thinkpool-pair 0.7.363 → 0.7.364

package/bridge.mjs

@@ -158,6 +158,8 @@ import { realtimeRecoveryDecision, turnInFlight } from './update-gate.mjs'
 import { saveSession, flushSession, deleteSession, loadAll, loadArchived, canResume, loadPtyId, savePtyId, loadNames, saveNames, appendDurableEvents, seedDurableEvents, readDurablePage, readDurableOldestSeq, hasDurableArchive, servesHistoryPage, acknowledgePendingScheduledOutcome, commitRecordedScheduledOutcome, deletePendingScheduledOutcome, loadPendingScheduledOutcome, loadPendingScheduledOutcomes, savePendingScheduledOutcome } from './session-store.mjs'
 import { stampEvent, makeSeqCounter, maxSeq, seqable, capReplayEvents, chunkReplayEvents, boundEventForBroadcast, inlineImageBlocks, ImageEventQueue, imageQueueConfig, uploadCodeImage as uploadCodeImageRequest, usageReportLine, codexUsageReportLine, appendCurrentPersonRequest, buildCheckpointFromLog, buildRecapFromLog, resolveCheckpointCarry, RECAP_CAP, trimmedBeforeSeq, firstSeq } from './event-id.mjs'
 import { createLatestReplayPump, requestedReplayIds } from './replay-transport.mjs'
+import { createEventDeliveryQueue } from './event-delivery-queue.mjs'
+import { formatMcpFlight, normalizeMcpFlight, recordMcpFlightEvent } from './mcp-flight-recorder.mjs'
 import { classifyCodeEvent, unknownCodeEventNotice } from './code-event-contract.mjs'
 import { fetchPairControlDeliveryAuthority } from './pair-control-authority.mjs'
 import { finishScheduledRunWithRetry, pollDueScheduledRuns, sanitizeScheduledOutcomeIntent, scheduledOutcomeReconciliationComplete, scheduledRunDeadlineRequired, scheduledRunSpawnAdmission, scheduledRunsCapability, scheduledRunsEnabled } from './scheduled-runs.mjs'
@@ -1343,7 +1345,21 @@ const bcastAwait = async (event, payload, ch = channel) => {
     }
   } catch { return null /* offline — replay covers it */ }
 }
-const bcast = (event, payload, ch = channel) => { void bcastAwait(event, payload, ch) }
+// Transcript frames are causally ordered. In particular, a small `result` must
+// never overtake a larger assistant/tool frame when the bridge falls back to
+// HTTP delivery or a socket is recovering. Retry the event at the head of this
+// stream until Realtime acknowledges it; the local JSONL archive is the crash
+// recovery backstop.
+const codeEventDelivery = createEventDeliveryQueue({
+  send: ({ event, payload, ch }) => bcastAwait(event, payload, ch),
+})
+const bcast = (event, payload, ch = channel) => {
+  if (event === 'code-event' && ch === channel) {
+    codeEventDelivery.enqueue({ event, payload, ch })
+    return
+  }
+  void bcastAwait(event, payload, ch)
+}
 const replayPump = createLatestReplayPump({
   send: ({ event, payload }) => bcastAwait(event, payload),
 })
@@ -2401,6 +2417,9 @@ function pushLog(entry, evt) {
   // stampEvent is idempotent, so the callers that already stamp are unaffected.
   stampStructuredTurn(entry, evt)
   stampEvent(evt)
+  // Observe the normalized runtime event, never the SDK/provider wire payload.
+  // The recorder retains only tool identity, timing, outcome, and argument count.
+  entry.mcpFlight = recordMcpFlightEvent(entry.mcpFlight, evt, { now: evt.ts })
   entry.lastActionAt = evt.ts
   entry.lastEvent = evt
   // Only TRANSCRIPT events consume a seq (seqable) — replay-chrome kinds the web
@@ -2561,7 +2580,7 @@ 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, models, effort, resume, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, pendingWorkerCompletions, workerCompletionsInFlight, rolePrompt, sliceType, flowSessionId, flowTaskKey, flowTaskContract, flowRole, flowReviewTarget, flowReviewTargets, flowReviewSnapshots, flowReviewRound, flowReviewConclusion, dispatchBaseSha, revertTarget, cwd, managedWorktree, reviewSliceRoots, openedAt, defer, provider, carryRecap, carryCheckpoint, lastUsage, receivedTurnCids, scheduleRunId, scheduleDeadlineAt, scheduleOutcomeRecorded, scheduleAdmissionLease }) {
+function openStructured({ id, runtime = 'claude', model, models, effort, resume, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, pendingWorkerCompletions, workerCompletionsInFlight, rolePrompt, sliceType, flowSessionId, flowTaskKey, flowTaskContract, flowRole, flowReviewTarget, flowReviewTargets, flowReviewSnapshots, flowReviewRound, flowReviewConclusion, dispatchBaseSha, revertTarget, cwd, managedWorktree, reviewSliceRoots, openedAt, defer, provider, carryRecap, carryCheckpoint, lastUsage, receivedTurnCids, mcpFlight, scheduleRunId, scheduleDeadlineAt, scheduleOutcomeRecorded, scheduleAdmissionLease }) {
   if (sessions.has(id)) return
   runtime = structuredRuntimeMetadata(runtime) ? runtime : 'claude'
   // Fail closed before exposing a native lane if its bridge semantic contract
@@ -2614,7 +2633,7 @@ function openStructured({ id, runtime = 'claude', model, models, effort, resume,
   const restoredScheduleOutcome = restoredScheduleOutcomeCandidate?.runId === scheduleRunId
     ? restoredScheduleOutcomeCandidate
     : null
-  const entry = { cmd: runtime, runtime, kind: 'structured', log: restoredLog, pending: new Map(), receivedTurnCids: new Set(Array.isArray(receivedTurnCids) ? receivedTurnCids.filter(Boolean).slice(-1000) : []), session: null, recovered: false, commands: commandCatalogForRuntime(runtime, commands), mode, effort, models: runtime === 'codex' ? codexModels : runtime === 'hermes' ? (Array.isArray(models) && models.length ? models : hostHermesModels) : undefined,
+  const entry = { cmd: runtime, runtime, kind: 'structured', log: restoredLog, pending: new Map(), receivedTurnCids: new Set(Array.isArray(receivedTurnCids) ? receivedTurnCids.filter(Boolean).slice(-1000) : []), mcpFlight: normalizeMcpFlight(mcpFlight), session: null, recovered: false, commands: commandCatalogForRuntime(runtime, commands), mode, effort, models: runtime === 'codex' ? codexModels : runtime === 'hermes' ? (Array.isArray(models) && models.length ? models : hostHermesModels) : 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
@@ -2990,7 +3009,7 @@ function openStructured({ id, runtime = 'claude', model, models, effort, resume,
   // 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, 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, scheduleRunId: entry.scheduleRunId || null, scheduleDeadlineAt: entry.scheduleDeadlineAt || null, scheduleOutcomeRecorded: entry.scheduleOutcomeRecorded === true, scheduleAdmissionLease: entry.scheduleAdmissionLease || null, pendingSideContexts: entry.pendingSideContexts, pendingWorkerCompletions: entry.pendingWorkerCompletions, workerCompletionsInFlight: entry.workerCompletionsInFlight, sliceType: entry.sliceType, flowSessionId: entry.flowSessionId, flowTaskKey: entry.flowTaskKey, flowTaskContract: entry.flowTaskContract, flowRole: entry.flowRole || null, flowReviewTargets: entry.flowReviewTargets || [], flowReviewSnapshots: entry.flowReviewSnapshots || [], flowReviewRound: entry.flowReviewRound || 0, flowReviewConclusion: entry.flowReviewConclusion, 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, carryCheckpoint: entry.pendingCheckpoint || null, receivedTurnCids: [...entry.receivedTurnCids].slice(-1000) })
+  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, scheduleRunId: entry.scheduleRunId || null, scheduleDeadlineAt: entry.scheduleDeadlineAt || null, scheduleOutcomeRecorded: entry.scheduleOutcomeRecorded === true, scheduleAdmissionLease: entry.scheduleAdmissionLease || null, pendingSideContexts: entry.pendingSideContexts, pendingWorkerCompletions: entry.pendingWorkerCompletions, workerCompletionsInFlight: entry.workerCompletionsInFlight, sliceType: entry.sliceType, flowSessionId: entry.flowSessionId, flowTaskKey: entry.flowTaskKey, flowTaskContract: entry.flowTaskContract, flowRole: entry.flowRole || null, flowReviewTargets: entry.flowReviewTargets || [], flowReviewSnapshots: entry.flowReviewSnapshots || [], flowReviewRound: entry.flowReviewRound || 0, flowReviewConclusion: entry.flowReviewConclusion, 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, carryCheckpoint: entry.pendingCheckpoint || null, receivedTurnCids: [...entry.receivedTurnCids].slice(-1000), mcpFlight: entry.mcpFlight })
   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
@@ -3129,6 +3148,12 @@ function openStructured({ id, runtime = 'claude', model, models, effort, resume,
     name: 'thinkpool',
     version: '1.0.0',
     tools: [
+      ...(!entry.flowRole && entry.sliceType !== 'review' ? [tool(
+        'read_mcp_flight_recorder',
+        'Read the privacy-filtered MCP flight recorder for THIS terminal. Use it to diagnose a tool path or audit recent MCP activity. It reports only the MCP tool name, start time, duration, and outcome; it never records or returns arguments, result bodies, prompts, secrets, or host paths.',
+        { limit: z.number().int().min(1).max(100).optional().describe('maximum recent calls to return (default 40)') },
+        async (args) => ({ content: [{ type: 'text', text: formatMcpFlight(entry.mcpFlight, { limit: args?.limit }) }] }),
+      )] : []),
       ...(runtime === 'hermes' && hermesRoleFor({ flowRole: entry.flowRole, sliceType: entry.sliceType }) === 'ordinary' ? [tool(
         'request_user_input',
         'Ask the people in this ThinkPool room one to three multiple-choice questions and wait for their answer. Use this for choices that genuinely block useful progress. Each question needs two or three mutually exclusive options; a free-text answer remains available in the room card.',
@@ -3900,7 +3925,7 @@ function openStructured({ id, runtime = 'claude', model, models, effort, resume,
               dispatchBaseSha: entry.dispatchBaseSha, revertTarget: entry.revertTarget, cwd: entry.cwd,
               managedWorktree: entry.managedWorktree, rolePrompt: entry.rolePrompt,
               reviewSliceRoots: entry.reviewSliceRoots, openedAt: entry.openedAt,
-              lastUsage: entry.lastUsage, carryRecap, carryCheckpoint: entry.pendingCheckpoint,
+              lastUsage: entry.lastUsage, mcpFlight: entry.mcpFlight, carryRecap, carryCheckpoint: entry.pendingCheckpoint,
             })
             return sessions.get(id) || null
           },
@@ -4423,7 +4448,7 @@ function respawnStructured(id, provider) {
   // openStructured seed from the TARGET provider's configured model, which is the
   // only model this lane was ever asked for. A same-env model change never reaches
   // here — that path is an in-place setModel (see provider-switch).
-  const { runtime, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, pendingWorkerCompletions, workerCompletionsInFlight, sliceType, flowSessionId, flowTaskKey, flowTaskContract, flowRole, flowReviewTarget, flowReviewTargets, flowReviewSnapshots, flowReviewRound, flowReviewConclusion, dispatchBaseSha, revertTarget, cwd, managedWorktree, rolePrompt, reviewSliceRoots, openedAt, lastUsage, pendingCheckpoint } = s
+  const { runtime, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, pendingWorkerCompletions, workerCompletionsInFlight, sliceType, flowSessionId, flowTaskKey, flowTaskContract, flowRole, flowReviewTarget, flowReviewTargets, flowReviewSnapshots, flowReviewRound, flowReviewConclusion, dispatchBaseSha, revertTarget, cwd, managedWorktree, rolePrompt, reviewSliceRoots, openedAt, lastUsage, mcpFlight, pendingCheckpoint } = s
   // Context-carry (2026-07-08): a provider switch is not an SDK resume — the new backend
   // starts blank mid-conversation. Synthesize a plain-text recap from the VISIBLE log NOW
   // (before teardown) and hand it to the fresh session as its first turn so the agent
@@ -4442,7 +4467,7 @@ function respawnStructured(id, provider) {
   // sessionData() (provider included) synchronously on open, so a bridge restart
   // restores the lane on its CURRENT provider, not the original — and its next
   // announce carries the new provider badge (additive {id,name} projection).
-  openStructured({ id, runtime, provider, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, pendingWorkerCompletions, workerCompletionsInFlight, sliceType, flowSessionId, flowTaskKey, flowTaskContract, flowRole, flowReviewTarget, flowReviewTargets, flowReviewSnapshots, flowReviewRound, flowReviewConclusion, dispatchBaseSha, revertTarget, cwd, managedWorktree, rolePrompt, reviewSliceRoots, openedAt, lastUsage, carryRecap, carryCheckpoint: pendingCheckpoint })
+  openStructured({ id, runtime, provider, log, commands, mode, spawnedBy, spawnDepth, cascadeRole, hop, sideParent, sideTask, pendingSideContexts, pendingWorkerCompletions, workerCompletionsInFlight, sliceType, flowSessionId, flowTaskKey, flowTaskContract, flowRole, flowReviewTarget, flowReviewTargets, flowReviewSnapshots, flowReviewRound, flowReviewConclusion, dispatchBaseSha, revertTarget, cwd, managedWorktree, rolePrompt, reviewSliceRoots, openedAt, lastUsage, mcpFlight, carryRecap, carryCheckpoint: pendingCheckpoint })
   return true
 }
 
@@ -5416,7 +5441,7 @@ 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, 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, scheduleRunId: rec.scheduleRunId, scheduleDeadlineAt: rec.scheduleDeadlineAt, scheduleOutcomeRecorded: rec.scheduleOutcomeRecorded, scheduleAdmissionLease: rec.scheduleAdmissionLease, pendingSideContexts: rec.pendingSideContexts, pendingWorkerCompletions: rec.pendingWorkerCompletions, workerCompletionsInFlight: rec.workerCompletionsInFlight, sliceType: rec.sliceType, flowSessionId: rec.flowSessionId, flowTaskKey: rec.flowTaskKey, flowTaskContract: rec.flowTaskContract, flowRole: rec.flowRole, flowReviewTarget: rec.flowReviewTarget, flowReviewTargets: rec.flowReviewTargets, flowReviewSnapshots: rec.flowReviewSnapshots, flowReviewRound: rec.flowReviewRound, flowReviewConclusion: rec.flowReviewConclusion, dispatchBaseSha: rec.dispatchBaseSha, revertTarget: rec.revertTarget, cwd: rec.cwd, managedWorktree: rec.managedWorktree, rolePrompt: rec.rolePrompt, reviewSliceRoots: rec.reviewSliceRoots, openedAt: rec.openedAt, lastUsage: rec.lastUsage, receivedTurnCids: rec.receivedTurnCids, carryRecap: wasInterrupted ? recoveryRecap : rec.carryRecap, carryCheckpoint: rec.carryCheckpoint,
+            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, scheduleRunId: rec.scheduleRunId, scheduleDeadlineAt: rec.scheduleDeadlineAt, scheduleOutcomeRecorded: rec.scheduleOutcomeRecorded, scheduleAdmissionLease: rec.scheduleAdmissionLease, pendingSideContexts: rec.pendingSideContexts, pendingWorkerCompletions: rec.pendingWorkerCompletions, workerCompletionsInFlight: rec.workerCompletionsInFlight, sliceType: rec.sliceType, flowSessionId: rec.flowSessionId, flowTaskKey: rec.flowTaskKey, flowTaskContract: rec.flowTaskContract, flowRole: rec.flowRole, flowReviewTarget: rec.flowReviewTarget, flowReviewTargets: rec.flowReviewTargets, flowReviewSnapshots: rec.flowReviewSnapshots, flowReviewRound: rec.flowReviewRound, flowReviewConclusion: rec.flowReviewConclusion, dispatchBaseSha: rec.dispatchBaseSha, revertTarget: rec.revertTarget, cwd: rec.cwd, managedWorktree: rec.managedWorktree, rolePrompt: rec.rolePrompt, reviewSliceRoots: rec.reviewSliceRoots, openedAt: rec.openedAt, lastUsage: rec.lastUsage, receivedTurnCids: rec.receivedTurnCids, mcpFlight: rec.mcpFlight, carryRecap: wasInterrupted ? recoveryRecap : rec.carryRecap, carryCheckpoint: rec.carryCheckpoint,
               // 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).
@@ -6175,6 +6200,7 @@ if (process.env.THINKPOOL_PAIR_AUTOUPDATE === '1' && VERSION) {
 async function shutdown(code = 0, farewell = true) {
   if (shuttingDown) return
   shuttingDown = true
+  codeEventDelivery.close()
   // Hard exit backstop, armed FIRST and independent of every await below — a wedged
   // realtime socket (the watchdog path) or a hung untrack must NEVER leave the
   // process alive with orphaned PTY children + stale "live" presence.

package/event-delivery-queue.mjs

@@ -0,0 +1,60 @@
+// Ordered, acknowledged delivery for bridge-produced transcript events.
+//
+// The bridge's local JSONL archive is durable, but a live `code-event` still
+// has to cross Realtime before either viewer can write its shared tx: snapshot.
+// Never let a failed (or merely late) assistant frame allow its terminal result
+// to overtake it. Keep the head queued until Realtime acknowledges it, then move
+// on in order. A fresh browser replay remains the crash-recovery backstop.
+
+const defaultWait = (ms) => new Promise((resolve) => setTimeout(resolve, ms))
+
+export const acceptedRealtimeDelivery = (result) =>
+  result === 'ok' || result?.success === true
+
+export function createEventDeliveryQueue ({ send, wait = defaultWait, retryDelay = (attempt) => Math.min(10_000, 250 * (2 ** Math.min(attempt, 6))) } = {}) {
+  if (typeof send !== 'function') throw new TypeError('event delivery queue requires send(frame)')
+  const frames = []
+  let running = false
+  let closed = false
+  let idle = Promise.resolve()
+  let settleIdle = null
+
+  const run = async () => {
+    if (running || closed) return
+    running = true
+    try {
+      while (!closed && frames.length) {
+        const head = frames[0]
+        let result = null
+        try { result = await send(head.frame) } catch { result = null }
+        if (acceptedRealtimeDelivery(result)) {
+          frames.shift()
+          continue
+        }
+        head.attempt += 1
+        await wait(retryDelay(head.attempt))
+      }
+    } finally {
+      running = false
+      if (!frames.length && settleIdle) { settleIdle(); settleIdle = null }
+      if (!closed && frames.length) void run()
+    }
+  }
+
+  return {
+    enqueue (frame) {
+      if (closed) return false
+      if (!frames.length && !running) idle = new Promise((resolve) => { settleIdle = resolve })
+      frames.push({ frame, attempt: 0 })
+      void run()
+      return true
+    },
+    pending: () => frames.length,
+    whenIdle: () => idle,
+    close () {
+      closed = true
+      frames.length = 0
+      if (settleIdle) { settleIdle(); settleIdle = null }
+    },
+  }
+}

package/hermes-policy.mjs

@@ -4,6 +4,7 @@ import { PEER_MCP_TOOLS, PEER_READ_MCP_TOOLS } from './cross-terminal.mjs'
 
 export const HERMES_POLICY_VERSION = 1
 export const HERMES_QUESTION_MCP_TOOL = 'request_user_input'
+export const MCP_FLIGHT_RECORDER_TOOL = 'read_mcp_flight_recorder'
 
 export const CODING_TOOLS = Object.freeze([
   'web_search', 'web_extract', 'terminal', 'process', 'read_file', 'write_file',
@@ -23,7 +24,7 @@ export const ESSENTIAL_CODING_TOOLS = Object.freeze([
   'terminal', 'process', 'read_file', 'write_file', 'patch', 'search_files',
 ])
 const ROLE_REQUIRED = Object.freeze({
-  ordinary: [...PEER_MCP_TOOLS, HERMES_QUESTION_MCP_TOOL, 'spawn_terminal', 'close_terminal'],
+  ordinary: [...PEER_MCP_TOOLS, HERMES_QUESTION_MCP_TOOL, MCP_FLIGHT_RECORDER_TOOL, 'spawn_terminal', 'close_terminal'],
   plan: [],
   conductor: ['submit_flow_plan'],
   builder: ['mark_flow_done'],
@@ -46,7 +47,7 @@ export function hermesPolicyForRole(role, { mcpTools } = {}) {
   // Ordinary worker leaves deliberately lack spawn/close. Main-lane proof is
   // enforced by requiredMcpTools at dispatch; keep this schema usable for a
   // non-delegating ordinary child without widening it.
-  const minimum = role === 'ordinary' ? [...PEER_MCP_TOOLS, HERMES_QUESTION_MCP_TOOL] : required
+  const minimum = role === 'ordinary' ? [...PEER_MCP_TOOLS, HERMES_QUESTION_MCP_TOOL, MCP_FLIGHT_RECORDER_TOOL] : required
   for (const tool of minimum) if (!supplied.includes(tool)) throw new Error(`Hermes ${role} policy is missing required ThinkPool tool ${tool}`)
   const restricted = role === 'plan' || role === 'conductor' || role === 'reviewer' || role === 'manual-review'
   const builtinTools = restricted ? READ_ONLY_TOOLS : CODING_TOOLS
@@ -64,7 +65,7 @@ export function hermesPolicyForRole(role, { mcpTools } = {}) {
 }
 
 export function hermesRequiredMcpTools(role, { canSpawnWorkers = false } = {}) {
-  if (role === 'ordinary') return [...PEER_MCP_TOOLS, HERMES_QUESTION_MCP_TOOL, ...(canSpawnWorkers ? ['spawn_terminal', 'close_terminal'] : [])]
+  if (role === 'ordinary') return [...PEER_MCP_TOOLS, HERMES_QUESTION_MCP_TOOL, MCP_FLIGHT_RECORDER_TOOL, ...(canSpawnWorkers ? ['spawn_terminal', 'close_terminal'] : [])]
   return [...ROLE_REQUIRED[role]]
 }
 

package/mcp-flight-recorder.mjs

@@ -0,0 +1,79 @@
+// Privacy-first MCP flight recorder. It retains enough operational evidence to
+// diagnose a tool path (tool identity, lifecycle, timing, and outcome), but it
+// intentionally never retains arguments, result bodies, prompts, or host paths.
+// The recorder is runtime-agnostic because every structured runtime normalizes
+// MCP activity into the common assistant/tool_result event contract.
+
+export const MCP_FLIGHT_MAX = 200
+
+const mcpToolUse = (event) => (Array.isArray(event?.blocks) ? event.blocks : [])
+  .filter((block) => block?.type === 'tool_use' && typeof block?.name === 'string' && block.name.startsWith('mcp__'))
+
+const safeToolName = (name) => String(name || '')
+  .replace(/[^a-zA-Z0-9_-]/g, '_')
+  .slice(0, 160)
+
+const safeId = (id) => typeof id === 'string' || typeof id === 'number'
+  ? String(id).slice(0, 200)
+  : null
+
+export function normalizeMcpFlight (value, { max = MCP_FLIGHT_MAX } = {}) {
+  const cap = Math.max(1, Math.min(MCP_FLIGHT_MAX, Number(max) || MCP_FLIGHT_MAX))
+  if (!Array.isArray(value)) return []
+  return value.slice(-cap).flatMap((item) => {
+    if (!item || typeof item !== 'object') return []
+    const toolUseId = safeId(item.toolUseId)
+    const tool = safeToolName(item.tool)
+    const startedAt = Number(item.startedAt)
+    if (!toolUseId || !tool || !Number.isFinite(startedAt)) return []
+    return [{
+      toolUseId,
+      tool,
+      startedAt,
+      finishedAt: Number.isFinite(Number(item.finishedAt)) ? Number(item.finishedAt) : null,
+      durationMs: Number.isFinite(Number(item.durationMs)) ? Math.max(0, Number(item.durationMs)) : null,
+      outcome: item.outcome === 'error' ? 'error' : item.outcome === 'ok' ? 'ok' : 'pending',
+      // Deliberately a count only. Never persist or return argument names/values:
+      // even a key can reveal a secret type or private filesystem layout.
+      argumentCount: Math.max(0, Math.min(1000, Number(item.argumentCount) || 0)),
+    }]
+  })
+}
+
+const argumentCount = (input) => input && typeof input === 'object' && !Array.isArray(input)
+  ? Object.keys(input).length
+  : input == null ? 0 : 1
+
+/** Record an already-normalized room event, returning a bounded safe trace. */
+export function recordMcpFlightEvent (flight, event, { now = Date.now(), max = MCP_FLIGHT_MAX } = {}) {
+  const next = normalizeMcpFlight(flight, { max })
+  for (const block of mcpToolUse(event)) {
+    const toolUseId = safeId(block.id)
+    const tool = safeToolName(block.name)
+    if (!toolUseId || !tool) continue
+    next.push({ toolUseId, tool, startedAt: now, finishedAt: null, durationMs: null, outcome: 'pending', argumentCount: argumentCount(block.input) })
+  }
+  if (event?.kind === 'tool_result') {
+    const toolUseId = safeId(event.toolUseId)
+    const item = toolUseId ? [...next].reverse().find((entry) => entry.toolUseId === toolUseId && entry.outcome === 'pending') : null
+    if (item) {
+      item.finishedAt = now
+      item.durationMs = Number.isFinite(Number(event.durationMs)) ? Math.max(0, Number(event.durationMs)) : Math.max(0, now - item.startedAt)
+      item.outcome = event.isError ? 'error' : 'ok'
+    }
+  }
+  return normalizeMcpFlight(next, { max })
+}
+
+export function formatMcpFlight (flight, { limit = 40 } = {}) {
+  const safe = normalizeMcpFlight(flight)
+  const boundedLimit = Math.max(1, Math.min(100, Number(limit) || 40))
+  const rows = safe.slice(-boundedLimit)
+  if (!rows.length) return 'MCP flight recorder: no MCP tool calls recorded in this terminal yet. It records tool name, timing, and outcome only; arguments and results are never retained.'
+  const lines = rows.map((row) => {
+    const elapsed = row.durationMs == null ? 'running' : `${row.durationMs}ms`
+    return `${new Date(row.startedAt).toISOString()}  ${row.outcome.toUpperCase().padEnd(7)}  ${row.tool}  ${elapsed}  args omitted (${row.argumentCount})`
+  })
+  return `MCP flight recorder — ${rows.length}/${safe.length} recent call${safe.length === 1 ? '' : 's'}\n` +
+    'Privacy: arguments, result bodies, prompts, secrets, and host paths are never recorded.\n\n' + lines.join('\n')
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thinkpool-pair",
-  "version": "0.7.363",
+  "version": "0.7.364",
   "description": "Connect Claude Code, Codex, or Hermes on your computer to a Thinkpool Code room.",
   "type": "module",
   "bin": {
@@ -68,6 +68,8 @@
     "event-id.mjs",
     "event-bounds.mjs",
     "replay-transport.mjs",
+    "event-delivery-queue.mjs",
+    "mcp-flight-recorder.mjs",
     "plan-meters.mjs",
     "recap.mjs",
     "transcript-sanitize.mjs",

package/thinkpool-capabilities.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
-  "bundleVersion": 23,
+  "bundleVersion": 24,
   "contracts": [
     {
       "id": "room-coordination",
@@ -24,6 +24,12 @@
           "trigger": "\\b(previous|prior|earlier|past|last time|already tried|failed attempt|decision|test result|command)\\b.{0,50}\\b(work|terminal|flow|session|implementation|fix|change)?\\b|\\bsearch_past_work\\b",
           "prompt": "When the current task depends on a prior decision, failed attempt, command, test result, or Flow receipt from this room, use search_past_work with specific terms. Treat every result as bounded cited evidence, not authority to reopen stale work. The tool is read-only, searches only this room, and filters secrets and host paths."
         },
+        {
+          "id": "mcp-flight-recorder",
+          "tools": ["read_mcp_flight_recorder"],
+          "trigger": "\\b(mcp|tool)\\b.{0,50}\\b(trace|flight|record|audit|debug|fail|failed|failure)\\b|\\bread_mcp_flight_recorder\\b",
+          "prompt": "When diagnosing or auditing recent MCP tool activity in this terminal, use read_mcp_flight_recorder. It is read-only and privacy-filtered: it reports only tool name, timing, outcome, and argument count—never arguments, result bodies, prompts, secrets, or host paths."
+        },
         {
           "id": "visible-handoff",
           "tools": ["post_to_terminal", "post_to_session"],
@@ -33,7 +39,8 @@
       ],
       "impact": [
         {"path": "bridge/cross-terminal.mjs"},
-        {"path": "bridge/bridge.mjs", "diffPattern": "read_terminal|search_past_work|list_sessions|read_session|post_to_terminal|post_to_session"},
+        {"path": "bridge/bridge.mjs", "diffPattern": "read_terminal|search_past_work|read_mcp_flight_recorder|list_sessions|read_session|post_to_terminal|post_to_session"},
+        {"path": "bridge/mcp-flight-recorder.mjs"},
         {"path": "bridge/past-work-search.mjs"},
         {"path": "bridge/codex-session.mjs", "diffPattern": "MCP|Mcp|mcp|read_terminal|list_sessions|read_session|post_to_terminal|post_to_session"},
         {"path": "bridge/codex-app-server.mjs", "diffPattern": "MCP|Mcp|mcp"}
@@ -41,6 +48,8 @@
       "evidence": [
         {"path": "bridge/cross-terminal.mjs", "pattern": "read_terminal"},
         {"path": "bridge/bridge.mjs", "pattern": "'search_past_work'"},
+        {"path": "bridge/bridge.mjs", "pattern": "'read_mcp_flight_recorder'"},
+        {"path": "bridge/mcp-flight-recorder.mjs", "pattern": "recordMcpFlightEvent"},
         {"path": "bridge/past-work-search.mjs", "pattern": "formatPastWorkSearch"},
         {"path": "bridge/cross-terminal.mjs", "pattern": "post_to_session"},
         {"path": "bridge/bridge.mjs", "pattern": "'list_sessions'"},