Codexhost Update Impact Audit

Codex usage gate probe

This surface lets an external Harness submit while the ChatGPT-signed-in Codex subscription is out of usage by projecting false for one Composer's two Desktop usage-gate subscriptions. It depends on private React/state-library internals, so check it on every update. npm run audit:codex-desktop reports its structure as the codex-usage-gate surface (see tools/codex-desktop-contract-audit/README.md); codex-usage-account-gate-dormant means the audit ran without a ChatGPT sign-in whose Account gate reads usage. The audit does not verify behavior.

Bundle evidence (evidence only, not contracts): the Composer owner combines submitDisabled from an Account rate-limit selector (authMethod !== "chatgpt" early return, then rate_limit.allowed === false, gated to host local) and a reserve hardBlocked selector. Confirm both still exist and still feed submitDisabled, and that no new Account-wide usage blocker was added beside them. If Desktop now exempts external models/Threads or stops blocking in the Renderer, report that the module can be removed.

Manual read-only live probe, when the audit surface is not no-impact or its counts need explaining, in a Composer inside the primary Renderer. Do not patch getSnapshot, subscribe to the store, or call effect functions:

1. From the editor's nearest React host, walk committed ancestors; exactly one owner has onLocalSubmitStart and boolean submitDisabled, in both a new draft and an existing Thread.
2. In its hook list, find subscriber memo candidates with dependencies [store, atom] or Desktop 26.928's [signal, undefined], where the readonly signal exposes atom, store, get, and subscribe. The next hook has queue = { value: boolean, getSnapshot }, followed by effect { create, deps: [subscriber.subscribe] }, without atom.write, and createRender() returning nothing. An instance getter may equal the subscriber getter or be a lazy wrapper; for an unbound wrapper, both snapshots must equal the native boolean store.get(atom).
3. Replay each candidate's atom.read with tracing proxies, following readonly boolean selector indirections with cycle and work bounds. Exactly one reads hardBlocked without active or Account gate fields (reserve). When signed in with ChatGPT, exactly one reads authMethod and rate_limit.allowed without reserve fields (Account). Each replayed boolean must equal its native value; reject mixed gates.
4. store.sub, subscriber.getSnapshot, and the instance getSnapshot are writable.

With codexhost running and an external Agent selected, bound gates show an instance getSnapshot of () => false for both hooks. The Agent control's hover title shows the "cannot separate from Codex usage" message when binding failed.

Behavioral check requires a ChatGPT account that is out of Codex usage; API-key sign-in never triggers the gate. With non-empty text and an external Agent, Send is enabled in a new draft and an existing Thread; after switching to Codex it returns to disabled. Do not submit a message unless the user asks. Remove any temporary text without touching text the user typed. Report unverified when no exhausted account is available.

Completion criterion: no surface receives a live verdict from bundle inspection alone.

5. Rank hypotheses and validate narrowly

Before proposing a fix, rank 3–5 falsifiable hypotheses by evidence and name the observation that would disprove each. Run the narrowest relevant tests first, for example:

npx vitest run <affected renderer tests> --config tests/vitest.config.js
npm run typecheck
npm run build:renderer
git diff --check

For a requested fix:

1. Add a regression test that models the observed DOM, Fiber, API-shape, visual relationship, or routing boundary.
2. Run it red for the observed reason.
3. Make one minimal production change in the owning module.
4. Run it green and repeat the relevant live probe.

A test that proves only that a control exists or a mock was called does not prove visual alignment, ownership, or routing.

Completion criterion: a fix is tied to one confirmed failing boundary, not to a changed bundle name or broad suspicion.

6. Report verdict first

The first line must answer the user's question directly:

结论：无影响 / 有确认影响 / 可能有影响 / 尚未验证。

Then give 3–5 decisive facts before detailed evidence. Classify every surface as:

• no impact: bundle contract and relevant live boundary both agree;
• confirmed impact: a live or focused regression check demonstrates failure;
• possible impact: evidence changed materially but the decisive boundary is unavailable;
• unverified: the boundary was not exercised or observable.

Include:

• reviewed baseline, direct predecessor if known, and current version;
• exact files/lines consuming each affected contract;
• old/new semantic evidence and whether it is relocation, styling, relationship, or API-shape change;
• live commands/checks and sanitized outcomes;
• focused tests actually run;
• skipped or blocked checks and why;
• smallest proposed change only for confirmed impact;
• confirmation that unrelated worktree changes were preserved.

Do not bury the answer in the evidence. “Frontend changed” and “codexhost is impacted” are separate conclusions.

Common failure modes

• Comparing current against whatever old bundle is easiest to find instead of the reviewed baseline.
• Losing the reviewed baseline and downloading hundreds of megabytes on every audit.
• Treating an asset hash, minified name, marker count, or chunk relocation as compatibility impact.
• Reloading a user's active Renderer before attempting a read-only inspection.
• Treating a state-hidden control or a pixel-height change as a visual regression without checking its parent layout and role.
• Declaring Host routing, title creation, or Fork healthy without exercising that boundary.
• Updating multiple controls before isolating the failed ownership contract.
• Reformatting, resetting, staging, or overwriting unrelated work while investigating.
• Declaring the Codex usage gate healthy from a Codex-selected, empty, API-key-signed-in, or not-exhausted Composer; only the external Agent path with exhausted ChatGPT usage exercises it.
• Checking only the Send button's DOM disabled; Desktop also guards Enter and the submit function with the same submitDisabled.