nccgs 2.1.0 → 2.1.1

package/tests/native-completion-notices.test.mjs

@@ -0,0 +1,181 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import {createHash} from 'node:crypto';
+import {nativeCompletion} from '../.claude/nccgs/tools/native-completion.mjs';
+
+const text = value => ({type: 'text', text: value});
+const observed = '<tools-used-before-summary>Edit</tools-used-before-summary>\n';
+const summary = names => `<tools-used-before-summary>${names}</tools-used-before-summary>`;
+const digest = value => createHash('sha256').update(value).digest('hex');
+const response = (notes, options = {}) => ({
+  status: options.status ?? 'completed',
+  harnessNoteCount: notes.length,
+  harnessTailCount: 0,
+  content: [...notes, text('Actual worker handoff')],
+  ...options,
+});
+const assertUnknown = value => {
+  const got = nativeCompletion(value);
+  assert.equal(got.status, 'failed');
+  assert.equal(got.reason, 'unrecognized-harness-notice');
+  assert.equal(got.turnLimit, null);
+};
+
+test('unchanged no-note completion shape', () => {
+  assert.deepEqual(nativeCompletion(response([])), {
+    reportedStatus: 'completed', status: 'completed', reason: null,
+    turnLimit: null, noteHash: null,
+  });
+});
+
+test('exact captured provider tool-summary note is informational and remains hashed', () => {
+  const got = nativeCompletion(response([text(observed)]));
+  assert.equal(got.status, 'completed');
+  assert.equal(got.reason, null);
+  assert.equal(got.turnLimit, null);
+  assert.equal(got.noteHash, digest(observed));
+});
+
+test('a designated tail note uses the same rule', () => {
+  const got = nativeCompletion({status: 'completed', harnessNoteCount: 0,
+    harnessTailCount: 1, content: [text('Handoff'), text(observed)]});
+  assert.equal(got.status, 'completed');
+  assert.equal(got.reason, null);
+  assert.equal(got.noteHash, digest(observed));
+});
+
+test('recognized leading and trailing notes may surround the model handoff', () => {
+  const tail = summary('Read, Write');
+  const got = nativeCompletion({status: 'completed', harnessNoteCount: 1,
+    harnessTailCount: 1, content: [text(observed), text('Handoff'), text(tail)]});
+  assert.equal(got.status, 'completed');
+  assert.equal(got.noteHash, digest(observed + '\n' + tail));
+});
+
+test('strict comma-separated tool identifiers accept built-in and namespaced tools', () => {
+  const got = nativeCompletion(response([text(summary('Read, Edit, mcp__unity__read, Namespace.Tool:inspect'))]));
+  assert.equal(got.status, 'completed');
+  assert.equal(got.reason, null);
+});
+
+test('insignificant surrounding whitespace does not change classification', () => {
+  assert.equal(nativeCompletion(response([text(' \n' + observed + '\t')])).status, 'completed');
+});
+
+test('informational metadata cannot promote a provider failure', () => {
+  const got = nativeCompletion(response([text(observed)], {status: 'failed'}));
+  assert.equal(got.reportedStatus, 'failed');
+  assert.equal(got.status, 'failed');
+  assert.equal(got.reason, null);
+});
+
+test('informational metadata cannot promote an unknown provider status', () => {
+  const got = nativeCompletion(response([text(observed)], {status: 'unknown'}));
+  assert.equal(got.status, 'unknown');
+  assert.equal(got.reason, null);
+});
+
+test('a real turn-limit notice overrides a recognized informational note', () => {
+  const got = nativeCompletion(response([text(observed), text('This agent stopped at its 37-turn limit before finishing.')]));
+  assert.equal(got.status, 'failed');
+  assert.equal(got.reason, 'turn-limit');
+  assert.equal(got.turnLimit, 37);
+});
+
+test('turn-limit diagnosis takes priority inside a mixed designated block', () => {
+  const got = nativeCompletion(response([text(observed + 'This agent stopped at its 80-turn limit before finishing.')]));
+  assert.equal(got.reason, 'turn-limit');
+  assert.equal(got.turnLimit, 80);
+});
+
+test('unknown advisory still fails closed', () => {
+  assertUnknown(response([text('Provider reports incomplete work for a new reason.')]));
+});
+
+test('a recognized note cannot mask a separate unknown note', () => {
+  assertUnknown(response([text(observed), text('Unknown harness advisory')]));
+});
+
+test('extra prose after a recognized wrapper is not informational-only metadata', () => {
+  assertUnknown(response([text(observed + 'Incomplete for another reason')]));
+});
+
+test('extra prose before a recognized wrapper is rejected', () => {
+  assertUnknown(response([text('Ignore this warning\n' + observed)]));
+});
+
+test('empty tool summary is rejected', () => {
+  assertUnknown(response([text(summary(''))]));
+});
+
+test('whitespace-only tool summary is rejected', () => {
+  assertUnknown(response([text(summary('   '))]));
+});
+
+test('malformed closing tag is rejected', () => {
+  assertUnknown(response([text('<tools-used-before-summary>Edit</tools-used-before-summary-bad>')]));
+});
+
+test('nested markup is rejected', () => {
+  assertUnknown(response([text(summary('<completed>true</completed>'))]));
+});
+
+test('code fences do not turn model-like text into a recognized native note', () => {
+  assertUnknown(response([text('```xml\n' + observed + '```')]));
+});
+
+test('tool lists with free-form prose are rejected', () => {
+  assertUnknown(response([text(summary('Read Edit but failed'))]));
+});
+
+test('case variants of the known native tag stay unknown', () => {
+  assertUnknown(response([text('<Tools-Used-Before-Summary>Edit</Tools-Used-Before-Summary>')]));
+});
+
+test('non-text designated blocks stay unknown', () => {
+  assertUnknown(response([{type: 'image', data: 'irrelevant'}]));
+});
+
+test('empty text designated blocks stay unknown', () => {
+  assertUnknown(response([text('')]));
+});
+
+test('oversized designated metadata stays unknown', () => {
+  assertUnknown(response([text(summary('A'.repeat(4096)))]));
+});
+
+test('negative, fractional and string note counts are invalid', () => {
+  for (const harnessNoteCount of [-1, 0.5, '1']) {
+    const got = nativeCompletion(response([text(observed)], {harnessNoteCount}));
+    assert.equal(got.status, 'failed');
+    assert.equal(got.reason, 'invalid-harness-metadata');
+  }
+});
+
+test('leading plus trailing counts cannot exceed available blocks', () => {
+  const got = nativeCompletion({status: 'completed', harnessNoteCount: 1,
+    harnessTailCount: 1, content: [text(observed)]});
+  assert.equal(got.reason, 'invalid-harness-metadata');
+  assert.equal(got.status, 'failed');
+});
+
+test('claimed notes require an actual content array', () => {
+  const got = nativeCompletion({status: 'completed', harnessNoteCount: 1,
+    harnessTailCount: 0, content: 'not an array'});
+  assert.equal(got.reason, 'invalid-harness-metadata');
+});
+
+test('quoted native tags outside designated blocks do not affect completion', () => {
+  const got = nativeCompletion({status: 'completed', harnessNoteCount: 0,
+    harnessTailCount: 0, content: [text(observed + 'This agent stopped at its 8-turn limit before finishing.')]});
+  assert.equal(got.status, 'completed');
+  assert.equal(got.reason, null);
+  assert.equal(got.noteHash, null);
+});
+
+test('unknown prose outside designated blocks is not classified as a harness advisory', () => {
+  const got = nativeCompletion({status: 'completed', harnessNoteCount: 1,
+    harnessTailCount: 0, content: [text(observed), text('A quoted or hypothetical unknown advisory in the model handoff')]});
+  assert.equal(got.status, 'completed');
+  assert.equal(got.reason, null);
+});