fix(claude-sdk-oauth): canonicalize eval summary continuity

packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/session-commit-boundary.ts

@@ -1,3 +1,4 @@
+import { clampEvalSummary } from "@code-yeongyu/senpi-codemode/eval-summary";
 import type { AssistantMessage } from "@earendil-works/pi-ai";
 import { CLAUDE_SDK_OAUTH_PROVIDER_ID } from "./account-management.ts";
 import { sessionSyncDigest } from "./session-sync.ts";
@@ -6,13 +7,35 @@ export type AssistantCommitOutcome = "clean" | "rewritten" | "not-resident";
 
 type ContentBlock = AssistantMessage["content"][number];
 
+function isRecord(value: unknown): value is Record<string, unknown> {
+	return typeof value === "object" && value !== null;
+}
+
+/**
+ * The eval tool's prepareArguments shim normalizes run summaries before schema
+ * validation. That harness-owned normalization mutates the committed tool-call
+ * arguments after the provider-final boundary, so continuity must fingerprint
+ * the same effective arguments on both sides. Keep other tools fail-closed.
+ */
+function semanticToolCallArguments(name: string, args: unknown): unknown {
+	if (name !== "eval" || !isRecord(args) || args.action === "peek" || args.action === "stop") return args;
+	if (typeof args.summary !== "string") return args;
+
+	const summary = clampEvalSummary(args.summary);
+	const canonical = { ...args };
+	if (summary === undefined) delete canonical.summary;
+	else canonical.summary = summary;
+	return canonical;
+}
+
 /**
  * Only the payload the model produced is fingerprinted. Everything the stream
  * pipeline stamps around it (thinking timing, content-block indices, partial
  * JSON) can legitimately differ between the last `message_update` and `message_end`
  * without any extension rewriting the answer; hashing such fields marked plain
  * turns `assistant_rewritten` and forced a full re-send on the next turn
- * (senpi#691, oh-my-openagent#7925). An unknown block shape stays fail-closed.
+ * (senpi#691, oh-my-openagent#7925). Harness-owned argument normalization is
+ * canonicalized to the effective tool input; unknown block shapes stay fail-closed.
  */
 function semanticContentBlock(block: ContentBlock): unknown {
 	switch (block.type) {
@@ -21,7 +44,12 @@ function semanticContentBlock(block: ContentBlock): unknown {
 		case "thinking":
 			return { type: block.type, thinking: block.thinking, thinkingSignature: block.thinkingSignature };
 		case "toolCall":
-			return { type: block.type, id: block.id, name: block.name, arguments: block.arguments };
+			return {
+				type: block.type,
+				id: block.id,
+				name: block.name,
+				arguments: semanticToolCallArguments(block.name, block.arguments),
+			};
 		default:
 			return block;
 	}

packages/coding-agent/test/suite/regressions/1472-claude-sdk-oauth-eval-summary-clamp.test.ts

@@ -0,0 +1,85 @@
+import type { AssistantMessage } from "@earendil-works/pi-ai";
+import { describe, expect, it } from "vitest";
+import { createEvalTool } from "../../../../senpi-codemode/src/tool/eval-tool.ts";
+import { CLAUDE_SDK_OAUTH_PROVIDER_ID } from "../../../src/core/extensions/builtin/claude-sdk-oauth/account-management.ts";
+import { AssistantCommitBoundary } from "../../../src/core/extensions/builtin/claude-sdk-oauth/session-commit-boundary.ts";
+
+const MODEL_ID = "claude-opus-5";
+
+const evalTool = createEvalTool({
+	enabledLanguages: { js: true, py: false, rb: false, jl: false },
+	kernelManager: {
+		getKernel: () => Promise.reject(new Error("unused in summary preparation regression")),
+	},
+	cellTimeoutSeconds: 30,
+	executeTool: () => Promise.reject(new Error("unused in summary preparation regression")),
+});
+
+function preparedEvalSummary(summary: string): string {
+	const prepare = evalTool.prepareArguments;
+	if (prepare === undefined) throw new Error("eval tool must define prepareArguments");
+	const prepared = prepare({ language: "js", code: "return 1", summary });
+	if (typeof prepared !== "object" || prepared === null || !("summary" in prepared))
+		throw new Error("prepared eval run must retain a summary");
+	if (typeof prepared.summary !== "string") throw new Error("prepared eval summary must be a string");
+	return prepared.summary;
+}
+
+function evalAssistant(summary: string, code = "return 1"): AssistantMessage {
+	return {
+		role: "assistant",
+		api: CLAUDE_SDK_OAUTH_PROVIDER_ID,
+		provider: CLAUDE_SDK_OAUTH_PROVIDER_ID,
+		model: MODEL_ID,
+		content: [
+			{
+				type: "toolCall",
+				id: "call-1",
+				name: "eval",
+				arguments: { language: "js", code, summary },
+			},
+		],
+		stopReason: "toolUse",
+		timestamp: 1,
+		usage: {
+			input: 1,
+			output: 1,
+			cacheRead: 0,
+			cacheWrite: 0,
+			totalTokens: 2,
+			cost: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				total: 0,
+			},
+		},
+	};
+}
+
+describe("issue #1472: eval summary normalization preserves Claude SDK continuity", () => {
+	it("treats the harness summary clamp as clean using eval prepareArguments", () => {
+		const streamedSummary = "s".repeat(81);
+		const streamed = evalAssistant(streamedSummary);
+		const committed = evalAssistant(preparedEvalSummary(streamedSummary));
+		const boundary = new AssistantCommitBoundary();
+
+		boundary.captureProviderFinal("eval-clamp", streamed);
+		expect(boundary.commit("eval-clamp", committed, MODEL_ID)).toBe("clean");
+	});
+
+	it("still detects a semantic summary rewrite within the schema limit", () => {
+		const boundary = new AssistantCommitBoundary();
+		boundary.captureProviderFinal("summary-rewrite", evalAssistant("inspect cache state"));
+
+		expect(boundary.commit("summary-rewrite", evalAssistant("delete cache state"), MODEL_ID)).toBe("rewritten");
+	});
+
+	it("still detects changes to other eval arguments", () => {
+		const boundary = new AssistantCommitBoundary();
+		boundary.captureProviderFinal("code-rewrite", evalAssistant("inspect result", "return 1"));
+
+		expect(boundary.commit("code-rewrite", evalAssistant("inspect result", "return 2"), MODEL_ID)).toBe("rewritten");
+	});
+});

packages/senpi-codemode/package.json

@@ -8,6 +8,10 @@
 	"types": "./src/index.ts",
 	"exports": {
 		".": "./src/index.ts",
+		"./eval-summary": {
+			"types": "./src/tool/eval-summary.d.mts",
+			"import": "./src/tool/eval-summary.mjs"
+		},
 		"./package.json": "./package.json"
 	},
 	"files": [

packages/senpi-codemode/src/tool/eval-request.ts

@@ -1,20 +1,10 @@
 import type { ExtensionContext } from "@code-yeongyu/senpi";
-import { EVAL_SUMMARY_MAX_LENGTH, type EvalControlInput, type EvalToolInput, type EvalToolRequest } from "./types.ts";
+import { clampEvalSummary } from "./eval-summary.mjs";
+import { type EvalControlInput, type EvalToolInput, type EvalToolRequest } from "./types.ts";
 
-const NON_INTERACTIVE_MODES = new Set(["print", "json"]);
-
-const ELLIPSIS = "...";
+export { clampEvalSummary };
 
-// Harness-side enforcement of the schema maxLength: the tool advertises the limit, but an
-// over-limit value is force-truncated here (prepareArguments runs before schema validation)
-// instead of failing the call.
-export function clampEvalSummary(value: unknown): string | undefined {
-	if (typeof value !== "string") return undefined;
-	const normalized = value.trim().replace(/\s+/gu, " ");
-	if (normalized.length === 0) return undefined;
-	if (normalized.length <= EVAL_SUMMARY_MAX_LENGTH) return normalized;
-	return `${normalized.slice(0, EVAL_SUMMARY_MAX_LENGTH - ELLIPSIS.length)}${ELLIPSIS}`;
-}
+const NON_INTERACTIVE_MODES = new Set(["print", "json"]);
 
 export function parseEvalRequest(params: unknown): EvalToolRequest {
 	if (!isRecord(params)) throw new TypeError("eval parameters must be an object");

packages/senpi-codemode/src/tool/eval-summary.d.mts

@@ -0,0 +1,3 @@
+export declare const EVAL_SUMMARY_MAX_LENGTH = 80;
+
+export declare function clampEvalSummary(value: unknown): string | undefined;

packages/senpi-codemode/src/tool/eval-summary.mjs

@@ -0,0 +1,11 @@
+export const EVAL_SUMMARY_MAX_LENGTH = 80;
+
+const ELLIPSIS = "...";
+
+export function clampEvalSummary(value) {
+	if (typeof value !== "string") return undefined;
+	const normalized = value.trim().replace(/\s+/gu, " ");
+	if (normalized.length === 0) return undefined;
+	if (normalized.length <= EVAL_SUMMARY_MAX_LENGTH) return normalized;
+	return `${normalized.slice(0, EVAL_SUMMARY_MAX_LENGTH - ELLIPSIS.length)}${ELLIPSIS}`;
+}

packages/senpi-codemode/src/tool/types.ts

@@ -8,6 +8,9 @@ import {
 	defaultCodemodeSettings,
 } from "../config/settings.ts";
 import type { TruncationMeta } from "../output/output-meta.ts";
+import { EVAL_SUMMARY_MAX_LENGTH } from "./eval-summary.mjs";
+
+export { EVAL_SUMMARY_MAX_LENGTH };
 
 export const evalLanguageOrder = ["js", "py", "rb", "jl"] as const;
 export type EvalLanguage = (typeof evalLanguageOrder)[number];
@@ -17,8 +20,6 @@ export function enabledLanguageList(enabled: EnabledEvalLanguages): EvalLanguage
 	return evalLanguageOrder.filter((language) => enabled[language]);
 }
 
-export const EVAL_SUMMARY_MAX_LENGTH = 80;
-
 /** The deadlines the schema teaches the model; every number comes from the resolved settings. */
 export interface EvalDeadlineSeconds {
 	readonly runBudgetSeconds: number;