Changelog

1.3.0

Agents window (experimental, opt-in). omnicopilot.exposeToAgentsWindow registers a second, agent-host-scoped set of entries for every tool-calling OmniRoute model, so they appear in the Copilot Agents window model picker (the regular chat picker is untouched). Requires the experimental VS Code setting chat.agentHost.byokModels.enabled and an agent-host restart; built on the proposed targetChatSessionType API, hence default-off. Requested in #9 by @Bulzi-Robb — implementation track #14.

The panel now shows your key's own spend. A My usage section reads OmniRoute's self-service endpoint (GET /api/usage/om-usage?format=json) and renders the daily/weekly spend against its limit, the reset time, and each provider connection's quota side by side — Codex, Claude, OpenCode — instead of one provider's number. The three normal states are surfaced, not hidden behind a 0%: usage not enabled for this key (the allowUsageCommand flag is off by default), nothing cached yet, and — on a server too old for the JSON form — the section simply doesn't render. Closes #12 → resolves #8 once the OmniRoute side ships.
Settings are findable. OmniRoute: Open Settings in the Command Palette and an Extension settings link in the panel — the discoverability half of #10.

Reasoning effort is forwarded to models that support extended thinking. When VS Code offers an effort control next to the model, the choice now reaches OmniRoute as reasoning_effort; a new omnicopilot.defaultReasoningEffort setting covers the case where the editor does not expose one. Values use OmniRoute's canonical vocabulary (none/low/medium/high/xhigh) and the server downshifts a tier a model does not support, so asking for the top tier is always safe. The default only applies to models the catalog marks as reasoning-capable — sending the field to a model without thinking support is ignored at best and a 400 at worst. Reasoning models are now labelled "extended thinking" in the picker tooltip. Requested in #7 by @aliaksandrsen.
Security: sharp 0.33 → 0.35 (high — inherited libvips CVE-2026-33327 / 33328 / 35590 / 35591) and esbuild 0.24 → 0.28 (moderate — dev server could be read cross-origin). Both are build-time-only dependencies and never shipped inside the .vsix, so no published version was exploitable; npm audit is now clean.

The in-editor dashboard no longer opens a broken tab. dashboardOpen: "editor" only guarded against the Simple Browser command being missing, which is the wrong failure mode: against a server that sends X-Frame-Options: DENY the command succeeds and the iframe renders a "refused to connect" page. The extension now checks the framing headers first and falls back to the external browser, explaining once that the server has to be built with DASHBOARD_ALLOW_EMBED=vscode — it is a build-time option, so setting the variable on an existing install is not enough.

Catalog and configuration fixes, validated against a live OmniRoute instance serving 2345 catalog entries.

No more duplicate models in the picker. OmniRoute defaults to MODELS_CATALOG_PREFIX_MODE=dual, which advertises every model twice — once under the short alias prefix and once under the canonical provider prefix (cc/claude-sonnet-4-6 and claude/claude-sonnet-4-6). The extension now requests ?prefix=alias, and independently drops mirror rows via their parent back-reference so servers too old for that parameter are covered as well. On the validation instance this removed 949 duplicates with zero models lost.
Only conversational models reach the picker. The catalog also lists image, video, audio and rerank registries, which the server rejects outright on /v1/chat/completions. Responses-API models (every Codex/GPT-5.x entry) are kept — OmniRoute translates those transparently.
/v1 no longer leaks into your settings. Configure the server root (http://localhost:20128); the extension appends /v1 per request. Values already stored with /v1 keep working and are never doubled.

First public release. 🎉

OmniRoute models in the Copilot Chat model picker (VS Code stable ≥1.104 and Insiders) — live discovery from GET /v1/models with per-model tool-calling and vision capabilities from the OmniRoute catalog, combos included
Streaming responses with full tool-calling round-trip (tool_choice: required honored) and image input for vision-capable models
Activity Bar panel: connection status, model count, server URL and API key (SecretStorage / OS keychain), quick actions
Status-bar connection dot with instant feedback from live requests
Localized in 42 languages — mirroring the OmniRoute catalog (pt-BR, es, fr, de, ja, ko, zh-CN, zh-TW, ru, ar, hi, …)
Open Dashboard in the external browser or in a VS Code editor tab (Simple Browser; editor mode requires an OmniRoute build that allows embedding)
Configure Coding CLI command driving omniroute setup-* (Codex, Claude Code, Cline, Continue, Cursor, Aider, OpenCode, Goose, Crush, Qwen, Kilo, Roo)
First-run onboarding with OmniRoute install pointers

Internal preview.