function issue(over: Partial<GiteaIssue>): GiteaIssue {
+    const labels = over.labels ?? []
+    return {
+      number: 1,
+      title: '#1',
+      body: '',
+      state: 'open',
+      labels,
+      facts: extractLabelFacts(labels),
+      milestone: null,
+      assignee: null,
+      assignees: [],
+      createdAt: '2026-01-05T09:00:00Z',
+      updatedAt: '2026-01-05T09:00:00Z',
+      closedAt: null,
+      url: '',
+      ...over,
+    }
+  }
+
+  const snap: ProjectSnapshot = {
+    issues: [
+      issue({ number: 2, title: 'ChangeSource', labels: ['est/3d', 'p/1'] }),
+      issue({ number: 3, title: 'SQLite cache', labels: ['est/2d', 'p/1'] }),
+      issue({ number: 9, title: 'Done thing', state: 'closed', labels: ['est/2d'], closedAt: '2026-01-12T09:00:00Z' }),
+    ],
+    timelines: { 9: [{ type: 'commit', at: '2026-01-07T09:00:00Z' }, { type: 'close', at: '2026-01-12T09:00:00Z' }] },
+    deps: [{ issue: 3, dependsOn: 2 }],
+  }
+
+  it('focus returns Now/Next/Later grounded in the scheduler', () => {
+    const v = buildProjectView('focus', undefined, snap, asOf) as {
+      now: { issue: number } | null
+      openCount: number
+    }
+    expect(v.now?.issue).toBe(2) // #3 depends on #2, so #2 comes first
+    expect(v.openCount).toBe(2)
+  })
+
+  it('board groups issues by lifecycle column', () => {
+    const v = buildProjectView('board', undefined, snap, asOf) as {
+      columns: { column: string; count: number }[]
+    }
+    const done = v.columns.find((c) => c.column === 'done')
+    expect(done?.count).toBe(1) // #9
+    const triage = v.columns.find((c) => c.column === 'triage')
+    expect(triage?.count).toBe(2) // #2, #3 labelled, no commits yet
+  })
+
+  it('calibration reports n + cold-start honestly', () => {
+    const v = buildProjectView('calibration', undefined, snap, asOf) as { n: number; coldStart: boolean }
+    expect(v.n).toBe(1) // one closed, estimated, with an actual
+    expect(v.coldStart).toBe(true)
+  })
+
+  it('issue returns intent + derived for a specific id', () => {
+    const v = buildProjectView('issue', { issueId: 3 }, snap, asOf) as {
+      issue: number
+      blockedBy: number[]
+    }
+    expect(v.issue).toBe(3)
+    expect(v.blockedBy).toEqual([2])
+  })
+
+  it('search matches title/labels', () => {
+    const v = buildProjectView('search', { query: 'sqlite' }, snap, asOf) as { results: { issue: number }[] }
+    expect(v.results.map((r) => r.issue)).toEqual([3])
+  })
+
+  it('unbuilt views return a notImplemented marker, not fabricated data', () => {
+    expect(buildProjectView('standup', undefined, snap, asOf)).toEqual({ notImplemented: 'standup' })
+  })
+})

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

@@ -0,0 +1,123 @@
+/**
+ * Reginald's model client — a thin OpenAI-wire chat-completions client over an
+ * injected fetch (the same seam the gitea client uses). `@commitea/core` stays
+ * pure: no network, no globals. The desktop main process passes the real
+ * `fetch`; tests pass a stub. Points at any OpenAI-compatible endpoint (LM
+ * Studio, Ollama, the OpenAI API) — the model router picks which.
+ */
+
+import type { FetchLike } from '../gitea/types.js'
+
+/** Where a model lives + which model to ask for. */
+export interface ModelConfig {
+  /** OpenAI-compatible base, including the version segment, e.g. `http://localhost:1234/v1`. */
+  baseUrl: string
+  model: string
+  /** Bearer key; omitted for local servers that don't check it. */
+  apiKey?: string
+}
+
+/** A model's request to run a tool. `arguments` is a raw JSON string (OpenAI shape). */
+export interface ToolCall {
+  id: string
+  name: string
+  arguments: string
+}
+
+/** One turn in the conversation, in our normalized shape. */
+export interface ChatMessage {
+  role: 'system' | 'user' | 'assistant' | 'tool'
+  content: string
+  /** assistant turns that requested tools. */
+  toolCalls?: ToolCall[]
+  /** tool turns: which call they answer. */
+  toolCallId?: string
+  /** tool turns: the function name. */
+  name?: string
+}
+
+/** A tool the model may call — name, description, and a JSON-Schema parameter spec. */
+export interface ToolDecl {
+  name: string
+  description: string
+  parameters: object
+}
+
+export interface CompletionResult {
+  content: string
+  toolCalls: ToolCall[]
+}
+
+export interface ChatClient {
+  complete(messages: ChatMessage[], tools?: ToolDecl[]): Promise<CompletionResult>
+}
+
+/** Map our message shape to the OpenAI wire shape. */
+function toWireMessage(m: ChatMessage): Record<string, unknown> {
+  if (m.role === 'assistant' && m.toolCalls?.length) {
+    return {
+      role: 'assistant',
+      content: m.content || null,
+      tool_calls: m.toolCalls.map((tc) => ({
+        id: tc.id,
+        type: 'function',
+        function: { name: tc.name, arguments: tc.arguments },
+      })),
+    }
+  }
+  if (m.role === 'tool') {
+    return { role: 'tool', tool_call_id: m.toolCallId, content: m.content }
+  }
+  return { role: m.role, content: m.content }
+}
+
+function toWireTool(t: ToolDecl): Record<string, unknown> {
+  return { type: 'function', function: { name: t.name, description: t.description, parameters: t.parameters } }
+}
+
+/** Shape of the one choice we read back. */
+interface RawChoiceMessage {
+  content: string | null
+  tool_calls?: { id: string; function: { name: string; arguments: string } }[]
+}
+
+export function createChatClient(config: ModelConfig, fetchImpl: FetchLike): ChatClient {
+  const url = `${config.baseUrl.replace(/\/+$/, '')}/chat/completions`
+
+  return {
+    async complete(messages, tools) {
+      const body: Record<string, unknown> = {
+        model: config.model,
+        messages: messages.map(toWireMessage),
+        temperature: 0,
+      }
+      if (tools?.length) {
+        body.tools = tools.map(toWireTool)
+        body.tool_choice = 'auto'
+      }
+      const res = await fetchImpl(url, {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json',
+          Accept: 'application/json',
+          ...(config.apiKey ? { Authorization: `Bearer ${config.apiKey}` } : {}),
+        },
+        body: JSON.stringify(body),
+      })
+      if (!res.ok) {
+        const text = await res.text().catch(() => '')
+        throw new Error(`model completion failed (${res.status}): ${text.slice(0, 200)}`)
+      }
+      const json = (await res.json()) as { choices?: { message: RawChoiceMessage }[] }
+      const msg = json.choices?.[0]?.message
+      return {
+        content: msg?.content ?? '',
+        toolCalls: (msg?.tool_calls ?? []).map((tc) => ({
+          id: tc.id,
+          name: tc.function.name,
+          arguments: tc.function.arguments,
+        })),
+      }
+    },
+  }
+}

packages/core/src/agent/model-router.ts

@@ -0,0 +1,22 @@
+/**
+ * Model router (per PLAN.md / decisions.md): prose + the read tool run on a
+ * small local model; decomposition/negotiation earns the big one. Reginald v0
+ * only exercises the small model (read + prose); the `big` slot is declared so
+ * `capture_work` / `record_directive` can route to it without a rewrite.
+ */
+
+import type { ModelConfig } from './chat-client.js'
+
+export type TaskKind = 'ritual' | 'plan'
+
+export interface ModelRouter {
+  /** Small local model — standups, focus prose, the read tool. */
+  small: ModelConfig
+  /** Big model — capture decomposition, directive negotiation. */
+  big: ModelConfig
+}
+
+/** `plan` → the big model; everything else → the small one. */
+export function pickModel(router: ModelRouter, kind: TaskKind): ModelConfig {
+  return kind === 'plan' ? router.big : router.small
+}

packages/core/src/agent/query-project.ts

@@ -0,0 +1,146 @@
+/**
+ * `query_project` — Reginald's single read tool. It selects a compact,
+ * model-friendly view over the reconciled backlog. Every number comes from
+ * deterministic code (scheduler, lifecycle inference, calibration); the model
+ * only requests a shape and narrates it — it never computes (decisions.md).
+ *
+ * v0 serves focus / board / calibration / issue. The remaining views
+ * (milestone / runway / standup / search) return a `notImplemented` marker so
+ * the model degrades honestly instead of inventing data.
+ */
+
+import { fitCalibration, calibrationSamples } from '../calibration/calibration-v0.js'
+import type { GiteaIssue } from '../gitea/types.js'
+import { inferLifecycle, type LifecycleColumn, type LifecycleEvent } from '../lifecycle/lifecycle-v0.js'
+import { type DependencyEdge, schedule, selectFocus } from '../scheduler/scheduler-v0.js'
+
+export type ProjectView =
+  | 'focus'
+  | 'board'
+  | 'calibration'
+  | 'issue'
+  | 'milestone'
+  | 'runway'
+  | 'standup'
+  | 'search'
+
+export interface QueryFilters {
+  issueId?: number
+  query?: string
+  limit?: number
+}
+
+/** The reconciled inputs a view is built from. */
+export interface ProjectSnapshot {
+  issues: GiteaIssue[]
+  timelines: Record<number, LifecycleEvent[]>
+  deps: DependencyEdge[]
+}
+
+const COLUMN_ORDER: LifecycleColumn[] = ['diagnosis', 'triage', 'steeping', 'review', 'done']
+
+function toSchedulable(issues: GiteaIssue[]) {
+  return issues
+    .filter((i) => i.state === 'open')
+    .map((i) => ({
+      number: i.number,
+      title: i.title,
+      labels: i.labels,
+      estimateDays: i.facts.estimateDays,
+      priority: i.facts.priority,
+    }))
+}
+
+function focusView(snap: ProjectSnapshot) {
+  const plan = schedule(toSchedulable(snap.issues), snap.deps)
+  const f = selectFocus(plan)
+  const slot = (item: (typeof plan.items)[number] | null) =>
+    item ? { issue: item.number, title: item.title, rationale: item.rationale } : null
+  return { now: slot(f.now), next: slot(f.next), later: slot(f.later), openCount: plan.items.length }
+}
+
+function boardView(snap: ProjectSnapshot, asOf: Date) {
+  const columns: Record<string, { issue: number; title: string; labels: string[] }[]> = {}
+  for (const key of COLUMN_ORDER) columns[key] = []
+  for (const issue of snap.issues) {
+    const inf = inferLifecycle(issue, snap.timelines[issue.number] ?? [], asOf)
+    columns[inf.column].push({ issue: issue.number, title: issue.title, labels: issue.labels })
+  }
+  return {
+    columns: COLUMN_ORDER.map((key) => ({ column: key, count: columns[key].length, issues: columns[key] })),
+  }
+}
+
+function calibrationView(snap: ProjectSnapshot, asOf: Date) {
+  const model = fitCalibration(calibrationSamples(snap.issues, snap.timelines, asOf))
+  return {
+    n: model.n,
+    coldStart: model.coldStart,
+    globalMultiplier: Number(Math.exp(model.global.mu).toFixed(2)),
+    byBucket: Object.entries(model.byBucket).map(([bucket, fit]) => ({
+      estimate: `est/${bucket}d`,
+      n: fit.n,
+      multiplier: Number(Math.exp(fit.mu).toFixed(2)),
+    })),
+  }
+}
+
+function issueView(snap: ProjectSnapshot, filters: QueryFilters, asOf: Date) {
+  const issue = snap.issues.find((i) => i.number === filters.issueId)
+  if (!issue) return { notFound: filters.issueId ?? null }
+  const inf = inferLifecycle(issue, snap.timelines[issue.number] ?? [], asOf)
+  const blockedBy = snap.deps.filter((d) => d.issue === issue.number).map((d) => d.dependsOn)
+  const blocks = snap.deps.filter((d) => d.dependsOn === issue.number).map((d) => d.issue)
+  return {
+    issue: issue.number,
+    title: issue.title,
+    state: issue.state,
+    column: inf.column,
+    labels: issue.labels,
+    assignee: issue.assignee,
+    milestone: issue.milestone?.title ?? null,
+    estimateDays: issue.facts.estimateDays,
+    priority: issue.facts.priority,
+    steepingDays: inf.steepingDays,
+    blockedBy,
+    blocks,
+  }
+}
+
+function searchView(snap: ProjectSnapshot, filters: QueryFilters) {
+  const q = (filters.query ?? '').toLowerCase().trim()
+  const limit = Math.min(filters.limit ?? 20, 100)
+  const hits = q
+    ? snap.issues.filter(
+        (i) => i.title.toLowerCase().includes(q) || i.labels.some((l) => l.toLowerCase().includes(q)),
+      )
+    : []
+  return {
+    query: filters.query ?? '',
+    results: hits.slice(0, limit).map((i) => ({ issue: i.number, title: i.title, state: i.state, labels: i.labels })),
+  }
+}
+
+/** Build the compact payload for one view. Unknown/unbuilt views return a marker. */
+export function buildProjectView(
+  view: ProjectView,
+  filters: QueryFilters | undefined,
+  snap: ProjectSnapshot,
+  asOf: Date,
+): unknown {
+  const f = filters ?? {}
+  switch (view) {
+    case 'focus':
+      return focusView(snap)
+    case 'board':
+      return boardView(snap, asOf)
+    case 'calibration':
+      return calibrationView(snap, asOf)
+    case 'issue':
+      return issueView(snap, f, asOf)
+    case 'search':
+      return searchView(snap, f)
+    default:
+      return { notImplemented: view }
+  }
+}

packages/core/src/index.ts

@@ -74,3 +74,13 @@ export type {
   CalibrationSample,
   PersonBias,
 } 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 { pickModel } from './agent/model-router.js'
+export type { ModelRouter, TaskKind } from './agent/model-router.js'
+export { runAgentTurn } from './agent/agent-loop.js'
+export type { AgentStep, AgentTurn, ToolExecutor } from './agent/agent-loop.js'
+export { QUERY_PROJECT_TOOL, REGINALD_SYSTEM, REGINALD_TOOLS } from './agent/agent-tools.js'
+export { buildProjectView } from './agent/query-project.js'
+export type { ProjectSnapshot, ProjectView, QueryFilters } from './agent/query-project.js'