Files
magpie/internal/provider/codex_switch.go
T
forever 8005f17551 provider, codex: a ChatGPT account at 100% with credits left keeps Codex signed in (#966)
* provider, codex: a ChatGPT account at 100% with credits left keeps Codex signed in — magpie becomes its provider only while the Codex app itself holds the account

magpie made itself Codex's provider (model_provider = "magpie", #540)
whenever any of the active ChatGPT account's windows read 100%. A Pro
account at 100% on its week with credits left keeps working in the Codex
app: the backend answers every turn (about 490 200s in one day's usage log
at 100%), yet magpie moved Codex off it, and as a provider without
requires_openai_auth the app lost its ChatGPT state ("Sign in to ChatGPT
to start a durable thread", no remote control).

CodexUsedUp now reads codexHeld, which mirrors the Codex app's own
composer block (ChatGPT.app webview bundle, hP and its helpers): the
backend says the account isn't allowed (rate_limit.allowed = false), it
has no credits (has_credits / unlimited), and it isn't a workspace still
within its overage (overage_limit_reached = false, no spend cap, an
ordinary rate_limit_reached). Window percent never decides it. Where the
app goes by what /wham/usage can't show (a statsig gate, the gpt-reserve
fallback, Team's reserve experiment that ignores overage), it counts as
held: a wrong move costs ChatGPT extras, a missed one every turn. A
reading kept through a failed read (keepLast) stops saying held once its
used-up window has started again, as its usage does. NextLogin's account
rotation and its thresholds are unchanged.

A user stuck on the old move (codex.out = 1) is put back beside the
sign-in by the next Sync, as #540 already does once the account has room.

Verification:
- TestCodexUsedUpIsWhatTheAppHolds (real /wham/usage replies, ids
  replaced) fails on the old trigger: 7 subcases, "CodexUsedUp = true,
  want false". Each workspace guard (spend cap, reason, overage, plan
  list, Team reserve) was mutated and fails its own subcase.
- TestCodexHeldKeptOnlyTillTheReset fails without the quota_last change.
- The real Pro account (7-day window 100%, ~28k credits, allowed false):
  CodexUsedUp true before, false after.
- go vet ./...; go build (darwin), GOOS=linux -tags nogui, GOOS=windows;
  go test -tags nogui ./... under a sandbox HOME; -race -count=20 on the
  new and related Codex tests.
- Not run: a live Codex app turn on an account truly held (none at hand),
  Windows/WSL runtime, GUI suites (no GUI change).

* provider: a Codex account at a spend cap or past its overage is held, credits or not; limit_reached counts as not allowed

codexHeld let credits through before the spend cap, so a Team seat at its
cap (spend_control.reached, allowed false) with has_credits read not held
and magpie left Codex on an account the backend refuses. The app's
readiness check (26.930.61225, app-primary) returns limited on
spend_control.reached or credits.overage_limit_reached before it looks at
credits, and treats rate_limit.limit_reached like allowed false; held now
reads the same. A spend cap with the rate limit still allowed (a
workspace's included usage) stays not held.
2026-10-06 14:50:39 +08:00

333 lines
11 KiB
Go

package provider
// Signing Codex or Claude Code in to the next of its accounts when the one
// it is on has used its allowance up. Codex's requests on its own models go
// through magpie while more of its accounts are on there, and move on to
// the next account when one is out — but the Codex app, once it knows the
// account it is signed in to is out, may not send at all, so no request
// reaches magpie to move on. Claude Code on its own models never comes
// through magpie: it asks Anthropic as the account it is signed in to, and
// that account being out ends its turns with "You've hit your session
// limit" while the gateway has long moved on to another (#208). So
// whichever magpie runs the gateway signs the agent in to the next account
// that is on and has room, as switching it on the Accounts page would: a
// session started after that is on it from the first turn.
//
// It moves at the share Smart routing counts an account spent at, so the
// account the agent is signed in to and the one the gateway goes to agree
// on which is out (#209): the sign-in follows Smart; Smart doesn't follow
// the sign-in. In order, the gateway goes to the first until it is used
// up or refused, so the sign-in moves only once it is used up (#530).
//
// The account it moved from is the one the user made first, and it stays
// that: once it is no longer low (backShare) the agent is signed back in
// to it, as Smart gives it requests again then (#408). A switch the user
// makes meanwhile ends that. With KeepLogin on the subscription, the agent
// stays signed in to the first, whatever it has left (#524); with
// KeepLoginAs, to that account, wherever it stands in the order.
import (
"context"
"encoding/json"
"log"
"os"
"path/filepath"
"slices"
"strings"
"sync"
"sync/atomic"
"time"
"github.com/yetone/magpie/internal/catalog"
)
// loginSwitchEvery is how often the account each agent is on is looked at.
const loginSwitchEvery = 5 * time.Minute
// SpentShare is the share of an allowance past which an account is all
// but used up: Smart routing keeps it for when no other can take a
// request, and the agent signed in to it is signed in to another.
const SpentShare = 98
// SpentShareOf is the share past which routing counts an account spent:
// SpentShare, but In order sends requests to the first until a window of it
// is used up or the vendor refuses it, so there an account at 98% is still
// tried in its turn (#530).
func SpentShareOf(routing string) float64 {
if routing == Ordered {
return 100
}
return SpentShare
}
// loginSwitching is how magpie moves agent's sign-in, as its subscription
// says: the share the account it is on is moved off at, and whether it is
// kept on the first instead (KeepLogin).
func loginSwitching(agent string) (share float64, keep bool) {
for _, p := range load().Providers {
if p.ID == agent {
return SpentShareOf(p.Routing), p.KeepLogin
}
}
return SpentShare, false
}
// backShare is the share below which the account magpie moved the agent
// off is signed back in to: no longer low, as Smart counts it (the
// gateway's lowShare), so it doesn't go back and forth at the edge.
const backShare = 90
// switchedAgents are the agents whose account magpie moves on.
var switchedAgents = []string{"codex", "claude"}
// usedUp reports whether an account's allowance is used up for now: a
// window that stops the account, for every model, at 100%.
func usedUp(q SubscriptionQuota) bool {
return usedPast(q, 100)
}
// UsedUp is usedUp for the CLI's magpie quota wait, so what it waits out
// is what stops an account here.
func UsedUp(q SubscriptionQuota) bool { return usedUp(q) }
// BackAt is when a used-up account has room again: when the last of the
// windows that stop it (usedUp's) starts again, a window's seconds to go
// counted from now. Zero when one of them doesn't say.
func BackAt(q SubscriptionQuota, now time.Time) time.Time {
var back time.Time
for _, w := range q.Windows {
if w.Aside || w.Model != "" || w.Used < 100 {
continue
}
var at time.Time
switch {
case w.ResetsAt != nil:
at = *w.ResetsAt
case w.ResetSecs > 0:
at = now.Add(time.Duration(w.ResetSecs) * time.Second)
default:
return time.Time{}
}
if at.After(back) {
back = at
}
}
return back
}
func usedPast(q SubscriptionQuota, share float64) bool {
for _, w := range q.Windows {
if !w.Aside && w.Model == "" && w.Used >= share {
return true
}
}
return false
}
// NextLogin is the account agent should be signed in to instead of the
// one it is on: the account magpie moved it off, once that has room again
// (back); else, when the one it is on is spent, the first of its other
// accounts that are on in magpie, in the order they were saved, whose
// allowance is known and not spent. ok is false when the agent should stay.
func NextLogin(ctx context.Context, agent string) (from, to string, back, ok bool) {
if accountRemoved(agent) {
return "", "", false, false // magpie has no say in its sign-in
}
var spares []Login
var first *Login
ls := Logins(agent)
for i, l := range ls {
switch {
case l.Active:
from = l.User
case l.Returns:
first = &ls[i]
}
if !l.Active && l.On && l.Lapsed == "" {
spares = append(spares, l)
}
}
// kept on an account of the user's choosing, the agent goes back to
// it whenever it is on another, and never moves off it (#524)
if as := keptAs(agent); as != "" {
i := slices.IndexFunc(ls, func(l Login) bool { return strings.EqualFold(l.User, as) })
if from == "" || strings.EqualFold(from, as) || i < 0 || ls[i].Lapsed != "" {
return "", "", false, false
}
return from, ls[i].User, true, true
}
if from == "" || len(spares) == 0 {
return "", "", false, false
}
// kept on the first, the agent goes back to it at once, however little
// that has left (#524)
share, keep := loginSwitching(agent)
if keep && first != nil && first.On && first.Lapsed == "" {
return from, first.User, true, true
}
u := LoginUsage(ctx, agent)
now := time.Now()
// an account at its usage cap is spent as one at the share below
// (account_caps.go): it is moved off, and none is moved to or back to
// till the window it filled renews
capped := func(user string, q SubscriptionQuota) bool {
return capReached(q, AccountCapOf(agent, user), now)
}
if first != nil && first.On && first.Lapsed == "" {
if q, known := u[first.User]; known && q.Error == "" && !usedPast(q, backShare) && !capped(first.User, q) {
return from, first.User, true, true
}
}
q, known := u[from]
if q.Provider == "claude" && q.AsOf != nil {
// An expired cached window cannot say whether this account is spent
// now. Keep other windows and the stored historical reading intact.
q.Windows = slices.DeleteFunc(slices.Clone(q.Windows), func(w QuotaWindow) bool {
return w.ResetsAt != nil && !w.ResetsAt.After(now)
})
}
// kept on the first, it stays there however little that has left; and
// it moves on at the share its routing counts the account spent at
if keep || !known || q.Error != "" || !usedPast(q, share) && !capped(from, q) {
return "", "", false, false
}
for _, l := range spares {
if q, known := u[l.User]; known && q.Error == "" && !usedPast(q, share) && !capped(l.User, q) {
return from, l.User, false, true
}
}
return "", "", false, false
}
// SwitchWhenSpent signs agent in to the next of its accounts when the one
// it is on is spent, or back to the one it moved off once that has room
// (NextLogin), and answers the account it signed it in to, "" when it
// stayed.
func SwitchWhenSpent(ctx context.Context, agent string) (string, error) {
from, to, back, ok := NextLogin(ctx, agent)
if !ok {
return "", nil
}
// the first moved off, not one moved to on the way
r, _ := loginReturnOf(agent, from)
if err := switchLogin(agent, to); err != nil {
return "", err
}
if back {
setLoginReturn(agent, loginReturn{})
if keptAs(agent) != "" {
log.Printf("%s: kept signed in to %s; signed it back in to it from %s", agent, to, from)
} else {
log.Printf("%s: %s has room again; signed it back in to it", agent, to)
}
} else {
if r.Back == "" {
r.Back = from
}
r.To = to
setLoginReturn(agent, r)
share, _ := loginSwitching(agent)
log.Printf("%s: %s has used %g%% or more of its allowance; signed it in to %s", agent, from, share, to)
}
// the models the agent is offered are the new account's plan's
catalog.Touched()
return to, nil
}
// loginReturn is the account magpie signed an agent out of when it was
// spent (Back), and the one it signed it in to (To): while the agent is
// still on To, Back is signed in again once it has room.
type loginReturn struct {
Back string `json:"back"`
To string `json:"to"`
}
var loginReturnsMu sync.Mutex
func loginReturnsPath() string { return filepath.Join(filepath.Dir(Path()), "login-returns.json") }
func readLoginReturns() map[string]loginReturn {
m := map[string]loginReturn{}
if b, err := os.ReadFile(loginReturnsPath()); err == nil {
_ = json.Unmarshal(b, &m)
}
return m
}
// loginReturnOf is the account to sign agent back in to, when it is still
// on the one magpie moved it to (active): one signed in otherwise since,
// by the user or the agent itself, is the user's choice.
func loginReturnOf(agent, active string) (loginReturn, bool) {
loginReturnsMu.Lock()
defer loginReturnsMu.Unlock()
r, ok := readLoginReturns()[agent]
if !ok || r.Back == "" || !strings.EqualFold(r.To, active) || strings.EqualFold(r.Back, active) {
return loginReturn{}, false
}
return r, true
}
// setLoginReturn keeps r for agent; an empty one forgets it.
func setLoginReturn(agent string, r loginReturn) {
loginReturnsMu.Lock()
defer loginReturnsMu.Unlock()
m := readLoginReturns()
if _, had := m[agent]; !had && r.Back == "" {
return
}
if r.Back == "" {
delete(m, agent)
} else {
m[agent] = r
}
b, err := json.MarshalIndent(m, "", " ")
if err == nil {
err = writePrivate(loginReturnsPath(), append(b, '\n'))
}
if err != nil {
log.Printf("%s: keeping the account to go back to: %v", agent, err)
}
}
// codexWasUsedUp is whether the account Codex is signed in to was out
// (CodexUsedUp) at the last look.
var codexWasUsedUp atomic.Bool
// noteCodexUsedUp tells the agents' files when the account Codex is signed
// in to runs out (CodexUsedUp: of its allowance and its credits), or has
// room again: the Codex app sends nothing at all for an account that is
// out, a magpie model's turn included, so the agent package then makes
// magpie Codex's provider, and puts it back beside the sign-in after
// (#540). Only a change is told.
func noteCodexUsedUp(now bool) {
if codexWasUsedUp.Swap(now) != now {
catalog.Touched()
}
}
// KeepOnAnAccountWithRoom runs SwitchWhenSpent for Codex and Claude Code a
// minute after it starts and every loginSwitchEvery after that, until ctx
// ends.
func KeepOnAnAccountWithRoom(ctx context.Context) {
t := time.NewTimer(time.Minute)
defer t.Stop()
for {
select {
case <-ctx.Done():
return
case <-t.C:
}
for _, agent := range switchedAgents {
c, cancel := context.WithTimeout(ctx, time.Minute)
if _, err := SwitchWhenSpent(c, agent); err != nil {
log.Printf("%s: switching to an account with room: %v", agent, err)
}
cancel()
}
c, cancel := context.WithTimeout(ctx, time.Minute)
noteCodexUsedUp(CodexUsedUp(c))
cancel()
t.Reset(loginSwitchEvery)
}
}