thinkpool-pair 0.7.237 → 0.7.239

package/bridge.mjs

@@ -50,8 +50,9 @@ import { validateProviderSwitch, providerSwitchPlan, BUILTIN_PROVIDER } from './
 import { createSdkMcpServer, tool } from '@anthropic-ai/claude-agent-sdk'
 import { z } from 'zod'
 import { readCodexDefaultModel, readCodexModels, codexConfigForMode, codexThreadCanResume } from './codex-session.mjs'
+import { withMcpSessionFactory } from './codex-mcp-http.mjs'
 import { startStructuredSession } from './runtime-session.mjs'
-import { defaultStructuredMode, structuredRuntimeForCommand, structuredRuntimeMetadata, structuredRuntimeSupportsMode } from './runtime-registry.mjs'
+import { defaultStructuredMode, shouldDeferStructuredRuntime, structuredRuntimeForCommand, structuredRuntimeMetadata, structuredRuntimeSupportsMode } from './runtime-registry.mjs'
 import { probeHermesRuntime } from './hermes-probe.mjs'
 import { canonicalRoomFilePath, waitForNativeImages } from './codex-images.mjs'
 import { createManagedLaneWorktree, removeManagedLaneWorktree } from './lane-worktree.mjs'
@@ -1694,7 +1695,7 @@ function worktreeSnapshot(cwd) {
 // relay STRUCTURED events. onEvent → broadcast `code-event` + print locally +
 // persist to the host file; tool calls round-trip through the perm card; the
 // rolling log replays to joiners and survives bridge restarts (session-store).
-function openStructured({ id, runtime = 'claude', model, effort, resume, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, rolePrompt, flowSessionId, flowTaskKey, flowRole, flowReviewTarget, flowReviewTargets, flowReviewRound, dispatchBaseSha, revertTarget, cwd, managedWorktree, reviewSliceRoots, openedAt, defer, provider, carryRecap, lastUsage }) {
+function openStructured({ id, runtime = 'claude', model, models, effort, resume, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, rolePrompt, flowSessionId, flowTaskKey, flowRole, flowReviewTarget, flowReviewTargets, flowReviewRound, dispatchBaseSha, revertTarget, cwd, managedWorktree, reviewSliceRoots, openedAt, defer, provider, carryRecap, lastUsage }) {
   if (sessions.has(id)) return
   runtime = structuredRuntimeMetadata(runtime) ? runtime : 'claude'
   // No explicit mode → a sensible default per runtime (see defaultModeForRuntime):
@@ -1733,7 +1734,7 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
     ? spawnDepth
     : (sideParent || (spawnedBy && !String(spawnedBy).startsWith('flow:')) ? 1 : 0)
   const initialHop = Number.isInteger(hop) && hop >= 0 ? hop : structuralDepth
-  const entry = { cmd: runtime, runtime, 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, effort, models: runtime === 'codex' ? codexModels : undefined,
+  const entry = { cmd: runtime, runtime, 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, effort, models: runtime === 'codex' ? codexModels : runtime === 'hermes' && Array.isArray(models) ? models : undefined,
     // 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
@@ -1996,7 +1997,7 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
   // restart. Without this, sessionData omitted it → on restart the resumed session
   // re-launched on the host default (Opus) regardless of the last switch, and the
   // switch looked like it "never changed the model" (Max 2026-07-02). Restored below.
-  const sessionData = () => ({ sessionId: entry.session?.sessionId || null, runtime: entry.runtime, log: entry.log, commands: entry.commands, mode: entry.mode, effort: entry.effort, model: entry.model || null, provider: entry.provider || null, spawnedBy: entry.spawnedBy, spawnDepth: entry.spawnDepth || 0, cascadeRole: entry.cascadeRole || null, hop: entry.hop || 0, sideParent: entry.sideParent, sideTask: entry.sideTask, pendingSideContexts: entry.pendingSideContexts, flowSessionId: entry.flowSessionId, flowTaskKey: entry.flowTaskKey, flowRole: entry.flowRole || null, flowReviewTargets: entry.flowReviewTargets || [], flowReviewRound: entry.flowReviewRound || 0, dispatchBaseSha: entry.dispatchBaseSha || null, cwd: entry.cwd, managedWorktree: entry.managedWorktree, rolePrompt: entry.rolePrompt, flowReviewTarget: entry.flowReviewTarget || null, revertTarget: entry.revertTarget || null, reviewSliceRoots: entry.reviewSliceRoots || [], openedAt: entry.openedAt || null, lastUsage: entry.lastUsage || null, carryRecap: entry.pendingRecap || null })
+  const sessionData = () => ({ sessionId: entry.session?.sessionId || null, runtime: entry.runtime, log: entry.log, commands: entry.commands, mode: entry.mode, effort: entry.effort, model: entry.model || null, models: entry.runtime === 'hermes' ? (entry.models || []) : undefined, provider: entry.provider || null, spawnedBy: entry.spawnedBy, spawnDepth: entry.spawnDepth || 0, cascadeRole: entry.cascadeRole || null, hop: entry.hop || 0, sideParent: entry.sideParent, sideTask: entry.sideTask, pendingSideContexts: entry.pendingSideContexts, flowSessionId: entry.flowSessionId, flowTaskKey: entry.flowTaskKey, flowRole: entry.flowRole || null, flowReviewTargets: entry.flowReviewTargets || [], flowReviewRound: entry.flowReviewRound || 0, dispatchBaseSha: entry.dispatchBaseSha || null, cwd: entry.cwd, managedWorktree: entry.managedWorktree, rolePrompt: entry.rolePrompt, flowReviewTarget: entry.flowReviewTarget || null, revertTarget: entry.revertTarget || null, reviewSliceRoots: entry.reviewSliceRoots || [], openedAt: entry.openedAt || null, lastUsage: entry.lastUsage || null, carryRecap: entry.pendingRecap || null })
   const persist = () => saveSession(room, id, sessionData())
   // Synchronous flush of this session's record. Used on open (so a brand-new session
   // has a file under its id BEFORE its first event — surviving a restart inside the
@@ -2052,7 +2053,7 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
   // own `id` (excluded) + `entry` (per-turn call counter, reset by the code-turn
   // handler). Auto-allowed by the PreToolUse gate (mcp__thinkpool__ prefix).
   // Spec: docs/specs/2026-06-28-cross-terminal-awareness.md
-  const peekServer = createSdkMcpServer({
+  const createPeekServer = () => createSdkMcpServer({
     name: 'thinkpool',
     version: '1.0.0',
     tools: [
@@ -2420,6 +2421,7 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
       )] : []),
     ],
   })
+  const peekServer = withMcpSessionFactory(createPeekServer)
   // S5 (slice 1b) — the review-lane write-block, wired into the live PreToolUse path.
   // Null for builder lanes (no reviewSliceRoots) → their toolset is UNCHANGED. For a review
   // lane, deny a mutating tool call if it targets ANY reviewed slice's worktree, allowing the
@@ -2486,7 +2488,7 @@ function openStructured({ id, runtime = 'claude', model, effort, resume, log, co
     },
     // defer: a restored-idle terminal cold-boots its query only on first use — restore is
     // then instant for display, and only used terminals pay the ~50s MCP boot (Max, 2026-07-02).
-    lazy: !!defer || (!cwd && !flowSessionId && !flowTaskKey),
+    lazy: shouldDeferStructuredRuntime({ runtime, defer, cwd, flowSessionId, flowTaskKey, models: entry.models }),
     reviewGate,
     // A Flow CONDUCTOR (flowSessionId set, no flowTaskKey) must DECOMPOSE, not explore via
     // a fan-out of Task/Explore subagents, and must not build. Hard-block Task/Bash/Write for
@@ -3808,7 +3810,7 @@ channel
             // FL-M6 — restore the flow context (id/role/cwd) so an in-flight flow survives a
             // bridge restart: the conductor keeps its subagent-block + plan interception, and
             // lanes keep their worktree cwd + the ability to mark done.
-            openStructured({ id: rec.id, runtime: rec.runtime || 'claude', model: rec.model || undefined, effort: rec.effort, provider: rec.runtime === 'claude' ? rec.provider || undefined : undefined, resume: resumable ? rec.sessionId : undefined, log: rec.log, commands: rec.commands, mode: rec.mode, spawnedBy: rec.spawnedBy, spawnDepth: rec.spawnDepth, cascadeRole: rec.cascadeRole, hop: rec.hop, sideParent: rec.sideParent, sideTask: rec.sideTask, pendingSideContexts: rec.pendingSideContexts, flowSessionId: rec.flowSessionId, flowTaskKey: rec.flowTaskKey, flowRole: rec.flowRole, flowReviewTarget: rec.flowReviewTarget, flowReviewTargets: rec.flowReviewTargets, flowReviewRound: rec.flowReviewRound, dispatchBaseSha: rec.dispatchBaseSha, revertTarget: rec.revertTarget, cwd: rec.cwd, managedWorktree: rec.managedWorktree, rolePrompt: rec.rolePrompt, reviewSliceRoots: rec.reviewSliceRoots, openedAt: rec.openedAt, lastUsage: rec.lastUsage, carryRecap: wasInterrupted ? recoveryRecap : rec.carryRecap,
+            openStructured({ id: rec.id, runtime: rec.runtime || 'claude', model: rec.model || undefined, models: rec.models, effort: rec.effort, provider: rec.runtime === 'claude' ? rec.provider || undefined : undefined, resume: resumable ? rec.sessionId : undefined, log: rec.log, commands: rec.commands, mode: rec.mode, spawnedBy: rec.spawnedBy, spawnDepth: rec.spawnDepth, cascadeRole: rec.cascadeRole, hop: rec.hop, sideParent: rec.sideParent, sideTask: rec.sideTask, pendingSideContexts: rec.pendingSideContexts, flowSessionId: rec.flowSessionId, flowTaskKey: rec.flowTaskKey, flowRole: rec.flowRole, flowReviewTarget: rec.flowReviewTarget, flowReviewTargets: rec.flowReviewTargets, flowReviewRound: rec.flowReviewRound, dispatchBaseSha: rec.dispatchBaseSha, revertTarget: rec.revertTarget, cwd: rec.cwd, managedWorktree: rec.managedWorktree, rolePrompt: rec.rolePrompt, reviewSliceRoots: rec.reviewSliceRoots, openedAt: rec.openedAt, lastUsage: rec.lastUsage, carryRecap: wasInterrupted ? recoveryRecap : rec.carryRecap,
               // Lazy-boot restored terminals that were IDLE + not part of a flow: their transcript
               // shows immediately; the query boots on first turn. Mid-turn + flow terminals boot now
               // (mid-turn needs auto-resume; flow needs its lane live).

package/codex-mcp-http.mjs

@@ -6,18 +6,47 @@ import http from 'node:http'
 import { randomUUID } from 'node:crypto'
 import { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/streamableHttp.js'
 
+export function withMcpSessionFactory(createSessionServer) {
+  if (typeof createSessionServer !== 'function') throw new TypeError('MCP session factory must be a function')
+  return Object.assign(createSessionServer(), { createSessionServer })
+}
+
 export async function startCodexMcpHttp({ sdkServer, host = '127.0.0.1' } = {}) {
-  const instance = sdkServer?.instance
-  if (!instance?.connect) throw new TypeError('Codex MCP requires an SDK MCP server instance')
+  const createSessionServer = sdkServer?.createSessionServer
+  const initialInstance = sdkServer?.instance
+  if (!initialInstance?.connect || typeof createSessionServer !== 'function') {
+    throw new TypeError('Codex MCP requires an SDK MCP server and createSessionServer factory')
+  }
 
   const secretPath = `/mcp/${randomUUID()}`
-  // Stateful sessions (not stateless): a stateless transport (sessionIdGenerator
-  // undefined) 500s on the `notifications/initialized` POST that every MCP client
-  // sends right after initialize — it has no session to attach the notification to.
-  // Each connecting Codex process gets its own session id; the bridge owns one
-  // server per lane, so the in-memory session map is bounded by the live lanes.
-  const transport = new StreamableHTTPServerTransport({ sessionIdGenerator: () => randomUUID() })
-  await instance.connect(transport)
+  // Stateful sessions are required: this SDK's stateless transport rejects the
+  // `notifications/initialized` POST sent by every MCP client. A transport is
+  // single-initialize, though, while Codex App Server -> exec fallback and Hermes
+  // crash recovery both start a new client against the same lane endpoint.
+  // Keep exactly one active child session per lane. A new initialize atomically
+  // retires the old server+transport before receiving its own isolated instance.
+  let active = null
+
+  const retireActive = async () => {
+    const current = active
+    active = null
+    if (!current) return
+    await Promise.allSettled([
+      current.instance.close(),
+      current.transport.close(),
+    ])
+  }
+
+  const beginSession = async () => {
+    await retireActive()
+    const next = createSessionServer()
+    const instance = next?.instance
+    if (!instance?.connect) throw new TypeError('Codex MCP session factory did not return an SDK MCP server instance')
+    const transport = new StreamableHTTPServerTransport({ sessionIdGenerator: () => randomUUID() })
+    await instance.connect(transport)
+    active = { instance, transport }
+    return active
+  }
 
   const server = http.createServer(async (req, res) => {
     if (req.url !== secretPath) {
@@ -25,7 +54,19 @@ export async function startCodexMcpHttp({ sdkServer, host = '127.0.0.1' } = {})
       return
     }
     try {
-      await transport.handleRequest(req, res)
+      const sessionId = req.headers['mcp-session-id']
+      let target = active
+      if (!sessionId) {
+        if (req.method !== 'POST') {
+          res.writeHead(400).end('missing session')
+          return
+        }
+        target = await beginSession()
+      } else if (!target?.transport?.sessionId || sessionId !== target.transport.sessionId) {
+        res.writeHead(404).end('session not found')
+        return
+      }
+      await target.transport.handleRequest(req, res)
     } catch {
       if (!res.headersSent) res.writeHead(500)
       res.end()
@@ -38,7 +79,7 @@ export async function startCodexMcpHttp({ sdkServer, host = '127.0.0.1' } = {})
       server.listen(0, host, resolve)
     })
   } catch (error) {
-    await transport.close().catch(() => {})
+    await retireActive()
     throw error
   }
   const address = server.address()
@@ -51,7 +92,7 @@ export async function startCodexMcpHttp({ sdkServer, host = '127.0.0.1' } = {})
       if (closed) return
       closed = true
       await Promise.allSettled([
-        transport.close(),
+        retireActive(),
         new Promise((resolve) => server.close(resolve)),
       ])
     },

package/hermes-isolation.mjs

@@ -4,6 +4,7 @@ import path from 'node:path'
 import { randomUUID } from 'node:crypto'
 import { createSdkMcpServer, tool } from '@anthropic-ai/claude-agent-sdk'
 import { startHermesSession } from './hermes-session.mjs'
+import { withMcpSessionFactory } from './codex-mcp-http.mjs'
 
 const waitFor = (predicate, timeoutMs = 25_000) => new Promise((resolve, reject) => {
   const started = Date.now()
@@ -36,10 +37,11 @@ export async function verifyHermesMcpIsolation({ rounds = 2, sessionFactory = st
       for (let i = 0; i < 2; i++) {
         const cwd = path.join(root, i === 0 ? 'alpha' : 'beta')
         fs.mkdirSync(cwd)
-        const sdkServer = createSdkMcpServer({
+        const createMarkerServer = () => createSdkMcpServer({
           name: 'thinkpool', version: '1.0.0',
           tools: [tool(markers[i], 'Synthetic ThinkPool lane-isolation marker.', {}, async () => ({ content: [{ type: 'text', text: markers[i] }] }))],
         })
+        const sdkServer = withMcpSessionFactory(createMarkerServer)
         sessions.push(sessionFactory({ cwd, lazy: true, mcpServers: { thinkpool: sdkServer }, onEvent: (event) => events[i].push(event) }))
       }
       sessions.forEach((session) => session.sendTurn('/tools'))

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thinkpool-pair",
-  "version": "0.7.237",
+  "version": "0.7.239",
   "description": "Share a local coding-agent CLI (Claude Code, Codex, Gemini, Aider, …) into a ThinkPool Code room, live.",
   "type": "module",
   "bin": {

package/runtime-registry.mjs

@@ -13,7 +13,7 @@ const RUNTIMES = Object.freeze({
   }),
   hermes: Object.freeze({
     id: 'hermes', command: 'thinkpool', label: 'Hermes Agent', protocol: 'acp',
-    structured: true, flow: false, canSteer: true, images: true, nativeModelCatalog: true, effortControl: false, defaultMode: 'default',
+    structured: true, flow: false, canSteer: true, images: true, nativeModelCatalog: true, catalogRequiresSession: true, effortControl: false, defaultMode: 'default',
     modes: Object.freeze(['default', 'acceptEdits']),
     beta: true,
   }),
@@ -28,3 +28,12 @@ export const structuredRuntimeForCommand = (command) => {
 export const defaultStructuredMode = (runtime) => structuredRuntimeMetadata(runtime)?.defaultMode || 'default'
 export const structuredRuntimeSupportsMode = (runtime, mode) => structuredRuntimeMetadata(runtime)?.modes?.includes(mode) === true
 export const structuredRuntimeSupportsFlow = (runtime) => structuredRuntimeMetadata(runtime)?.flow === true
+export const shouldDeferStructuredRuntime = ({ runtime, defer, cwd, flowSessionId, flowTaskKey, models } = {}) => {
+  const metadata = structuredRuntimeMetadata(runtime)
+  // Some ACP agents publish their model catalog only from session/new or
+  // session/resume. A brand-new lane must perform that zero-inference handshake
+  // before its first prompt or the model picker has no truthful choices to show.
+  // Once the exact catalog is persisted, the normal lazy-start policy resumes.
+  if (metadata?.catalogRequiresSession && !(Array.isArray(models) && models.length)) return false
+  return !!defer || (!cwd && !flowSessionId && !flowTaskKey)
+}