Keep the client responsible for echoing turn state across requests. Scope WebSocket opening state to the first request and retain account isolation when moving a request to another account.
Co-authored-by: Codex <noreply@openai.com>
Let plugins page through retained account observations and find a prior baseline without reading raw provider documents or refreshing upstream. Reuse provider window identities and the existing completion index, preserving resets and gaps.
Co-authored-by: Codex <noreply@openai.com>
Move ProxyEndpoint into the neutral upstream module so error observations no longer depend on account types. Keep proxy matching in the account adapter and verify host/port matching, redaction, and the owner dependency graph.
Co-authored-by: Codex <noreply@openai.com>
Migrate the session binding before requesting a full client replay when safe native continuation recovery exceeds the configured wait limit. Remove stale capacity retry headers from client responses while preserving upstream diagnostics and account rotation limits.
Co-authored-by: Codex <noreply@openai.com>
Reserve a connection attempt for another eligible account when a new WebSocket request repeatedly fails before sending. Use the same rotation budget as HTTP while preserving session and replay restrictions.
Co-authored-by: Codex <noreply@openai.com>
Rank quota scheduling strictly by configured reset windows and persist its settings independently from smart scheduling.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Resolve explicit Claude session identities before account selection and
revalidate them after attempt middleware using the existing affinity policy.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Apply the global request profile to account catalog discovery, warmup,
and diagnostic routing. Freeze each operation's identity through the
existing provider profile resolver.
Refresh verified offline releases and centralize their version, artifact,
and environment defaults. Cover ZIP64 MSIX metadata and keep release
update tests independent of the current bundled versions.
Fixes#703
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Make event delivery respond to downstream cancellation and pool shutdown. Return or discard sockets before publishing terminal events while preserving normal connection reuse.
Co-authored-by: Codex <noreply@openai.com>
Use the last successful catalog for the same account, credential revision, and plan until a refresh replaces it. Apply the same exclusions in memory and in PostgreSQL while preserving explicit allowlists and accounts with no catalog.
Co-authored-by: Codex <noreply@openai.com>
* fix(egress): honor deployment proxies for unbound outbound requests
Prefer the proxy selected for the request. Otherwise apply the deployment proxy and bypass rules consistently to HTTP and WebSocket connections, including background downloads. Keep explicit direct probes separate.
Co-authored-by: Codex <noreply@openai.com>
* fix(openai): preserve WebSocket dial errors
Recover the OS error or I/O kind from Hyper without retaining the endpoint. Keep refused connections distinguishable for both direct and proxied WebSocket dials.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Use each upstream handshake's message-size limit to select HTTP before
uploading oversized response.create frames. Remove fixed thresholds and
thread size hysteresis while preserving continuation replay and cleanup.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Separate protocol identifiers from Provider identities and centralize shared
request facts and observations in gateway-protocol. Keep endpoint models
and private continuation state owned by Providers.
Reorganize the architecture documentation and clarify the shutdown guarantees
for buffered charges and observations.
Validated with workspace formatting, strict Clippy, 3013 affected-module
tests and 16 architecture checks. One reference-binary TLS test is ignored.
Co-authored-by: Codex <noreply@openai.com>
* fix(openai): rotate new round-robin sessions without affinity hits
Strict affinity keeps a session on one account, but the shared request
cursor advanced on every scheduling load, including affinity hits. A new
session then took whatever slot that request volume happened to leave.
Advance a separate session cursor only when round-robin assigns an
account, and keep the request cursor for other strategies.
* fix(openai): advance round robin when an account assignment succeeds
Commit the session binding and round-robin cursor together after acquiring the account lease. Release a conflicting lease before retrying so concurrent first requests share a binding without consuming extra turns. Count delegated selections and preserve explicit plugin picks.
Verified with OpenAI provider tests, store tests against Redis and PostgreSQL, architecture checks, and workspace Clippy.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: koeria <riennie@qq.com>
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Treat x-retry-metadata: NO_MORE_RETRY as a hard signal to skip
same-account backoff and transport recovery, even when Retry-After
is within maxRetryAfterSeconds. Capacity failures still rotate under
the existing policy; server_error still returns the upstream error.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Reuse connection-scoped peer discovery for native HTTP and WebSocket forwarding. Strip the route header for ordinary upstreams and leave explicit plugin forwarding changes to plugins.
Co-authored-by: Codex <noreply@openai.com>
Persist validated directories per credential revision. Restore them in bulk and refresh only missing, expired, or changed accounts with paced background requests. Filter large pools using current per-account evidence while preserving unknown and explicitly allowed candidates.
Co-authored-by: Codex <noreply@openai.com>
* feat(accounts): redeem reset credits when weekly quota runs low
Co-authored-by: Codex <noreply@openai.com>
* fix(accounts): observe weekly quota recovery after the last reset credit
Keep checking the local quota snapshot while recovery is unconfirmed. Once recovery is observed with no cards left, return to daily discovery.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Discover models per account so the Codex picker includes models unique to an eligible group member. Use account-specific catalog evidence when selecting a credential for native Responses requests.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Support Responses, Chat Completions and Messages ingress with Responses
or Messages API upstreams in the existing OpenAI provider.
Share generation admission and delivery, preserve native protocol payloads,
and translate cross-protocol tools, reasoning, streams and usage. Add
Messages token counting and editable upstream protocol settings.
Co-authored-by: Codex <noreply@openai.com>
* feat(keys): add explicit budget windows with safe retries
Add revision-checked daily and weekly windows across admission, settlement
and queries. Reuse the existing budget ledger for atomic window updates,
latest-operation replay and audit records.
Expose typed SDK callbacks while keeping account associations and cycle
decisions in plugins.
Refs #658
Co-authored-by: Codex <noreply@openai.com>
* fix(keys): preserve budget accounting across window changes
Keep restored balances when historical charges are unavailable and reject window changes that would require those charges. Honor future start times after release and report actual spending windows to peers.
Require an explicit null to release a window through the plugin SDK.
Co-authored-by: Codex <noreply@openai.com>
* fix(plugins): preserve budget window precision across SDK calls
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Supervise scheduling lease release tasks and drain them within a bounded shutdown budget. Preserve other instances and the existing TTL fallback.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Read update candidates from the published channel table and advance it only after the release assets are public. Keep the tagged preparation commit separate from the publication commit.
Co-authored-by: Codex <noreply@openai.com>
Publish an immutable channel snapshot with each release and advance a fixed release pointer only after public assets pass verification. Check the exact platform package before notifying clients.
Co-authored-by: Codex <noreply@openai.com>
Use the configured affinity mode for content and cache-key bindings, including strict waiting and binding store failures. Cover queue rejection, cancellation, migration, and preferred spillover, and document the scheduling boundaries.
Refs #623
Co-authored-by: Codex <noreply@openai.com>
Use the client cache key or the opening messages as a tenant-scoped scheduling preference. Keep explicit sessions and continuation ownership separate, and allow spillover when the preferred account is busy.
Refs #622
Co-authored-by: Codex <noreply@openai.com>
Let OpenAI accounts send explicitly allowlisted model IDs upstream while retaining catalog checks for other accounts. Update the account editor and connection test to support custom IDs.
Co-authored-by: Codex <noreply@openai.com>
Read the local running version during bootstrap without waiting for a
remote release check. Reuse the restart handoff pipe on Unix so the
replacement starts as soon as the old service releases its resources.
Keep shutdown cleanup budgets and bounded listener retries. Update the
handoff regression coverage and lifecycle documentation.
Validation: Rustfmt, strict workspace Clippy, 252 Host and bootstrap
tests, and isolated Linux startup and consecutive restart measurements.
Co-authored-by: Codex <noreply@openai.com>