All notable changes to this project are documented in this file.
Dates use ISO format (YYYY-MM-DD).
This repository's current stable release line is 2.x. Full release notes live in docs/releases/ — this file is the short version. Pre-0.1.0 iteration history is archived in docs/releases/legacy-pre-0.1-history.md. Conventions: entry headings are ## [x.y.z] - YYYY-MM-DD; section headings come from Added, Changed, Fixed, Removed, Security, Internal, Notes; issue and pull-request references link as [#N](https://github.com/ndycode/codex-multi-auth/pull/N).
The runtime proxy and local bridge route native Decisions API requests to
API and ZDR credentials, with failover inside the selected pool. A successful
resets redeem now reaches a running router, including a hard pin, without a
restart. Reset-credit state moves to version 3, so upgrade the CLI and router
together.
Full notes.
Codex voice and web search work with runtime rotation on, the packaged-app router can chain a second local proxy, macOS app binds are managed through launchd, GPT-6.1 Terra is a first-class model, and API routes gain US endpoints and a PSP pool. Saved re-logins take effect in a running router, and refresh coordination fails closed. Full notes.
Regular accounts are used up before emergency backups, last-resort reset
recovery spends the soonest-expiring credit first, resets list --refresh
shows each credit's expiry, and cached rate limits are rechecked after an
hour so pinned accounts recover promptly.
Full notes.
Proxy environment variables reach every outbound request, status defaults
to a compact view with a live codex-ma web dashboard, resume --last
keeps the right provider, and native WebSocket forwarding recovers safely
from upstream drops.
Full notes.
Native Codex dictation rides the managed OAuth pool, sequential scheduling honors saved account order within priority tiers, and API/ZDR routes get daily capability checks plus WebSocket disconnect diagnostics. Full notes.
Tracking release: merges the community-submitted client_version fix
for OAuth model discovery. Runtime behavior is identical to 2.20.0.
Full notes.
- #734 independently
diagnosed and fixed the same upstream
client_versionrequirement as #735 (which shipped in2.20.0). Both PRs resolved to the same shipped implementation, so the merge records the contributor's work without changing behavior.
Luna Reserve arrives as a first-class routing fallback, direct model discovery self-heals its now-required client version, and usage rows stop losing token counts on headerless SSE streams. Full notes.
- First-class Luna Reserve (
gpt-reserve) support: registered as a Luna-compatible model with the Luna reasoning ladder and tool capabilities, parsed from the official multi-bucketrateLimitsByLimitIdresponse bylimitName: "gpt-reserve", and reported through an additivelunaReserveobject inlimits --json --refresh(offered/available, used/remaining %, window duration, reset timestamp, limit ID, normal Luna slug). An absent bucket reports unknown rather than 0%, and a genuine quota 429 on a Luna model falls back togpt-reservewithout spending reset credits or feeding ordinary quota bookkeeping (#732)
checkOAuth model discovery no longer fails with HTTP 400 on fresh installs: the now-requiredclient_versionquery parameter falls back to the installedcodex --versionwhen the inventory has no stored version, and the detected value is persisted for the next check (#733, #735)- Responses SSE bodies without a reliable
Content-Typeare sniffed via a bounded 4 KiB prefix instead of being parsed as one JSON document, so usage-ledger rows carry real token counts and stream completion checks still run (#731)
Light-terminal readability plus a full dependency and toolchain refresh:
the login dashboard gains a terminal background mode
(auto/dark/light), all five open Dependabot advisories are
resolved, and the runtime now requires Node.js 22.19.0 or later.
Full notes.
- Node.js floor raised to
>=22.19.0(was>=18.17.0): undici 8 and Vitest 5 both require Node 22, and Node 20 is end-of-life. CI now tests on Node 22.x and 24.x (#730) - Runtime dependencies:
hono4.13.9 → 4.13.13,ws8.21.3 → 8.22.0,zod4.4.3 → 4.6.5,undici6.29.0 → 8.11.2 (#730) - Dev toolchain: Vitest 4 → 5.0.2, ESLint 10.0 → 10.12,
typescript-eslint 8.56 → 8.71.1, lint-staged 16 → 17.6, fast-check
4.5 → 4.10,
@types/node20 → 26. TypeScript is split deliberately:typescript6.0.3 keeps the JS compiler API (eslint, tsserver, tests) whiletsgo(typescript 7.0.2, native) runsbuildandtypecheckvia an explicit path (#730) - GitHub Actions pins moved to current SHAs: checkout v7, setup-node v7, upload/download-artifact v7/v8, codeql-action v4.38.2 (#730)
ProxyAgentpinsallowH2: falseso undici 8's H2-by-default ALPN negotiation keeps the upstream wire protocol at HTTP/1.1 (#730)
- Dashboard title, account labels, section labels, and help text are
readable on light terminal themes: a persisted
uiColorModedashboard setting (auto/dark/light, defaultauto), aCODEX_TUI_COLOR_MODEprocess override, andCOLORFGBGauto-detection pick light-mode foreground tokens across truecolor, ansi256, and ansi16 profiles (#729) - Muted prompt text no longer stacks the
dimattribute on the muted foreground in light mode, which previously made secondary text faint again (#729) - Light-mode ansi16
muteduses normal black: bright black (ansi16(90)) maps to mid-gray in common palettes, which was the original contrast failure (#729)
- All five open Dependabot advisories resolved,
npm auditclean:source-map-js<1.2.2 (high, indexed source-map event-loop DoS),brace-expansion4.x <5.0.12 (quadratic-expansion ReDoS),vitest/@vitest/mocker2.1.0–4.1.11 (redirect-mock path traversal / arbitrary file read), and@humanfs/node<0.16.8 (recursive copy follows symlinks outside the source tree) — all dev-graph only (#730) honostays pinned at 4.13.13,brace-expansionoverride at 5.0.12, andundiciis now the exact-pinned8.11.2(seeSECURITY.mdfor rationale) (#730)
Subscription visibility and launch performance: status, list, and
check accounts report the selected workspace, observed plan, and first-use
priming state per account, with opt-in priming for unused Pro Max
subscriptions — and the shadow Codex home no longer reads multi-gigabyte
linked SQLite files into memory on every launch and cleanup.
Full notes.
- Per-account subscription status on
status(text and--json),list, andcheck accounts: selected workspace, observed plan label,autoPrimesetting, first-use priming state, and observation freshness (#726) check accounts --primeopt-in first-use priming for unused personal subscriptions; automatic priming forautoPrimepolicy accounts now recognizespromax/Pro 500 (#726)byWorkspacequota-cache namespace keyed by hashed record+workspace identity so organization members sharing anaccountIdcannot collide; token-only accounts derive workspace identity from the access token (#726)
- Linked shadow-home SQLite files (main database plus
-wal/-shmsidecars) are no longer read into memory for launch snapshots or sync-back — the shadow path is the original file itself (#727) - Forked rollouts index and display by their first
session_meta; the parent's appended metadata no longer hijacks the rollout id (#727) - Concurrent quota-cache writes no longer lose priming outcomes:
primingObservedAtmerges evidence independently of the quota timestamp, an overlapping timeout cannot erase a completion, and a genuinely newer observation still retires stale evidence (#726) - Strict policy reads reject malformed
autoPrime/paused/drainedrows; load failures surface asunknowninstead ofoff(#726) - Disabled or invalid workspace selections never fall back to a sibling workspace or the stored binding, including token-only accounts (#726)
undici6.28.0 → 6.29.0: high-severity fixes for WebSocket permessage-deflate decompression DoS, response splitting via the retry interceptor, and unrequested WebSocket subprotocol DoS
Model-catalog update tracking upstream openai/codex@b1e72963: GPT-6.1 Sol
becomes the default model, the retiring gpt-5.5 family migrates to its named
replacements ahead of the 2026-10-14 OAuth shutdown, and the leaked
gpt-6-astra-aeon slug retires onto the flagship.
Full notes.
gpt-6.1-solsupport across the model map, resolver, reasoning-effort ladder (low–ultra, defaultlow), capability profile, pricing card, config templates, and the standalonescripts/codex.jswrapper.gpt-6.1-solis nowDEFAULT_MODEL, mirroring upstream commitb1e72963which made it the catalog's priority-1 model.
- Retired
gpt-5.5andgpt-5.5-pro(including dated snapshot ids): they now resolve togpt-6-solandgpt-6-astrarespectively, matching OpenAI's migration targets ahead of the 2026-10-14 OAuth shutdown. - Retired
gpt-6-astra-aeon(never a durable catalog id): resolves togpt-6-astra. - Unsupported-model fallback chain re-floored away from
gpt-5.5:gpt-6.1-sol → gpt-6-sol → gpt-5.6-sol → gpt-6-luna, with the pro tier steppinggpt-6-astra → gpt-6-solfirst. gpt-5and unknown GPT-5-ish requests now land ongpt-5.6-sol;gpt-5.4routes togpt-6-sol,gpt-5.4-mini/nanotogpt-6-luna, andgpt-5.4-protogpt-6-astra.- Pricing for retired ids is resolved through the replacement model's rate card.
- Context-window estimates table emptied deliberately: models with
uncertain Codex-vs-API limits (
gpt-6.1-sol,gpt-6-*,gpt-5.6-*, Daybreak) returnnullunless a user override is supplied, rather than inventing a window.
- Usage pricing resolves bare aliases (
gpt-6.1,gpt-6,gpt-5.6,astra) and effort-suffixed ids through the exact alias map, so raw ledger rows price at the canonical card instead of failing amaxCostUsdbudget closed. - Wrapper forward bound raised to 5 spawns so the deepest supported-model
staircase (
gpt-6.1-sol → … → gpt-5.6-luna) completes on the last allowed forward, matching the plugin host's outbound-attempt budget.
Eighteen-PR review-and-hardening stack: OAuth redaction and boundary-trust fixes, cross-process CAS for governance stores, clock-skew-safe merges, credential-signature backup rotation, V1 migration preservation, reset-backend discovery for the relocated ChatGPT layout, plus perf, packaging, property and chaos test coverage. Full notes.
- Close OAuth log-redaction gaps in body and URL scrubbers — bearer ordering, fragments, nested query values, encoded separators, token-like hostnames, JSON key names, chunked entropy scrubbing (#719)
- Enforce trust at library boundaries: audit-log
0600re-assertion, malformed token-digest rejection, Windows root-relative override refusal,AbortSignal.anylistener-leak fix (#710) - Close regex DoS on overlong email labels and timer-overflow near-zero failover clamp (#721)
- Cross-process CAS for JSON governance stores under the directory lock (#713)
- Hybrid
updatedAtfloor so clock skew cannot lose writes (#723) - Backup rotation keyed by credential signature; chained pending-auth healing (#712)
- V1→V3 migration preserves hybrid-era fields and per-family indexes (#717)
constructor/__proto__keys safe in budget and routing merges (#714)- Wrapper signal relay and dead-owner shadow-home sweep safety (#720)
- Active-account sync fingerprints canonical plus mirror files (#711)
- Relocated ChatGPT backend discovery; ambiguous reset identities rejected (#724)
- Linear account dedup via incremental identity indexes (#708)
- Cached runtime policy stores with fingerprint validation (#709)
- Fewer transform allocations; non-throwing breaker probe (#716)
- Ship
skills/; livebrace-expansionpin via self-ref override (#722) - Smaller tarball, faster typecheck, slimmer forward-path imports (#711)
- Property coverage for storage/transformer/SSE/budget/redaction (#714)
- Real fs/net fault injection and live concurrency chaos suites (#715)
- Living documentation rewritten from audited architecture (#718)
- Release hygiene: v2.17.1 notes restored, links re-aligned (#707)
Version resync over 2.17.1 — no functional changes. 2.17.1 was prepared but never published; the fixes below shipped in the published 2.17.2 artifact. This release re-aligns the package version, changelog, and release-notes filenames. Full notes.
Automatic subscription priming observes each selected workspace on its own record, shared-organization accounts no longer share one quota entry, and concurrent quota writes merge instead of clobbering. App-bind cleanup now works on non-English POSIX systems. Full notes.
- Automatic first-use checks prime each account's selected workspace independently; quota observations are keyed by credential record and exact workspace, so shared-organization accounts no longer collide and a Personal-workspace check no longer deletes a valid organization entry (#706)
- Quota cache saves merge against a pre-edit baseline under the file lock, so an older session's save can no longer discard another process's workspace observation (#706)
- Pending rotated credentials carry
recordId, keeping the account's identity key stable through a refresh that only persisted through the journal (#706) unbind-appand helper cleanup now stop owned processes on non-English POSIX locales; theps -o lstart=identity probe pinsLC_ALL=C, so it parses on every system instead of failing and leaking live routers
Codex CLI 0.156+ no longer refuses org-sourced accounts, and an opt-in native desktop binding adds capability-aware routing, API/ZDR pools, reset credits and Responses WebSockets. Full notes.
- With runtime rotation on, interactive
mcodex resumelists sessions across providers, including desktop sessions saved underopenaithat Codex's own picker hides. Explicit ids,--last, remote and non-interactive invocations keep native behaviour, and a failed discovery falls back to the native picker. Needs a Codex version with app-serverthread/list(#699) rotation bind-app --nativekeeps the desktop app on its real OpenAI login (Remote Control pairing keeps working) while inference routes through the account pool. Requirescli_auth_credentials_store = "file";--custom-providerandunbind-approll it back (#702)- Native mode routes each request only to an account and workspace whose catalog advertises the requested model, effort and speed, serves Responses over WebSockets, and schedules subscriptions by a 5% reserve, priority tier and earliest reset. Catalog outages never block inference (#705)
- A
switchpin stays strict in native mode: tiers, the reserve and automatic reset redemption apply only when nothing is pinned, and a model the pinned account doesn't advertise is refused with 403model_not_available_in_account_catalog(#705) login --apiadds API and ZDR credentials as explicitapi/<model>andzdr/<model>pools that fail closed and are never a fallback for subscription requests.api-routes.jsonstores the keys in plain text (mode 0600 on macOS/Linux; Windows relies on your user profile's ACLs) (#705)account priority <index> <0..9>(default 1) andresets list|redeem|autofor earned subscription reset credits;resets auto last-resortredeems only when every eligible subscription is blocked,manual(the default) never does (#705)- Subscription priming, opt-in only:
check --prime, or per-accountaccount auto-prime <index> on|off(default off, every 15 minutes while a router runs). Priming uses subscription quota (#705) check accounts|resets|capabilitiesruns one part ofcheck(#705)status --jsonaddsapiAccounts,totalAccountCount,selectionMode,modelInventory, and per-account priority,autoPrime,lastInferenceRequestAt, reset-credit and forecast fields (#705)fix --liverebinds an org-sourced account the backend no longer authorizes, reportsrebound-unauthorized-workspace, and resyncs~/.codex/auth.jsonwhen the active account changed (#701)
- New logins pick a unique Personal workspace over an organization default.
With several workspaces and no unique Personal one, an interactive login asks
and a non-interactive login keeps the automatic choice and warns with
--org(#705) loginchecks the chosen workspace againstwham/accounts/check(10-second timeout, fails open). An unauthorized automatic choice falls back to the backend's default; an explicit--orgis saved as chosen butauth.jsongets the authorized default (#701)- Plain
checkalso refreshes reset credits and model discovery. Billable API capability probes reuse results younger than 15 minutes; onlycheck capabilitiesforces them (#705) checkrejects unknown arguments with exit 1 (2.16.0 ignored them), so scripts passing extra flags tocheckneed updating (#705)ws(8.21.3) is a new runtime dependency. Downgrading below 2.17.0 drops account priority tiers on the next policy write; seedocs/upgrade.md(#705)
~/.codex/auth.jsonno longer gets anorg-...id astokens.account_id, which Codex CLI 0.156+ refuses on every request ("selected workspace missing from routing discovery"). The token'schatgpt_account_idis written instead, a stale org id from an older release is rewritten, and a switch no longer carries the previous account's workspace id onto new tokens (#703, for #700)best --live,forecast --liveandreport --livekeep an explicit or org-sourced workspace binding across a token refresh instead of overwriting it with the token's default (#701)- A refreshed token whose
accounts.jsonwrite keeps failing (Windows EBUSY/EPERM) is kept live and journaled to<accounts-file>.pending-auth.jsoninstead of being discarded after the old refresh token was already spent (#705) - On macOS and Linux, the
codex appapp-server shim in~/.codex/multi-auth/app-server-shims/runs the original Node in place instead of a hard link or copy, which current Homebrew Node could not start from, so the desktop app's app-server launch failed (#705) - A quota reset header too large to represent (e.g. a 308-digit
reset-after-seconds) no longer reads as a window that never resets: reset times are capped at 7 days ahead, andformatWaitTimeprintsunknowninstead ofNaNs/Infinitym NaNs(45b7dd20) - Windows: a
codex.cmdnpm shim on PATH is launchable when package andnpm root -glookup both miss (ad59405c) - Windows:
CODEX_MULTI_AUTH_REAL_CODEX_BINpointing atcodex.cmdor its sh shim resolves to the package'scodex.jsinstead of failing every launch (933c1ff7)
honopinned to4.13.9(2.16.0 shipped4.12.33, inside the<=4.13.4range of seven advisories including GHSA-8j4g-w8fx-2239) and the dev-onlynanoidto3.3.18for GHSA-2v37-7h3g-55p8 (#704)
GPT-6 Sol and Luna are supported, and every retired model now runs on its named replacement. Full notes.
gpt-6-solandgpt-6-lunawith the upstream catalog's reasoning ladders, published standard and Fast pricing, and config template entries. Before this, both silently ran GPT-6 Astra (#696)
- Retired models (every codex model,
gpt-5.1,gpt-5.2, thegpt-5.4family,gpt-5-mini/nano, the*-chat-latestsnapshots) now resolve to the replacement OpenAI names, are removed from the templates, and are priced at the replacement's rate. The fallback floor isgpt-5.5and the quota probe no longer tries retired models. Codex ids now use thegpt-5.2prompt family and rotation bucket (#697, #698)
- GPT-6 requests with more than 272K input tokens are priced at the published long-context rate instead of the cheaper short-context rate (#696)
status -m <codex model>reads the account family the runtime proxy actually keys, not the unusedcodexslot (#698)
Image generation and editing now run through the authenticated runtime rotation proxy, and built-in image_gen is exposed to Codex 0.154.0 sessions without native-auth fallback. Full notes.
- The runtime proxy accepts
POST /images/generationsandPOST /images/edits(with/v1aliases) and forwards request bodies unchanged to the upstream image endpoints, reusing bearer authentication, account rotation, managed OAuth refresh, and usage accounting. Image requests get a minimum five-minute upstream-header timeout and are not replayed on ambiguous transport failures (#694) - Provider configuration adds a non-secret
x-openai-actor-authorization=codex-multi-auth-localcompatibility marker, stripped case-insensitively at the upstream boundary, so Codex 0.154.0 exposes its built-inimage_gentool for runtime-rotation sessions (#695)
A long-running task no longer dies when the backend reports that the selected model is at capacity: the runtime proxy waits it out and re-sends instead of burning through the account pool. Full notes.
- The runtime proxy waits out a "selected model is at capacity" response instead
of rotating. There was no handling for this error before: it fell into the
generic
status >= 500branch, which cools the account down and rotates. That is the wrong shape, because capacity is a property of the MODEL rather than of an account, so every account in the pool fails identically, the rotation leaves server-error cooldowns on healthy accounts on the way past, the transient-attempt budget runs out within seconds and the request ends as a pool-exhausted 503. The wait backs off 2s, 5s, 15s, 30s then 60s, an upstreamretry-afterwins when one is sent, and the whole thing stops at a wall-clock deadline that covers the upstream calls too, not just the time asleep. The responding account is left completely unpenalized: its pool token is refunded and it is never marked rate limited or cooled down. A client disconnect during a wait abandons the request rather than sending another upstream call for a response nobody is reading (#692, closing #689) CODEX_MULTI_AUTH_MODEL_CAPACITY_RETRY_MSsets that deadline: 10 minutes by default, 1 hour maximum,0to restore the previous rotate-and-fail behaviour. An unparseable value falls back to the default rather than disabling, so a typo cannot silently switch the feature off (#692)
A machine-readable quota snapshot, an explicit loopback upstream for the runtime proxy, and a manual switch that actually revalidates a rate-limited account. Full notes.
codex-multi-auth limits --json [--refresh]prints a stable quota snapshot for local integrations, so nothing has to scrape terminal output or read the internal cache file. Cached mode makes no network requests;--refreshreuses the dashboard's sequential refresh and its five-minute freshness floor. The payload carriesschemaVersion: 1, aselectionblock naming the pin, the per-family active index and the resultingroutedIndex, and one row per account. Absent provider values are explicit JSONnull, access tokens and probe-model names are never emitted, and account emails are masked inlabelexactly asforecast --jsonmasks them (#688, closing #687)CODEX_MULTI_AUTH_RUNTIME_PROXY_UPSTREAM_BASE_URLroutes the runtime rotation proxy through an explicit local upstream for operators chaining it into a local inspection proxy. The value must be HTTP with a numeric loopback host (127.0.0.0/8or[::1]) and an explicit port; a name such aslocalhostis rejected because the OS resolves it at connect time and the request carries the managed OAuth bearer token. It fails closed: an invalid value, a proxy that cannot start, or runtime rotation being disabled all exit non-zero on a request-bearing invocation rather than falling back to the direct backend (#690)
switch <n>clears the selected account's rate-limit windows in the same atomic write that sets the pin. After an out-of-band quota reset the markers survived, so the runtime proxy kept treating the account as blocked and deferring or returning 503 until the recorded window expired on its own: the one command meant to say "use this account now" could not unblock it. A 429 that raced the switch is still preserved; one recorded earlier is not (#691)- The runtime proxy drops the switched account's cached quota observations when it observes a new selection generation. It previously cleared session affinity and nothing else, so a pre-switch reading kept driving preemptive deferrals against the account just selected (#691)
- A failed debounced save re-arms with backoff and a bounded retry count. The
debounce timer clears itself before the save runs, so one transient failure,
such as a Windows antivirus scanner holding
accounts.jsonopen, silently discarded every rate-limit window, cooldown and rotated refresh token recorded since the last successful write (#691) flushPendingSave()logs a failure and resolves instead of rejecting.RuntimeRotationProxy.closeawaits it with no catch, so a save failure during teardown aborted the rest of proxy shutdown (#691)- Every reader of
affinityGeneration, and the increment itself, now require a safe integer. The counter is how a running proxy notices a selection change, and past 2^53Math.max(counter, disk) + 1equals the counter, so one hand edit or corrupted write to a 16-digit value froze it: each later switch wrote the same number and no proxy ever observed another selection change. A poisoned counter is dropped, so the next write repairs it to 1
Effort suffixes resolve to one key per model, usage is priced by service tier instead of assuming standard rates, and a class of lookup defect where a prototype member name was mistaken for a model id is closed. Full notes.
- Service-tier aware pricing. Every rate in
MODEL_PRICINGis a STANDARD-tier rate and the estimator had no tier input, so a session billed at OpenAI's Fast tier was priced at half its cost and amaxCostUsdcap could overrun without tripping. The tier is read off the Responses payload (both the bare object and the stream-event shape), carried onUsageTokenCountsso it reaches the ledger through every existing path, and applied at Astra's published Fast rate of $20 / $100 per 1M. A tier with no published rate is reported as unknown cost rather than approximated, so a budget fails closed exactly as it does for an unpriced model (#686) gpt-6-astracarries a cached-input rate of $1 standard and $2 on Fast. It shipped without one in 2.11.0, which made Astra the only row in the table billing cached tokens at the FULL input rate: a tenfold over-statement on a cache-heavy session that trips a cost cap far too early. The 90% cached discount is a uniform platform rate every other row already encodes at input/10 (#686)
maxandultraeffort suffixes are stripped from a model id, sogpt-6-astra-ultrano longer keys separately from the model it routes to. Four call sites stripped onlynone|minimal|low|medium|high|xhigh, a set that predates GPT-5.6: the entitlement cache blocked the same account and model pair twice under two keys, the capability policy recorded outcomes under a key the matrix never reads, and the fallback chain found no row for an effort-suffixed model. The sharedstripModelEffortSuffixderives its set fromREASONING_EFFORTSand names the one real exception,codex-maxandgpt-5.1-codex-max, whose final segment is part of the model NAME (#685)- The service tier survives the runtime recorder and a ledger round-trip.
createRuntimeUsageRecorderrebuilds the ledger input field by field andnormalizeParsedUsageRowrebuilds only the numeric token fields, so a Fast response through the runtime proxy, the default path, priced at the standard rate, and a persisted Fast row read back as standard (#686) - A response reporting its tier as
fastis recognised. The upstream catalog names the tierpriorityand displays it as "Fast", so either spelling can arrive; unmapped it becameunknownand failed a cost budget closed on a tier that does have a rate (#686) normalizeUsageLedgerRowvalidates the service tier instead of accepting any truthy value. Every other field in that normaliser is checked against its allowed set, and the ledger read path already rejected an unknown tier, which left the write path as the only unvalidated way onto a row
GPT-6 Astra and the Daybreak cyber models are first-class, and five defects found while wiring them up are fixed. Full notes.
- GPT-6 Astra is a first-class model family.
gpt-6-astraand the long-horizongpt-6-astra-aeoneach resolve to their own canonical id with the frontier reasoning ladder (lowthroughmax, plusultra, which is rewritten tomaxon the wire exactly as upstream Codex does). Baregpt-6andastraresolve to the flagship, andnone/minimalare coerced up tolowbecause Astra does not accept them. A dedicated GPT-6 resolver claims every id the alias table does not name, including thegpt-6-astra-proplan tier, dated snapshots, and any tier OpenAI adds later, so an unrecognised GPT-6 id can never fall through to GPT-5.5. That silent downgrade is the exact failure v2.5.0 had to fix for GPT-5.6, one major version down.aeondeliberately keeps its own id rather than collapsing into the flagship: it is a behaviourally different model, not a rename. Both shipped config templates carry the two models, and thecodex-multi-auth-codexwrapper mirrors the same map, with a model-by-effort parity suite pinning the two together - The Daybreak cyber models from the upstream Codex catalog
(
gpt-daybreak-blue-latest,gpt-daybreak-red-latest, plusdaybreak-blueanddaybreak-redshorthands) resolve to their own ids. Their slugs carry neither acodexnor agpt 5token, so every resolver declined them and asking for the cyber-permissive model silently ran GPT-5.5. An unrecognised Daybreak id resolves toblue, the defensive variant, so a typo cannot upgrade a caller into the permissive one. They stay out of the picker templates, matching theirvisibility: hideupstream gpt-6-astrais priced at its published launch rate, $10 per 1M input and $50 per 1M output on the standard service tier. No cached-input rate was published, so cached tokens bill at the full input rate, which over-states cost and makes amaxCostUsdbudget trip early rather than late. OpenAI's Fast service tier is 2x standard, and no row in this table has ever carried a service-tier dimension, so a session billed at that tier is under-counted by half for Astra exactly as it already is forgpt-5.6-soland every other priced model. Nothing in this repo selects that tier:fastSessionis a local latency setting that lowers reasoning effort, not OpenAI's billed Fast mode.gpt-6-astra-aeonand both Daybreak models have no published rate and are listed inUNPRICED_ROUTABLE_MODELSrather than guessed at, so a cost budget fails closed while they are in the window- An unsupported-model fallback chain for Astra:
gpt-6-astrasteps togpt-5.6-sol, andgpt-6-astra-aeonsteps to the flagship first, since that is still GPT-6 Astra without the long-horizon behaviour. Astra rolls out org by org, so an account that is not entitled yet gets a real unsupported-model response for it. The chain is opt-in (fallbackOnUnsupportedCodexModeldefaults tofalse) and fires only on that response, so it never swaps the model out from under a request the account could have served gpt-5.6-solgains its own chain entry togpt-5.5. The chain is walked one hop at a time (index.tsreassigns the model and passes that as the nextrequestedModel), so a model with no entry of its own ends the walk. GPT-5.6 shipped without one, which would have stranded the Astra path a rung short of a model every account has. Terra and Luna stay without entries: nothing steps into them- All four new models are listed in
UNESTIMATED_ROUTABLE_MODELS. OpenAI published two different context windows for Astra on launch day, 1.05M for the API surface and 272K for Codex, and this wrapper talks to the Codex backend. SetcontextBudgetGuardModelWindowOverridesonce you know your real ceiling
-
The
codex-multi-auth-codexwrapper resolves codex model ids that its explicit list does not name, matching lib. Its final clause testedmodel === "codex"whereresolveCodexCatalogModeltests for thecodexsubstring, so ids such asgpt-6-codexresolved to nothing in the wrapper while lib resolved them to the current codex model -
The wrapper buckets Daybreak ids into the
gpt-5.2prompt family for status instead of returning no family at all. Their slug matched none ofresolveModelFamilyForStatus's branches -
resolveModelFamilyForStatusstrips the provider prefix before classifying. Every branch is astartsWith, andresolveStatusModelforwards the raw--modelvalue, soopenai/gpt-6matched none of them and status routing fell back toactiveIndexinstead of the family index. This affectedopenai/gpt-5.6-solandopenai/gpt-5.1the same way before GPT-6 existed -
gpt6, written with no separator, resolves to Astra rather than GPT-5.5.tokenizeModelIdsplits on non-alphanumeric runs, so it produced a singlegpt6token, thegpt+6pair never formed, and the id fell toDEFAULT_MODELwhile still passingcapability-policy's catalog gate. Gate and resolver now agree on it -
The
codex-multi-auth-codexwrapper retries the GPT-6 rows too. It keeps its ownWRAPPER_UNSUPPORTED_MODEL_FALLBACK_CHAIN, which shipped without them, socodex-multi-auth-codex --model gpt-6-astraon an unentitled account exited with the failure code while the plugin-host path retried, and the docs advertised the retry for both. A parity test now pins the GPT-6 rows across the two tables. Note the wrapper's retry has always been unconditional rather than gated onfallbackOnUnsupportedCodexModel; it prints the swap to stderr, and the docs now say which path gates it and which does not -
The bare-
astrabranch of the GPT-6 resolver is anchored to a GPT-6 version token or to an id with no GPT version at all. It existed for picker labels such asAstra Pro, but it also claimed any id merely containingastra, so agpt-4-astra-xwould have run the frontier model for a caller who named GPT-4 -
capability-policy's catalog gate matches the tokenizer's separator rule. It tested a hand-picked[-_\s]wheretokenizeModelIdsplits on any run of non-alphanumerics, sogpt.6-turboandgpt---5kept their raw string as the policy key while routing resolved them to a canonical model, leaving the store tracking a key no request reads. It is still narrow enough thatgpt-4o,gpt-4.5-previewandgpt.4-turbodo not reach the resolver -
Prototype member names are no longer mistaken for model ids. Model strings reach several lookup tables raw from the caller, so
constructor,__proto__andtoStringindexed them and came back truthy. Three consequences, all fixed:estimateUsageCostUsdreturnedNaNrather thannull, which silently made amaxCostUsdcap unenforceable becauseNaN >= limitis false; the unsupported-model fallback chain threwTypeError: targets is not iterableinside the request path; andmodels --model <name>emitted a row whose normalizedModel, promptFamily and capabilities were all undefined. The fallback chains are now null-prototype objects, which also stops anunsupportedCodexFallbackChainkey of__proto__reassigning the object's prototype
A pinned account gets a real bounded retry instead of a single attempt, and a new opt-in guard can pause a session before it overflows the model's context window. Full notes.
- Context Budget Guard, experimental and shipping disabled. Enable it from Settings → Experimental or with
contextBudgetGuardEnabled. It tracks how full each session's context window is from the tokens each turn actually carries (input_tokensplus the reasoning-stripped output, nevertotal_tokens) and answers the next request locally with a/compactnotice once usage crossescontextBudgetGuardHardPercent, before that request costs an upstream round-trip. A soft threshold attaches a non-blockingx-codex-context-budget-percentheader instead. Window sizes are best-effort estimates, not published facts, socontextBudgetGuardModelWindowOverridesalways wins and a model with no estimate is skipped rather than guessed at. The pause is one-shot per measurement: it drops what it recorded as it fires, so the/compactturn it asks for is never blocked by it (#681)
- A pinned request retries a transient upstream failure instead of returning
codex_pinned_account_unavailableon the first one. Every transient branch cools the account down before the next selection pass, and selection refuses a cooling-down pin, so the retry budget could never be spent. A pinned retry now waives that account's own cooldown, and only from the second pass of the same request, with 250ms/500ms/1s backoff and a cap ofmin(retryAllAccountsMaxRetries + 1, 4)upstream attempts. Rate limits, open circuits, disabled accounts and policy blocks all still stop it, the cooldown still applies to every other request and still setsretry_after_ms, and a pin already cooling down on arrival is still refused without an upstream call (#683) - A pinned request is no longer refused by the local token bucket. That bucket spreads load across the selectable pool, and a pin has no alternative account, so a drained bucket could only reject a request the pinned account could serve. Circuit-breaker admission still applies, and the bucket is neither consumed nor refunded for a pin (#682)
account_skip_reasonsnames the admission gate that actually rejected a request. The reason was re-derived from account state afterwards, which reports the first blocker it finds — so a drained bucket on a cooling-down account was reported as the cooldown, and a rejection from a half-open circuit whose probe another request had just claimed was reported as an exhausted bucket (#682)- A context-window override below 1 token is dropped instead of being stored as a window of zero.
contextBudgetGuardModelWindowOverridesvalidated positivity on the raw value and floored afterwards, so any value in(0, 1)was accepted by the schema and then silently ignored (#681)
A maintenance release with no runtime, CLI, or configuration changes: npm run typecheck:scripts failed on every clean checkout. Full notes.
npm run typecheck:scriptsbuilds generated output before it runs the compiler.scripts/codex-multi-auth.jsimports../dist/lib/codex-manager.jsandtsconfig.scripts.jsontypechecks that file withcheckJs, so on a fresh clone the script stopped withTS2307: Cannot find module '../dist/lib/codex-manager.js'. The build now lives in the script itself, matchingpack:check,coverage,bench:runtime-pathandgenerate:schema, which covers all three CI call sites and a clean local tree in one place (#680, thanks @smpark00)
CONTRIBUTING.mdlistsnpm run typecheck:scriptsin the documented local gate, now that it works from a clean checkout (#680)- The
validateCI job drops a build step made redundant by the script change, leaving per-job compile counts unchanged, andtest/ci-workflows.test.tsderives its build-ordering guard over every job inci.ymlinstead of a hardcoded pair, reading steps by YAML key position so a commented-out orif:-guarded build cannot satisfy it (#680)
A --cost budget can no longer be silently defeated by the models most worth capping, and refreshing one account from the login dashboard stops disturbing the others. Full notes.
codex-multi-auth login --account <index|email|account_id>re-authenticates exactly one saved account, leaving the active selection, every per-family rotation position, and a manualswitchpin untouched. The write is refused if the OAuth identity is not the account you named, so the previous credentials stay intact. Implies--preserve-selection, and cannot be combined with--org(#679, thanks @fnmendez)
- A
--costbudget is enforced for models with no published price instead of counting them as free. 2.9.0 fixed the guard comparing0 >= limit, but eight of the fifteen routable model ids (gpt-5.4-pro,gpt-5.5-pro,gpt-5.2-pro,gpt-5.4-mini,gpt-5.4-nano,gpt-5-mini,gpt-5-nano,gpt-5.1— everyprotier among them) had noMODEL_PRICINGentry, so their rows priced tonull, aggregated as$0.00, and a cost cap never fired: two million tokens ongpt-5.5-propassed a--cost 1cap. Usage summaries now carryunpricedRequestsand a--costlimit fails closed withcost limit cannot be evaluated. The missing rates are deliberately not guessed — a wrong figure would move the trip point rather than fix it (fe9b0d1) - A plain
loginno longer resets every model family's rotation position onto the account just signed in. Families track which account they are on independently, so one that had rotated away from a rate-limited account kept its own position; adding an account discarded that. Families that were following the global selection still follow it (fe9b0d1) - Refreshing an account from the login dashboard no longer rewrites the native
~/.codex/auth.jsonunconditionally. Refreshing account 3 while account 1 was active handed plaincodexaccount 3's credentials, silently switching accounts; the sync now runs only when the refreshed account is the one Codex is using — which also means refreshing the active account does update it, where a selection-preserving refresh previously left it on expired tokens (#679) - Refreshing a disabled account no longer returns it to rotation. The merge set
enabled: trueunconditionally, so topping up a credential silently undid a deliberate disable; a targeted refresh now preserves the flag and says the account is still out of rotation (#679)
budget set --cost and --tokens were accepted but never enforced, a dead network path counted against account health, and account removal left live sessions routed to the wrong account. Full notes.
- The usage ledger records real token counts and cost. Both the runtime rotation proxy and the plugin host read the upstream
usageobject out of the response as it streams past, socodex-multi-auth usagereports actual consumption instead of zeros (#678) - Transport failures emit a trace-correlated
proxyLog.errorline. That path never throws, so it previously produced no structured log entry at all (#677)
maxTokensandmaxCostUsdbudget limits are enforced. Every ledger row recorded zero tokens and zero cost, so the guard compared0 >= limitand never fired —budget set --cost 50allowed unlimited spend, with only--requestsdoing anything (#678)- Reasoning tokens are subtracted out of the output bucket before recording. The upstream reports
output_tokensinclusive ofoutput_tokens_details.reasoning_tokenswhile the cost estimator prices the two as disjoint, so recording both raw would bill reasoning twice and trip a cost cap early (#678) - A streaming request's ledger row is written when the stream ends rather than when the handler returns, which is the only point at which its token counts exist. A client that disconnects mid-stream still records a row, without counts (#678)
- An upstream transport failure before response headers no longer feeds the account's circuit breaker or health tracker. A socket reset, TLS drop, or fetch timeout is a property of the network path, not of the account, and three of them used to open that account's breaker for ~30s during an outage that affected every account equally (#677)
- Deleting an account, restoring a backup, and the automatic removal of revoked-token accounts bump
affinityGeneration. Session affinity is keyed by account index, so without the bump a running proxy kept routing in-flight conversations to whatever account slid into the vacated slot (#678) - Two accounts with neither an
accountIdnor an email no longer share one account-policy entry. Both hashed the literal"unknown", so pausing, draining, or tagging one applied to all of them; the key now falls back to the refresh token and remains independent of account position (#678) withStreamingFailoverrespects consumer backpressure. It enqueued every chunk as fast as the upstream delivered it, so a slow client buffered the whole response in memory — a 50-chunk stream was fully drained for a consumer that had read one chunk (#678)- A transport failure records a runtime skip reason for the account it happened on, instead of clearing the skip-reason overlay that
forecastandreportread (#677)
codex "your prompt" took the slow shadow-home startup path, prompt text after -- was parsed as configuration, and the pinned-account 503 called every recovery deadline a limit reset. Full notes.
- A root TUI launch carrying the optional initial prompt (
codex "your prompt") keeps the canonicalCODEX_HOME. It was classified as a subcommand launch and took the shadow home, whose omitted SQLite state made Codex rebuild its whole thread index from rollouts before submitting the prompt you had already typed (#674) - A prompt forced with
--is treated as a prompt, socodex -- execopens a session whose prompt is the word "exec" rather than dispatching the exec subcommand — matching how Codex itself parses it (#674) - The launcher no longer reads flags out of prompt text:
--account,-m/--model, and--config=written after a--are left alone instead of being stripped, treated as settings, or rewritten in place on a retry (#674) - Wrapper-injected options are placed on the option side of
--instead of appended. Appending put them in the subcommand's positional list, socodex exec -- ...failed withunexpected argument '-c' found,codex sandbox -- echo hipassed the pair to the sandboxed command, andcodex mcp add t -- echo hiwrote it into that server's storedargsinconfig.toml(#674) - The pinned-account 503 words its recovery deadline by the blocker actually holding the account — circuit breaker, cooldown, or rate limit — instead of calling every deadline a recorded limit reset, which made provider outages read as blown subscription quotas (#676)
- That wording is taken from the pin's live runtime state rather than the retry loop's selection verdict, so the common case where the pinned account 429s mid-request no longer reports the internal token
already-attemptedwith a deadline attributed to nothing (#676) - The quota phrasing is used only when the rate limit is what bounds recovery; when a cooldown or breaker ends later, the deadline is worded neutrally (#676)
docs/development/ARCHITECTURE.mdanddocs/development/CONFIG_FLOW.mddescribe the interactive-TUI branch as covering prompt-bearing launches, matching the routing the wrapper now performs (#674)
forecast --model reported an account ready while every request to it failed, and the pinned-account 503 told forced pins to run a command that clears nothing. Full notes.
forecast,best, andreportwith--modelcheck that model's prompt family instead of a hardwiredcodexfamily. An account held down by a live limit for the family you asked about no longer reportsreadywith no reasons (#670)- Those commands no longer report a delay caused by a rate limit on a different model in the same family — availability now uses exactly the two keys account selection consults (#670)
- When a family-wide and a model-specific limit are both active, the reported wait is the later one rather than the earlier, so the stated time is actually usable (#670)
- A recorded
rate-limitedstate backed by a live limit for the requested family is no longer discarded as stale (#670) - The pinned-account 503 no longer advises
unpinfor a pin set by--account/CODEX_MULTI_AUTH_FORCE_ACCOUNT, whichunpincannot clear. Forced pins are told to relaunch (#671) - That 503 now reports when the account recovers, taking the latest of its rate limit, cooldown, and circuit-breaker deadline (#671)
- A permanently blocked pin — disabled, no enabled workspace, invalidated token, policy block — reports no recovery time instead of a deadline that expires into another 503 (#671)
- That suppression now also holds when the request itself is what disabled the account, which previously surfaced as
already-attemptedand hid the block (#671) - A corrupt timestamp in stored account state no longer turns the pinned 503 into a generic 500 carrying no account index, reason, or skip map (#671)
pin_source,reset_at, andretry_after_msoncodex_pinned_account_unavailableresponses (#671)
docs/reference/error-contracts.mddocuments the new fields, both pin kinds, and howretry_after_msdiffers between the pinned and pool-exhausted codes (#671)
Background rotation helpers never shut down — 183 of them at 5.58 GB on one machine, the oldest 33 hours past a 12-hour timeout. Full notes.
- Helpers no longer treat a reused process ID as a live launcher. Owner identity is now process ID plus start time (#663, #664)
- Each helper writes its own status file instead of all of them overwriting one, so
rotation statusno longer reports a random helper (#663, #664) - Helper metadata is cleaned up on exit and swept on the next launch, instead of accumulating forever — 701 leftover files on one machine (#663, #664)
- Helpers left behind by a short-lived command are reaped in 15 minutes rather than 12 hours (#665)
- That reaper no longer kills a live
codex appsession that is simply idle between messages (#669) rotation unbind-appno longer strands owner files it can't pair with a status record (#666, #669)rotation statuscan no longer name one helper on its status line while marking a different helper's account as current (#667, #669)- A stale status file with a reused process ID no longer counts as a running helper (#667, #669)
- Malformed process IDs in status files are rejected instead of being probed (#667, #669)
CODEX_MULTI_AUTH_APP_ROTATION_MAX_LIFETIME_MS— hard cap on helper lifetime, default 24h,0disables (#664)CODEX_MULTI_AUTH_APP_ROTATION_DETACHED_IDLE_MS— shorter timeout for a helper whose launcher died and that never served a request, default 15m,0keeps the full idle timeout (#665, #669)
- Helper status files are now per-process:
runtime-rotation-app-helper.<pid>.json. The old shared path is still read but no longer written (#664) - The stale-metadata sweep runs after the helper spawns, so it no longer delays
codex appor TUI startup (#669) rotation unbind-appstops helpers in parallel instead of one at a time (#669)
- Removed the deprecated
@openauthjs/openauthdependency. PKCE is now generated directly fromnode:cryptoper RFC 7636 (#661) - Test-only failure hooks in the published wrapper now require an explicit opt-in, so a stray environment variable can't arm one (#668, #669)
codex app-server didn't work at all with rotation on, which is the default. Full notes.
app-serverrefused to start on any machine that had run one before, and served a frozen thread list when it did start. It now uses your real Codex home rather than a mirrored copy (#659, #662)- The CLI shim
app-serverdidn't need was settingCODEX_MULTI_AUTH_RUNTIME_ROTATION_PROXY=0on everything the server spawned, so nested commands ran unrotated and billed the wrong account (#662) - A short-lived
app-serverleft a helper process running for 12 hours after exit (#662) - A helper that couldn't start surfaced as an unhandled-rejection stack trace, leaking a temporary directory. It's now a one-line error and exit 1 (#662)
- A rotation proxy that can't start now fails
app-serveroutright rather than silently running it unrotated — a wrong-account bill is invisible, a failed start isn't (#662)
- Quota windows with no label, duration, or reset time are hidden instead of rendering as a permanent
100%(#657)
Five fixes across login, uninstall, diagnostics, and quota. Full notes.
- Manual and incognito login now show the complete authorization URL, so pasting the callback back no longer fails with a state-mismatch error (#652)
- Uninstall instructions pointed at the old scoped package name, which can't remove what's installed (#653)
why-selectedshowed a winning account the rotation proxy would never have picked. It now runs the same gates as the real thing (#654)- Accounts with a revoked token are marked
token-invalid — re-login neededand stay out of rotation instead of rejoining after a cooldown (#655) - Exhausted weekly and monthly quotas wait for the real reset instead of retrying after a flat two hours and hitting an immediate
429(#656) - Rotation helpers verify they own a process before shutting it down, so an unrelated process is never mistaken for one of ours (#658)
mcodex resumeandforkhung on a blank screen. They were using a temporary copy of your Codex home that deliberately omits the session database, so the thread you asked for genuinely wasn't there (#647, #648)- The shell prompt didn't always come back after an interrupted Codex exit (#648)
--helpno longer starts a rotation proxy and leaves a background helper behind (#648)
hono4.12.21 → 4.12.33: JSX context leaking between requests, XSS via thecx()escaping bypass (#650)undici6.25.0 → 6.28.0: Set-Cookie injection, WebSocket fragment DoS, keep-alive queue poisoning, SameSite downgrade. Staying on 6.x deliberately — 7.x would raise the Node floor to 20 (#650)
- macOS stopped asking for the login keychain on every launch. The official CLI reads
cli_auth_credentials_storefromconfig.toml, and we only wrote it during a login or switch — so a third-party front-end running the official binary kept hitting the keychain (#641, #642) - Interactive sessions no longer reindex your session history on every launch. They run against your real Codex home instead of a temporary copy (#639, #643)
doctor --fixcan repair the credential store directly, and reports correctly on a machine with no accounts yet (#642)
cli_auth_credentials_storeis written on first run and re-checked on startup, not just during a login or switch. Opt out withCODEX_MULTI_AUTH_ENFORCE_CLI_FILE_AUTH_STORE=0(#642)- Two interactive sessions can now run at once against the same state, the same as running the official CLI twice (#639)
A bug hunt. Two of these could route to the wrong account or stall refreshes machine-wide. Full notes.
- A manual pin could route to a different account than the one you pinned. Pins are held by position, and account removal only remapped the active index — so removing a revoked account silently shifted your pin (#638)
- One failed refresh blocked refreshes for every process for 20 seconds, because failures were cached alongside successes (#637)
- A refresh waiting on the shared lease could be evicted and start a second refresh for the same token, which fails with
invalid_granton a healthy account (#637) - Monthly (Codex Business) windows were labelled
5h(#635, #636) - An account at 100% used with no reset time could still be recommended as your best option (#637)
- Exhaustion was decided on a rounded percentage, so
99.6%used benched an account that still had quota (#637) importand legacy migration dropped your pin and reset the affinity counter (#637)- Session trimming cut the leading instructions it had just decided to keep (#637)
- A timed-out request never got its token back, gradually starving the bucket (#637)
- Project- and profile-scoped budgets never applied — stored under a normalised key, looked up with the raw one (#637)
checknow shows when each quota window resets, not just how much is left (#633)
- Quota percentages kept their red/yellow/green colouring once a reset time was appended (#633)
- OAuth login on WSL. No browser ever opened, because WSL reports itself as Linux and
xdg-openisn't installed there. It now opens the Windows browser (#630) - The manual-paste clipboard copied to the distro's clipboard rather than the Windows one you paste from (#630)
- A lost OAuth callback now explains the Windows/WSL port conflict instead of just timing out (#630)
- Diagnostics probe with GPT-5.6, falling back for accounts without access (#627)
- GPT-5.6 now appears in older Codex builds' model pickers (#626)
- Upgrades no longer miss newly shipped models — the config installer merged the provider block shallowly, so an upgraded config never gained anything new (#626)
- The wrapper understands GPT-5.6. It re-implements the model map and never got the v2.5.0 work, so
gpt-5.6-*requests silently resolved togpt-5.5(#626)
pidOffsetEnableddefaults on, so parallel agents spread across accounts instead of all picking the same one and cascading into429s (#628)
- GPT-5.6 (Sol, Terra, Luna) as first-class models, plus the
maxandultrareasoning tiers abovexhigh.ultrais rewritten tomaxbefore the request is sent, matching upstream
gpt-5.1-codex-maxis a model id, not Codex at max effort — it was nearly rerouted onto plaingpt-5.1-codex- Unrecognised
gpt-5.6-*ids silently ran GPT-5.5
codex-multi-auth-codex --account <index|email|id>forces one invocation onto a single account. Ephemeral (never touches your saved pin) and fail-hard (never silently uses a different account) (#623, #624)
Nine fixes across rate limiting, refresh, streaming, and storage. Full notes.
- A single bad
retry-aftercould bench an account for 31 years. Windows are now clamped to 7 days (#617) - A slow refresh could delete a lock it no longer owned, leaving two processes refreshing at once — and since refresh tokens are single-use, logging the account out (#617)
- A routine save could revert a freshly rotated token, permanently breaking that account's next refresh (#617)
- A transient
429benched an account for hours by folding the healthy weekly window into the retry delay (#617) forecastrecommended a strictly worse account, because it took the longest of both quota windows rather than the binding one (#617)- Upstream streaming failures were reported to the client as success, with the account recorded as healthy and retry suppressed (#617, #618)
- The SSE parser required a space after
data:, so a spec-valid line parsed as zero events (#617) - The V1→V3 storage migration discarded migrated account bodies, so a rate-limited account read as available and could burst
429s (#619) - OAuth
expires_inaccepted zero and negative values, minting an already-expired token and driving a tight refresh loop (#619)
- An orphaned app-bind left
config.tomlpointing at a dead proxy with no CLI way out.rotation unbind-appnow self-heals without a backup, androtation statusreports "bound but unmanaged" instead of "not configured" (#614)
codex-multi-auth historylists every local Codex session regardless of provider.codex resumefilters by the provider recorded in each session, so rotation hides sessions created under the native provider — the files were always there (#612)
- A permanent
503that wedged rotation even with healthy accounts. Recovery reloaded from disk, restoring the same state that had wedged the pool, and the guard refused to run in exactly the situation it existed for (#606, #607) - Two cooldown paths changed account state without scheduling the write, so a restart dropped it (#608, #609)
- Streaming stalled indefinitely for slow clients, buffering without bound
- Account deduplication needed more than one pass; it now loops until stable
- Atomic writes use
crypto.randomBytesfor temporary filenames instead ofMath.random()(#517)
- Opt-in
sequentialscheduling drains one account before moving to the next, so quota windows stagger instead of resetting together (#509)
- Drain-first mode advanced its pointer on a transient fallback, breaking the invariant it exists for
- An expired token could be forwarded after a successful refresh, producing a
401and a wrong invalidation cooldown - Flagged-account restore dropped workspaces, so a multi-workspace account lost its list permanently
- A transient disk failure during quota-cache reload wiped every other account's data
- Login reported
Added accountwhen it had updated or rebound an existing one (#512) login --manualreported every failure asCancelled, including a malformed URL and a state mismatch (#512)
- CI workflow steps pinned to exact commit SHAs (#519)
- Response headers under our own namespace are blocked by prefix, so a header added later is blocked by default (#546)
forecastreported working accounts as unavailable.token-exhaustedhas no natural expiry and nothing cleared it, so it lingered indefinitely — a successful request now clears it
- Wrapper-launched sessions repair
session_index.jsonlafter known thread-store write damage, serialised so concurrent sessions can't collide - App-bind status warnings resolve the active status path before printing guidance
- The global
codexbin, which collided with the official Codex npm, native, and Homebrew installs.codex-multi-authandcodex-multi-auth-codexare unchanged
usage,account,budget,bridge,integrations,models, andmonitorwere falling through to the official Codex CLI instead of being handled locallybridge token list --jsonreturned invalid JSON on an empty store, and included token hashes
- Local-only usage ledger, account policy metadata, routing profiles, and budget guards
- Model and account capability matrix, with policy evaluated before account selection
- Optional loopback-only bridge with
/health,/v1/models, and/v1/responses rotation reset-rate-limitsto clear stale runtime timers
2.0.1 - 2026-04-25
- Runtime rotation is enabled by default. Opt out with
rotation disable,codexRuntimeRotationProxy=false, orCODEX_MULTI_AUTH_RUNTIME_ROTATION_PROXY=0
- Install and update self-heal the packaged app bind and launcher routing
2.0.0 - 2026-04-25
- The runtime rotation proxy: failover across accounts mid-session, without restarting Codex. Loopback-only with an authenticated local client key
rotation enable|disable|status|bind-app|unbind-app, with app binding fully reversible
- The proxy left a stale
content-encodingheader on responses it had decoded
1.3.2 - 2026-04-24
- Interactive forwarded sessions were capturing output, which broke TTY use. TTY sessions now inherit terminal stdio; non-TTY runs still capture so model-fallback can inspect errors (#437)
1.3.1 - 2026-04-24
- GPT-5.5 and GPT-5.5 Pro as first-attempt models, with deterministic fallback for older runtimes and accounts without access (#435)
- Model fallback now only triggers on a real upstream unsupported-model or no-access response (#435)
1.3.0 - 2026-04-17
why-selectedexplains which account rotation picked and why (#410)verify --pathsfor path-safety inspection (#410)- A feature-flagged routing mutex for safer concurrent selection, off by default (#412)
- Post-audit hardening across OAuth URL redaction, redirect-URI handling, selector null handling, short-
429retry ordering, recovery writes, and malformed SSE warnings
- Codex CLI sync: target detection, import adapters, named backup wrappers, and sync orchestration. Codex CLI state stays mirror-only (#72)
- Cleared accounts could revive themselves when the initial delete partially failed (#71)
0.1.7 - 2026-03-03
- The first consolidated
codex-multi-authrelease, bringing runtime, TUI, account management, and docs into one package (#4, #14) - Rotating backup fallback, and automatic promotion of a real backup when primary storage holds fixture data (#29)
- Hardened Windows command routing across
codex.bat,codex.cmd, andcodex.ps1(#27)
- Codex CLI account switching stabilised, and storage identity fixed across worktree branch changes (#27, #28)
0.1.6 - 2026-03-03
- Updates could reset your accounts when storage was only reachable through recovery artifacts
- Codex CLI sync wrote auth to a different profile directory than
CODEX_HOME - Account switches now fail fast when the required Codex auth write doesn't complete
0.1.5 - 2026-03-03
- A Windows crash (
UV_HANDLE_CLOSING) on wrapper shutdown, after the command had already completed
0.1.4 - 2026-03-03
- Stuck refresh lanes and duplicate refresh churn, via token normalisation and stale/timeout recovery
switch <index>keeps local selection deterministic and reports sync failures clearly
0.1.3 - 2026-03-03
switch <index>reported failure when only the optional Codex host-state sync was unavailable — the local switch had succeeded
0.1.2 - 2026-03-03
- Rotating backup recovery across
.bak,.bak.1, and.bak.2, staged then committed so a failure can't leave a partial history chain - Startup cleanup for orphaned staging files from interrupted writes, which could otherwise sit around holding token material
- Windows home resolution now tries
USERPROFILE→HOME→HOMEDRIVE+HOMEPATH→homedir()
0.1.1 - 2026-03-01
- The OAuth callback is pinned to
127.0.0.1:1455rather thanlocalhost, which never arrived on machines resolving IPv6 first - A clean global install could fail at startup; the
codexbin now lazy-loads the auth runtime
- Account emails dedupe case-insensitively
- Per-project storage is shared across linked Git worktrees
- The settings hub was split into five focused modules;
Qnow cancels without saving
0.1.0 - 2026-02-27
- The first stable release: multi-account OAuth for Codex
- The
codex auth ...command family for login, switching, diagnostics, and reporting - Dashboard settings hub and backend reliability controls
- Rotation and resilience modules for refresh, quota deferral, and failover
Pre-0.1.0 iteration history is archived in docs/releases/legacy-pre-0.1-history.md.
Native catalog discovery, login synchronization, and runtime rotation now handle recent Codex CLI changes safely. The release includes fixes from #765, #766, and #767. Full notes.
- Rotation provider arguments are inserted before the command separator, so forwarded native CLI arguments remain intact. (#765)
- Native login changes are adopted by running state without losing account synchronization. (#766)
- Catalog probes use the installed Codex version, recover from rejected probes, and retry shared requests without leaking stale version state. (#767)
GPT-6.2 model support and runtime-helper lifetime hardening are now available in the latest release. The release includes fixes from #769 and #770. Full notes.
- GPT-6.2 Sol and Astra model families are recognized across model resolution, request transformation, context limits, pricing, usage redaction, catalog fixtures, and the runtime wire path. (#769)