sorokin.oleg.edu/learnharness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

fix: enforce Claude CLI MCP admission and safe continuation

Browse Source
This commit is contained in:
amosblomqvist
2026-10-28 08:01:50 +02:00
parent 812ea23419
commit 6ca431644d
13 changed files with 534 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/claude-code-harness.test.ts
+29 -1
View File
@@ -55,7 +55,7 @@ function toAskScenario(): string[] {
streamEvent({ type: "content_block_stop", index: 0 }),
streamEvent({ type: "message_delta", delta: { stop_reason: "tool_use" }, usage: {} }),
streamEvent({ type: "message_stop" }),
JSON.stringify({ fake: "call", name: "ask_user_question", arguments: ASK_INPUT }),
JSON.stringify({ fake: "call", id: ASK_ID, name: "ask_user_question", arguments: ASK_INPUT }),
JSON.stringify({ fake: "reply" }),
];
}
@@ -168,6 +168,34 @@ test("a question answered before the idle timeout continues the waiting CLI proc
await waitForExit(waiting);
});
test("a question pauses later calls from the same CLI response until the learner answers", async (t) => {
const secondId = "toolu_second_question";
const secondQuestion = "What do you want to learn next?";
const scenario = toAskScenario();
scenario.splice(4, 0,
streamEvent({ type: "content_block_start", index: 1, content_block: { type: "tool_use", id: secondId, name: "mcp__learnharness__ask_user_question", input: {} } }),
streamEvent({ type: "content_block_delta", index: 1, delta: { type: "input_json_delta", partial_json: JSON.stringify({ question: secondQuestion }) } }),
streamEvent({ type: "content_block_stop", index: 1 }),
);
scenario.splice(scenario.length - 1, 0, JSON.stringify({ fake: "call", id: secondId, name: "ask_user_question", arguments: { question: secondQuestion } }));
const fakeRun = createFakeRun(t, [scenario]);
const factory = createAgentFactory(fakeRun);
const chatId = createClaudeCodeChat(t);
await runChatTurn(chatId, { type: "text", text: "teach me sockets" }, factory);
assert.equal(danglingQuestionOf(chatStore.getChat(chatId)?.messages ?? [])?.toolCallId, ASK_ID);
assert.equal(JSON.stringify(chatStore.getChat(chatId)?.messages).includes(secondId), false, "unadmitted second question must not be posed or persisted");
await runChatTurn(chatId, toAnswerInput(), factory);
assert.equal(danglingQuestionOf(chatStore.getChat(chatId)?.messages ?? [])?.toolCallId, secondId);
await runChatTurn(chatId, {
type: "answer", toolCallId: secondId,
answer: { kind: "ask", questionText: secondQuestion, dontKnow: false, answers: [], text: "HTTP next", answeredAt: new Date().toISOString() },
}, factory);
assert.equal(fakeRun.children.length, 1);
assert.ok(getLastAssistantText(chatId).includes(ANSWER_TEXT));
assert.ok(getLastAssistantText(chatId).includes("HTTP next"));
await waitForExit(fakeRun.children[0]!);
});
test("a question answered after the idle timeout starts a fresh CLI process that replays the question and answer", async (t) => {
liveProcessIdleTimeoutMs.current = 300;
t.after(() => (liveProcessIdleTimeoutMs.current = LIVE_PROCESS_IDLE_TIMEOUT_MS));