1.3.0 — Agents window (experimental)

1.3.0

• Agents window (experimental, opt-in). New setting omnicopilot.exposeToAgentsWindow (default off): 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 Copilot Chat picker keeps working unchanged. 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.

Full changelog: https://github.com/diegosouzapw/OmniCopilot/blob/main/CHANGELOG.md

1.3.0 preview — Agents window (experimental)

Preview build for the Agents window feature (#14, requested in #9). Not on the Marketplace/Open VSX yet — the store release follows once the preview is confirmed against a real Agents window setup.

What's in it

• New setting omnicopilot.exposeToAgentsWindow (default off): 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 Copilot Chat picker keeps working unchanged.

How to test

1. Download omnicopilot-1.3.0.vsix below → in VS Code: Extensions → ⋯ → Install from VSIX…
2. Enable the experimental VS Code setting chat.agentHost.byokModels.enabled.
3. Enable omnicopilot.exposeToAgentsWindow in the extension settings.
4. Restart the agent host (easiest: reload the window), open the Agents window, and check its model picker for the OmniRoute entries.

What to report

• Whether the OmniRoute models appear in the Agents window picker.
• Whether a chat + tool call actually works there.
• Whether anything shows up duplicated in the regular picker (that would mean your VS Code build ignores the targetChatSessionType proposal — turn the setting back off).

Feedback → #14.

1.2.0 — Usage panel & settings

Your key's own spend, in the panel

The Activity Bar panel gains a My usage section — your key's 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.

It reads OmniRoute's self-service endpoint (GET /api/usage/om-usage?format=json) with your
connection's own key — never the management surface.

The three normal states are shown, not flattened into 0%:

• "Usage is not enabled for this key" — the allowUsageCommand flag is off by default; an
  admin turns it on per key in OmniRoute's API-key manager. A setting, not an error.
• "nothing is cached yet" — the key is allowed, the server just hasn't learned a quota.
• No section — the OmniRoute server predates ?format=json; the panel hides rather than
  parse the older text form.

Requires OmniRoute v3.8.50+ for the usage section; everything else works on any version.
Resolves #8 (via #12).

Settings are findable

OmniRoute: Open Settings in the Command Palette, and an Extension settings link in the
panel — the discoverability half of #10.

Install

VS Code Marketplace · Open VSX — or search "OmniRoute" in the Extensions view.

Full changelog: https://github.com/diegosouzapw/OmniCopilot/blob/main/CHANGELOG.md

1.1.0 — Reasoning effort

Reasoning effort reaches the model

Models with extended thinking — GPT-5.x, Claude with thinking, DeepSeek R1, Gemini Thinking — accept a reasoning tier, and VS Code can show an effort control next to the model. Nothing connected the two: only temperature was read out of modelOptions, so every request arrived with no effort at all.

The provider now resolves an effort per request and sends it as reasoning_effort.

• The editor's choice always wins. A new omnicopilot.defaultReasoningEffort setting covers the case where no control is shown.
• Canonical values (none / low / medium / high / xhigh), normalized rather than forwarded verbatim — minimal is OpenAI-only and other providers reject it; max / extra / ultra are synonyms for the top tier. OmniRoute downshifts a tier a model does not implement, so asking for the top one is always safe.
• The default never leaks across the catalog. It applies only to models the catalog marks reasoning-capable; sending the field to a model without thinking support is ignored at best and a 400 at worst. An explicit request is still forwarded when the flag is missing, since the catalog can be stale.
• Reasoning models are labelled extended thinking in the picker tooltip — the VS Code LM API (1.104) has no capability flag for it.

Requested in #7 by @aliaksandrsen.

Security

Package	From	To	Severity
sharp	0.33	0.35.3	high — inherited libvips CVE-2026-33327 / 33328 / 35590 / 35591
esbuild	0.24	0.28.2	moderate — dev server readable cross-origin

Both are devDependencies (icon generator, bundler) and neither ships inside the .vsix, so no published version was exploitable through them. npm audit is now clean.

Install

VS Code Marketplace · Open VSX — or search "OmniRoute" in the Extensions view.

Full changelog: https://github.com/diegosouzapw/OmniCopilot/blob/main/CHANGELOG.md

v1.0.2 — editor dashboard fallback

Fixes dashboardOpen: "editor" opening a "refused to connect" tab against a server that blocks framing. The extension now probes the framing headers first and falls back to the external browser, explaining once that OmniRoute has to be built with DASHBOARD_ALLOW_EMBED=vscode (build-time — setting the variable on an existing install is not enough).

Install: Marketplace · Open VSX

v1.0.1 — catalog fixes

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

• No more duplicate models in the picker. OmniRoute defaults to MODELS_CATALOG_PREFIX_MODE=dual, advertising 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 older servers are covered too. On the validation instance: 949 duplicates removed, zero models lost.
• Only conversational models reach the picker. Image, video, audio and rerank registries are filtered out — the server rejects those on /v1/chat/completions with HTTP 400. Responses-API models (every Codex/GPT-5.x entry) are kept, since OmniRoute translates them transparently.
• /v1 no longer leaks into your settings. Configure the server root (http://localhost:20128); /v1 is appended per request. Values already stored with /v1 keep working and are never doubled.

Install: VS Code Marketplace · Open VSX

OmniCopilot 1.0.0

First public release 🎉

1200+ AI models in your GitHub Copilot Chat — free & forever free.

• OmniRoute models in the native Copilot Chat model picker (VS Code ≥1.104, stable & Insiders)
• Live catalog discovery with per-model tool calling & vision (combos included)
• Activity Bar panel: status, model count, server URL + API key (OS keychain)
• Localized in 42 languages
• Configure Coding CLI for Codex, Claude Code, Cline, Continue, Cursor, Aider & more
• Dashboard in browser or VS Code tab

Install: download omnicopilot.vsix below → VS Code → Extensions → ⋯ → Install from VSIX, or code --install-extension omnicopilot.vsix. Marketplace listing coming next.

Requires an OmniRoute server (npm i -g omniroute).