Tags

v2.8.18

- Clean abandoned managed inference helpers and their anonymous volumes.
- Remove unused inference core images after successful updates and startup recovery.
- Prevent duplicate not-found notifications after deleting a Pages project.

v2.8.17

- Stream inference core state backups and restores without buffering archives in Gateway memory.
- Resume the unchanged inference core when an update is interrupted before container replacement.

v2.8.16

- Preserve inference compaction across core WebSocket turns with bounded upstream session reuse and valid terminal errors on interrupted streams.
- Allow safe provider failover before substantive output while preserving continuation and accounting boundaries.
- Stabilize non-proxied DNS checks and enable additional routes for custom templates through the canonical placeholder.
- Add secure Pages artifact upload and managed Docker, deployment, and Pages route parity to AI and MCP tools.
- Fix permission-aware settings workflows, scope cleanup, Pages deletion, empty permission groups, and inference reasoning-order persistence.
- Stabilize status-page initial rendering and health-bar layout.

v2.8.15

- Make Save & Recreate actually recreate containers when only secrets changed.
- Apply created and updated secret values to runtime env and remove deleted secret keys from the recreated container.

v2.8.14

- Correct Codex usage wrapper normalization and remove overrides that altered native usage behavior.
- Route inference threads across subscription accounts above their configured quota reserve and enforce a minimum 1% reserve.
- Keep failed or cancelled Core attempts from being recorded as completed zero-token requests.
- Clear the container updating state live after environment-driven recreation, including when the replacement ID was already adopted.

v2.8.13

- Add authenticated Gateway quota and token-usage views for Codex Desktop and a separate gateway-codex CLI, including 5-hour, 7-day, 30-day, and monthly API windows.
- Integrate optional Desktop and CLI usage activation into setup codex with live old-Gateway compatibility checks, current Codex schema validation, fail-closed initial reads, and non-blocking refreshes.
- Support ownership-aware macOS and native/AppImage Linux activation without replacing system codex, plus offline desktop, CLI, or complete usage-wrapper removal that preserves the base Codex integration.
- Prepare @wiolett/gateway-inference 0.3.4 for manual npm publication with corrected package version and executable manifest.

v2.8.12

- Keep established inference WebSocket sessions alive until the final shutdown boundary during Gateway updates.
- Decode zstd, gzip, and deflate compressed Codex Responses HTTP fallback bodies before routing them to the inference core.
- Anchor inference activity account tooltips to the rendered model text instead of the entire table cell.

v2.8.11

- Rescale public subscription credits to one credit per 1,000,000 weighted tokens while preserving internal accounting units.
- Add 30-day inference usage totals and color-coded history charts, with integer credit summaries outside activity history.
- Group repeated inference provider accounts with animated collapse, stable columns, and whole-row reordering constrained within each provider.
- Allow permission groups without direct scopes and fix production group creation service wiring and modal toast layering.
- Prefetch volume metrics, preserve samples across tab changes, refresh them with volume snapshots, and color volume metric charts.
- Align public and internal AI documentation with current relay, integration, Compose, licensing, alerting, and inference contracts, and remove the stale competitive landscape document.

v2.8.10

- Rescale public subscription credits to one credit per 100,000 weighted tokens without migrating internal accounting data.
- Preserve existing enforcement, ledger entries, limits, and active reservations through boundary conversion.
- Show absolute subscription quota reset dates and times in provider settings.
- Align internal AI documentation with current Docker volume, file-scope, image-update, model-ordering, and inference billing contracts.

v2.8.9

- Preserve Codex continuation identity across inference-core WebSocket boundaries.
- Record failed and incomplete WebSocket turns as failures instead of completed zero-output requests.
- Exclude Codex WebSocket warmup turns from inference activity and accounting.
- Keep request-history data visible during background refreshes and show the provider account used by each request.
- Group repeated provider accounts, constrain account reordering to each provider, and support whole-row model and provider dragging.
- Preserve configurable model and reasoning-level order in client catalogs and selectors.
- Show API billing correctly and remove subscription multipliers from API-backed model forms.
- Fix reasoning mapping reordering and duplicate bottom borders.
- Align internal AI and inference documentation with the current runtime, billing, continuation, and client contracts.

v2.8.8

- Preserve Codex continuations across inference-core socket boundaries.
- Mark failed and incomplete WebSocket turns as failed instead of completed with zero credits.
- Refresh the local Codex model catalog from the proxy daemon without requiring MCP.

gateway-inference-v0.3.1

- Advertise code-mode deferred tool discovery for routed Codex models to prevent full tool-catalog expansion in the first request

v2.8.7

- Correct inference credit accounting for cached reads, cache writes, and reasoning tokens while clearing false estimated-usage labels for provider-reported settlements
- Advertise routed deferred Codex tool discovery and prepare @wiolett/gateway-inference 0.3.1 so large tool catalogs no longer inflate the first request
- Fix double page scrolling and Linux toggle rendering across the application shell
- Refine disk-image volume creation with stable node selection, animated height, clean dropdown labels, and capability-aware availability
- Require the Personal plan or higher for disk-image volume creation with shared upgrade-dialog and server-side enforcement
- Highlight available inference-core updates with a warning border and active updates with the matching info border

v2.8.6-docker

- Add capability-gated fixed-capacity disk-image Docker volumes backed by ext4 images
- Persist disk-image mounts across daemon restarts and safely reconcile existing managed images
- Add online volume growth with loop-device and filesystem resizing
- Report regular and disk-image volume usage, inode capacity, and running-container attachments
- Preserve managed volume metadata across label updates, renames, and deletion

v2.8.6

- Add ordered inference models and reasoning levels so Gateway, Codex manifests, and AI Workspace selectors preserve administrator-defined ordering
- Automatically replace the AI Workspace default model when its configured inference model is removed
- Restore direct assistant delta rendering with smooth tail animations and remove the artificial word-reveal loop that caused slow, uneven streaming
- Increase default inference admission limits to 1,800 requests per minute and 32 concurrent requests per token, with Docker Compose environment overrides
- Add fixed-capacity disk-image Docker volumes with node capability detection, safe creation, persistent mounting, online growth, and cleanup
- Add volume space, inode, and running-container metrics with history charts and resize controls for managed disk-image volumes
- Add reusable selected/unselected scope filtering and group-name search across permission selectors
- Replace OAuth consent scopes with the full resource-aware selector, hide unavailable scopes, and restore loopback callback popup delivery
- Split container filesystem permissions into independent read and write scopes while migrating existing grants automatically
- Consolidate legacy Cloudflare DNS permissions into standard connector and domain permissions
- Rename AI Workspace and MCP proxy-host tools to route terminology consistently across tools, permissions, tests, and documentation

v2.8.5

- Stabilize pooled subscription accounts so newly connected and removed providers immediately update existing model routes and account counts across ChatGPT, Claude, and other pooled providers
- Restore missing provider model capabilities, including Claude reasoning, tools, vision, modalities, and supported reasoning efforts
- Advertise Codex Fast mode for models exposing the OpenAI priority service tier
- Respect configured quota reserves while treating only fully exhausted accounts as unavailable and allowing low-quota fallback when no normal-capacity account remains
- Recover ChatGPT accounts already added to the core pool instead of leaving OAuth authorization stuck or reporting a duplicate-account failure
- Fix provider authorization dialogs so successful connections close reliably, callback completion remains loading until settled, and cancellation cleans up pending sessions
- Preserve unsaved inference provider and model settings during realtime catalog refreshes and remove provider-switch form flicker
- Remove internal scrolling from inference dialogs while keeping animated content and dropdowns visible
- Restore AI Workspace setup and navigation after configuration, including automatic default inference-limit initialization
- Smooth assistant response streaming, prevent previously rendered words from flickering, delay incomplete Markdown headings, and keep user messages stable during generated-title updates
- Surface rejected inference-core WebSocket upgrades clearly without retrying the same core transport failure across provider accounts
- Consolidate Relay image, supervisor, worker, connector, manifest, checksum, and update artifacts into one GitLab release

v2.8.4-relay

- Add multi-instance Relay Pool operation with multiple active relay assignments, per-connection distribution, failover, and existing-stream pinning
- Add the Relay Supervisor and managed Relay Worker for enrolling and operating relay instances on additional physical hosts
- Add signed policy snapshots with revision checks, expiry enforcement, endpoint authorization, and workload-scoped grants
- Add staged endpoint registration and source verification before activating new assignment generations
- Add graceful drain, forced disconnect, active tunnel, registered endpoint, and pressure telemetry
- Add signed one-instance-at-a-time updates for Relay Supervisor and Relay Worker binaries on amd64 and arm64
- Bundle digest-pinned Relay, Database Connector, and Secure Link Connector images in one signed Relay release manifest
- Preserve protocol v1 and singleton compatibility for existing Gateway installations during migration to Relay Pool

v2.8.4

- Add a fault-domain-aware Relay Pool with local and remote relay instances, enrollment, health, drain, resume, rebalance, and staged assignment activation
- Add global and workload-level Relay Spread controls with inherited, fixed-count, and all-ready-relays distribution behind a single Secure Link
- Balance new workload connections across active relay assignments while preserving existing TCP and WebSocket streams during drain and failover
- Add signed Relay Pool policies, scoped grants, source probes, assignment generations, policy expiry handling, and durable update state
- Add one-instance-at-a-time Relay Pool updates with automatic drain, worker and supervisor verification, resume, and failure pause
- Migrate existing singleton Relay installations in place without recreating Secure Links or changing their public topology
- Add support for multiple service addresses per managed node
- Add PEM and PKCS#12 certificate exports and restore certificate export actions
- Preserve unsaved route settings during background refresh and consistently highlight modified configuration blocks
- Improve Relay settings with stable instance ordering, workload spread controls, responsive sizing, and clearer local-node status
- Synchronize realtime resource mutations, recreated environments, generated images, inference settings, OpenRouter balances, and accounting updates

v2.8.4-docker

v2.8.4-monitoring