trantor 0.18.110 → 0.18.111

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "trantor",
-  "version": "0.18.110",
+  "version": "0.18.111",
   "description": "Trantor — the hub-world for AI agent crews: live message bus, presence, project Kanban/flow board + crew orchestration for independent AI coding agents (Claude, Codex, Gemini, Kimi, DeepSeek)",
   "mcpServers": {
     "relay": {

package/bin/crew/codex-relay.mjs

@@ -0,0 +1,8 @@
+// Codex's MCP env table overrides ambient runner variables and may pin another seat.
+export function codexRelayFlags({ persona, session, project, hub }) {
+  const env = { RELAY_AGENT: persona, RELAY_SESSION: session, RELAY_PROJECT: project, RELAY_URL: hub };
+  return Object.entries(env).map(([key, value]) => {
+    const setting = `mcp_servers.relay.env.${key}=${JSON.stringify(value)}`;
+    return ` -c '${setting.replaceAll("'", "'\"'\"'")}'`;
+  }).join("");
+}

package/bin/crew-runner.mjs

@@ -48,6 +48,7 @@ import { costLine } from "../lib/state/cost.mjs";
 import { resolveStateFlags, stateFeatureOn } from "../lib/state/flags.mjs";
 
 import { awaitSwapRelease } from "./crew/stage.mjs";
+import { codexRelayFlags } from "./crew/codex-relay.mjs";
 
 const releaseState = process.env.CREW_RELEASE_STATE ? JSON.parse(process.env.CREW_RELEASE_STATE) : null;
 delete process.env.CREW_RELEASE_STATE;
@@ -243,8 +244,8 @@ let MODEL = process.env.CREW_MODEL || "";
 // ZAI-coding-plan models into deepseek/…). `scrooge route` returns qualified ids.
 if (MODEL && !MODEL.includes("/") && AGENT === "deepseek") MODEL = `deepseek/${MODEL}`;
 const CLI = {
-  codex:    { first: `codex exec{M} --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox "$(cat {P})" < /dev/null`,
-              next:  `codex exec resume --last{M} --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox "$(cat {P})" < /dev/null`, mflag: " -m " },
+  codex:    { first: `codex exec{M}{R} --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox "$(cat {P})" < /dev/null`,
+              next:  `codex exec resume --last{M}{R} --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox "$(cat {P})" < /dev/null`, mflag: " -m " },
   gemini:   { first: `gemini --yolo{M} -p "$(cat {P})"`,
               next:  `gemini --yolo{M} -r latest -p "$(cat {P})"`, mflag: " -m " },
   // kimi-code (successor to kimi-cli) has no --print (-p alone is non-interactive), REJECTS
@@ -863,7 +864,8 @@ async function runTurn(prompt, isFirst, trigger = "kickoff", opts = {}) {
   const turnEffort = effortForTurn({ agent: DRIVER, model: MODEL, launchEffort: EFFORT, cardDifficulty });
   const mfrag = (MODEL && cli.mflag ? `${cli.mflag}${MODEL}` : "") + turnEffort.flag;
   cmd = cmd.replaceAll("{M}", mfrag).replaceAll("{P}", pf).replaceAll("{SID}", sid).replaceAll("{DIR}", TURN_DIR)
-    .replaceAll("{S}", STATE_SCHEMA_FILE);
+    .replaceAll("{S}", STATE_SCHEMA_FILE)
+    .replaceAll("{R}", () => codexRelayFlags({ persona: PERSONA, session: SESSION, project: PROJ, hub: HUB }));
   // PRECEDENCE: each file is PREPENDED, so the list is iterated in written order, highest priority
   // first, and the CREW layer (~/.agent-bus/.env) wins. test-crew-env.mjs runs the real shell.
   const envs = [join(homedir(), ".agent-bus", ".env"), cli.env].filter(f => f && existsSync(f));

package/bin/keys.mjs

@@ -8,7 +8,7 @@ import { join } from "node:path";
 import { homedir } from "node:os";
 import { load, migrateKeysToKeychain, keyLocationReport, IDENTITY_SERVICE } from "../lib/identity.mjs";
 import { sfetch } from "../lib/signed-fetch.mjs";
-import { pruneTestKeys } from "../lib/key-prune.mjs";
+import { pruneTestKeys, pruneStaleTmp } from "../lib/key-prune.mjs";
 
 const args = process.argv.slice(2);
 const verb = args[0] || "status";
@@ -65,6 +65,10 @@ async function prune() {
   const idleDays = at < 0 ? 7 : Number(args[at + 1]);
   if (!Number.isFinite(idleDays) || idleDays <= 0) throw new Error("--idle-days must be a positive number");
   if (args.includes("--apply") && args.includes("--dry-run")) throw new Error("choose --apply or --dry-run");
+  const apply = args.includes("--apply");
+  // The tmp sweep is local (a stranded atomic-write temp is not a key and needs no hub) and runs
+  // first, so it works even when the hub is unreachable; --apply gates it like the rest of prune.
+  const tmp = pruneStaleTmp({ apply });
   const bus = process.env.AGENT_BUS_DIR || join(homedir(), ".agent-bus");
   let config = {};
   try { config = JSON.parse(readFileSync(join(bus, "config.json"), "utf8")); }
@@ -79,8 +83,11 @@ async function prune() {
       snapshots.push(await response.json());
     } catch (error) { throw new Error(`cannot check hub ${hub}: ${error.message}; no keys pruned`); }
   }
-  const result = pruneTestKeys({ snapshots, idleDays, apply: args.includes("--apply") });
-  if (JSON_MODE) { console.log(JSON.stringify(result, null, 2)); return; }
+  const result = pruneTestKeys({ snapshots, idleDays, apply });
+  if (JSON_MODE) { console.log(JSON.stringify({ ...result, tmp }, null, 2)); return; }
+  const tmpDid = tmp.dryRun ? "dry run — would remove" : "removed";
+  console.log(`${tmpDid} ${tmp.removed.length} stale key-write tmp file(s) older than 1h${tmp.dryRun && tmp.removed.length ? "; use --apply to remove them" : ""}`);
+  for (const f of tmp.removed) console.log(`  ${f}`);
   console.log(`${result.dryRun ? "dry run — would remove" : "removed"} ${result.removed.length} test key file(s), idle over ${idleDays} days`);
   for (const name of result.removed) console.log(`  ${name}`);
   console.log(`kept ${result.kept.length} key file(s)${result.dryRun ? "; use --apply to remove the listed files" : ""}`);

package/bin/pair.mjs

@@ -9,12 +9,14 @@ import { spawnSync } from "node:child_process";
 
 const ROOT = dirname(dirname(fileURLToPath(import.meta.url)));
 const args = process.argv.slice(2);
+const USAGE = "usage: trantor pair [--role owner] [--sim] [--json]";
+if (args.includes("--help") || args.includes("-h")) { console.log(USAGE); process.exit(0); }
 const asJson = args.includes("--json");
 const openSim = args.includes("--sim");
 const roleIdx = args.indexOf("--role");
 const role = roleIdx >= 0 ? (args[roleIdx + 1] || "") : "approver";
 if (!role || role.startsWith("--")) {
-  console.error("usage: trantor pair [--role owner] [--sim] [--json]");
+  console.error(USAGE);
   process.exit(1);
 }
 

package/hub/auth.mjs

@@ -53,9 +53,9 @@ function cmpSemver(a, b) {
 }
 const AUTH_HEADERS = ["x-trantor-pubkey", "x-trantor-sig", "x-trantor-ts", "x-trantor-nonce"];
 const PUBLIC_ENDPOINTS = new Set(["/", "/ui", "/health", "/enroll"]);
-const OWNER_ENDPOINTS = new Set(["/project/delete", "/sweep", "/reconcile", "/invite", "/import", "/policy"]);
+const OWNER_ENDPOINTS = new Set(["/project/delete", "/sweep", "/reconcile", "/invite", "/import", "/policy", "/session/prompts", "/session/prompt/claim", "/session/prompt/ack"]);
 const DEVICE_ENDPOINTS = new Set(["/device/register", "/device/unregister", "/devices"]);
-const READ_ENDPOINTS = new Set(["/peers", "/tasks", "/events", "/inbox", "/peer", "/card", "/stream", "/history", "/projects", "/catchup", "/phases", "/recent", "/handoffs", "/verify-gates", "/claims", "/proposals", "/grants", "/overseer/context", "/overseer/status"]);
+const READ_ENDPOINTS = new Set(["/peers", "/tasks", "/events", "/inbox", "/peer", "/card", "/stream", "/history", "/projects", "/catchup", "/phases", "/recent", "/handoffs", "/verify-gates", "/claims", "/proposals", "/grants", "/overseer/context", "/overseer/status", "/session/chat", "/session/prompt"]);
 const roleRank = { read: 1, write: 2, approver: 3, owner: 4 };
 const hasAuthHeaders = (req) => AUTH_HEADERS.some(h => !!req.headers[h]);
 const authPath = (u) => `${u.pathname}${u.search || ""}`;
@@ -116,6 +116,13 @@ function projectFromRequest(P, q, b) {
     const p = state.proposals.find(x => x.id === Number(b?.id ?? q?.id));
     return p?.project || "";
   }
+  // claim/ack and the single-prompt GET carry only an id — authorization runs against the
+  // PROMPT's project, or a project-scoped key could claim any project's queued prompt (#12129).
+  if (P === "/session/prompt" || P === "/session/prompt/claim" || P === "/session/prompt/ack") {
+    const pr = (state.sessionPrompts || []).find(x => x.id === Number(b?.id ?? q?.id));
+    if (pr) return pr.project || "";
+    if (P !== "/session/prompt") return "";
+  }
   return canon(String(b?.project || q?.project || "").slice(0, 80));
 }
 // Cross-project actions require a declared link or the operator identity, even with wildcard scopes (#6228).

package/hub/routes/session.mjs

@@ -0,0 +1,187 @@
+// #12129: the hub is the meeting point between the desktop relay and the phone. The chat mirror
+// is MEMORY-ONLY (a restart answers have:0, the desktop resends its window); the prompt queue
+// lives in state so a restart never strands a queued prompt. Live frames ride pushEventToStreams
+// (scope-filtered, never logged); only chat.prompt.queued/.delivered/.failed are audit-appended.
+import { setInterval } from "node:timers";
+
+const CHAT_CAP = 400;            // turns kept per project, highest idx wins
+const CHAT_PAGE = 200;           // max turns per GET
+const PROMPT_CAP = 200;
+const PROMPT_TEXT_MAX = 8000;
+const EVENT_TEXT_MAX = 500;
+const PROMPT_TTL_MS = Number(process.env.RELAY_PROMPT_TTL_MS || 120000);
+const PROMPT_SWEEP_MS = Number(process.env.RELAY_PROMPT_SWEEP_MS || 30000);
+const EXPIRED_REASON = "No Mac is relaying this project. Is the Trantor app open?";
+const ACK_OUTCOMES = new Set(["delivered", "blocked", "not_ready", "stalled", "no_agent"]);
+
+export function createSessionRuntime({ state, markDirty, appendEvent, pushEventToStreams }) {
+  const now = () => Date.now();
+  const mirrors = new Map();     // project -> { session, total, turns: Map(idx->turn), have, meta, updatedAt, host }
+  state.sessionPrompts ||= [];
+  state.sessionPromptSeq = Number(state.sessionPromptSeq || 0);
+
+  const findPrompt = (id) => state.sessionPrompts.find(p => p.id === Number(id));
+
+  // A prompt still queued past the TTL means no desktop relay claimed it. Expiry is lazy (on
+  // read) plus a sweep that is a no-op when nothing is queued, so an idle hub does no work.
+  function expireQueued() {
+    let changed = false;
+    for (const p of state.sessionPrompts) {
+      if (p.status !== "queued" || now() - p.ts <= PROMPT_TTL_MS) continue;
+      p.status = "expired"; p.outcome = "expired"; p.error = EXPIRED_REASON;
+      appendEvent("chat.prompt.failed", p.project, p.by, { promptId: p.id, outcome: "expired", error: EXPIRED_REASON });
+      changed = true;
+    }
+    if (changed) markDirty();
+    return changed;
+  }
+  const sweep = setInterval(() => {
+    if (state.sessionPrompts.some(p => p.status === "queued")) expireQueued();
+  }, PROMPT_SWEEP_MS);
+  sweep.unref?.();
+
+  function mergeChat(b, auth) {
+    const project = String(b.project || "").slice(0, 80);
+    const session = String(b.session || "").slice(0, 200);
+    if (!project) return { code: 400, body: { error: "project required" } };
+    if (!session) return { code: 400, body: { error: "session required" } };
+    let m = mirrors.get(project);
+    // A different session resets the mirror (handoff/compaction): the desktop then sends its
+    // whole window, and stale turns from the previous session must not survive the switch.
+    if (!m || m.session !== session) {
+      m = { session, total: 0, turns: new Map(), have: 0, meta: {}, updatedAt: 0, host: "" };
+      mirrors.set(project, m);
+    }
+    // The last turn grows while the assistant writes, so the desktop resends from the lowest
+    // changed idx and the hub REPLACES by idx — re-sends are idempotent by construction.
+    for (const t of Array.isArray(b.turns) ? b.turns : []) {
+      const idx = Number(t?.idx);
+      if (!Number.isInteger(idx) || idx < 0) continue;
+      const clean = {
+        idx,
+        role: String(t.role || "").slice(0, 40),
+        blocks: Array.isArray(t.blocks) ? t.blocks : [],
+      };
+      if (t.queued) clean.queued = true;
+      m.turns.set(idx, clean);
+      if (idx + 1 > m.have) m.have = idx + 1;
+    }
+    if (m.turns.size > CHAT_CAP) {
+      const drop = [...m.turns.keys()].sort((a, b) => a - b).slice(0, m.turns.size - CHAT_CAP);
+      for (const k of drop) m.turns.delete(k);
+    }
+    m.total = Math.max(0, Number(b.total) || 0);
+    // Property access on a primitive is safe (undefined); only null/undefined would throw.
+    const meta = b.meta || {};
+    m.meta = {
+      working: !!meta.working,
+      model: String(meta.model || "").slice(0, 120),
+      context: meta.context ?? null,
+    };
+    m.updatedAt = now();
+    m.host = String(b.host || auth?.identity?.name || "").slice(0, 120);
+    pushEventToStreams({ ts: m.updatedAt, type: "chat.updated", project, total: m.total, working: m.meta.working });
+    return { code: 200, body: { ok: true, have: m.have } };
+  }
+
+  async function route({ req, res, q, P, auth, ctx }) {
+    const isChat = P === "/session/chat";
+    const isPrompts = P === "/session/prompts";
+    const isPrompt = P === "/session/prompt" || P === "/session/prompt/claim" || P === "/session/prompt/ack";
+    if (!isChat && !isPrompts && !isPrompt) return false;
+    const { body, json, canon } = ctx;
+
+    if (req.method === "POST" && isChat) {
+      const b = await body(req);
+      // the same canon the GET reads with, or a publish and a read could name two mirrors
+      b.project = canon(String(b.project || "").slice(0, 80));
+      const r = mergeChat(b, auth);
+      return json(res, r.code, r.body);
+    }
+
+    if (req.method === "GET" && isChat) {
+      const project = canon(String(q.project || "").slice(0, 80));
+      const m = mirrors.get(project);
+      if (!m) return json(res, 200, { session: "", total: 0, turns: [], meta: {}, updatedAt: 0, host: "" });
+      const afterRaw = Number(q.after);
+      const after = Number.isFinite(afterRaw) ? afterRaw : -1;
+      const turns = [...m.turns.values()].filter(t => t.idx > after).sort((a, b) => a.idx - b.idx).slice(0, CHAT_PAGE);
+      return json(res, 200, { session: m.session, total: m.total, turns, meta: m.meta, updatedAt: m.updatedAt, host: m.host });
+    }
+
+    if (req.method === "POST" && P === "/session/prompt") {
+      const b = await body(req);
+      const project = canon(String(b.project || "").slice(0, 80));
+      const text = String(b.text || "");
+      if (!project || !text.trim()) return json(res, 400, { error: "project and non-empty text required" });
+      if (text.length > PROMPT_TEXT_MAX) return json(res, 400, { error: `text must be <= ${PROMPT_TEXT_MAX} chars` });
+      expireQueued();
+      state.sessionPromptSeq += 1;
+      const p = { id: state.sessionPromptSeq, project, text, by: String(auth?.identity?.name || "").slice(0, 120), ts: now(), status: "queued" };
+      state.sessionPrompts.push(p);
+      // Cap: evict oldest terminal rows first so live work is never the first casualty.
+      while (state.sessionPrompts.length > PROMPT_CAP) {
+        const ti = state.sessionPrompts.findIndex(x => x.status !== "queued" && x.status !== "claimed");
+        state.sessionPrompts.splice(ti >= 0 ? ti : 0, 1);
+      }
+      markDirty();
+      appendEvent("chat.prompt.queued", project, p.by, { promptId: p.id, text: text.slice(0, EVENT_TEXT_MAX) });
+      pushEventToStreams({ ts: p.ts, type: "chat.prompt", id: p.id, project });
+      return json(res, 200, { id: p.id, status: "queued" });
+    }
+
+    if (req.method === "GET" && isPrompts) {
+      expireQueued();
+      const project = canon(String(q.project || "").slice(0, 80));
+      const status = String(q.status || "").slice(0, 40);
+      const prompts = state.sessionPrompts
+        .filter(p => (!project || p.project === project) && (!status || p.status === status))
+        .map(p => ({ id: p.id, project: p.project, text: p.text, by: p.by, ts: p.ts, status: p.status }));
+      return json(res, 200, { prompts });
+    }
+
+    if (req.method === "POST" && P === "/session/prompt/claim") {
+      const b = await body(req);
+      expireQueued();
+      const p = findPrompt(b.id);
+      if (!p) return json(res, 404, { error: "prompt not found" });
+      // The first claim wins; a second claim — even from the same host — gets 409.
+      if (p.status !== "queued") return json(res, 409, { error: `prompt already ${p.status}`, status: p.status });
+      p.status = "claimed"; p.host = String(b.host || "").slice(0, 120); p.claimedAt = now();
+      markDirty();
+      // The full text goes only to the claimer.
+      return json(res, 200, { ok: true, prompt: { id: p.id, project: p.project, text: p.text, by: p.by, ts: p.ts } });
+    }
+
+    if (req.method === "POST" && P === "/session/prompt/ack") {
+      const b = await body(req);
+      const p = findPrompt(b.id);
+      if (!p) return json(res, 404, { error: "prompt not found" });
+      const outcome = String(b.outcome || "");
+      if (!ACK_OUTCOMES.has(outcome)) return json(res, 400, { error: `outcome must be one of: ${[...ACK_OUTCOMES].join(", ")}` });
+      if (p.status === "delivered" || p.status === "failed" || p.status === "expired") {
+        return json(res, 409, { error: `prompt already ${p.status}`, status: p.status });
+      }
+      const error = String(b.error || "").slice(0, EVENT_TEXT_MAX);
+      p.outcome = outcome; p.error = error; p.ackedAt = now();
+      p.status = outcome === "delivered" ? "delivered" : "failed";
+      markDirty();
+      const extra = { promptId: p.id, outcome };
+      if (error) extra.error = error;
+      appendEvent(outcome === "delivered" ? "chat.prompt.delivered" : "chat.prompt.failed", p.project,
+        String(auth?.identity?.name || ""), extra);
+      return json(res, 200, { ok: true, status: p.status });
+    }
+
+    if (req.method === "GET" && P === "/session/prompt") {
+      expireQueued();
+      const p = findPrompt(q.id);
+      if (!p) return json(res, 404, { error: "prompt not found" });
+      return json(res, 200, { id: p.id, status: p.status, outcome: p.outcome || "", error: p.error || "" });
+    }
+
+    return false;
+  }
+
+  return { route, expireQueued };
+}

package/hub/store.mjs

@@ -6,7 +6,7 @@ import { DEFAULT_ORG } from "../lib/store-contract.mjs";
 
 export async function createStoreRuntime({ STORE_KIND, PG_URL, ORG_ID, DATA }) {
 function emptyState() {
-  return { messages: [], peers: {}, seq: 0, tasks: [], taskSeq: 0, projectMeta: {}, lessons: [], events: [], cardEventsBackfilled: false, aliases: {}, phaseMeta: {}, verifyGates: [], verifyGateSeq: 0, proposals: [], proposalSeq: 0, balances: { ts: 0, by: "", entries: [] }, balancesByMachine: {}, subagentCostReset: false, handoffLog: [], identities: {}, devices: [], inviteTokens: {}, focus: {}, orgPolicy: {}, overseerState: {}, instances: {}, dutySession: "", contractReap: {}, eventSeq: 0 };
+  return { messages: [], peers: {}, seq: 0, tasks: [], taskSeq: 0, projectMeta: {}, lessons: [], events: [], cardEventsBackfilled: false, aliases: {}, phaseMeta: {}, verifyGates: [], verifyGateSeq: 0, proposals: [], proposalSeq: 0, balances: { ts: 0, by: "", entries: [] }, balancesByMachine: {}, subagentCostReset: false, handoffLog: [], identities: {}, devices: [], inviteTokens: {}, focus: {}, orgPolicy: {}, overseerState: {}, instances: {}, dutySession: "", contractReap: {}, eventSeq: 0, sessionPrompts: [], sessionPromptSeq: 0 };
 }
 
 const CARD_LOG_MAX = 40;
@@ -115,6 +115,10 @@ function normalizeState(loaded = {}) {
   s.overseerState = loaded.overseerState || {};
   s.dutySession = String(loaded.dutySession || "");
   s.contractReap = loaded.contractReap && typeof loaded.contractReap === "object" ? loaded.contractReap : {};
+  // #12129: the phone's prompt queue survives a hub restart (JSON store); the id high-water mark
+  // is rebuilt from the rows so a reloaded queue never mints a colliding id.
+  s.sessionPrompts = Array.isArray(loaded.sessionPrompts) ? loaded.sessionPrompts : [];
+  s.sessionPromptSeq = Number(loaded.sessionPromptSeq || Math.max(0, ...s.sessionPrompts.map(p => Number(p.id) || 0))) || 0;
   // The event id high-water mark. Seeded from the store's own MAX(id) where it supplied one, and
   // otherwise from the largest id in the array — never from the array TAIL, which is exactly the
   // assumption that let a clobbered id mint colliding event ids and kill the log.

package/hub.mjs

@@ -18,6 +18,7 @@ import { routeCards } from "./hub/routes/cards.mjs";
 import { routeInsights } from "./hub/routes/insights.mjs";
 import { routeMessages } from "./hub/routes/messages.mjs";
 import { routeDevices } from "./hub/routes/devices.mjs";
+import { createSessionRuntime } from "./hub/routes/session.mjs";
 import { createNotifier } from "./hub/notifier.mjs";
 
 const PORT = Number(process.env.RELAY_PORT || 4477);
@@ -125,7 +126,11 @@ const context = {
   hubSend: duty.hubSend, assertNoSecrets, UI, AUTH_MODE, ONLINE_MS, PEER_TTL_MS,
   REAP_GRACE_MS, CONTRACT_WINDOW_MS,
 };
-const routes = [routeDevices, routeAdmin, routeCards, routeInsights, routeMessages];
+const sessionChat = createSessionRuntime({
+  state: store.state, markDirty: store.markDirty,
+  appendEvent: events.appendEvent, pushEventToStreams: events.pushEventToStreams,
+});
+const routes = [routeDevices, routeAdmin, routeCards, routeInsights, routeMessages, sessionChat.route];
 
 const server = http.createServer(async (req, res) => {
   const u = new URL(req.url, "http://x");

package/lib/identity.mjs

@@ -28,7 +28,7 @@ export const HDR_INST = {
 };
 
 const busDir = () => process.env.AGENT_BUS_DIR || join(homedir(), ".agent-bus");
-const keysDir = () => join(busDir(), "keys");
+export const keysDir = () => join(busDir(), "keys");
 // A name is cosmetic, but it becomes a filename, so it gets sanitised twice over: dot-runs collapse
 // FIRST (so no `..` survives even in principle), then anything outside the safe charset goes. Relying
 // on "we stripped the separators" alone would make traversal one regex edit away.
@@ -187,17 +187,17 @@ export function loadOrCreate(name, kind = "agent") {
   const f = keyPath(name);
   const tmp = `${f}.${process.pid}.${randomBytes(4).toString("hex")}.tmp`;
   try {
-    writeFileSync(tmp, JSON.stringify(id), { mode: 0o600 });
-    if (existsSync(f)) {                                     // someone beat us between check and write
-      try { unlinkSync(tmp); } catch {}                      // the loser's tmp still holds a private key
-      return load(name) || id;
+    try {
+      writeFileSync(tmp, JSON.stringify(id), { mode: 0o600 });
+      if (existsSync(f)) return load(name) || id;             // someone beat us between check and write
+      renameSync(tmp, f);
+      chmodSync(f, 0o600);
+      processCache.set(cacheKey(String(name)), id);
+      return id;
+    } finally {
+      try { unlinkSync(tmp); } catch {}                       // unconditional: no early exit leaves a tmp holding a private key
     }
-    renameSync(tmp, f);
-    chmodSync(f, 0o600);
-    processCache.set(cacheKey(String(name)), id);
-    return id;
   } catch {
-    try { unlinkSync(tmp); } catch {}
     return load(name) || id;                                  // lost the race, or unwritable — never throw
   }
 }
@@ -220,8 +220,12 @@ export function listKeyNames() {
 function writeKeyStub(f, id) {
   const stub = { name: id.name, kind: id.kind, pubkey: id.pubkey, createdAt: id.createdAt, keychain: IDENTITY_SERVICE };
   const tmp = `${f}.${process.pid}.${randomBytes(4).toString("hex")}.tmp`;
-  writeFileSync(tmp, JSON.stringify(stub), { mode: 0o600 });
-  renameSync(tmp, f);
+  try {
+    writeFileSync(tmp, JSON.stringify(stub), { mode: 0o600 });
+    renameSync(tmp, f);
+  } finally {
+    try { unlinkSync(tmp); } catch {}                         // a failed rename must not strand its tmp
+  }
   try { chmodSync(f, 0o600); } catch {}
 }
 
@@ -353,11 +357,15 @@ export function loadOrCreateInstance(durableIdentity, instanceId) {
       durablePubkey: durableIdentity.pubkey,
     };
     const tmp = `${f}.${process.pid}.${randomBytes(4).toString("hex")}.tmp`;
-    writeFileSync(tmp, JSON.stringify(inst), { mode: 0o600 });
-    if (existsSync(f)) { try { return JSON.parse(readFileSync(f, "utf8")); } catch { return inst; } }
-    renameSync(tmp, f);
-    chmodSync(f, 0o600);
-    return inst;
+    try {
+      writeFileSync(tmp, JSON.stringify(inst), { mode: 0o600 });
+      if (existsSync(f)) { try { return JSON.parse(readFileSync(f, "utf8")); } catch { return inst; } }
+      renameSync(tmp, f);
+      chmodSync(f, 0o600);
+      return inst;
+    } finally {
+      try { unlinkSync(tmp); } catch {}                       // the race loser's tmp still holds a private key — never strand it
+    }
   } catch { return null; }                                  // unwritable — caller falls back to durable
 }
 

package/lib/key-prune.mjs

@@ -1,5 +1,6 @@
-import { lstatSync, readFileSync, unlinkSync } from "node:fs";
-import { keyPath, listKeyNames } from "./identity.mjs";
+import { lstatSync, readFileSync, unlinkSync, readdirSync } from "node:fs";
+import { join } from "node:path";
+import { keyPath, listKeyNames, keysDir } from "./identity.mjs";
 
 import { isTestIdentity } from "./test-identity-names.mjs";
 export { isTestIdentity } from "./test-identity-names.mjs";
@@ -47,3 +48,30 @@ export function pruneTestKeys({ snapshots, idleDays = 7, apply = false, now = Da
   }
   return { dryRun: !apply, idleDays, removed, kept };
 }
+
+// An interrupted atomic key write strands `<name>.json.<pid>.<rand>.tmp` — and a tmp from a lost
+// race holds a full PRIVATE key. Real keys always end in `.json`, so anything ending in `.tmp` in
+// the keys tree is by construction an incomplete write; an hour past its mtime no live writer is
+// still between write and rename, so it is safe to remove (--apply only, like the rest of prune).
+export function pruneStaleTmp({ maxAgeMs = 3_600_000, apply = false, now = Date.now() } = {}) {
+  if (!Number.isFinite(maxAgeMs) || maxAgeMs <= 0) throw new Error("tmp age must be a positive number");
+  const cutoff = now - maxAgeMs;
+  const removed = [], kept = [];
+  for (const dir of [keysDir(), join(keysDir(), "instances")]) {
+    let entries = [];
+    try { entries = readdirSync(dir, { withFileTypes: true }); } catch { continue; }
+    for (const entry of entries) {
+      if (!entry.name.endsWith(".tmp")) continue;              // a real key file never matches
+      const file = join(dir, entry.name);
+      let stat;
+      try { stat = lstatSync(file); } catch { continue; }      // vanished mid-sweep
+      if (!stat.isFile() || stat.isSymbolicLink()) continue;   // never follow a link out of the keys tree
+      if (stat.mtimeMs >= cutoff) { kept.push(file); continue; }
+      if (apply) {
+        try { unlinkSync(file); } catch { kept.push(file); continue; }
+      }
+      removed.push(file);
+    }
+  }
+  return { dryRun: !apply, maxAgeMs, removed, kept };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "trantor",
-  "version": "0.18.110",
+  "version": "0.18.111",
   "type": "module",
   "bin": {
     "trantor": "bin/cli.mjs"