@@ -0,0 +1,24 @@
+---
+"@objectstack/cli": minor
+---
+
+`os serve`'s runtime state file is keyed by the PROJECT, not by the environment id alone — so two projects on one machine stop overwriting each other's supervision record.
+
+<!-- adr-0087: not-required (no-migration-prescription) Nothing authorable moves: no spec key, export, config field or stored metadata changes spelling or shape, `packages/spec` is untouched, and `objectstack migrate meta` has nothing to rewrite. What is retired is the NAME of a best-effort supervision file that `os serve` writes under the ObjectStack home — a path on disk, not a metadata surface the ledger can project into `spec-changes.json` or the generated upgrade guide. The affected party is an out-of-tree supervisor that opens that path, and the one action it takes is deleting a single stale file; there is no authored artifact for a metadata upgrader to rewrite, and no ledger entry could reach the party that is affected. -->
+
+**BREAKING** for anything that opens the runtime state file by its old name. Shipped as `minor` under the launch-window convention: while the whole workspace versions in lockstep the bump level carries no breaking-ness, so this banner and the ADR-0087 disposition above are the carriers. The file `os serve` writes under the ObjectStack home was named `runtime.<environment>.json` and is now named `runtime.<environment>.<project>.json`.
+
+`os serve` publishes `{ pid, port, url, environmentId, startedAt }` to a file under the ObjectStack home, so a supervisor can answer *"is my server running, and where?"*. That file was named `runtime.<environment>.json`, and both halves of where it lived were machine-global: `resolveObjectStackHome()` takes no arguments (it reads `OS_HOME`, else `~/.objectstack`), and an environment id is not a project identity. Two different projects on one machine, both in the ordinary `local` environment, therefore wrote one file.
+
+Driven with two real boots, two project roots and one home, that produced two failures with one cause:
+
+- project B's boot replaced project A's record, so a reader asking about A's server was answered `pid`/`port`/`url` belonging to **B** — confidently, while A's own server was still alive and still listening elsewhere;
+- project A's shutdown then deleted the file that by that point described **B**, leaving a running server with no supervision record at all.
+
+The file is now `runtime.<environment>.<project>.json`, where the project component is a sanitised basename plus a short digest of the served app's root — the same root `serve` already resolves for host-anchored package loads. The payload is unchanged: no new key, and in particular no database path (which #15374 ruled out deliberately, because it would turn a best-effort supervision file into an identity contract).
+
+**If you read this file:** a reader that hard-codes `runtime.<environment>.json` now gets `ENOENT` rather than a stale or foreign record — a loud, correct answer to "is my server running", where the old name could only give a confident wrong one. Readers that glob `runtime.*.json` inside a home they pinned themselves (as `scripts/publish-smoke.sh` does) are unaffected. A `runtime.<environment>.json` left over from an earlier version is no longer written or cleaned up by `os serve`; delete it once.
+
+**Which root the project component is taken from**, for a supervisor that has to reconstruct the name out of tree: it is the app root `serve` anchors at, which is the config file's own directory when that file exists and that directory carries a `package.json`, and the process's working directory otherwise. Two boundaries follow, stated rather than fixed: the same app served from two working directories without a manifest keys two files, and the key is the resolved path rather than the realpath, so two symlinked spellings of one project key differently — each spelling gets its own file, and each is internally consistent.
+
+Two boots of the *same* project from the *same* anchor still share one file, which is the same-project case and unchanged here.
@@ -50,6 +50,7 @@ import {
   publishBoundPort,
   resolveBoundPort,
   runtimeBoundPortChannels,
+  runtimeStateFileName,
   type BoundPortChannels,
 } from './serve.js';
 import { MAX_PORT } from '../utils/port-contract.js';
@@ -287,7 +288,12 @@ describe('#13062 all THREE channels publish that one number', () => {
       publishBoundPort(45063, runtimeBoundPortChannels(() => { /* banner not under test here */ }));
     });
 
-    const runtimeFile = join(home, 'runtime.env_local.json');
+    // ⛔ Not the literal `runtime.env_local.json`: the name is keyed by the
+    // PROJECT as well as the environment (#15733), and a literal here would be
+    // a second copy of that rule — free to keep passing against a writer that
+    // had drifted off it. `runtimeBoundPortChannels` reached outside a boot
+    // anchors on the CWD, which is this runner's.
+    const runtimeFile = join(home, runtimeStateFileName('env_local', process.cwd()));
     expect(existsSync(runtimeFile), 'no runtime state file was written at all').toBe(true);
     const state = JSON.parse(readFileSync(runtimeFile, 'utf8'));
     expect(state.port, 'the state file does not publish the bound port').toBe(45063);
 
@@ -47,6 +47,7 @@ import { isEmailTransportProvider, emailProviderRequiresApiKey, unsupportedProvi
 // bare `import()` resolved against this CLI's own realpath, so its bundled
 // copy wins, never the host's (#10909).
 import { isSmsTransportProvider, SMS_TRANSPORT_PROVIDERS } from '@objectstack/service-sms';
+import { createHash } from 'node:crypto';
 import { resolveObjectStackHome } from '@objectstack/runtime';
 import { LOG_LEVELS, resolveLogLevel, readLogLevelEnv } from '../utils/log-level.js';
 // The ONE port contract — range, reader and refusal prose — shared with the two
@@ -391,9 +392,9 @@ export function formatExhaustedPortSearchNotice(requestedPort: number, cause: un
  * `serve` used to publish the port it was ASKED for on all three of the
  * channels that ANNOUNCE an address — the `objectstack:listening` IPC message,
  * the ready banner's `API:` row (through {@link resolveAuthBaseUrl}), and
- * `runtime.<environment>.json`. For every port but one the requested and the
- * bound value coincide, which is why it stayed invisible; for `0` they CANNOT
- * coincide. `MIN_PORT = 0` is legal on purpose — `utils/port-contract.ts` says
+ * `runtime.<environment>.<project>.json`. For every port but one the requested
+ * and the bound value coincide, which is why it stayed invisible; for `0` they
+ * CANNOT coincide. `MIN_PORT = 0` is legal on purpose — `utils/port-contract.ts` says
  * so in its own words, from its own measurement, that 0 is "a REQUEST, not an
  * error", and `listen(0)` binds a kernel-assigned port. So `os serve --port 0`
  * announced `{ port: 0 }`, printed `API: http://localhost:0/` and wrote
@@ -459,8 +460,8 @@ export interface ListeningMessage {
  */
 export interface BoundPortChannels {
   /**
-   * Writes `runtime.<environment>.json`. ⛔ Must COMPLETE before either
-   * announcement below: it is the file both of them send a consumer to.
+   * Writes `runtime.<environment>.<project>.json`. ⛔ Must COMPLETE before
+   * either announcement below: it is the file both of them send a consumer to.
    */
   writeRuntimeState: (published: { port: number; url: string }) => void;
   /** Sends {@link ListeningMessage}, when an IPC channel is open. */
@@ -475,14 +476,14 @@ export interface BoundPortChannels {
  * ## The bug this shape exists to make impossible
  *
  * `os serve` announces its address on three channels: the runtime state file
- * `runtime.<environment>.json`, the `objectstack:listening` IPC message, and
- * the ready banner. Two of those are ANNOUNCEMENTS a consumer reacts to; the
- * third is the FILE those consumers then open. Published in the order they
+ * `runtime.<environment>.<project>.json`, the `objectstack:listening` IPC
+ * message, and the ready banner. Two of those are ANNOUNCEMENTS a consumer
+ * reacts to; the third is the FILE those consumers then open. Published in the order they
  * happened to be written — banner, IPC, file — every consumer that believes an
  * announcement races a file that is not there yet:
  *
  * ```text
- * banner  ─▶ a supervisor sees "ready" and opens runtime.env_local.json
+ * banner  ─▶ a supervisor sees "ready" and opens runtime.env_local.<project>.json
  * IPC     ─▶ the `os dev` parent sees the port
  * file    ─────────────────────────▶ ...written here. The ENOENT already happened.
  * ```
@@ -491,10 +492,11 @@ export interface BoundPortChannels {
  * claim (`serve-publishes-bound-port.e2e.test.ts`) is an ORDINARY consumer — it
  * waits for the banner AND the IPC message, then reads the file — and it
  * ejected 14 PRs from the shared merge queue in a rolling 24 hours (10
- * independent hits, #13158) with `ENOENT: ... runtime.env_local.json`. A real
- * supervisor written the same way loses the same race; all a loaded machine
- * does is deschedule the child between the announcement and the write, which is
- * why it read as a flake for a day.
+ * independent hits, #13158) with `ENOENT: ... runtime.env_local.json` — the
+ * name that file carried then, before {@link runtimeStateFileName} keyed it by
+ * project as well. A real supervisor written the same way loses the same race;
+ * all a loaded machine does is deschedule the child between the announcement
+ * and the write, which is why it read as a flake for a day.
  *
  * ⛔ The repair is NOT to make the reader poll. A consumer that must poll after
  * being told "ready" was told "ready" too early — polling spreads the defect
@@ -521,6 +523,92 @@ export function publishBoundPort(boundPort: number, channels: BoundPortChannels)
   channels.printBanner();
 }
 
+/**
+ * The identity a runtime state file is keyed by: the ROOT OF THE PROJECT this
+ * process is serving, folded into one filename-safe component.
+ *
+ * ## The collision this ends, DRIVEN rather than reasoned (#15733)
+ *
+ * The state file used to be `runtime.<environment>.json`, and both halves of
+ * that name are machine-global. {@link resolveObjectStackHome} takes NO
+ * arguments — it reads `OS_HOME`, else `~/.objectstack` — and an environment
+ * id is not a project identity, so two different projects on one machine, both
+ * in the ordinary `local` environment, wrote ONE file. Two real boots, two
+ * project roots, one home:
+ *
+ * ```text
+ * project A boots  → runtime.env_local.json = { pid 9301, port 42693 }
+ * project B boots  → runtime.env_local.json = { pid 9345, port 46175 }
+ * A asks "is my server running, and where?"
+ *                  → answered pid 9345 on :46175 — which is B — while A's own
+ *                    server is still alive and still listening on 42693.
+ * A shuts down     → its exit handler deletes the file that by then described
+ *                    B, so B serves on with no supervision record at all.
+ * ```
+ *
+ * Those are one defect seen twice: the NAME carried no answer to *whose*
+ * server this is, so every project addressed the same file. Keying it by the
+ * served app's root gives each project its own — a reader that finds a file
+ * has found its own, and a process cleaning up on exit removes only its own.
+ *
+ * ⛔ NOT repaired by putting an identity in the PAYLOAD instead. #15374 ruled
+ * that one deliberately: the identity of the file a process is serving is a
+ * property of THAT PROCESS, and a payload key would turn a best-effort
+ * supervision file into an identity contract while STILL leaving two projects
+ * overwriting and deleting each other's records. The key is the defect; the
+ * payload is unchanged here.
+ *
+ * ## Why the name carries a readable half as well as a digest
+ *
+ * The digest is what makes the name unique; the slug is what makes a home
+ * directory legible to whoever is standing in front of it. A directory
+ * answering `runtime.env_local.<12 hex>.json` twice tells a reader nothing
+ * about which one is theirs, and answering that question is this file's whole
+ * job. The slug is DERIVED, never trusted: uniqueness rests on the digest
+ * alone, so a project root whose basename sanitises away to nothing is still
+ * keyed correctly — it just reads as the digest.
+ *
+ * ⚠️ Boundary, stated rather than fixed: the key is the resolved path, not the
+ * realpath, so two symlinked spellings of ONE project key differently (each
+ * spelling gets its own file, and each is internally consistent). Two boots of
+ * the SAME project also still share a file — that is the same-project case,
+ * which is #15374's in-process watch, not this one.
+ *
+ * ⚠️ Second boundary, and the one an OUT-OF-TREE reader has to replicate to
+ * find the record: WHICH root this is handed is {@link servedAppRootOrCwd},
+ * which {@link anchorServedApp} sets to the CONFIG'S OWN DIRECTORY only when
+ * that config exists and that directory carries a `package.json`, and to
+ * `process.cwd()` otherwise. So one app served from two working directories
+ * with no manifest beside its config keys TWO files, and a supervisor that
+ * reconstructs the name out of tree has to apply that same rule rather than
+ * assume the config's directory. That fallback is #11185's and is deliberate:
+ * a directory that declares nothing is not anchored at, because anchoring
+ * there could only turn a working boot into an `undeclared` refusal.
+ */
+export function projectStateKey(servedAppRoot: string): string {
+  const absolute = path.resolve(servedAppRoot);
+  const digest = createHash('sha256').update(absolute).digest('hex').slice(0, 12);
+  const slug = path.basename(absolute).toLowerCase().replace(/[^a-z0-9]+/g, '-').replace(/^-+|-+$/g, '').slice(0, 24);
+  return slug.length > 0 ? `${slug}-${digest}` : digest;
+}
+
+/**
+ * The runtime state file's name: `runtime.<environment>.<project>.json`.
+ *
+ * Both components are identities the file has to carry, and neither is
+ * sufficient alone — the environment id keeps a staging boot from answering
+ * for a local one, and {@link projectStateKey} keeps ANOTHER PROJECT's local
+ * boot from answering for this one.
+ *
+ * ⚠️ A reader that hard-codes the old `runtime.<environment>.json` now gets
+ * ENOENT rather than a stale or foreign record. That is the intended trade: a
+ * missing file is a loud, correct answer to "is my server running", where the
+ * name it replaces could only give a confident wrong one.
+ */
+export function runtimeStateFileName(environmentId: string, servedAppRoot: string): string {
+  return `runtime.${environmentId}.${projectStateKey(servedAppRoot)}.json`;
+}
+
 /**
  * The real channels: the same three writes this command has always done, with
  * their failure handling unchanged.
@@ -534,7 +622,15 @@ export function runtimeBoundPortChannels(printBanner: () => void): BoundPortChan
     writeRuntimeState: ({ port, url }) => {
       try {
         const environmentId = process.env.OS_ENVIRONMENT_ID ?? 'env_local';
-        const runtimeFile = path.join(resolveObjectStackHome(), `runtime.${environmentId}.json`);
+        // Keyed by the SERVED APP'S ROOT as well as the environment, so two
+        // projects on one machine stop addressing one file (#15733). The root
+        // is the one this command already anchored for host resolution; a
+        // caller reaching these channels outside a boot gets the CWD, which is
+        // what every path in that situation already resolves against.
+        const runtimeFile = path.join(
+          resolveObjectStackHome(),
+          runtimeStateFileName(environmentId, servedAppRootOrCwd()),
+        );
         fs.mkdirSync(path.dirname(runtimeFile), { recursive: true });
         fs.writeFileSync(runtimeFile, JSON.stringify({
           pid: process.pid,
@@ -4657,9 +4753,9 @@ export default class Serve extends Command {
       // ── The port this process ACTUALLY bound (#13062) ─────────────
       // Read ONCE, here, and handed to every channel that announces an address:
       // the ready banner below, the `objectstack:listening` IPC message and
-      // `runtime.<environment>.json`. Those three were three outputs of ONE
-      // number, and that number was the port that had been REQUESTED — equal to
-      // the bound one for every value except the one where it can never be
+      // `runtime.<environment>.<project>.json`. Those three were three outputs
+      // of ONE number, and that number was the port that had been REQUESTED
+      // — equal to the bound one for every value except the one where it can never be
       // (`--port 0`), which is how all three came to announce `localhost:0`
       // with nothing erroring.
       //
 
@@ -0,0 +1,64 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * ONE project publishing its runtime state file, as a REAL separate process.
+ *
+ * `serve-runtime-state-project-key.test.ts` drives several of these at once to
+ * ask a question a single process cannot answer: what happens on a machine
+ * where TWO projects are serving. Two things about that are only observable
+ * across processes —
+ *
+ *   • each child has its OWN pid, so "the second boot's record replaced the
+ *     first's" is a reading rather than an artefact of one process writing
+ *     twice; and
+ *   • `runtimeBoundPortChannels` registers an `exit` cleanup per file it
+ *     writes, so "whose record does a shutdown delete" needs a process that
+ *     can shut down while another keeps running.
+ *
+ * ⛔ It writes through the REAL channel — `runtimeBoundPortChannels`, the
+ * object `os serve` itself publishes with — reached by a RELATIVE import into
+ * `packages/cli/src`, never through the package name. A bare
+ * `@objectstack/cli` specifier would resolve through `exports` to `dist/` and
+ * turn every verdict below into a statement about build state.
+ *
+ * ⛔ It does NOT boot a server. What is under test is the NAME the state file
+ * is keyed by; a listening socket adds cost and a port race to a question that
+ * has neither.
+ *
+ * Protocol: argv[2] is the port to publish, argv[3] a fixed project root the
+ * child names a file for so the caller can compare that naming with its own.
+ * One JSON line is written to stdout once the file is on disk. The child then
+ * stays alive until its stdin closes, and exits CLEANLY on that — a signal
+ * would skip the `exit` listener whose behaviour is half of what the caller is
+ * measuring.
+ *
+ * ⛔ This module EXPORTS nothing and is never imported: its body runs on
+ * import, so an `export` here would invite a caller to pull the constant in and
+ * take a `process.exit(2)` with it. Both arguments come over argv instead.
+ */
+
+import { runtimeBoundPortChannels, runtimeStateFileName } from '../../src/commands/serve.js';
+
+const port = Number(process.argv[2]);
+if (!Number.isInteger(port) || port <= 0) {
+  process.stderr.write('runtime-state-child: argv[2] must be a positive integer port\n');
+  process.exit(2);
+}
+
+runtimeBoundPortChannels(() => { /* no banner: this child publishes one channel */ })
+  .writeRuntimeState({ port, url: `http://localhost:${port}` });
+
+process.stdout.write(`${JSON.stringify({
+  pid: process.pid,
+  port,
+  cwd: process.cwd(),
+  home: process.env.OS_HOME,
+  // The name this child's OWN copy of the writer produces, for a fixed root.
+  // The caller recomputes it from its own import and compares: two independent
+  // resolutions of the same source agreeing is what rules out a stale build
+  // answering for either side.
+  namingControl: runtimeStateFileName('env_local', process.argv[3] ?? ''),
+})}\n`);
+
+process.stdin.resume();
+process.stdin.on('end', () => { process.exit(0); });