@ngockhoale/ukit 3.5.19 → 3.5.20
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/CHANGELOG.md +12 -0
- package/package.json +1 -1
- package/scripts/audit/field-replay.mjs +4 -4
- package/scripts/audit/field-telemetry.mjs +24 -6
- package/template_project/.claude/hooks/block-dangerous.mjs +34 -2
- package/template_project/.claude/ukit/runtime/execution-ledger.mjs +2 -0
- package/template_project/.claude/ukit/runtime/mission/classify.mjs +20 -1
- package/template_project/.claude/ukit/runtime/output-compression.mjs +21 -11
- package/template_project/.claude/ukit/runtime/stop-coordinator.mjs +36 -5
package/CHANGELOG.md
CHANGED
|
@@ -2,6 +2,18 @@
|
|
|
2
2
|
|
|
3
3
|
All notable changes to UKit are documented here.
|
|
4
4
|
|
|
5
|
+
## 3.5.20 - 2026-11-21
|
|
6
|
+
|
|
7
|
+
Field-telemetry fix round 6 (live session + field data since 3.5.19).
|
|
8
|
+
|
|
9
|
+
- **Stop gates honour background waits (R6-1).** When the completion gate releases a Stop because a background job or subagent is still running (WAIT), the mission gate no longer blocks that Stop. The WAIT now travels through the stop coordinator (`outcome: waiting`, `waitingOn`), and the mission gate releases it with a note that it is not completion.
|
|
10
|
+
- **Status questions are not requirements (R6-2).** Progress checks such as "đã xong tới đâu rồi", "vậy là xong chưa", "còn lỗi không", "…rồi sao", "sao vẫn chưa …" and "is it done yet" are classified as questions, so they no longer become mission requirements that need acceptance evidence. Work requests ("sửa lỗi login đi", "thêm nút export CSV") are unchanged, and a status check appended to a work clause ("thêm nút export, chạy test chưa", "sửa login rồi chạy test chưa") keeps the work requirement.
|
|
11
|
+
- **Old pauses no longer release every later stop (R6-7).** The stop coordinator used to scan past newer owner turns that were not control phrases. One "dừng đi" therefore released 83 later stops in one BAGuide session over 17 hours. Now a later owner work request (short or long) supersedes an older pause/cancel. A read-only question asked during a pause ("health check: is everything ok?") keeps the pause, and peer-envelope text is still skipped. Ref: anthropics/claude-code#101528.
|
|
12
|
+
- **Same-command temp cleanup is allowed (R6-4).** A recursive delete of `"$P"` is no longer blocked when `P` was assigned from `mktemp -d` in an earlier segment of the same command, quoted, and never reassigned; the exemption also requires every other segment to start with an allow-listed command that cannot assign shell variables, and no substitution, arithmetic, grouping, `eval` or `printf` anywhere in the command. Every other recursive delete of a variable stays blocked.
|
|
13
|
+
- **Ranked multi-section reports pass through uncompressed (R6-10).** Output compression recognises `N.` rank prefixes and sections whose numbering restarts. The field-telemetry report no longer comes back reordered.
|
|
14
|
+
- **Field audit tools (R6-3, R6-5).** `field-telemetry` normalises job/task/agent ids, so one class no longer splits per id. `--since` accepts `YYYY-MM-DDTHH:MM[:SS][Z]` for post-release windows; `field-replay` uses the same cutoff.
|
|
15
|
+
- R6-6 (batch planner per-file history) was checked against data: after 3.5.19, all 798/798 files have measured durations, so no code change was needed.
|
|
16
|
+
|
|
5
17
|
## 3.5.19 - 2026-11-21
|
|
6
18
|
|
|
7
19
|
Field-telemetry fix round 5 and faster batched test runs.
|
package/package.json
CHANGED
|
@@ -10,7 +10,7 @@
|
|
|
10
10
|
* sample commands per class (redacted + truncated; values never printed).
|
|
11
11
|
*
|
|
12
12
|
* Usage: node scripts/audit/field-replay.mjs [--projects-dir DIR] [--match SUBSTR]...
|
|
13
|
-
* [--since YYYY-MM-DD] [--limit N] [--sample-chars N] [--hook sensitive] [--mode inproc|cli] [--json]
|
|
13
|
+
* [--since YYYY-MM-DD[THH:MM]] [--limit N] [--sample-chars N] [--hook sensitive] [--mode inproc|cli] [--json]
|
|
14
14
|
*
|
|
15
15
|
* Hooks are registered in HOOKS below; add block-dangerous / execution-ledger
|
|
16
16
|
* (PostToolUse) by adding an entry with its own payload builder + verdict parser.
|
|
@@ -22,7 +22,7 @@ import path from 'node:path';
|
|
|
22
22
|
import readline from 'node:readline';
|
|
23
23
|
import { spawn } from 'node:child_process';
|
|
24
24
|
import { fileURLToPath, pathToFileURL } from 'node:url';
|
|
25
|
-
import { parseArgs, listTranscripts, normalizeMessage, redact, truncate } from './field-telemetry.mjs';
|
|
25
|
+
import { parseArgs, listTranscripts, normalizeMessage, redact, truncate, beforeSince } from './field-telemetry.mjs';
|
|
26
26
|
|
|
27
27
|
const REPO = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', '..');
|
|
28
28
|
const HOOK_DIR = path.join(REPO, 'template_project', '.claude', 'hooks');
|
|
@@ -55,7 +55,7 @@ export function extractBashCommands(line, out, { since } = {}) {
|
|
|
55
55
|
return;
|
|
56
56
|
}
|
|
57
57
|
if (o?.type !== 'assistant' || !Array.isArray(o.message?.content)) return;
|
|
58
|
-
if (
|
|
58
|
+
if (beforeSince(o.timestamp, since)) return;
|
|
59
59
|
for (const c of o.message.content) {
|
|
60
60
|
if (c && c.type === 'tool_use' && c.name === 'Bash' && typeof c.input?.command === 'string' && c.input.command.trim()) {
|
|
61
61
|
out.set(c.input.command, (out.get(c.input.command) || 0) + 1);
|
|
@@ -129,7 +129,7 @@ export function classifyResults(results, hook, { samples = 40, sampleChars = 200
|
|
|
129
129
|
export async function main(argv = process.argv.slice(2)) {
|
|
130
130
|
const opts = parseArgs(argv);
|
|
131
131
|
if (opts.help) {
|
|
132
|
-
process.stdout.write('Usage: field-replay.mjs [--projects-dir DIR] [--match SUBSTR]... [--since YYYY-MM-DD] [--limit N] [--sample-chars N] [--hook sensitive] [--mode inproc|cli] [--json]\n');
|
|
132
|
+
process.stdout.write('Usage: field-replay.mjs [--projects-dir DIR] [--match SUBSTR]... [--since YYYY-MM-DD[THH:MM]] [--limit N] [--sample-chars N] [--hook sensitive] [--mode inproc|cli] [--json]\n');
|
|
133
133
|
return 0;
|
|
134
134
|
}
|
|
135
135
|
const hook = HOOKS[opts.hook];
|
|
@@ -17,7 +17,7 @@
|
|
|
17
17
|
* the assistant tool_use block with the same id.
|
|
18
18
|
*
|
|
19
19
|
* Usage: node scripts/audit/field-telemetry.mjs [--projects-dir DIR]
|
|
20
|
-
* [--match SUBSTR]... [--since YYYY-MM-DD] [--json]
|
|
20
|
+
* [--match SUBSTR]... [--since YYYY-MM-DD[THH:MM]] [--json]
|
|
21
21
|
*/
|
|
22
22
|
import fs from 'node:fs';
|
|
23
23
|
import fsp from 'node:fs/promises';
|
|
@@ -33,11 +33,29 @@ const HEX_RE = /\b(?=[0-9a-f]*[0-9])(?=[0-9a-f]*[a-f])[0-9a-f]{7,}\b/gi;
|
|
|
33
33
|
const PATH_RE = /(?<![\w.-])(?:~|\$[A-Z_]+|\.{1,2})?(?:\/[^\s"'`:,;()<>\]\[{}]+)+\/?|\b[\w.-]+(?:\/[\w.-]+){2,}\b|\b[\w.-]+(?:\/[\w.-]+)*\/[\w-]+\.[A-Za-z]\w{0,5}\b/g;
|
|
34
34
|
// ^ relative paths need >=2 slashes or a file extension, so "dotenv/secret" stays readable.
|
|
35
35
|
const NUM_RE = /\b\d+(?:\.\d+)?(?:ms|s|m|h|d|%|x|MB|KB|GB)?\b/g;
|
|
36
|
+
// Host background job / agent ids (`b0efx3x02`, `tasks/b9t9a0mz3.output`) after a job-ish word: one class, not one per id.
|
|
37
|
+
const JOB_ID_RE = /\b(job|verification|agent|task|tasks\/|task-id>?)(\s*)(?=[a-z0-9]*\d)(?=[a-z0-9]*[a-z])[a-z0-9]{6,20}\b/gi;
|
|
38
|
+
|
|
39
|
+
// `--since` accepts YYYY-MM-DD or an ISO datetime (YYYY-MM-DDTHH:MM[:SS][Z]); a zone-less value is UTC.
|
|
40
|
+
export const SINCE_RE = /^\d{4}-\d{2}-\d{2}(?:T\d{2}:\d{2}(?::\d{2}(?:\.\d+)?)?Z?)?$/;
|
|
41
|
+
export function sinceToMs(since) {
|
|
42
|
+
if (!since) return 0;
|
|
43
|
+
const iso = since.length === 10 ? `${since}T00:00:00Z` : (since.endsWith('Z') ? since : `${since}Z`);
|
|
44
|
+
const ms = Date.parse(iso);
|
|
45
|
+
return Number.isFinite(ms) ? ms : 0;
|
|
46
|
+
}
|
|
47
|
+
// True when a transcript timestamp is before the window start (unknown timestamps stay in the window).
|
|
48
|
+
export function beforeSince(ts, since) {
|
|
49
|
+
if (!since || typeof ts !== 'string') return false;
|
|
50
|
+
const ms = Date.parse(ts);
|
|
51
|
+
return Number.isFinite(ms) ? ms < sinceToMs(since) : false;
|
|
52
|
+
}
|
|
36
53
|
|
|
37
54
|
export function normalizeMessage(text, max = 140) {
|
|
38
55
|
let s = String(text ?? '');
|
|
39
56
|
s = s.replace(/<\/?system-reminder>/g, ' ');
|
|
40
57
|
s = s.replace(UUID_RE, '#').replace(TOOLU_RE, '#');
|
|
58
|
+
s = s.replace(JOB_ID_RE, '$1$2#');
|
|
41
59
|
s = s.replace(PATH_RE, '<p>');
|
|
42
60
|
s = s.replace(HEX_RE, '#');
|
|
43
61
|
s = s.replace(NUM_RE, 'N');
|
|
@@ -95,12 +113,12 @@ export function parseTranscriptLine(line, acc, { since } = {}) {
|
|
|
95
113
|
}
|
|
96
114
|
if (!o || typeof o !== 'object') return;
|
|
97
115
|
const ts = o.timestamp;
|
|
98
|
-
if (
|
|
116
|
+
if (beforeSince(ts, since)) {
|
|
99
117
|
// still learn tool_use ids so later blocks can resolve their command
|
|
100
118
|
if (o.type !== 'assistant') return;
|
|
101
119
|
}
|
|
102
120
|
const day = isoDay(ts);
|
|
103
|
-
const inWindow = !since || day === 'unknown' ||
|
|
121
|
+
const inWindow = !since || day === 'unknown' || !beforeSince(ts, since);
|
|
104
122
|
|
|
105
123
|
// assistant tool_use → remember Bash commands by id
|
|
106
124
|
if (o.type === 'assistant' && Array.isArray(o.message?.content)) {
|
|
@@ -251,12 +269,12 @@ export function parseArgs(argv) {
|
|
|
251
269
|
else if (a === '--help' || a === '-h') opts.help = true;
|
|
252
270
|
}
|
|
253
271
|
if (!opts.match.length) opts.match = ['KHOA-EXTENAL'];
|
|
254
|
-
if (opts.since &&
|
|
272
|
+
if (opts.since && (!SINCE_RE.test(opts.since) || !sinceToMs(opts.since))) throw new Error(`--since must be YYYY-MM-DD or YYYY-MM-DDTHH:MM[:SS][Z], got ${opts.since}`);
|
|
255
273
|
return opts;
|
|
256
274
|
}
|
|
257
275
|
|
|
258
276
|
export async function listTranscripts({ projectsDir, match, since }) {
|
|
259
|
-
const sinceMs =
|
|
277
|
+
const sinceMs = sinceToMs(since);
|
|
260
278
|
const out = [];
|
|
261
279
|
const walk = async (dir) => {
|
|
262
280
|
let entries;
|
|
@@ -311,7 +329,7 @@ function renderText(rep, meta) {
|
|
|
311
329
|
export async function main(argv = process.argv.slice(2)) {
|
|
312
330
|
const opts = parseArgs(argv);
|
|
313
331
|
if (opts.help) {
|
|
314
|
-
process.stdout.write('Usage: field-telemetry.mjs [--projects-dir DIR] [--match SUBSTR]... [--since YYYY-MM-DD] [--json]\n');
|
|
332
|
+
process.stdout.write('Usage: field-telemetry.mjs [--projects-dir DIR] [--match SUBSTR]... [--since YYYY-MM-DD[THH:MM]] [--json]\n');
|
|
315
333
|
return 0;
|
|
316
334
|
}
|
|
317
335
|
const t0 = Date.now();
|
|
@@ -245,7 +245,37 @@ async function evaluate({ command, projectRoot, env, deadlineMs, signal }) {
|
|
|
245
245
|
let verdictGeneric = false;
|
|
246
246
|
let directSeen = false;
|
|
247
247
|
|
|
248
|
-
|
|
248
|
+
// R6-4: a quoted `"$P"` target is a directory this same command just created with
|
|
249
|
+
// `P=$(mktemp -d …)` — safe to remove. Only when that assignment is the ONLY place the
|
|
250
|
+
// name appears unexpanded. Builtins can assign through COMPUTED names (read/getopts/
|
|
251
|
+
// printf -v/eval/((…))/wait -p …), so enumerating them is unsound: instead EVERY other
|
|
252
|
+
// segment must start with a literal allow-listed command that cannot assign shell
|
|
253
|
+
// variables, and the command may hold no other substitution/arithmetic/grouping syntax.
|
|
254
|
+
const segments = scan.split(/[;&|\n]/);
|
|
255
|
+
const mktempVars = new Map(); // name -> index of the segment that assigned it
|
|
256
|
+
const MKTEMP_ASSIGN = /^\s*([A-Za-z_]\w*)=\$\(\s*mktemp((?:\s+[-\w.\/]+)*)\s*\)\s*$/;
|
|
257
|
+
const SAFE_HEAD = new Set(['cd', 'echo', 'ls', 'cat', 'cp', 'mv', 'mkdir', 'touch', 'git', 'node', 'npm', 'npx', 'yarn', 'pnpm', 'tar', 'unzip', 'grep', 'find', 'true', 'pwd', 'rm', 'sed', 'head', 'tail', 'wc', 'diff', 'cmp']);
|
|
258
|
+
const otherSegmentsSafe = segments.every((seg) => {
|
|
259
|
+
if (!seg.trim() || MKTEMP_ASSIGN.test(seg)) return true;
|
|
260
|
+
const head = seg.trim().split(/\s+/)[0];
|
|
261
|
+
return SAFE_HEAD.has(head);
|
|
262
|
+
});
|
|
263
|
+
const scanWithoutAssigns = scan.split('\n').join(';').split(/(?=[;&|])/).map((s) => s.replace(/^([;&|]+\s*)?[A-Za-z_]\w*=\$\(\s*mktemp[^)]*\)/, '$1')).join('');
|
|
264
|
+
const plainBraces = scanWithoutAssigns.replace(/\$\{[A-Za-z_]\w*\}/g, '$V');
|
|
265
|
+
const exoticSyntax = /\$\(|`|\$'|\$"|\(\(|[(){}]|\beval\b|\bprintf\b/.test(plainBraces);
|
|
266
|
+
if (otherSegmentsSafe && !exoticSyntax) segments.forEach((seg, idx) => {
|
|
267
|
+
const m = MKTEMP_ASSIGN.exec(seg);
|
|
268
|
+
if (!m || !/(^|\s)-[a-zA-Z]*d/.test(m[2]) || /(^|\s)-[a-zA-Z]*u/.test(m[2])) return;
|
|
269
|
+
const bare = scan.match(new RegExp(`(^|[^$\\w{])${m[1]}(?!\\w)`, 'g')) || [];
|
|
270
|
+
if (bare.length === 1 && !mktempVars.has(m[1])) mktempVars.set(m[1], idx);
|
|
271
|
+
});
|
|
272
|
+
const isFreshTempTarget = (target, idx) => {
|
|
273
|
+
const v = /^"\$(?:\{([A-Za-z_]\w*)\}|([A-Za-z_]\w*))"$/.exec(target);
|
|
274
|
+
const name = v && (v[1] || v[2]);
|
|
275
|
+
return Boolean(name) && mktempVars.has(name) && mktempVars.get(name) < idx;
|
|
276
|
+
};
|
|
277
|
+
|
|
278
|
+
for (const [segIdx, rawSegment] of segments.entries()) {
|
|
249
279
|
const segment = rawSegment;
|
|
250
280
|
if (!segment) continue;
|
|
251
281
|
const tokens = segment.trimStart().split(/\s+/).filter(Boolean);
|
|
@@ -295,7 +325,9 @@ async function evaluate({ command, projectRoot, env, deadlineMs, signal }) {
|
|
|
295
325
|
for (let target of targets) {
|
|
296
326
|
// Strip ONE trailing slash (never shrink a bare `/` or `~` to '').
|
|
297
327
|
if (target.length > 1 && target.endsWith('/')) target = target.slice(0, -1);
|
|
298
|
-
if (
|
|
328
|
+
if (isFreshTempTarget(target, segIdx)) {
|
|
329
|
+
// created by this command's own `mktemp -d`: always inside the OS temp area
|
|
330
|
+
} else if (UNSAFE_ONE_TARGET_RE.test(target)) {
|
|
299
331
|
anyUnsafe = true;
|
|
300
332
|
allSafe = false;
|
|
301
333
|
} else if (SAFE_ONE_TARGET_RE.test(target)) {
|
|
@@ -6078,6 +6078,8 @@ async function runEvaluateStop() {
|
|
|
6078
6078
|
process.stderr.write(`[ukit-completion] ${result.reason}\n`);
|
|
6079
6079
|
emit({ systemMessage: advisory ? `${message} ${advisory}` : message,
|
|
6080
6080
|
...(result.outcome === 'blocked' || result.outcome === 'handback' ? { outcome: 'blocked', complete: false } : {}),
|
|
6081
|
+
// R6-1: a WAIT release names the pending job/agent so later Stop lanes (mission gate) honour it.
|
|
6082
|
+
...(result.wait === true && result.waitingOn ? { outcome: 'waiting', complete: false, waitingOn: result.waitingOn } : {}),
|
|
6081
6083
|
});
|
|
6082
6084
|
return;
|
|
6083
6085
|
}
|
|
@@ -40,6 +40,10 @@ export function overlapScore(a, b) {
|
|
|
40
40
|
|
|
41
41
|
const RE = {
|
|
42
42
|
question: /^(?:what|how|why|where|when|which|who|whom|is|are|was|were|does|do|did|can|could|should|would|will|la gi|tai sao|nhu the nao|o dau|bao nhieu|co phai|lam sao)\b|\?\s*$/,
|
|
43
|
+
// R6-2: Vietnamese status questions rarely end with '?' — a trailing question particle marks them.
|
|
44
|
+
viQuestionTail: /\b(?:chua|khong|ko|nhi|chua nhi|duoc chua)\s*[?!.]*$/,
|
|
45
|
+
// Progress/status checks are questions even when they name a work verb ("đã fix chưa", "xong tới đâu rồi").
|
|
46
|
+
statusQuestion: /\b(?:da|xong|fix|sua|chay|lam|done)\b.*\bchua\s*[?!.]*$|\b(?:toi dau|den dau|the nao|ra sao|sao) roi\s*[?!.]*$|\btien do\b.*\b(?:roi|the nao|ra sao)\b|\broi sao\s*[?!.]*$|^(?:sao|tai sao) (?:van|lai|chua|khong|ko|con)\b|^(?:is it|are we|is this) (?:done|finished|ready)\b/,
|
|
43
47
|
imperative: /\b(?:implement|add|fix|create|make|build|change|update|remove|delete|write|refactor|include|replace|rename|move|thay|doi|them|sua|xoa|tao|lam|cap nhat|viet)\b/,
|
|
44
48
|
supersede: /\b(?:instead of|no longer|rather than|replace|replaced|switch from|change\b.+\bto|thay vi|khong con|doi\b.+\bsang|doi\b.+\bthanh|bo\b.+\bdung)\b/,
|
|
45
49
|
defect: /\b(?:fix|bug|bugs|broken|crash|crashes|regress\w*|fails?|failing|loi|hong|khong chay|sua loi)\b/,
|
|
@@ -55,9 +59,24 @@ const RE = {
|
|
|
55
59
|
const FILE_RE = /(?:[\w@.-]+\/)+[\w@.-]+|\b[\w-]+\.(?:mjs|cjs|js|jsx|ts|tsx|json|md|css|scss|html|py|sh|yml|yaml|go|rs|java|sql)\b/g;
|
|
56
60
|
|
|
57
61
|
const cleanFiles = (a) => [...new Set((a || []).map((x) => x.replace(/[.,;:]+$/, '')))];
|
|
62
|
+
// A status check only owns the segment when no other clause carries a work request:
|
|
63
|
+
// "da fix chua?" is a question, "them nut export, chay test chua" is still work.
|
|
64
|
+
// Clauses split on punctuation and conjunctions; a clause is a work order only when it
|
|
65
|
+
// STARTS with an imperative verb ("sua login", "them nut export") — "da sua" (past marker)
|
|
66
|
+
// reports status, so "da sua va chay test chua" stays a question.
|
|
67
|
+
const CLAUSE_SPLIT = /\s*[,;]\s*|\s+(?:va|roi|and|then|thi|sau do)\s+/;
|
|
68
|
+
const IMPERATIVE_START = new RegExp(`^(?:(?:please|hay|giup|lam on|ok|oke|cu|tiep tuc)\\s+)*${RE.imperative.source}`);
|
|
69
|
+
function isStatusQuestion(f) {
|
|
70
|
+
if (!RE.statusQuestion.test(f)) return false;
|
|
71
|
+
const clauses = f.split(CLAUSE_SPLIT).map((c) => c.trim()).filter(Boolean);
|
|
72
|
+
if (clauses.length < 2) return true;
|
|
73
|
+
return !clauses.some((c) => !RE.statusQuestion.test(c) && IMPERATIVE_START.test(c));
|
|
74
|
+
}
|
|
75
|
+
|
|
58
76
|
function kindOf(segment) {
|
|
59
77
|
const f = foldText(segment);
|
|
60
|
-
if (
|
|
78
|
+
if (isStatusQuestion(f)) return 'question';
|
|
79
|
+
if ((RE.question.test(f) || RE.viQuestionTail.test(f)) && !RE.imperative.test(f)) return 'question';
|
|
61
80
|
if (RE.prohibition.test(f)) return 'prohibition';
|
|
62
81
|
if (RE.defect.test(f)) return 'defect';
|
|
63
82
|
if (RE.test.test(f)) return 'test';
|
|
@@ -1471,22 +1471,32 @@ export function isVerbatimOutputCommand(command) {
|
|
|
1471
1471
|
// monotonically) up to RANKED_TABLE_MAX_TOKENS passes through raw. The configured
|
|
1472
1472
|
// outputMaxTokens budget still governs every other output.
|
|
1473
1473
|
const RANKED_TABLE_MAX_TOKENS = 6000;
|
|
1474
|
-
|
|
1474
|
+
// R6-10: a rank prefix `N.` / `N)` counts too (" 1. 38 class…"), and a multi-section report
|
|
1475
|
+
// (sections split by non-numbered header lines, numbering restarting per section) is ranked when
|
|
1476
|
+
// every contiguous numbered run is itself monotonic.
|
|
1477
|
+
const RANKED_LINE_RE = /^\s*(\d+(?:\.\d+)?)(?:\s|[:|,x×.)]|$)/;
|
|
1475
1478
|
export function isPassThroughOutput(text) {
|
|
1476
1479
|
const lines = String(text ?? '').split('\n').filter((l) => l.trim());
|
|
1477
1480
|
if (estimateTokenCount(String(text ?? '')) > RANKED_TABLE_MAX_TOKENS) return false;
|
|
1478
|
-
const
|
|
1481
|
+
const runs = [];
|
|
1482
|
+
let run = null;
|
|
1483
|
+
let numbered = 0;
|
|
1479
1484
|
for (const line of lines) {
|
|
1480
1485
|
const m = RANKED_LINE_RE.exec(stripAnsi(line));
|
|
1481
|
-
if (m)
|
|
1482
|
-
|
|
1483
|
-
|
|
1484
|
-
|
|
1485
|
-
|
|
1486
|
-
|
|
1487
|
-
|
|
1488
|
-
|
|
1489
|
-
|
|
1486
|
+
if (!m) { run = null; continue; }
|
|
1487
|
+
if (!run) { run = []; runs.push(run); }
|
|
1488
|
+
run.push(Number(m[1]));
|
|
1489
|
+
numbered += 1;
|
|
1490
|
+
}
|
|
1491
|
+
if (numbered < 8 || numbered < lines.length * 0.6) return false;
|
|
1492
|
+
return runs.every((nums) => {
|
|
1493
|
+
let asc = true; let desc = true;
|
|
1494
|
+
for (let i = 1; i < nums.length; i += 1) {
|
|
1495
|
+
if (nums[i] < nums[i - 1]) asc = false;
|
|
1496
|
+
if (nums[i] > nums[i - 1]) desc = false;
|
|
1497
|
+
}
|
|
1498
|
+
return asc || desc;
|
|
1499
|
+
});
|
|
1490
1500
|
}
|
|
1491
1501
|
|
|
1492
1502
|
function mergeDeep(base, override) {
|
|
@@ -247,6 +247,8 @@ export function normalizeEvaluatorStdout(stdout) {
|
|
|
247
247
|
if (parsed && typeof parsed.systemMessage === 'string' && parsed.systemMessage) {
|
|
248
248
|
return { kind: 'advisory', systemMessage: parsed.systemMessage,
|
|
249
249
|
...(parsed.outcome === 'blocked' && parsed.complete === false ? { outcome: 'blocked', complete: false } : {}),
|
|
250
|
+
...(parsed.outcome === 'waiting' && parsed.complete === false && typeof parsed.waitingOn === 'string' && parsed.waitingOn
|
|
251
|
+
? { outcome: 'waiting', complete: false, waitingOn: parsed.waitingOn.slice(0, 80) } : {}),
|
|
250
252
|
};
|
|
251
253
|
}
|
|
252
254
|
// A message-less non-completion release (the finite handback was already delivered) is never a pass.
|
|
@@ -374,6 +376,7 @@ export function coordinateStopDecisions({ completion = null, watchdog = null, ha
|
|
|
374
376
|
policyOwner: 'none',
|
|
375
377
|
policies,
|
|
376
378
|
...(completion?.outcome === 'blocked' && completion?.complete === false ? { blockedOutcome: true } : {}),
|
|
379
|
+
...(completion?.outcome === 'waiting' && completion?.waitingOn ? { waitingOn: completion.waitingOn } : {}),
|
|
377
380
|
};
|
|
378
381
|
}
|
|
379
382
|
|
|
@@ -2631,7 +2634,22 @@ const RESUME_RE = new RegExp(
|
|
|
2631
2634
|
'iu',
|
|
2632
2635
|
);
|
|
2633
2636
|
const SYNTHETIC_USER_PREFIX = /^(?:<system-reminder|<task-notification|<local-command|<user-prompt-submit-hook|Stop hook feedback|\[ukit|Caveat:|\[Request interrupted)/i;
|
|
2634
|
-
|
|
2637
|
+
// Lazy: the mission classifier is optional for this module (partial installs / sandboxes); without it a
|
|
2638
|
+
// trailing '?' is the only question signal.
|
|
2639
|
+
let classifyPromptFn;
|
|
2640
|
+
async function loadClassifyPrompt() {
|
|
2641
|
+
if (classifyPromptFn === undefined) {
|
|
2642
|
+
try { classifyPromptFn = (await import('./mission/classify.mjs')).classifyPrompt || null; } catch { classifyPromptFn = null; }
|
|
2643
|
+
}
|
|
2644
|
+
return classifyPromptFn;
|
|
2645
|
+
}
|
|
2646
|
+
async function isReadOnlyOwnerQuestion(text) {
|
|
2647
|
+
const classify = await loadClassifyPrompt();
|
|
2648
|
+
if (typeof classify !== 'function') return /\?\s*$/.test(text);
|
|
2649
|
+
try { return classify({ text }).relation?.type === 'question'; } catch { return false; }
|
|
2650
|
+
}
|
|
2651
|
+
const PEER_ENVELOPE_START =/^(?:Another Claude session sent a message|<(?:agent|teammate)-message[\s>])/;
|
|
2652
|
+
const RESUME_COMMAND_TAG =/^<command-name>\/?(?:ukit:)?handoff-(?:fullstack|implement)(?![\w-])/;
|
|
2635
2653
|
|
|
2636
2654
|
function genuineUserText(entry) {
|
|
2637
2655
|
if (!entry || entry.type !== 'user' || entry.isSidechain === true || entry.isMeta === true) return null;
|
|
@@ -2693,10 +2711,18 @@ export async function detectUserInterruption(transcriptPath, {
|
|
|
2693
2711
|
const text = genuineUserText(entry);
|
|
2694
2712
|
if (text === null) continue;
|
|
2695
2713
|
if (RESUME_COMMAND_TAG.test(text)) return { kind: null };
|
|
2696
|
-
|
|
2697
|
-
if (
|
|
2698
|
-
if (
|
|
2699
|
-
|
|
2714
|
+
// Peer-envelope-shaped text (merged peer block) is not a decisive owner turn: keep scanning.
|
|
2715
|
+
if (PEER_ENVELOPE_START.test(text)) continue;
|
|
2716
|
+
if (text.length <= INTERRUPT_MAX_TEXT) {
|
|
2717
|
+
if (CANCEL_RE.test(text)) return { kind: 'cancel' };
|
|
2718
|
+
if (PAUSE_RE.test(text)) return { kind: 'pause' };
|
|
2719
|
+
if (RESUME_RE.test(text)) return { kind: null };
|
|
2720
|
+
}
|
|
2721
|
+
// R6-7: a read-only question during a pause ("health check: is everything ok?") keeps the pause;
|
|
2722
|
+
// any later owner WORK request (long or short) supersedes an older pause/cancel — one "dừng đi"
|
|
2723
|
+
// must not release every later stop of the session.
|
|
2724
|
+
if (await isReadOnlyOwnerQuestion(text)) continue;
|
|
2725
|
+
return { kind: null };
|
|
2700
2726
|
}
|
|
2701
2727
|
return { kind: null };
|
|
2702
2728
|
} catch {
|
|
@@ -3178,6 +3204,11 @@ export async function evaluateMissionGate({ merged, projectRoot, payload = {}, o
|
|
|
3178
3204
|
const g = await gm.loadGraph({ projectRoot, missionId: mid });
|
|
3179
3205
|
const live = (r) => (r.status === 'active' || r.status === 'conflict') && r.kind !== 'question';
|
|
3180
3206
|
if (!Object.values(g.state.reqs).some(live)) return merged;
|
|
3207
|
+
// R6-1: the completion gate released this Stop as a WAIT on a pending host background job/agent.
|
|
3208
|
+
// Blocking cannot produce that result, so the mission gate releases too (advisory, no block budget).
|
|
3209
|
+
if (typeof merged.waitingOn === 'string' && merged.waitingOn) {
|
|
3210
|
+
return { ...merged, systemMessage: [merged.systemMessage, `[ukit-mission] waiting on ${merged.waitingOn}; the mission continues when its task-notification arrives (not completion).`].filter(Boolean).join('\n') };
|
|
3211
|
+
}
|
|
3181
3212
|
|
|
3182
3213
|
// Ledger evidence: a passing verification receipt newer than the task's done claim verifies its ACs.
|
|
3183
3214
|
const receipt = await newestPassingVerification(projectRoot);
|