Add gpt-5.4-mini, gemini-3.1-flash-lite, tzafon 1.6/1.7; drop gemini-3-pro-preview

[rgarcia]

Summary

Follow-up to the model-annotation audit. Updates the @onkernel/cua-ai computer-use support table to match what the discovery smoke tests actually verify, and bumps versions so merging sets up main for a tag-triggered release.

Annotations added (all passed the computer-use smoke test):

• gpt-5.4-mini (openai) — family match; in pi-ai's registry, no override needed. Its docs page reports Computer use: Supported.
• gemini-3.1-flash-lite (google) — exact match; in registry, no override.
• tzafon.northstar-cua-fast-1.6 and tzafon.northstar-cua-fast-1.7-experiment (tzafon) — exact match plus CUA_MODEL_OVERRIDES entries, since pi-ai's registry does not carry them.

Annotation removed:

• gemini-3-pro-preview (google) — Google retired it; the API now returns 404 "This model is no longer available." for every request, so the annotation pointed at a dead model. No Pro-tier successor currently passes the computer-use smoke (gemini-3.1-pro-preview / gemini-3.5-flash return "Computer Use is not enabled" / EAP-only), so nothing replaces it.

Deliberately excluded:

• gemini-3.1-flash-lite-preview — superseded by the stable gemini-3.1-flash-lite above.

Version bumps

• @onkernel/cua-ai 0.2.1 → 0.2.2
• @onkernel/cua-agent 0.3.1 → 0.3.2 (its @onkernel/cua-ai dependency bumped to 0.2.2)
• Changelogs + package-lock.json updated.

After merge, cut the release by pushing the tags per the release skill (AI first): cua-ai/v0.2.2, then cua-agent/v0.3.2.

Note on bump size: the release skill classifies new-model support as a minor bump; this PR uses patch for consistency with the prior release. Easy to switch to 0.3.0 / 0.4.0 before merge if you'd prefer minor.

Test plan

• npm run typecheck (exit 0)
• npm test --workspace @onkernel/cua-ai (88 passed; updated models.test.ts to cover the new ids and the gemini-3-pro-preview removal)
• Runtime check: listCuaModels includes the four new ids; the two tzafon ids resolve via overrides; getCuaModel("google:gemini-3-pro-preview") now throws

🤖 Generated with Claude Code

---

Note

Low Risk
Catalog and version-bump only; no runtime or auth changes, though callers pinned to gemini-3-pro-preview will need another Google CUA model.

Overview
@onkernel/cua-ai 0.2.2 refreshes the computer-use model catalog in CUA_MODEL_ANNOTATIONS and related docs/tests: adds gpt-5.4-mini (OpenAI family), gemini-3.1-flash-lite (Google exact), and tzafon.northstar-cua-fast-1.6 / 1.7-experiment (Tzafon exact + CUA_MODEL_OVERRIDES where pi-ai lacks them). Removes retired gemini-3-pro-preview so listCuaModels / getCuaModel no longer advertise it.

@onkernel/cua-agent 0.3.2 bumps its @onkernel/cua-ai dependency to 0.2.2; lockfile and changelogs updated for release.

^{Reviewed by Cursor Bugbot for commit dfab56d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[firetiger-agent]

Created a monitoring plan for this PR.

What this PR does: Expands the set of AI models available for computer-use automation by adding gpt-5.4-mini, gemini-3.1-flash-lite, and two Tzafon Northstar variants; removes the retired gemini-3-pro-preview (Google now returns 404 for it on every request).

Intended effect:

• listCuaModels() output: baseline (0.2.1) excludes these four IDs; confirmed if all four resolve without error after consumers upgrade to 0.2.2.
• gemini-3-pro-preview Google 404 errors: baseline is pre-existing (Google retired the model before this PR); confirmed fixed if Google API 404s for this model drop to 0 post-upgrade.

Risks:

• Hard-coded gemini-3-pro-preview consumer — API error logs (unsupported CUA model log line), alert if any match after a consumer upgrades to 0.2.2.
• npm tag not cut — release requires manual tag push (cua-ai/v0.2.2) after merge; alert if npm registry does not show 0.2.2 within 30 minutes of merge.
• API error rate regression — kernel_api_request_duration_seconds error rate; alert if it rises above 2% sustained for any hour after a consumer service upgrades.

Status updates will be posted automatically on this PR as monitoring progresses.

View monitor

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Agent tests reference removed model
  • Updated all packages/agent/test/agent.test.ts runtime-resolution and setModel cases from retired google:gemini-3-pro-preview to supported google:gemini-3-flash-preview so they no longer trigger unsupported-model errors.

Or push these changes by commenting:

@cursor push 986ccf3879

Preview (986ccf3879)

diff --git a/packages/agent/test/agent.test.ts b/packages/agent/test/agent.test.ts
--- a/packages/agent/test/agent.test.ts
+++ b/packages/agent/test/agent.test.ts
@@ -145,7 +145,7 @@
 	});
 
 	it("refreshes CUA runtime state when state.model changes", () => {
-		const runtime = resolveCuaRuntimeSpec("google:gemini-3-pro-preview");
+		const runtime = resolveCuaRuntimeSpec("google:gemini-3-flash-preview");
 		const agent = new CuaAgent({
 			browser,
 			client,
@@ -154,7 +154,7 @@
 			},
 		});
 
-		agent.state.model = "google:gemini-3-pro-preview";
+		agent.state.model = "google:gemini-3-flash-preview";
 
 		expect(agent.state.model.id).toBe(runtime.model.id);
 		expect(agent.state.systemPrompt).toBe(runtime.defaultSystemPrompt);
@@ -173,9 +173,9 @@
 			},
 		});
 
-		agent.state.model = "google:gemini-3-pro-preview";
+		agent.state.model = "google:gemini-3-flash-preview";
 
-		const runtime = resolveCuaRuntimeSpec("google:gemini-3-pro-preview");
+		const runtime = resolveCuaRuntimeSpec("google:gemini-3-flash-preview");
 		expect(agent.state.tools.map((item) => item.name)).toEqual([...runtime.toolExecutors.map((item) => item.definition.name), "custom"]);
 		expect(agent.state.systemPrompt).toBe("custom prompt");
 	});
@@ -284,7 +284,7 @@
 	});
 
 	it("refreshes CUA runtime state through setModel", async () => {
-		const runtime = resolveCuaRuntimeSpec("google:gemini-3-pro-preview");
+		const runtime = resolveCuaRuntimeSpec("google:gemini-3-flash-preview");
 		const harness = new CuaAgentHarness({
 			...(await createHarnessServices()),
 			browser,
@@ -292,7 +292,7 @@
 			model: "openai:gpt-5.5",
 		});
 
-		await harness.setModel("google:gemini-3-pro-preview");
+		await harness.setModel("google:gemini-3-flash-preview");
 
 		expect(harness.getModel().id).toBe(runtime.model.id);
 		expect(harness.getTools()).toHaveLength(runtime.toolExecutors.length);
@@ -324,7 +324,7 @@
 		});
 
 		await harness.setActiveTools([]);
-		await harness.setModel("google:gemini-3-pro-preview");
+		await harness.setModel("google:gemini-3-flash-preview");
 
 		expect(harness.getActiveTools()).toEqual([]);
 	});
@@ -340,10 +340,10 @@
 		});
 
 		await harness.setActiveTools(["click", "screenshot"]);
-		await harness.setModel("google:gemini-3-pro-preview");
+		await harness.setModel("google:gemini-3-flash-preview");
 
 		expect(harness.getTools()).toHaveLength(
-			resolveCuaRuntimeSpec("google:gemini-3-pro-preview").toolExecutors.length,
+			resolveCuaRuntimeSpec("google:gemini-3-flash-preview").toolExecutors.length,
 		);
 		expect(harness.getActiveTools().map((tool) => tool.name)).toEqual(["click", "screenshot"]);

_{You can send follow-ups to the cloud agent here.}

^{Reviewed by Cursor Bugbot for commit dfab56d. Configure here.}

[cursor]

Agent tests reference removed model

Medium Severity

Removing gemini-3-pro-preview from CUA_MODEL_ANNOTATIONS while bumping @onkernel/cua-agent to @onkernel/cua-ai 0.2.2 leaves packages/agent/test/agent.test.ts calling resolveCuaRuntimeSpec and setModel with google:gemini-3-pro-preview. Those paths invoke getCuaModel, which now throws for that ref, so @onkernel/cua-agent tests fail against the updated workspace.

 

^{Reviewed by Cursor Bugbot for commit dfab56d. Configure here.}