mlai-code 0.3.2 → 0.3.4
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 +178 -31
- package/src/cloud.mjs +2 -0
- package/src/engine.mjs +1 -1
- package/src/headless.mjs +25 -9
- 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 +20 -1
- package/src/relay.mjs +1 -1
- package/src/render.mjs +12 -3
- package/src/runtime.mjs +23 -1
- package/src/sandbox.mjs +19 -3
- package/src/server.mjs +46 -7
- package/src/session.mjs +53 -10
- package/src/shared-data.mjs +1 -1
- package/src/tui/acp.mjs +121 -19
- package/src/tui/app.mjs +24 -5
- package/src/tui/chat.mjs +43 -16
- package/src/tui/cmd-config.mjs +3 -1
- package/src/tui/cmd-core.mjs +11 -7
- package/src/tui/cmd-extensibility.mjs +13 -2
- package/src/tui/commands.mjs +2 -1
- 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/statusline.mjs +2 -1
- package/src/two-words.mjs +14 -8
- package/src/view-sync.mjs +48 -0
- package/src/wakeups.mjs +2 -2
- package/src/winenv.mjs +1 -1
- package/src/workspace-api.mjs +21 -4
- package/vendor/client-sdk/index.cjs +5 -4
- package/vendor/cloud-cli/mlai.mjs +184 -28
- package/worker/architecture_worker.py +33 -4
- package/worker/browser_tool.py +19 -1
- package/worker/build_planner.py +75 -7
- package/worker/chat_fork.py +4 -1
- package/worker/checkpoints.py +71 -8
- package/worker/cloud_pty.py +91 -2
- package/worker/command_safety.py +80 -3
- package/worker/compaction_brief.py +186 -0
- package/worker/connected_turn.py +30 -0
- package/worker/context_manager.py +6 -0
- package/worker/conversation.py +139 -12
- package/worker/deck_render.js +11 -0
- package/worker/deck_tools.py +825 -0
- package/worker/git_sync.py +42 -17
- package/worker/goals.py +4 -0
- package/worker/image_generation.py +102 -23
- package/worker/kimi_bridge.py +580 -52
- package/worker/kimi_gates.py +74 -6
- package/worker/kimi_hook.py +213 -19
- package/worker/media_tools.py +182 -0
- package/worker/memory.py +2 -2
- package/worker/mlai_hooks.py +314 -5
- package/worker/mlai_mcp.py +49 -11
- 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/model_stream.py +21 -7
- package/worker/offline_guard.py +1 -1
- package/worker/preview.py +62 -4
- package/worker/product_facts.py +20 -0
- package/worker/project_context.py +33 -7
- package/worker/reply_language.py +92 -10
- package/worker/repo_policy.py +5 -0
- package/worker/roles.json +199 -37
- package/worker/roles.py +271 -9
- package/worker/runner.py +366 -53
- package/worker/runtime_control.py +64 -3
- package/worker/secret_scan.py +3 -2
- package/worker/server_memory.py +122 -0
- package/worker/site_check.py +279 -0
- package/worker/skill_library.py +15 -3
- package/worker/skill_pack.py +170 -16
- package/worker/subagents.py +6 -2
- package/worker/tool_catalog.py +4 -3
- package/worker/verification.py +156 -19
- package/worker/wakeups.py +56 -4
package/worker/kimi_gates.py
CHANGED
|
@@ -10,7 +10,7 @@ really done also run here, after the engine exits, and a failed check resumes th
|
|
|
10
10
|
and the bridge runs them. Only assertion failures count; a broken test file is treated as the tester's fault.
|
|
11
11
|
|
|
12
12
|
Also: deterministic trimming of oversized tool output in model requests (cache-friendly, saves tokens)."""
|
|
13
|
-
import json, os, re, subprocess, sys, time
|
|
13
|
+
import json, os, re, shutil, subprocess, sys, time
|
|
14
14
|
from pathlib import Path
|
|
15
15
|
|
|
16
16
|
ACCEPTANCE_DIR = 'acceptance'
|
|
@@ -71,6 +71,8 @@ Rules:
|
|
|
71
71
|
- 3 to 8 focused test cases. No network, no sleeping, no servers, no new dependencies, no file writes outside a temporary directory.
|
|
72
72
|
- Python: a unittest module. The repository root is on sys.path, so import modules as the repository does (e.g. `from shop.cart import total`).
|
|
73
73
|
- JavaScript: an ES module using node:test and node:assert/strict. Import repository files with absolute file URLs built from REPO_ROOT, e.g. `const m = await import(REPO_ROOT + '/calc.js')` where `const REPO_ROOT = 'file://' + process.env.ACCEPTANCE_REPO;`. If the module is CommonJS, use createRequire.
|
|
74
|
+
- To read a repository file, use a path, not a URL: `fs.readFileSync(path.join(process.env.ACCEPTANCE_REPO, 'src/styles.css'), 'utf8')`. ACCEPTANCE_REPO is a plain folder path; for URLs use process.env.ACCEPTANCE_REPO_URL (a file:// URL ending in /) or pathToFileURL from node:url. Never pass the plain path to new URL().
|
|
75
|
+
- Colours, sizes and other CSS values have many equal spellings (#141, #141210, #141210ff, rgb(20 18 16), a named colour). Never match a colour or CSS value with one spelling; if the task says "near-black", parse the colour and check its lightness.
|
|
74
76
|
- Prefer expected values stated in the task or in the existing tests. When you compute one, write the arithmetic in a comment and check it twice.
|
|
75
77
|
- If the task is not testable this way (UI-only, docs, config, research, or behaviour you cannot pin down from the task), say so.
|
|
76
78
|
|
|
@@ -113,9 +115,17 @@ def tester_invalid(language, output):
|
|
|
113
115
|
if re.search(r'^FAIL: ', output, re.M): return False
|
|
114
116
|
return bool(re.search(r'(ImportError|ModuleNotFoundError|SyntaxError|IndentationError|NameError)', output)) or not re.search(r'^ERROR: ', output, re.M)
|
|
115
117
|
if re.search(r'AssertionError|ERR_ASSERTION', output): return False
|
|
116
|
-
|
|
118
|
+
# ERR_INVALID_URL: new URL('./src/styles.css', ACCEPTANCE_REPO + '/') with a plain folder path (CLI bug list, 0.3.3).
|
|
119
|
+
return bool(re.search(r'ERR_MODULE_NOT_FOUND|SyntaxError|Cannot find module|ReferenceError|does not provide an export|ERR_INVALID_URL|TypeError: Invalid URL|ERR_INVALID_FILE_URL_PATH|ENOENT: no such file', output)) or '# fail 0' in output
|
|
117
120
|
|
|
118
|
-
|
|
121
|
+
TESTER_RETRY = '''
|
|
122
|
+
|
|
123
|
+
This is a second try. Your previous test could not run; the error is in `error`. Fix the test itself (imports, paths,
|
|
124
|
+
module format, a missing loader) and return the same JSON shape. Test through files and commands that run in plain
|
|
125
|
+
Node or Python here: for a React or other bundled UI, test the server, data and pure modules, or read the built output,
|
|
126
|
+
instead of importing JSX. If no test can run in this repository, return {"applicable": false, "reason": "..."}.'''
|
|
127
|
+
|
|
128
|
+
REVIEW_SYSTEM = '''You check failing acceptance tests before they are reported to a builder. For each failing test, decide whether the test itself is right: does the task text really require this behaviour, and is the expected value correct? Recompute every expected number step by step from the task. A test is wrong if it matches a colour or CSS value by one spelling (a 3-digit hex pattern that rejects #141210), or if it fails while loading a file (Invalid URL, ENOENT) instead of on the behaviour. A test is wrong if it asserts something the task does not require, or if its expected value is miscalculated.
|
|
119
129
|
|
|
120
130
|
Reply with JSON only: {"genuine": ["test_name", ...], "wrong": [{"test": "test_name", "why": "..."}]}'''
|
|
121
131
|
|
|
@@ -141,14 +151,26 @@ def run_acceptance(root, repo, language, code, timeout=180):
|
|
|
141
151
|
name = 'test_acceptance.py' if language == 'python' else 'acceptance.test.mjs'
|
|
142
152
|
path = d / name; path.write_text(code)
|
|
143
153
|
env = {k: v for k, v in os.environ.items() if k not in ('MODEL_API_KEY', 'MINI_VAULT', 'EXA_API_KEY', 'FIRECRAWL_API_KEY', 'VERCEL_OIDC_TOKEN')}
|
|
144
|
-
env.update({'PYTHONPATH': str(repo), 'ACCEPTANCE_REPO': str(repo), 'PYTHONDONTWRITEBYTECODE': '1'})
|
|
145
|
-
cmd = [sys.executable, str(path), '-v'] if language == 'python' else ['node', '--test', str(path)]
|
|
154
|
+
env.update({'PYTHONPATH': str(repo), 'ACCEPTANCE_REPO': str(repo), 'ACCEPTANCE_REPO_URL': Path(repo).resolve().as_uri().rstrip('/') + '/', 'PYTHONDONTWRITEBYTECODE': '1'})
|
|
146
155
|
if language == 'python': path.write_text(code if 'unittest.main' in code else code + '\n\nif __name__ == "__main__":\n import unittest; unittest.main()\n')
|
|
156
|
+
# BUG-0221: run the test from a scratch folder inside the project, so relative imports, package imports and
|
|
157
|
+
# __file__-based paths resolve against the project, not the engine's own folder. The copy in root stays for the log.
|
|
158
|
+
scratch = Path(repo) / ('.mlai-acceptance-' + os.urandom(4).hex())
|
|
159
|
+
try:
|
|
160
|
+
scratch.mkdir()
|
|
161
|
+
run_path = scratch / name; run_path.write_text(path.read_text())
|
|
162
|
+
except OSError:
|
|
163
|
+
scratch, run_path = None, path
|
|
164
|
+
cmd = [sys.executable, str(run_path), '-v'] if language == 'python' else ['node', '--test', str(run_path)]
|
|
147
165
|
try:
|
|
148
166
|
r = subprocess.run(cmd, cwd=repo, env=env, capture_output=True, text=True, timeout=timeout)
|
|
149
|
-
|
|
167
|
+
out = (r.stdout + '\n' + r.stderr)
|
|
168
|
+
if scratch is not None: out = out.replace(str(scratch) + '/', '').replace(str(scratch), '.')
|
|
169
|
+
return r.returncode, out[-6000:]
|
|
150
170
|
except subprocess.TimeoutExpired:
|
|
151
171
|
return 124, 'Acceptance check timed out after %ds' % timeout
|
|
172
|
+
finally:
|
|
173
|
+
if scratch is not None: shutil.rmtree(scratch, ignore_errors=True)
|
|
152
174
|
|
|
153
175
|
def acceptance(config, st, model, root, repo, changed, event):
|
|
154
176
|
"""Write and run the hidden acceptance test once. Returns {status, ...}: passed | failed | skipped | invalid."""
|
|
@@ -172,6 +194,24 @@ def acceptance(config, st, model, root, repo, changed, event):
|
|
|
172
194
|
event('acceptance', status='skipped', content='Acceptance check not applicable: ' + str(reason)[:300])
|
|
173
195
|
return {'status': 'skipped', 'reason': str(reason)[:300]}
|
|
174
196
|
code, output = run_acceptance(root, repo, spec['language'], spec['code'])
|
|
197
|
+
if code not in (0, 124) and tester_invalid(spec['language'], output):
|
|
198
|
+
# Web bug list Bug-017 and Bug-021: a test that cannot even load (JSX imported into plain Node, a wrong path)
|
|
199
|
+
# says nothing about the build. Give the tester its own error once and run the corrected test.
|
|
200
|
+
retry = json.dumps({'task': str(config.get('task', ''))[:12000], 'languages_changed': sorted(langs), 'repository_files': files,
|
|
201
|
+
'previous_test': spec['code'][:20000], 'error': output[-3000:]})
|
|
202
|
+
try:
|
|
203
|
+
fixed_text, _ = model.request([{'role': 'system', 'content': TESTER_SYSTEM + TESTER_RETRY},
|
|
204
|
+
{'role': 'user', 'content': retry}], max_tokens=12000, allow_tools=False)
|
|
205
|
+
fixed = _parse(fixed_text)
|
|
206
|
+
except Exception:
|
|
207
|
+
fixed = None
|
|
208
|
+
if fixed and fixed.get('applicable') and fixed.get('language') == spec['language'] and isinstance(fixed.get('code'), str) and len(fixed['code']) <= 40000:
|
|
209
|
+
spec = {**spec, **fixed}
|
|
210
|
+
code, output = run_acceptance(root, repo, spec['language'], spec['code'])
|
|
211
|
+
elif isinstance(fixed, dict) and fixed.get('applicable') is False:
|
|
212
|
+
reason = str(fixed.get('reason') or 'the tester could not write a test that runs here')[:300]
|
|
213
|
+
event('acceptance', status='skipped', content='Acceptance check not applicable: ' + reason)
|
|
214
|
+
return {'status': 'skipped', 'reason': reason}
|
|
175
215
|
covers = [str(c)[:200] for c in spec.get('covers') or []][:8]
|
|
176
216
|
if code == 0: status = 'passed'
|
|
177
217
|
elif code != 124 and tester_invalid(spec['language'], output): status = 'invalid'
|
|
@@ -250,6 +290,10 @@ def final_issues(config, st, answer, done, model, root, repo, event, allow_accep
|
|
|
250
290
|
issues = []
|
|
251
291
|
try: changed = role_engine.changed_files(repo, st.get('base'))
|
|
252
292
|
except Exception: changed = []
|
|
293
|
+
try:
|
|
294
|
+
from runtime_control import turn_changes
|
|
295
|
+
changed = turn_changes(root, repo, changed) # BUG-0495: not files that were already untracked or changed at the start
|
|
296
|
+
except Exception: pass
|
|
253
297
|
code_changed = [p for p in changed if role_engine.CODE_EXT.search(p)]
|
|
254
298
|
progress = {}
|
|
255
299
|
try: progress = json.loads((Path(root) / 'kimi-progress.json').read_text())
|
|
@@ -291,6 +335,30 @@ def final_issues(config, st, answer, done, model, root, repo, event, allow_accep
|
|
|
291
335
|
except Exception as exc:
|
|
292
336
|
issue = 'Verification report could not complete: ' + str(exc)[:200]
|
|
293
337
|
event('note', content=issue); issues.append(issue)
|
|
338
|
+
if allow_acceptance and changed and not st.get('inspect'):
|
|
339
|
+
# The running preview is the user's view of the work: a page that answers 500, or a linked page that 404s, is not
|
|
340
|
+
# done, whatever the final message says. Nothing listening on the preview port means there is nothing to judge.
|
|
341
|
+
try:
|
|
342
|
+
import site_check
|
|
343
|
+
routes = sorted(site_check.internal_links(repo, changed))
|
|
344
|
+
bad = site_check.live_routes(routes)
|
|
345
|
+
found = site_check.live_findings(bad)
|
|
346
|
+
if found:
|
|
347
|
+
event('verification_report', status='failed', reasons=[found[0]['message'].split('.')[0]], findings=found, checks=0, passed=0)
|
|
348
|
+
issues += [f['message'] for f in found if f['message'] not in issues]
|
|
349
|
+
except Exception as exc:
|
|
350
|
+
event('note', content='Could not check the preview routes: ' + str(exc)[:200])
|
|
351
|
+
if allow_acceptance and config.get('conversation_route') in (None, 'build'):
|
|
352
|
+
# A run must not end on "I'll now make the slides", and a presentation or PDF request must leave a file behind.
|
|
353
|
+
# Each note is sent once; if the agent still cannot do it, its own explanation stands.
|
|
354
|
+
try:
|
|
355
|
+
from conversation import deliverable_issue, stalled_promise, STALLED_NOTE
|
|
356
|
+
if 'stalled' not in done and not changed and stalled_promise(answer):
|
|
357
|
+
done.add('stalled'); issues.append(STALLED_NOTE)
|
|
358
|
+
elif 'deliverable' not in done:
|
|
359
|
+
note = deliverable_issue(config.get('task', ''), changed, repo)
|
|
360
|
+
if note: done.add('deliverable'); issues.append(note)
|
|
361
|
+
except Exception: pass # these hints never block a run
|
|
294
362
|
if config.get('selected_design'):
|
|
295
363
|
from design_director import browser_review
|
|
296
364
|
from quality import fingerprint
|
package/worker/kimi_hook.py
CHANGED
|
@@ -30,7 +30,7 @@ def save(name, data):
|
|
|
30
30
|
def event(event_type, **data):
|
|
31
31
|
with (ROOT / 'events.jsonl').open('a') as f: f.write(json.dumps({'type': event_type, 'time': time.time(), **data}) + '\n')
|
|
32
32
|
|
|
33
|
-
HOOK_KEYS = ('recent', 'landed', 'stuck_warned', 'gates_done', 'planned', 'plan_reviewed', 'plan_dismissed', 'brief_ok', 'brief_rounds', 'brief_nudged', 'mode', 'nudged', 'diagnosed_for', 'scope', 'scope_ok', 'approved_kinds', 'calls', 'agents', 'budget_warned', 'touched', 'policy_noted', 'policy_ok_paths', 'token_stage', 'account_stop', 'account_stop_at', 'helpers', 'budget_next', 'budget_declined')
|
|
33
|
+
HOOK_KEYS = ('steps_used', 'step_limit_noted', 'recent', 'landed', 'stuck_warned', 'gates_done', 'planned', 'plan_reviewed', 'plan_dismissed', 'brief_ok', 'brief_rounds', 'brief_nudged', 'mode', 'nudged', 'diagnosed_for', 'scope', 'scope_ok', 'approved_kinds', 'calls', 'agents', 'budget_warned', 'touched', 'policy_noted', 'policy_ok_paths', 'token_stage', 'account_stop', 'account_stop_at', 'helpers', 'budget_next', 'budget_declined', 'plan_items', 'brief_q_nudged', 'plan_nudged', 'question_changes')
|
|
34
34
|
def save_state(st):
|
|
35
35
|
"""The hook owns only these keys; the bridge owns kimi-state.json (static) and kimi-progress.json."""
|
|
36
36
|
save('kimi-hook.json', {k: st[k] for k in HOOK_KEYS if k in st})
|
|
@@ -38,6 +38,22 @@ def save_state(st):
|
|
|
38
38
|
def state():
|
|
39
39
|
return {**load('kimi-state.json', {}), **load('kimi-progress.json', {}), **load('kimi-hook.json', {})}
|
|
40
40
|
|
|
41
|
+
# Tools that change the project, run code or reach outside it. Before a Brief task list is approved, and after the user
|
|
42
|
+
# dismissed a plan (Not now), none of them run: the agent reads, answers and waits (BUG-0102).
|
|
43
|
+
WORK_TOOLS = frozenset(('Write', 'Edit', 'Bash', 'Agent', 'AgentSwarm'))
|
|
44
|
+
WORK_MCP = frozenset(('scope', 'preview', 'browser', 'job', 'deploy', 'integration_call', 'network', 'schedule_wakeup', 'send_message', 'ios_simulator', 'generate_image', 'stock_photo', 'brand_lookup', 'deck_check', 'deck_import', 'find_image', 'find_icon', 'image_prep'))
|
|
45
|
+
DISMISSED_BLOCK = ('The user dismissed the plan (Not now). Do not start the work in this task: no edits, commands, images, scope '
|
|
46
|
+
'requests or helper agents, and do not look for another way to start. In one or two sentences, say the plan is on hold and ask '
|
|
47
|
+
'what they would like changed, then finish. The user starts the work again from the chat.')
|
|
48
|
+
|
|
49
|
+
def is_work_tool(name):
|
|
50
|
+
name = str(name or '')
|
|
51
|
+
if name in WORK_TOOLS: return True
|
|
52
|
+
m = re.match(r'^mcp__(\w+?)__(\w+)$', name)
|
|
53
|
+
if not m: return False
|
|
54
|
+
if m.group(1) != 'mlai': return True # connected-app and other MCP tools can change outside data
|
|
55
|
+
return m.group(2) in WORK_MCP
|
|
56
|
+
|
|
41
57
|
def block(reason):
|
|
42
58
|
sys.stderr.write(reason); sys.stderr.flush(); sys.exit(2)
|
|
43
59
|
|
|
@@ -45,13 +61,16 @@ def steering():
|
|
|
45
61
|
"""Consume queued user instructions (oldest first)."""
|
|
46
62
|
d = ROOT / 'steering'
|
|
47
63
|
if not d.is_dir(): return []
|
|
48
|
-
out = []
|
|
64
|
+
out, notes = [], []
|
|
49
65
|
for f in sorted(d.glob('*.json')):
|
|
50
|
-
try:
|
|
51
|
-
except (OSError, ValueError):
|
|
66
|
+
try: note = json.loads(f.read_text())
|
|
67
|
+
except (OSError, ValueError): note = {}
|
|
68
|
+
msg = note.get('message', '') if isinstance(note, dict) else ''
|
|
52
69
|
f.unlink(missing_ok=True)
|
|
53
|
-
if isinstance(msg, str) and msg.strip():
|
|
54
|
-
|
|
70
|
+
if isinstance(msg, str) and msg.strip():
|
|
71
|
+
out.append(msg.strip()[:5000]); notes.append({'note': f.stem, **({'source': str(note['source'])} if note.get('source') else {})})
|
|
72
|
+
# The note id lets the app match a delivered instruction to the message it sent.
|
|
73
|
+
for m, n in zip(out, notes): event('steering', content=m, **n)
|
|
55
74
|
return out
|
|
56
75
|
|
|
57
76
|
def platform_notes():
|
|
@@ -73,11 +92,19 @@ def steer_block(messages, stopping=False):
|
|
|
73
92
|
def wait_decision(request, seconds=540, strict=False):
|
|
74
93
|
"""One approval at a time even when Kimi runs tool calls in parallel. Strict (repository policy) reviews ignore "Approve all"."""
|
|
75
94
|
import fcntl
|
|
95
|
+
queued = time.time()
|
|
76
96
|
with (ROOT / 'approval.lock').open('a') as lock:
|
|
77
97
|
fcntl.flock(lock, fcntl.LOCK_EX)
|
|
98
|
+
# BUG-0251: a call that queued behind an approval the user then rejected (parallel tool calls of one step)
|
|
99
|
+
# is rejected too, so a leftover card can never run a change the user just said no to.
|
|
100
|
+
if rejected_since(queued): block('The user rejected a related action in this step, so this call was not run. Do not retry it; explain what you would change and ask.')
|
|
78
101
|
if not strict and (ROOT / 'kimi-approve-all').exists(): return
|
|
79
102
|
_wait(request, seconds)
|
|
80
103
|
|
|
104
|
+
def rejected_since(t):
|
|
105
|
+
try: return float((ROOT / 'last-reject').read_text().strip() or 0) >= t
|
|
106
|
+
except (OSError, ValueError): return False
|
|
107
|
+
|
|
81
108
|
def user_hooks(fn, *args):
|
|
82
109
|
"""Project hooks and hard-deny rules (mlai_hooks). They fail open on an internal error, like the rest of this hook."""
|
|
83
110
|
try:
|
|
@@ -88,6 +115,9 @@ def user_hooks(fn, *args):
|
|
|
88
115
|
def mlai_run(st):
|
|
89
116
|
return user_hooks('kimi_run', st) or {}
|
|
90
117
|
|
|
118
|
+
REJECTED = ('The user rejected this action. Do not make the same change another way (another tool, a shell command or a retry) '
|
|
119
|
+
'and do not claim the permission mode changed. Stop now, say plainly what was not done, and ask the user how to proceed.')
|
|
120
|
+
|
|
91
121
|
def _wait(request, seconds):
|
|
92
122
|
ident = request['id']
|
|
93
123
|
if request.get('kind') not in ('plan', 'brief'):
|
|
@@ -106,8 +136,21 @@ def _wait(request, seconds):
|
|
|
106
136
|
verdict = decision.read_text().strip(); decision.unlink(missing_ok=True); (ROOT / 'pending.json').unlink(missing_ok=True)
|
|
107
137
|
if verdict == 'approve_all':
|
|
108
138
|
(ROOT / 'kimi-approve-all').write_text('1'); event('mode', mode='auto', content='Approved all remaining steps for this task.')
|
|
139
|
+
if verdict == 'approve_chat':
|
|
140
|
+
# BUG-0594: remember this exact command for the chat, only when it is a safe kind; otherwise it is a one-time approval.
|
|
141
|
+
try:
|
|
142
|
+
from command_safety import chat_allow_ok, chat_allow_key
|
|
143
|
+
if request.get('kind') == 'command' and not request.get('strict') and chat_allow_ok(request.get('effect'), request.get('command'), REPO):
|
|
144
|
+
allowed = [c for c in load('chat-allow.json', []) if isinstance(c, str)]
|
|
145
|
+
key = chat_allow_key(request.get('command'))
|
|
146
|
+
if key not in allowed: save('chat-allow.json', (allowed + [key])[-50:])
|
|
147
|
+
event('note', content='Always allowed in this chat: ' + key[:200])
|
|
148
|
+
except Exception: pass
|
|
149
|
+
verdict = 'approve'
|
|
109
150
|
if verdict not in ('approve', 'approve_all'):
|
|
110
|
-
|
|
151
|
+
(ROOT / 'last-reject').write_text(str(time.time()))
|
|
152
|
+
# BUG-0491: "Choose another approach" made the agent retry the same write with Write, printf and touch.
|
|
153
|
+
block(REJECTED)
|
|
111
154
|
|
|
112
155
|
def wait_plan(request, seconds=540):
|
|
113
156
|
"""Plan review: returns the raw verdict (approve | approve:0,2 | reject), or None on timeout."""
|
|
@@ -127,11 +170,29 @@ def briefing(st):
|
|
|
127
170
|
"""Brief mode before the user has approved a task list: read-only, every TodoList opens a review."""
|
|
128
171
|
return bool(st.get('brief')) and not st.get('brief_ok') and not st.get('plan_dismissed')
|
|
129
172
|
|
|
173
|
+
def brief_label(st):
|
|
174
|
+
"""BUG-0591: an ask-first brief in an Auto or Guided chat is not "Brief mode" (the composer still says Auto), so the
|
|
175
|
+
agent is not told, and does not tell the user, that the workspace is in Brief mode."""
|
|
176
|
+
back = st.get('brief_return')
|
|
177
|
+
if back in ('auto', 'approve'): return 'Ask first (the chat stays in %s)' % ('Auto' if back == 'auto' else 'Guided')
|
|
178
|
+
return 'Brief mode'
|
|
179
|
+
|
|
130
180
|
def brief_unlock(st):
|
|
131
181
|
back = st.get('brief_return') if st.get('brief_return') in ('approve', 'auto') else 'approve'
|
|
132
182
|
st['brief_ok'] = True; st['mode'] = back; st['scope_ok'] = True; save_state(st)
|
|
133
183
|
event('mode', mode=back, content='Plan agreed. Building the approved list in %s mode.' % ('Auto' if back == 'auto' else 'Guided'))
|
|
134
184
|
|
|
185
|
+
def plan_key(text):
|
|
186
|
+
return ' '.join(str(text).lower().split())[:300]
|
|
187
|
+
|
|
188
|
+
def plan_added(items, st):
|
|
189
|
+
"""Items in this task list that the user has not approved yet. Status updates of the same list add nothing."""
|
|
190
|
+
known = set(st.get('plan_items') or [])
|
|
191
|
+
return [i for i in items if plan_key(i) not in known]
|
|
192
|
+
|
|
193
|
+
def remember_plan(items, st):
|
|
194
|
+
st['plan_items'] = sorted(set(st.get('plan_items') or []) | {plan_key(i) for i in items})[:200]
|
|
195
|
+
|
|
135
196
|
def plan_gate(args, st):
|
|
136
197
|
"""First task list of a run: optional user review (Guided mode by default, or always when the user asked).
|
|
137
198
|
Brief mode: every task list is reviewed until the user approves; replies come back to the agent as feedback."""
|
|
@@ -142,9 +203,24 @@ def plan_gate(args, st):
|
|
|
142
203
|
st['planned'] = True; save_state(st)
|
|
143
204
|
policy = st.get('plan_approval') or 'guided'
|
|
144
205
|
wants = brief or policy == 'always' or (policy == 'guided' and st.get('mode') == 'approve')
|
|
206
|
+
if wants and not brief and question_only(st):
|
|
207
|
+
# BUG-0345: a task list for answering a question needs no review. It does not count as the plan, so a later
|
|
208
|
+
# edit in this turn still asks for the plan first.
|
|
209
|
+
st['planned'] = False; save_state(st); return
|
|
145
210
|
if brief:
|
|
146
|
-
if len(items) < 2: block('
|
|
147
|
-
|
|
211
|
+
if len(items) < 2: block(brief_label(st) + ': the task list needs at least two concrete items before the user can review it. Call TodoList again with the full plan.')
|
|
212
|
+
if st.get('mode') == 'brief' and not int(st.get('brief_rounds') or 0) and not st.get('brief_q_nudged') and not load('kimi-brief-questions.json', []):
|
|
213
|
+
# BUG-0124: Brief mode promises to clarify requirements before the task list. Ask once for the open questions.
|
|
214
|
+
st['brief_q_nudged'] = True; st['planned'] = False; save_state(st)
|
|
215
|
+
event('note', content='Brief mode: asked for clarifying questions before the task list')
|
|
216
|
+
block('Brief mode clarifies requirements first. Before this task list goes to the user, record one to three short questions about '
|
|
217
|
+
'what is still open in the request (for example style, content, images or scope) with mcp__mlai__brief_questions '
|
|
218
|
+
'{"questions": [{"question": "...", "options": [{"label": "...", "detail": "..."}], "recommended": 0}]}, each with your '
|
|
219
|
+
'recommended answer. Then call TodoList again with the same list. If the request truly leaves nothing open, call TodoList again unchanged.')
|
|
220
|
+
elif st.get('readonly') or len(items) < 2 or not wants or (ROOT / 'kimi-approve-all').exists(): return
|
|
221
|
+
elif not first or st.get('plan_reviewed'):
|
|
222
|
+
# Guided: an approved list is not asked again, but a later list that adds work asks again (BUG-0039).
|
|
223
|
+
if not st.get('plan_reviewed') or not plan_added(items, st): return
|
|
148
224
|
st['plan_reviewed'] = True
|
|
149
225
|
rounds = int(st.get('brief_rounds') or 0) + 1
|
|
150
226
|
if brief: st['brief_rounds'] = rounds
|
|
@@ -173,17 +249,19 @@ def plan_gate(args, st):
|
|
|
173
249
|
block('The user replied to your plan: "' + reply + '"\nAnswer briefly if they asked something, then call TodoList again with the revised task list (3 to 10 items) and, if anything is still open, record your questions with mcp__mlai__brief_questions {"questions": [...]} before the TodoList. Nothing is built until they approve.')
|
|
174
250
|
if verdict in ('approve', 'approve_all'):
|
|
175
251
|
if brief: brief_unlock(st)
|
|
252
|
+
remember_plan(items, st)
|
|
176
253
|
st['scope_ok'] = True; save_state(st); event('plan_review', content='Plan approved'); return
|
|
177
254
|
if verdict.startswith('approve:'):
|
|
178
255
|
keep = {int(x) for x in verdict[8:].split(',') if x.strip().isdigit()}
|
|
179
256
|
kept = [t for i, t in enumerate(items) if i in keep]; removed = [t for i, t in enumerate(items) if i not in keep]
|
|
180
257
|
if kept:
|
|
181
258
|
if brief: brief_unlock(st)
|
|
259
|
+
remember_plan(kept, st); save_state(st)
|
|
182
260
|
event('plan_review', content='%d of %d tasks approved' % (len(kept), len(items)))
|
|
183
261
|
block('The user approved only part of the task list.\nApproved: ' + '; '.join(kept) + '\nRemoved (keep these out of scope; do not do them): ' + '; '.join(removed) + '\nCall TodoList again with only the approved items, then build.')
|
|
184
262
|
event('plan_review', content='Plan dismissed')
|
|
185
263
|
st['plan_dismissed'] = True; save_state(st)
|
|
186
|
-
block(
|
|
264
|
+
block(DISMISSED_BLOCK)
|
|
187
265
|
|
|
188
266
|
def resolve(path):
|
|
189
267
|
p = Path(path)
|
|
@@ -291,6 +369,19 @@ def secret_gate(name, args, rel, pol):
|
|
|
291
369
|
'Read it from an environment variable instead (for example process.env.' + f['name'] + ' or os.environ["' + f['name'] + '"]), '
|
|
292
370
|
'add the name to .env.example with an empty value, and tell the user to store the real value in the Vault as ' + f['name'] + '. Re-issue the edit without the literal value.')
|
|
293
371
|
|
|
372
|
+
# BUG-0224: installed dependencies are not the project's code. Editing them in place (often to get past a check) is
|
|
373
|
+
# refused unless the task itself asks for it; the change would also vanish on the next install.
|
|
374
|
+
DEP_DIRS = ('node_modules', 'site-packages', 'dist-packages', 'bower_components')
|
|
375
|
+
DEP_WRITE = re.compile(r'(?:\bsed\s+(?:-\w+\s+)*-i|\bperl\s+-\w*i|>>?|\btee\b|\bpatch\b|write_text|writeFileSync)[^|;&\n]*\b(?:node_modules|site-packages)/')
|
|
376
|
+
def dependency_gate(what, rel, st):
|
|
377
|
+
parts = rel.replace('\\', '/').split('/')
|
|
378
|
+
if not any(p in DEP_DIRS for p in parts): return
|
|
379
|
+
if re.search(r'node_modules|site-packages|patch-package|vendored? (?:dependency|library)', str(st.get('task', '')), re.I): return
|
|
380
|
+
event('safety', effect='write', decision='deny', command=what[:300], content='Refused: installed dependencies are not edited in place')
|
|
381
|
+
block('Refused: this changes an installed dependency (' + rel[:120] + '), which is outside the project and is lost on the next install. '
|
|
382
|
+
'Do not edit dependencies to get past a check. Fix the project code or its configuration, or explain the problem and ask the user. '
|
|
383
|
+
'This step was not run.')
|
|
384
|
+
|
|
294
385
|
def protected_path_gate(name, args, rel, pol, st):
|
|
295
386
|
"""Paths the policy marks for review always ask, even in Auto mode (once per file per task)."""
|
|
296
387
|
if pol is None or rel.startswith('..'): return
|
|
@@ -324,11 +415,73 @@ def command_policy_gate(command, pol, st, mode, auto):
|
|
|
324
415
|
policy_event(pat, 'allow', 'Allowed by policy: ' + command[:300], command=command[:300]); return True
|
|
325
416
|
return False
|
|
326
417
|
|
|
418
|
+
BOM = b'\xef\xbb\xbf'
|
|
419
|
+
|
|
420
|
+
def remember_bom(name, args):
|
|
421
|
+
"""BUG-0546: the file tools drop a UTF-8 byte order mark; note which edited files had one so PostToolUse puts it back."""
|
|
422
|
+
if name not in ('Write', 'Edit') or not args.get('path'): return
|
|
423
|
+
try:
|
|
424
|
+
target = resolve(str(args['path']))
|
|
425
|
+
with open(target, 'rb') as f: had = f.read(3) == BOM
|
|
426
|
+
except OSError: return
|
|
427
|
+
if had:
|
|
428
|
+
known = load('kimi-bom.json', [])
|
|
429
|
+
if str(target) not in known: save('kimi-bom.json', (known + [str(target)])[-200:])
|
|
430
|
+
|
|
431
|
+
def restore_bom():
|
|
432
|
+
"""Runs before the next tool call and at Stop (PostToolUse is only registered for project hooks), so it never races an edit."""
|
|
433
|
+
for path in load('kimi-bom.json', []):
|
|
434
|
+
try:
|
|
435
|
+
target = Path(path)
|
|
436
|
+
data = target.read_bytes()
|
|
437
|
+
if data and not data.startswith(BOM): target.write_bytes(BOM + data)
|
|
438
|
+
except OSError: pass
|
|
439
|
+
|
|
440
|
+
def step_limit_gate(st):
|
|
441
|
+
"""BUG-0589: Settings > Agent > Maximum steps per task. Each command, edit or app call is one step; at the limit
|
|
442
|
+
the agent stops working and reports instead of running more (the model loop alone never went below 10)."""
|
|
443
|
+
try: limit = int(st.get('step_limit') or 0)
|
|
444
|
+
except (TypeError, ValueError): limit = 0
|
|
445
|
+
if limit <= 0: return
|
|
446
|
+
used = int(st.get('steps_used') or 0)
|
|
447
|
+
if used >= limit:
|
|
448
|
+
if not st.get('step_limit_noted'):
|
|
449
|
+
st['step_limit_noted'] = True; save_state(st)
|
|
450
|
+
event('note', content='Step limit reached (%d). Asked the agent to wrap up' % limit)
|
|
451
|
+
block('Step limit reached: the user allows %d step%s per task (Settings > Agent > Maximum steps per task), and all are used. '
|
|
452
|
+
'This call was not run. Do not call more tools: reply now with what is done, what is left, and that they can raise the limit or send a follow-up.' % (limit, '' if limit == 1 else 's'))
|
|
453
|
+
st['steps_used'] = used + 1; save_state(st)
|
|
454
|
+
|
|
455
|
+
def question_only(st):
|
|
456
|
+
"""BUG-0345: this turn answers a question and has not tried to change anything yet."""
|
|
457
|
+
return bool(st.get('question_turn') and not st.get('question_changes'))
|
|
458
|
+
|
|
459
|
+
def plan_first_due(st):
|
|
460
|
+
"""The task list must be approved before the first change: Guided (plan_first), or Always in Auto too (BUG-0590)."""
|
|
461
|
+
mode = st.get('mode', 'approve')
|
|
462
|
+
return bool(st.get('plan_first') and (mode == 'approve' or (mode == 'auto' and st.get('plan_approval') == 'always'))
|
|
463
|
+
and not st.get('planned') and not st.get('plan_reviewed') and not st.get('readonly') and not (ROOT / 'kimi-approve-all').exists())
|
|
464
|
+
|
|
465
|
+
def always_plan_gate(command, st):
|
|
466
|
+
"""With "Approve the task list first: Always", a command that changes files waits for the plan in Auto, like Write and Edit
|
|
467
|
+
(in Guided the command approval already asks). Reads and checks still run (BUG-0590)."""
|
|
468
|
+
if st.get('mode') != 'auto' or st.get('plan_approval') != 'always' or not plan_first_due(st): return
|
|
469
|
+
from command_safety import classify
|
|
470
|
+
if classify(command, REPO, cwd=REPO, home=os.path.dirname(REPO)).get('effect') in ('read', 'check'): return
|
|
471
|
+
event('note', content='Approve the task list first: asked for the plan before the first change')
|
|
472
|
+
block('The user approves your task list before any change, even in Auto. Call TodoList with 2 to 6 short, concrete items '
|
|
473
|
+
'(what you will change and how you will check it), then wait for the review. This command was not run; re-issue it after the plan is approved.')
|
|
474
|
+
|
|
327
475
|
def pre(payload, st):
|
|
328
476
|
name = str(payload.get('tool_name') or '')
|
|
329
477
|
args = payload.get('tool_input') or {}
|
|
330
478
|
if not isinstance(args, dict): args = {}
|
|
479
|
+
try: restore_bom(); remember_bom(name, args)
|
|
480
|
+
except Exception: pass
|
|
331
481
|
if (ROOT / 'cancel').exists(): block('Task stopped by user.')
|
|
482
|
+
if st.get('plan_dismissed') and is_work_tool(name): # before any approval card: Not now means nothing else asks or runs
|
|
483
|
+
event('plan_review', content='Plan on hold: ' + re.sub(r'^mcp__\w+?__', '', name) + ' was not run')
|
|
484
|
+
block(DISMISSED_BLOCK)
|
|
332
485
|
hooked = user_hooks('kimi_pre', name, args, st) # hard-deny, credential block, before_tool / before_command hooks
|
|
333
486
|
if hooked: block(hooked)
|
|
334
487
|
queued = steering()
|
|
@@ -350,16 +503,20 @@ def pre(payload, st):
|
|
|
350
503
|
block('You have made this exact call three times recently. Repeating it will not give new information. Step back: re-read the latest error, form one hypothesis about the root cause, and take a different action.')
|
|
351
504
|
save_state(st)
|
|
352
505
|
if ACCEPTANCE in json.dumps(args) or HIDDEN.search(json.dumps(args)): block('That path holds the independent acceptance check, which is hidden from the builder by design. Work from the task and the failure output instead.')
|
|
506
|
+
if is_work_tool(name): step_limit_gate(st)
|
|
353
507
|
if name == 'TodoList': return plan_gate(args, st)
|
|
354
|
-
|
|
355
|
-
|
|
356
|
-
if briefing(st) and name in ('Write', 'Edit', 'Agent', 'AgentSwarm'):
|
|
357
|
-
block('Brief mode: nothing is built until the user approves your task list. Read the repository, then call TodoList with the proposed items and state your open questions. This call was not run.')
|
|
508
|
+
if briefing(st) and (name in ('Write', 'Edit', 'Agent', 'AgentSwarm') or (name.startswith('mcp__') and is_work_tool(name))):
|
|
509
|
+
block(brief_label(st) + ': nothing is built until the user approves your task list. Read the repository, then call TodoList with the proposed items and state your open questions. This call was not run.')
|
|
358
510
|
if st.get('inspect') and name in ('Write', 'Edit', 'Agent', 'AgentSwarm'):
|
|
359
511
|
event('safety', effect='write', decision='deny', command=name + ' ' + str(args.get('path', ''))[:300], content='Refused: the user asked for no changes in this turn')
|
|
360
512
|
block(INSPECT_BLOCK)
|
|
361
513
|
if name in ('Write', 'Edit') and not st.get('readonly'):
|
|
362
514
|
nudged = set(st.get('nudged') or [])
|
|
515
|
+
if (plan_first_due(st) and ('guided_plan' not in nudged or st.get('plan_approval') == 'always')):
|
|
516
|
+
nudged.add('guided_plan'); st['nudged'] = sorted(nudged); st['question_changes'] = True; save_state(st)
|
|
517
|
+
event('note', content='Guided mode: asked for the plan before the first edit')
|
|
518
|
+
block(('The user approves your task list before any change, even in Auto.' if st.get('mode') == 'auto' else 'Guided mode: the user approves your plan before any file changes.') + ' Call TodoList with 2 to 6 short, concrete items '
|
|
519
|
+
'(what you will change and how you will check it), then wait for the review. This edit was not applied; re-issue it after the plan is approved.')
|
|
363
520
|
if st.get('large') and not st.get('planned') and 'plan' not in nudged:
|
|
364
521
|
nudged.add('plan'); st['nudged'] = sorted(nudged); save_state(st)
|
|
365
522
|
event('note', content='Large task: asked for a plan before editing')
|
|
@@ -386,6 +543,7 @@ def pre(payload, st):
|
|
|
386
543
|
helper_gate(name, args, st)
|
|
387
544
|
if name == 'Bash':
|
|
388
545
|
command = str(args.get('command', ''))
|
|
546
|
+
always_plan_gate(command, st)
|
|
389
547
|
try:
|
|
390
548
|
from command_policy import offline_network_wait_reason
|
|
391
549
|
reason = offline_network_wait_reason(command, st.get('task', ''))
|
|
@@ -393,6 +551,7 @@ def pre(payload, st):
|
|
|
393
551
|
if reason: event('retry', content=reason); block(reason)
|
|
394
552
|
masked = pipe_gate(command)
|
|
395
553
|
if masked: event('note', content='Check would report a later command\'s exit status; asked for pipefail'); block(masked)
|
|
554
|
+
if DEP_WRITE.search(command): dependency_gate(command, 'node_modules', st)
|
|
396
555
|
if command_policy_gate(command, pol, st, mode, auto): return
|
|
397
556
|
return command_gate(command, st, mode, auto)
|
|
398
557
|
if name in ('Write', 'Edit'):
|
|
@@ -406,6 +565,10 @@ def pre(payload, st):
|
|
|
406
565
|
old, new = str(args.get('old_string', '')), str(args.get('new_string', ''))
|
|
407
566
|
if old and old in original: after = original.replace(old, new) if args.get('replace_all') else original.replace(old, new, 1)
|
|
408
567
|
rel = os.path.relpath(str(target), REPO)
|
|
568
|
+
if re.search(r'(^|/)\.git/hooks(/|$)', rel.replace(os.sep, '/')): # BUG-0551: the team's git hooks are not the agent's to change
|
|
569
|
+
from mlai_hooks import GIT_HOOK_REASON
|
|
570
|
+
event('safety', effect='write', decision='deny', command=name + ' ' + rel[:300], content='Refused: editing a git hook')
|
|
571
|
+
block(GIT_HOOK_REASON)
|
|
409
572
|
if not rel.startswith('..'): secret_gate(name, args, rel, pol)
|
|
410
573
|
if after is not None:
|
|
411
574
|
try:
|
|
@@ -413,6 +576,7 @@ def pre(payload, st):
|
|
|
413
576
|
problem = gate(rel, original, after, REPO) if not rel.startswith('..') else None
|
|
414
577
|
except Exception: problem = None
|
|
415
578
|
if problem: event('edit_guard', path=str(args.get('path', ''))[:200], content=problem[:600]); block(problem)
|
|
579
|
+
dependency_gate(name + ' ' + rel, rel, st)
|
|
416
580
|
protected_path_gate(name, args, rel, pol, st)
|
|
417
581
|
return edit_gate(name, args, rel, str(target), st, auto)
|
|
418
582
|
if name in ('Agent', 'AgentSwarm') and not auto:
|
|
@@ -491,6 +655,13 @@ def edit_gate(name, args, rel, target, st, auto):
|
|
|
491
655
|
wait_decision({'id': os.urandom(8).hex(), 'kind': 'scope', 'area': area(rel), 'command': summary})
|
|
492
656
|
grant_area(st, rel)
|
|
493
657
|
|
|
658
|
+
def chat_allowed(command, effect):
|
|
659
|
+
"""BUG-0594: the person chose "Always allow in this chat" for this exact command (safe kinds only, re-checked)."""
|
|
660
|
+
try:
|
|
661
|
+
from command_safety import chat_allow_ok, chat_allow_key
|
|
662
|
+
return chat_allow_key(command) in load('chat-allow.json', []) and chat_allow_ok(effect, command, REPO)
|
|
663
|
+
except Exception: return False
|
|
664
|
+
|
|
494
665
|
def command_gate(command, st, mode, auto):
|
|
495
666
|
"""Classify the command; allow, ask or refuse according to the mode budget and the approved scope."""
|
|
496
667
|
from command_safety import classify, decide, describe, relative
|
|
@@ -504,15 +675,20 @@ def command_gate(command, st, mode, auto):
|
|
|
504
675
|
if decision['decision'] == 'deny':
|
|
505
676
|
event('safety', effect=effect, decision='deny', command=command[:2000], content='Refused. ' + (decision.get('reason') or describe(decision)))
|
|
506
677
|
if effect == 'catastrophic': block('Refused: ' + (decision.get('reason') or describe(decision)) + '. This command is never run, in any mode. Choose a contained alternative (delete specific files inside the project, keep the harness state alone, never touch devices or system directories).')
|
|
507
|
-
if briefing(st): block('
|
|
678
|
+
if briefing(st): block(brief_label(st) + ': only read-only commands run before the user approves the task list. ' + describe(decision) + ' was not run; put it in the plan instead.')
|
|
508
679
|
if st.get('inspect'): block('Not allowed: ' + describe(decision) + '. ' + INSPECT_BLOCK)
|
|
509
680
|
block('Not allowed in this mode: ' + describe(decision) + '. Plan mode only reads; propose the command in your answer instead.')
|
|
510
681
|
if decision['decision'] == 'allow':
|
|
511
682
|
if effect not in ('read', 'check'): event('safety', effect=effect, decision='allow', command=command[:300], content=describe(decision))
|
|
512
683
|
return
|
|
684
|
+
if effect not in ('destructive', 'privilege') and chat_allowed(command, effect):
|
|
685
|
+
event('safety', effect=effect, decision='allow', command=command[:300], content='Always allowed in this chat')
|
|
686
|
+
return
|
|
513
687
|
request = {'id': os.urandom(8).hex(), 'kind': 'command', 'effect': effect, 'reason': decision.get('reason') or '', 'command': command[:20000]}
|
|
688
|
+
# Deletes and sudo ask every time, even after Approve all (BUG-0101). The card says so and offers no Approve all for them.
|
|
689
|
+
if effect in ('destructive', 'privilege'): request['strict'] = True
|
|
514
690
|
event('safety', effect=effect, decision='ask', command=command[:300], content=describe(decision))
|
|
515
|
-
wait_decision(request)
|
|
691
|
+
wait_decision(request, strict=effect in ('destructive', 'privilege')) # "Approve all" never covers deletes or sudo (BUG-0101)
|
|
516
692
|
if effect in ('install', 'service'):
|
|
517
693
|
kinds = sorted(set(st.get('approved_kinds') or []) | {effect}); st['approved_kinds'] = kinds; save_state(st)
|
|
518
694
|
elif effect in ('write', 'exec') and not st.get('readonly'):
|
|
@@ -574,16 +750,23 @@ def readonly_stop(st):
|
|
|
574
750
|
event('done_gate', content='Research claims not recorded yet')
|
|
575
751
|
block('Before answering, record the material claims with mcp__mlai__evidence {claims:[{claim, kind, source_ids, quote, uncertainty}]}. Facts need an exact quote from a retrieved source; label inferences and assumptions. Link the source URLs in the answer.')
|
|
576
752
|
|
|
753
|
+
def _laptop_run():
|
|
754
|
+
"""BUG-0531: the run is on the person's own computer (CLI or Mac app)."""
|
|
755
|
+
if os.getenv('MINI_LAPTOP') == '1': return True
|
|
756
|
+
try: return json.loads((ROOT / 'request.json').read_text()).get('laptop') is True
|
|
757
|
+
except (OSError, ValueError, AttributeError): return False
|
|
758
|
+
|
|
577
759
|
def stop(payload, st):
|
|
578
760
|
if (ROOT / 'cancel').exists(): return
|
|
579
761
|
if st.get('account_stop'): return # out of credits: the summary is the last turn, no more gates
|
|
580
762
|
queued = steering()
|
|
581
763
|
if queued: steer_block(queued, stopping=True)
|
|
582
764
|
if st.get('readonly') or st.get('inspect'): return readonly_stop(st)
|
|
765
|
+
if st.get('plan_dismissed'): return # Not now: the run ends here; no done gate may send the agent back to work (BUG-0102)
|
|
583
766
|
if briefing(st) and not st.get('planned') and not st.get('brief_nudged'):
|
|
584
767
|
st['brief_nudged'] = True; save_state(st)
|
|
585
|
-
event('done_gate', content='
|
|
586
|
-
block('
|
|
768
|
+
event('done_gate', content=brief_label(st) + ': asked for the task list before finishing')
|
|
769
|
+
block(brief_label(st) + ': the user is waiting to review a task list. Call TodoList with 3 to 10 concrete items (and mcp__mlai__brief_questions for anything open) before finishing. If the request is already answered and nothing needs building, say so plainly and finish.')
|
|
587
770
|
if briefing(st): return
|
|
588
771
|
count = int(load('kimi-stop-count', 0) or 0)
|
|
589
772
|
if count >= 3 or time.time() > st.get('deadline', 1e18) - 90: return
|
|
@@ -592,6 +775,10 @@ def stop(payload, st):
|
|
|
592
775
|
import roles as role_engine
|
|
593
776
|
try: changed = role_engine.changed_files(REPO, st.get('base'))
|
|
594
777
|
except Exception: changed = []
|
|
778
|
+
try:
|
|
779
|
+
from runtime_control import turn_changes
|
|
780
|
+
changed = turn_changes(ROOT, REPO, changed) # BUG-0495: not files that were already untracked or changed at the start
|
|
781
|
+
except Exception: pass
|
|
595
782
|
code_changed = [p for p in changed if role_engine.CODE_EXT.search(p)]
|
|
596
783
|
if st.get('require_changes') and not changed and 'patch' not in done:
|
|
597
784
|
done.add('patch'); issues.append('This task requires a repository change, but the working tree is unchanged. Make the fix, or explain the blocker if it truly cannot be fixed.')
|
|
@@ -616,9 +803,12 @@ def stop(payload, st):
|
|
|
616
803
|
if blocking: done.add('verification'); issues += blocking[:3]
|
|
617
804
|
except Exception: pass
|
|
618
805
|
role = st.get('role')
|
|
806
|
+
if role and 'role' not in done and (ROOT / 'last-reject').exists() and float((ROOT / 'last-reject').read_text().strip() or 0) >= float(st.get('started') or 0):
|
|
807
|
+
done.add('role') # BUG-0529: after a denial the done-check must not push unrequested work; the agent stops and asks
|
|
808
|
+
event('role_check', role=role['id'], title=role['title'], status='skipped', issues=[])
|
|
619
809
|
if role and 'role' not in done:
|
|
620
810
|
done.add('role')
|
|
621
|
-
try: gaps = role_engine.gate(role, REPO, st.get('base'), changed, st.get('task', ''))
|
|
811
|
+
try: gaps = role_engine.gate(role, REPO, st.get('base'), changed, st.get('task', ''), laptop=_laptop_run())
|
|
622
812
|
except Exception: gaps = []
|
|
623
813
|
event('role_check', role=role['id'], title=role['title'], status='blocked' if gaps else 'passed', issues=[{'kind': i['kind'], 'label': i['label'], 'message': i['message'][:600]} for i in gaps])
|
|
624
814
|
if gaps: issues.append('Role done-check (' + role['title'] + '): ' + ' '.join(i['message'] for i in gaps) + ' Fix these for real (real data paths, real artifacts, verified). If one truly cannot apply, say why in the final summary.')
|
|
@@ -654,9 +844,13 @@ def main():
|
|
|
654
844
|
st = state()
|
|
655
845
|
if kind == 'PreToolUse': pre(payload, st)
|
|
656
846
|
elif kind == 'PostToolUse':
|
|
847
|
+
try: restore_bom()
|
|
848
|
+
except Exception: pass
|
|
657
849
|
feedback = user_hooks('kimi_post', payload, st) # after_tool / after_edit hooks
|
|
658
850
|
if feedback: block(feedback)
|
|
659
851
|
elif kind == 'Stop':
|
|
852
|
+
try: restore_bom()
|
|
853
|
+
except Exception: pass
|
|
660
854
|
stop(payload, st)
|
|
661
855
|
# Turn accepted: tell the bridge, which ends the engine if it lingers waiting on background processes (preview servers).
|
|
662
856
|
save('kimi-stopped', {'time': time.time()})
|