Add Moonshot Kimi K3 computer-use provider

[rgarcia]

Summary

Adds moonshotai as a CUA provider so Kimi K3 can drive Kernel browsers: cua -p --model moonshotai:kimi-k3 "..." (moonshot: accepted as an alias, bare kimi-k3 resolves too). Requires MOONSHOT_API_KEY.

How it works

• Moonshot has no native computer tool or coordinate protocol — its API is OpenAI-compatible chat completions with ordinary function tools, so this follows the xAI/Meta custom-function-tool pattern.
• No custom stream adapter: kimi-k3 rides pi-ai 0.80.10's builtin moonshotai provider (openai-completions), whose registry entry already carries K3's compat quirks (deepseek-style thinking format, reasoning-content replay for K3's preserved-thinking-history requirement, max-only thinking effort).
• The provider module declares a normalized 0–1 fractional coordinate contract. Live probes show Kimi grounding operates in width/height fractions: K2.5 emits 0–1 fractions regardless of prompt wording, and K3 follows the declared contract exactly — (0.927, 0.293) for a link centered at that exact fraction of the screenshot, consistent across repeated runs. The translator scales fractions to viewport pixels at execution.
• Payload middleware sets parallel_tool_calls: false since browser actions mutate shared state.

pi bump 0.80.6 → 0.80.10

Needed for the Kimi K3 registry entry. Adaptations:

• Models decorators in cua-agent delegate the new checkAuth/getAvailable/login/logout methods and the widened getAuth overload.
• sendSessionIdHeader compat flag renamed to sessionAffinityFormat.
• Grok 4.5's registry api moved to openai-responses (cua reroutes it anyway; test expectation updated).
• CLI --thinking accepts pi's new max level (the only level K3 supports).

Testing

• All unit tests pass across workspaces (cua-ai 164, cua-agent 160, cua-cli 88). ptywright dist tests fail locally on main too (native build) — unrelated.
• Live K3 validation (all passing):
  • update-models discovery smoke: kimi-k3 emits a click function call.
  • computer-tool.integration.test.ts moonshotai case: canonical click with in-range fractional coordinates.
  • e2e.live.test.ts: CuaAgent and CuaAgentHarness both execute browser steps against a live Kernel browser with moonshotai:kimi-k3.
  • Coordinate convention confirmed: K3 returns 0–1 fractions under the shipped instructions, correctly grounded, stable across runs.
• Also validated the same path against kimi-k2.5/k2.6/k2.7 via the discovery script (all pass the click smoke).

CI uses the new org-wide MOONSHOT_API_KEY secret for the integration/e2e jobs.

---

Note

Medium Risk
Touches multi-package model routing and a pi-ai bump with new live-provider CI paths; coordinate scaling and Kimi-specific thinking/tool-call behavior need correct runtime behavior.

Overview
Adds Moonshot Kimi K3 as a first-class CUA provider so agents and cua can run moonshotai:kimi-k3 (alias moonshot:, bare kimi-k3) with MOONSHOT_API_KEY.

In @onkernel/cua-ai, the new moonshot module follows the Meta/xAI pattern: canonical browser function tools (no native computer API), a 0–1 fractional coordinate contract with prompts/middleware, and parallel_tool_calls: false. kimi-k3 is annotated in models.ts and uses pi-ai’s builtin openai-completions transport (no custom stream adapter). Auth and ref aliases are wired through api-keys, runtime-spec, exports, docs, and tests.

@earendil-works/pi-ai / pi-agent-core go 0.80.6 → 0.80.10 for the Kimi registry entry. Agent Models wrappers forward checkAuth / getAvailable / login / logout and the widened getAuth; compat uses sessionAffinityFormat instead of sendSessionIdHeader. CLI --thinking gains max for K3.

CI, README/CLI help, and update-models discovery/drift scripts include Moonshot. Workspace versions bump (cua-ai 0.7.0, cua-agent 0.7.0, cua-cli 0.5.0) with integration and live e2e coverage for Kimi.

^{Reviewed by Cursor Bugbot for commit 70404a8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

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

Autofix Details

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

• ✅ Fixed: K3 thinking map uses nulls
  • Added a Kimi K3 route override that maps all non-max thinking levels to "max" and updated the unit test to assert the new mapping.

Or push these changes by commenting:

@cursor push d66ee77a5a

Preview (d66ee77a5a)

diff --git a/packages/ai/src/models.ts b/packages/ai/src/models.ts
--- a/packages/ai/src/models.ts
+++ b/packages/ai/src/models.ts
@@ -234,6 +234,12 @@
 			compat: { supportsDeveloperRole: false, sessionAffinityFormat: "openai-nosession", supportsLongCacheRetention: false },
 		};
 	}
+	if (model.provider === "moonshotai" && model.id === "kimi-k3") {
+		return {
+			...model,
+			thinkingLevelMap: { ...model.thinkingLevelMap, off: "max", minimal: "max", low: "max", medium: "max", high: "max", xhigh: "max" },
+		};
+	}
 	return model;
 }
 

diff --git a/packages/ai/test/models.test.ts b/packages/ai/test/models.test.ts
--- a/packages/ai/test/models.test.ts
+++ b/packages/ai/test/models.test.ts
@@ -87,9 +87,10 @@
 		expect(kimi.input).toContain("image");
 		expect(kimi.contextWindow).toBe(1_048_576);
 		expect(kimi.maxTokens).toBe(131_072);
-		// K3 launched with max-only thinking effort; pi-ai clamps the rest away.
+		// K3 only accepts max reasoning effort, so every level maps to max.
 		expect(kimi.thinkingLevelMap?.max).toBe("max");
-		expect(kimi.thinkingLevelMap?.high).toBeNull();
+		expect(kimi.thinkingLevelMap?.high).toBe("max");
+		expect(kimi.thinkingLevelMap?.low).toBe("max");
 	});
 
 	it("accepts the moonshot: alias for moonshotai refs", () => {

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

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{Bugbot Autofix is ON, but it could not run because the branch was deleted or merged before autofix could start.}

^{Reviewed by Cursor Bugbot for commit 70404a8. Configure here.}

[cursor]

Quickstart starves K3 thinking budget

Medium Severity

The quickstart now advertises moonshotai:kimi-k3 but still caps maxTokens at 1024. K3 always runs max-only thinking, so reasoning can consume the budget before a tool call is emitted. The integration case already raises this to 8192 for the same reason; following the documented quickstart with Kimi is likely to return text-only or truncated replies instead of a click.

 

^{Reviewed by Cursor Bugbot for commit 70404a8. Configure here.}

[cursor]

Example skips Moonshot payload hook

Medium Severity

Moonshot’s serial-tool constraint lives only in providerModule.onPayload (parallel_tool_calls: false), unlike Meta/xAI where the transport enforces it. The quickstart resolves a runtime spec that includes that hook, then calls cuaModels().complete without passing spec.onPayload, so the advertised Moonshot path never applies the constraint.

Additional Locations (1)

• packages/ai/src/providers/moonshot/index.ts#L44-L49

 

^{Reviewed by Cursor Bugbot for commit 70404a8. Configure here.}