Merge remote-tracking branch 'origin/farm/36563f95/strip-context-management-on-forced-tool-choice'

2026-06-09 00:26:38 +02:00
parent 521efb385d 85891a5784
commit 8a32d2ab33
2 changed files with 108 additions and 0 deletions
@@ -12,6 +12,10 @@

 - Added AIML API as an OpenAI-compatible provider preset with `AIMLAPI_API_KEY` discovery ([#2105](https://github.com/can1357/oh-my-pi/issues/2105)).

+### Fixed
+
+- Pinned a regression test against issue [#2123](https://github.com/can1357/oh-my-pi/issues/2123): OAuth requests to adaptive-thinking Claude Opus models (4.6+) ship a `context_management.edits[clear_thinking_20251015]` block paired with the `thinking` field, but the eager-todo prelude (and other paths that force `tool_choice` to `tool`/`any` on the first user turn) route through `disableThinkingIfToolChoiceForced`, which would strip `params.thinking` while leaving the orphan `context_management` behind. The Anthropic API then rejected the request with `400 ... clear_thinking_20251015 strategy requires thinking to be enabled or adaptive`. The fix that lands in [15.10.5] now drops both fields together; the new test locks the contract so the strategy can never outlive its enabling `thinking` payload again.
+
 ## [15.10.5] - 2026-06-08

 ### Breaking Changes
@@ -0,0 +1,104 @@
+/**
+ * Issue #2123 — `Error while using Claude Opus models`
+ *
+ * Reporter (omp 15.10.4, Windows, Claude Pro/Max OAuth): a fresh chat on
+ * Claude Opus 4.6 fails immediately with
+ *   `400 ... clear_thinking_20251015 strategy requires thinking to be enabled or adaptive`.
+ *
+ * Root cause: OAuth requests on adaptive-thinking Opus models attach a
+ * `context_management.edits[clear_thinking_20251015]` block. On the very
+ * first user turn the coding-agent's eager-todo prelude pins a forced
+ * `tool_choice: { type: "tool", name: "todo" }`, which routes through
+ * `disableThinkingIfToolChoiceForced` in `providers/anthropic.ts`. The
+ * 15.10.4 implementation stripped `params.thinking` but left
+ * `params.context_management` in place, sending an orphan strategy that
+ * the Anthropic API rejects with the exact message above.
+ *
+ * Fix: when forced tool_choice removes `thinking`, also remove
+ * `context_management` so the wire payload satisfies the
+ * "thinking enabled or adaptive" precondition (i.e. neither is sent, and
+ * the strategy goes with them).
+ */
+import { describe, expect, it } from "bun:test";
+import { Effort } from "@oh-my-pi/pi-ai/effort";
+import { streamAnthropic } from "@oh-my-pi/pi-ai/providers/anthropic";
+import type { Context, Model, Tool } from "@oh-my-pi/pi-ai/types";
+
+const OPUS_46_OAUTH: Model<"anthropic-messages"> = {
+	id: "claude-opus-4-6",
+	name: "Claude Opus 4.6",
+	api: "anthropic-messages",
+	provider: "anthropic",
+	baseUrl: "https://api.anthropic.com",
+	reasoning: true,
+	input: ["text", "image"],
+	cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
+	contextWindow: 1_000_000,
+	maxTokens: 128_000,
+	thinking: { mode: "anthropic-adaptive", minLevel: Effort.Minimal, maxLevel: Effort.XHigh },
+};
+
+const todoTool: Tool = {
+	name: "todo",
+	description: "Manage a phased task list",
+	parameters: {
+		type: "object",
+		properties: { ops: { type: "array", items: { type: "object" } } },
+		required: ["ops"],
+	} as unknown as Tool["parameters"],
+};
+
+const firstTurnContext: Context = {
+	systemPrompt: ["Stay concise."],
+	messages: [{ role: "user", content: "Plan my migration", timestamp: Date.now() }],
+	tools: [todoTool],
+};
+
+function abortedSignal(): AbortSignal {
+	const controller = new AbortController();
+	controller.abort();
+	return controller.signal;
+}
+
+type CapturedPayload = {
+	thinking?: unknown;
+	context_management?: unknown;
+	tool_choice?: { type?: string; name?: string };
+};
+
+function capturePayload(toolChoice: "any" | { type: "tool"; name: string }): Promise<CapturedPayload> {
+	const { promise, resolve } = Promise.withResolvers<CapturedPayload>();
+	void streamAnthropic(OPUS_46_OAUTH, firstTurnContext, {
+		apiKey: "sk-ant-oat-test",
+		isOAuth: true,
+		signal: abortedSignal(),
+		thinkingEnabled: true,
+		reasoning: Effort.High,
+		toolChoice,
+		onPayload: payload => {
+			resolve(payload as CapturedPayload);
+			return undefined;
+		},
+	});
+	return promise;
+}
+
+describe("issue #2123 — OAuth Opus + forced tool_choice must strip context_management with thinking", () => {
+	it("strips context_management when forced named tool_choice deletes adaptive thinking", async () => {
+		// Mirrors the eager-todo prelude: first user turn on Opus pins
+		// tool_choice: { type: "tool", name: "<active-tool>" } while adaptive
+		// thinking is still requested at default `high` effort.
+		const payload = await capturePayload({ type: "tool", name: "todo" });
+		expect(payload.thinking).toBeUndefined();
+		expect(payload.context_management).toBeUndefined();
+	});
+
+	it("strips context_management when forced `any` tool_choice deletes adaptive thinking", async () => {
+		// Plan-mode and other "must call something" enforcement paths use
+		// the broader tool_choice "any" (Anthropic's `any`). Same invariant:
+		// orphan context_management must NOT survive the strip.
+		const payload = await capturePayload("any");
+		expect(payload.thinking).toBeUndefined();
+		expect(payload.context_management).toBeUndefined();
+	});
+});