thinkpool-pair 0.7.176 → 0.7.178

package/account.mjs

@@ -16,7 +16,7 @@ import os from 'node:os'
 import { saveAuth, loadAuth, loadDirs, bindDir, loadServed, rememberServed, loadDefaultDir, saveDefaultDir, claudeRecentProjectDir } from './auth-store.mjs'
 import { isSafeToRestart } from './update-gate.mjs'
 import { makeThrottledTrack, presenceSelfEchoVerdict } from './presence.mjs'
-import { publicKeyB64, announceProviders, listProviders, addProvider, removeProvider, unseal } from './providers.mjs'
+import { publicKeyB64, announceProviders, listProviders, addProvider, addProviderModel, removeProvider, unseal } from './providers.mjs'
 import { supervisorServes } from './serve-consent.mjs'
 import { resolveServeDir } from './serve-dir.mjs'
 import { pairKeyFor, pairTopic, CROSSROOM_BUS } from './cross-terminal.mjs'
@@ -579,6 +579,17 @@ export async function runAccount(SUPABASE_URL, SUPABASE_ANON) {
     if (r.ok) { pushPresence(); process.stderr.write(`\n  ◆ provider added (${String(fields?.name || '').slice(0, 40)}) — re-announced.\n`) }
     provReply('provider-add-res', { nonce, ok: r.ok, error: r.ok ? undefined : r.error, id: r.ok ? r.id : undefined })
   })
+  // provider-add-model {id, model, name?, nonce, jwt} → clone the source row's
+  // baseUrl + key HOST-SIDE onto a new model, re-announce. Deliberately carries NO
+  // `sealed` envelope: the key never leaves the bridge for this op, so there is
+  // nothing to decrypt. Owner-authed exactly like its siblings.
+  acct.on('broadcast', { event: 'provider-add-model' }, async ({ payload }) => {
+    const nonce = payload?.nonce
+    if (!(await isOwner(payload?.jwt))) return provReply('provider-add-model-res', { nonce, ok: false, error: 'unauthorized' })
+    const r = addProviderModel({ id: payload?.id, model: payload?.model, name: payload?.name })
+    if (r.ok) { pushPresence(); process.stderr.write(`\n  ◆ model added to an existing provider (${String(payload?.model || '').slice(0, 40)}) — re-announced.\n`) }
+    provReply('provider-add-model-res', { nonce, ok: r.ok, error: r.ok ? undefined : r.error, id: r.ok ? r.id : undefined })
+  })
   // provider-remove {id, nonce, jwt} → remove (built-in refuses), re-announce.
   acct.on('broadcast', { event: 'provider-remove' }, async ({ payload }) => {
     const nonce = payload?.nonce

package/agent-notify.mjs

@@ -0,0 +1,131 @@
+/* ─────────────────────────────────────────────────────────────
+   bridge/agent-notify.mjs — the PURE decision layer for agent push events.
+
+   The bridge is the only process that knows a turn ended or a permission
+   card went unanswered. It can't call /api/notify-code directly (that
+   endpoint is webhook-secret-gated and this is a published npm package), so
+   it writes a compact `__agent` row into code_messages and lets the existing
+   push_notify_on_code_message trigger fan it out.
+
+   Everything that DECIDES whether to write such a row lives here, with no
+   I/O, so bridge/agent-notify.test.mjs can exercise it without a room, a
+   socket, or a wall clock.
+
+   Spec: docs/specs/2026-07-09-code-notifications-rework.md (Slice 3)
+   ───────────────────────────────────────────────────────────── */
+
+// A turn shorter than this means somebody was almost certainly watching it —
+// pushing "done" 4 seconds after they hit enter is noise, not news.
+export const DEFAULT_MIN_TURN_MS = 30_000
+
+// How long a permission card may sit unanswered before we assume nobody is
+// looking at the room and push. Short enough to be useful, long enough that
+// the common case (you're right there, you click it) never fires.
+export const DEFAULT_PERM_GRACE_MS = 20_000
+
+// Push copy is a banner, not a transcript.
+export const SUMMARY_MAX = 200
+
+/* A lane is USER-FACING when a human opened it and is waiting on it.
+   - spawnedBy: a Dispatch worker lane an agent fanned out. Its owner is
+     another agent; the human never asked for it directly.
+   - flowTaskKey: a Flow slice lane. Same reasoning — the conductor collects
+     it, and a 12-lane flow would fan 12 banners at one person.
+   Both run in inherited/bypass modes almost always, so they also rarely
+   raise permission cards. The conductor lane itself (flowSessionId but no
+   flowTaskKey) IS user-facing — that's the one a human is watching. */
+export function isUserFacingLane(entry) {
+  if (!entry) return false
+  return !entry.spawnedBy && !entry.flowTaskKey
+}
+
+/* Collapse whitespace, clip, ellipsize. Never throws on non-strings. */
+export function clipSummary(text, max = SUMMARY_MAX) {
+  if (text == null) return ''
+  const s = String(text).replace(/\s+/g, ' ').trim()
+  if (s.length <= max) return s
+  return s.slice(0, max).replace(/\s+\S*$/, '').trimEnd() + '…'
+}
+
+/* Should a settled turn produce a `turn-done` event row?
+
+   No when: the lane isn't user-facing · the user pressed Stop (subtype
+   'aborted' — they know it's over, they're the one who ended it) · the turn
+   was short enough that somebody was plainly watching · we never saw the
+   turn start (startedAt null → a stray `error` event outside any turn, e.g.
+   the SDK-regression warning emitted at init).
+
+   `error` turns DO notify — with the same neutral copy as a success. Max's
+   call: "the agent stopped" is the news; the reason is in the room. */
+export function shouldNotifyTurnDone({ entry, subtype, startedAt, now, minTurnMs = DEFAULT_MIN_TURN_MS }) {
+  if (!isUserFacingLane(entry)) return false
+  if (subtype === 'aborted') return false
+  if (!startedAt) return false
+  if (!Number.isFinite(now)) return false
+  return (now - startedAt) >= minTurnMs
+}
+
+/* The one-line "what is it asking me" for a permission card.
+   AskUserQuestion cards carry `questions`; ordinary tool cards degrade to
+   `<ToolName> <first meaningful arg>`, matching the local stderr line. */
+export function permissionSummary(payload) {
+  const p = payload || {}
+  const qs = Array.isArray(p.questions) ? p.questions.map(q => q?.question).filter(Boolean) : []
+  if (qs.length) return clipSummary(qs.join(' / '))
+  if (p.risk === 'plan') return clipSummary(p.plan ? `Plan ready: ${p.plan}` : 'A plan is ready for review')
+  const i = p.input
+  const arg = i && (i.command || i.file_path || i.path || i.pattern || i.url || i.query || i.description)
+  const name = p.toolName || 'A tool'
+  return clipSummary(arg ? `${name} ${arg}` : name)
+}
+
+/* Per-session permission-card notifier.
+
+   arm(id, payload)  — a card was raised. Start the grace timer.
+   resolve(id)       — the card was answered (human click, mode-flip
+                       auto-approval, or a drain-to-deny). If its timer had
+                       already FIRED we pushed a banner, so retract it.
+   clearAll()        — turn ended / session died. Every outstanding card is
+                       dead; retract anything we announced.
+
+   Timers are injected so the unit test drives them synchronously. onFire /
+   onRetract are fire-and-forget (they end up as REST inserts); a throw in
+   either must never take the bridge down. */
+export function createPermNotifier({
+  graceMs = DEFAULT_PERM_GRACE_MS,
+  onFire = () => {},
+  onRetract = () => {},
+  setTimer = setTimeout,
+  clearTimer = clearTimeout,
+} = {}) {
+  const cards = new Map()   // id → { timer, fired, payload }
+  const safe = (fn, id, payload) => { try { fn(id, payload) } catch { /* never break the turn */ } }
+
+  const settle = (id) => {
+    const rec = cards.get(id)
+    if (!rec) return
+    if (rec.timer) { try { clearTimer(rec.timer) } catch { /* noop */ } }
+    cards.delete(id)
+    if (rec.fired) safe(onRetract, id, rec.payload)
+  }
+
+  return {
+    arm(id, payload) {
+      if (!id || cards.has(id)) return
+      const rec = { timer: null, fired: false, payload }
+      rec.timer = setTimer(() => {
+        rec.timer = null
+        rec.fired = true
+        safe(onFire, id, payload)
+      }, graceMs)
+      // Never hold the process open on a permission timer.
+      try { rec.timer?.unref?.() } catch { /* not a node timer (test double) */ }
+      cards.set(id, rec)
+    },
+    resolve: settle,
+    clearAll() { for (const id of [...cards.keys()]) settle(id) },
+    // test/debug introspection only
+    size: () => cards.size,
+    fired: (id) => !!cards.get(id)?.fired,
+  }
+}

package/bridge.mjs

@@ -39,10 +39,13 @@ import { randomUUID } from 'node:crypto'
 import { createClient } from '@supabase/supabase-js'
 import { serveDecision, fetchServeRow, refusalMessage, gateFailAction } from './serve-consent.mjs'
 import { reapTerminalRow as _reapTerminalRow } from './reap-terminal.mjs'
+// Slice 3 — the pure decision layer for agent push events (turn-done / needs-input).
+// Unit-tested in bridge/agent-notify.test.mjs; everything with I/O stays here.
+import { createPermNotifier, shouldNotifyTurnDone, permissionSummary, clipSummary, isUserFacingLane, DEFAULT_MIN_TURN_MS } from './agent-notify.mjs'
 // resolveProviderEnv(id) → {ANTHROPIC_BASE_URL,ANTHROPIC_AUTH_TOKEN,ANTHROPIC_MODEL} for a
 // registered custom provider, or null for the built-in/unknown (leave the default env intact).
 // Multi-provider BYOK slice 1: a lane spawned with a `provider` id runs on that endpoint.
-import { resolveProviderEnv, providerNameMap, publicKeyB64, announceProviders, listProviders, addProvider, removeProvider, unseal } from './providers.mjs'
+import { resolveProviderEnv, providerNameMap, publicKeyB64, announceProviders, listProviders, addProvider, removeProvider, unseal, effectiveLaneModel } from './providers.mjs'
 import { validateProviderSwitch, BUILTIN_PROVIDER } from './switch-provider.mjs'
 import { createSdkMcpServer, tool } from '@anthropic-ai/claude-agent-sdk'
 import { z } from 'zod'
@@ -763,6 +766,46 @@ async function budgetRpc (fn, args) {
   const t = await r.text()
   return t ? JSON.parse(t) : null
 }
+
+// ── Agent push events — spec docs/specs/2026-07-09-code-notifications-rework.md (Slice 3).
+// The bridge is the only process that knows a turn ended or a permission card went
+// unanswered, and it CANNOT call /api/notify-code (webhook-secret-gated; this is a
+// published npm package). So it writes a compact `__agent` row into code_messages with
+// the SAME authed REST path budgetRpc uses; the existing push_notify_on_code_message
+// trigger fans it out and api/notify-code.js routes on payload.__agent.
+//
+// Durable by construction: the row is also the transport-neutral event log the iPhone
+// sender will read, and it feeds get_code_unread_counts for free.
+//
+// Fire-and-forget in the strictest sense — an anon bridge (no codeAuthToken/uid), a
+// dead network, or a 4xx must never surface in the turn. TP_AGENT_NOTIFY_OFF=1 kills
+// the whole path. The web client filters `ae:` rows out of its seed/replay (Slice 2),
+// so an old client paired with a new bridge renders nothing new.
+const AGENT_NOTIFY_OFF = process.env.TP_AGENT_NOTIFY_OFF === '1'
+const MIN_TURN_MS = Number(process.env.TP_NOTIFY_MIN_TURN_MS) > 0
+  ? Number(process.env.TP_NOTIFY_MIN_TURN_MS) : DEFAULT_MIN_TURN_MS
+function persistAgentEvent (payload) {
+  if (AGENT_NOTIFY_OFF || !codeAuthToken || !myServeUid || !room) return
+  const row = {
+    id: `ae:${randomUUID()}`,
+    session_code: room,
+    author_id: myServeUid,
+    payload: { __agent: true, ts: Date.now(), ...payload },
+  }
+  fetch(`${SUPABASE_URL}/rest/v1/code_messages`, {
+    method: 'POST',
+    headers: {
+      apikey: SUPABASE_ANON,
+      Authorization: `Bearer ${codeAuthToken}`,
+      'Content-Type': 'application/json',
+      Prefer: 'return=minimal',
+    },
+    body: JSON.stringify(row),
+  }).then((r) => {
+    if (!r.ok && process.env.TP_DEBUG) process.stderr.write(`\n  ◆ agent-event insert ${r.status}\n`)
+  }).catch(() => { /* best-effort — a missed banner is not worth a stack trace */ })
+}
+
 // Fire-and-forget: fold a completed turn's output tokens into the room's durable monthly
 // ledger. output_tokens = the billed reasoning+output spend the indicator already tracks.
 function recordRoomSpend (tokens) {
@@ -1525,23 +1568,40 @@ function worktreeSnapshot(cwd) {
 // rolling log replays to joiners and survives bridge restarts (session-store).
 function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rolePrompt, flowSessionId, flowTaskKey, cwd, reviewSliceRoots, openedAt, defer, provider, carryRecap }) {
   if (sessions.has(id)) return
+  // The lane's EFFECTIVE model — computed ONCE and used for BOTH the truthful display
+  // label (entry.model) and the SDK's `model` option below. These used to be computed
+  // separately: the label resolved the provider's configured model while the SDK got the
+  // RAW inherited `model`, which OVERRIDES ANTHROPIC_MODEL from resolveProviderEnv. A GLM
+  // lane opened while focused on Fable therefore showed `glm-4.6` and sent
+  // `claude-fable-5` → 400 [1211][Unknown Model]. One value, one truth.
+  // undefined → omit the SDK option entirely and let the provider env/endpoint default.
+  const laneModel = effectiveLaneModel({ provider, model, configuredModel: resolveProviderEnv(provider)?.ANTHROPIC_MODEL })
   // spawnedBy: set when this lane was Dispatched (spawn_terminal). Restored from the
   // session store so the Ensemble flag survives a bridge restart (else a respin
   // stripped it and the lane reverted to a plain tab — the t6 "no chip" bug).
   const entry = { cmd: 'claude', kind: 'structured', log: Array.isArray(log) ? log.slice(-STRUCTURED_LOG_MAX) : [], pending: new Map(), session: null, recovered: false, commands: Array.isArray(commands) ? commands : undefined, mode,
-    // model: truthful active-model label. When this lane runs on a custom (non-anthropic)
-    // registered provider the SDK id is impersonated (see the onEvent guard below), so seed
-    // from THAT provider's CONFIGURED model — glm-5.2, not a Claude label — before any turn.
-    // No configured model → the provider's name, never a false Claude id. Built-in Anthropic
-    // keeps the passed model. Keyed on the lane's own provider id, not the bridge default.
+    // model: truthful active-model label — now the SAME `laneModel` the SDK is given, so the
+    // chip cannot disagree with the wire. When this lane runs on a custom (non-anthropic)
+    // registered provider the SDK id is impersonated (see the onEvent guard below), so
+    // laneModel resolves from THAT provider's CONFIGURED model — glm-4.6, not a Claude label.
+    // No configured model → the provider's NAME, never a false Claude id (an inherited
+    // claudeish id was dropped upstream by effectiveLaneModel; it must not sneak back in as a
+    // label fallback here). Built-in Anthropic keeps the passed model.
     model: (!provider || provider === 'anthropic')
-      ? (model || null)
-      : (resolveProviderEnv(provider)?.ANTHROPIC_MODEL || providerNameMap()[provider] || model || provider),
+      ? (laneModel || null)
+      : (laneModel || providerNameMap()[provider] || provider),
     provider: provider || null, spawnedBy: spawnedBy || undefined, flowSessionId: flowSessionId || null, flowTaskKey: flowTaskKey || null, cwd: cwd || null,
     // Stable creation order — persisted so a bridge restart restores tabs in the SAME
     // order (not readdir/filesystem order). Legacy recs (no openedAt) derive it from the
     // first transcript event ts, so even the first post-fix restart is ordered right.
     openedAt: openedAt || (Array.isArray(log) ? (log.find((e) => e?.ts)?.ts || 0) : 0) || Date.now() }
+  // Slice 3 — a permission card left unanswered past the grace window pushes
+  // "<lane> — needs you: <what>"; answering it anywhere retracts the banner
+  // everywhere. Worker/flow lanes are excluded at arm() time (isUserFacingLane).
+  entry.permNotifier = createPermNotifier({
+    onFire:    (_cardId, summary) => persistAgentEvent({ kind: 'needs-input',    term: id, termName: termNames[id] || null, summary }),
+    onRetract: (_cardId)          => persistAgentEvent({ kind: 'needs-resolved', term: id, termName: termNames[id] || null }),
+  })
   // C1 (RT-2): per-term contiguous seq counter. Seeded from the restored log's max
   // so a bridge restart resumes ABOVE every persisted seq — fresh events never reuse
   // an old seq, which is what keeps the between-turns restart dup-free (see
@@ -2045,7 +2105,15 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
   // is silently lost (the 2026-07-02 auto-resume bug). Empty string → nothing to carry.
   entry.pendingRecap = (typeof carryRecap === 'string' && carryRecap.trim()) ? carryRecap : null
   entry.session = startClaudeSession({
-    cwd: cwd || process.cwd(), model, resume, mode, rolePrompt,
+    // laneModel, NOT the raw `model` param: the SDK's `model` option OVERRIDES the
+    // ANTHROPIC_MODEL supplied by resolveProviderEnv() in `env` below, so an inherited
+    // Claude id here defeats the whole per-provider env. undefined → option omitted.
+    cwd: cwd || process.cwd(), model: laneModel, resume, mode, rolePrompt,
+    // Haiku composer-prefill is an Anthropic-subscription freebie. This lane's env carries a
+    // CUSTOM provider's base URL + auth token, so on a non-anthropic lane that call would bill
+    // the user's own BYOK key for a feature they didn't ask for (and 400 on strict providers).
+    // Same lane-provider key as the model-truth guard in onEvent below.
+    suggest: !provider || provider === 'anthropic',
     // Live ROOM NOW tail appended to every turn's <system-reminder>: sibling lanes
     // (same registries the read_terminal roster uses) + active git worktrees (only
     // when >1 — a solo checkout has nothing to coordinate about). Keeps every lane
@@ -2135,8 +2203,11 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
       // roster broadcast (delivered to every subscriber) instead of sniffing the live
       // code-event stream a passive dashboard channel misses. Covers every turn path
       // (normal, spawn, flow lane, cross-room) since they all flow through onEvent.
+      // The RISING edge is also where a turn's wall clock starts — the only place the
+      // bridge sees "this lane just began working". `result`/`error` reads it to decide
+      // whether the turn was long enough that nobody was watching (Slice 3).
       { const busyNow = entry.session?.turnActive ?? false
-        if (busyNow !== entry._busyAnn) { entry._busyAnn = busyNow; announce() } }
+        if (busyNow !== entry._busyAnn) { if (busyNow) entry._turnStart = Date.now(); entry._busyAnn = busyNow; announce() } }
       // Capture this session's current MODEL (resolved full id, with version — from
       // system/result `model` or usage `ctx.model`) and the SDK's supported-model
       // LIST (from the `models` event) so the announce carries them to EVERY device.
@@ -2247,6 +2318,30 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
         bcast('code-event', { term: id, evt: { kind: 'compact', status: 'done', ts: Date.now() } })
         entry.compactBy = null
       }
+      // ── Slice 3: the turn settled. Two things happen, in this order.
+      //   1. Every outstanding permission card on this lane is now dead (the SDK
+      //      won't ask again for a turn that ended) — clearAll retracts any
+      //      "needs you" banner we already pushed. Do this even for aborted turns
+      //      and non-user-facing lanes: a stale card can never fire later.
+      //   2. If the turn ran long enough that nobody was plausibly watching, and
+      //      the lane is one a human opened, and they didn't press Stop — write a
+      //      `turn-done` event row. An `error` turn uses the same neutral copy.
+      // _turnStart is cleared unconditionally so a stray post-settle event can't
+      // re-notify off a stale stamp.
+      if (evt.kind === 'result' || evt.kind === 'error') {
+        const startedAt = entry._turnStart
+        entry._turnStart = null
+        entry.permNotifier?.clearAll()
+        if (shouldNotifyTurnDone({ entry, subtype: evt.subtype, startedAt, now: Date.now(), minTurnMs: MIN_TURN_MS })) {
+          persistAgentEvent({
+            kind: 'turn-done',
+            term: id,
+            termName: termNames[id] || null,
+            summary: clipSummary(evt.kind === 'result' ? evt.resultText : evt.message),
+            subtype: evt.subtype || (evt.kind === 'error' ? 'error' : null),
+          })
+        }
+      }
       // Single tail: stamp into the log (transcript kinds only), broadcast live,
       // print locally, persist. Shared so a deferred image event re-enters it once
       // its base64 has been lifted to a Storage path (see deferImageEvent).
@@ -2281,6 +2376,13 @@ function openStructured({ id, model, resume, log, commands, mode, spawnedBy, rol
       const payload = { term: id, id: req.id, toolName: req.toolName, input: req.input, risk: req.risk, plan: req.plan, questions: req.questions }
       entry.pending.set(req.id, { resolve, payload })
       bcast('code-perm-req', payload)
+      // Slice 3 — arm the grace timer. Fires a "needs you" push only if the card is
+      // STILL unanswered 20s from now (you're not in the room). Every resolve path —
+      // the code-perm broadcast, drainPending, and the bypass/acceptEdits mode-flip
+      // auto-approvals — calls permNotifier.resolve(id), which retracts the banner
+      // if it had fired. Worker + flow-slice lanes are skipped: their cards belong to
+      // an agent's fan-out, and a 12-lane flow would fan 12 banners at one person.
+      if (isUserFacingLane(entry)) entry.permNotifier?.arm(req.id, permissionSummary(payload))
       process.stderr.write(req.risk === 'plan'
         ? `\n  ${A.mag}◆ plan ready — approve in the room.${A.rst}\n`
         : req.risk === 'ask'
@@ -2310,6 +2412,8 @@ function drainPending(s) {
   if (!s?.pending) return
   for (const [, p] of s.pending) { try { p.resolve('deny') } catch { /* noop */ } }
   s.pending.clear()
+  // Every card just died → retract any "needs you" banner they raised (Slice 3).
+  s.permNotifier?.clearAll()
 }
 
 // Switching a terminal to bypassPermissions must RETROACTIVELY clear the cards the
@@ -2329,6 +2433,7 @@ function allowPending(s) {
     if (p?.payload?.risk === 'plan' || p?.payload?.risk === 'ask') continue
     try { p.resolve('allow') } catch { /* noop */ }
     s.pending.delete(key)
+    s.permNotifier?.resolve(key)   // auto-approved → retract the "needs you" banner
   }
 }
 
@@ -2347,6 +2452,7 @@ function acceptEditsPending(s) {
     if (!ACCEPT_EDIT_TOOLS.has(tn)) continue
     try { p.resolve('allow') } catch { /* noop */ }
     s.pending.delete(key)
+    s.permNotifier?.resolve(key)   // auto-approved → retract the "needs you" banner
   }
 }
 
@@ -2742,7 +2848,9 @@ channel
     const p = s && payload.id && s.pending.get(payload.id)
     // Pass the raw decision through — normal tools use allow/deny, plan cards
     // use run/accept/keep (claude-session interprets). Default deny on missing.
-    if (p) { s.pending.delete(payload.id); p.resolve(payload.decision || 'deny') }
+    // Answered from a device → retract the "needs you" banner on every OTHER device
+    // (permNotifier.resolve is a no-op unless its grace timer already fired).
+    if (p) { s.pending.delete(payload.id); s.permNotifier?.resolve(payload.id); p.resolve(payload.decision || 'deny') }
   })
   .on('broadcast', { event: 'code-abort' }, ({ payload }) => {
     const s = payload?.term && sessions.get(payload.term)

package/claude-session.mjs

@@ -176,6 +176,8 @@ const simplifyBlocks = (blocks = []) => blocks.map((b) => {
  * @param {(req)=>Promise<'allow'|'deny'>} o.requestPermission
  *        called for EVERY tool call with { id, toolName, input, risk };
  *        resolve 'allow'/'deny'. (Caller implements any auto-allow policy.)
+ * @param {boolean=} o.suggest          allow the Haiku composer-prefill fallback.
+ *        MUST be false on a lane running a custom provider — see haikuSuggest.
  * @returns {{ sendTurn(text), abort(), end(), readonly sessionId }}
  */
 const MODES = new Set(['default', 'acceptEdits', 'plan', 'bypassPermissions'])
@@ -203,7 +205,7 @@ const TP_ROOM_REMINDER = [
   'Verify before claiming done — show runtime evidence you produced, not "should work, go test it".',
 ].join(' ')
 
-export function startClaudeSession({ cwd, model, resume, env, mode: initialMode = 'default', onEvent, requestPermission, mcpServers, crossPostGate, crossRoomPostGate, didSpawnTarget = null, rolePrompt, blockSubagents = false, onSubmitPlan = null, onLaneDone = null, onReviewVerdict = null, reviewGate = null, lazy = false, roomContext = null }) {
+export function startClaudeSession({ cwd, model, resume, env, mode: initialMode = 'default', onEvent, requestPermission, mcpServers, crossPostGate, crossRoomPostGate, didSpawnTarget = null, rolePrompt, blockSubagents = false, onSubmitPlan = null, onLaneDone = null, onReviewVerdict = null, reviewGate = null, lazy = false, roomContext = null, suggest = true }) {
   // Per-turn reminder + live ROOM NOW tail. roomContext (bridge-supplied) returns the
   // room's CURRENT state — sibling lanes, active git worktrees — or null. The static
   // rules keep the agent aware of the room's FEATURES; the live tail keeps it aware of
@@ -746,7 +748,16 @@ export function startClaudeSession({ cwd, model, resume, env, mode: initialMode
   // (opts.env carries the OAuth — no API key needed), but as a BARE model call: no
   // settingSources, no MCP, no tools — so it's a fast cold call, not a full session boot.
   // Fed only `lastAssistantText` (trimmed). Never blocks a turn; any failure is silent.
+  //
+  // BUILT-IN ANTHROPIC PATH ONLY (`suggest`). On a lane running a registered custom
+  // provider, opts.env carries that provider's ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN —
+  // so this hardcoded `claude-haiku-4-5` call would spend the USER'S OWN BYOK credit on a
+  // composer prefill they never asked for, at whatever (wrong, flagship) model the endpoint
+  // maps the unknown id to. z.ai silently serves glm for any `claude-*` id, which is exactly
+  // why this went unnoticed; a stricter provider 400s. This is an Anthropic-subscription
+  // freebie — it does not get a per-provider small model, it gets switched off.
   const haikuSuggest = async () => {
+    if (!suggest) return
     const seed = (lastAssistantText || '').trim().slice(-1500)
     if (!seed || closed) return
     const ac2 = new AbortController()
@@ -978,7 +989,13 @@ export function startClaudeSession({ cwd, model, resume, env, mode: initialMode
               }
               if (best) curModel = best
             }
-            emit({ kind: 'result', subtype: m.subtype, sessionId, model: curModel, costUsd: m.total_cost_usd, usage: m.usage, numTurns: m.num_turns, durationMs: m.duration_ms ?? null, denials: Array.isArray(m.permission_denials) ? m.permission_denials.length : 0 })
+            // resultText: the SDK's final assistant text, clipped. The bridge folds it
+            // into the `turn-done` push body ("<lane> — done: <summary>") so a banner
+            // says what happened, not just that something did. Clipped + whitespace-
+            // collapsed here so neither the broadcast frame nor the persisted log grows
+            // by a full turn's prose. Absent on aborted/error results — nothing to say.
+            { const rt = typeof m.result === 'string' ? m.result.replace(/\s+/g, ' ').trim().slice(0, 200) : ''
+              emit({ kind: 'result', subtype: m.subtype, sessionId, model: curModel, costUsd: m.total_cost_usd, usage: m.usage, numTurns: m.num_turns, durationMs: m.duration_ms ?? null, denials: Array.isArray(m.permission_denials) ? m.permission_denials.length : 0, resultText: rt || null }) }
             // Arm the Haiku fallback: if Claude's own prompt_suggestion doesn't arrive within
             // the window, generate one from the last assistant reply. (Skip aborted turns.)
             if (sugTimer) clearTimeout(sugTimer)

package/event-id.mjs

@@ -215,6 +215,14 @@ export function summarizeUsage(logEvents) {
   return { turns: results.length, per: [...per.values()], totalIn, totalOut, totalCost, haveCost }
 }
 
+// Does this model id name a Claude/Anthropic model? Used by two callers with the
+// same question and NO room for a second opinion:
+//   • usageReportLine — token/cost reporting is Claude-only.
+//   • providers.mjs/effectiveLaneModel — a lane on a CUSTOM (non-anthropic) provider
+//     must never be handed an inherited Claude id; the endpoint 400s on it.
+// Exported (not a local const) precisely so there is exactly one definition.
+export const claudeish = (m) => /claude|anthropic|opus|sonnet|haiku|fable/i.test(String(m || ''))
+
 // model id → friendly label for the /usage report: claude-opus-4-8 → "Opus 4.8".
 // Bridge-local twin of the web's prettyModel (structured.jsx); falls back to the raw id.
 export function prettyModelId (m) {
@@ -229,7 +237,6 @@ export function prettyModelId (m) {
 // handler is a thin wrapper: ctlLine(usageReportLine(s.model, s.log)). `sessionModel` is
 // the resolved model id the /model state tracks; `logEvents` the session's retained log.
 export function usageReportLine (sessionModel, logEvents) {
-  const claudeish = (m) => /claude|anthropic|opus|sonnet|haiku|fable/i.test(String(m || ''))
   const sum = summarizeUsage(logEvents)
   // Gate (Max's ask: "make /usage work if users bridge is on claude llms"). The Code bridge
   // runs the Claude Agent SDK, so a normal lane IS a Claude model — but a session pointed at a

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thinkpool-pair",
-  "version": "0.7.176",
+  "version": "0.7.178",
   "description": "Share a local coding-agent CLI (Claude Code, Codex, Gemini, Aider, …) into a ThinkPool Code room, live.",
   "type": "module",
   "bin": {
@@ -15,6 +15,7 @@
     "claude-session.mjs",
     "turn-stall.mjs",
     "update-gate.mjs",
+    "agent-notify.mjs",
     "event-id.mjs",
     "recap.mjs",
     "transcript-sanitize.mjs",

package/providers.mjs

@@ -42,6 +42,7 @@ import os from 'node:os'
 import fs from 'node:fs'
 import path from 'node:path'
 import crypto from 'node:crypto'
+import { claudeish } from './event-id.mjs'
 
 const DIR       = path.join(os.homedir(), '.thinkpool-pair')
 const REG_FILE  = path.join(DIR, 'providers.json')
@@ -180,6 +181,59 @@ export function addProvider(fields) {
   return { ok: true, id }
 }
 
+/**
+ * Add ANOTHER model to an already-registered provider — without the user ever
+ * retyping the key (Max, 2026-07-09). The registry's unit is a {provider,model}
+ * pair, not a provider, and `buildSwitchMenu` already groups rows by brand
+ * ("Z.ai GLM 4.5" / "Z.ai GLM Air" collapse under one "z.ai" header) — so a new
+ * model is simply a CLONE of the source row with a fresh id and a new model.
+ * No schema change, no new menu shape.
+ *
+ * The clone copies `key` + `baseUrl` HOST-SIDE from the row that already holds
+ * them, so no key material crosses the wire for this operation — the caller's
+ * RPC carries no `sealed` envelope at all (see account.mjs `provider-add-model`).
+ *
+ * @param {{id:string, model:string, name?:string}} fields  id = the SOURCE provider
+ * @returns {{ok:true,id:string} | {ok:false,error:string}}
+ */
+export function addProviderModel({ id, model, name } = {}) {
+  if (!id || id === BUILTIN_ID) return { ok: false, error: 'the built-in Anthropic provider has no key to reuse' }
+  if (!model || typeof model !== 'string' || !model.trim()) return { ok: false, error: 'model is required' }
+  const arr = loadProviders()
+  const src = arr.find((p) => p.id === id)
+  if (!src) return { ok: false, error: 'no such provider' }
+
+  const nextModel = String(model).trim()
+  // An endpoint+model pair is the identity of a row — a second row for the same
+  // pair would render as a duplicate menu entry that switches to nothing new.
+  if (arr.some((p) => p.baseUrl === src.baseUrl && (p.model || '') === nextModel)) {
+    return { ok: false, error: `${nextModel} is already registered on this provider` }
+  }
+
+  const nextName = (name && String(name).trim()) || `${brandOf(src.name)} ${nextModel}`.trim()
+  const newId = crypto.randomUUID()
+  arr.push({
+    id: newId,
+    name: nextName,
+    baseUrl: src.baseUrl,   // copied host-side
+    model: nextModel,
+    key: src.key,           // copied host-side — never re-sent by the client
+    addedAt: Date.now(),
+  })
+  saveProviders(arr)
+  return { ok: true, id: newId }
+}
+
+/**
+ * First name token — the brand, matching code-constants.js `brandKey`. "Z.ai GLM
+ * 4.5" → "Z.ai". Used to derive a default name for a cloned row so the switch
+ * menu groups it under the SAME header as its source.
+ */
+function brandOf(name) {
+  const first = String(name || '').trim().split(/\s+/)[0]
+  return first || 'Provider'
+}
+
 /**
  * Remove a provider by id. The built-in `anthropic` refuses.
  * @returns {{ok:true} | {ok:false,error:string}}
@@ -244,3 +298,35 @@ export function resolveProviderEnv(id) {
   if (p.model) env.ANTHROPIC_MODEL = p.model
   return env
 }
+
+/**
+ * The ONE model id a lane actually runs on — used for BOTH the display label and the
+ * SDK's `model` option, which is the whole point: they drifted, and the SDK option
+ * silently OVERRIDES the ANTHROPIC_MODEL env resolveProviderEnv() supplies.
+ *
+ * The bug this closes: a new lane inherits the focused lane's model (client-side), so
+ * opening a GLM lane while on Fable sent `model: 'claude-fable-5'` to
+ * https://api.z.ai/api/anthropic → `API Error: 400 [1211][Unknown Model]`, while the
+ * chip truthfully read `glm-4.6`. Label and wire disagreed; the wire lost.
+ *
+ * Rules (custom = any registered, non-built-in provider):
+ *   • built-in anthropic / no provider → passthrough, unchanged. Default path is sacred.
+ *   • custom + model equals the configured model  → honour it.
+ *   • custom + model is a Claude id (claudeish)   → DROP it. It leaked across a provider
+ *       boundary and the endpoint cannot serve it. Fall back to the configured model, or
+ *       to undefined (no `model` option → the env / endpoint default decides).
+ *   • custom + explicit non-Claude model (glm-4.5, deepseek-…) → honour it. This is a
+ *       deliberate pick (/model, spawn_terminal({provider, model})), not an inheritance leak.
+ *
+ * NEVER returns an inherited Claude id on a custom provider — that is the invariant.
+ * Pure: `configuredModel` is passed in (not read from disk) so it is directly testable.
+ *
+ * @param {{provider?:string|null, model?:string|null, configuredModel?:string|null}} a
+ * @returns {string|undefined}  the model to send; undefined → omit the SDK `model` option
+ */
+export function effectiveLaneModel({ provider, model, configuredModel } = {}) {
+  if (!provider || provider === BUILTIN_ID) return model || undefined
+  if (model && model === configuredModel) return model
+  if (model && !claudeish(model)) return model
+  return configuredModel || undefined
+}