feat(ai): Gemini 4 Argon, sized as it is — its window, its reasoning, and room to think

[ndemianc]

This is the editor half of the Gemini 4 Argon launch (Step 2 of the launch plan). The catalog row is what puts Argon in front of Pro users: systemu-net/thin.ly#451. That row works without this PR, because the picker reads /account/models live. This PR makes the editor treat Argon as what it is, and it can ship in any release.

What changes

• Argon is sized exactly. Until now it fell to the gemini heuristic: a 1,000,000 window instead of 1,048,576, and no reasoning flag. providers/catalog.js now has rows for google/gemini-4-argon and the bare gemini-4-argon: 1,048,576 context, tools, vision, reasoning, never fast, so it never becomes the ghost-text model.
• A reasoning effort is sent. For Argon, chat and the agent send reasoning: { effort: "medium" }, and compaction sends "low". Before, no effort was sent, so Google's default applied, which is unknown.
  • Only the vendor-prefixed id, which is how OpenRouter and the gateway name Argon, carries the effort. Both take that field for this model: OpenRouter lists reasoning for every Argon endpoint, and the gateway passes the body through.
  • A bare id may be some other endpoint, so it sends none.
• Reasoning models get room to answer. Argon's reasoning comes out of the same max_tokens as its answer, and the agent's 8,192 default can be spent on thinking alone. A model whose caps say it reasons now gets the 32,768 per-turn ceiling, which is also the gateway's PAID_MAX_TOKENS.
  • This applies to o-series and Kimi K3 as well. The existing nudge loop still handles a turn that stops at the limit.
  • It raises the run's worst-case output bound for these models (8 × the per-turn cap) from 65,536 to 262,144 tokens. On Argon that is about $2.60 of output at the intro rate, against $0.66 before.
• Compaction no longer comes back empty. On a reasoning model it summarizes with 8,192 tokens instead of 1,500, which reasoning alone could use up.

The web build runs the same extensions/levelcode-ai source, so it picks all of this up.

Tests

• New and extended cases:
  • catalog.test.js: the two Argon rows, and reasoningEffortFor.
  • providers.test.js: the buildChatBody field, and that the effort reaches the request body on the chat, one-shot and agent paths (index.js copies options field by field).
  • agentReasoningBudget.test.js: runs the real agent loop with a stand-in provider and reads what each turn asked for.
  • compactReasoningHost.test.js: runs the real compactAgentMemory, sliced out of extension.js.
  • sessionExpiredHost.test.js: now hands handleSend the real catalog, which it needs, and checks the chat's effort.
• Seen to fail when they should: each of the 14 changes was undone one at a time with the network refused, and every undo was caught by a test.
• Gate: scripts/test-extensions.sh passes on macOS with Node 24, 126 files, and again with every setTimeout delayed by 15 ms, as on the Linux runner. tsc --checkJs finds no new undefined names.
• Live: scripts/kimi-k3-spike.js against google/gemini-4-argon passed with this branch's request builder and effort: medium. Turn 1 called the tool; turn 2 answered without reasoning carried back.

Not run

• The suites on Linux. Docker was not answering on this Mac.
• The editor itself. Neither a desktop build nor the browser build has been run with this branch, with Argon or otherwise.
• A real multi-step agent task on Argon. It is on the launch plan's checklist for after the thin.ly deploy.

For the next release notes

RELEASE-NOTES.md on develop still holds v1.3.0's notes and is replaced each release, so I did not touch it. Suggested entry:

Gemini 4 Argon, sized as it is. Argon is in the model picker for Pro plans (thin.ly side). The editor now knows its real 1,048,576-token window, asks for a medium reasoning effort, and gives reasoning models room to think and still answer: 32,768 tokens a turn in the agent, and a Compact that no longer comes back empty. About 100 turns a month on Pro at the introductory price.

[Copilot]

🔵 Needs a closer look

Higher billable token ceilings across reasoning models warrant human validation, with desktop, browser, and real multi-step Argon checks still outstanding.

0 open findings

What changed in this PR

Adds editor-side Gemini 4 Argon support to LevelCode’s shared desktop and web AI implementation, complementing the live model roster.

Changes:

• Defines Argon’s exact context window and capabilities.
• Forwards reasoning effort through chat, agent, and compaction requests.
• Raises reasoning-model token budgets and adds regression coverage.

File	Description
extensions/​levelcode-ai/​test/​sessionExpiredHost.test.js	Checks chat reasoning effort.
extensions/​levelcode-ai/​test/​providers.test.js	Tests effort forwarding into request bodies.
extensions/​levelcode-ai/​test/​compactReasoningHost.test.js	Tests compaction budgets and effort.
extensions/​levelcode-ai/​test/​catalog.test.js	Checks Argon capabilities and effort selection.
extensions/​levelcode-ai/​test/​agentReasoningBudget.test.js	Tests agent token budgets.
extensions/​levelcode-ai/​providers/​openaiCompat.js	Serializes reasoning effort.
extensions/​levelcode-ai/​providers/​index.js	Forwards effort through provider dispatch.
extensions/​levelcode-ai/​providers/​catalog.js	Adds Argon metadata and effort helper.
extensions/​levelcode-ai/​extension.js	Applies chat effort and compaction settings.
extensions/​levelcode-ai/​agent.js	Raises reasoning-model budgets and selects effort.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.