throughline 0.3.24 → 0.4.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.claude/commands/tl.md +6 -21
- package/.codex-sidecar.yml +62 -0
- package/CHANGELOG.md +632 -0
- package/README.ja.md +71 -46
- package/README.md +420 -76
- package/bin/throughline.mjs +169 -7
- package/codex/skills/throughline/SKILL.md +157 -0
- package/codex/skills/throughline/agents/openai.yaml +7 -0
- package/docs/INHERITANCE_ON_CLEAR_ONLY.md +159 -0
- package/docs/L1_L2_L3_REDESIGN.md +415 -0
- package/docs/PUBLIC_RELEASE_PLAN.md +185 -0
- package/docs/THROUGHLINE_CLEAR_AUTO_HANDOFF_PLAN.md +286 -0
- package/docs/THROUGHLINE_CODEX_DUAL_SUPPORT.md +249 -0
- package/docs/THROUGHLINE_CODEX_FIRST_ROADMAP.md +555 -0
- package/docs/THROUGHLINE_CODEX_MONITOR_IMPLEMENTATION_PLAN.md +220 -0
- package/docs/THROUGHLINE_CODEX_TRIM_IMPLEMENTATION_PLAN.md +528 -0
- package/docs/THROUGHLINE_CODEX_TRIM_ROLLBACK_FIX_PLAN.md +672 -0
- package/docs/archive/CONCEPT.md +476 -0
- package/docs/archive/EXPERIMENT.md +371 -0
- package/docs/archive/README.md +22 -0
- package/docs/archive/SESSION_LINKING_DESIGN.md +231 -0
- package/docs/archive/THROUGHLINE_NEXT_STEPS.md +134 -0
- package/docs/throughline-codex-trim-rollback-incident-report.md +306 -0
- package/docs/throughline-handoff-context.example.json +57 -0
- package/docs/throughline-rollback-context-trim-insight.md +455 -0
- package/package.json +6 -2
- package/src/baton.mjs +17 -45
- package/src/baton.test.mjs +4 -41
- package/src/cli/codex-capture.mjs +95 -0
- package/src/cli/codex-handoff-model-smoke.mjs +292 -0
- package/src/cli/codex-handoff-model-smoke.test.mjs +262 -0
- package/src/cli/codex-handoff-smoke.mjs +163 -0
- package/src/cli/codex-handoff-smoke.test.mjs +149 -0
- package/src/cli/codex-handoff-start.mjs +291 -0
- package/src/cli/codex-handoff-start.test.mjs +194 -0
- package/src/cli/codex-hook.mjs +276 -0
- package/src/cli/codex-hook.test.mjs +293 -0
- package/src/cli/codex-host-primitive-audit.mjs +110 -0
- package/src/cli/codex-host-primitive-audit.test.mjs +75 -0
- package/src/cli/codex-restore-smoke.mjs +357 -0
- package/src/cli/codex-restore-source-audit.mjs +304 -0
- package/src/cli/codex-resume.mjs +138 -0
- package/src/cli/codex-rollback-model-visible-smoke.mjs +373 -0
- package/src/cli/codex-rollback-model-visible-smoke.test.mjs +255 -0
- package/src/cli/codex-sidecar-diagnostics.mjs +48 -0
- package/src/cli/codex-sidecar-dry-run.mjs +85 -0
- package/src/cli/codex-summarize.mjs +224 -0
- package/src/cli/codex-threads.mjs +89 -0
- package/src/cli/codex-visibility-smoke.mjs +196 -0
- package/src/cli/codex-vscode-restore-smoke.mjs +226 -0
- package/src/cli/codex-vscode-rollback-smoke.mjs +114 -0
- package/src/cli/doctor.mjs +503 -1
- package/src/cli/doctor.test.mjs +542 -3
- package/src/cli/handoff-preview.mjs +78 -0
- package/src/cli/help.test.mjs +64 -0
- package/src/cli/install.mjs +226 -3
- package/src/cli/install.test.mjs +205 -4
- package/src/cli/trim.mjs +564 -0
- package/src/codex-app-server.mjs +1816 -0
- package/src/codex-app-server.test.mjs +512 -0
- package/src/codex-auto-refresh.mjs +194 -0
- package/src/codex-auto-refresh.test.mjs +182 -0
- package/src/codex-capture.mjs +235 -0
- package/src/codex-capture.test.mjs +393 -0
- package/src/codex-handoff-model-smoke.mjs +114 -0
- package/src/codex-handoff-model-smoke.test.mjs +89 -0
- package/src/codex-handoff-smoke.mjs +124 -0
- package/src/codex-handoff-smoke.test.mjs +103 -0
- package/src/codex-handoff.mjs +331 -0
- package/src/codex-handoff.test.mjs +220 -0
- package/src/codex-host-primitive-audit.mjs +374 -0
- package/src/codex-host-primitive-audit.test.mjs +208 -0
- package/src/codex-restore-smoke.test.mjs +639 -0
- package/src/codex-restore-source-audit.mjs +1348 -0
- package/src/codex-restore-source-audit.test.mjs +623 -0
- package/src/codex-resume.test.mjs +242 -0
- package/src/codex-rollout-memory.mjs +711 -0
- package/src/codex-rollout-memory.test.mjs +610 -0
- package/src/codex-sidecar-cli.test.mjs +75 -0
- package/src/codex-sidecar.mjs +246 -0
- package/src/codex-sidecar.test.mjs +172 -0
- package/src/codex-summarize.test.mjs +143 -0
- package/src/codex-thread-identity.mjs +23 -0
- package/src/codex-thread-index.mjs +173 -0
- package/src/codex-thread-index.test.mjs +164 -0
- package/src/codex-usage.mjs +110 -0
- package/src/codex-usage.test.mjs +140 -0
- package/src/codex-visibility-smoke.test.mjs +222 -0
- package/src/codex-vscode-restore-smoke.mjs +206 -0
- package/src/codex-vscode-restore-smoke.test.mjs +325 -0
- package/src/codex-vscode-rollback-smoke.mjs +90 -0
- package/src/codex-vscode-rollback-smoke.test.mjs +290 -0
- package/src/db-schema.test.mjs +96 -0
- package/src/db.mjs +14 -1
- package/src/haiku-summarizer.mjs +267 -26
- package/src/haiku-summarizer.test.mjs +282 -0
- package/src/handoff-preview.test.mjs +108 -0
- package/src/handoff-record.mjs +294 -0
- package/src/handoff-record.test.mjs +226 -0
- package/src/hook-entrypoints.test.mjs +286 -0
- package/src/package-files.test.mjs +19 -0
- package/src/prompt-submit.mjs +9 -6
- package/src/resume-context.mjs +58 -171
- package/src/resume-context.test.mjs +177 -0
- package/src/session-start.mjs +85 -26
- package/src/state-file.mjs +50 -6
- package/src/state-file.test.mjs +50 -0
- package/src/token-monitor.mjs +14 -10
- package/src/token-monitor.test.mjs +27 -0
- package/src/trim-cli.test.mjs +1584 -0
- package/src/trim-model.mjs +584 -0
- package/src/trim-model.test.mjs +568 -0
- package/src/turn-processor.mjs +17 -10
- package/src/vscode-task.mjs +33 -10
- package/src/vscode-task.test.mjs +19 -9
- package/src/cli/save-inflight.mjs +0 -81
|
@@ -0,0 +1,639 @@
|
|
|
1
|
+
import { test } from 'node:test';
|
|
2
|
+
import assert from 'node:assert/strict';
|
|
3
|
+
import { spawnSync } from 'node:child_process';
|
|
4
|
+
import { chmodSync, mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
|
|
5
|
+
import { tmpdir } from 'node:os';
|
|
6
|
+
import { dirname, join } from 'node:path';
|
|
7
|
+
import { fileURLToPath } from 'node:url';
|
|
8
|
+
|
|
9
|
+
const REPO_ROOT = dirname(dirname(fileURLToPath(import.meta.url)));
|
|
10
|
+
|
|
11
|
+
function makeTempHome() {
|
|
12
|
+
return mkdtempSync(join(tmpdir(), 'tl-codex-restore-home-'));
|
|
13
|
+
}
|
|
14
|
+
|
|
15
|
+
function makeTempProject() {
|
|
16
|
+
return mkdtempSync(join(tmpdir(), 'tl-codex-restore-project-'));
|
|
17
|
+
}
|
|
18
|
+
|
|
19
|
+
function makeFakeCodexAppServer(
|
|
20
|
+
dir,
|
|
21
|
+
{
|
|
22
|
+
turnCounts = [22, 22],
|
|
23
|
+
turnsListCounts = null,
|
|
24
|
+
supportsTurnsList = true,
|
|
25
|
+
retainedText = null,
|
|
26
|
+
retainedTextLocation = 'itemText',
|
|
27
|
+
} = {},
|
|
28
|
+
) {
|
|
29
|
+
const script = join(dir, 'fake-codex-restore-app-server.mjs');
|
|
30
|
+
const state = join(dir, 'fake-codex-restore-state.json');
|
|
31
|
+
writeFileSync(
|
|
32
|
+
script,
|
|
33
|
+
`#!/usr/bin/env node
|
|
34
|
+
import { existsSync, readFileSync, writeFileSync } from 'node:fs';
|
|
35
|
+
import { createInterface } from 'node:readline';
|
|
36
|
+
|
|
37
|
+
const statePath = ${JSON.stringify(state)};
|
|
38
|
+
const turnCounts = ${JSON.stringify(turnCounts)};
|
|
39
|
+
const turnsListCounts = ${JSON.stringify(turnsListCounts)};
|
|
40
|
+
const supportsTurnsList = ${JSON.stringify(supportsTurnsList)};
|
|
41
|
+
const retainedText = ${JSON.stringify(retainedText)};
|
|
42
|
+
const retainedTextLocation = ${JSON.stringify(retainedTextLocation)};
|
|
43
|
+
let launches = 0;
|
|
44
|
+
if (existsSync(statePath)) launches = JSON.parse(readFileSync(statePath, 'utf8')).launches;
|
|
45
|
+
const turnCount = turnCounts[Math.min(launches, turnCounts.length - 1)];
|
|
46
|
+
const turnsListCountSource = Array.isArray(turnsListCounts) ? turnsListCounts : turnCounts;
|
|
47
|
+
const turnsListCount = turnsListCountSource[Math.min(launches, turnsListCountSource.length - 1)];
|
|
48
|
+
writeFileSync(statePath, JSON.stringify({ launches: launches + 1 }));
|
|
49
|
+
function buildTurn(index) {
|
|
50
|
+
const turn = { id: 'turn-' + (index + 1) };
|
|
51
|
+
if (retainedText && index === 0) {
|
|
52
|
+
if (retainedTextLocation === 'aggregatedOutput') {
|
|
53
|
+
turn.items = [{ type: 'commandExecution', aggregatedOutput: retainedText }];
|
|
54
|
+
} else if (retainedTextLocation === 'replacementHistory') {
|
|
55
|
+
turn.replacement_history = [
|
|
56
|
+
{
|
|
57
|
+
type: 'message',
|
|
58
|
+
role: 'user',
|
|
59
|
+
content: [{ type: 'input_text', text: retainedText }],
|
|
60
|
+
},
|
|
61
|
+
];
|
|
62
|
+
} else {
|
|
63
|
+
turn.items = [{ type: 'userMessage', text: retainedText }];
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
return turn;
|
|
67
|
+
}
|
|
68
|
+
const turns = Array.from({ length: turnCount }, (_, index) => buildTurn(index));
|
|
69
|
+
const listedTurns = Array.from({ length: turnsListCount }, (_, index) => buildTurn(index));
|
|
70
|
+
const rl = createInterface({ input: process.stdin });
|
|
71
|
+
function send(message) { process.stdout.write(JSON.stringify(message) + '\\n'); }
|
|
72
|
+
rl.on('line', (line) => {
|
|
73
|
+
const msg = JSON.parse(line);
|
|
74
|
+
if (msg.method === 'initialized') return;
|
|
75
|
+
if (msg.method === 'initialize') {
|
|
76
|
+
send({ id: msg.id, result: { userAgent: 'fake-codex' } });
|
|
77
|
+
} else if (msg.method === 'thread/read' || msg.method === 'thread/resume') {
|
|
78
|
+
send({ id: msg.id, result: { thread: { id: msg.params.threadId, turns } } });
|
|
79
|
+
} else if (supportsTurnsList && msg.method === 'thread/turns/list') {
|
|
80
|
+
send({ id: msg.id, result: { data: listedTurns, nextCursor: null, backwardsCursor: null } });
|
|
81
|
+
} else {
|
|
82
|
+
send({ id: msg.id, error: { code: -32601, message: 'unknown method' } });
|
|
83
|
+
}
|
|
84
|
+
});
|
|
85
|
+
`,
|
|
86
|
+
);
|
|
87
|
+
chmodSync(script, 0o755);
|
|
88
|
+
return script;
|
|
89
|
+
}
|
|
90
|
+
|
|
91
|
+
function runRestoreSmoke(home, codexHome, project, args = [], extraEnv = {}) {
|
|
92
|
+
return spawnSync(
|
|
93
|
+
process.execPath,
|
|
94
|
+
[join(REPO_ROOT, 'bin/throughline.mjs'), 'codex-restore-smoke', ...args],
|
|
95
|
+
{
|
|
96
|
+
cwd: project,
|
|
97
|
+
env: {
|
|
98
|
+
...process.env,
|
|
99
|
+
HOME: home,
|
|
100
|
+
USERPROFILE: home,
|
|
101
|
+
CODEX_HOME: codexHome,
|
|
102
|
+
...extraEnv,
|
|
103
|
+
},
|
|
104
|
+
encoding: 'utf8',
|
|
105
|
+
},
|
|
106
|
+
);
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
function writeCodexRollout(codexHome, { project, threadId, turnCount, restoreRisk = false }) {
|
|
110
|
+
const dir = join(codexHome, 'sessions', '2026', '05', '07');
|
|
111
|
+
mkdirSync(dir, { recursive: true });
|
|
112
|
+
const path = join(dir, `rollout-2026-05-07T00-00-00-${threadId}.jsonl`);
|
|
113
|
+
const rows = [
|
|
114
|
+
{
|
|
115
|
+
timestamp: '2026-05-07T00:00:00.000Z',
|
|
116
|
+
type: 'session_meta',
|
|
117
|
+
payload: {
|
|
118
|
+
id: threadId,
|
|
119
|
+
timestamp: '2026-05-07T00:00:00.000Z',
|
|
120
|
+
cwd: project,
|
|
121
|
+
source: 'vscode',
|
|
122
|
+
cli_version: '0.128.0-alpha.1',
|
|
123
|
+
},
|
|
124
|
+
},
|
|
125
|
+
];
|
|
126
|
+
|
|
127
|
+
for (let turn = 1; turn <= turnCount; turn++) {
|
|
128
|
+
rows.push({
|
|
129
|
+
timestamp: `2026-05-07T00:00:${String(turn).padStart(2, '0')}.000Z`,
|
|
130
|
+
type: 'event_msg',
|
|
131
|
+
payload: {
|
|
132
|
+
type: 'user_message',
|
|
133
|
+
message: `restore smoke user turn ${turn}`,
|
|
134
|
+
},
|
|
135
|
+
});
|
|
136
|
+
rows.push({
|
|
137
|
+
timestamp: `2026-05-07T00:00:${String(turn).padStart(2, '0')}.100Z`,
|
|
138
|
+
type: 'event_msg',
|
|
139
|
+
payload: { type: 'task_started' },
|
|
140
|
+
});
|
|
141
|
+
rows.push({
|
|
142
|
+
timestamp: `2026-05-07T00:00:${String(turn).padStart(2, '0')}.200Z`,
|
|
143
|
+
type: 'event_msg',
|
|
144
|
+
payload: {
|
|
145
|
+
type: 'agent_message',
|
|
146
|
+
message: `restore smoke assistant turn ${turn}`,
|
|
147
|
+
},
|
|
148
|
+
});
|
|
149
|
+
rows.push({
|
|
150
|
+
timestamp: `2026-05-07T00:00:${String(turn).padStart(2, '0')}.300Z`,
|
|
151
|
+
type: 'event_msg',
|
|
152
|
+
payload: { type: 'task_complete' },
|
|
153
|
+
});
|
|
154
|
+
}
|
|
155
|
+
|
|
156
|
+
if (restoreRisk) {
|
|
157
|
+
const riskyText = `restore smoke user turn ${turnCount}`;
|
|
158
|
+
rows.push({
|
|
159
|
+
timestamp: '2026-05-07T00:01:00.000Z',
|
|
160
|
+
type: 'compacted',
|
|
161
|
+
payload: {
|
|
162
|
+
message: '',
|
|
163
|
+
replacement_history: [
|
|
164
|
+
{
|
|
165
|
+
type: 'message',
|
|
166
|
+
role: 'user',
|
|
167
|
+
content: [{ type: 'input_text', text: riskyText }],
|
|
168
|
+
},
|
|
169
|
+
],
|
|
170
|
+
},
|
|
171
|
+
});
|
|
172
|
+
rows.push({
|
|
173
|
+
timestamp: '2026-05-07T00:01:00.100Z',
|
|
174
|
+
type: 'event_msg',
|
|
175
|
+
payload: { type: 'thread_rolled_back', num_turns: 1 },
|
|
176
|
+
});
|
|
177
|
+
rows.push({
|
|
178
|
+
timestamp: '2026-05-07T00:01:00.200Z',
|
|
179
|
+
type: 'event_msg',
|
|
180
|
+
payload: { type: 'user_message', message: riskyText },
|
|
181
|
+
});
|
|
182
|
+
}
|
|
183
|
+
|
|
184
|
+
writeFileSync(path, rows.map((row) => JSON.stringify(row)).join('\n') + '\n');
|
|
185
|
+
return path;
|
|
186
|
+
}
|
|
187
|
+
|
|
188
|
+
test('codex-restore-smoke refuses without explicit experimental env', () => {
|
|
189
|
+
const home = makeTempHome();
|
|
190
|
+
const codexHome = makeTempHome();
|
|
191
|
+
const project = makeTempProject();
|
|
192
|
+
try {
|
|
193
|
+
const result = runRestoreSmoke(home, codexHome, project, ['--codex-thread-id', 'thread-restore', '--json']);
|
|
194
|
+
|
|
195
|
+
assert.equal(result.status, 1);
|
|
196
|
+
const payload = JSON.parse(result.stdout);
|
|
197
|
+
assert.equal(payload.status, 'refused');
|
|
198
|
+
assert.equal(payload.reason, 'experimental_env_required');
|
|
199
|
+
assert.equal(payload.restartSafe, false);
|
|
200
|
+
} finally {
|
|
201
|
+
rmSync(project, { recursive: true, force: true });
|
|
202
|
+
rmSync(home, { recursive: true, force: true });
|
|
203
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
204
|
+
}
|
|
205
|
+
});
|
|
206
|
+
|
|
207
|
+
test('codex-restore-smoke reports stable fresh app-server restore counts', () => {
|
|
208
|
+
const home = makeTempHome();
|
|
209
|
+
const codexHome = makeTempHome();
|
|
210
|
+
const project = makeTempProject();
|
|
211
|
+
const threadId = '019dfdef-0000-7000-8000-000000000001';
|
|
212
|
+
try {
|
|
213
|
+
writeCodexRollout(codexHome, { project, threadId, turnCount: 22 });
|
|
214
|
+
const fake = makeFakeCodexAppServer(project, { turnCounts: [22, 22] });
|
|
215
|
+
const result = runRestoreSmoke(
|
|
216
|
+
home,
|
|
217
|
+
codexHome,
|
|
218
|
+
project,
|
|
219
|
+
['--codex-thread-id', threadId, '--codex-app-server-bin', fake, '--json'],
|
|
220
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
221
|
+
);
|
|
222
|
+
|
|
223
|
+
assert.equal(result.status, 0, result.stderr);
|
|
224
|
+
const payload = JSON.parse(result.stdout);
|
|
225
|
+
assert.equal(payload.status, 'app-server-restart-stable');
|
|
226
|
+
assert.equal(payload.reason, 'fresh_app_server_restore_counts_stable');
|
|
227
|
+
assert.equal(payload.proofScope, 'app_server_process_restart_only');
|
|
228
|
+
assert.equal(payload.restartSafe, false);
|
|
229
|
+
assert.equal(payload.expectedTurns, 22);
|
|
230
|
+
assert.equal(payload.observations.length, 2);
|
|
231
|
+
assert.equal(payload.observations[0].turnCountCheck.status, 'match');
|
|
232
|
+
assert.equal(payload.observations[1].turnCountCheck.status, 'match');
|
|
233
|
+
assert.equal(payload.observations[0].turnsListTurns, 22);
|
|
234
|
+
assert.equal(payload.observations[0].turnsListComplete, true);
|
|
235
|
+
} finally {
|
|
236
|
+
rmSync(project, { recursive: true, force: true });
|
|
237
|
+
rmSync(home, { recursive: true, force: true });
|
|
238
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
239
|
+
}
|
|
240
|
+
});
|
|
241
|
+
|
|
242
|
+
test('codex-restore-smoke reports mismatch when fresh app-server restore counts drift', () => {
|
|
243
|
+
const home = makeTempHome();
|
|
244
|
+
const codexHome = makeTempHome();
|
|
245
|
+
const project = makeTempProject();
|
|
246
|
+
const threadId = '019dfdef-0000-7000-8000-000000000002';
|
|
247
|
+
try {
|
|
248
|
+
writeCodexRollout(codexHome, { project, threadId, turnCount: 22 });
|
|
249
|
+
const fake = makeFakeCodexAppServer(project, { turnCounts: [22, 21] });
|
|
250
|
+
const result = runRestoreSmoke(
|
|
251
|
+
home,
|
|
252
|
+
codexHome,
|
|
253
|
+
project,
|
|
254
|
+
['--codex-thread-id', threadId, '--codex-app-server-bin', fake, '--json'],
|
|
255
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
256
|
+
);
|
|
257
|
+
|
|
258
|
+
assert.equal(result.status, 1);
|
|
259
|
+
const payload = JSON.parse(result.stdout);
|
|
260
|
+
assert.equal(payload.status, 'app-server-restart-mismatch');
|
|
261
|
+
assert.equal(payload.reason, 'fresh_app_server_restore_counts_mismatch');
|
|
262
|
+
assert.equal(payload.restartSafe, false);
|
|
263
|
+
assert.equal(payload.observations[1].turnCountCheck.status, 'mismatch');
|
|
264
|
+
} finally {
|
|
265
|
+
rmSync(project, { recursive: true, force: true });
|
|
266
|
+
rmSync(home, { recursive: true, force: true });
|
|
267
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
268
|
+
}
|
|
269
|
+
});
|
|
270
|
+
|
|
271
|
+
test('codex-restore-smoke reports mismatch when thread turns/list differs from read and resume', () => {
|
|
272
|
+
const home = makeTempHome();
|
|
273
|
+
const codexHome = makeTempHome();
|
|
274
|
+
const project = makeTempProject();
|
|
275
|
+
const threadId = '019dfdef-0000-7000-8000-000000000004';
|
|
276
|
+
try {
|
|
277
|
+
writeCodexRollout(codexHome, { project, threadId, turnCount: 22 });
|
|
278
|
+
const fake = makeFakeCodexAppServer(project, { turnCounts: [22, 22], turnsListCounts: [21, 21] });
|
|
279
|
+
const result = runRestoreSmoke(
|
|
280
|
+
home,
|
|
281
|
+
codexHome,
|
|
282
|
+
project,
|
|
283
|
+
['--codex-thread-id', threadId, '--codex-app-server-bin', fake, '--json'],
|
|
284
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
285
|
+
);
|
|
286
|
+
|
|
287
|
+
assert.equal(result.status, 1);
|
|
288
|
+
const payload = JSON.parse(result.stdout);
|
|
289
|
+
assert.equal(payload.status, 'app-server-restart-mismatch');
|
|
290
|
+
assert.equal(payload.observations[0].readTurns, 22);
|
|
291
|
+
assert.equal(payload.observations[0].resumedTurns, 22);
|
|
292
|
+
assert.equal(payload.observations[0].turnsListTurns, 21);
|
|
293
|
+
assert.equal(payload.observations[0].turnCountCheck.status, 'mismatch');
|
|
294
|
+
} finally {
|
|
295
|
+
rmSync(project, { recursive: true, force: true });
|
|
296
|
+
rmSync(home, { recursive: true, force: true });
|
|
297
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
298
|
+
}
|
|
299
|
+
});
|
|
300
|
+
|
|
301
|
+
test('codex-restore-smoke reports structured error when thread turns/list is unavailable', () => {
|
|
302
|
+
const home = makeTempHome();
|
|
303
|
+
const codexHome = makeTempHome();
|
|
304
|
+
const project = makeTempProject();
|
|
305
|
+
const threadId = '019dfdef-0000-7000-8000-000000000005';
|
|
306
|
+
try {
|
|
307
|
+
writeCodexRollout(codexHome, { project, threadId, turnCount: 22 });
|
|
308
|
+
const fake = makeFakeCodexAppServer(project, {
|
|
309
|
+
turnCounts: [22, 22],
|
|
310
|
+
supportsTurnsList: false,
|
|
311
|
+
});
|
|
312
|
+
const result = runRestoreSmoke(
|
|
313
|
+
home,
|
|
314
|
+
codexHome,
|
|
315
|
+
project,
|
|
316
|
+
['--codex-thread-id', threadId, '--codex-app-server-bin', fake, '--json'],
|
|
317
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
318
|
+
);
|
|
319
|
+
|
|
320
|
+
assert.equal(result.status, 1);
|
|
321
|
+
const payload = JSON.parse(result.stdout);
|
|
322
|
+
assert.equal(payload.status, 'app-server-restore-smoke-error');
|
|
323
|
+
assert.equal(payload.reason, 'app_server_restore_request_failed');
|
|
324
|
+
assert.match(payload.error, /thread\/turns\/list/);
|
|
325
|
+
assert.equal(payload.restartSafe, false);
|
|
326
|
+
} finally {
|
|
327
|
+
rmSync(project, { recursive: true, force: true });
|
|
328
|
+
rmSync(home, { recursive: true, force: true });
|
|
329
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
330
|
+
}
|
|
331
|
+
});
|
|
332
|
+
|
|
333
|
+
test('codex-restore-smoke refuses before app-server when rollout restore safety is risky', () => {
|
|
334
|
+
const home = makeTempHome();
|
|
335
|
+
const codexHome = makeTempHome();
|
|
336
|
+
const project = makeTempProject();
|
|
337
|
+
const threadId = '019dfdef-0000-7000-8000-000000000003';
|
|
338
|
+
try {
|
|
339
|
+
writeCodexRollout(codexHome, {
|
|
340
|
+
project,
|
|
341
|
+
threadId,
|
|
342
|
+
turnCount: 22,
|
|
343
|
+
restoreRisk: true,
|
|
344
|
+
});
|
|
345
|
+
const fake = makeFakeCodexAppServer(project, { turnCounts: [22, 22] });
|
|
346
|
+
const result = runRestoreSmoke(
|
|
347
|
+
home,
|
|
348
|
+
codexHome,
|
|
349
|
+
project,
|
|
350
|
+
['--codex-thread-id', threadId, '--codex-app-server-bin', fake, '--json'],
|
|
351
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
352
|
+
);
|
|
353
|
+
|
|
354
|
+
assert.equal(result.status, 1);
|
|
355
|
+
const payload = JSON.parse(result.stdout);
|
|
356
|
+
assert.equal(payload.status, 'refused');
|
|
357
|
+
assert.equal(payload.reason, 'restore_safety_risk');
|
|
358
|
+
assert.equal(payload.restoreSafety.status, 'risk');
|
|
359
|
+
assert.equal(payload.restartSafe, false);
|
|
360
|
+
} finally {
|
|
361
|
+
rmSync(project, { recursive: true, force: true });
|
|
362
|
+
rmSync(home, { recursive: true, force: true });
|
|
363
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
364
|
+
}
|
|
365
|
+
});
|
|
366
|
+
|
|
367
|
+
test('codex-restore-smoke can inspect risky rollout read-only and report app-server response text matches', () => {
|
|
368
|
+
const home = makeTempHome();
|
|
369
|
+
const codexHome = makeTempHome();
|
|
370
|
+
const project = makeTempProject();
|
|
371
|
+
const threadId = '019dfdef-0000-7000-8000-000000000006';
|
|
372
|
+
const retainedText = 'restore smoke user turn 22';
|
|
373
|
+
try {
|
|
374
|
+
writeCodexRollout(codexHome, {
|
|
375
|
+
project,
|
|
376
|
+
threadId,
|
|
377
|
+
turnCount: 22,
|
|
378
|
+
restoreRisk: true,
|
|
379
|
+
});
|
|
380
|
+
const fake = makeFakeCodexAppServer(project, {
|
|
381
|
+
turnCounts: [22, 22],
|
|
382
|
+
turnsListCounts: [22, 22],
|
|
383
|
+
retainedText,
|
|
384
|
+
});
|
|
385
|
+
const result = runRestoreSmoke(
|
|
386
|
+
home,
|
|
387
|
+
codexHome,
|
|
388
|
+
project,
|
|
389
|
+
[
|
|
390
|
+
'--codex-thread-id',
|
|
391
|
+
threadId,
|
|
392
|
+
'--codex-app-server-bin',
|
|
393
|
+
fake,
|
|
394
|
+
'--inspect-risky-rollout',
|
|
395
|
+
'--json',
|
|
396
|
+
],
|
|
397
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
398
|
+
);
|
|
399
|
+
|
|
400
|
+
assert.equal(result.status, 1);
|
|
401
|
+
const payload = JSON.parse(result.stdout);
|
|
402
|
+
assert.equal(payload.status, 'app-server-restore-text-retained');
|
|
403
|
+
assert.equal(payload.reason, 'restore_text_seen_in_app_server_response');
|
|
404
|
+
assert.equal(payload.restoreSafety.status, 'risk');
|
|
405
|
+
assert.equal(payload.restoreSafetyRiskInspected, true);
|
|
406
|
+
assert.equal(payload.restartSafe, false);
|
|
407
|
+
assert.equal(payload.restoreTextNeedles.length, 1);
|
|
408
|
+
assert.equal(payload.restoreTextMatchCheck.status, 'matches-found');
|
|
409
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[0], {
|
|
410
|
+
source: 'thread_read',
|
|
411
|
+
cycles: [1, 2],
|
|
412
|
+
matchedNeedleIds: ['retained_rollback_text_1'],
|
|
413
|
+
samplePaths: ['$.thread.turns[0].items[0].text'],
|
|
414
|
+
locationKinds: ['item_text_field'],
|
|
415
|
+
locationRisks: ['direct_turn_text_candidate'],
|
|
416
|
+
blockingKinds: ['item_text_field'],
|
|
417
|
+
nonBlockingKinds: [],
|
|
418
|
+
hasBlockingCandidates: true,
|
|
419
|
+
});
|
|
420
|
+
assert.equal(payload.restoreTextMatchCheck.hasBlockingCandidates, true);
|
|
421
|
+
assert.deepEqual(payload.restoreTextMatchCheck.blockingKinds, ['item_text_field']);
|
|
422
|
+
assert.deepEqual(payload.restoreTextMatchCheck.locationRisks, [
|
|
423
|
+
'direct_turn_text_candidate',
|
|
424
|
+
]);
|
|
425
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[1].samplePaths, [
|
|
426
|
+
'$.thread.turns[0].items[0].text',
|
|
427
|
+
]);
|
|
428
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[1].locationKinds, [
|
|
429
|
+
'item_text_field',
|
|
430
|
+
]);
|
|
431
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[2].samplePaths, [
|
|
432
|
+
'$[0].items[0].text',
|
|
433
|
+
]);
|
|
434
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[2].locationKinds, [
|
|
435
|
+
'item_text_field',
|
|
436
|
+
]);
|
|
437
|
+
assert.equal(payload.observations[0].responseTextMatches.status, 'matches-found');
|
|
438
|
+
assert.deepEqual(
|
|
439
|
+
payload.observations[0].responseTextMatches.sources.map((source) => source.source),
|
|
440
|
+
['thread_read', 'thread_resume', 'thread_turns_list'],
|
|
441
|
+
);
|
|
442
|
+
assert.equal(
|
|
443
|
+
payload.observations[0].responseTextMatches.matchedNeedles[0].id,
|
|
444
|
+
'retained_rollback_text_1',
|
|
445
|
+
);
|
|
446
|
+
assert.deepEqual(
|
|
447
|
+
payload.observations[0].responseTextMatches.sources[0].matches[0].locations.map(
|
|
448
|
+
(location) => location.path,
|
|
449
|
+
),
|
|
450
|
+
['$.thread.turns[0].items[0].text'],
|
|
451
|
+
);
|
|
452
|
+
assert.deepEqual(
|
|
453
|
+
payload.observations[0].responseTextMatches.sources[0].matches[0].locations.map(
|
|
454
|
+
(location) => location.kind,
|
|
455
|
+
),
|
|
456
|
+
['item_text_field'],
|
|
457
|
+
);
|
|
458
|
+
assert.deepEqual(
|
|
459
|
+
payload.observations[0].responseTextMatches.sources[0].matches[0].locations.map(
|
|
460
|
+
(location) => location.risk,
|
|
461
|
+
),
|
|
462
|
+
['direct_turn_text_candidate'],
|
|
463
|
+
);
|
|
464
|
+
assert.deepEqual(
|
|
465
|
+
payload.observations[0].responseTextMatches.sources[0].matches[0].locations.map(
|
|
466
|
+
(location) => location.blockingCandidate,
|
|
467
|
+
),
|
|
468
|
+
[true],
|
|
469
|
+
);
|
|
470
|
+
} finally {
|
|
471
|
+
rmSync(project, { recursive: true, force: true });
|
|
472
|
+
rmSync(home, { recursive: true, force: true });
|
|
473
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
474
|
+
}
|
|
475
|
+
});
|
|
476
|
+
|
|
477
|
+
test('codex-restore-smoke text output summarizes risky response text matches', () => {
|
|
478
|
+
const home = makeTempHome();
|
|
479
|
+
const codexHome = makeTempHome();
|
|
480
|
+
const project = makeTempProject();
|
|
481
|
+
const threadId = '019dfdef-0000-7000-8000-000000000007';
|
|
482
|
+
const retainedText = 'restore smoke user turn 22';
|
|
483
|
+
try {
|
|
484
|
+
writeCodexRollout(codexHome, {
|
|
485
|
+
project,
|
|
486
|
+
threadId,
|
|
487
|
+
turnCount: 22,
|
|
488
|
+
restoreRisk: true,
|
|
489
|
+
});
|
|
490
|
+
const fake = makeFakeCodexAppServer(project, {
|
|
491
|
+
turnCounts: [22, 22],
|
|
492
|
+
turnsListCounts: [22, 22],
|
|
493
|
+
retainedText,
|
|
494
|
+
});
|
|
495
|
+
const result = runRestoreSmoke(
|
|
496
|
+
home,
|
|
497
|
+
codexHome,
|
|
498
|
+
project,
|
|
499
|
+
['--codex-thread-id', threadId, '--codex-app-server-bin', fake, '--inspect-risky-rollout'],
|
|
500
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
501
|
+
);
|
|
502
|
+
|
|
503
|
+
assert.equal(result.status, 1);
|
|
504
|
+
assert.match(result.stdout, /status:\s+app-server-restore-text-retained/);
|
|
505
|
+
assert.match(result.stdout, /restore safety risk inspected read-only: yes/);
|
|
506
|
+
assert.match(
|
|
507
|
+
result.stdout,
|
|
508
|
+
/restore text match check: matches-found 1 needle sources=thread_read,thread_resume,thread_turns_list/,
|
|
509
|
+
);
|
|
510
|
+
assert.match(result.stdout, /paths=thread_read:\$\.thread\.turns\[0\]\.items\[0\]\.text/);
|
|
511
|
+
assert.match(result.stdout, /kinds=item_text_field/);
|
|
512
|
+
assert.match(result.stdout, /risks=direct_turn_text_candidate/);
|
|
513
|
+
assert.match(result.stdout, /blocking-candidates=item_text_field/);
|
|
514
|
+
assert.match(
|
|
515
|
+
result.stdout,
|
|
516
|
+
/response text matches: matches-found 1 needle sources=thread_read,thread_resume,thread_turns_list/,
|
|
517
|
+
);
|
|
518
|
+
} finally {
|
|
519
|
+
rmSync(project, { recursive: true, force: true });
|
|
520
|
+
rmSync(home, { recursive: true, force: true });
|
|
521
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
522
|
+
}
|
|
523
|
+
});
|
|
524
|
+
|
|
525
|
+
test('codex-restore-smoke separates quoted output matches from blocking restore matches', () => {
|
|
526
|
+
const home = makeTempHome();
|
|
527
|
+
const codexHome = makeTempHome();
|
|
528
|
+
const project = makeTempProject();
|
|
529
|
+
const threadId = '019dfdef-0000-7000-8000-000000000008';
|
|
530
|
+
const retainedText = 'restore smoke user turn 22';
|
|
531
|
+
try {
|
|
532
|
+
writeCodexRollout(codexHome, {
|
|
533
|
+
project,
|
|
534
|
+
threadId,
|
|
535
|
+
turnCount: 22,
|
|
536
|
+
restoreRisk: true,
|
|
537
|
+
});
|
|
538
|
+
const fake = makeFakeCodexAppServer(project, {
|
|
539
|
+
turnCounts: [22, 22],
|
|
540
|
+
turnsListCounts: [22, 22],
|
|
541
|
+
retainedText,
|
|
542
|
+
retainedTextLocation: 'aggregatedOutput',
|
|
543
|
+
});
|
|
544
|
+
const result = runRestoreSmoke(
|
|
545
|
+
home,
|
|
546
|
+
codexHome,
|
|
547
|
+
project,
|
|
548
|
+
[
|
|
549
|
+
'--codex-thread-id',
|
|
550
|
+
threadId,
|
|
551
|
+
'--codex-app-server-bin',
|
|
552
|
+
fake,
|
|
553
|
+
'--inspect-risky-rollout',
|
|
554
|
+
'--json',
|
|
555
|
+
],
|
|
556
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
557
|
+
);
|
|
558
|
+
|
|
559
|
+
assert.equal(result.status, 1);
|
|
560
|
+
const payload = JSON.parse(result.stdout);
|
|
561
|
+
assert.equal(payload.status, 'app-server-restore-text-quoted');
|
|
562
|
+
assert.equal(
|
|
563
|
+
payload.reason,
|
|
564
|
+
'restore_text_seen_only_in_quoted_or_output_response_fields',
|
|
565
|
+
);
|
|
566
|
+
assert.equal(payload.restoreTextMatchCheck.status, 'matches-found');
|
|
567
|
+
assert.equal(payload.restoreTextMatchCheck.hasBlockingCandidates, false);
|
|
568
|
+
assert.deepEqual(payload.restoreTextMatchCheck.blockingKinds, []);
|
|
569
|
+
assert.deepEqual(payload.restoreTextMatchCheck.nonBlockingKinds, ['aggregated_output']);
|
|
570
|
+
assert.deepEqual(payload.restoreTextMatchCheck.locationRisks, [
|
|
571
|
+
'quoted_or_tool_output_context',
|
|
572
|
+
]);
|
|
573
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[0].samplePaths, [
|
|
574
|
+
'$.thread.turns[0].items[0].aggregatedOutput',
|
|
575
|
+
]);
|
|
576
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[0].locationKinds, [
|
|
577
|
+
'aggregated_output',
|
|
578
|
+
]);
|
|
579
|
+
} finally {
|
|
580
|
+
rmSync(project, { recursive: true, force: true });
|
|
581
|
+
rmSync(home, { recursive: true, force: true });
|
|
582
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
583
|
+
}
|
|
584
|
+
});
|
|
585
|
+
|
|
586
|
+
test('codex-restore-smoke classifies replacement_history matches as durable restore sources', () => {
|
|
587
|
+
const home = makeTempHome();
|
|
588
|
+
const codexHome = makeTempHome();
|
|
589
|
+
const project = makeTempProject();
|
|
590
|
+
const threadId = '019dfdef-0000-7000-8000-000000000009';
|
|
591
|
+
const retainedText = 'restore smoke user turn 22';
|
|
592
|
+
try {
|
|
593
|
+
writeCodexRollout(codexHome, {
|
|
594
|
+
project,
|
|
595
|
+
threadId,
|
|
596
|
+
turnCount: 22,
|
|
597
|
+
restoreRisk: true,
|
|
598
|
+
});
|
|
599
|
+
const fake = makeFakeCodexAppServer(project, {
|
|
600
|
+
turnCounts: [22, 22],
|
|
601
|
+
turnsListCounts: [22, 22],
|
|
602
|
+
retainedText,
|
|
603
|
+
retainedTextLocation: 'replacementHistory',
|
|
604
|
+
});
|
|
605
|
+
const result = runRestoreSmoke(
|
|
606
|
+
home,
|
|
607
|
+
codexHome,
|
|
608
|
+
project,
|
|
609
|
+
[
|
|
610
|
+
'--codex-thread-id',
|
|
611
|
+
threadId,
|
|
612
|
+
'--codex-app-server-bin',
|
|
613
|
+
fake,
|
|
614
|
+
'--inspect-risky-rollout',
|
|
615
|
+
'--json',
|
|
616
|
+
],
|
|
617
|
+
{ THROUGHLINE_EXPERIMENTAL_CODEX_RESTORE_SMOKE: '1' },
|
|
618
|
+
);
|
|
619
|
+
|
|
620
|
+
assert.equal(result.status, 1);
|
|
621
|
+
const payload = JSON.parse(result.stdout);
|
|
622
|
+
assert.equal(payload.status, 'app-server-restore-text-retained');
|
|
623
|
+
assert.equal(payload.restoreTextMatchCheck.hasBlockingCandidates, true);
|
|
624
|
+
assert.deepEqual(payload.restoreTextMatchCheck.blockingKinds, ['replacement_history']);
|
|
625
|
+
assert.deepEqual(payload.restoreTextMatchCheck.locationRisks, [
|
|
626
|
+
'durable_restore_source',
|
|
627
|
+
]);
|
|
628
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[0].samplePaths, [
|
|
629
|
+
'$.thread.turns[0].replacement_history[0].content[0].text',
|
|
630
|
+
]);
|
|
631
|
+
assert.deepEqual(payload.restoreTextMatchCheck.sources[0].locationKinds, [
|
|
632
|
+
'replacement_history',
|
|
633
|
+
]);
|
|
634
|
+
} finally {
|
|
635
|
+
rmSync(project, { recursive: true, force: true });
|
|
636
|
+
rmSync(home, { recursive: true, force: true });
|
|
637
|
+
rmSync(codexHome, { recursive: true, force: true });
|
|
638
|
+
}
|
|
639
|
+
});
|