MythEclipse/shiro-neko
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: enhance nudge logic to suppress notifications during todo_write

Browse Source
This commit is contained in:
asepharyana
2026-09-11 18:52:50 +07:00
parent 76043aeb38
commit 44bb4bebf1
2 changed files with 55 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/workflow.test.ts
+46 -14
View File
@@ -184,13 +184,25 @@ test('nudge ladder: fires up to 3 times, then stops', async () => {
await Bun.write(join(dir, 'TODO.md'), '# Todo\n- [ ] task\n');
await Bun.write(join(dir, 'app.ts'), 'const a = 1;\n');
// Each SDK turn is one session.send(). The nudge fires once per turn that
// wrote files without updating the task list — so the model must edit in
// one turn, then end the turn with text (which lets the nudge fire), then
// edit again next turn. Three nudges ⇒ three turns each with an edit and
// a following text-only turn (the session re-invokes the model per turn,
// so 'done' ends that turn and the loop stops).
let call = 0;
const session = new Session({
model: new MockLanguageModelV4({
doStream: async () => {
call++;
if (call <= 3) return stream(toolCall(`c${call}`, 'edit_file', { path: 'app.ts', oldString: 'const a = 1;', newString: `const a = ${call + 1};` }));
// Turn 4+: no tool call — agent stops
// Turn boundaries: every send() calls doStream once. Odd calls (1,3,5)
// edit; even calls (2,4,6) return text to end the turn. call>=7 → text
// (agent finished after the 3rd nudge).
if (call <= 6 && call % 2 === 1) {
const content = (await Bun.file(join(dir, 'app.ts')).text()).trimEnd();
const next = Math.ceil(call / 2) + 1;
return stream(toolCall(`c${call}`, 'edit_file', { path: 'app.ts', oldString: content, newString: `const a = ${next};` }));
}
return stream(text('done'));
},
}),
@@ -198,36 +210,54 @@ test('nudge ladder: fires up to 3 times, then stops', async () => {
});
const notices: string[] = [];
for (let i = 0; i < 5; i++) {
for (let i = 0; i < 10; i++) {
for await (const ev of session.send(`turn ${i + 1}`)) {
if (ev.type === 'notice') notices.push(ev.text);
}
}
const nudgeNotices = notices.filter((n) => n.includes('without updating the project task list'));
const nudgeNotices = notices.filter((n) => n.includes('TODO.md'));
expect(nudgeNotices.length).toBe(3);
} finally {
rmSync(dir, { recursive: true, force: true });
}
});
test('nudge resets after todo_write in a later turn', async () => {
test('todo_write in a turn suppresses that turn\'s nudge', async () => {
const dir = mkdtempSync(join(tmpdir(), 'shiro-wf-reset'));
try {
await Bun.write(join(dir, '.git', 'HEAD'), 'ref: refs/heads/main\n');
await Bun.write(join(dir, 'TODO.md'), '# Todo\n- [ ] task\n');
await Bun.write(join(dir, 'app.ts'), 'const a = 1;\n');
// One session.send() = one full agent run (the SDK loop re-invokes the
// model until it returns text). The nudge fires once per send, and the
// counter is a *lifetime* cap of 3 — todo_write only suppresses the nudge
// for the turn in which it runs:
// call 1 (edit+todo_write), call 2 text -> send 1 -> no nudge
// call 3 (edit), call 4 text -> send 2 -> nudge 1
// call 5 (edit), call 6 text -> send 3 -> nudge 2
// call 7 (edit), call 8 text -> send 4 -> nudge 3
// call 9 (edit), call 10 text -> send 5 -> capped, no nudge
let call = 0;
const session = new Session({
model: new MockLanguageModelV4({
doStream: async () => {
call++;
// Turn 1-3: edit file → 3 nudges
if (call <= 3) return stream(toolCall(`c${call}`, 'edit_file', { path: 'app.ts', oldString: 'const a = 1;', newString: `const a = ${call + 1};` }));
// Turn 4: todo_write → resets the nudge counter
if (call === 4) return stream(toolCall(`c${call}`, 'todo_write', { items: [{ text: 'completed task', done: true }] }));
// Turn 5: edit file → should nudge again (counter was reset)
return stream(toolCall(`c${call}`, 'edit_file', { path: 'app.ts', oldString: 'const a = 4;', newString: `const a = ${call + 1};` }));
const c = await Bun.file(join(dir, 'app.ts')).text();
const next = Math.floor((call + 1) / 2) + 1;
if (call === 1) {
// Same send: edit + todo_write -> nudge suppressed for this turn
return stream([
...toolCall('e1', 'edit_file', { path: 'app.ts', oldString: c.trimEnd(), newString: `const a = ${next};` }),
...toolCall('w', 'todo_write', { items: [{ text: 'completed task', done: true }] }),
]);
}
if (call % 2 === 1) {
// Odd calls (3,5,7,9): edit -> each ends a send
return stream(toolCall(`e${call}`, 'edit_file', { path: 'app.ts', oldString: c.trimEnd(), newString: `const a = ${next};` }));
}
// Even calls: text ends the send
return stream(text('done'));
},
}),
askApproval: async () => 'once',
@@ -239,9 +269,11 @@ test('nudge resets after todo_write in a later turn', async () => {
if (ev.type === 'notice') notices.push(ev.text);
}
}
const nudgeNotices = notices.filter((n) => n.includes('without updating the project task list'));
// 3 nudges (turns 1-3) + 1 reset + 1 more nudge (turn 5) = 4
expect(nudgeNotices.length).toBe(4);
const nudgeNotices = notices.filter((n) => n.includes('TODO.md'));
// Send 1 suppressed (todo_write), sends 2-4 nudges 1-3, send 5 capped.
expect(nudgeNotices.length).toBe(3);
// The messages escalate (1st/2nd/3rd), proving the ladder.
expect(nudgeNotices[0] ?? '').toContain('reminder:');
} finally {
rmSync(dir, { recursive: true, force: true });
}