@oar-sh/oar 0.9.19 → 0.9.20

package/server/relay-self-restart.mjs

@@ -1,7 +1,8 @@
-import { spawn } from 'child_process';
+import { spawn, spawnSync } from 'child_process';
 import path from 'path';
 import { RELAY_RESTART_EXIT_CODE, RELAY_STOP_EXIT_CODE } from './relay-exit-codes.mjs';
 import { relayRuntimeNodeFlags, relayRuntimeNodeFlagsOrNone } from './relay-heap-guard.mjs';
+import { detectRelayHosting } from './services/relay-restart-picture.mjs';
 
 // Role selection travels on argv, never in the environment. The runtime's env is
 // inherited by tmux worker sessions and by the Copilot CLI it launches, and the
@@ -76,6 +77,64 @@ function waitForChildExit(child) {
   });
 }
 
+// A systemctl --user call of the handover: its result, never a throw.
+function runSystemctl(spawnSyncImpl, args) {
+  try {
+    return spawnSyncImpl('systemctl', ['--user', ...args], { encoding: 'utf8', stdio: ['ignore', 'pipe', 'pipe'], timeout: 15_000 }) || { status: null };
+  } catch (error) {
+    return { status: null, error };
+  }
+}
+
+function systemctlFailure(result) {
+  if (result.error) return result.error.message || String(result.error);
+  const said = String(result.stderr || '').trim().split('\n')[0];
+  return `status ${result.status ?? 'none'}${said ? `: ${said}` : ''}`;
+}
+
+/**
+ * The runtime asked for a restart (exit 75). Under a user service whose main
+ * process this supervisor is, systemd restarts the whole service, so the new
+ * supervisor and runtime run the code on disk with fresh Node flags: true
+ * once systemd's stop (SIGTERM) has come. False, with one log line, wherever
+ * that cannot be done (not the unit's main process: a relay started inside a
+ * worker of the service), systemctl fails, or no stop comes within
+ * `handoverWaitMs`; the caller relaunches the runtime itself then.
+ */
+async function handRestartToSystemd({ hosting, spawnSyncImpl, pid, handoverWaitMs, shutdownRequested, isShuttingDown, delay, logger }) {
+  if (!hosting?.systemd || !hosting.userManager || !hosting.unit) return false;
+  const { unit } = hosting;
+  const main = runSystemctl(spawnSyncImpl, ['show', '-p', 'MainPID', '--value', unit]);
+  const mainPid = Number(String(main.stdout ?? '').trim());
+  if (main.status !== 0 || mainPid !== pid) {
+    const whose = main.status === 0 ? `its main process is ${mainPid || 'none'}` : `its main process could not be read (${systemctlFailure(main)})`;
+    logger?.log?.(`[relay] restart: not handed to systemd, ${unit} is not this supervisor's (pid ${pid}; ${whose}).`);
+    return false;
+  }
+  // Clears a failed state and the restart backoff; restart's answer alone decides.
+  runSystemctl(spawnSyncImpl, ['reset-failed', unit]);
+  // A stop that came in meanwhile wins: a queued restart could replace its job.
+  // The calls above block; a signal that arrived during them waits on the
+  // signal pipe, which only the loop's poll phase reads. This runs from the
+  // runtime's 'exit' callback, itself in a poll phase, so one setImmediate
+  // would run in the check phase right after it, before the pipe is read;
+  // the second runs after the next poll phase has delivered the signal.
+  await new Promise((resolve) => setImmediate(() => setImmediate(resolve)));
+  if (isShuttingDown()) {
+    logger?.log?.(`[relay] restart: a stop came in; not restarting ${unit}.`);
+    return true;
+  }
+  const restart = runSystemctl(spawnSyncImpl, ['restart', '--no-block', unit]);
+  if (restart.status !== 0) {
+    logger?.log?.(`[relay] restart: systemctl --user restart --no-block ${unit} failed (${systemctlFailure(restart)}).`);
+    return false;
+  }
+  logger?.log?.(`[relay] restart handed to systemd: ${unit} restarts with the code on disk.`);
+  const stopped = await Promise.race([shutdownRequested.then(() => true), Promise.resolve(delay(handoverWaitMs)).then(isShuttingDown)]);
+  if (!stopped) logger?.log?.(`[relay] restart: systemd did not stop ${unit} within ${Math.round(handoverWaitMs / 1000)} s.`);
+  return stopped;
+}
+
 export async function runDirectRelaySupervisor({
   scriptPath,
   args = [],
@@ -96,16 +155,26 @@ export async function runDirectRelaySupervisor({
   // crashes days apart are not a crash loop.
   crashCountResetMs = 10 * 60 * 1000,
   now = Date.now,
+  // How this process is run (detectRelayHosting); read at the first restart
+  // request when not given.
+  hosting = null,
+  spawnSyncImpl = spawnSync,
+  pid = process.pid,
+  // How long a restart handed to systemd waits for its SIGTERM.
+  handoverWaitMs = 30_000,
 } = {}) {
   let runtimeProc = null;
   let shuttingDown = false;
   let shutdownExitCode = 0;
+  let shutdownArrived;
+  const shutdownRequested = new Promise((resolve) => { shutdownArrived = resolve; });
 
   const requestShutdown = (signal = 'SIGTERM', exitCode = 0) => {
     if (shuttingDown) return;
     shuttingDown = true;
     shutdownExitCode = exitCode;
     stopChildProcess(runtimeProc, { signal });
+    shutdownArrived();
   };
 
   const handleSigInt = () => requestShutdown('SIGINT', 0);
@@ -165,8 +234,20 @@ export async function runDirectRelaySupervisor({
         : null;
       const signal = exitCode === null ? (result.signal || null) : null;
       if (exitCode === RELAY_RESTART_EXIT_CODE) {
-        logger?.log?.(`[relay] runtime requested restart; relaunching ${path.basename(String(scriptPath || ''))}...`);
         crashCount = 0;
+        const handedOver = await handRestartToSystemd({
+          hosting: hosting || detectRelayHosting({ env }),
+          spawnSyncImpl,
+          pid,
+          handoverWaitMs,
+          shutdownRequested,
+          isShuttingDown: () => shuttingDown,
+          delay,
+          logger,
+        });
+        // systemd's stop: this supervisor ends, systemd starts the next one.
+        if (handedOver || shuttingDown) return await exitImpl(shutdownExitCode);
+        logger?.log?.(`[relay] runtime requested restart; relaunching ${path.basename(String(scriptPath || ''))}...`);
         await delay(restartDelayMs);
         continue;
       }

package/server/relay-tools.md

@@ -39,10 +39,11 @@ A paired relay stays locked until the user mentions it in this conversation (`@n
 
 This relay itself is listed with "self": true when its owner switched agent sessions on, and is addressed by its name or as "this"; it needs no mention. Use it to hand parts of your task to other sessions:
 
-- `create_session` starts a session with any provider that `relay_info` lists: a local one in a folder ("cwd"), or provider "claude-cloud" with "repo" (a GitHub URL or owner/repo) and an optional "branch". The provider ids are the target relay's, as `relay_info` shows them; an id it does not know is refused with the list it has. A session started before a relay update can use providers the update added, without a restart. At most 4 sessions you started may work at the same time; a fifth is refused until one finishes.
+- `create_session` starts a session with any provider that `relay_info` lists: a local one in a folder ("cwd"), or provider "claude-cloud" with "repo" (a GitHub URL or owner/repo) and an optional "branch". The provider ids are the target relay's, as `relay_info` shows them; an id it does not know is refused with the list it has. A session started before a relay update can use providers the update added, without a restart. Without "mode" the new session gets your session's relay mode. At most 4 sessions you started may work at the same time; a fifth is refused until one finishes.
 - The first `create_session` on this relay shows the user an approval card, once per conversation. A session that an agent created cannot create sessions itself (one level only), unless the owner allowed it for that session: `list_sessions` and `read_session` then show "mayCreateSessions": true. Only the owner sets that.
 - Sessions you start here are grouped under this session in the user's sidebar. `list_sessions` lists in the sidebar's order (a group sorts by its newest session, head first) and shows the group as a "group" field (id, name, role); `read_session` adds the group's live members with their state. `create_session` takes "group" (a group id, or its name in any case, from `relay_info`'s "groups") to put the new session into an existing group instead, here or on a paired relay; an unknown group is refused with the groups listed and nothing is created, and the result carries the "group". Renaming and dissolving groups is the user's.
-- `update_session` ("session" plus any of "model", "effort", "mode", "auto_compact_window", "title", "group", "thinking_enabled", "thinking_display", "cwd") changes another session's settings for its next turns without prompting it, here or on a paired relay, and answers the settings and "changed"; a group name that does not exist makes one, "none" takes the session out (a group the user made stays, even empty; one update_session made goes once empty), a new "cwd" restarts its CLI there (refused while it works), "auto_compact_window" 0 is the model default. The model is checked as for `send`. A failed call still lists in "changed" what it wrote before the failing step. A mode set this way lasts until the next `send` that names another mode or comes from a caller in another mode: a `send` without "mode" passes the caller's own mode.
+- `update_session` ("session" plus any of "model", "effort", "mode", "auto_compact_window", "title", "group", "thinking_enabled", "thinking_display", "cwd", "planOnly") changes another session's settings for its next turns without prompting it, here or on a paired relay, and answers the settings and "changed"; a group name that does not exist makes one, "none" takes the session out (a group the user made stays, even empty; one update_session made goes once empty), a new "cwd" restarts its CLI there (refused while it works), "auto_compact_window" 0 is the model default. The model is checked as for `send`. A failed call still lists in "changed" what it wrote before the failing step. A mode set this way lasts until the next `update_session` that names another mode; a `send` that names a mode runs only that turn in it, and one without "mode" leaves the session in its own mode.
+- "planOnly": true makes a plan-only session: agents can't take it out of plan mode. Set it only when the user asks; plan mode alone never sets it. `create_session` with "planOnly": true and no "mode" starts in plan; with another mode named it is refused and nothing is created. A target relay too old for planOnly creates the session without the lock and says so in "planOnlyWarning". `update_session` with "planOnly": true sets it on an existing session (with "mode": "plan" unless it is in plan already). While it is set, a `send` or `update_session` that names another mode is refused with 409, code "plan_only" ("This session is plan-only; agents can't change its mode. Clear planOnly first, or ask the user."); a `send` without "mode" runs in the session's saved mode. To clear it, pass "planOnly": false in the same `update_session` that names the new mode (`{"planOnly":false,"mode":"agent"}`); alone, or with "mode": "plan", it is ignored, and on a plan-only session the answer notes "planOnly unchanged: clear it in the same call as the new mode." Use `update_session` for all this, not the preferences route (it refuses an agent's mode change with 409 and any planOnly change with 403). `list_sessions` and `read_session` show "planOnly": true. In a plan-only conversation you count as a plan-mode caller, also in the one agent turn after "Approve plan": your `create_session` defaults to plan and your write actions ask the user first.
 - Only a session your conversation started may be updated freely; any other (the user's, another agent's, the one that started you) needs the user's approval card once per conversation and relay (refused: `REMOTE_RELAY_PERMISSION`); never your own conversation.
 - A "model" for `create_session` or `send` must be an id `relay_info` lists for the provider (Claude also takes "opus", "sonnet", "haiku"). An unknown id is refused with the close matches and the valid ids (`Unknown model 'gpt-5-4' for Copilot on relay "…"; did you mean 'gpt-5.4'? Valid ids: … Nothing was sent.`), and nothing is created or sent: pick from the list instead of guessing. Copilot ids use dots (`claude-opus-5.5`, not the Claude form `claude-opus-5-5`); a hyphen form is refused with the close match named, and a Copilot session never moves to another provider because of its model name. If the answer has a "modelNote" saying the model could not be checked, the relay had no model list yet.
 - A new session knows nothing of this conversation: write a self-contained prompt. A Claude Cloud session works on a fresh clone in a sandbox at Anthropic, so name the branch it should work on and push.
@@ -51,7 +52,7 @@ This relay itself is listed with "self": true when its owner switched agent sess
 - Pass only the arguments an action takes. `create_session` and `update_session` refuse any other one, list their valid arguments, and change nothing; every other action still runs and names the arguments it did not take in "ignored".
 - `read_task_output` ("session", "task_id" from `read_session`'s tasks, optional "lines") reads the end of a background shell or Monitor task's output, also for an hour after it ended: the last lines, minutes since its last output, and pass/fail counts for a test run. Credentials are scrubbed out.
 
-On this relay and on a paired one, `subscribe` ("sessions": ids, or ["all"]) wakes you when something happens in the sessions you watch, instead of polling them. Leave "events" out unless you need more: the default (question_asked, error, turn_end, subagent_finished, task_failed, stalled, all_idle) wakes you for questions, errors, replies, finished sub-agents, failed tasks (a non-zero exit included), stalls and all_idle, but not for each shell command that ended well. To get those too, pass that list plus "task_finished" (and "task_started" for the starts); each one costs you a turn, so do it only if you must follow every command. "relay_restarted" (this relay only, never in the default) wakes you when the relay runs again after a restart, with who requested it and why; named alone it needs no sessions.
+On this relay and on a paired one, `subscribe` ("sessions": ids, or ["all"]) wakes you when something happens in the sessions you watch, instead of polling them. Leave "events" out unless you need more: the default (question_asked, error, turn_end, subagent_finished, task_failed, stalled, all_idle, board_settled) wakes you for questions, errors, replies, finished sub-agents, failed tasks (a non-zero exit included), stalls, all_idle and a plan card the user settled (board_settled: the button chosen and the message it queued, if any), but not for each shell command that ended well. To get those too, pass that list plus "task_finished" (and "task_started" for the starts); each one costs you a turn, so do it only if you must follow every command. "relay_restarted" (this relay only, never in the default) wakes you when the relay runs again after a restart, with who requested it and why; named alone it needs no sessions.
 
 Each wake-up costs you a turn; these `subscribe` options cut them (questions and errors always come at once):
 - "digest_minutes": N gathers the subscription's other events into one message every N minutes (1 to 1440).

package/server/repositories/message-repository.mjs

@@ -5,6 +5,33 @@ import { createMessagePinStatements } from './message-pin-statements.mjs';
 import { createConversationHistoryPageStatements } from '../services/conversation-history-page.mjs';
 import { sessionWorkerProviderSqlList } from '../../shared/provider-routing.mjs';
 
+/**
+ * The claim's plan-boundary hold, part of every findPending*
+ * statement: a pending row waits while its conversation has a live turn on
+ * the other side of plan mode, because handing it to the worker would steer
+ * it into that turn and switch what the turn may do. The conversation's later
+ * rows wait behind it whatever their mode, so one conversation keeps its
+ * order, while the claim goes on to other conversations' rows.
+ * `PLAN_BOUNDARY_HOLD_MARK` lets a test find any claim without it.
+ */
+export const PLAN_BOUNDARY_HOLD_MARK = '/* plan-boundary-hold */';
+
+export function planBoundaryHoldSql(alias) {
+  return `${PLAN_BOUNDARY_HOLD_MARK} NOT EXISTS (
+    SELECT 1
+    FROM queue held
+    JOIN queue live
+      ON live.conversation_id = held.conversation_id
+     AND live.status = 'processing'
+     AND live.id != held.id
+    WHERE held.conversation_id = ${alias}.conversation_id
+      AND held.status = 'pending'
+      AND COALESCE(held.kind, '') != 'continuation'
+      AND (held.timestamp < ${alias}.timestamp OR (held.timestamp = ${alias}.timestamp AND held.rowid <= ${alias}.rowid))
+      AND (LOWER(TRIM(COALESCE(live.relay_mode, ''))) = 'plan') != (LOWER(TRIM(COALESCE(held.relay_mode, ''))) = 'plan')
+  )`;
+}
+
 export function createMessageRepository(db) {
     const shareVisibility = createShareVisibilityStatements(db);
     const messagePins = createMessagePinStatements(db);
@@ -187,7 +214,7 @@ export function createMessageRepository(db) {
         // steer-recovered row waited 5 minutes while newer messages jumped
         // it). Head-of-line protection is the WHERE clause's next_attempt_at
         // backoff plus MAX_REQUEUE_RETRIES, which fails a row for good.
-        findPending:    db.prepare(`SELECT * FROM queue WHERE status = 'pending' AND COALESCE(kind, '') != 'continuation' AND (next_attempt_at IS NULL OR next_attempt_at <= ?) ORDER BY timestamp ASC, rowid ASC LIMIT 1`),
+        findPending:    db.prepare(`SELECT q.* FROM queue q WHERE q.status = 'pending' AND COALESCE(q.kind, '') != 'continuation' AND (q.next_attempt_at IS NULL OR q.next_attempt_at <= ?) AND ${planBoundaryHoldSql('q')} ORDER BY q.timestamp ASC, q.rowid ASC LIMIT 1`),
         // Global fallback for requesters without a session identity (the legacy
         // Copilot relay CLI). Conversations bound to a session-worker provider
         // (claude/cursor/grok) are excluded outright: handing one of their turns
@@ -218,6 +245,7 @@ export function createMessageRepository(db) {
               NULLIF(rsc.provider_type, ''),
               'github'
             )) NOT IN (${sessionWorkerProviderSqlList()})
+            AND ${planBoundaryHoldSql('q')}
           ORDER BY
             q.timestamp ASC,
             q.rowid ASC
@@ -254,6 +282,7 @@ export function createMessageRepository(db) {
                 NULLIF(c.sdk_session_id, '')
               ) = ?
             )
+            AND ${planBoundaryHoldSql('q')}
           ORDER BY
             CASE
               WHEN COALESCE(
@@ -282,6 +311,7 @@ export function createMessageRepository(db) {
               NULLIF(rs.sdk_session_id, ''),
               NULLIF(c.sdk_session_id, '')
             ) = ?
+            AND ${planBoundaryHoldSql('q')}
           ORDER BY
             q.timestamp ASC,
             q.rowid ASC

package/server/repositories/remote-relay-repository.mjs

@@ -75,6 +75,12 @@ export function createRemoteRelayRepository(db) {
   const getMayCreateSessions = hasMayCreateColumn
     ? db.prepare(`SELECT may_create_sessions FROM conversations WHERE id = ?`)
     : null;
+  // Plan-only (migration 0035). Without the column no session is.
+  const hasPlanOnlyColumn = db.prepare(`PRAGMA table_info(conversations)`).all()
+    .some((column) => column.name === 'plan_only');
+  const getPlanOnly = hasPlanOnlyColumn
+    ? db.prepare(`SELECT plan_only FROM conversations WHERE id = ?`)
+    : null;
 
   return {
     hasUnlock(conversationId, remoteRelayId) {
@@ -161,6 +167,10 @@ export function createRemoteRelayRepository(db) {
     mayCreateSessions(conversationId) {
       return Number(getMayCreateSessions?.get(String(conversationId || ''))?.may_create_sessions) === 1;
     },
+    /** Is this conversation plan-only (agents cannot take it out of plan mode)? */
+    planOnly(conversationId) {
+      return Number(getPlanOnly?.get(String(conversationId || ''))?.plan_only) === 1;
+    },
   };
 }
 

package/server/repositories/session-repository.mjs

@@ -45,6 +45,10 @@ export function createSessionRepository(db) {
     const conversationsHaveMayCreateSessions = db.prepare(`PRAGMA table_info(conversations)`).all()
       .some((column) => String(column?.name || '').trim() === 'may_create_sessions');
     const conversationMayCreateSelect = conversationsHaveMayCreateSessions ? 'c.may_create_sessions' : 'NULL AS may_create_sessions';
+    // Plan-only (migration 0035), the same way.
+    const conversationsHavePlanOnly = db.prepare(`PRAGMA table_info(conversations)`).all()
+      .some((column) => String(column?.name || '').trim() === 'plan_only');
+    const conversationPlanOnlySelect = conversationsHavePlanOnly ? 'c.plan_only' : 'NULL AS plan_only';
     return {
         // conversations
         getConv:        db.prepare(`SELECT * FROM conversations WHERE id = ? AND status != 'deleted'`),
@@ -53,7 +57,8 @@ export function createSessionRepository(db) {
         listConvIdsMissingRuntimeSession: db.prepare(`SELECT c.id AS id FROM conversations c LEFT JOIN runtime_sessions rs ON rs.conversation_id = c.id WHERE rs.id IS NULL AND c.status != 'deleted'`),
         runtimeSessionsSupportProviders,
         conversationsHaveMayCreateSessions,
-        listConvs:      db.prepare(`SELECT c.id, c.title, c.title_source, c.archived, c.compacted_into, c.compacted_from, c.sdk_session_id, c.preferred_relay_mode, c.preferred_model, c.preferred_reasoning_effort, c.configured_workspace_root_path, c.runtime_workspace_root_path, c.draft_text, c.draft_updated_at, c.draft_updated_by_client_id, c.draft_attachments, ${conversationOriginSelect}, ${conversationCloudSelect}, ${conversationGroupSelect}, ${conversationMayCreateSelect}, c.created_at, c.updated_at, rs.id AS runtime_session_id, rs.strategy AS runtime_strategy, rs.status AS runtime_status, rs.last_used_at AS runtime_last_used_at, ${runtimeProviderSelect}, COUNT(m.id) as message_count FROM conversations c LEFT JOIN messages m ON m.conversation_id = c.id LEFT JOIN runtime_sessions rs ON rs.conversation_id = c.id WHERE c.status != 'deleted' AND (? = 1 OR c.archived = ?) GROUP BY c.id ORDER BY CASE WHEN c.sdk_session_id IS NULL OR c.sdk_session_id = '' THEN 1 ELSE 0 END ASC, c.updated_at DESC`),
+        conversationsHavePlanOnly,
+        listConvs:      db.prepare(`SELECT c.id, c.title, c.title_source, c.archived, c.compacted_into, c.compacted_from, c.sdk_session_id, c.preferred_relay_mode, c.preferred_model, c.preferred_reasoning_effort, c.configured_workspace_root_path, c.runtime_workspace_root_path, c.draft_text, c.draft_updated_at, c.draft_updated_by_client_id, c.draft_attachments, ${conversationOriginSelect}, ${conversationCloudSelect}, ${conversationGroupSelect}, ${conversationMayCreateSelect}, ${conversationPlanOnlySelect}, c.created_at, c.updated_at, rs.id AS runtime_session_id, rs.strategy AS runtime_strategy, rs.status AS runtime_status, rs.last_used_at AS runtime_last_used_at, ${runtimeProviderSelect}, COUNT(m.id) as message_count FROM conversations c LEFT JOIN messages m ON m.conversation_id = c.id LEFT JOIN runtime_sessions rs ON rs.conversation_id = c.id WHERE c.status != 'deleted' AND (? = 1 OR c.archived = ?) GROUP BY c.id ORDER BY CASE WHEN c.sdk_session_id IS NULL OR c.sdk_session_id = '' THEN 1 ELSE 0 END ASC, c.updated_at DESC`),
         insertConv:     db.prepare(`INSERT OR IGNORE INTO conversations (id, title, created_at, updated_at) VALUES (?, ?, ?, ?)`),
         updateConvTime: db.prepare(`UPDATE conversations SET updated_at = ? WHERE id = ?`),
         updateConvTitle: db.prepare(`UPDATE conversations SET title = ?, title_source = 'manual' WHERE id = ?`),

package/server/routes/messages-routes.mjs

@@ -15,6 +15,7 @@ import {
 } from '../services/relay-queue-gate-service.mjs';
 import {
   persistConversationModelPreference as persistConversationModelPreferenceTx,
+  readPlanOnly,
 } from '../services/conversation-preferences-service.mjs';
 import { fallbackReasoningEffort, inheritedConversationModel } from '../services/conversation-model-fallback.mjs';
 import { killTmuxSession } from '../services/session-worker-launch-service.mjs';
@@ -86,8 +87,11 @@ import {
 import { TASK_OUTPUT_NONE_MESSAGES } from '../public/app/task-console-render.mjs';
 import { recordClaudeCliStart } from '../services/claude-session-leftovers.mjs';
 import { normalizeClaudeBilling } from '../services/claude-billing-switch-service.mjs';
+import { normalizeTurnTokenUsageReport } from '../../shared/turn-token-usage.mjs';
 import { createClaudeApiStopNoteLedger } from '../services/plan-usage-claude-api.mjs';
 import { normalizeClaudeApiUsageReport } from '../services/claude-api-usage-service.mjs';
+import { planBoundaryHoldSql } from '../repositories/message-repository.mjs';
+import { PLAN_ONLY_REFUSAL, planOnlyRefusesMode } from '../../shared/plan-only.mjs';
 import {
   admitRemoteRelayRequest,
   readRemoteRelayRequest,
@@ -921,6 +925,23 @@ export function resolveResponderConversationMismatch({
   };
 }
 
+/**
+ * Whether a pending row must wait for its conversation's live turn to end
+ * instead of being handed to the worker, which would steer it into that turn.
+ * A steer switches the live turn's mode (the Claude worker re-applies the
+ * permission mode on every delivery), so across plan mode it would change
+ * what the running turn may do: an "Approve plan" follow-up would give a
+ * running plan turn write access, and a plan message would take it from a
+ * running write turn. Rows on the same side of plan mode steer as before.
+ */
+export function crossesPlanBoundaryOfLiveTurn({ stmts, row } = {}) {
+  const conversationId = String(row?.conversation_id || '').trim();
+  if (!conversationId || typeof stmts?.listProcessingQueueByConversation?.all !== 'function') return false;
+  const isPlan = (mode) => String(mode || '').trim().toLowerCase() === 'plan';
+  const pendingIsPlan = isPlan(row.relay_mode);
+  return stmts.listProcessingQueueByConversation.all(conversationId)
+    .some((live) => String(live?.id || '') !== String(row.id || '') && isPlan(live?.relay_mode) !== pendingIsPlan);
+}
 export function dequeuePendingMessage({
   db,
   stmts,
@@ -970,6 +991,10 @@ export function dequeuePendingMessage({
         : stmts.findPending.get(currentIso);
     }
     if (!next) return null;
+    // Across plan mode, a row waits for the live turn instead of steering in.
+    // The claim statements leave such rows out (planBoundaryHoldSql); this is
+    // the fail-closed backstop for one that does not.
+    if (crossesPlanBoundaryOfLiveTurn({ stmts, row: next })) return null;
     // One processing attempt = one attempt_id. Minted here for every claim —
     // even for callers that never echo it back — so finalization, requeue, and
     // question writes can be fenced to the attempt that actually owns the row.
@@ -2049,6 +2074,8 @@ export function registerMessagesRoutes(app, deps) {
     claudeBillingSwitchService = null,
     // The Claude API booking (claude-api-usage-service.mjs).
     claudeApiUsageService = null,
+    // Tokens per turn for every provider (token-usage-service.mjs).
+    tokenUsageService = null,
     // The delivery fence's runtime half for GET /api/pending, as on the
     // socket: the Claude API claim check (null: no refusal) and what a fenced
     // row does (the swap, or fail closed).
@@ -2640,6 +2667,7 @@ export function registerMessagesRoutes(app, deps) {
           AND rs.sdk_session_id = ?
       )
       AND (q.next_attempt_at IS NULL OR q.next_attempt_at <= ?)
+      AND ${planBoundaryHoldSql('q')}
     ORDER BY q.timestamp ASC, q.rowid ASC
     LIMIT 1
   `);
@@ -4804,6 +4832,12 @@ export function registerMessagesRoutes(app, deps) {
       ? normalizeRelayMode(existingConversation.preferred_relay_mode)
       : null;
     const requestedRelayMode = savedRelayMode || normalizeRelayMode(relayMode || mode);
+    // Plan-only: an agent's send that names another mode is refused
+    // and nothing is queued (nor a billing switched). Without a mode it runs
+    // in the session's own. The owner's own sends are not limited.
+    if (remoteRequest.remote && readPlanOnly(existingConversation) && planOnlyRefusesMode(relayMode || mode)) {
+      return res.status(409).json({ ...PLAN_ONLY_REFUSAL, conversationId });
+    }
     const trimmedText = stripRelayPromptContext(text, requestedRelayMode || relayMode || mode);
     const normalizedAttachments = normalizeAttachments(rawAttachments);
     const requestedAttachmentCount = Math.min(
@@ -6611,6 +6645,59 @@ export function registerMessagesRoutes(app, deps) {
     res.json({ ok: true, billing: booked.billing, basis: booked.basis, models: booked.booked.map((entry) => entry.model) });
   });
 
+  // POST /api/turn-token-usage — every worker reports, after a turn, the
+  // turn's tokens per model (`models`), or a Claude cloud worker the
+  // session's running totals (`sessionTotals`), keyed by turn, run and model
+  // (shared/turn-token-usage.mjs). Advisory and idempotent: a retried or
+  // reordered report books only what it adds (token-usage-service.mjs).
+  // Who ran the turn is the relay's to say: the chat's provider, and for a
+  // Claude chat the posting worker's registered kind (the billing can switch
+  // between turns). A body that says otherwise is logged. A deleted chat
+  // takes no report, so a late retry cannot book its turn a second time; an
+  // archived one does.
+  app.post('/api/turn-token-usage', auth, (req, res) => {
+    touchCli();
+    if (!tokenUsageService) return res.status(501).json({ error: 'Token usage storage is unavailable' });
+    const parsed = normalizeTurnTokenUsageReport(req.body);
+    if (!parsed.ok) return res.status(400).json({ error: parsed.error });
+    const { report } = parsed;
+    const conversationId = report.conversationId;
+    if (!stmts.getConv?.get?.(conversationId)) return res.status(404).json({ error: 'Conversation not found' });
+    // The binding's provider wins over the report's. A chat whose binding is
+    // gone for a moment (a cache rebuild) still books, as the provider the
+    // report states.
+    const runtimeSession = stmts.getRuntimeSessionByConversation?.get?.(conversationId) || null;
+    const chatProvider = String(runtimeSession?.provider_type || report.provider || '').trim().toLowerCase();
+    let provider = chatProvider;
+    let billing = '';
+    if (chatProvider === 'openai') {
+      // A chat on the user's own OpenAI key runs on the Copilot SDK.
+      provider = 'github';
+      billing = 'byok-openai';
+    } else if (chatProvider === 'github') {
+      billing = report.billing;
+    } else if (normalizeClaudeBilling(chatProvider)) {
+      const requesterSid = normalizeSessionWorkerId(readBridgeIdentity(req)?.sessionId);
+      const registeredKind = normalizeClaudeBilling(requesterSid ? sessionWorkerRegistry?.getWorker?.(requesterSid)?.kind : '');
+      // The worker that ran the turn, then the chat's billing; the report's
+      // own word only when the chat has no binding.
+      provider = registeredKind || normalizeClaudeBilling(chatProvider);
+    } else if (!['claude-cloud', 'deepseek', 'cursor', 'grok'].includes(chatProvider)) {
+      return res.status(409).json({ error: 'Conversation keeps no token usage' });
+    }
+    if (report.provider && report.provider !== provider) {
+      console.warn(`[${ts()}] token usage conv=${conversationId.slice(0, 8)} stated=${report.provider} booked=${provider}`);
+    }
+    if (report.sessionTotals && provider !== 'claude-cloud') {
+      return res.status(400).json({ error: 'sessionTotals is for Claude cloud only' });
+    }
+    const booked = report.sessionTotals
+      ? tokenUsageService.recordCloudSessionTotals({ conversationId, provider, billing, report })
+      : tokenUsageService.recordTurn({ conversationId, provider, billing, report });
+    if (!booked) return res.status(501).json({ error: 'Token usage storage is unavailable' });
+    res.json({ ok: true, provider, billing, models: booked.map((entry) => entry.model) });
+  });
+
   // POST /api/claude-plan-usage — Claude worker reports the session's
   // structured /usage data (plan rate-limit windows + session cost totals).
   //

package/server/routes/relay-ask-routes.mjs

@@ -27,12 +27,15 @@
 //                                          relay from its Inbox: asks only;
 //                                          an approval or any other card is
 //                                          never answered over the pairing
-//   GET  /api/relay-asks/remote/inbox      what its Inbox shows of this relay
+//   GET  /api/relay-asks/remote/inbox      what its Inbox shows of this relay;
+//                                          `?push=1`: it can push, and each
+//                                          card a turn waits on says who does
+//                                          (`notice`)
 //
 // Every text of an ask or card is an agent's: it goes out as data, and the
 // browser shows it as plain text.
 
-import { REMOTE_RELAY_ERROR_CODES } from '../../shared/remote-relay-contract.mjs';
+import { REMOTE_INBOX_CLAIMS, REMOTE_INBOX_NOTICES, REMOTE_RELAY_ERROR_CODES } from '../../shared/remote-relay-contract.mjs';
 import { RELAY_ASK_LIMITS, relayQuestionKind } from '../../shared/relay-asks.mjs';
 import { schemaFields } from '../../shared/question-schema.mjs';
 import { findOriginRemoteRelay, admitRemoteRelayRequest, readRemoteRelayRequest } from '../services/remote-relay-inbound.mjs';
@@ -53,6 +56,8 @@ const RELAY_ASK_ERROR_STATUS = Object.freeze({
 const INBOX_REMOTE_TIMEOUT_MS = 5_000;
 // Cards answered in the last day are listed under the divider, at most these.
 const INBOX_ANSWERED_CARDS_MAX = 50;
+// Plan cards, pending (one per session) and settled in the last day, at most these each.
+const INBOX_PLAN_CARDS_MAX = 50;
 // Registry states that mean "do not even try" for the Inbox.
 const UNREACHABLE_STATES = new Set(['offline', 'unauthorized']);
 
@@ -90,7 +95,15 @@ export function inboxItemOfCard(question, session, { forRemote = false } = {}) {
     createdAt: question?.createdAt || null,
     ...(question?.status === 'answered' ? { answeredAt: question?.answeredAt || null } : {}),
   };
-  if (forRemote && kind === 'approval') return { ...base, question: '' };
+  // Its kind and summary say what it is about, never what it would send.
+  if (forRemote && kind === 'approval') {
+    return {
+      ...base,
+      question: '',
+      ...(toText(context.variant) ? { variant: toText(context.variant) } : {}),
+      ...(toText(context.summary) ? { summary: toText(context.summary) } : {}),
+    };
+  }
   const fields = schemaFields(question?.requestSchema);
   return {
     ...base,
@@ -107,6 +120,37 @@ export function inboxItemOfCard(question, session, { forRemote = false } = {}) {
   };
 }
 
+/**
+ * One Inbox item of a plan card (a formatted relay board). A paired
+ * relay's Inbox gets its title and status alone, "Open there": a card is
+ * never answered over the pairing. `question` repeats the title for one
+ * that does not know plan items (0.9.19 shows it as a card's question).
+ */
+export function inboxItemOfPlanBoard(board, session, { forRemote = false } = {}) {
+  const title = toText(board?.title) || 'Plan ready for review';
+  const status = toText(board?.status) || 'pending';
+  const base = {
+    id: toText(board?.id),
+    kind: 'plan',
+    session,
+    title,
+    status,
+    createdAt: board?.createdAt || null,
+  };
+  if (forRemote) return { ...base, question: title };
+  return {
+    ...base,
+    ...(status !== 'pending' ? { actedAt: board?.actedAt || null } : {}),
+    body: String(board?.body || ''),
+    actions: (Array.isArray(board?.actions) ? board.actions : [])
+      .map((action) => ({ id: toText(action?.id), label: toText(action?.label) || toText(action?.id) }))
+      .filter((action) => action.id),
+    recommendedAction: toText(board?.recommendedAction) || null,
+    planOnly: board?.planOnly === true,
+    selectedAction: toText(board?.selectedAction) || null,
+  };
+}
+
 /** One Inbox item of an ask (relay-ask-service.mjs shape). */
 export function inboxItemOfAsk(ask) {
   const {
@@ -139,6 +183,8 @@ export function registerRelayAskRoutes(app, deps) {
     db,
     stmts,
     formatQuestionRow,
+    // Plan cards (relay boards) in the Inbox, formatted as their routes do.
+    formatRelayBoardRow = null,
     relayAskService = null,
     relayAskLinkService = null,
     remoteRelayInbound = null,
@@ -146,6 +192,9 @@ export function registerRelayAskRoutes(app, deps) {
     // calls them (remote-relay-client.mjs): the Inbox of paired relays.
     remoteRelayRegistry = null,
     remoteRelayClient = null,
+    // push-dispatch-service.mjs questionNotice: whether this relay's own push
+    // of a card reached a device.
+    questionNotice = null,
     now = () => Date.now(),
   } = deps;
 
@@ -154,6 +203,26 @@ export function registerRelayAskRoutes(app, deps) {
     SELECT * FROM relay_questions WHERE status = 'answered' AND answered_at >= ? ORDER BY answered_at DESC LIMIT ?
   `) || null;
 
+  // Plan cards: of each session that is neither deleted nor
+  // archived the newest pending one (an older one stays in its session),
+  // and the ones settled in the last day.
+  const planBoardSessionFilter = `c.status != 'deleted' AND COALESCE(c.archived, 0) = 0`;
+  const pendingPlanBoardsStmt = db?.prepare?.(`
+    SELECT b.* FROM relay_boards b JOIN conversations c ON c.id = b.conversation_id
+    WHERE b.status = 'pending' AND b.board_type = 'plan_ready' AND ${planBoardSessionFilter}
+      AND NOT EXISTS (
+        SELECT 1 FROM relay_boards newer
+        WHERE newer.conversation_id = b.conversation_id AND newer.status = 'pending' AND newer.board_type = 'plan_ready'
+          AND (newer.created_at > b.created_at OR (newer.created_at = b.created_at AND newer.rowid > b.rowid))
+      )
+    ORDER BY b.created_at DESC LIMIT ?
+  `) || null;
+  const settledPlanBoardsStmt = db?.prepare?.(`
+    SELECT b.* FROM relay_boards b JOIN conversations c ON c.id = b.conversation_id
+    WHERE b.status IN ('acted', 'dismissed') AND b.board_type = 'plan_ready' AND b.acted_at >= ? AND ${planBoardSessionFilter}
+    ORDER BY b.acted_at DESC LIMIT ?
+  `) || null;
+
   function sessionOf(id) {
     let row = null;
     try {
@@ -204,13 +273,20 @@ export function registerRelayAskRoutes(app, deps) {
     return { relayId, name: toText(paired?.name) || claimed || toText(request.origin?.relayName) || relayId };
   }
 
-  /** What an Inbox shows of this relay: its asks, and the cards turns wait on. */
-  function localInboxItems({ forRemote = false } = {}) {
+  /**
+   * What an Inbox shows of this relay: its asks, and the cards turns wait on.
+   * `read`, when given, learns whether the pending cards could be read
+   * (`read.pendingCards`): a failed read lists none, which is not "none pending".
+   */
+  function localInboxItems({ forRemote = false, read = null } = {}) {
     const items = (relayAskService?.isAvailable?.() ? relayAskService.listHere() : []).map(inboxItemOfAsk);
     let pending = [];
     try {
       pending = stmts.listQuestions.all('pending', null, null).map(formatQuestionRow).filter(Boolean);
-    } catch {}
+      if (read) read.pendingCards = true;
+    } catch {
+      if (read) read.pendingCards = false;
+    }
     let answered = [];
     try {
       const since = new Date(now() - RELAY_ASK_LIMITS.answeredKeepMs).toISOString();
@@ -219,6 +295,19 @@ export function registerRelayAskRoutes(app, deps) {
     for (const question of [...pending, ...answered]) {
       items.push(inboxItemOfCard(question, sessionOf(question.conversationId), { forRemote }));
     }
+    if (typeof formatRelayBoardRow === 'function') {
+      let boards = [];
+      try {
+        const since = new Date(now() - RELAY_ASK_LIMITS.answeredKeepMs).toISOString();
+        boards = [
+          ...(pendingPlanBoardsStmt?.all(INBOX_PLAN_CARDS_MAX) || []),
+          ...(settledPlanBoardsStmt?.all(since, INBOX_PLAN_CARDS_MAX) || []),
+        ].map(formatRelayBoardRow).filter(Boolean);
+      } catch {}
+      for (const board of boards) {
+        items.push(inboxItemOfPlanBoard(board, sessionOf(board.conversationId), { forRemote }));
+      }
+    }
     return items.sort((a, b) => String(b.createdAt || '').localeCompare(String(a.createdAt || '')));
   }
 
@@ -375,11 +464,52 @@ export function registerRelayAskRoutes(app, deps) {
     }
   });
 
-  // GET /api/relay-asks/remote/inbox — what a paired relay's Inbox shows of this one.
+  // Who pushes a card a turn waits on: question id → the paired
+  // relay (its id) that claimed it. In memory: after a restart a
+  // card is just claimable again.
+  const claims = new Map();
+
+  /**
+   * `notice` (and with `you`, `claimed`) for each pending approval and
+   * ask_user item, for a caller that can push: this relay's own push
+   * decides first, then the first caller to find it missed claims it.
+   */
+  function addNotices(items, callerRelayId) {
+    const pendingIds = new Set();
+    const noticed = items.map((item) => {
+      if (item.status !== 'pending' || (item.kind !== 'approval' && item.kind !== 'ask_user')) return item;
+      pendingIds.add(item.id);
+      let own = null;
+      try {
+        own = typeof questionNotice === 'function' ? questionNotice(item.id) : null;
+      } catch {}
+      if (own === 'reached') return { ...item, notice: REMOTE_INBOX_NOTICES.here };
+      if (own === 'pending') return { ...item, notice: REMOTE_INBOX_NOTICES.wait };
+      const claim = claims.get(item.id);
+      if (!claim) {
+        claims.set(item.id, callerRelayId);
+        return { ...item, notice: REMOTE_INBOX_NOTICES.you, claimed: REMOTE_INBOX_CLAIMS.now };
+      }
+      if (claim === callerRelayId) return { ...item, notice: REMOTE_INBOX_NOTICES.you, claimed: REMOTE_INBOX_CLAIMS.before };
+      return { ...item, notice: REMOTE_INBOX_NOTICES.other };
+    });
+    // A card no longer pending needs no claim.
+    for (const id of claims.keys()) {
+      if (!pendingIds.has(id)) claims.delete(id);
+    }
+    return noticed;
+  }
+
+  // GET /api/relay-asks/remote/inbox — what a paired relay's Inbox shows of
+  // this one. `?push=1`: the caller can push, and learns who pushes each card.
   app.get('/api/relay-asks/remote/inbox', auth, (req, res) => {
     const caller = admitPairedRelay(req, res);
     if (!caller) return;
-    const items = localInboxItems({ forRemote: true });
+    const read = {};
+    let items = localInboxItems({ forRemote: true, read });
+    // Pending cards that could not be read are not "no card pending": the
+    // claims stay, and no card is claimed or offered on this read.
+    if (toText(req.query?.push) === '1' && read.pendingCards) items = addNotices(items, caller.relayId);
     res.json({ ok: true, items, pending: pendingOf(items) });
   });