Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
# 120 — wp5 remainder: #2497 security review (native-main token refresh)

PR #2497 is the last open `bug`-labelled PR. It is the credential boundary `AGENTS.md`
places under explicit security review, so it got one: an independent adversarial reviewer,
then my own verification of each finding against the tree.

Rebased cleanly onto dev (1/1). ~2,600 lines across 20 files.

## Verdict: not merged. Three security blockers plus a red suite.

### B1 — publication is not atomic (`main-account.ts:319`)

`replaceMainAuthJsonWithoutClobbering` renames the canonical `auth.json` to a backup, then
later `linkSync`s the temp into place. A crash between the two leaves no `auth.json`, and
there is no startup recovery for the `.ocx-main-auth.*.bak` file.
Comment on lines +11 to +15

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Move the unfixed security review out of devlog

Because #2497 remains unmerged and these credential-boundary fixes have not shipped, this tracked file publicly records exact vulnerable functions, failure windows, credential-adoption behavior, and replay mechanics while the defects are still open. Move the review to .tmp/ or other scratch space and publish only the fixed outcome after it ships; an open PR containing the affected code does not make a new analysis of its weaknesses post-disclosure material.

AGENTS.md reference: AGENTS.md:L115-L119

Useful? React with 👍 / 👎.


Partially defensible: the code DOES restore the backup on both failure paths it can observe
(`restoreAuthJsonBackupWithoutReplacing` at the snapshot-mismatch and link-failure branches),
and it re-verifies the published snapshot afterwards. What it cannot cover is a process death
between rename and link, and an external writer holding the old inode. The reviewer is right
that the window exists; the code is more careful than "non-atomic" alone suggests.

### B2 — same-account fallback adopts a different grant (`main-account.ts:403`) — ACCEPTED

`freshStoredCredentialForMain` first looks for the SAME refresh grant, which is correct. It
then falls back to `findUniqueFreshCredentialForChatgptAccount` — any fresh pool credential
sharing the ChatGPT account id — and writes that pool refresh token into native-main
`auth.json`. The added test blesses replacing `native-main-refresh` with `pool-main-refresh`.

This is the same hazard `anthropic-routing.ts` fails closed on: a background slot must not
have its credential adopted into an active one merely because the identity matches. Account-id
equivalence is not grant ownership.

### B3 — the "exactly one" replay is one LOGICAL replay, not one physical send

Verified: the guard is set before refresh/rebuild/relay on all three paths
(`core.ts:3460`, `core.ts:4998`, `compact.ts:620`), and the compact path explicitly uses
single-send mode (`compact.ts:647`). But the passthrough post-401 send goes through
`fetchWithTransientRetry` (`core.ts:3532`), whose ladder is 3 transient attempts x 3 reset
attempts. So one 401 recovery can be up to nine physical sends. Nothing reaches the client
twice, but upstream work can be committed more than once.

### B4 — the PR's own new suites are red on this head — REPRODUCED

`bun test tests/responses-native-main-refresh.test.ts tests/responses-compact-native-main-refresh.test.ts`
→ **1 pass / 9 fail**. The fixtures request account-gated `gpt-5.6-sol` without an
authenticated `/models` roster, so dev's entitlement gate (which landed in #2550, after this
PR was written) fails closed before the replay path is reached.

That is a stale-base artefact rather than a defect in the replay logic — but it means the
regressions this PR relies on prove nothing at the head being merged.

## Disposition: NEEDS_HUMAN, not merged

B2 is a credential-ownership decision, not a bug I should silently pick a side on: tightening
it to grant-only changes what happens to an operator who re-logged in through the pool and
expects main to follow. B1 needs a publication redesign. B3 needs the auth replay to bypass
the nested retry ladder.

Fixing all three inside someone else's 2,600-line credential PR and admin-merging it is
exactly the shortcut `AGENTS.md` §"Security working notes" and `MAINTAINERS.md` exist to
prevent. The findings are posted to the PR for the author and the maintainer; the branch stays
unmerged.

Non-blocking, recorded for the author: persistence preserves the old `id_token` while reads
prioritize its account id over a refreshed `access_token`/`account_id`, which can reintroduce
a stale account header (`main-account.ts:173`).

Loading