feat(codex-local): give each Codex account its own home and path secret (#12709)

## Thinking Path

> - Paperclip is the control plane for companies that use AI agents for
work
> - Local adapters connect Paperclip agents to provider command line
tools
> - The Codex adapter stores login data in a shared company home
> - A shared home cannot keep credentials for more than one Codex
account
> - This pull request gives each account a safe home and a matching
company secret
> - The benefit is that one company can use multiple Codex accounts at
the same time

## Linked Issues or Issue Description

**Problem or motivation**

A company can hold only one Codex subscription credential because device
login uses one shared home. A second account cannot log in without
replacing or conflicting with the first credential.

**Proposed solution**

This change validates the vendor account identifier, stores each
credential in its own home, and creates a company secret that points to
that home. Repeat login calls return success when the matching secret
already exists.

**Roadmap alignment**

The change supports the roadmap goal for centrally managed secrets with
scoped access and audited resolution.

**Additional context**

The security review returned approve with no blocking finding. The
branch adds shared account-handle validation and tests for device login
and the Codex local adapter.

## What Changed

- Add strict allowlist validation for Codex account handles.
- Store each Codex account credential in a separate home under the Codex
cache root.
- Verify that the resolved account home stays inside the cache root.
- Create the `CODEX_HOME_<handle>` company secret for each account.
- Keep repeat and concurrent login calls safe and idempotent.
- Add shared helper and route, adapter, and validation tests.

## Verification

- `pnpm --filter @paperclipai/adapter-codex-local test` passes with 343
tests.
- `pnpm --filter @paperclipai/server test
src/__tests__/agent-device-login-routes.test.ts` passes with 25 tests.
- The adapter suite passes with 23 tests.
- The shared package and Codex adapter typechecks pass.
- Continuous integration must pass on every check before merge.

## Risks

The account handle becomes part of a directory path and secret name. The
strict allowlist and root containment check reduce path traversal risk.
Existing single-account homes remain unchanged unless a new device login
creates an account-specific home.

## Model Used

OpenAI GPT-5 (exact runtime model ID: gpt-5), with tool use and code
execution. The runtime context window is not exposed in this run.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>
This commit is contained in:
Nicky Leach 2026-09-02 14:46:53 -07:00 committed by GitHub
parent dfdfc8664e
commit 9064cfd09e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
15 changed files with 3261 additions and 711 deletions

View File

@ -23,8 +23,8 @@ const TOKEN_SENTINEL = "SENTINEL_TOKEN_XYZ";
// This suite proves the device-login credential promotion helper. The helper
// runs an independent readiness check on the exact staged credential first, then
// validates it with the export rules, then writes only the company-scoped
// credential home and the company-scoped cache. It writes only while the session
// validates it with the export rules, then writes this account's own home and,
// as a fallback, the company default home. It writes only while the session
// holds the sole active claim on the slot (the conditional check). It never
// writes the instance-global host, and it never logs secret bytes.
describe("device-login credential promotion", () => {
@ -83,11 +83,7 @@ describe("device-login credential promotion", () => {
return path.join(resolveManagedCodexHomeDir(env, companyId), "auth.json");
}
async function readIfPresent(target: string): Promise<string | null> {
return readFile(target, "utf8").catch(() => null);
}
it("a failed readiness check rejects and writes neither the company home nor the cache", async () => {
it("a failed readiness check rejects and writes neither the company home nor the account home", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const logs: string[] = [];
@ -119,10 +115,10 @@ describe("device-login credential promotion", () => {
expect(logs.join("\n")).not.toContain(ACCOUNT);
});
it("a user login seeds an empty company home and cache slot for the new identity", async () => {
it("a user login seeds this account's own home and the company default home", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
@ -131,22 +127,22 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
expect(outcome).toBe("promoted");
expect(result.outcome).toBe("promoted");
const homeAuth = await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8");
expect(JSON.parse(homeAuth).tokens.account_id).toBe(ACCOUNT);
const cacheAuth = await readFile(
const accountAuth = await readFile(
resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A),
"utf8",
);
expect(JSON.parse(cacheAuth).tokens.account_id).toBe(ACCOUNT);
expect(JSON.parse(accountAuth).tokens.account_id).toBe(ACCOUNT);
// The instance-global host was never seeded.
await expect(
lstat(path.join(resolveSharedCodexHomeDir(env), "auth.json")),
).rejects.toThrow();
});
it("a strictly-newer same-identity login updates the company home", async () => {
it("a strictly-newer login updates this account's own home", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
await promoteDeviceLoginCredential({
@ -158,7 +154,7 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER, marker: "new" }),
companyId: COMPANY_A,
userInitiated: true,
@ -167,13 +163,15 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
expect(outcome).toBe("promoted");
const homeAuth = JSON.parse(await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8"));
expect(homeAuth.last_refresh).toBe(NEWER);
expect(homeAuth.tokens.refresh_token).toContain("new");
expect(result.outcome).toBe("promoted");
const accountAuth = JSON.parse(
await readFile(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A), "utf8"),
);
expect(accountAuth.last_refresh).toBe(NEWER);
expect(accountAuth.tokens.refresh_token).toContain("new");
});
it("an older same-identity login keeps the company home", async () => {
it("an older login keeps this account's own home", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
await promoteDeviceLoginCredential({
@ -185,7 +183,7 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: OLDER, marker: "older" }),
companyId: COMPANY_A,
userInitiated: true,
@ -194,12 +192,14 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
expect(outcome).toBe("kept");
const homeAuth = JSON.parse(await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8"));
expect(homeAuth.tokens.refresh_token).toContain("keep");
expect(result.outcome).toBe("kept");
const accountAuth = JSON.parse(
await readFile(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A), "utf8"),
);
expect(accountAuth.tokens.refresh_token).toContain("keep");
});
it("a different-identity login keeps the home and reports a foreign-identity outcome", async () => {
it("promotion writes the account home for a second, different account", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
await promoteDeviceLoginCredential({
@ -211,7 +211,7 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: OTHER_ACCOUNT, lastRefresh: NEWER, marker: "other" }),
companyId: COMPANY_A,
userInitiated: true,
@ -220,23 +220,19 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
// The home keeps the first identity. The login did not install the other
// account, so the outcome is `kept_foreign_identity`, not a plain `kept`: the
// caller must fail the session instead of a report of `authenticated`. The
// other identity still lands in its own per-identity cache slot.
expect(outcome).toBe("kept_foreign_identity");
const homeAuth = JSON.parse(await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8"));
expect(homeAuth.tokens.account_id).toBe(ACCOUNT);
const otherCache = JSON.parse(
// A second, different account gets its own home, so the second login
// succeeds instead of failing behind the first account's home.
expect(result.outcome).toBe("promoted");
const otherAccountAuth = JSON.parse(
await readFile(resolveCodexAuthCacheEntryPath(env, OTHER_ACCOUNT, COMPANY_A), "utf8"),
);
expect(otherCache.tokens.account_id).toBe(OTHER_ACCOUNT);
expect(otherAccountAuth.tokens.account_id).toBe(OTHER_ACCOUNT);
});
it("the cache off-switch skips the cache slot but still seeds the company home", async () => {
it("promotion returns the account id and the account home path", async () => {
const home = await makeInstanceRoot();
const env = envFor(home, { PAPERCLIP_CODEX_AUTH_CACHE: "off" });
const outcome = await promoteDeviceLoginCredential({
const env = envFor(home);
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
@ -245,29 +241,231 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
expect(outcome).toBe("promoted");
// The company home was seeded.
expect(await readIfPresent(companyHomeAuthPath(env, COMPANY_A))).toContain(ACCOUNT);
// The cache slot was not written.
await expect(
lstat(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A)),
).rejects.toThrow();
expect(result.accountId).toBe(ACCOUNT);
expect(result.accountHomeDir).toBe(
path.dirname(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A)),
);
});
it("a cache write failure keeps the promotion successful and the company home durable", async () => {
it("promotion reports accountHomeCreated true for a first login of an account", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
// Force the per-identity cache write to fail. Plant a regular file where the
// cache expects the identity directory, so the private-directory guard throws
// before the cache slot is written. The company home write runs first, so the
// credential is already durable when the cache write fails.
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
expect(result.outcome).toBe("promoted");
expect(result.accountHomeCreated).toBe(true);
});
it("promotion reports accountHomeCreated false for a repeat login of the same account", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: OLDER, marker: "first" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER, marker: "second" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
expect(result.outcome).toBe("promoted");
expect(result.accountHomeCreated).toBe(false);
});
it("two concurrent promotions for the same account report only one accountHomeCreated true", async () => {
// Two different logins for the SAME Codex account run their own
// promotion slot, so they can promote at the same time. Without a lock
// around the absence check and the directory creation, both calls could
// see the directory as absent and both report `created: true`; a caller
// that later deletes the directory on `created: true` would then delete
// a home the other call's login still uses.
const home = await makeInstanceRoot();
const env = envFor(home);
const [first, second] = await Promise.all([
promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER, marker: "racer-a" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
}),
promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER, marker: "racer-b" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
}),
]);
// Both racers still authenticate; only the timestamp-comparison outcome
// (`promoted` vs. `kept`) can differ, because they carry the same
// `lastRefresh` timestamp.
expect(["promoted", "kept"]).toContain(first.outcome);
expect(["promoted", "kept"]).toContain(second.outcome);
const createdFlags = [first.accountHomeCreated, second.accountHomeCreated];
expect(createdFlags.filter(Boolean)).toHaveLength(1);
await expect(
lstat(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A)),
).resolves.toBeDefined();
});
it("promotion keeps the company default home when it holds another account", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER, marker: "first" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
const before = await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8");
await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: OTHER_ACCOUNT, lastRefresh: NEWER, marker: "other" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
// The company default home fallback still names the first account: a
// second account never clobbers it once some account has claimed it.
expect(await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8")).toBe(before);
});
it("promotion seeds the company default home when it holds no usable credential", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
expect(result.outcome).toBe("promoted");
const homeAuth = await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8");
expect(JSON.parse(homeAuth).tokens.account_id).toBe(ACCOUNT);
});
it("promotion fails the login when the account identifier cannot become a handle", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
await expect(
promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: "acct 42", lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
}),
).rejects.toThrow();
await expect(lstat(companyHomeAuthPath(env, COMPANY_A))).rejects.toThrow();
});
it("a whitespace-padded account identifier never shares a home with, or reports as authenticated against, the unpadded account", async () => {
// A distinct identifier with surrounding whitespace must reject, not
// alias onto the unpadded account's home. Promote the unpadded account
// first, then attempt a second login whose identifier differs only by a
// leading space. The second login must fail, and it must never read or
// report a "kept" outcome against the first account's home.
const home = await makeInstanceRoot();
const env = envFor(home);
await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER, marker: "original" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
});
await expect(
promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ` ${ACCOUNT}`, lastRefresh: NEWER, marker: "padded" }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
}),
).rejects.toThrow();
// The unpadded account's home still holds only the credential the first
// login wrote. The padded login never read it, never wrote it, and never
// received a "kept" outcome that would report it as authenticated.
const accountAuth = JSON.parse(
await readFile(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A), "utf8"),
);
expect(accountAuth.tokens.refresh_token).toContain("original");
});
it("a broken account-home directory fails the whole login", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
// Force this account's own home write to fail. Plant a regular file where
// the account expects its own directory, so the private-directory guard
// throws before any write. That write is now fail-loud (it is the durable
// result), so the whole login fails, and the best-effort company default
// home fallback is never reached.
const cacheDir = resolveCodexAuthCacheDir(env, COMPANY_A);
await mkdir(cacheDir, { recursive: true, mode: 0o700 });
const entryPath = resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A);
await writeFile(path.dirname(entryPath), "not-a-directory");
await expect(
promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
checkReadiness: ready,
isSoleActiveOwner: soleOwner,
env,
log: noopLog,
}),
).rejects.toThrow();
await expect(lstat(companyHomeAuthPath(env, COMPANY_A))).rejects.toThrow();
});
it("a company default home write failure keeps the promotion successful and this account's own home durable", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
// Force the company default home write to fail. Plant a regular file where
// the company home directory belongs, so mkdir fails. This account's own
// home write already ran and is durable, so the login still succeeds.
const companyHome = resolveManagedCodexHomeDir(env, COMPANY_A);
await mkdir(path.dirname(companyHome), { recursive: true });
await writeFile(companyHome, "not-a-directory");
const logs: string[] = [];
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
@ -279,14 +477,13 @@ describe("device-login credential promotion", () => {
},
});
// The company home holds a usable credential, so the promotion still reports
// success rather than a failed login.
expect(outcome).toBe("promoted");
const homeAuth = JSON.parse(await readFile(companyHomeAuthPath(env, COMPANY_A), "utf8"));
expect(homeAuth.tokens.account_id).toBe(ACCOUNT);
// The cache failure is observable and carries no secret bytes.
expect(result.outcome).toBe("promoted");
const accountAuth = JSON.parse(
await readFile(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A), "utf8"),
);
expect(accountAuth.tokens.account_id).toBe(ACCOUNT);
const haystack = logs.join("\n");
expect(haystack).toContain("the per-identity cache write failed");
expect(haystack).toContain("seeding the company default home failed");
expect(haystack).not.toContain(TOKEN_SENTINEL);
expect(haystack).not.toContain(ACCOUNT);
});
@ -324,7 +521,7 @@ describe("device-login credential promotion", () => {
it("a promotion whose session is no longer the sole active owner writes nothing", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
userInitiated: true,
@ -334,17 +531,17 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
expect(outcome).toBe("not_sole_owner");
expect(result.outcome).toBe("not_sole_owner");
await expect(lstat(companyHomeAuthPath(env, COMPANY_A))).rejects.toThrow();
await expect(
lstat(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A)),
).rejects.toThrow();
});
it("an automatic background login never seeds an empty company slot", async () => {
it("an automatic background login never seeds an empty home", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const outcome = await promoteDeviceLoginCredential({
const result = await promoteDeviceLoginCredential({
authBytes: subscriptionAuth({ accountId: ACCOUNT, lastRefresh: NEWER }),
companyId: COMPANY_A,
// Not a user-initiated login.
@ -354,7 +551,7 @@ describe("device-login credential promotion", () => {
env,
log: noopLog,
});
expect(outcome).toBe("background_skipped");
expect(result.outcome).toBe("background_skipped");
await expect(lstat(companyHomeAuthPath(env, COMPANY_A))).rejects.toThrow();
await expect(
lstat(resolveCodexAuthCacheEntryPath(env, ACCOUNT, COMPANY_A)),

View File

@ -1,11 +1,10 @@
import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
import os from "node:os";
import path from "node:path";
import { toAccountHandle } from "@paperclipai/shared";
import {
ensureCodexAuthCacheEntryDir,
isCodexAuthCacheEnabled,
ensureCodexAuthCacheEntryDirExclusive,
readSubscriptionAccountId,
writeCodexAuthCacheEntry,
} from "./codex-auth-cache.js";
import { writeCredentialSeedOrNewer } from "./codex-auth-seed-write.js";
import { codexHomeHasUsableAuth, resolveManagedCodexHomeDir } from "./codex-home.js";
@ -18,13 +17,18 @@ import { assertUsableSubscriptionShape } from "./device-login-export.js";
//
// The helper never writes the instance-global host (`CODEX_HOME` or `~/.codex`),
// because that source has no `companyId`. It writes two company-scoped targets:
// the company credential home and the company per-identity cache slot. The
// company home write is the first-login host rule: the cache vend never seeds an
// empty host, so a first login must seed the company home directly.
// this account's own home, and — only as a fallback for an agent with no bound
// secret — the company default home.
//
// This account's own home is the durable result of a login: the write is fail
// loud, and the caller names it with a company secret so any agent can bind to
// it. The company default home is a best-effort fallback: it is seeded only the
// first time any account logs in for the company (while it holds no usable
// credential yet), and a write failure there never fails the login.
//
// Two decisions gate the write:
// - Decision C: only a user-initiated login seeds the company slot. An
// automatic background path never seeds an empty slot.
// - Decision C: only a user-initiated login seeds a home. An automatic
// background path never seeds one.
// - Decision H: the helper writes only while the session still holds the sole
// active claim on `(company_id, adapter_type)`. This is defense in depth over
// the partial unique index, so a second session cannot race the same slot.
@ -93,30 +97,45 @@ export async function checkStagedCredentialReadiness(
/**
* The promotion outcome.
*
* - `promoted`: the helper wrote the company home (a seed or a strictly-newer
* same-identity update). The helper then tries the per-identity cache slot when
* the cache is on. The cache write is best-effort: a cache failure keeps the
* `promoted` outcome, because the company home is already durable.
* - `kept`: the login carried the SAME identity as the company home, and the home
* already held a newer same-identity credential, so the home was kept. This is a
* successful authentication: a later run vends and uses the same account. The
* helper still tries the best-effort cache write.
* - `kept_foreign_identity`: the login carried a DIFFERENT identity than the
* company home. The helper never clobbers an occupied home, so it kept the other
* account and installed nothing durable for this login. The identity-anchored
* vend reads the home identity first, so a later run can never select this
* login. This is NOT a successful authentication; the caller must fail the
* session instead of a report of `authenticated`.
* - `promoted`: the helper wrote this account's own home (a seed for a first
* login of this account, or a strictly-newer update for a repeat login). The
* helper also seeds the company default home when it holds no usable
* credential yet.
* - `kept`: the login carried a credential that is not newer than what this
* account's own home already holds, so the home was kept as-is. This is
* still a successful authentication: a later run reads the same home.
* - `not_sole_owner`: Decision H rejected the write. Nothing was written.
* - `background_skipped`: Decision C rejected the write (an automatic background
* path never seeds a company slot). Nothing was written.
* path never seeds a home). Nothing was written.
*/
export type PromoteDeviceLoginCredentialOutcome =
| "promoted"
| "kept"
| "kept_foreign_identity"
| "not_sole_owner"
| "background_skipped";
export type PromoteDeviceLoginCredentialOutcome = "promoted" | "kept" | "not_sole_owner" | "background_skipped";
/**
* The promotion result. `accountId` and `accountHomeDir` are set once the
* credential's identity is known every outcome below the handle-validation
* step carries `accountId`; only `promoted` and `kept` reach the write step and
* carry `accountHomeDir`.
*
* `accountHomeCreated` is true only when this account's own home directory was
* absent right before this call created it. The check and the creation run
* under one lock, so a concurrent promotion for the same account (a second
* login for the same Codex account) never also reports `created: true` for a
* directory the first call already made. A caller that must undo a later
* failure (for example, a failed secret write) should remove the directory
* only when `accountHomeCreated` is true, and should still re-check that no
* company secret now names this account's home before it deletes: the
* absence of a company secret at the time this call started is not proof
* that this login is the only login that used the directory. A user can also
* delete the secret and keep the account home, so a repeat login then reads
* no secret. `accountHomeCreated` is false for every outcome that does not
* reach the write step (`not_sole_owner`, `background_skipped`).
*/
export interface PromoteDeviceLoginCredentialResult {
outcome: PromoteDeviceLoginCredentialOutcome;
accountId: string | null;
accountHomeDir: string | null;
accountHomeCreated: boolean;
}
export interface PromoteDeviceLoginCredentialInput {
/** The exact staged credential bytes the login sandbox produced. */
@ -167,13 +186,14 @@ function requireSafeCompanyId(companyId: string): string {
/**
* Promotes a device-login credential into the company scope. The order is fixed:
* readiness check, credential validation, Decision C, Decision H, then the writes.
* The readiness check and the writes run while the caller still holds the active
* claim, so a second session cannot race the same slot.
* readiness check, credential validation, account-handle validation, Decision C,
* Decision H, then the writes. The readiness check and the writes run while the
* caller still holds the active claim, so a second session cannot race the same
* slot.
*/
export async function promoteDeviceLoginCredential(
input: PromoteDeviceLoginCredentialInput,
): Promise<PromoteDeviceLoginCredentialOutcome> {
): Promise<PromoteDeviceLoginCredentialResult> {
const { authBytes, userInitiated, checkReadiness, isSoleActiveOwner, log } = input;
const env = input.env ?? process.env;
const companyId = requireSafeCompanyId(input.companyId);
@ -191,84 +211,92 @@ export async function promoteDeviceLoginCredential(
const accountId = readSubscriptionAccountId(authBytes);
if (!accountId) {
// The shape gate above already guarantees a subscription identity; this guard
// keeps the account_id non-null for the cache key without a non-null cast.
// keeps the account_id non-null for the handle conversion without a non-null
// cast.
throw new Error("device-login promotion: the credential has no subscription identity");
}
// 3. Decision C: only a user-initiated login seeds the company slot.
// 2b. Convert the identity into a safe account handle. The handle names both
// this account's own home directory and its company secret, so a login
// whose identity cannot form one must fail before any write.
const accountHandle = toAccountHandle(accountId);
if (!accountHandle) {
throw new Error("device-login promotion: the account identifier cannot form a valid account handle");
}
// 3. Decision C: only a user-initiated login seeds a home.
if (!userInitiated) {
await log("[paperclip] Codex device-login promotion: skipped (an automatic background login never seeds a company slot).");
return "background_skipped";
await log("[paperclip] Codex device-login promotion: skipped (an automatic background login never seeds a home).");
return { outcome: "background_skipped", accountId, accountHomeDir: null, accountHomeCreated: false };
}
// 4. Decision H: write only while the session still owns the active slot.
const soleOwner = await isSoleActiveOwner();
if (!soleOwner) {
await log("[paperclip] Codex device-login promotion: skipped (the session no longer holds the sole active claim on the slot).");
return "not_sole_owner";
return { outcome: "not_sole_owner", accountId, accountHomeDir: null, accountHomeCreated: false };
}
// 5a. First-login host rule: seed or update the company credential home. The
// shared writer seeds an empty home and applies a strictly-newer
// same-identity update; it keeps a newer same-identity or a different
// identity. It never touches the instance-global host.
const companyHome = resolveManagedCodexHomeDir(env, companyId);
await mkdir(companyHome, { recursive: true, mode: PRIVATE_DIR_MODE });
const companyHomeAuthPath = path.join(companyHome, AUTH_FILE_NAME);
const homeOutcome = await writeCredentialSeedOrNewer({
// 5a. This account's own home is the durable result of a login: each account
// handle addresses exactly one home, so this write can never collide with
// a different identity, and a write failure here fails the whole
// promotion (fail loud, unlike the company default home fallback below).
// Two different logins can promote the same account at the same time
// (each login owns its own promotion slot), so the absence check and the
// directory creation run inside one lock: only the caller that truly
// finds the directory absent gets `created: true`. A plain
// check-then-create sequence here would let two concurrent callers for
// the same account both see the directory as absent and both believe
// they created it.
const { entryPath: accountHomeAuthPath, created: accountHomeCreated } =
await ensureCodexAuthCacheEntryDirExclusive(env, accountHandle, companyId);
const accountHomeDir = path.dirname(accountHomeAuthPath);
const accountHomeOutcome = await writeCredentialSeedOrNewer({
sourceBytes: authBytes,
destinationPath: companyHomeAuthPath,
destinationPath: accountHomeAuthPath,
seedIfDestAbsent: true,
log,
writtenLine: "[paperclip] Codex device-login promotion: wrote the company credential home at mode 0600.",
keptLine: "[paperclip] Codex device-login promotion: kept the company credential home (the login is not a seed or a strictly-newer same-identity credential).",
tempPrefix: "auth.json.promotion-home",
writtenLine: "[paperclip] Codex device-login promotion: wrote this account's own home at mode 0600.",
keptLine: "[paperclip] Codex device-login promotion: kept this account's own home (the login is not a seed or a strictly-newer credential).",
tempPrefix: "auth.json.promotion-account-home",
errorLabel: "codex device-login promotion",
env,
});
// A kept home has two very different meanings. The writer keeps the home when
// it already holds a newer SAME-identity credential (a genuine success: a later
// run vends and uses the same account), and it also keeps the home when the home
// holds a DIFFERENT identity (the writer never clobbers an occupied home). The
// second case installed nothing durable for this login: the home still holds the
// other account, and the identity-anchored vend reads the home identity first,
// so it can never select this login. Compare the login identity with the kept
// home identity, so the caller can fail the session instead of a report of
// `authenticated`. A home that this helper cannot read as the same identity is
// treated as a foreign identity; the caller fails closed.
let foreignIdentityKeep = false;
if (homeOutcome === "kept") {
const homeBytes = await readFile(companyHomeAuthPath).catch(() => null);
const homeAccountId = homeBytes ? readSubscriptionAccountId(homeBytes) : null;
foreignIdentityKeep = homeAccountId !== accountId;
}
// 5b. Record the credential in its per-identity company cache slot, so a later
// run can vend a strictly-newer copy. The cache write respects the
// off-switch. It is company-scoped and per identity, so it never crosses a
// company boundary and never clobbers a different identity.
//
// The company home write above already made the credential durable, so a
// later run authenticates from the home even when the cache slot is absent.
// The cache is only a vend optimization. So a cache write failure (a
// permission error, a full disk, or a lock timeout) must not fail the
// promotion, or the operator sees a failed login for a credential that is
// already usable. Log the failure and keep the home outcome; a later run
// re-seeds the cache slot from the durable home.
if (isCodexAuthCacheEnabled(env)) {
// 5b. Company default home fallback, for an agent with no bound secret. Seed
// it only the first time any account logs in for the company, i.e. only
// while it holds no usable credential yet; a login for a second account
// must never touch it once some account has claimed it. This write is
// best-effort: this account's own home above is already durable, so a
// failure here (a permission error, a full disk, a lock timeout) must not
// fail the promotion.
const companyHome = resolveManagedCodexHomeDir(env, companyId);
if (!(await codexHomeHasUsableAuth(companyHome))) {
try {
const cacheEntryPath = await ensureCodexAuthCacheEntryDir(env, accountId, companyId);
await writeCodexAuthCacheEntry({ sandboxAuthBytes: authBytes, cacheEntryPath, log, env });
await mkdir(companyHome, { recursive: true, mode: PRIVATE_DIR_MODE });
const companyHomeAuthPath = path.join(companyHome, AUTH_FILE_NAME);
await writeCredentialSeedOrNewer({
sourceBytes: authBytes,
destinationPath: companyHomeAuthPath,
seedIfDestAbsent: true,
log,
writtenLine: "[paperclip] Codex device-login promotion: seeded the company default home.",
keptLine: "[paperclip] Codex device-login promotion: kept the company default home.",
tempPrefix: "auth.json.promotion-home",
errorLabel: "codex device-login promotion",
env,
});
} catch {
await log(
"[paperclip] Codex device-login promotion: the per-identity cache write failed; the company credential home is durable, so the login stays successful.",
"[paperclip] Codex device-login promotion: seeding the company default home failed; this account's own home is durable, so the login stays successful.",
);
}
}
if (homeOutcome === "written") {
return "promoted";
}
return foreignIdentityKeep ? "kept_foreign_identity" : "kept";
return {
outcome: accountHomeOutcome === "written" ? "promoted" : "kept",
accountId,
accountHomeDir,
accountHomeCreated,
};
}

View File

@ -4,6 +4,7 @@ import path from "node:path";
import { afterEach, describe, expect, it } from "vitest";
import {
assertAccountHomeCacheDirStillValid,
clearCodexAuthCache,
clearCodexAuthCacheEntry,
ensureCodexAuthCacheEntryDir,
@ -12,6 +13,8 @@ import {
resolveCodexAuthCacheEntryPath,
selectVendCredential,
toCacheKey,
withAccountHomeSecretMutationLock,
withCodexAccountHomePromotionLock,
} from "./codex-auth-cache.js";
import { resolveSharedCodexHomeDir } from "./codex-home.js";
@ -127,6 +130,17 @@ describe("codex auth cache store", () => {
);
});
it("resolveCodexAuthCacheEntryPath rejects an identifier that toAccountHandle rejects", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
// A space, a plus sign, and a leading hyphen all fail the stricter
// account-handle allowlist, even though the older denylist alone would
// accept some of them.
expect(() => resolveCodexAuthCacheEntryPath(env, "acct 42", "company-a")).toThrow();
expect(() => resolveCodexAuthCacheEntryPath(env, "acct+42", "company-a")).toThrow();
expect(() => resolveCodexAuthCacheEntryPath(env, "-rf", "company-a")).toThrow();
});
it("resolveCodexAuthCacheEntryPath verifies the resolved path stays under the cache root", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
@ -335,4 +349,215 @@ describe("codex auth cache store", () => {
expect(isCodexAuthCacheEnabled({ PAPERCLIP_CODEX_AUTH_CACHE: "no" })).toBe(false);
});
});
describe("Phase 6: whole-promotion serialization for one company", () => {
it("withCodexAccountHomePromotionLock never overlaps two concurrent callers for the same company", async () => {
// The device-login route holds this lock across its whole promotion
// sequence: the account-home directory decision, the credential write,
// and the secret bind or cleanup that follows. Two different logins for
// the SAME Codex account must run that whole sequence one at a time, so
// a login can never delete the shared account-home directory while
// another login's own sequence is still writing its credential or
// binding its own secret to that same directory. This proves the lock
// itself enforces that: two concurrent callers for one company never run
// their callbacks at the same time, whichever caller goes first.
const home = await makeInstanceRoot();
const env = envFor(home);
const events: string[] = [];
let releaseFirstCaller!: () => void;
const firstCallerGate = new Promise<void>((resolve) => {
releaseFirstCaller = resolve;
});
const firstCall = withCodexAccountHomePromotionLock(env, "company-shared", async () => {
events.push("first-enter");
await firstCallerGate;
events.push("first-exit");
return "first";
});
// Give the first caller a chance to acquire the lock and enter its
// callback before the second caller starts racing for the same lock.
await new Promise((resolve) => setTimeout(resolve, 20));
const secondCall = withCodexAccountHomePromotionLock(env, "company-shared", async () => {
events.push("second-enter");
events.push("second-exit");
return "second";
});
// The second caller must stay blocked on the lock while the first
// caller still holds it: it must never log `second-enter` before the
// first caller's `first-exit`.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["first-enter"]);
releaseFirstCaller();
const [first, second] = await Promise.all([firstCall, secondCall]);
expect(first).toBe("first");
expect(second).toBe("second");
expect(events).toEqual(["first-enter", "first-exit", "second-enter", "second-exit"]);
});
it("withCodexAccountHomePromotionLock keeps two different companies independent", async () => {
// The lock is per company, so two different companies' device logins
// never wait on each other.
const home = await makeInstanceRoot();
const env = envFor(home);
const events: string[] = [];
let releaseCompanyA!: () => void;
const companyAGate = new Promise<void>((resolve) => {
releaseCompanyA = resolve;
});
const companyACall = withCodexAccountHomePromotionLock(env, "company-a", async () => {
events.push("a-enter");
await companyAGate;
events.push("a-exit");
});
await new Promise((resolve) => setTimeout(resolve, 20));
const companyBCall = withCodexAccountHomePromotionLock(env, "company-b", async () => {
events.push("b-enter");
events.push("b-exit");
});
await companyBCall;
// Company B's callback ran and finished while company A's callback was
// still waiting on its own gate, so the two locks never contended.
expect(events).toEqual(["a-enter", "b-enter", "b-exit"]);
releaseCompanyA();
await companyACall;
expect(events).toEqual(["a-enter", "b-enter", "b-exit", "a-exit"]);
});
});
describe("Phase 7: account-home secret-mutation serialization for one company", () => {
it("withAccountHomeSecretMutationLock never overlaps two concurrent callers for the same company", async () => {
// The secrets service holds this lock for the whole of a `local_encrypted`
// secret's create or rotate call, and an account-home cleanup's claimant
// scan holds it for the whole of its final check-and-delete step. This
// proves the lock itself enforces mutual exclusion between any two
// holders for one company, whichever caller goes first.
const home = await makeInstanceRoot();
const env = envFor(home);
const events: string[] = [];
let releaseFirstCaller!: () => void;
const firstCallerGate = new Promise<void>((resolve) => {
releaseFirstCaller = resolve;
});
const firstCall = withAccountHomeSecretMutationLock(env, "company-shared", async () => {
events.push("first-enter");
await firstCallerGate;
events.push("first-exit");
return "first";
});
await new Promise((resolve) => setTimeout(resolve, 20));
const secondCall = withAccountHomeSecretMutationLock(env, "company-shared", async () => {
events.push("second-enter");
events.push("second-exit");
return "second";
});
// The second caller must stay blocked on the lock while the first
// caller still holds it: it must never log `second-enter` before the
// first caller's `first-exit`.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["first-enter"]);
releaseFirstCaller();
const [first, second] = await Promise.all([firstCall, secondCall]);
expect(first).toBe("first");
expect(second).toBe("second");
expect(events).toEqual(["first-enter", "first-exit", "second-enter", "second-exit"]);
});
it("withAccountHomeSecretMutationLock keeps two different companies independent", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const events: string[] = [];
let releaseCompanyA!: () => void;
const companyAGate = new Promise<void>((resolve) => {
releaseCompanyA = resolve;
});
const companyACall = withAccountHomeSecretMutationLock(env, "company-a", async () => {
events.push("a-enter");
await companyAGate;
events.push("a-exit");
});
await new Promise((resolve) => setTimeout(resolve, 20));
const companyBCall = withAccountHomeSecretMutationLock(env, "company-b", async () => {
events.push("b-enter");
events.push("b-exit");
});
await companyBCall;
expect(events).toEqual(["a-enter", "b-enter", "b-exit"]);
releaseCompanyA();
await companyACall;
expect(events).toEqual(["a-enter", "b-enter", "b-exit", "a-exit"]);
});
it("withAccountHomeSecretMutationLock does not contend with withCodexAccountHomePromotionLock for the same company", async () => {
// The two locks use separate lock directories (Security condition: no
// shared lock key), so a device-login promotion that holds
// `withCodexAccountHomePromotionLock` for its whole sequence can still
// call into a secrets-service write that takes this lock without
// deadlocking on itself.
const home = await makeInstanceRoot();
const env = envFor(home);
const events: string[] = [];
await withCodexAccountHomePromotionLock(env, "company-shared", async () => {
events.push("promotion-enter");
await withAccountHomeSecretMutationLock(env, "company-shared", async () => {
events.push("mutation-enter");
events.push("mutation-exit");
});
events.push("promotion-exit");
});
expect(events).toEqual(["promotion-enter", "mutation-enter", "mutation-exit", "promotion-exit"]);
});
});
describe("Phase 8: account-home directory validity before a secret write commits", () => {
it("resolves for a value naming a directory that still exists under the cache root", async () => {
const home = await makeInstanceRoot();
const env = envFor(home);
const companyId = "company-still-exists";
const entryPath = await ensureCodexAuthCacheEntryDir(env, "acct-still-exists", companyId);
const accountHomeDir = path.dirname(entryPath);
await expect(
assertAccountHomeCacheDirStillValid(env, companyId, accountHomeDir),
).resolves.toBeUndefined();
});
it("rejects a value naming a directory under the cache root that no longer exists", async () => {
// This is the shape an account-home cleanup leaves behind: a value that
// once named a real directory, now removed. A create or a rotate that
// is about to commit this exact value must fail instead of writing a
// secret that points at nothing.
const home = await makeInstanceRoot();
const env = envFor(home);
const companyId = "company-removed";
const entryPath = await ensureCodexAuthCacheEntryDir(env, "acct-removed", companyId);
const accountHomeDir = path.dirname(entryPath);
await rm(accountHomeDir, { recursive: true, force: true });
await expect(assertAccountHomeCacheDirStillValid(env, companyId, accountHomeDir)).rejects.toThrow(
/no longer exists/,
);
});
it("leaves a value alone when it does not sit under this company's cache root", async () => {
// Most `local_encrypted` secret values (an API key, a token, a
// hand-typed string) never sit under the cache root, so this check must
// never reject a write that only happens to name a missing path.
const home = await makeInstanceRoot();
const env = envFor(home);
await expect(
assertAccountHomeCacheDirStillValid(env, "company-unrelated", "/some/unrelated/missing/path"),
).resolves.toBeUndefined();
});
});
});

View File

@ -3,15 +3,19 @@ import path from "node:path";
import { randomUUID } from "node:crypto";
import { resolvePaperclipInstanceRootForAdapter } from "@paperclipai/adapter-utils/server-utils";
import { withDirectoryMergeLock } from "@paperclipai/adapter-utils/workspace-restore-merge";
import { toAccountHandle } from "@paperclipai/shared";
import { USE_SOURCE_EXIT, decideCodexAuthMerge } from "./codex-auth-merge-decision.js";
import { writeCredentialSeedOrNewer } from "./codex-auth-seed-write.js";
// The identity-keyed host credential cache keeps one usable subscription
// credential per identity (`account_id`) in a SEPARATE host store, outside the
// shared Codex home and outside the symlink allowlist. The cache is additive: it
// The identity-keyed host credential store keeps one usable subscription
// credential per identity (`account_id`) in a SEPARATE host tree, outside the
// shared Codex home and outside the symlink allowlist. Each identity's entry
// directory doubles as that account's addressable home: the device-login
// promotion writes the durable credential there, and a company secret names the
// directory by path. The store also still backs the identity-anchored vend: it
// never changes the copy-back path, the fail-closed decision predicate, or the
// host default store overwrite. It only refreshes an identity the host already
// holds. It never seeds an empty host store and never picks a credential at
// host default store overwrite, it only refreshes an identity the host already
// holds, it never seeds an empty host store, and it never picks a credential at
// random.
const CACHE_DIR_NAME = "codex-auth-cache";
@ -110,18 +114,25 @@ export function resolveCodexAuthCacheDir(
/**
* Resolves the entry path for one identity: `<cacheRoot>/<safeAccountId>/auth.json`.
* The `account_id` is sanitized by {@link toCacheKey}. After the join, this
* verifies the resolved entry path stays under the cache root and ends at exactly
* `<safeAccountId>/auth.json`. This function does no filesystem work; it is safe
* for a read path (the vend and the clear). (Security condition 3.)
* The `account_id` is validated first by {@link toAccountHandle} (a strict
* allowlist), the entry point of this function, then sanitized again by
* {@link toCacheKey} (a denylist) as a second, independent layer. After the
* join, this verifies the resolved entry path stays under the cache root and
* ends at exactly `<safeAccountId>/auth.json`. This function does no filesystem
* work; it is safe for a read path (the vend and the clear). (Security
* condition 3.)
*/
export function resolveCodexAuthCacheEntryPath(
env: NodeJS.ProcessEnv = process.env,
accountId: string,
companyId: string,
): string {
const handle = toAccountHandle(accountId);
if (!handle) {
throw new Error("codex auth cache: account_id is not a valid account handle");
}
const resolvedRoot = resolveCodexAuthCacheDir(env, companyId);
const safeKey = toCacheKey(accountId);
const safeKey = toCacheKey(handle);
const entryDir = path.resolve(resolvedRoot, safeKey);
const entryPath = path.resolve(entryDir, CACHE_ENTRY_FILE);
const expectedEntryPath = path.join(resolvedRoot, safeKey, CACHE_ENTRY_FILE);
@ -156,8 +167,9 @@ async function ensurePrivateDir(dir: string): Promise<void> {
/**
* Resolves the entry path and creates the cache root and the entry directory
* private (mode 0700), each guarded by `lstat`. Use this on the write path
* before the cache slot is written.
* private (mode 0700), each guarded by `lstat`. The `account_id` is validated
* at the entry point of this function through {@link resolveCodexAuthCacheEntryPath}.
* Use this on the write path before the entry is written.
*/
export async function ensureCodexAuthCacheEntryDir(
env: NodeJS.ProcessEnv = process.env,
@ -170,11 +182,194 @@ export async function ensureCodexAuthCacheEntryDir(
return entryPath;
}
export interface EnsureCodexAuthCacheEntryDirExclusiveResult {
/** The entry path: `<accountHomeDir>/auth.json`. */
entryPath: string;
/** True only when this exact call found the account's own home directory
* absent and created it. */
created: boolean;
}
/**
* Same as {@link ensureCodexAuthCacheEntryDir}, but the absence check and the
* directory creation run inside one lock keyed to the company-scoped cache
* root. Two different logins for the SAME Codex account can promote at the
* same time (each login owns its own promotion slot), so a plain
* check-then-create sequence lets both concurrent callers see the slot as
* absent and both report `created: true`. Under this lock, only the caller
* that truly finds the slot absent gets `created: true`; the other caller
* correctly reports `created: false` and so never deletes a home the first
* caller's login already wrote to and named with a company secret.
*/
export async function ensureCodexAuthCacheEntryDirExclusive(
env: NodeJS.ProcessEnv = process.env,
accountId: string,
companyId: string,
): Promise<EnsureCodexAuthCacheEntryDirExclusiveResult> {
const cacheRoot = resolveCodexAuthCacheDir(env, companyId);
await ensurePrivateDir(cacheRoot);
return withDirectoryMergeLock(
cacheRoot,
async () => {
const entryPath = resolveCodexAuthCacheEntryPath(env, accountId, companyId);
const entryDir = path.dirname(entryPath);
const created = await lstat(entryDir)
.then(() => false)
.catch((error: NodeJS.ErrnoException) => {
if (error.code === "ENOENT") return true;
throw error;
});
await ensurePrivateDir(entryDir);
return { entryPath, created };
},
env,
);
}
/**
* Resolves a per-company, named lock target directory under the same
* instance-scoped `companies/<companyId>/` tree the cache root and the
* promotion lock use. Two callers that pass the same `lockName` for the same
* `companyId` always resolve the same directory, so they share one lock; two
* different `lockName` values never share a lock key, so a caller that holds
* one of these named locks can still call {@link ensureCodexAuthCacheEntryDirExclusive}
* (which locks the cache root) or another named lock without nesting a lock
* inside itself.
*/
function resolveCodexAuthCacheNamedLockDir(
env: NodeJS.ProcessEnv = process.env,
companyId: string,
lockName: string,
): string {
const safeCompanyId = toSafePathSegment(companyId, "companyId");
const instanceRoot = resolvePaperclipInstanceRootForAdapter({
homeDir: nonEmpty(env.PAPERCLIP_HOME) ?? undefined,
instanceId: nonEmpty(env.PAPERCLIP_INSTANCE_ID) ?? undefined,
env,
});
return path.resolve(instanceRoot, "companies", safeCompanyId, lockName);
}
/**
* Serializes one company's whole device-login promotion sequence: the
* account-home directory create-or-reuse decision, the credential write, and
* the secret bind or cleanup that follows. Two different logins for the SAME
* Codex account run this sequence one at a time under this lock, so no login
* can decide to delete the shared account-home directory while another
* login's flow is still mid-way through writing its credential or binding
* its own secret to that same directory.
*
* `ensureCodexAuthCacheEntryDirExclusive` alone is not enough: it locks only
* the short directory-creation step, so its lock is already released by the
* time a login reaches the secret bind. A second login can then write its
* credential and be about to bind its own secret while the first login's
* later, unrelated secret-write failure decides to remove the directory both
* logins now share, deleting a credential home the second login still needs.
* Holding this lock across the full sequence for every login closes that
* window: a login that must clean up its own directory always finishes that
* cleanup, including the directory-recreation of any later login that lands
* on the now-empty slot, before the next login's sequence starts.
*/
export async function withCodexAccountHomePromotionLock<T>(
env: NodeJS.ProcessEnv = process.env,
companyId: string,
fn: () => Promise<T>,
): Promise<T> {
const lockDir = resolveCodexAuthCacheNamedLockDir(env, companyId, "codex-auth-cache-promotion-lock");
await ensurePrivateDir(lockDir);
return withDirectoryMergeLock(lockDir, fn, env);
}
/**
* Serializes one company's writes to the literal value of any
* `local_encrypted` company secret against the account-home cleanup's
* claimant scan in {@link withCodexAccountHomePromotionLock}'s caller. Only a
* `local_encrypted` secret can hold a literal directory path, so this is the
* one provider a cleanup claimant scan must trust. Without this lock, a
* secret create or rotation can commit its new value after the scan's last
* pass finds no claimant but before the cleanup deletes the directory: the
* scan and the write never observe each other, so the write's new value
* survives past a directory the delete has already removed.
*
* The secrets service holds this lock for the full duration of a
* `local_encrypted` secret's create or rotate call. A cleanup claimant scan
* holds it for the full duration of its final check-and-delete step. The two
* critical sections can then never interleave: a write that starts first
* finishes (and becomes visible to the scan) before the scan's lock-holding
* check runs, and a write that starts after the scan's lock-holding check
* finishes only commits once the scan (and any delete it decided on) is
* done, so it can never name a directory the delete just removed without the
* scan having had a chance to see it first.
*
* This is a separate lock directory from
* {@link withCodexAccountHomePromotionLock}'s, so a device-login promotion
* (which holds that lock for its whole sequence, including its own secret
* create) can still call into a secrets-service write that takes this lock
* without deadlocking on itself.
*/
export async function withAccountHomeSecretMutationLock<T>(
env: NodeJS.ProcessEnv = process.env,
companyId: string,
fn: () => Promise<T>,
): Promise<T> {
const lockDir = resolveCodexAuthCacheNamedLockDir(env, companyId, "codex-account-home-secret-mutation-lock");
await ensurePrivateDir(lockDir);
return withDirectoryMergeLock(lockDir, fn, env);
}
/**
* Confirms a `local_encrypted` secret value that names a directory under this
* company's Codex account-home cache root still exists on disk. Call this
* inside {@link withAccountHomeSecretMutationLock}, before a create or a
* rotate commits the value.
*
* The lock alone stops a write and the account-home cleanup's
* check-and-delete from interleaving; it does not stop them from running in
* either order. When the cleanup's check-and-delete runs first it finds no
* secret names the directory, because this write had not committed yet, and
* removes the directory a write that was only queued behind the lock still
* goes on, once the lock frees, to commit the very directory the cleanup
* just removed. That would leave an active secret naming a directory that
* does not exist. This check closes that window: a write whose value would
* name a directory the cleanup already removed fails instead of committing.
*
* A value outside this company's cache root is left alone. Only a value
* under the cache root can ever be an account-home directory, so this never
* rejects an unrelated `local_encrypted` secret write, such as an API key or
* a hand-typed token.
*/
export async function assertAccountHomeCacheDirStillValid(
env: NodeJS.ProcessEnv = process.env,
companyId: string,
value: string,
): Promise<void> {
const cacheRoot = resolveCodexAuthCacheDir(env, companyId);
if (!value.startsWith(cacheRoot + path.sep)) return;
const exists = await lstat(value)
.then(() => true)
.catch((error: NodeJS.ErrnoException) => {
if (error.code === "ENOENT") return false;
throw error;
});
if (!exists) {
throw new Error(
"codex auth cache: account-home directory no longer exists; refusing to write a secret that names it",
);
}
}
/**
* Reads the usable subscription `account_id` from an `auth.json` payload. Returns
* `null` for an absent, unusable, or api-key credential (no subscription
* identity). This mirrors `parseAuth` in `codex-auth-merge-decision.cjs`; keep
* the two in step when the auth format changes.
*
* The returned value is the exact, untrimmed `account_id` string. A caller
* passes it on to {@link toAccountHandle} unchanged: that function is the
* one place that decides whether surrounding whitespace is acceptable, and it
* must see the raw value to reject an identifier that differs from another
* one only by whitespace. Trimming here, before that check runs, would let
* two distinct identifiers collapse onto the same account handle.
*/
export function readSubscriptionAccountId(bytes: Buffer): string | null {
let parsed: unknown;
@ -196,7 +391,10 @@ export function readSubscriptionAccountId(bytes: Buffer): string | null {
return null;
}
const tokenRecord = tokens as Record<string, unknown>;
const accountId = typeof tokenRecord.account_id === "string" ? tokenRecord.account_id.trim() : "";
const rawAccountId = typeof tokenRecord.account_id === "string" ? tokenRecord.account_id : "";
// A blank-or-whitespace-only value is absent; a value with real content
// keeps its exact, untrimmed form.
const accountId = rawAccountId.trim().length > 0 ? rawAccountId : "";
const hasTokenMaterial = ["id_token", "access_token", "refresh_token"].some((key) => {
const value = tokenRecord[key];
return typeof value === "string" && value.trim().length > 0;

View File

@ -40,7 +40,14 @@ export {
type CredentialReadinessResult,
type PromoteDeviceLoginCredentialInput,
type PromoteDeviceLoginCredentialOutcome,
type PromoteDeviceLoginCredentialResult,
} from "./adapter-auth-promotion.js";
export {
withCodexAccountHomePromotionLock,
withAccountHomeSecretMutationLock,
assertAccountHomeCacheDirStillValid,
resolveCodexAuthCacheDir,
} from "./codex-auth-cache.js";
export { parseCodexJsonl, isCodexHarnessCrash, isCodexProviderQuotaError, isCodexTransientUpstreamError, isCodexUnknownSessionError } from "./parse.js";
export {
getQuotaWindows,

View File

@ -2806,7 +2806,13 @@ describe("Daytona sandbox provider plugin", () => {
],
});
// Let syncIn register on the activity gate and reach the hung upload.
await new Promise((resolve) => setTimeout(resolve, 0));
// The real `fs.stat` and `mkdir` round trip run before the upload call,
// so a fixed tick count can race ahead of them on a slower or busier
// host. Poll for the actual upload call instead of guessing a tick
// count, so this assertion never fires before syncIn reaches the hang.
await vi.waitFor(() => {
expect(sandbox.fs.uploadFiles).toHaveBeenCalled();
});
const cancelPromise = plugin.definition.onEnvironmentCancelInteractiveSetup?.({
driverKey: "daytona",

View File

@ -0,0 +1,73 @@
import { describe, expect, it } from "vitest";
import { ACCOUNT_HANDLE_MAX_LENGTH, toAccountHandle } from "./account-handle.js";
import { createUserSecretDefinitionSchema } from "./validators/secret.js";
describe("toAccountHandle", () => {
it("returns the value for a plain identifier", () => {
expect(toAccountHandle("acct-42")).toBe("acct-42");
});
it("returns null for a value that holds a space, such as \"a b\"", () => {
expect(toAccountHandle("a b")).toBeNull();
});
it("returns null for a value with surrounding whitespace instead of trimming it", () => {
// Trimming a value before validation would let " acct-42" and "acct-42"
// resolve to the same handle. Two distinct identifiers must never share
// one handle, so surrounding whitespace is a rejection, not something to
// strip.
expect(toAccountHandle(" acct-42")).toBeNull();
expect(toAccountHandle("acct-42 ")).toBeNull();
expect(toAccountHandle(" acct-42 ")).toBeNull();
expect(toAccountHandle("\tacct-42\n")).toBeNull();
// The plain identifier with no surrounding whitespace still passes.
expect(toAccountHandle("acct-42")).toBe("acct-42");
});
it("returns null for an empty string", () => {
expect(toAccountHandle("")).toBeNull();
});
it("returns null for a value that holds a plus sign, such as \"a+b\"", () => {
expect(toAccountHandle("a+b")).toBeNull();
});
it("returns null for a value that holds a colon, such as \"a:b\"", () => {
expect(toAccountHandle("a:b")).toBeNull();
});
it("returns null for a value longer than 100 characters", () => {
const tooLong = "a".repeat(ACCOUNT_HANDLE_MAX_LENGTH + 1);
expect(toAccountHandle(tooLong)).toBeNull();
const atLimit = "a".repeat(ACCOUNT_HANDLE_MAX_LENGTH);
expect(toAccountHandle(atLimit)).toBe(atLimit);
});
it("returns null for \".\" and for \"..\"", () => {
expect(toAccountHandle(".")).toBeNull();
expect(toAccountHandle("..")).toBeNull();
});
it("returns null for a value that starts with a hyphen, such as \"-rf\"", () => {
expect(toAccountHandle("-rf")).toBeNull();
});
it("returns null for a value that holds a shell metacharacter, such as \"$(id)\"", () => {
expect(toAccountHandle("$(id)")).toBeNull();
});
it("returns null for a value that holds a path separator or a NUL byte", () => {
expect(toAccountHandle("a/b")).toBeNull();
expect(toAccountHandle("a\\b")).toBeNull();
expect(toAccountHandle("a\0b")).toBeNull();
});
it("every accepted handle also passes the secret key schema", () => {
const candidates = ["acct-42", "acct_42", "ACCT.42", "a".repeat(ACCOUNT_HANDLE_MAX_LENGTH)];
for (const candidate of candidates) {
const handle = toAccountHandle(candidate);
expect(handle).not.toBeNull();
expect(createUserSecretDefinitionSchema.shape.key.safeParse(handle).success).toBe(true);
}
});
});

View File

@ -0,0 +1,44 @@
// A canonical account handle names one login account inside one company. Two
// downstream sinks each guard the raw account identifier with a different
// validator: a directory-segment denylist, and the secret-key allowlist
// (`packages/shared/src/validators/secret.ts`). `toAccountHandle` accepts a
// value only when it satisfies both, so one handle can safely name both a
// directory and a secret.
/** The longest accepted account handle. The secret key limit is 120
* characters, and the longest name prefix a caller adds is `CODEX_HOME_` at
* 11 characters, so 100 leaves headroom. */
export const ACCOUNT_HANDLE_MAX_LENGTH = 100;
// The same character class the secret key schema uses.
const ACCOUNT_HANDLE_PATTERN = /^[a-zA-Z0-9_.-]+$/;
/**
* Converts a raw account identifier into a safe account handle. Returns
* `null` when the value cannot become one instead of throwing, so a caller
* decides how to fail the operation.
*
* The identifier is the account's identity key: it names one directory and
* one secret, so two distinct identifiers must never resolve to the same
* handle. A caller must never normalize the value before this check (for
* example, by trimming it) normalizing first can alias a distinct
* identifier onto an already-claimed handle and let a login accept a
* different account's credential home as its own.
*
* An input becomes a handle only when every rule below is true:
*
* - It carries no leading or trailing whitespace.
* - It matches `/^[a-zA-Z0-9_.-]+$/`.
* - It is {@link ACCOUNT_HANDLE_MAX_LENGTH} characters or shorter.
* - It is neither `.` nor `..`.
* - It does not start with `-` (a leading `-` reads as a command-line option).
*/
export function toAccountHandle(rawAccountId: string): string | null {
if (typeof rawAccountId !== "string" || rawAccountId.length === 0) return null;
if (rawAccountId.trim() !== rawAccountId) return null;
if (!ACCOUNT_HANDLE_PATTERN.test(rawAccountId)) return null;
if (rawAccountId.length > ACCOUNT_HANDLE_MAX_LENGTH) return null;
if (rawAccountId === "." || rawAccountId === "..") return null;
if (rawAccountId.startsWith("-")) return null;
return rawAccountId;
}

View File

@ -2726,3 +2726,4 @@ export {
isPaperclipDevRunnerCommand,
rewriteUrlHostToLoopback,
} from "./runtime-exposure/loopback-bind.js";
export { ACCOUNT_HANDLE_MAX_LENGTH, toAccountHandle } from "./account-handle.js";

View File

@ -1,6 +1,10 @@
import express from "express";
import request from "supertest";
import { lstat, mkdtemp, rm } from "node:fs/promises";
import os from "node:os";
import path from "node:path";
import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
import { withAccountHomeSecretMutationLock } from "@paperclipai/adapter-codex-local/server";
import { AdapterAuthSessionConflictError } from "../services/device-login-service.js";
import type {
AdapterAuthSessionRow,
@ -68,6 +72,26 @@ const mockSecretService = vi.hoisted(() => ({
secretKeys: new Set<string>(),
manifest: [],
})),
// The account-home secret lookup and creation. Absent by default, so a
// successful promotion creates it; a test overrides either to drive the
// idempotent-repeat-login and secret-creation-failure paths.
getByName: vi.fn(async (_companyId: string, _name: string) => null as Record<string, unknown> | null),
create: vi.fn(async (_companyId: string, _input: Record<string, unknown>) => ({ id: "secret-1" })),
// The company-wide secret listing the cleanup-safety scan reads. Empty by
// default, so a cleanup that finds no other claimant removes the
// directory; a test overrides it with a differently named secret to prove
// the scan finds a claimant a name-only check would miss.
list: vi.fn(
async (_companyId: string) => [] as Array<{ id: string; name: string; provider: string }>,
),
// The pre-existing-secret value check. Defaults to a value that never
// matches an account home, so a test that leaves this unset and still
// reaches an existing-secret branch fails loud instead of passing by
// accident; a test overrides it to the account home under test.
resolveSecretValueForDeviceLoginCheck: vi.fn(
async (_companyId: string, _secretId: string, _context: { configPath: string }) =>
"/unset-mock-value" as string,
),
}));
const mockEnvironmentService = vi.hoisted(() => ({
@ -341,10 +365,20 @@ const loginPath = (companyId: string, type = "codex_local") =>
`/api/companies/${companyId}/adapters/${type}/login-sessions`;
describe("adapter device-login routes", () => {
// Real temp directories a test plants as an account home, cleaned up in
// `afterEach` whether or not the route removed them itself.
const accountHomeTestDirs: string[] = [];
beforeEach(() => {
vi.resetModules();
vi.clearAllMocks();
mockDeviceLoginPromotion.mockResolvedValue("promoted");
mockDeviceLoginPromotion.mockResolvedValue({
outcome: "promoted",
accountId: "acct-default",
accountHomeDir: "/tmp/paperclip-codex-account-home/acct-default",
});
mockSecretService.getByName.mockResolvedValue(null);
mockSecretService.create.mockResolvedValue({ id: "secret-1" });
harness.store = createMemoryStore();
harness.runtime = createFakeRuntime();
harness.acquisitions = [];
@ -376,9 +410,14 @@ describe("adapter device-login routes", () => {
);
});
afterEach(() => {
afterEach(async () => {
// Release the gate, so every in-flight login run ends and clears its timer.
harness.releaseGate();
while (accountHomeTestDirs.length > 0) {
const dir = accountHomeTestDirs.pop();
if (!dir) continue;
await rm(dir, { recursive: true, force: true }).catch(() => undefined);
}
});
it("requires a board actor and rejects an agent-token start", async () => {
@ -726,6 +765,12 @@ describe("adapter device-login routes", () => {
});
it("lets a second owner start an active login in the same company", async () => {
// Both logins bind the default account home, and the login service
// reconfirms the bound secret's value a second time right before it
// commits the terminal `authenticated` state.
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(
"/tmp/paperclip-codex-account-home/acct-default",
);
const app = await createApp();
const first = await request(app)
@ -744,9 +789,31 @@ describe("adapter device-login routes", () => {
expect(second.body.sessionId).not.toBe(first.body.sessionId);
// Each owner's start acquires its own lease.
expect(harness.acquisitions).toHaveLength(2);
// Drain both sessions to a terminal status before the test ends. Neither
// session's promotion has run yet (both still wait on the gate), so an
// undrained session keeps running after this test returns and can consume
// a later test's own mocked promotion result. A status read is owner
// scoped, so read each session back as the owner that started it.
harness.releaseGate();
await vi.waitFor(async () => {
currentActor = boardActor(OWNER_A);
const firstStatus = await request(app).get(`${loginPath(COMPANY_1)}/${first.body.sessionId}`);
expect(firstStatus.body.status).toBe("authenticated");
});
await vi.waitFor(async () => {
currentActor = boardActor(OWNER_B);
const secondStatus = await request(app).get(`${loginPath(COMPANY_1)}/${second.body.sessionId}`);
expect(secondStatus.body.status).toBe("authenticated");
});
});
it("returns 409 for a second active start in a different environment", async () => {
// The login binds the default account home, and the login service
// reconfirms the bound secret's value a second time right before it
// commits the terminal `authenticated` state.
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(
"/tmp/paperclip-codex-account-home/acct-default",
);
const app = await createApp();
const first = await request(app)
@ -761,12 +828,24 @@ describe("adapter device-login routes", () => {
.send({ environmentId: SANDBOX_ENV_2 });
expect(second.status, JSON.stringify(second.body)).toBe(409);
expect(harness.acquisitions).toHaveLength(1);
// Drain the first session to a terminal status before the test ends. An
// undrained session keeps running after this test returns and can consume
// a later test's own mocked promotion result.
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${first.body.sessionId}`);
expect(status.body.status).toBe("authenticated");
});
});
it("fails closed when promotion loses the sole-owner claim", async () => {
// This is the expiry/reaper-race result from Decision H. It is a normal
// resolved adapter outcome, but it must never be accepted as authentication.
mockDeviceLoginPromotion.mockResolvedValueOnce("not_sole_owner");
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "not_sole_owner",
accountId: null,
accountHomeDir: null,
});
const app = await createApp();
const start = await request(app)
@ -790,12 +869,176 @@ describe("adapter device-login routes", () => {
expect(mockDeviceLoginPromotion).toHaveBeenCalledTimes(1);
});
it("fails closed when the login is a different account than the company home", async () => {
// The promotion keeps the occupied company home and installs nothing durable
// for a different-identity login. The identity-anchored vend can never select
// this login, so a later run keeps the existing account. The route must fail
// the session instead of a report of `authenticated`.
mockDeviceLoginPromotion.mockResolvedValueOnce("kept_foreign_identity");
it("a successful Codex login creates the company secret naming the account home", async () => {
// A second, different account no longer fails behind an occupied company
// home: each account gets its own home, named by a company secret.
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(
"/tmp/paperclip-codex-account-home/acct-second",
);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-second",
accountHomeDir: "/tmp/paperclip-codex-account-home/acct-second",
});
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("authenticated");
});
expect(mockSecretService.getByName).toHaveBeenCalledWith(COMPANY_1, "CODEX_HOME_acct-second");
expect(mockSecretService.create).toHaveBeenCalledWith(
COMPANY_1,
expect.objectContaining({
name: "CODEX_HOME_acct-second",
value: "/tmp/paperclip-codex-account-home/acct-second",
}),
expect.anything(),
);
});
it("a repeat login for the same account creates no second secret", async () => {
// The secret already names this account's home, so the route reads it,
// confirms the value still matches, and creates nothing new: a repeat
// login is idempotent.
const accountHomeDir = "/tmp/paperclip-codex-account-home/acct-default";
mockSecretService.getByName.mockResolvedValue({ id: "existing-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "kept",
accountId: "acct-default",
accountHomeDir,
});
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("authenticated");
});
expect(mockSecretService.create).not.toHaveBeenCalled();
expect(mockSecretService.resolveSecretValueForDeviceLoginCheck).toHaveBeenCalledWith(
COMPANY_1,
"existing-secret",
expect.objectContaining({ configPath: "secrets.CODEX_HOME_acct-default" }),
);
});
it("holds the account-home secret mutation lock while checking a pre-existing secret's value", async () => {
// A `local_encrypted` secret rotate takes `withAccountHomeSecretMutationLock`
// for the whole of its write. The pre-existing-secret check must run
// inside the same lock and return right after: otherwise a rotate could
// commit a new value in the gap between the check and the login
// reporting success, and the login would report success for a value
// that no longer names the account's own home. This proves the check
// waits for a lock held elsewhere, and only reads the secret's value
// once that lock frees.
const accountHomeDir = "/tmp/paperclip-codex-account-home/acct-default";
mockSecretService.getByName.mockResolvedValue({ id: "existing-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "kept",
accountId: "acct-default",
accountHomeDir,
});
let releaseHeldLock!: () => void;
const heldLockGate = new Promise<void>((resolve) => {
releaseHeldLock = resolve;
});
const lockHolder = withAccountHomeSecretMutationLock(undefined, COMPANY_1, () => heldLockGate);
// Give the held lock a chance to actually acquire before the login
// starts racing for the same lock.
await new Promise((resolve) => setTimeout(resolve, 20));
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
// The login's check must stay blocked behind the held lock.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(mockSecretService.resolveSecretValueForDeviceLoginCheck).not.toHaveBeenCalled();
releaseHeldLock();
await lockHolder;
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("authenticated");
});
expect(mockSecretService.resolveSecretValueForDeviceLoginCheck).toHaveBeenCalled();
});
it("reconfirms the account-home secret a second time immediately before the terminal authenticated commit", async () => {
// The early check inside `promote` releases its lock well before the
// login service records the terminal `authenticated` state, so it alone
// cannot prove the value still matches at that later point. The service
// must run the same check again, right before that commit, under a
// fresh lock acquisition: this test proves the second check runs, not
// only the first.
const accountHomeDir = "/tmp/paperclip-codex-account-home/acct-default";
mockSecretService.getByName.mockResolvedValue({ id: "existing-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "kept",
accountId: "acct-default",
accountHomeDir,
});
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("authenticated");
});
expect(mockSecretService.resolveSecretValueForDeviceLoginCheck).toHaveBeenCalledTimes(2);
});
it("fails closed when the account-home secret is rotated after validation but before the login service commits its terminal state", async () => {
// A board user can rotate the matching `CODEX_HOME` secret after
// `promote`'s own check passes and before the login service commits its
// terminal state, because `promote`'s lock is fully released by then.
// The service must catch this with a second check it runs right before
// that commit, or it would report `authenticated` for a value that no
// longer names the account's own home.
const accountHomeDir = "/tmp/paperclip-codex-account-home/acct-default";
mockSecretService.getByName.mockResolvedValue({ id: "existing-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck
// `promote`'s own early check: still matches.
.mockResolvedValueOnce(accountHomeDir)
// The rotation lands in the gap. The final check, immediately before
// the terminal commit, now reads the rotated value.
.mockResolvedValueOnce("/rotated/away/from/the/account/home");
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "kept",
accountId: "acct-default",
accountHomeDir,
});
const app = await createApp();
const start = await request(app)
@ -811,12 +1054,471 @@ describe("adapter device-login routes", () => {
expect(status.body.failure?.reason).toBe("promotion_failed");
});
const row = await (harness.store as ReturnType<typeof createMemoryStore>).getByPublicId(
sessionId,
COMPANY_1,
expect(mockSecretService.resolveSecretValueForDeviceLoginCheck).toHaveBeenCalledTimes(2);
});
it("fails closed when a pre-existing secret names a different account home", async () => {
// A stale or hand-entered secret at this name must not let the login
// report success while a bound agent reads the wrong credential home.
mockSecretService.getByName.mockResolvedValue({ id: "stale-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(
"/some/other/path/left-over-from-before",
);
expect(row?.status).toBe("failed");
expect(mockDeviceLoginPromotion).toHaveBeenCalledTimes(1);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "kept",
accountId: "acct-stale",
accountHomeDir: "/tmp/paperclip-codex-account-home/acct-stale",
});
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
expect(mockSecretService.create).not.toHaveBeenCalled();
});
it("fails closed and removes the account home when secret creation fails", async () => {
// This login created the account home, so a secret-creation failure must
// fail the login and remove the directory it just created, so the
// operation stays atomic.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-broken-secret",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).rejects.toThrow();
});
it("a secret creation failure keeps an account home that already existed", async () => {
// The account home directory pre-dates this login (a user deleted the
// secret but kept the directory, or an earlier login already wrote it).
// A secret-creation failure must still fail the login, but it must never
// remove a directory this login did not create.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-existing-home",
accountHomeDir,
accountHomeCreated: false,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("a secret name conflict succeeds and keeps the account home", async () => {
// Two logins for one account can both read no secret and then race the
// create call. The loser gets a 409 conflict, which proves the secret
// already exists. A repeat login for the same account must be idempotent,
// so the loser must confirm the winning secret's value, report a
// successful login, and must never remove the account home the winner
// just wrote.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-conflict",
accountHomeDir,
accountHomeCreated: true,
});
// `vi.resetModules()` in `beforeEach` gives the route module a fresh copy
// of `../errors.js`. Import the conflict helper from that same fresh copy,
// so the route's `err instanceof HttpError` check sees a matching class.
const { conflict: freshConflict } =
await vi.importActual<typeof import("../errors.js")>("../errors.js");
mockSecretService.create.mockRejectedValueOnce(
freshConflict("a secret with this name already exists"),
);
// The first read (before `create`) finds no secret; the second read
// (after the conflict) finds the winner's secret.
mockSecretService.getByName
.mockResolvedValueOnce(null)
.mockResolvedValueOnce({ id: "winning-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(accountHomeDir);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("authenticated");
});
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("fails closed when the secret that won the create race names a different account home", async () => {
// The 409 conflict proves some secret now holds this name, but that
// secret could belong to unrelated stale state, not the winner of a
// genuine same-account race. The login must not report success without
// checking the winning secret's value.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-conflict-mismatch",
accountHomeDir,
accountHomeCreated: true,
});
const { conflict: freshConflict } =
await vi.importActual<typeof import("../errors.js")>("../errors.js");
mockSecretService.create.mockRejectedValueOnce(
freshConflict("a secret with this name already exists"),
);
mockSecretService.getByName
.mockResolvedValueOnce(null)
.mockResolvedValueOnce({ id: "winning-secret" });
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValue(
"/some/other/path/left-over-from-before",
);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
// The mismatch is a validation failure on the winner's secret, not this
// call's own directory, so this call must still never delete it.
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("does not remove an account home a concurrent login already claimed", async () => {
// This call created the account home directory (`accountHomeCreated:
// true`), but a concurrent login for the same account can still create
// the company secret before this call's own, unrelated secret-creation
// attempt fails. The directory now belongs to that other login's secret,
// so the cleanup scan must find it by value and keep the directory.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-concurrent-claim",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
// The first read (before `create`) finds no secret.
mockSecretService.getByName.mockResolvedValueOnce(null);
// The cleanup scan lists every company secret and resolves each one's
// value. The concurrent login's secret carries this call's own generated
// name, so this proves the scan still catches a same-name claimant.
mockSecretService.list.mockResolvedValueOnce([
{ id: "concurrent-secret", name: "CODEX_HOME_acct-concurrent-claim", provider: "local_encrypted" },
]);
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValueOnce(accountHomeDir);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("does not remove an account home a differently named secret still references", async () => {
// A user can bind a hand-named secret to an account home directly, with
// no relation to the generated `CODEX_HOME_<handle>` name. A cleanup that
// checks only the generated name would miss this secret and delete a
// directory it still needs; the scan must catch it by value instead.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-aliased",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
mockSecretService.getByName.mockResolvedValueOnce(null);
mockSecretService.list.mockResolvedValueOnce([
// An unrelated secret that resolves to a different value. The scan
// must not stop at the first entry and must not treat every secret as
// a match.
{ id: "unrelated-secret", name: "SOME_OTHER_SECRET", provider: "local_encrypted" },
{ id: "hand-named-secret", name: "MY_CODEX_HOME", provider: "local_encrypted" },
]);
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockImplementation(
async (_companyId: string, secretId: string, _context: { configPath: string }) =>
secretId === "hand-named-secret" ? accountHomeDir : "/some/unrelated/path",
);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("does not remove an account home an AWS Secrets Manager-backed secret still references", async () => {
// A secret's value is a plain string regardless of its provider, so a
// hand-named secret bound to this account home through AWS Secrets
// Manager is just as real a claimant as a `local_encrypted` one. The
// scan must resolve it too, not skip it for its provider.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-aws-claim",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
mockSecretService.getByName.mockResolvedValueOnce(null);
mockSecretService.list.mockResolvedValueOnce([
{ id: "aws-secret", name: "MY_CODEX_HOME", provider: "aws_secrets_manager" },
]);
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValueOnce(accountHomeDir);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
expect(mockSecretService.resolveSecretValueForDeviceLoginCheck).toHaveBeenCalledWith(
COMPANY_1,
"aws-secret",
expect.objectContaining({ configPath: "secrets.MY_CODEX_HOME" }),
);
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("removes the account home when the cleanup scan finds no claimant", async () => {
// No secret, under any name, resolves to this directory, so the cleanup
// must still remove it — the broader scan must not make cleanup any less
// eager than the old name-only check when nothing claims the directory.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-unclaimed",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
mockSecretService.getByName.mockResolvedValueOnce(null);
mockSecretService.list.mockResolvedValueOnce([
{ id: "unrelated-secret", name: "SOME_OTHER_SECRET", provider: "local_encrypted" },
]);
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockResolvedValueOnce("/some/unrelated/path");
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).rejects.toThrow();
});
it("preserves the account home when the cleanup scan cannot resolve a secret's value", async () => {
// A resolve failure on one secret is not proof that secret names a
// different directory: the failure can hit the exact secret that would
// have matched. The cleanup must keep the directory rather than treat an
// unresolved secret as a non-match, even when every secret that DID
// resolve named a different directory.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-unresolvable",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
mockSecretService.getByName.mockResolvedValueOnce(null);
mockSecretService.list.mockResolvedValueOnce([
{ id: "unrelated-secret", name: "SOME_OTHER_SECRET", provider: "local_encrypted" },
{ id: "unresolvable-secret", name: "MAYBE_CODEX_HOME", provider: "local_encrypted" },
]);
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockImplementation(
async (_companyId: string, secretId: string, _context: { configPath: string }) => {
if (secretId === "unresolvable-secret") throw new Error("decryption key unavailable");
return "/some/unrelated/path";
},
);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("preserves the account home when a secret appears after the scan starts", async () => {
// The scan lists secrets once, then resolves each one in turn. A new
// secret can appear after that first list call, while the scan is still
// resolving an earlier secret. The scan must re-list and check that new
// secret too, instead of trusting its first, now-stale list.
const accountHomeDir = await mkdtemp(path.join(os.tmpdir(), "paperclip-codex-account-home-test-"));
accountHomeTestDirs.push(accountHomeDir);
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct-late-claim",
accountHomeDir,
accountHomeCreated: true,
});
mockSecretService.create.mockRejectedValueOnce(new Error("db unavailable"));
mockSecretService.getByName.mockResolvedValueOnce(null);
// The first pass sees only the unrelated secret. By the time the scan
// re-lists, a second, differently named secret now names this account
// home; the scan must catch it on that later pass.
mockSecretService.list
.mockResolvedValueOnce([{ id: "unrelated-secret", name: "SOME_OTHER_SECRET", provider: "local_encrypted" }])
.mockResolvedValueOnce([
{ id: "unrelated-secret", name: "SOME_OTHER_SECRET", provider: "local_encrypted" },
{ id: "late-secret", name: "LATE_CODEX_HOME", provider: "local_encrypted" },
]);
mockSecretService.resolveSecretValueForDeviceLoginCheck.mockImplementation(
async (_companyId: string, secretId: string, _context: { configPath: string }) =>
secretId === "late-secret" ? accountHomeDir : "/some/unrelated/path",
);
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
await expect(lstat(accountHomeDir)).resolves.toBeDefined();
});
it("fails closed when the account identifier cannot form a valid account home name", async () => {
// The promotion helper itself already rejects an identifier that cannot
// become an account handle; a defensive re-check in the route covers a
// promotion result the route did not fully trust.
mockDeviceLoginPromotion.mockResolvedValueOnce({
outcome: "promoted",
accountId: "acct with space",
accountHomeDir: "/tmp/paperclip-codex-account-home/unused",
});
const app = await createApp();
const start = await request(app)
.post(loginPath(COMPANY_1))
.send({ environmentId: SANDBOX_ENV_1 });
expect(start.status, JSON.stringify(start.body)).toBe(201);
const sessionId = start.body.sessionId as string;
harness.releaseGate();
await vi.waitFor(async () => {
const status = await request(app).get(`${loginPath(COMPANY_1)}/${sessionId}`);
expect(status.body.status).toBe("failed");
expect(status.body.failure?.reason).toBe("promotion_failed");
});
expect(mockSecretService.create).not.toHaveBeenCalled();
});
it("omits the URL, code, credential bytes, and lease id from logs and activity", async () => {

View File

@ -382,6 +382,84 @@ describe("device login service", () => {
expect(JSON.stringify(activity)).not.toContain(PROMPT_CODE);
});
it("wraps the terminal authenticated commit in a promotion's runTerminalCommit", async () => {
// A promotion that defines `runTerminalCommit` gets the chance to wrap the
// service's own terminal write, so it can hold a lock across a check it
// already ran once earlier in `promote` and the write that publishes
// `authenticated`.
const store = createMemoryStore();
const { runtime } = createFakeRuntime({
exec: execSuccess,
authBytes: Buffer.from('{"token":"secret"}'),
});
const companyId = randomUUID();
const wrapCalls: unknown[] = [];
const service = makeService({
store,
runtime,
promotion: {
promote: () => {},
async runTerminalCommit(commit, context) {
wrapCalls.push(context);
return commit();
},
},
});
const { session, completed } = await service.start({
companyId,
environmentId: randomUUID(),
adapterType: ADAPTER_TYPE,
startedByUserId: OWNER_A,
});
const outcome = await completed;
expect(outcome.status).toBe("authenticated");
expect(wrapCalls).toHaveLength(1);
const row = await store.getByPublicId(session.sessionId, companyId);
expect(row?.status).toBe("authenticated");
});
it("records a failed terminal, and never authenticates, when runTerminalCommit rejects", async () => {
// A promotion's `runTerminalCommit` rejects when its own re-check, run
// immediately before the terminal write, finds the value it validated
// earlier no longer holds (for example, a rotation landed in the gap).
// The service must fail the login the same way a `promote` rejection
// does, never publishing `authenticated` for the stale value.
const store = createMemoryStore();
const { runtime, deleteCalls } = createFakeRuntime({
exec: execSuccess,
authBytes: Buffer.from('{"token":"secret"}'),
});
const companyId = randomUUID();
const service = makeService({
store,
runtime,
promotion: {
promote: () => {},
async runTerminalCommit() {
// A real promotion would run its re-check here, find the bound
// value stale, and throw before ever calling `commit`.
throw new Error("the bound value no longer matches");
},
},
});
const { session, completed } = await service.start({
companyId,
environmentId: randomUUID(),
adapterType: ADAPTER_TYPE,
startedByUserId: OWNER_A,
});
const outcome = await completed;
expect(outcome.status).toBe("failed");
const row = await store.getByPublicId(session.sessionId, companyId);
expect(row?.status).toBe("failed");
expect(row?.failureReason).toBe("promotion_failed");
// The service still deletes the sandbox on this failure path, the same
// as a `promote` rejection.
expect(deleteCalls).toHaveLength(1);
});
it("gives a Grok prompt to a grok_local session, and the session surfaces the Grok code and URL", async () => {
// This proves the Grok parser ran: the profile map resolves the parser from
// the trusted adapter type, so a `grok_local` session runs the Grok parser,

View File

@ -1,9 +1,11 @@
import { randomUUID } from "node:crypto";
import { mkdirSync, rmSync } from "node:fs";
import { mkdir, rm } from "node:fs/promises";
import os from "node:os";
import path from "node:path";
import { afterAll, afterEach, beforeAll, describe, expect, it, vi } from "vitest";
import { and, eq } from "drizzle-orm";
import { resolveCodexAuthCacheDir, withAccountHomeSecretMutationLock } from "@paperclipai/adapter-codex-local/server";
import {
activityLog,
agents,
@ -88,6 +90,16 @@ describeEmbeddedPostgres("secretService", () => {
return companyId;
}
// Creates a real directory under this company's Codex account-home cache
// root, so a test can prove the write-time directory-validity check reads
// an actual filesystem entry, the same way the production cleanup and the
// production secret write do.
async function makeAccountHomeDir(companyId: string, accountHandle: string): Promise<string> {
const accountHomeDir = path.join(resolveCodexAuthCacheDir(undefined, companyId), accountHandle);
await mkdir(accountHomeDir, { recursive: true });
return accountHomeDir;
}
async function seedCompanyMember(
companyId: string,
userId: string,
@ -360,6 +372,446 @@ describeEmbeddedPostgres("secretService", () => {
).rejects.toThrow(/already exists/i);
});
it("serializes two local_encrypted secret creates in the same company so their writes never overlap", async () => {
// An account-home cleanup's claimant scan and a `local_encrypted` secret
// write share one lock (`withAccountHomeSecretMutationLock`), so a write
// can never commit inside the exact window the scan already used to
// decide no secret claims a directory it is about to delete. This proves
// the lock itself enforces that: two `local_encrypted` creates in the
// SAME company never run their provider write at the same time,
// whichever caller goes first.
const companyId = await seedCompany();
const svc = secretService(db);
const events: string[] = [];
let releaseFirstWrite!: () => void;
const firstWriteGate = new Promise<void>((resolve) => {
releaseFirstWrite = resolve;
});
const originalCreateSecret = localEncryptedProvider.createSecret.bind(localEncryptedProvider);
vi.spyOn(localEncryptedProvider, "createSecret")
.mockImplementationOnce(async (input) => {
events.push("first-provider-enter");
await firstWriteGate;
events.push("first-provider-exit");
return originalCreateSecret(input);
})
.mockImplementationOnce(async (input) => {
events.push("second-provider-enter");
return originalCreateSecret(input);
});
const firstCreate = svc.create(companyId, {
name: `account-home-${randomUUID()}`,
provider: "local_encrypted",
value: "/company/codex-home/acct-a",
});
// Give the first call a chance to acquire the lock and enter its provider
// write before the second call starts racing for the same lock.
await new Promise((resolve) => setTimeout(resolve, 20));
const secondCreate = svc.create(companyId, {
name: `hand-named-${randomUUID()}`,
provider: "local_encrypted",
value: "/company/codex-home/acct-a",
});
// The second call must stay blocked on the lock while the first call
// still holds it: it must never enter its own provider write before the
// first call's provider write exits.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["first-provider-enter"]);
releaseFirstWrite();
await Promise.all([firstCreate, secondCreate]);
expect(events).toEqual(["first-provider-enter", "first-provider-exit", "second-provider-enter"]);
});
it("serializes a local_encrypted secret rotate against a concurrent create of a different secret in the same company", async () => {
// Same lock, the other write path: a rotate that writes a new
// `local_encrypted` value must serialize against a concurrent create the
// same way a create serializes against another create.
const companyId = await seedCompany();
const svc = secretService(db);
const existing = await svc.create(companyId, {
name: `existing-${randomUUID()}`,
provider: "local_encrypted",
value: "/company/codex-home/acct-b",
});
const events: string[] = [];
let releaseRotateWrite!: () => void;
const rotateWriteGate = new Promise<void>((resolve) => {
releaseRotateWrite = resolve;
});
const originalCreateVersion = localEncryptedProvider.createVersion.bind(localEncryptedProvider);
vi.spyOn(localEncryptedProvider, "createVersion").mockImplementationOnce(async (input) => {
events.push("rotate-provider-enter");
await rotateWriteGate;
events.push("rotate-provider-exit");
return originalCreateVersion(input);
});
const originalCreateSecret = localEncryptedProvider.createSecret.bind(localEncryptedProvider);
vi.spyOn(localEncryptedProvider, "createSecret").mockImplementationOnce(async (input) => {
events.push("create-provider-enter");
return originalCreateSecret(input);
});
const rotateCall = svc.rotate(existing.id, { value: "/company/codex-home/acct-b-rotated" });
await new Promise((resolve) => setTimeout(resolve, 20));
const createCall = svc.create(companyId, {
name: `hand-named-${randomUUID()}`,
provider: "local_encrypted",
value: "/company/codex-home/acct-b",
});
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["rotate-provider-enter"]);
releaseRotateWrite();
await Promise.all([rotateCall, createCall]);
expect(events).toEqual(["rotate-provider-enter", "rotate-provider-exit", "create-provider-enter"]);
});
it("fails a queued local_encrypted create when an account-home cleanup removes its directory first", async () => {
// The mutation lock alone stops a write and an account-home cleanup's
// check-and-delete from interleaving; it does not stop them from running
// in either order. When the cleanup wins the lock first, deletes the
// directory, and releases the lock, a create that was only queued behind
// it must not go on to commit that now-deleted directory as a secret
// value. It must fail instead.
const companyId = await seedCompany();
const svc = secretService(db);
const accountHomeDir = await makeAccountHomeDir(companyId, "acct-queued-create");
const events: string[] = [];
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
events.push("cleanup-enter");
await cleanupGate;
await rm(accountHomeDir, { recursive: true, force: true });
events.push("cleanup-exit");
});
// Give the cleanup a chance to acquire the lock before the create starts
// racing for the same lock.
await new Promise((resolve) => setTimeout(resolve, 20));
const createCall = svc.create(companyId, {
name: `account-home-${randomUUID()}`,
provider: "local_encrypted",
value: accountHomeDir,
});
// The create must stay queued behind the held lock.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["cleanup-enter"]);
releaseCleanup();
await cleanupCall;
await expect(createCall).rejects.toThrow(/no longer exists/);
});
it("fails a queued local_encrypted rotate when an account-home cleanup removes its directory first", async () => {
// Same race, the other write path: a rotate that would commit a
// now-deleted account-home directory must fail the same way a create
// does.
const companyId = await seedCompany();
const svc = secretService(db);
const existing = await svc.create(companyId, {
name: `hand-named-${randomUUID()}`,
provider: "local_encrypted",
value: "/some/unrelated/placeholder/value",
});
const accountHomeDir = await makeAccountHomeDir(companyId, "acct-queued-rotate");
const events: string[] = [];
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
events.push("cleanup-enter");
await cleanupGate;
await rm(accountHomeDir, { recursive: true, force: true });
events.push("cleanup-exit");
});
await new Promise((resolve) => setTimeout(resolve, 20));
const rotateCall = svc.rotate(existing.id, { value: accountHomeDir });
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["cleanup-enter"]);
releaseCleanup();
await cleanupCall;
await expect(rotateCall).rejects.toThrow(/no longer exists/);
});
it("serializes an aws_secrets_manager secret create against a concurrent account-home mutation lock holder", async () => {
// A plain string value can equal a Codex account-home path regardless of
// which provider stores it. Prove a non-local (`aws_secrets_manager`)
// create now takes the SAME `withAccountHomeSecretMutationLock` a
// `local_encrypted` create takes, not only when the create's own
// provider is `local_encrypted`.
const companyId = await seedCompany();
const svc = secretService(db);
const awsVault = await svc.createProviderConfig(companyId, {
provider: "aws_secrets_manager",
displayName: "AWS production",
config: { region: "us-east-1", namespace: "prod-use1" },
});
vi.spyOn(awsSecretsManagerProvider, "createSecret").mockResolvedValue({
material: {
scheme: "aws_secrets_manager_v1",
secretId: "arn:aws:secretsmanager:us-east-1:123456789012:secret:paperclip/prod-use1/company/aws-secret",
versionId: "aws-version-1",
source: "managed",
},
valueSha256: "value-sha-1",
fingerprintSha256: "fingerprint-sha-1",
externalRef: "arn:aws:secretsmanager:us-east-1:123456789012:secret:paperclip/prod-use1/company/aws-secret",
providerVersionRef: "aws-version-1",
});
const events: string[] = [];
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
events.push("cleanup-enter");
await cleanupGate;
events.push("cleanup-exit");
});
// Give the cleanup a chance to acquire the lock before the AWS create
// starts racing for the same lock.
await new Promise((resolve) => setTimeout(resolve, 20));
const createCall = svc.create(companyId, {
name: `aws-secret-${randomUUID()}`,
provider: "aws_secrets_manager",
providerConfigId: awsVault.id,
value: "runtime-secret",
});
// The AWS create must stay queued behind the held lock, the same way a
// `local_encrypted` create does.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["cleanup-enter"]);
releaseCleanup();
await cleanupCall;
const created = await createCall;
expect(events).toEqual(["cleanup-enter", "cleanup-exit"]);
expect(created.status).toBe("active");
});
it("serializes an aws_secrets_manager secret rotate against a concurrent account-home mutation lock holder", async () => {
// Same reasoning as the create test above, the other write path: a
// non-local rotate must also take the lock, not only a
// `local_encrypted` rotate.
const companyId = await seedCompany();
const svc = secretService(db);
const awsVault = await svc.createProviderConfig(companyId, {
provider: "aws_secrets_manager",
displayName: "AWS production",
config: { region: "us-east-1", namespace: "prod-use1" },
});
vi.spyOn(awsSecretsManagerProvider, "createSecret").mockResolvedValue({
material: {
scheme: "aws_secrets_manager_v1",
secretId: "arn:aws:secretsmanager:us-east-1:123456789012:secret:paperclip/prod-use1/company/aws-secret-rotate",
versionId: "aws-version-1",
source: "managed",
},
valueSha256: "value-sha-1",
fingerprintSha256: "fingerprint-sha-1",
externalRef: "arn:aws:secretsmanager:us-east-1:123456789012:secret:paperclip/prod-use1/company/aws-secret-rotate",
providerVersionRef: "aws-version-1",
});
vi.spyOn(awsSecretsManagerProvider, "createVersion").mockResolvedValue({
material: {
scheme: "aws_secrets_manager_v1",
secretId: "arn:aws:secretsmanager:us-east-1:123456789012:secret:paperclip/prod-use1/company/aws-secret-rotate",
versionId: "aws-version-2",
source: "managed",
},
valueSha256: "value-sha-2",
fingerprintSha256: "fingerprint-sha-2",
externalRef: "arn:aws:secretsmanager:us-east-1:123456789012:secret:paperclip/prod-use1/company/aws-secret-rotate",
providerVersionRef: "aws-version-2",
});
const existing = await svc.create(companyId, {
name: `aws-secret-rotate-${randomUUID()}`,
provider: "aws_secrets_manager",
providerConfigId: awsVault.id,
value: "runtime-secret",
});
const events: string[] = [];
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
events.push("cleanup-enter");
await cleanupGate;
events.push("cleanup-exit");
});
await new Promise((resolve) => setTimeout(resolve, 20));
const rotateCall = svc.rotate(existing.id, { value: "rotated-runtime-secret" });
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["cleanup-enter"]);
releaseCleanup();
await cleanupCall;
const rotated = await rotateCall;
expect(events).toEqual(["cleanup-enter", "cleanup-exit"]);
expect(rotated.latestVersion).toBe(2);
});
it("serializes a rename, archive, or disable update against a concurrent account-home mutation lock holder", async () => {
// A rename or a status change to `archived` or `disabled` can stop the
// generated account-home secret from resolving to the value a
// device-login promotion already validated. Prove `update` now holds
// the SAME lock a cleanup (or a promotion's terminal-commit re-check)
// holds for its whole critical section, so an update can never land
// inside that section.
const companyId = await seedCompany();
const svc = secretService(db);
const secret = await svc.create(companyId, {
name: `CODEX_HOME_update-${randomUUID()}`,
provider: "local_encrypted",
value: "/company/codex-home/acct-update",
});
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
await cleanupGate;
});
await new Promise((resolve) => setTimeout(resolve, 20));
let updateResolved = false;
const updateCall = svc.update(secret.id, { status: "archived" }).then((result) => {
updateResolved = true;
return result;
});
// The update must stay queued behind the held lock.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(updateResolved).toBe(false);
releaseCleanup();
await cleanupCall;
const updated = await updateCall;
expect(updateResolved).toBe(true);
expect(updated?.status).toBe("archived");
});
it("serializes a secret delete against a concurrent account-home mutation lock holder", async () => {
// Same reasoning as the update test above: a delete must also take the
// lock, so it can never land inside a cleanup's, or a promotion's
// terminal-commit re-check's, critical section.
const companyId = await seedCompany();
const svc = secretService(db);
const secret = await svc.create(companyId, {
name: `CODEX_HOME_delete-${randomUUID()}`,
provider: "local_encrypted",
value: "/company/codex-home/acct-delete",
});
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
await cleanupGate;
});
await new Promise((resolve) => setTimeout(resolve, 20));
let removeResolved = false;
const removeCall = svc.remove(secret.id).then((result) => {
removeResolved = true;
return result;
});
// The delete must stay queued behind the held lock.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(removeResolved).toBe(false);
releaseCleanup();
await cleanupCall;
await removeCall;
expect(removeResolved).toBe(true);
const remaining = await svc.getById(secret.id);
expect(remaining).toBeNull();
});
it("commits a local_encrypted secret naming an account-home directory that still exists", async () => {
// A regression guard for the check above: a create whose value names a
// directory that is still present must keep succeeding, unblocked.
const companyId = await seedCompany();
const svc = secretService(db);
const accountHomeDir = await makeAccountHomeDir(companyId, "acct-still-present");
const created = await svc.create(companyId, {
name: `account-home-${randomUUID()}`,
provider: "local_encrypted",
value: accountHomeDir,
});
expect(created.status).toBe("active");
});
it("fails a queued adapter-schema-secret create when an account-home cleanup removes its directory first", async () => {
// `normalizeAdapterConfigForPersistence` creates an adapter config secret
// through a separate managed-create path (`createManagedLocalSecret`),
// not `svc.create`. It must run the same queued-behind-the-lock
// directory check as `svc.create` and `svc.rotate`, so an adapter
// schema secret can never commit an account-home directory a cleanup
// already removed while this call waited for the lock.
const companyId = await seedCompany();
const svc = secretService(db);
const accountHomeDir = await makeAccountHomeDir(companyId, "acct-queued-adapter-secret");
const events: string[] = [];
let releaseCleanup!: () => void;
const cleanupGate = new Promise<void>((resolve) => {
releaseCleanup = resolve;
});
const cleanupCall = withAccountHomeSecretMutationLock(undefined, companyId, async () => {
events.push("cleanup-enter");
await cleanupGate;
await rm(accountHomeDir, { recursive: true, force: true });
events.push("cleanup-exit");
});
// Give the cleanup a chance to acquire the lock before the adapter
// config normalization starts racing for the same lock.
await new Promise((resolve) => setTimeout(resolve, 20));
const normalizeCall = svc.normalizeAdapterConfigForPersistence(
companyId,
{ apiKey: accountHomeDir },
{ adapterType: "hermes_gateway" },
);
// The queued create must stay blocked behind the held lock.
await new Promise((resolve) => setTimeout(resolve, 20));
expect(events).toEqual(["cleanup-enter"]);
releaseCleanup();
await cleanupCall;
await expect(normalizeCall).rejects.toThrow(/no longer exists/);
});
it("keeps an adapter-schema-secret create working for a value outside the account-home cache root", async () => {
// A regression guard for the check above: an adapter schema secret whose
// value is not an account-home directory (an ordinary API key, for
// example) must keep succeeding, unblocked.
const companyId = await seedCompany();
const svc = secretService(db);
const normalized = await svc.normalizeAdapterConfigForPersistence(
companyId,
{ apiKey: `plain-api-key-${randomUUID()}` },
{ adapterType: "hermes_gateway" },
);
const apiKeyBinding = (normalized as Record<string, unknown>).apiKey as { type: string; secretId: string };
expect(apiKeyBinding.type).toBe("secret_ref");
const secret = await svc.getById(apiKeyBinding.secretId);
expect(secret?.status).toBe("active");
});
it("validates the access namespace as agent-only with env-style aliases", async () => {
const companyId = await seedCompany();
const svc = secretService(db);

View File

@ -1,5 +1,6 @@
import { Router, type NextFunction, type Request, type Response } from "express";
import { generateKeyPairSync, randomUUID } from "node:crypto";
import { rm } from "node:fs/promises";
import path from "node:path";
import type { Db } from "@paperclipai/db";
import { agents as agentsTable, companies, heartbeatRuns, issues as issuesTable, projects as projectsTable } from "@paperclipai/db";
@ -32,6 +33,7 @@ import {
startAdapterAuthSessionRequestSchema,
startClaudeSetupTokenSessionRequestSchema,
submitBrowserCodeRequestSchema,
toAccountHandle,
type AgentAdapterType,
} from "@paperclipai/shared";
import {
@ -174,6 +176,8 @@ import {
import {
checkStagedCredentialReadiness,
promoteDeviceLoginCredential,
withAccountHomeSecretMutationLock,
withCodexAccountHomePromotionLock,
} from "@paperclipai/adapter-codex-local/server";
import {
checkStagedGrokCredentialReadiness,
@ -273,6 +277,119 @@ function readRunIssueId(context: Record<string, unknown> | null) {
return typeof nestedIssueId === "string" && isUuidLike(nestedIssueId) ? nestedIssueId : null;
}
// Confirms a pre-existing `CODEX_HOME_<handle>` secret still names this
// account's own home before a device login treats the secret's presence as a
// successful, idempotent login. The secret name alone is not proof of a
// match: a stale value from before the cache root moved, or a value a user
// entered by hand, would otherwise let the login report success while a
// bound agent reads the wrong (or a missing) credential home. Fails loud on
// a mismatch, so the login fails instead of silently pointing agents at the
// wrong home.
//
// Each caller must run this function inside `withAccountHomeSecretMutationLock`,
// the same lock a `local_encrypted` secret rotate holds for its whole write.
// A caller that only checks once, early in the promotion, and then reports
// success later is not enough on its own: the lock this call held is fully
// released by the time it returns, so a rotate queued behind it can commit a
// new value before the login service records its terminal `authenticated`
// state, which happens well after this call returns (see `runTerminalCommit`
// below, which runs this same check again, under a fresh lock acquisition,
// immediately before that terminal state commits).
async function assertAccountHomeSecretMatches(
secretsSvc: { resolveSecretValueForDeviceLoginCheck: (companyId: string, secretId: string, context: { configPath: string }) => Promise<string> },
companyId: string,
secret: { id: string },
secretName: string,
expectedAccountHomeDir: string,
): Promise<void> {
const storedValue = await secretsSvc.resolveSecretValueForDeviceLoginCheck(companyId, secret.id, {
configPath: `secrets.${secretName}`,
});
if (storedValue !== expectedAccountHomeDir) {
throw new Error(
`device-login credential promotion rejected: the existing ${secretName} secret does not name this account's own home`,
);
}
}
// Confirms no company secret, under any name or provider, still names this
// account home before a failed promotion deletes the directory. The
// generated `CODEX_HOME_<handle>` name is not the only secret that can
// reference this directory: a user can bind a hand-named secret to the same
// account home, so a check that reads only the generated name misses that
// secret and deletes a directory it still needs. A bound agent then reads a
// `CODEX_HOME` value that points at nothing. A secret's value is a plain
// string regardless of its provider, so an AWS Secrets Manager-backed secret
// (or any other provider) can equal this directory's path just as a
// `local_encrypted` secret can; the scan resolves every secret's value, not
// only `local_encrypted` ones.
//
// A secret whose value fails to resolve is NOT proof that secret names a
// different directory: the resolve call can fail for a secret that would
// have matched. Treating that failure as a non-match would let the cleanup
// delete a directory a secret still needs. So the scan fails closed: any
// resolution failure makes the whole scan report a claim, even when every
// secret that DID resolve named a different directory.
//
// The scan lists every secret once, then resolves each secret's value in
// turn, and each resolve call is its own round trip. A new secret can enter
// the company between the initial list and the last resolve call, so a
// single pass can finish, find no claimant among the secrets it read, and
// still miss a secret that named this directory moments later. So the scan
// re-lists after every pass and resolves only the secrets it has not yet
// checked, and it only reports "no claimant" once a pass finds nothing new
// to check. A scan that keeps finding new secrets on every pass fails
// closed after a bounded number of passes, so a fast stream of concurrent
// secret creation cannot force an unsafe delete.
//
// The scan alone still cannot rule out a secret write that commits after the
// scan's own last pass finishes but before the caller's delete runs: the
// scan and that write are two separate operations with no shared state, so
// neither can see the other. The caller closes that window by running the
// scan and the delete inside `withAccountHomeSecretMutationLock`, the same
// lock every `local_encrypted` secret create or rotate holds for its whole
// write. That lock is the atomic protection; the multi-pass scan above stays
// as a defense-in-depth check for a write path that has not taken the lock.
const ACCOUNT_HOME_CLAIM_SCAN_MAX_PASSES = 5;
async function anySecretNamesAccountHome(
secretsSvc: {
list: (companyId: string) => Promise<Array<{ id: string; name: string; provider: string }>>;
resolveSecretValueForDeviceLoginCheck: (
companyId: string,
secretId: string,
context: { configPath: string },
) => Promise<string>;
},
companyId: string,
accountHomeDir: string,
): Promise<boolean> {
const checkedSecretIds = new Set<string>();
for (let pass = 0; pass < ACCOUNT_HOME_CLAIM_SCAN_MAX_PASSES; pass += 1) {
const secrets = await secretsSvc.list(companyId);
const uncheckedSecrets = secrets.filter((secret) => !checkedSecretIds.has(secret.id));
if (uncheckedSecrets.length === 0) return false;
let resolutionFailed = false;
for (const secret of uncheckedSecrets) {
checkedSecretIds.add(secret.id);
const storedValue = await secretsSvc
.resolveSecretValueForDeviceLoginCheck(companyId, secret.id, {
configPath: `secrets.${secret.name}`,
})
.catch(() => {
resolutionFailed = true;
return null;
});
if (storedValue === accountHomeDir) return true;
}
if (resolutionFailed) return true;
}
// Every pass found a secret it had not yet checked. Fail closed: an
// endless stream of new secrets is not proof that none of them claims
// this directory.
return true;
}
export function agentRoutes(
db: Db,
options: {
@ -542,6 +659,21 @@ export function agentRoutes(
// process owns one instance, so the in-memory prompt and the cancellation
// controllers persist across requests.
const adapterLoginStore = createDbAdapterAuthSessionStore(db);
// The account-home secret a Codex login validated and bound, keyed by
// session id, queued for `runTerminalCommit` to reconfirm right before the
// login service commits its terminal `authenticated` write. `promote`'s own
// check runs under a lock that is fully released by the time `promote`
// returns, so a rotation can still land after that check and before the
// terminal write; `runTerminalCommit` closes that gap by re-running the
// same check under a fresh lock acquisition that it holds across the
// terminal write itself. `runTerminalCommit` deletes the entry it reads, so
// nothing outlives one login attempt.
const pendingAccountHomeSecretCommits = new Map<
string,
{ secretId: string; secretName: string; accountHomeDir: string }
>();
const adapterLoginService = createDeviceLoginService({
store: adapterLoginStore,
runtime: createProductionLoginSessionRuntime({
@ -577,59 +709,222 @@ export function agentRoutes(
// promotion function.
promotionByAdapterType: {
codex_local: {
// Hold one lock across the whole promotion sequence below: the
// credential write, the existing-secret check, the secret create, and
// the cleanup a create failure can trigger. Two different logins for
// the SAME Codex account run this whole sequence one at a time, so a
// login can never decide to delete the shared account-home directory
// while another login's own sequence is still mid-way through writing
// its credential or binding its own secret to that same directory. A
// lock around only the directory-creation step is not enough: that
// lock is already released by the time a login reaches the secret
// bind, so a second login can write its credential and be about to
// bind its own secret while the first login's later, unrelated
// secret-write failure removes the directory both logins now share.
async promote(authBytes, context) {
// Hold the promotion critical-section lock across the ownership check and
// the credential write. The reaper takes the same lock before it reclaims
// a stale `promoting` row. So a reclaim never interleaves with a live
// write: the reaper either wins the lock first and the ownership check
// then reads a reclaimed row and writes nothing, or the write finishes
// first under the lock and the reaper reclaims only after it completes. A
// read-only fence is not enough, because the filesystem write can start
// after the fence; the lock spans the whole section.
const outcome = await adapterLoginStore.withCompanyAdapterPromotionLock(
context.companyId,
context.startedByUserId,
context.adapterType,
() =>
promoteDeviceLoginCredential({
authBytes,
companyId: context.companyId,
userInitiated: true,
checkReadiness: (bytes) => checkStagedCredentialReadiness(bytes),
isSoleActiveOwner: async () => {
// The partial unique index allows one active row per company and
// adapter. So a `promoting` row for this session is the sole
// active owner of the company credential slot. The read runs
// inside the lock, so it observes a reaper reclaim that committed
// before this section acquired the lock.
const row = await adapterLoginStore.get(context.sessionId);
return row?.status === "promoting" && row.companyId === context.companyId;
},
log: (line) => {
// The promotion lines carry no token bytes and no raw account id,
// so it is safe to log them with the session identifier.
logger.info({ sessionId: context.sessionId }, line);
},
}),
);
// A resolved promotion is not necessarily an accepted promotion. In
// particular, a reaper/expiry race can revoke this session's sole
// ownership between the service transition and Decision H. Fail closed:
// only a credential write or a deliberate safe keep can authenticate.
if (outcome === "kept_foreign_identity") {
// The login produced a different account than the one the company
// credential home already holds. The promotion never clobbers an
// occupied home, so this login installed nothing durable, and the
// identity-anchored vend can never select it: a later run keeps the
// existing account. Fail the session, so the operator never sees a
// false `authenticated` for an account the system will not use.
throw new Error(
"device-login credential promotion rejected: the login is a different account than the one already set for this company; the existing account was kept",
return withCodexAccountHomePromotionLock(undefined, context.companyId, async () => {
// Hold the promotion critical-section lock across the ownership check
// and the credential write. The reaper takes the same lock before it
// reclaims a stale `promoting` row. So a reclaim never interleaves with
// a live write: the reaper either wins the lock first and the
// ownership check then reads a reclaimed row and writes nothing, or
// the write finishes first under the lock and the reaper reclaims only
// after it completes. A read-only fence is not enough, because the
// filesystem write can start after the fence; the lock spans the whole
// section.
const result = await adapterLoginStore.withCompanyAdapterPromotionLock(
context.companyId,
context.startedByUserId,
context.adapterType,
() =>
promoteDeviceLoginCredential({
authBytes,
companyId: context.companyId,
userInitiated: true,
checkReadiness: (bytes) => checkStagedCredentialReadiness(bytes),
isSoleActiveOwner: async () => {
// The partial unique index allows one active row per company and
// adapter. So a `promoting` row for this session is the sole
// active owner of the company credential slot. The read runs
// inside the lock, so it observes a reaper reclaim that committed
// before this section acquired the lock.
const row = await adapterLoginStore.get(context.sessionId);
return row?.status === "promoting" && row.companyId === context.companyId;
},
log: (line) => {
// The promotion lines carry no token bytes and no raw account id,
// so it is safe to log them with the session identifier.
logger.info({ sessionId: context.sessionId }, line);
},
}),
);
// A resolved promotion is not necessarily an accepted promotion. In
// particular, a reaper/expiry race can revoke this session's sole
// ownership between the service transition and Decision H. Fail closed:
// only a credential write or a deliberate safe keep can authenticate.
if (result.outcome !== "promoted" && result.outcome !== "kept") {
throw new Error(`device-login credential promotion rejected: ${result.outcome}`);
}
// The account's own home is durable at this point (the promotion above
// wrote it fail-loud). Name it with a company secret, so any agent can
// bind to it. Reading the secret by name first keeps a repeat login for
// the same account idempotent: `create` throws a conflict when the name
// already exists.
const handle = result.accountId ? toAccountHandle(result.accountId) : null;
if (!handle || !result.accountHomeDir) {
throw new Error(
"device-login credential promotion rejected: the promotion carried no account home",
);
}
const secretName = `CODEX_HOME_${handle}`;
const accountHomeDir = result.accountHomeDir;
const existingSecret = await secretsSvc.getByName(context.companyId, secretName);
if (existingSecret) {
// A same-name secret already exists. Confirm it still names this
// account's own home before treating a repeat login as a success:
// the name alone is not proof of a match.
//
// Run the check inside the same lock a `local_encrypted` secret
// rotate holds for its whole write, the same lock a rotate
// takes. This is an early fail-fast only: the lock is fully
// released once this call returns, well before the login
// service commits its terminal state, so queue the same check
// for `runTerminalCommit` to run again right before that
// commit, under a fresh lock acquisition it holds across the
// commit itself.
await withAccountHomeSecretMutationLock(undefined, context.companyId, () =>
assertAccountHomeSecretMatches(secretsSvc, context.companyId, existingSecret, secretName, accountHomeDir),
);
pendingAccountHomeSecretCommits.set(context.sessionId, {
secretId: existingSecret.id,
secretName,
accountHomeDir,
});
return;
}
try {
const createdSecret = await secretsSvc.create(
context.companyId,
{
name: secretName,
provider: "local_encrypted",
value: accountHomeDir,
description: `CODEX_HOME generated by logging into account ${handle}`,
},
{ userId: context.startedByUserId, agentId: null },
);
// The value just committed is correct at this instant, but a
// rotate queued behind the create's own lock can still commit a
// different value before the login service records its
// terminal state. Queue the same reconfirm `runTerminalCommit`
// runs for the two branches above.
pendingAccountHomeSecretCommits.set(context.sessionId, {
secretId: createdSecret.id,
secretName,
accountHomeDir,
});
} catch (err) {
if (err instanceof HttpError && err.status === 409) {
// A conflict means a concurrent login for the same account won the
// create race. Confirm the winning secret still names this
// account's own home before treating the race as a successful,
// idempotent login.
const winningSecret = await secretsSvc.getByName(context.companyId, secretName);
if (!winningSecret) {
throw new Error(
`device-login credential promotion rejected: the ${secretName} secret conflict could not be resolved`,
);
}
// Same lock and the same reasoning as the pre-existing-secret
// check above: an early fail-fast only, so also queue the
// same check for `runTerminalCommit` to run again, under a
// fresh lock acquisition it holds across the terminal commit.
await withAccountHomeSecretMutationLock(undefined, context.companyId, () =>
assertAccountHomeSecretMatches(secretsSvc, context.companyId, winningSecret, secretName, accountHomeDir),
);
pendingAccountHomeSecretCommits.set(context.sessionId, {
secretId: winningSecret.id,
secretName,
accountHomeDir,
});
return;
}
// The account home write failed for a reason other than a naming
// conflict. Remove the directory only when this exact login created
// it AND no secret, under any name, still names it. The lock
// around this whole method already rules out another
// SAME-ACCOUNT LOGIN from being mid-sequence here, but it does
// not rule out a secret a different path created at this exact
// directory in between the read above and this failure — for
// example, a concurrent login for the same account that won a
// race on this exact name, or a user who names a secret by hand
// under a different name entirely. The re-check is this call's
// only signal for that case, so it stays even under the lock:
// `accountHomeCreated` alone is not proof no such secret now
// claims the directory, and a same-name check alone is not proof
// either, because the claiming secret can carry any name.
//
// The check and the delete run inside `withAccountHomeSecretMutationLock`,
// the same lock the secrets service holds for the whole of a
// `local_encrypted` secret's create or rotate call. That closes the
// window `anySecretNamesAccountHome`'s own multi-pass scan cannot: a
// secret write that commits after this check's last pass but before
// the delete runs. Under the shared lock, a write either finishes
// (and becomes visible to the check) before this section acquires the
// lock, or it waits for this section to finish before it can commit.
if (result.accountHomeCreated) {
await withAccountHomeSecretMutationLock(undefined, context.companyId, async () => {
const claimed = await anySecretNamesAccountHome(secretsSvc, context.companyId, accountHomeDir);
if (!claimed) {
await rm(accountHomeDir, { recursive: true, force: true }).catch(() => undefined);
}
});
}
throw new Error(
"device-login credential promotion rejected: failed to record the account home secret",
);
}
});
},
// The login service calls this immediately before it commits its
// terminal `authenticated` write, wrapping that write in the
// callback it hands in as `commit`. `promote` above already
// validated the bound account-home secret once, early, but its own
// lock is fully released by the time `promote` returns — well before
// this runs. Re-run the same check here, and hold the SAME lock
// across both the check and `commit`, so a rotate cannot land in the
// gap between the validated value and the terminal write that
// reports it as authenticated: a rotate either finishes (and this
// check reads its new value, and rejects) before this section
// acquires the lock, or it waits for this section — including the
// terminal commit — to finish first.
async runTerminalCommit(commit, context) {
const pending = pendingAccountHomeSecretCommits.get(context.sessionId);
pendingAccountHomeSecretCommits.delete(context.sessionId);
if (!pending) {
// `promote` never reached a secret bind for this session (for
// example, a rejected promotion already failed the login before
// the service ever reaches this call). Nothing to reconfirm.
return commit();
}
if (outcome !== "promoted" && outcome !== "kept") {
throw new Error(`device-login credential promotion rejected: ${outcome}`);
}
return withAccountHomeSecretMutationLock(undefined, context.companyId, async () => {
// Resolve by the secret's id, not its name: a rotate changes the
// value under the same id, so re-resolving this id picks up a
// rotation the same way the very first check would have, with no
// need to re-look the secret up by name. A deleted secret makes
// this resolve call itself fail (unlike a value mismatch, which
// `assertAccountHomeSecretMatches` turns into its own error), and
// that failure propagates the same way: the login never
// authenticates.
await assertAccountHomeSecretMatches(
secretsSvc,
context.companyId,
{ id: pending.secretId },
pending.secretName,
pending.accountHomeDir,
);
return commit();
});
},
},
grok_local: {

View File

@ -163,6 +163,16 @@ export interface CredentialPromotionContext {
*/
export interface CredentialPromotion {
promote(authBytes: Buffer, context: CredentialPromotionContext): void | Promise<void>;
/**
* Wraps the service's terminal "authenticated" commit, immediately before
* the service runs it. A promotion that already bound and validated a
* value in `promote` can hold the SAME lock across this call, so a write
* that would invalidate that value cannot land in the gap between the
* earlier validation and this terminal commit. When the wrapper throws,
* the service records a failed login instead of `authenticated` and never
* runs `commit`. A promotion that omits this runs `commit` directly.
*/
runTerminalCommit?<T>(commit: () => Promise<T>, context: CredentialPromotionContext): Promise<T>;
}
/** The redacted lifecycle phases. Each phase carries no secret data. */
@ -1082,15 +1092,41 @@ export function createDeviceLoginService(deps: DeviceLoginServiceDeps) {
// Publish `authenticated` only when the final conditional write still finds
// the held claim. A lost write means the claim expired and the reaper
// reclaimed the row, so the service never publishes `authenticated`.
return await terminate({
const commitAuthenticated = () =>
terminate({
sessionId,
lease,
terminal: "authenticated",
reason: null,
expectedStatuses: ["promoting"],
conditionalTransition,
activity,
});
const promotionContext: CredentialPromotionContext = {
sessionId,
lease,
terminal: "authenticated",
reason: null,
expectedStatuses: ["promoting"],
conditionalTransition,
activity,
});
companyId: input.companyId,
startedByUserId: input.startedByUserId,
adapterType: input.adapterType,
};
try {
return profile.promotion.runTerminalCommit
? await profile.promotion.runTerminalCommit(commitAuthenticated, promotionContext)
: await commitAuthenticated();
} catch {
// The wrapped final check rejected the value `promote` bound earlier
// (for example, a rotation landed after that validation). The login
// did not finish, so fail closed the same way a `promote` rejection
// does: never publish `authenticated`, still delete the sandbox.
return await terminate({
sessionId,
lease,
terminal: "failed",
reason: "promotion_failed",
expectedStatuses: ["promoting"],
conditionalTransition,
activity,
});
}
}
const terminal: AdapterAuthSessionStatus =

File diff suppressed because it is too large Load Diff