mlai-code 0.3.3 → 0.3.5
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/README.md +1 -1
- package/package.json +1 -1
- package/src/checkpoints.mjs +217 -0
- package/src/cli.mjs +153 -26
- package/src/cloud.mjs +2 -0
- package/src/engine.mjs +1 -1
- package/src/headless.mjs +39 -11
- package/src/home.mjs +3 -2
- package/src/host.mjs +141 -13
- package/src/ide-quick.mjs +147 -0
- package/src/ide.mjs +182 -0
- package/src/netbox.mjs +1 -1
- package/src/parity.mjs +17 -0
- package/src/relay.mjs +1 -1
- package/src/render.mjs +16 -5
- package/src/runtime.mjs +2 -0
- package/src/server.mjs +46 -7
- package/src/session.mjs +31 -8
- package/src/tui/acp.mjs +121 -19
- package/src/tui/app.mjs +2 -1
- package/src/tui/chat.mjs +35 -10
- package/src/tui/cmd-config.mjs +3 -1
- package/src/tui/cmd-core.mjs +9 -5
- package/src/tui/cmd-extensibility.mjs +5 -0
- package/src/tui/extensions-local.mjs +2 -0
- package/src/tui/mcp-config.mjs +22 -4
- package/src/tui/review-cli.mjs +15 -4
- package/src/tui/snapshots.mjs +57 -16
- package/src/tui/transcript.mjs +13 -2
- package/src/two-words.mjs +14 -8
- package/src/view-sync.mjs +48 -0
- package/src/winenv.mjs +1 -1
- package/src/workspace-api.mjs +21 -4
- package/vendor/client-sdk/index.cjs +2 -1
- package/vendor/cloud-cli/mlai.mjs +23 -11
- package/worker/architecture_worker.py +33 -4
- package/worker/browser_tool.py +19 -1
- package/worker/chat_fork.py +4 -1
- package/worker/checkpoints.py +71 -8
- package/worker/command_safety.py +56 -3
- package/worker/connected_turn.py +30 -0
- package/worker/conversation.py +125 -13
- package/worker/git_sync.py +4 -4
- package/worker/image_generation.py +2 -2
- package/worker/kimi_bridge.py +251 -31
- package/worker/kimi_gates.py +25 -6
- package/worker/kimi_hook.py +154 -13
- package/worker/media_tools.py +4 -1
- package/worker/mlai_hooks.py +222 -3
- package/worker/mlai_mcp.py +26 -6
- package/worker/mlai_review/README.md +35 -0
- package/worker/mlai_review/__init__.py +1 -0
- package/worker/mlai_review/__main__.py +261 -0
- package/worker/mlai_review/_compat.py +29 -0
- package/worker/mlai_review/analysis/__init__.py +3 -0
- package/worker/mlai_review/analysis/detect.py +50 -0
- package/worker/mlai_review/analysis/main.py +128 -0
- package/worker/mlai_review/analysis/messages.py +8 -0
- package/worker/mlai_review/analysis/tools.py +194 -0
- package/worker/mlai_review/application/__init__.py +1 -0
- package/worker/mlai_review/application/asking.py +53 -0
- package/worker/mlai_review/application/challengers.py +142 -0
- package/worker/mlai_review/application/context_builder.py +194 -0
- package/worker/mlai_review/application/errors.py +17 -0
- package/worker/mlai_review/application/feedback.py +52 -0
- package/worker/mlai_review/application/matching.py +46 -0
- package/worker/mlai_review/application/messages.py +41 -0
- package/worker/mlai_review/application/ports.py +222 -0
- package/worker/mlai_review/application/prompts.py +342 -0
- package/worker/mlai_review/application/review_service.py +200 -0
- package/worker/mlai_review/application/reviewers.py +438 -0
- package/worker/mlai_review/application/streaming.py +91 -0
- package/worker/mlai_review/application/tools.py +138 -0
- package/worker/mlai_review/application/usage.py +145 -0
- package/worker/mlai_review/application/write_check.py +46 -0
- package/worker/mlai_review/contracts/__init__.py +5 -0
- package/worker/mlai_review/contracts/analysis.py +57 -0
- package/worker/mlai_review/contracts/review.py +42 -0
- package/worker/mlai_review/domain/__init__.py +4 -0
- package/worker/mlai_review/domain/changes.py +77 -0
- package/worker/mlai_review/domain/context.py +147 -0
- package/worker/mlai_review/domain/diagnostics.py +239 -0
- package/worker/mlai_review/domain/feedback.py +216 -0
- package/worker/mlai_review/domain/findings.py +505 -0
- package/worker/mlai_review/domain/lifecycle.py +85 -0
- package/worker/mlai_review/domain/partition.py +42 -0
- package/worker/mlai_review/domain/relations.py +166 -0
- package/worker/mlai_review/domain/relevance.py +202 -0
- package/worker/mlai_review/domain/report.py +99 -0
- package/worker/mlai_review/domain/structure.py +56 -0
- package/worker/mlai_review/domain/writes.py +86 -0
- package/worker/mlai_review/eval.py +221 -0
- package/worker/mlai_review/eval_cases/clean-rename.json +11 -0
- package/worker/mlai_review/eval_cases/missing-ownership-check.json +22 -0
- package/worker/mlai_review/eval_cases/off-by-one.json +22 -0
- package/worker/mlai_review/eval_cases/sql-injection.json +20 -0
- package/worker/mlai_review/feedback_report.py +60 -0
- package/worker/mlai_review/infrastructure/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/analysis/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/analysis/none.py +12 -0
- package/worker/mlai_review/infrastructure/analysis/translate.py +17 -0
- package/worker/mlai_review/infrastructure/github/__init__.py +0 -0
- package/worker/mlai_review/infrastructure/github/client.py +445 -0
- package/worker/mlai_review/infrastructure/github/feedback.py +212 -0
- package/worker/mlai_review/infrastructure/github/http.py +31 -0
- package/worker/mlai_review/infrastructure/github/rendering.py +398 -0
- package/worker/mlai_review/infrastructure/github/token.py +20 -0
- package/worker/mlai_review/infrastructure/local/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/local/analyzer.py +48 -0
- package/worker/mlai_review/infrastructure/local/events.py +100 -0
- package/worker/mlai_review/infrastructure/local/git_reader.py +233 -0
- package/worker/mlai_review/infrastructure/local/publisher.py +138 -0
- package/worker/mlai_review/infrastructure/messages.py +54 -0
- package/worker/mlai_review/infrastructure/models/__init__.py +0 -0
- package/worker/mlai_review/infrastructure/models/openai_compatible.py +269 -0
- package/worker/mlai_review/infrastructure/structure/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/structure/heuristic.py +186 -0
- package/worker/mlai_review/infrastructure/structure/tree_sitter.py +186 -0
- package/worker/mlai_review/infrastructure/workspace/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/workspace/local.py +105 -0
- package/worker/mlai_review/infrastructure/workspace/memory.py +77 -0
- package/worker/mlai_review/infrastructure/workspace/paths.py +17 -0
- package/worker/mlai_review/pipeline.py +89 -0
- package/worker/mlai_review/runtime.py +194 -0
- package/worker/mlai_review/settings.py +139 -0
- package/worker/preview.py +15 -0
- package/worker/product_facts.py +3 -1
- package/worker/repo_policy.py +5 -0
- package/worker/roles.py +29 -2
- package/worker/runner.py +164 -23
- package/worker/runtime_control.py +95 -9
- package/worker/secret_scan.py +3 -2
- package/worker/server_memory.py +122 -0
- package/worker/subagents.py +6 -2
- package/worker/tool_catalog.py +5 -2
- package/worker/verification.py +113 -8
package/worker/mlai_hooks.py
CHANGED
|
@@ -192,6 +192,127 @@ def outside_repo(p, repo=None):
|
|
|
192
192
|
return not (full == base or full.startswith(base.rstrip(os.sep) + os.sep))
|
|
193
193
|
|
|
194
194
|
|
|
195
|
+
# BUG-0499: on This Mac/CLI, reading outside the project asks first unless the path is under a folder the person added
|
|
196
|
+
# with --add-dir (request.json `readDirs`). A yes covers that folder for the rest of the run.
|
|
197
|
+
OUTSIDE_READ_PROGRAMS = frozenset(('cat', 'head', 'tail', 'less', 'more', 'nl', 'tac', 'grep', 'rg', 'egrep', 'fgrep', 'ag', 'ack', 'ls', 'tree', 'wc',
|
|
198
|
+
'stat', 'file', 'xxd', 'od', 'strings', 'diff', 'cmp', 'awk', 'sed', 'sort', 'uniq', 'cut', 'jq', 'yq', 'find', 'fd',
|
|
199
|
+
'md5sum', 'sha256sum', 'sha1sum', 'shasum', 'base64', 'zcat', 'du', 'cp', 'tar', 'zip', 'rsync'))
|
|
200
|
+
_OUTSIDE_OK = set()
|
|
201
|
+
# Installed software and system settings a shell command may inspect (which, ls /usr/local/bin, cat /etc/os-release).
|
|
202
|
+
SYSTEM_READ = ('/usr', '/bin', '/sbin', '/lib', '/opt', '/etc', '/nix', '/snap', '/System', '/Library', '/Applications', '/private/etc')
|
|
203
|
+
|
|
204
|
+
|
|
205
|
+
def _read_roots(cfg, root=None):
|
|
206
|
+
import tempfile
|
|
207
|
+
out = [os.path.realpath(str(root or ROOT)), os.path.realpath(os.path.dirname(os.path.abspath(__file__))), os.path.realpath(tempfile.gettempdir()), '/tmp', '/private/tmp', '/dev']
|
|
208
|
+
out += [os.path.realpath(os.path.expanduser(d)) for d in (cfg.get('readDirs') or []) if isinstance(d, str) and d]
|
|
209
|
+
return out + sorted(_OUTSIDE_OK)
|
|
210
|
+
|
|
211
|
+
|
|
212
|
+
def _under(full, roots):
|
|
213
|
+
return any(full == r or full.startswith(r.rstrip(os.sep) + os.sep) for r in roots)
|
|
214
|
+
|
|
215
|
+
|
|
216
|
+
def _command_read_paths(command):
|
|
217
|
+
"""Path-looking arguments of read programs in a shell command (absolute, ~, $HOME or ../ paths only)."""
|
|
218
|
+
out = []
|
|
219
|
+
try:
|
|
220
|
+
import command_safety as cs
|
|
221
|
+
for st in cs.split_statements(str(command or ''))[:40]:
|
|
222
|
+
try: words, redirects = cs._words(st); words, _f = cs._strip_wrappers(words)
|
|
223
|
+
except Exception: continue
|
|
224
|
+
targets = [t for kind, t in redirects if kind == 'read']
|
|
225
|
+
if words and words[0].rsplit('/', 1)[-1] in OUTSIDE_READ_PROGRAMS: targets += [w for w in words[1:] if not w.startswith('-')]
|
|
226
|
+
for w in targets:
|
|
227
|
+
w = w.replace('$HOME', '~').replace('${HOME}', '~')
|
|
228
|
+
if w.startswith(('/', '~', '../')) or w == '..': out.append(w)
|
|
229
|
+
except Exception: return []
|
|
230
|
+
return out[:20]
|
|
231
|
+
|
|
232
|
+
|
|
233
|
+
def outside_read_paths(tool, command, path, cfg, root=None, repo=None):
|
|
234
|
+
"""Paths this call would read outside the project and outside every allowed folder."""
|
|
235
|
+
if command: paths = _command_read_paths(command)
|
|
236
|
+
elif canonical(tool) == 'read': paths = [path] if isinstance(path, str) else list(path or [])
|
|
237
|
+
else: return []
|
|
238
|
+
base = os.path.realpath(str(repo or REPO)); roots = [base] + _read_roots(cfg, root) + (list(SYSTEM_READ) if command else [])
|
|
239
|
+
out = []
|
|
240
|
+
for p in paths:
|
|
241
|
+
if not p: continue
|
|
242
|
+
s = os.path.expanduser(str(p))
|
|
243
|
+
full = os.path.realpath(s if os.path.isabs(s) else os.path.join(base, s))
|
|
244
|
+
if not _under(full, roots): out.append(full)
|
|
245
|
+
return out
|
|
246
|
+
|
|
247
|
+
|
|
248
|
+
def read_allowed(full, cfg=None, root=None):
|
|
249
|
+
"""True when an absolute path outside the project may be read (an --add-dir folder, or approved this run)."""
|
|
250
|
+
cfg = cfg if cfg is not None else settings(root)
|
|
251
|
+
dirs = [os.path.realpath(os.path.expanduser(d)) for d in (cfg.get('readDirs') or []) if isinstance(d, str) and d]
|
|
252
|
+
return _under(os.path.realpath(str(full)), dirs + sorted(_OUTSIDE_OK))
|
|
253
|
+
|
|
254
|
+
|
|
255
|
+
def outside_read_gate(tool, args, command, path, cfg, root=None, mode=None):
|
|
256
|
+
if not cfg.get('laptop') or cfg.get('restricted'): return None
|
|
257
|
+
if (cfg.get('perms') or {}).get('bypass') or (Path(root or ROOT) / 'kimi-approve-all').exists(): return None
|
|
258
|
+
outside = outside_read_paths(tool, command, path, cfg, root)
|
|
259
|
+
if not outside: return None
|
|
260
|
+
import mlai_permissions as mp
|
|
261
|
+
shown = ', '.join(clip(p, 200) for p in outside[:3])
|
|
262
|
+
if not mp.ask_user('Read outside the project: ' + shown, 'outside the project', root or ROOT,
|
|
263
|
+
reason='This reads a file outside the project folder. Start mlai with --add-dir <folder> to allow a folder without asking.'):
|
|
264
|
+
emit('policy', root, rule='outside the project', decision='deny', content='Not read: ' + shown, permission=True)
|
|
265
|
+
return ('The user did not approve reading outside the project (' + shown + '). Do not try another way to read it; '
|
|
266
|
+
'tell the user they can allow the folder with --add-dir <folder>.')
|
|
267
|
+
for p in outside: _OUTSIDE_OK.add(p)
|
|
268
|
+
return None
|
|
269
|
+
|
|
270
|
+
|
|
271
|
+
# BUG-0531: asked to fix one function, the agent created tests/test_x.py and pip-installed pytest into a new .venv.
|
|
272
|
+
# Running the project's existing tests is fine; adding test files or installing tools needs the user's say-so.
|
|
273
|
+
_INSTALL = re.compile(r'(?:^|[;&|(]\s*|\s)(?:sudo\s+)?(?:(?:python3?|py)\s+-m\s+pip|pip3?|uv\s+pip|pipx|conda|mamba|poetry|brew|apt(?:-get)?|gem|cargo|go)\s+(?:install|add)\b'
|
|
274
|
+
r'|(?:^|[;&|(]\s*|\s)(?:npm|pnpm|yarn|bun)\s+(?:install|i|add)\s+(?:-\S+\s+)*[@\w]', re.I)
|
|
275
|
+
_TEST_FILE = re.compile(r'(?:^|/)(?:tests?|__tests__|spec|e2e)/|(?:^|/)test_[^/]+\.py$|_test\.(?:py|go)$|\.(?:test|spec)\.[cm]?[jt]sx?$', re.I)
|
|
276
|
+
_ASKED_INSTALL = re.compile(r'\b(?:install|installing|dependenc(?:y|ies)|packages?|set ?up|setup|venv|virtual ?env|requirements|pip|npm i|add (?:a |the )?(?:library|lib|package|dependency))\b', re.I)
|
|
277
|
+
_ASKED_TESTS = re.compile(r'\b(?:tests?|testing|unit ?tests?|spec|specs|coverage|pytest|jest|vitest|tdd)\b', re.I)
|
|
278
|
+
_SCOPE_OK = set()
|
|
279
|
+
|
|
280
|
+
|
|
281
|
+
def install_command(command):
|
|
282
|
+
return bool(command) and bool(_INSTALL.search(str(command)))
|
|
283
|
+
|
|
284
|
+
|
|
285
|
+
def new_test_file(tool, path, root=None):
|
|
286
|
+
if canonical(tool) != 'edit': return None
|
|
287
|
+
for p in ([path] if isinstance(path, str) else list(path or [])):
|
|
288
|
+
if not p or not _TEST_FILE.search(str(p).replace('\\', '/')): continue
|
|
289
|
+
full = Path(p) if os.path.isabs(str(p)) else Path(REPO) / str(p)
|
|
290
|
+
if not full.exists(): return str(p)
|
|
291
|
+
return None
|
|
292
|
+
|
|
293
|
+
|
|
294
|
+
def scope_gate(tool, command, path, cfg, run=None, root=None):
|
|
295
|
+
"""Laptop runs: adding a new test file or installing packages the request did not ask for needs the user first.
|
|
296
|
+
Unattended runs (auto, approve all, bypass) cannot ask, so the agent is told to ask in its answer instead."""
|
|
297
|
+
if not cfg.get('laptop') or cfg.get('restricted'): return None
|
|
298
|
+
task = str((run or {}).get('task') or _request(root or ROOT).get('task') or '')
|
|
299
|
+
what = None
|
|
300
|
+
if command and install_command(command) and not _ASKED_INSTALL.search(task): what = ('install', 'Install packages: ' + clip(command, 300))
|
|
301
|
+
elif not command:
|
|
302
|
+
f = new_test_file(tool, path, root)
|
|
303
|
+
if f and not _ASKED_TESTS.search(task): what = ('test file', 'Add a new test file: ' + clip(f, 200))
|
|
304
|
+
if not what or what[1] in _SCOPE_OK: return None
|
|
305
|
+
unattended = (run or {}).get('mode') in ('auto', 'plan') or (cfg.get('perms') or {}).get('bypass') or (Path(root or ROOT) / 'kimi-approve-all').exists()
|
|
306
|
+
stop = ('Not done: ' + what[1] + '. The user did not ask for this. You may run the tests the project already has, but do not add new '
|
|
307
|
+
'test files or install tools on your own. Finish what was asked and, in your answer, ask whether they want this.')
|
|
308
|
+
if unattended:
|
|
309
|
+
emit('policy', root, rule='not requested', decision='deny', content=what[1]); return stop
|
|
310
|
+
import mlai_permissions as mp
|
|
311
|
+
if not mp.ask_user(what[1], 'not requested', root or ROOT, reason='Not part of what you asked for. Allow the agent to do this?'):
|
|
312
|
+
emit('policy', root, rule='not requested', decision='deny', content=what[1], permission=True); return stop
|
|
313
|
+
_SCOPE_OK.add(what[1]); return None
|
|
314
|
+
|
|
315
|
+
|
|
195
316
|
def restricted_block(tool, command, path, cfg, root=None):
|
|
196
317
|
if not cfg.get('restricted'): return None
|
|
197
318
|
if command or canonical(tool) in ('shell', 'agent'):
|
|
@@ -245,7 +366,8 @@ def settings(root=None, repo=None, config=None):
|
|
|
245
366
|
'vault': {}, 'plugins': {p['id']: p for p in (ext.get('plugins') or []) if isinstance(p, dict) and isinstance(p.get('id'), str)},
|
|
246
367
|
'files': ext.get('files') if isinstance(ext.get('files'), dict) else {}, 'perms': _perms_config(ext.get('permissions')),
|
|
247
368
|
'offline': isinstance(ext.get('network'), dict) and ext['network'].get('enabled') is False,
|
|
248
|
-
'restricted': isinstance(ext.get('restricted'), dict) and ext['restricted'].get('enabled') is True
|
|
369
|
+
'restricted': isinstance(ext.get('restricted'), dict) and ext['restricted'].get('enabled') is True,
|
|
370
|
+
'readDirs': [d for d in (config.get('readDirs') or []) if isinstance(d, str)][:20] if isinstance(config.get('readDirs'), list) else []}
|
|
249
371
|
_CACHE.clear(); _CACHE[key] = out
|
|
250
372
|
return out
|
|
251
373
|
|
|
@@ -341,11 +463,33 @@ def _glob_hit(c, q):
|
|
|
341
463
|
|
|
342
464
|
|
|
343
465
|
def path_match(path, patterns, repo=None):
|
|
344
|
-
"""Glob match for repo-relative patterns (src/**, **/*.pem) and absolute or ~ patterns (~/.aws/**).
|
|
466
|
+
"""Glob match for repo-relative patterns (src/**, **/*.pem) and absolute or ~ patterns (~/.aws/**).
|
|
467
|
+
BUG-0526: a symlink alias (docs -> secret, cfg -> ~/.aws) is matched by the real path it points to as well."""
|
|
345
468
|
repo = os.path.normpath(str(repo or REPO))
|
|
346
469
|
raw = _expand(str(path or '').strip().strip('"\''))
|
|
347
470
|
if not raw: return None
|
|
348
471
|
absolute = os.path.normpath(raw if os.path.isabs(raw) else os.path.join(repo, raw))
|
|
472
|
+
hit = _path_match(absolute, patterns, repo)
|
|
473
|
+
if hit or not patterns: return hit
|
|
474
|
+
try:
|
|
475
|
+
real, real_repo = os.path.realpath(absolute), os.path.realpath(repo)
|
|
476
|
+
except (OSError, ValueError):
|
|
477
|
+
return None
|
|
478
|
+
if real == absolute and real_repo == repo: return None
|
|
479
|
+
hit = _path_match(real, patterns, real_repo)
|
|
480
|
+
if hit: return hit
|
|
481
|
+
# Expanded pattern roots can themselves be symlinks (macOS /tmp -> /private/tmp): compare against the real ones too.
|
|
482
|
+
for pat in patterns:
|
|
483
|
+
e = _expand(pat)
|
|
484
|
+
if not os.path.isabs(e): continue
|
|
485
|
+
parts = e.split('/'); i = next((k for k, x in enumerate(parts) if re.search(r'[*?\[]', x)), len(parts))
|
|
486
|
+
fixed = '/'.join(parts[:i]) or '/'
|
|
487
|
+
moved = os.path.realpath(fixed)
|
|
488
|
+
if moved != os.path.normpath(fixed) and _path_match(real, [moved + ('/' + '/'.join(parts[i:]) if i < len(parts) else '')], real_repo): return pat
|
|
489
|
+
return None
|
|
490
|
+
|
|
491
|
+
|
|
492
|
+
def _path_match(absolute, patterns, repo):
|
|
349
493
|
rel = os.path.relpath(absolute, repo) if absolute == repo or absolute.startswith(repo + os.sep) else None
|
|
350
494
|
for pat in patterns:
|
|
351
495
|
q = os.path.normpath(_expand(pat)) + ('/**' if str(pat).endswith('/**') and not os.path.normpath(_expand(pat)).endswith('**') else '')
|
|
@@ -448,7 +592,7 @@ def _run_stamp(root=None):
|
|
|
448
592
|
|
|
449
593
|
|
|
450
594
|
def remember_denied(command, pat, root=None, repo=None):
|
|
451
|
-
if pat in ('credentials', 'network'): return
|
|
595
|
+
if pat in ('credentials', 'network', 'git-hooks'): return
|
|
452
596
|
targets = _deny_targets(command, pat, repo)
|
|
453
597
|
if not targets: return
|
|
454
598
|
f = Path(root or ROOT) / DENIED_FILE
|
|
@@ -496,11 +640,36 @@ def denied_again(tool, command=None, paths=(), root=None, repo=None):
|
|
|
496
640
|
return None, None
|
|
497
641
|
|
|
498
642
|
|
|
643
|
+
# BUG-0551: the project's git hooks (pre-commit lint, pre-push tests) are the team's rules. The agent never skips them
|
|
644
|
+
# (--no-verify, -n, core.hooksPath) or rewrites them; when a hook blocks, it reports that and asks the user.
|
|
645
|
+
GIT_HOOK_REASON = ('Refused: this would skip or change the project\'s git hooks (for example a pre-commit lint check). Agents never use '
|
|
646
|
+
'--no-verify or edit .git/hooks. Do not get the same result another way. Tell the user what the hook reported and ask how '
|
|
647
|
+
'they want to proceed (fix the problem, or they can commit themselves).')
|
|
648
|
+
|
|
649
|
+
|
|
650
|
+
def git_hook_reason(command):
|
|
651
|
+
for seg in segments(command):
|
|
652
|
+
try: words = shlex.split(seg)
|
|
653
|
+
except ValueError: words = seg.split()
|
|
654
|
+
if not words: continue
|
|
655
|
+
if any(re.search(r'(^|/)\.git/hooks(/|$)', w) for w in words) and (re.search(r'(^|\s)(>|>>|tee|cp|mv|rm|ln|chmod|sed\s+-i|truncate|install)\b', seg) or '>' in seg):
|
|
656
|
+
return GIT_HOOK_REASON
|
|
657
|
+
if os.path.basename(words[0]) != 'git': continue
|
|
658
|
+
if any(re.match(r'core\.hookspath\s*=', w, re.I) for w in words): return GIT_HOOK_REASON
|
|
659
|
+
sub = next((w for w in words[1:] if not w.startswith('-') and '=' not in w), '')
|
|
660
|
+
if sub in ('commit', 'push', 'merge', 'am', 'rebase', 'cherry-pick', 'revert') and '--no-verify' in words: return GIT_HOOK_REASON
|
|
661
|
+
if sub == 'commit' and any(re.fullmatch(r'-[A-Za-z]*n[A-Za-z]*', w) for w in words[words.index('commit') + 1:]): return GIT_HOOK_REASON
|
|
662
|
+
if sub == 'config' and any(w.lower() == 'core.hookspath' for w in words): return GIT_HOOK_REASON
|
|
663
|
+
return None
|
|
664
|
+
|
|
665
|
+
|
|
499
666
|
def deny_command(command, cfg=None, repo=None):
|
|
500
667
|
"""(rule, reason) when the command may never run, else (None, None)."""
|
|
501
668
|
cfg = cfg or settings(repo=repo)
|
|
502
669
|
reason = credential_reason(command, repo)
|
|
503
670
|
if reason: return 'credentials', reason
|
|
671
|
+
reason = git_hook_reason(command)
|
|
672
|
+
if reason: return 'git-hooks', reason
|
|
504
673
|
d = cfg['deny']
|
|
505
674
|
pat = match_command(command, d['commands'])
|
|
506
675
|
if pat: return pat, HARD_DENY_COMMAND.format(pat=pat)
|
|
@@ -525,6 +694,26 @@ def deny_path(path, cfg=None, repo=None):
|
|
|
525
694
|
return None, None
|
|
526
695
|
|
|
527
696
|
|
|
697
|
+
_LISTED_PATH = re.compile(r'^(?:\./)?([^\s:][^:\n]*?)(?::\d+)?(?::|$)')
|
|
698
|
+
|
|
699
|
+
|
|
700
|
+
def filter_denied_lines(text, cfg=None, repo=None):
|
|
701
|
+
"""BUG-0527: drops search/list output lines that name a hard-denied or credential file (path:line: text, or a bare
|
|
702
|
+
path), so protected text never reaches the model. Returns (text, removed_count)."""
|
|
703
|
+
if not isinstance(text, str) or not text: return text, 0
|
|
704
|
+
cfg = cfg or settings(repo=repo)
|
|
705
|
+
pats = list(cfg['deny']['paths'])
|
|
706
|
+
if not pats: return text, 0
|
|
707
|
+
keep, removed = [], 0
|
|
708
|
+
for line in text.split('\n'):
|
|
709
|
+
m = _LISTED_PATH.match(line)
|
|
710
|
+
if m and ('/' in m.group(1) or '.' in m.group(1)) and path_match(m.group(1), pats, repo):
|
|
711
|
+
removed += 1; continue
|
|
712
|
+
keep.append(line)
|
|
713
|
+
if removed: keep.append('(%d result%s from protected paths left out by a hard-deny rule.)' % (removed, '' if removed == 1 else 's'))
|
|
714
|
+
return '\n'.join(keep), removed
|
|
715
|
+
|
|
716
|
+
|
|
528
717
|
def deny_url(url, cfg=None):
|
|
529
718
|
cfg = cfg or settings()
|
|
530
719
|
try: host = urllib.parse.urlparse(str(url)).hostname or ''
|
|
@@ -781,6 +970,8 @@ def before(tool, args, run=None, command=None, path=None, cfg=None, root=None):
|
|
|
781
970
|
r = dispatch('before_command', payload('before_command', run, tool={'name': tool, 'canonical': name}, command=command), command, (), cfg, root)
|
|
782
971
|
if r['decision'] == 'block': return block_text(r, 'before')
|
|
783
972
|
blocked = permission(tool, args, command, path, cfg, root, mode=(run or {}).get('mode'))
|
|
973
|
+
if not blocked and _PERM.get('grant') not in ('allow', 'asked', 'bypass'): blocked = outside_read_gate(tool, args, command, path, cfg, root)
|
|
974
|
+
if not blocked and _PERM.get('grant') not in ('allow', 'asked', 'bypass'): blocked = scope_gate(tool, command, path, cfg, run, root)
|
|
784
975
|
if blocked or (run or {}).get('mode') != 'auto' or _PERM.get('grant') in ('allow', 'bypass'): return blocked
|
|
785
976
|
try: import mlai_auto_review
|
|
786
977
|
except ImportError: return None
|
|
@@ -861,6 +1052,34 @@ def tool_paths(args):
|
|
|
861
1052
|
return out[:20]
|
|
862
1053
|
|
|
863
1054
|
|
|
1055
|
+
# ------------------------------------------------------------------ structured "cannot complete" outcome (BUG-0523)
|
|
1056
|
+
# The agent reports a request it decided cannot be done (no git remote, an unsupported feature, a refused action) with
|
|
1057
|
+
# the cannot_complete tool. The CLI reads cannot-complete.json after the run and exits with its own code (6).
|
|
1058
|
+
CANNOT_COMPLETE_FILE = 'cannot-complete.json'
|
|
1059
|
+
CANNOT_COMPLETE_TOOL = {'name': 'cannot_complete', 'description': (
|
|
1060
|
+
'Report that the request cannot be done as asked (for example: pushing with no git remote configured, a feature this '
|
|
1061
|
+
'product does not have, an action the user denied). Call it once, before your final answer, with a short reason the user '
|
|
1062
|
+
'can act on. Do not call it when the work is done or only partly verified.'),
|
|
1063
|
+
'inputSchema': {'type': 'object', 'properties': {'reason': {'type': 'string', 'minLength': 1, 'maxLength': 600}}, 'required': ['reason']}}
|
|
1064
|
+
|
|
1065
|
+
|
|
1066
|
+
def record_cannot_complete(args, root=None):
|
|
1067
|
+
reason = ' '.join(str((args or {}).get('reason') or '').split())[:600] if isinstance(args, dict) else ''
|
|
1068
|
+
if not reason: raise ValueError('Give a short reason')
|
|
1069
|
+
root = Path(root or ROOT)
|
|
1070
|
+
tmp = root / (CANNOT_COMPLETE_FILE + '.tmp'); tmp.write_text(json.dumps({'reason': reason, 'time': time.time()})); tmp.replace(root / CANNOT_COMPLETE_FILE)
|
|
1071
|
+
emit('cannot_complete', root, content=reason)
|
|
1072
|
+
return {'recorded': True, 'note': 'Recorded. Now finish with a short answer that says what could not be done and why, and what the user can do instead.'}
|
|
1073
|
+
|
|
1074
|
+
|
|
1075
|
+
# BUG-0523/BUG-0529/BUG-0531: rules for runs on the person's own computer (CLI and Mac app), both engines.
|
|
1076
|
+
LAPTOP_RULES = (
|
|
1077
|
+
'Outcome rules: (1) If the request cannot be done as asked, call cannot_complete {reason} once and then answer; never report it as done. '
|
|
1078
|
+
'(2) When the user denies an action, stop that line of work: do not retry it another way and do not start other work they did not ask '
|
|
1079
|
+
'for. Say what was not done and ask what they want next. (3) Do only what was asked. You may run the project\'s existing tests, but '
|
|
1080
|
+
'ask the user before adding new test files or installing packages or test tools (pip install, npm install, a new virtualenv) just to verify a change.')
|
|
1081
|
+
|
|
1082
|
+
|
|
864
1083
|
# ------------------------------------------------------------------ built-in runner (worker/runner.py)
|
|
865
1084
|
class _Globals:
|
|
866
1085
|
"""Attribute access over a module's globals() dict."""
|
package/worker/mlai_mcp.py
CHANGED
|
@@ -34,7 +34,7 @@ from repo_memory import TOOL as MEMORY_PROPOSE_TOOL
|
|
|
34
34
|
import computer_use
|
|
35
35
|
|
|
36
36
|
WRITE_TOOLS = [IMAGE_TOOL, STOCK_TOOL, BRAND_TOOL, *deck_tools.TOOLS, MEMORY_PROPOSE_TOOL,
|
|
37
|
-
{'name': 'browser', 'description': 'Open the local app (http://127.0.0.1:3000 by default, start it with preview first; it is browsed under a preview-style *.vercel.run host, the way the user reaches it, so dev-server host and origin blocks show up here) in headless Chromium at a given viewport, run up to 12 steps and take a screenshot. Reports page errors, console errors, failed requests, horizontal overflow, visible text and whether the page became interactive (hydrated). For UI work check a phone width (<=480) and a desktop width (>=1024) with at least one click/fill/press and an assert_text. Each call starts a fresh browser profile: localStorage, cookies and sessions from an earlier call are gone, so do every step that depends on saved state (sign in, add items, then a goto step to the same URL to reload, then assert_text) inside one call.',
|
|
37
|
+
{'name': 'browser', 'description': 'Built-in Playwright browser checks (BUG-0202: this is the browser tool; no Playwright install is needed). Open the local app (http://127.0.0.1:3000 by default, start it with preview first; it is browsed under a preview-style *.vercel.run host, the way the user reaches it, so dev-server host and origin blocks show up here) in headless Chromium at a given viewport, run up to 12 steps and take a screenshot. Reports page errors, console errors, failed requests, horizontal overflow, visible text and whether the page became interactive (hydrated). For UI work check a phone width (<=480) and a desktop width (>=1024) with at least one click/fill/press and an assert_text. Each call starts a fresh browser profile: localStorage, cookies and sessions from an earlier call are gone, so do every step that depends on saved state (sign in, add items, then a goto step to the same URL to reload, then assert_text) inside one call.',
|
|
38
38
|
'inputSchema': {'type': 'object', 'properties': {'url': {'type': 'string'}, 'width': {'type': 'integer', 'minimum': 320, 'maximum': 1920}, 'height': {'type': 'integer', 'minimum': 400, 'maximum': 1200},
|
|
39
39
|
'steps': {'type': 'array', 'maxItems': 12, 'items': {'type': 'object', 'properties': {'action': {'type': 'string', 'enum': ['goto', 'click', 'fill', 'press', 'assert_text']}, 'role': {'type': 'string'}, 'name': {'type': 'string'}, 'selector': {'type': 'string'}, 'value': {'type': 'string'}, 'key': {'type': 'string'}, 'text': {'type': 'string'}, 'url': {'type': 'string'}}, 'required': ['action']}}}}},
|
|
40
40
|
{'name': 'preview', 'description': 'Start (or check) the managed preview of this project on port 3000: Next.js and Vite dev/start scripts, or a static index.html. Installs missing packages first and runs the dev server (hot reload) during the task; after the task it is swapped for the production build the user opens and shares. Servers started with Bash in the background stop when the task ends.',
|
|
@@ -161,8 +161,13 @@ def ask(request, always=False):
|
|
|
161
161
|
"""Wait for the user's verdict on one request (one at a time, shared with the hook). Returns the verdict."""
|
|
162
162
|
if not always and auto_mode(): return 'approve'
|
|
163
163
|
import fcntl
|
|
164
|
+
queued = time.time()
|
|
164
165
|
with (ROOT / 'approval.lock').open('a') as lock:
|
|
165
166
|
fcntl.flock(lock, fcntl.LOCK_EX)
|
|
167
|
+
# BUG-0251: queued behind an approval the user rejected in the same step: rejected too, never shown later.
|
|
168
|
+
try:
|
|
169
|
+
if float((ROOT / 'last-reject').read_text().strip() or 0) >= queued: return None
|
|
170
|
+
except (OSError, ValueError): pass
|
|
166
171
|
request = {'id': os.urandom(8).hex(), **request}
|
|
167
172
|
tmp = ROOT / 'pending.json.tmp'; tmp.write_text(json.dumps(request)); tmp.replace(ROOT / 'pending.json'); emit('approval', **request)
|
|
168
173
|
decision = ROOT / ('decision-' + request['id']); deadline = time.time() + 540
|
|
@@ -172,6 +177,7 @@ def ask(request, always=False):
|
|
|
172
177
|
raise ValueError('Task stopped' if (ROOT / 'cancel').exists() else 'The user did not approve this call in time. Do not retry it; report what is waiting.')
|
|
173
178
|
time.sleep(0.4)
|
|
174
179
|
verdict = decision.read_text().strip(); decision.unlink(missing_ok=True); (ROOT / 'pending.json').unlink(missing_ok=True)
|
|
180
|
+
if verdict not in ('approve', 'approve_all'): (ROOT / 'last-reject').write_text(str(time.time()))
|
|
175
181
|
if verdict == 'approve_all' and not always:
|
|
176
182
|
(ROOT / 'kimi-approve-all').write_text('1'); emit('mode', mode='auto', content='Approved all remaining steps for this task.')
|
|
177
183
|
return request['id'] if verdict in ('approve', 'approve_all') else None
|
|
@@ -183,9 +189,9 @@ def integration(args):
|
|
|
183
189
|
if not match: raise ValueError('This tool is not enabled for this task. Use integration_search to find enabled tools.')
|
|
184
190
|
if os.getenv('MINI_READONLY') == '1' and match.get('write') is not False:
|
|
185
191
|
raise ValueError('This turn is read-only, so tools that change data in a connected app cannot run now. Tell the user what you would do and ask them to confirm.')
|
|
186
|
-
req = {'kind': 'integration', 'connection': connection, 'name': name, 'arguments': args.get('arguments') or {}, 'command': 'Connected tool: ' + name}
|
|
187
|
-
# The app runs the call with the user's sign-in, so the request
|
|
188
|
-
|
|
192
|
+
req = {'kind': 'integration', 'connection': connection, 'name': name, 'arguments': args.get('arguments') or {}, 'command': 'Connected tool: ' + name, 'write': match.get('write') is not False} # BUG-0457
|
|
193
|
+
# The app runs the call with the user's sign-in, so the request goes through the app. Connected-app calls ask even in
|
|
194
|
+
# Auto mode, as Connectors and the docs promise (BUG-0485); "Allow for this task" on the card covers that app afterwards.
|
|
189
195
|
ident = ask(req, always=True)
|
|
190
196
|
if not ident: raise ValueError('The user rejected this integration call.')
|
|
191
197
|
result = ROOT / ('result-' + ident)
|
|
@@ -224,6 +230,12 @@ def fingerprint_now():
|
|
|
224
230
|
from quality import fingerprint
|
|
225
231
|
return fingerprint(REPO, run_state().get('base') or 'HEAD')
|
|
226
232
|
|
|
233
|
+
def readonly_browser():
|
|
234
|
+
"""BUG-0202: the Browser panel says Playwright can open the local app, so the agent must have the tool in question and
|
|
235
|
+
no-change turns too. A browser check only looks at the running app (each scenario still asks first); it never edits files."""
|
|
236
|
+
if os.getenv('MINI_SUBAGENT') == '1': return []
|
|
237
|
+
return [t for t in WRITE_TOOLS if t.get('name') == 'browser']
|
|
238
|
+
|
|
227
239
|
def browser(args):
|
|
228
240
|
args = {k: v for k, v in args.items() if k in ('url', 'width', 'height', 'steps')}
|
|
229
241
|
installed = False
|
|
@@ -337,7 +349,15 @@ def simulator(args):
|
|
|
337
349
|
if not seen: return text
|
|
338
350
|
return {'_mcp_content': [{'type': 'text', 'text': text}, {'type': 'image', 'mimeType': 'image/jpeg', 'data': base64.b64encode(Path(seen['path']).read_bytes()).decode()}]}
|
|
339
351
|
|
|
352
|
+
try: # BUG-0523: structured "cannot complete" outcome
|
|
353
|
+
from mlai_hooks import CANNOT_COMPLETE_TOOL as _CC_TOOL
|
|
354
|
+
TOOLS.append(_CC_TOOL)
|
|
355
|
+
except Exception: pass
|
|
356
|
+
|
|
340
357
|
def call(name, args):
|
|
358
|
+
if name == 'cannot_complete':
|
|
359
|
+
from mlai_hooks import record_cannot_complete
|
|
360
|
+
return record_cannot_complete(args, ROOT)
|
|
341
361
|
if name == 'schedule_wakeup':
|
|
342
362
|
import wakeups
|
|
343
363
|
return wakeups.tool(ROOT, args, emit)
|
|
@@ -404,7 +424,7 @@ def call(name, args):
|
|
|
404
424
|
return web('web_fetch', {'urls': [str(u) for u in urls]})
|
|
405
425
|
if name == 'evidence': return ledger().run({'claims': args['claims']} if args.get('claims') else {})
|
|
406
426
|
if name in ('browser', 'preview', 'job', 'network'):
|
|
407
|
-
if os.getenv('MINI_READONLY') == '1': raise ValueError('Not available in read-only mode')
|
|
427
|
+
if os.getenv('MINI_READONLY') == '1' and name != 'browser': raise ValueError('Not available in read-only mode')
|
|
408
428
|
return {'browser': browser, 'preview': preview, 'job': job, 'network': network}[name](args)
|
|
409
429
|
if name in ('list_sessions', 'read_messages', 'send_message'): return hub(name, args)
|
|
410
430
|
if name in ('repo_map', 'symbols'):
|
|
@@ -441,7 +461,7 @@ def main():
|
|
|
441
461
|
if msg_id is None: continue # notifications
|
|
442
462
|
if method == 'initialize':
|
|
443
463
|
reply(msg_id, {'protocolVersion': (msg.get('params') or {}).get('protocolVersion') or '2025-06-18', 'capabilities': {'tools': {}}, 'serverInfo': {'name': 'mlai', 'version': '1.0'}})
|
|
444
|
-
elif method == 'tools/list': reply(msg_id, {'tools': TOOLS + (
|
|
464
|
+
elif method == 'tools/list': reply(msg_id, {'tools': TOOLS + (readonly_browser() if os.getenv('MINI_READONLY') == '1' else WRITE_TOOLS + (INTEGRATION_TOOLS if enabled_tools() else []) + ([DEPLOY_TOOL] if deploy_info() else []) + ([computer_use.TOOL] if computer_use.available() else []) + sim_tools()) + hub_tools()})
|
|
445
465
|
elif method == 'tools/call':
|
|
446
466
|
p = msg.get('params') or {}
|
|
447
467
|
try:
|
|
@@ -0,0 +1,35 @@
|
|
|
1
|
+
# ML AI Review
|
|
2
|
+
|
|
3
|
+
The verified reviewer (relevance filter, outline context and related code, linter diagnostics, three
|
|
4
|
+
independent readings, grouping and matching, refuter, noise budget, streaming inline comments,
|
|
5
|
+
"looks resolved" lifecycle, feedback from dismissed comments, security specialist, eval harness),
|
|
6
|
+
running inside our worker instead of Temporal and Azure.
|
|
7
|
+
|
|
8
|
+
python -m mlai_review run REQUEST.json # one review; events to request.out (JSONL)
|
|
9
|
+
python -m mlai_review analyse --root DIR --out FILE [--no-eslint] [--no-install]
|
|
10
|
+
python -m mlai_review feedback owner/repo --days 30
|
|
11
|
+
python -m mlai_review eval [--cases DIR] [--mode deep] [--record --request REQUEST.json]
|
|
12
|
+
|
|
13
|
+
The model key comes from `MODEL_API_KEY` (our model route or a run token for the gateway) and,
|
|
14
|
+
for pull requests, `GITHUB_TOKEN`; both are removed from the environment before anything else runs.
|
|
15
|
+
|
|
16
|
+
Where it runs:
|
|
17
|
+
|
|
18
|
+
- Chat in the cloud: `lib/code-review.ts` (actions `review_start`, `review_status`, `review_dismiss`).
|
|
19
|
+
- Chat on This Mac: `apps/laptop/src/review.mjs`, same actions, run token from `/api/laptop`.
|
|
20
|
+
- Pull requests: `lib/code-review-run.ts`, started by the "Review pull requests" routine
|
|
21
|
+
(`lib/code-review-routine.ts`). Linters run first with no secrets present; the review then posts
|
|
22
|
+
inline comments and an "ML AI Review" check run, or a commit status when the token cannot
|
|
23
|
+
create check runs. `workflows/watch-pr-review.ts` stops the computer as soon as the review ends
|
|
24
|
+
(or the reviewer exits); its 40-minute lifetime is only a ceiling.
|
|
25
|
+
|
|
26
|
+
Metering: cloud chat reviews are runs in the ledger (`lib/code-review-billing.ts`): gated like a task
|
|
27
|
+
start, capped by the request's `budget` (billable tokens left) and weighted by `creditRates`, exactly
|
|
28
|
+
as the task worker and model gateway weigh tokens. The final `usage` event carries `billable`; past
|
|
29
|
+
the cap no model call starts and a `limited` event explains why the review is partial.
|
|
30
|
+
|
|
31
|
+
Modes (`settings.for_mode`): `standard` (3 readings), `deep` (5 readings, refuter with more effort
|
|
32
|
+
and lookups, write-path check), `security` (security specialist, refuter). Scope `repo` reviews
|
|
33
|
+
the source files of the whole repository, sensitive paths first.
|
|
34
|
+
|
|
35
|
+
Without the tree-sitter wheel, outlines fall back to `infrastructure/structure/heuristic.py`.
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
"""ML AI orchestrator: runs each pull request review as a durable workflow."""
|