Scheduler v4 reports session ownership separately from native selection preferences and pins. Only eligible candidates are annotated, and cache preferences do not imply a session binding.
BREAKING CHANGE: scheduler plugins must declare capability version 4 and accept the new candidate fields.
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>
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>
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>
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>
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>
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>
Carry the freeze recovery target into provider account selection so probe responses cannot update credentials or clear the active protection. Let the recovery worker postpone or release the same freeze after observing the upstream result.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Use a fresh upstream key concurrency limit as the account default for scheduling, leases, and capacity reporting. Show disclosed limits alongside the upstream pool summary.
BREAKING CHANGE: Both peers must support CPR Link schema v3.
Refs: #587
Co-authored-by: Codex <noreply@openai.com>
Export authoritative budget access flags and keep monetary fields display-only
for peer consumers. Use the official usage_limit_reached error type for budget
exhaustion while preserving ordinary rate-limit cooldowns.
Apply fresh quota facts atomically with peer snapshots through the shared quota
writer, retaining connection, claim, and observation-time guards. Reuse the
existing sync worker and document the recovery contract.
Validated with 234 selected tests, strict workspace Clippy, and Rustfmt.
Co-authored-by: Codex <noreply@openai.com>
Return non-retryable client version errors and correct provider error types.
Keep safe user messages separate from operational diagnostics.
Unify error wording and document the conventions in the development skill.
Co-authored-by: Codex <noreply@openai.com>
* feat(accounts): support configurable error protection rules
Replace capacity-specific freezing with bounded, configurable account error rules and shared attempt feedback.
Drop legacy settings without conversion and expose rule editing and freeze reasons in the admin UI.
Co-authored-by: Codex <noreply@openai.com>
* fix(accounts): preserve error protection scheduling boundaries
Keep xAI accounts excluded while recovery probes are pending, and cover
selection before and after recovery.
Separate scheduling cooldown reads from admin display details, share the
atomic Redis migration and selection logic, and use one policy revision.
Remove the unused adaptive concurrency reduction path and its tests.
Validation: 393 targeted backend tests, workspace Clippy with -D warnings,
Rustfmt, frontend type checking and build, and local service health checks.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
* feat(plugin): publish account-scoped adapted model catalogs
* fix(core): preserve candidate account scope during execution
Apply the narrowed candidate scope before provider account selection, including affinity and account rotation. Cover adapted-only models and same-name models with incompatible tool capabilities through the real OpenAI selector and local HTTP fixtures.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Resolve identity and frozen access facts without constructing model execution
contexts. Coordinate parent task suspension, resumption and retained callbacks
at the SDK and Runtime boundary, with bounded admission and cancellation.
Expose account HTTP, client verification, access facts and quota details through
the existing host owners.
BREAKING CHANGE: Plugin sessions require RPC v7; rebuild and repackage plugins
with the matching SDK. No legacy protocol negotiation or fallback is provided.
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Remove upstream response model tracking across providers, storage, APIs,
and plugin facts without migrating historical records.
Refine sidebar hints, locale loading indicators, model columns, and usage
page labels and filters.
Co-authored-by: Codex <noreply@openai.com>
Use request, attempt, and outbound middleware for plugin-owned identity
selection and complete request rewriting. Preserve native account identity
and partition WebSocket connections through existing outbound opening changes.
Remove the construction parameter channel and its obsolete tests and docs.
Keep middleware at v6 without a compatibility shim.
Co-authored-by: Codex <noreply@openai.com>
* feat(plugins): select upstream adapters within account scopes
* fix(plugins): defer adapter decisions until account selection
Keep scoped adapters pending until an account is known so unavailable
fallbacks cannot mask a healthy earlier match. Filter native xAI accounts
when only an adapter can accept the request, preserving normal selection
and cold-stream behavior.
Cover fallback precedence and mixed-pool prewarm eligibility regressions.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Remove redundant contract cases and express fixtures in terms of current behavior.
Reduce isolated database migration contention, run independent scenarios concurrently,
and replace fixed waits with clock and I/O synchronization. Check absolute Redis
session expiry and wait for PostgreSQL migration session locks to be released.
Document nextest for local test execution.
Validated 4060 tests, 70 Redis ACL tests, and five repeat runs of 26 focused tests.
The 10 plugin state tests also pass 50 repeat runs with four workers.
Rustfmt and strict Clippy pass. Local full-suite execution with 16 workers falls
from 162.760s to 89.072s, excluding compilation.
Co-authored-by: Codex <noreply@openai.com>
Preserve request evidence and cold middleware boundaries across cancellation. Make backup auditing atomic, retain committed account-group results, and serialize restart operations.
Unify frontend stream authentication and session recovery, clear stale secrets, and remove redundant runtime paths. Correct CI permissions and database test URL handling.
Align documentation with current module responsibilities, APIs, and validation workflows, including both submodules.
* fix(retry): rotate accounts after connection failures
Reliable NotSent HTTP connection failures stay on the same account and
egress until the shared 30-second opening window is exhausted. When an
account's proxy does not respond, each opening waits for the full
15-second connect timeout on the same route, so requests that selected
that account fail after about 30 seconds while other accounts stay idle.
After a connection failure, exclude the account and reselect through
the existing rotation path, consuming routing and rotation budget and
sharing the same opening window. Rotation happens after the first
failure because a second same-account opening on an unresponsive
egress would leave too little of the window to rotate. If no other
account is eligible, restore the exclusion and continue the original
same-account HTTP recovery, so single-account deployments keep their
current behavior. Required accounts and continuation attempts are
unchanged.
Refs #485
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(retry): retain same-account connection recovery
Keep existing account rotation rules and cover rotation after a recovered
connection receives a replay-safe rate-limit rejection.
Co-authored-by: Codex <noreply@openai.com>
* fix(core): reserve connection budget for account rotation
Retry the original account before rotating on unsent connection failures.
Split the retry window and retain an opening for HTTP recovery on another
account, with bounded fallback when no alternative can be selected.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Preserve response.create field order and align OAuth query ordering with
current official Codex sources.
Stop adding the timing-metrics request header by default and remove the
four unused engine TTFT and token interval fields throughout the backend,
plugin observation contracts, and frontend. Keep rate calculations unchanged.
Co-authored-by: Codex <noreply@openai.com>
Move business JSON out of control metadata, encode request snapshot headers
compactly, and reuse encoded frames. Cache compiled configuration schemas
and prepare plugin instances with bounded concurrency.
Ignore unused plugin result fields and remove redundant business validation
while preserving host authorization, resource ownership, and capacity limits.
Keep response decoding failures local to their calls.
Seal callback side effects before completing or cancelling RPC calls, and
recover abandoned state migrations without allowing old owners to promote.
Validation: SDK 112 tests and its doctest, Runtime 203 tests, Core 542 tests,
40 PostgreSQL plugin tests, and 16 architecture checks passed. Strict Clippy,
formatting, and diff checks passed.
BREAKING CHANGE: RPC protocol v4 changes frame payload layout and request
snapshot header encoding. Plugins must update their SDK and protocol
manifest version and be rebuilt.
Co-authored-by: Codex <noreply@openai.com>
* feat(scheduling): add plugin filtering and remove automatic switchback
Add scheduler v2 filtered delegation across native selection, affinity,
capacity waits, and lease retries. Keep plugin configuration and business
rules owned by plugins without legacy scheduler compatibility.
Remove automatic switchback from the runtime, SDK, and settings UI.
Migrate the existing settings JSON without creating new tables.
Co-authored-by: Codex <noreply@openai.com>
* fix(dev): align plugin source checks with scheduler v2
Update the workbench submodule to the v2-capable SDK and scheduler,
and document the explicit scheduler manifest version.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Collect request, retry, error and recovery evidence with bounded reads and
credential redaction. Keep ordinary detail responses small and expose a
shared download action next to the close button.
Preserve export metadata when limits are reached and deduplicate selected
errors without discarding business context based on generic field names.
Co-authored-by: Codex <noreply@openai.com>
Track short-lived failures by client key, thread, and turn. Skip RetryThenReturnError retries on subsequent requests, clear markers on success with version checks, and preserve other retry policies.
Co-authored-by: Codex <noreply@openai.com>
Retry ordinary 503 server errors up to five times on the same account with capped backoff and positive jitter. Preserve upstream Retry-After and return the final error without rotating accounts; retain rotation for explicit capacity rejections.
Co-authored-by: Codex <noreply@openai.com>
Deliver HTTP/SSE and WebSocket events immediately while preserving bounded
recovery after replaceable preludes. Share prelude classification so response
metadata does not hide WebSocket connection-limit recovery.
Keep irreversible delivery, cancellation, middleware, and retry-budget guards.
Document the distinction between transport keepalive and event idle timeouts.
Refs #405
Co-authored-by: Codex <noreply@openai.com>
Start downstream keepalives while admitted streaming requests await their first output, including middleware processing. Separate transport commitment from the replay barrier so heartbeat frames preserve safe retries and execution accounting.
Co-authored-by: Codex <noreply@openai.com>
* fix(openai): preserve Lite interrupts across prelude retries
Deliver known Responses Lite WebSocket preludes before semantic output while retaining existing recovery guards. Keep interrupt intent across retries and retarget it only after the replacement response is delivered.
Co-authored-by: Codex <noreply@openai.com>
* fix(openai): forward Lite events without waiting for output
Deliver unclassified Lite WebSocket events in order and reuse the existing commit boundary to stop replay after their delivery. Preserve error classification, retry budgets, account recovery, and interrupt handling.
Co-authored-by: Codex <noreply@openai.com>
---------
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Build durable request facts and bounded aggregate queries, freeze historical
identities, and move recovery lease updates out of request history rows.
Align token fields across storage, APIs and pages.
Invalidate analytics only after actual request deletions and track embedded
migration inputs during builds.
Co-authored-by: Codex <noreply@openai.com>