mlai-code 0.3.1 → 0.3.3

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.
Files changed (56) hide show
  1. package/README.md +1 -1
  2. package/package.json +2 -2
  3. package/src/cli.mjs +27 -7
  4. package/src/distribution.mjs +5 -5
  5. package/src/engine-key.mjs +18 -4
  6. package/src/engine.mjs +4 -9
  7. package/src/home.mjs +1 -1
  8. package/src/parity.mjs +3 -1
  9. package/src/render.mjs +3 -1
  10. package/src/runtime.mjs +21 -1
  11. package/src/sandbox.mjs +19 -3
  12. package/src/session.mjs +22 -2
  13. package/src/shared-data.mjs +1 -1
  14. package/src/tui/app.mjs +22 -4
  15. package/src/tui/chat.mjs +8 -6
  16. package/src/tui/cmd-core.mjs +2 -2
  17. package/src/tui/cmd-extensibility.mjs +8 -2
  18. package/src/tui/commands.mjs +2 -1
  19. package/src/tui/statusline.mjs +2 -1
  20. package/src/wakeups.mjs +2 -2
  21. package/vendor/client-sdk/index.cjs +3 -3
  22. package/vendor/cloud-cli/mlai.mjs +169 -25
  23. package/worker/build_planner.py +75 -7
  24. package/worker/cloud_pty.py +91 -2
  25. package/worker/command_safety.py +25 -1
  26. package/worker/compaction_brief.py +186 -0
  27. package/worker/context_manager.py +9 -3
  28. package/worker/conversation.py +58 -1
  29. package/worker/deck_render.js +11 -0
  30. package/worker/deck_tools.py +825 -0
  31. package/worker/git_sync.py +38 -13
  32. package/worker/goals.py +4 -0
  33. package/worker/image_generation.py +102 -23
  34. package/worker/kimi_bridge.py +406 -40
  35. package/worker/kimi_gates.py +84 -1
  36. package/worker/kimi_hook.py +113 -12
  37. package/worker/media_tools.py +179 -0
  38. package/worker/memory.py +2 -2
  39. package/worker/mlai_hooks.py +93 -3
  40. package/worker/mlai_mcp.py +23 -5
  41. package/worker/model_stream.py +21 -7
  42. package/worker/obs_mask.py +97 -0
  43. package/worker/offline_guard.py +1 -1
  44. package/worker/preview.py +47 -4
  45. package/worker/product_facts.py +18 -0
  46. package/worker/project_context.py +33 -7
  47. package/worker/reply_language.py +92 -10
  48. package/worker/roles.json +199 -37
  49. package/worker/roles.py +243 -8
  50. package/worker/runner.py +253 -42
  51. package/worker/site_check.py +279 -0
  52. package/worker/skill_library.py +15 -3
  53. package/worker/skill_pack.py +170 -16
  54. package/worker/tool_catalog.py +3 -2
  55. package/worker/verification.py +43 -11
  56. package/worker/wakeups.py +56 -4
@@ -14,7 +14,7 @@ import json, os, re, subprocess, sys, time
14
14
  from pathlib import Path
15
15
 
16
16
  ACCEPTANCE_DIR = 'acceptance'
17
- TRIM_LIMIT = int(os.getenv('MINI_KIMI_TRIM_CHARS', '16000'))
17
+ TRIM_LIMIT = int(os.getenv('MINI_KIMI_TRIM_CHARS', '12000'))
18
18
  ERROR_LINE = re.compile(r'error|fail|exception|traceback|assert|panic|cannot|denied|not found|undefined|expected', re.I)
19
19
 
20
20
  # ---------------------------------------------------------------- output trimming
@@ -115,6 +115,13 @@ def tester_invalid(language, output):
115
115
  if re.search(r'AssertionError|ERR_ASSERTION', output): return False
116
116
  return bool(re.search(r'ERR_MODULE_NOT_FOUND|SyntaxError|Cannot find module|ReferenceError|does not provide an export', output)) or '# fail 0' in output
117
117
 
118
+ TESTER_RETRY = '''
119
+
120
+ This is a second try. Your previous test could not run; the error is in `error`. Fix the test itself (imports, paths,
121
+ module format, a missing loader) and return the same JSON shape. Test through files and commands that run in plain
122
+ Node or Python here: for a React or other bundled UI, test the server, data and pure modules, or read the built output,
123
+ instead of importing JSX. If no test can run in this repository, return {"applicable": false, "reason": "..."}.'''
124
+
118
125
  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 asserts something the task does not require, or if its expected value is miscalculated.
119
126
 
120
127
  Reply with JSON only: {"genuine": ["test_name", ...], "wrong": [{"test": "test_name", "why": "..."}]}'''
@@ -172,6 +179,24 @@ def acceptance(config, st, model, root, repo, changed, event):
172
179
  event('acceptance', status='skipped', content='Acceptance check not applicable: ' + str(reason)[:300])
173
180
  return {'status': 'skipped', 'reason': str(reason)[:300]}
174
181
  code, output = run_acceptance(root, repo, spec['language'], spec['code'])
182
+ if code not in (0, 124) and tester_invalid(spec['language'], output):
183
+ # Web bug list Bug-017 and Bug-021: a test that cannot even load (JSX imported into plain Node, a wrong path)
184
+ # says nothing about the build. Give the tester its own error once and run the corrected test.
185
+ retry = json.dumps({'task': str(config.get('task', ''))[:12000], 'languages_changed': sorted(langs), 'repository_files': files,
186
+ 'previous_test': spec['code'][:20000], 'error': output[-3000:]})
187
+ try:
188
+ fixed_text, _ = model.request([{'role': 'system', 'content': TESTER_SYSTEM + TESTER_RETRY},
189
+ {'role': 'user', 'content': retry}], max_tokens=12000, allow_tools=False)
190
+ fixed = _parse(fixed_text)
191
+ except Exception:
192
+ fixed = None
193
+ if fixed and fixed.get('applicable') and fixed.get('language') == spec['language'] and isinstance(fixed.get('code'), str) and len(fixed['code']) <= 40000:
194
+ spec = {**spec, **fixed}
195
+ code, output = run_acceptance(root, repo, spec['language'], spec['code'])
196
+ elif isinstance(fixed, dict) and fixed.get('applicable') is False:
197
+ reason = str(fixed.get('reason') or 'the tester could not write a test that runs here')[:300]
198
+ event('acceptance', status='skipped', content='Acceptance check not applicable: ' + reason)
199
+ return {'status': 'skipped', 'reason': reason}
175
200
  covers = [str(c)[:200] for c in spec.get('covers') or []][:8]
176
201
  if code == 0: status = 'passed'
177
202
  elif code != 124 and tester_invalid(spec['language'], output): status = 'invalid'
@@ -215,6 +240,35 @@ def pre_existing_failure(st, checks, repo, done, event):
215
240
  return True
216
241
  return False
217
242
 
243
+ # Off by default: every code change gets the independent acceptance check. MINI_SMALL_CHANGE_LINES=N skips it for
244
+ # one-file edits of at most N lines that already have a passing check after the last edit.
245
+ SMALL_CHANGE_LINES = int(os.getenv('MINI_SMALL_CHANGE_LINES', '0') or 0)
246
+
247
+ def changed_lines(repo, base, paths):
248
+ """Added + removed lines in tracked files plus the length of new untracked files."""
249
+ total = 0
250
+ try:
251
+ out = subprocess.run(['git', 'diff', '--numstat', base or 'HEAD', '--', *paths], cwd=repo, capture_output=True, text=True, timeout=15)
252
+ seen = set()
253
+ for line in out.stdout.splitlines():
254
+ parts = line.split('\t')
255
+ if len(parts) == 3:
256
+ seen.add(parts[2]); total += sum(int(x) if x.isdigit() else 200 for x in parts[:2])
257
+ for p in paths:
258
+ if p not in seen:
259
+ try: total += len((Path(repo) / p).read_text(errors='replace').splitlines())
260
+ except OSError: pass
261
+ except (OSError, subprocess.SubprocessError, ValueError): return 10 ** 6
262
+ return total
263
+
264
+ def small_change(repo, base, code_changed, checks, last_edit, last_check):
265
+ """One code file, at most SMALL_CHANGE_LINES changed lines, and a passing check that ran after the last edit."""
266
+ if SMALL_CHANGE_LINES <= 0 or len(code_changed) != 1 or not last_edit or last_check < last_edit: return False
267
+ try: from runtime_control import check_status
268
+ except Exception: return False
269
+ if not any(isinstance(c, dict) and check_status(c) == 'passed' for c in checks or []): return False
270
+ return changed_lines(repo, base, code_changed) <= SMALL_CHANGE_LINES
271
+
218
272
  def final_issues(config, st, answer, done, model, root, repo, event, allow_acceptance=True):
219
273
  """Blockers persist until repaired; expensive new acceptance tests are written once."""
220
274
  import roles as role_engine
@@ -262,6 +316,30 @@ def final_issues(config, st, answer, done, model, root, repo, event, allow_accep
262
316
  except Exception as exc:
263
317
  issue = 'Verification report could not complete: ' + str(exc)[:200]
264
318
  event('note', content=issue); issues.append(issue)
319
+ if allow_acceptance and changed and not st.get('inspect'):
320
+ # The running preview is the user's view of the work: a page that answers 500, or a linked page that 404s, is not
321
+ # done, whatever the final message says. Nothing listening on the preview port means there is nothing to judge.
322
+ try:
323
+ import site_check
324
+ routes = sorted(site_check.internal_links(repo, changed))
325
+ bad = site_check.live_routes(routes)
326
+ found = site_check.live_findings(bad)
327
+ if found:
328
+ event('verification_report', status='failed', reasons=[found[0]['message'].split('.')[0]], findings=found, checks=0, passed=0)
329
+ issues += [f['message'] for f in found if f['message'] not in issues]
330
+ except Exception as exc:
331
+ event('note', content='Could not check the preview routes: ' + str(exc)[:200])
332
+ if allow_acceptance and config.get('conversation_route') in (None, 'build'):
333
+ # A run must not end on "I'll now make the slides", and a presentation or PDF request must leave a file behind.
334
+ # Each note is sent once; if the agent still cannot do it, its own explanation stands.
335
+ try:
336
+ from conversation import deliverable_issue, stalled_promise, STALLED_NOTE
337
+ if 'stalled' not in done and not changed and stalled_promise(answer):
338
+ done.add('stalled'); issues.append(STALLED_NOTE)
339
+ elif 'deliverable' not in done:
340
+ note = deliverable_issue(config.get('task', ''), changed, repo)
341
+ if note: done.add('deliverable'); issues.append(note)
342
+ except Exception: pass # these hints never block a run
265
343
  if config.get('selected_design'):
266
344
  from design_director import browser_review
267
345
  from quality import fingerprint
@@ -277,6 +355,11 @@ def final_issues(config, st, answer, done, model, root, repo, event, allow_accep
277
355
  if research_required(config) or ledger.load()['sources']:
278
356
  issue = ledger.completion_issue(answer or '')
279
357
  if issue: issues.append(issue.replace('evidence {claims', 'mcp__mlai__evidence {claims'))
358
+ if allow_acceptance and code_changed and not issues and 'acceptance' not in done and small_change(repo, st.get('base'), code_changed, checks, last_edit, last_check):
359
+ # A small edit with a passing check after it: writing and running a hidden test costs a model call and a test run
360
+ # for little extra signal. Larger changes still get the independent check.
361
+ done.add('acceptance'); st['_acceptance_skipped'] = True
362
+ event('acceptance', status='skipped', content='Independent acceptance check skipped: small change with a passing check after the last edit')
280
363
  if allow_acceptance and code_changed and not issues:
281
364
  previous = st.get('_acceptance')
282
365
  if 'acceptance' not in done:
@@ -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')
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', 'plan_items', 'brief_q_nudged', 'plan_nudged')
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: msg = json.loads(f.read_text()).get('message', '')
51
- except (OSError, ValueError): msg = ''
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(): out.append(msg.strip()[:5000])
54
- for m in out: event('steering', content=m)
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():
@@ -132,6 +151,17 @@ def brief_unlock(st):
132
151
  st['brief_ok'] = True; st['mode'] = back; st['scope_ok'] = True; save_state(st)
133
152
  event('mode', mode=back, content='Plan agreed. Building the approved list in %s mode.' % ('Auto' if back == 'auto' else 'Guided'))
134
153
 
154
+ def plan_key(text):
155
+ return ' '.join(str(text).lower().split())[:300]
156
+
157
+ def plan_added(items, st):
158
+ """Items in this task list that the user has not approved yet. Status updates of the same list add nothing."""
159
+ known = set(st.get('plan_items') or [])
160
+ return [i for i in items if plan_key(i) not in known]
161
+
162
+ def remember_plan(items, st):
163
+ st['plan_items'] = sorted(set(st.get('plan_items') or []) | {plan_key(i) for i in items})[:200]
164
+
135
165
  def plan_gate(args, st):
136
166
  """First task list of a run: optional user review (Guided mode by default, or always when the user asked).
137
167
  Brief mode: every task list is reviewed until the user approves; replies come back to the agent as feedback."""
@@ -144,7 +174,18 @@ def plan_gate(args, st):
144
174
  wants = brief or policy == 'always' or (policy == 'guided' and st.get('mode') == 'approve')
145
175
  if brief:
146
176
  if len(items) < 2: block('Brief mode: the task list needs at least two concrete items before the user can review it. Call TodoList again with the full plan.')
147
- elif not first or st.get('readonly') or st.get('plan_reviewed') or len(items) < 2 or not wants or (ROOT / 'kimi-approve-all').exists(): return
177
+ 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', []):
178
+ # BUG-0124: Brief mode promises to clarify requirements before the task list. Ask once for the open questions.
179
+ st['brief_q_nudged'] = True; st['planned'] = False; save_state(st)
180
+ event('note', content='Brief mode: asked for clarifying questions before the task list')
181
+ block('Brief mode clarifies requirements first. Before this task list goes to the user, record one to three short questions about '
182
+ 'what is still open in the request (for example style, content, images or scope) with mcp__mlai__brief_questions '
183
+ '{"questions": [{"question": "...", "options": [{"label": "...", "detail": "..."}], "recommended": 0}]}, each with your '
184
+ 'recommended answer. Then call TodoList again with the same list. If the request truly leaves nothing open, call TodoList again unchanged.')
185
+ elif st.get('readonly') or len(items) < 2 or not wants or (ROOT / 'kimi-approve-all').exists(): return
186
+ elif not first or st.get('plan_reviewed'):
187
+ # Guided: an approved list is not asked again, but a later list that adds work asks again (BUG-0039).
188
+ if not st.get('plan_reviewed') or not plan_added(items, st): return
148
189
  st['plan_reviewed'] = True
149
190
  rounds = int(st.get('brief_rounds') or 0) + 1
150
191
  if brief: st['brief_rounds'] = rounds
@@ -173,17 +214,19 @@ def plan_gate(args, st):
173
214
  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
215
  if verdict in ('approve', 'approve_all'):
175
216
  if brief: brief_unlock(st)
217
+ remember_plan(items, st)
176
218
  st['scope_ok'] = True; save_state(st); event('plan_review', content='Plan approved'); return
177
219
  if verdict.startswith('approve:'):
178
220
  keep = {int(x) for x in verdict[8:].split(',') if x.strip().isdigit()}
179
221
  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
222
  if kept:
181
223
  if brief: brief_unlock(st)
224
+ remember_plan(kept, st); save_state(st)
182
225
  event('plan_review', content='%d of %d tasks approved' % (len(kept), len(items)))
183
226
  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
227
  event('plan_review', content='Plan dismissed')
185
228
  st['plan_dismissed'] = True; save_state(st)
186
- block('The user dismissed this task list. Do not start the work. Briefly ask what they want changed, then finish.')
229
+ block(DISMISSED_BLOCK)
187
230
 
188
231
  def resolve(path):
189
232
  p = Path(path)
@@ -214,6 +257,52 @@ def account_gate(st, name):
214
257
  block('Your account has used all its credits. Do not start new work or call more tools. Write the final summary now in plain text: what changed, '
215
258
  'what was verified, what is left. The user can continue after picking a plan or buying a top-up. This call was not run.')
216
259
 
260
+ # Helper agents re-read the repository from scratch: on small tasks they multiplied tokens (one trivial test started
261
+ # four) without finishing sooner. Unless the user asked for them, small tasks get none and large tasks at most three.
262
+ HELPER_ASK = re.compile(r'\b(sub-?agents?|helpers?|in parallel|parallel agents?|delegat\w*|swarm|agents)\b', re.I)
263
+ HELPER_CAP_LARGE = int(os.getenv('MINI_HELPER_CAP', '3') or 0)
264
+
265
+ def helper_gate(name, args, st):
266
+ if name not in ('Agent', 'AgentSwarm') or os.getenv('MINI_HELPER_GATE', '1') == '0': return
267
+ if args.get('resume') or HELPER_ASK.search(st.get('task') or ''): return
268
+ used = int(st.get('helpers') or 0)
269
+ cap = HELPER_CAP_LARGE if st.get('large') else 0
270
+ if used >= cap:
271
+ event('note', content='Helper agent not started: %s' % ('small task, done directly' if not cap else 'limit of %d helpers reached' % cap))
272
+ block(('This task is small enough to do directly. ' if not cap else 'You already started %d helper agents for this task. ' % used) +
273
+ 'Do the work yourself with Read, Grep, Edit and Bash (several reads can go in one response). Helpers start cold and cost far more tokens. This call was not run.')
274
+ st['helpers'] = used + 1; save_state(st)
275
+
276
+ # Optional soft per-run budget, off by default (users are not asked about tokens). With MINI_RUN_SOFT_BUDGET set, past
277
+ # that many raw tokens the user is asked once whether to continue; each approval allows another budget's worth.
278
+ def soft_budget():
279
+ try: return max(0, int(os.getenv('MINI_RUN_SOFT_BUDGET', '0') or 0))
280
+ except ValueError: return 0
281
+
282
+ def budget_gate(st, name):
283
+ budget = soft_budget()
284
+ if not budget or name in ('TodoList', 'WaitFor', 'TaskOutput', 'TaskList'): return
285
+ if st.get('budget_declined'):
286
+ if name in ('Write', 'Edit', 'Bash', 'Agent', 'AgentSwarm'):
287
+ block('The user chose to stop this task at its token budget. Do not start new work. Write the final summary now: what changed, what was verified, what is left and the exact next step.')
288
+ return
289
+ u = load('usage.json', {}); used = u.get('raw') or u.get('tokens') or 0
290
+ threshold = st.get('budget_next') or budget
291
+ if used < threshold: return
292
+ request = {'id': os.urandom(8).hex(), 'kind': 'budget', 'command': 'This task has used %s tokens (budget %s). Continue working?' % (format(int(used), ','), format(int(threshold), ',')),
293
+ 'title': 'Token budget reached', 'used': int(used), 'budget': int(threshold)}
294
+ policy_event('soft_budget', 'ask', 'Asked the user whether to continue past %s tokens' % format(int(threshold), ','))
295
+ try:
296
+ wait_decision(request, seconds=540, strict=True)
297
+ except SystemExit:
298
+ st['budget_declined'] = True; save_state(st)
299
+ try: (ROOT / 'budget-declined').write_text('1')
300
+ except OSError: pass
301
+ policy_event('soft_budget', 'stop', 'Stopped at the token budget')
302
+ block('The user chose to stop this task at its token budget (or did not answer). Do not start new work. Write the final summary now: what changed, what was verified, what is left and the exact next step, so the task can be resumed.')
303
+ st['budget_next'] = int(used) + budget; save_state(st)
304
+ policy_event('soft_budget', 'continue', 'User approved another %s tokens' % format(budget, ','))
305
+
217
306
  def token_gate(pol, st, name):
218
307
  """Per-task token budget from policy: at 80% wrap up (Plan-style, no new work), at 100% stop and summarize."""
219
308
  account_gate(st, name)
@@ -283,6 +372,9 @@ def pre(payload, st):
283
372
  args = payload.get('tool_input') or {}
284
373
  if not isinstance(args, dict): args = {}
285
374
  if (ROOT / 'cancel').exists(): block('Task stopped by user.')
375
+ if st.get('plan_dismissed') and is_work_tool(name): # before any approval card: Not now means nothing else asks or runs
376
+ event('plan_review', content='Plan on hold: ' + re.sub(r'^mcp__\w+?__', '', name) + ' was not run')
377
+ block(DISMISSED_BLOCK)
286
378
  hooked = user_hooks('kimi_pre', name, args, st) # hard-deny, credential block, before_tool / before_command hooks
287
379
  if hooked: block(hooked)
288
380
  queued = steering()
@@ -305,15 +397,19 @@ def pre(payload, st):
305
397
  save_state(st)
306
398
  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.')
307
399
  if name == 'TodoList': return plan_gate(args, st)
308
- writes = name in ('Write', 'Edit', 'Bash', 'Agent', 'AgentSwarm')
309
- if writes and st.get('plan_dismissed'): block('The user dismissed the task list, so no work may start. Ask what they want changed and finish.')
310
- if briefing(st) and name in ('Write', 'Edit', 'Agent', 'AgentSwarm'):
400
+ if briefing(st) and (name in ('Write', 'Edit', 'Agent', 'AgentSwarm') or (name.startswith('mcp__') and is_work_tool(name))):
311
401
  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.')
312
402
  if st.get('inspect') and name in ('Write', 'Edit', 'Agent', 'AgentSwarm'):
313
403
  event('safety', effect='write', decision='deny', command=name + ' ' + str(args.get('path', ''))[:300], content='Refused: the user asked for no changes in this turn')
314
404
  block(INSPECT_BLOCK)
315
405
  if name in ('Write', 'Edit') and not st.get('readonly'):
316
406
  nudged = set(st.get('nudged') or [])
407
+ if (st.get('plan_first') and st.get('mode', 'approve') == 'approve' and not st.get('planned') and not st.get('plan_reviewed')
408
+ and 'guided_plan' not in nudged and not (ROOT / 'kimi-approve-all').exists()):
409
+ nudged.add('guided_plan'); st['nudged'] = sorted(nudged); save_state(st)
410
+ event('note', content='Guided mode: asked for the plan before the first edit')
411
+ block('Guided mode: the user approves your plan before any file changes. Call TodoList with 2 to 6 short, concrete items '
412
+ '(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.')
317
413
  if st.get('large') and not st.get('planned') and 'plan' not in nudged:
318
414
  nudged.add('plan'); st['nudged'] = sorted(nudged); save_state(st)
319
415
  event('note', content='Large task: asked for a plan before editing')
@@ -336,6 +432,8 @@ def pre(payload, st):
336
432
  auto = mode == 'auto' or (ROOT / 'kimi-approve-all').exists()
337
433
  runaway(name, args, st)
338
434
  token_gate(pol, st, name)
435
+ budget_gate(st, name)
436
+ helper_gate(name, args, st)
339
437
  if name == 'Bash':
340
438
  command = str(args.get('command', ''))
341
439
  try:
@@ -463,8 +561,10 @@ def command_gate(command, st, mode, auto):
463
561
  if effect not in ('read', 'check'): event('safety', effect=effect, decision='allow', command=command[:300], content=describe(decision))
464
562
  return
465
563
  request = {'id': os.urandom(8).hex(), 'kind': 'command', 'effect': effect, 'reason': decision.get('reason') or '', 'command': command[:20000]}
564
+ # Deletes and sudo ask every time, even after Approve all (BUG-0101). The card says so and offers no Approve all for them.
565
+ if effect in ('destructive', 'privilege'): request['strict'] = True
466
566
  event('safety', effect=effect, decision='ask', command=command[:300], content=describe(decision))
467
- wait_decision(request)
567
+ wait_decision(request, strict=effect in ('destructive', 'privilege')) # "Approve all" never covers deletes or sudo (BUG-0101)
468
568
  if effect in ('install', 'service'):
469
569
  kinds = sorted(set(st.get('approved_kinds') or []) | {effect}); st['approved_kinds'] = kinds; save_state(st)
470
570
  elif effect in ('write', 'exec') and not st.get('readonly'):
@@ -532,6 +632,7 @@ def stop(payload, st):
532
632
  queued = steering()
533
633
  if queued: steer_block(queued, stopping=True)
534
634
  if st.get('readonly') or st.get('inspect'): return readonly_stop(st)
635
+ if st.get('plan_dismissed'): return # Not now: the run ends here; no done gate may send the agent back to work (BUG-0102)
535
636
  if briefing(st) and not st.get('planned') and not st.get('brief_nudged'):
536
637
  st['brief_nudged'] = True; save_state(st)
537
638
  event('done_gate', content='Brief mode: asked for the task list before finishing')
@@ -0,0 +1,179 @@
1
+ """Stock photos (Pixabay) and brand lookup (Brandfetch) for decks and sites.
2
+ The loopback gateway in image_generation.py owns the provider keys (popped from the environment before any tool
3
+ starts); tools call it with the run-scoped gateway token. Laptop runs have no keys: the gateway forwards to the app
4
+ server's /api/model/v1/media endpoint with the run token, which holds the keys and caps use per run.
5
+ Pixabay terms: search results are cached for 24 hours and chosen images are downloaded into the project (no
6
+ permanent hotlinking); the photographer credit is recorded next to the files.
7
+ """
8
+ import hashlib, json, re, threading, time, urllib.parse, urllib.request, urllib.error
9
+ from pathlib import Path
10
+
11
+ PIXABAY = 'https://pixabay.com/api/'
12
+ BRANDFETCH = 'https://api.brandfetch.io/v2/brands/'
13
+ CACHE_SECONDS = 24 * 3600
14
+ MAX_PHOTO_BYTES = 12_000_000
15
+ MAX_LOGO_BYTES = 3_000_000
16
+ PHOTO_HOSTS = ('pixabay.com', 'cdn.pixabay.com')
17
+ LOGO_HOSTS = ('asset.brandfetch.io', 'cdn.brandfetch.io', 'asset.brandfetch.com')
18
+ CAPS = {'stock': 40, 'brand': 6}
19
+
20
+ STOCK_TOOL = {'name': 'stock_photo', 'description': 'Find a real photograph on Pixabay and save it under public/stock/ (free, no credits). Use for people, places, products, teams and textures; use generate_image for bespoke art. Returns the saved path, size, photographer credit and other matches; call again with pick=<id> to use a different match.',
21
+ 'inputSchema': {'type': 'object', 'properties': {'query': {'type': 'string', 'maxLength': 100, 'description': 'A few concrete English words, e.g. "team meeting laptop"'},
22
+ 'filename': {'type': 'string', 'description': 'New two-word kebab filename, e.g. team-meeting'},
23
+ 'orientation': {'type': 'string', 'enum': ['horizontal', 'vertical', 'all']}, 'pick': {'type': 'integer', 'description': 'Pixabay id from an earlier result'}},
24
+ 'required': ['query', 'filename'], 'additionalProperties': False}}
25
+ BRAND_TOOL = {'name': 'brand_lookup', 'description': "Look up a company's brand by domain (e.g. stripe.com): name, colours, fonts and logos saved under public/brand/. Use before designing a deck or site for a named company, so it carries their real logo and palette.",
26
+ 'inputSchema': {'type': 'object', 'properties': {'domain': {'type': 'string', 'maxLength': 120}}, 'required': ['domain'], 'additionalProperties': False}}
27
+
28
+
29
+ def _stem(name):
30
+ stem = re.sub(r'[^a-z0-9-]+', '-', str(name or '').lower().rsplit('.', 1)[0]).strip('-')[:48]
31
+ if not stem: raise ValueError('filename is required, e.g. team-meeting')
32
+ return stem
33
+
34
+ def clean_domain(value):
35
+ raw = str(value or '').strip().lower()
36
+ if '://' in raw: raw = urllib.parse.urlparse(raw).hostname or ''
37
+ raw = raw.split('/')[0].split('?')[0].removeprefix('www.')
38
+ if not re.fullmatch(r'(?=.{3,120}$)([a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?\.)+[a-z]{2,24}', raw): raise ValueError('Give a company domain such as stripe.com')
39
+ return raw
40
+
41
+ def _get(url, headers=None, limit=3_000_000, timeout=30):
42
+ req = urllib.request.Request(url, headers={'User-Agent': 'ml-ai-code/1', **(headers or {})})
43
+ with urllib.request.build_opener(urllib.request.ProxyHandler({})).open(req, timeout=timeout) as r:
44
+ data = r.read(limit + 1)
45
+ if len(data) > limit: raise ValueError('Response too large')
46
+ return data
47
+
48
+ def _host_ok(url, hosts):
49
+ p = urllib.parse.urlparse(str(url or ''))
50
+ return p.scheme == 'https' and (p.hostname or '') in hosts
51
+
52
+ def trim_hits(data):
53
+ """The fields a deck needs from a Pixabay search; nothing else leaves the provider response."""
54
+ out = []
55
+ for h in (data or {}).get('hits', [])[:12]:
56
+ url = h.get('largeImageURL') or h.get('webformatURL')
57
+ if not _host_ok(url, PHOTO_HOSTS): continue
58
+ out.append({'id': int(h.get('id') or 0), 'tags': str(h.get('tags') or '')[:120], 'url': url, 'width': int(h.get('imageWidth') or 0), 'height': int(h.get('imageHeight') or 0),
59
+ 'user': str(h.get('user') or '')[:60], 'page': str(h.get('pageURL') or '')[:300]})
60
+ return out
61
+
62
+ def trim_brand(data):
63
+ """Name, colours, fonts and logo URLs from a Brandfetch brand; logos only from Brandfetch's asset hosts."""
64
+ d = data or {}
65
+ colors = [{'hex': c.get('hex'), 'type': c.get('type'), 'brightness': c.get('brightness')} for c in d.get('colors', [])[:8] if re.fullmatch(r'#[0-9a-fA-F]{6}', str(c.get('hex') or ''))]
66
+ fonts = [{'name': str(f.get('name') or '')[:60], 'type': f.get('type'), 'origin': f.get('origin')} for f in d.get('fonts', [])[:4] if f.get('name')]
67
+ logos = []
68
+ for l in d.get('logos', [])[:8]:
69
+ for f in (l.get('formats') or [])[:6]:
70
+ if _host_ok(f.get('src'), LOGO_HOSTS) and f.get('format') in ('svg', 'png', 'jpeg', 'webp'):
71
+ logos.append({'type': l.get('type'), 'theme': l.get('theme'), 'format': f.get('format'), 'src': f.get('src'), 'width': f.get('width'), 'height': f.get('height')})
72
+ return {'name': str(d.get('name') or '')[:120], 'domain': str(d.get('domain') or '')[:120], 'description': str(d.get('description') or '')[:400], 'colors': colors, 'fonts': fonts, 'logos': logos}
73
+
74
+ def search_pixabay(key, query, orientation='horizontal', pick=None):
75
+ params = {'key': key, 'image_type': 'photo', 'safesearch': 'true', 'per_page': 12, 'order': 'popular'}
76
+ if pick: params['id'] = int(pick)
77
+ else: params.update(q=' '.join(str(query).split())[:100], orientation=orientation if orientation in ('horizontal', 'vertical') else 'all')
78
+ return trim_hits(json.loads(_get(PIXABAY + '?' + urllib.parse.urlencode(params))))
79
+
80
+ def fetch_brand(key, domain):
81
+ return trim_brand(json.loads(_get(BRANDFETCH + urllib.parse.quote(domain), {'Authorization': 'Bearer ' + key})))
82
+
83
+ class Media:
84
+ """Runs inside the gateway process. `remote` is (url, token) for laptop runs; otherwise the keys are used directly."""
85
+ def __init__(self, root, repo, stock_key='', brand_key='', remote=None, emit=lambda *a, **kw: None):
86
+ self.root, self.repo = Path(root), Path(repo).resolve()
87
+ self.stock_key, self.brand_key, self.remote, self.emit = stock_key, brand_key, remote, emit
88
+ self.used = {'stock': 0, 'brand': 0}; self.lock = threading.Lock()
89
+ @property
90
+ def available(self): return bool(self.remote or self.stock_key or self.brand_key)
91
+ def _cache(self, kind, key, fetch):
92
+ path = self.root / 'media-cache.json'
93
+ try: cache = json.loads(path.read_text())
94
+ except (FileNotFoundError, ValueError): cache = {}
95
+ ident = kind + ':' + hashlib.sha256(json.dumps(key, sort_keys=True).encode()).hexdigest()[:24]
96
+ hit = cache.get(ident)
97
+ if hit and time.time() - hit.get('at', 0) < CACHE_SECONDS: return hit['value']
98
+ with self.lock:
99
+ if self.used[kind] >= CAPS[kind]: raise ValueError('%s limit reached (%d per task). Reuse what you found.' % ('Stock photo search' if kind == 'stock' else 'Brand lookup', CAPS[kind]))
100
+ self.used[kind] += 1
101
+ value = fetch()
102
+ cache = {k: v for k, v in cache.items() if time.time() - v.get('at', 0) < CACHE_SECONDS}
103
+ cache[ident] = {'at': time.time(), 'value': value}
104
+ tmp = path.with_suffix('.tmp'); tmp.write_text(json.dumps(cache)[:4_000_000]); tmp.replace(path)
105
+ return value
106
+ def _call(self, op, body, direct):
107
+ if self.remote:
108
+ url, token = self.remote
109
+ req = urllib.request.Request(url, data=json.dumps({'op': op, **body}).encode(), headers={'Authorization': 'Bearer ' + token, 'Content-Type': 'application/json'})
110
+ try:
111
+ with urllib.request.urlopen(req, timeout=40) as r: return json.load(r)['result']
112
+ except urllib.error.HTTPError as exc:
113
+ try: message = json.load(exc).get('error', {}).get('message')
114
+ except Exception: message = None
115
+ raise ValueError(message or 'The media service had an error') from None
116
+ return direct()
117
+ def _dir(self, name):
118
+ public = self.repo / 'public'; d = public / name
119
+ for p in (public, d):
120
+ if p.is_symlink(): raise ValueError('Asset directory cannot be a symlink')
121
+ d.mkdir(parents=True, exist_ok=True); return d
122
+ def stock(self, args):
123
+ query = ' '.join(str(args.get('query') or '').split())[:100]
124
+ if not query: raise ValueError('query is required')
125
+ stem = _stem(args.get('filename')); orientation = args.get('orientation') or 'horizontal'; pick = args.get('pick')
126
+ if pick is not None and (isinstance(pick, bool) or not isinstance(pick, int) or pick <= 0): raise ValueError('pick is a Pixabay id from an earlier result')
127
+ if not (self.remote or self.stock_key): raise ValueError('Stock photos are not configured on this server (PIXABAY_API_KEY).')
128
+ hits = self._cache('stock', [query, orientation, pick], lambda: self._call('stock', {'query': query, 'orientation': orientation, **({'pick': pick} if pick else {})},
129
+ lambda: search_pixabay(self.stock_key, query, orientation, pick)))
130
+ if not hits: return {'saved': None, 'note': 'No photos matched. Try simpler, concrete words.', 'matches': []}
131
+ chosen = next((h for h in hits if pick and h['id'] == pick), hits[0])
132
+ if not _host_ok(chosen['url'], PHOTO_HOSTS): raise ValueError('Unexpected image host')
133
+ d = self._dir('stock'); ext = '.png' if chosen['url'].lower().endswith('.png') else '.jpg'; path = d / (stem + ext)
134
+ if any(d.glob(stem + '.*')): raise ValueError('Asset filename already exists. Choose a new name.')
135
+ data = _get(chosen['url'], limit=MAX_PHOTO_BYTES, timeout=60)
136
+ if not (data[:3] == b'\xff\xd8\xff' or data[:8] == b'\x89PNG\r\n\x1a\n'): raise ValueError('The photo download was not an image')
137
+ with path.open('xb') as f: f.write(data)
138
+ credit = 'Photo by %s on Pixabay' % (chosen['user'] or 'a Pixabay contributor')
139
+ with (d / 'CREDITS.md').open('a') as f: f.write('- %s: %s (%s)\n' % (path.name, credit, chosen['page']))
140
+ rel = str(path.relative_to(self.repo))
141
+ self.emit('stock_photo', content='Saved ' + rel, path=rel, credit=credit)
142
+ return {'saved': {'path': rel, 'url': '/stock/' + path.name, 'width': chosen['width'], 'height': chosen['height'], 'credit': credit, 'tags': chosen['tags']},
143
+ 'matches': [{k: h[k] for k in ('id', 'tags', 'width', 'height')} for h in hits if h is not chosen][:8]}
144
+ def brand(self, args):
145
+ domain = clean_domain(args.get('domain'))
146
+ if not (self.remote or self.brand_key): raise ValueError('Brand lookup is not configured on this server (BRANDFETCH_API_KEY).')
147
+ info = self._cache('brand', [domain], lambda: self._call('brand', {'domain': domain}, lambda: fetch_brand(self.brand_key, domain)))
148
+ d = self._dir('brand'); slug = re.sub(r'[^a-z0-9]+', '-', domain).strip('-'); saved = []
149
+ # One logo per background, SVG first; then an icon as a fallback mark. Brandfetch's theme names the logo's own
150
+ # tone: a "light" logo is white (for dark slides), a "dark" logo is coloured or black (for light slides).
151
+ order = sorted(info.get('logos', []), key=lambda l: (l['type'] != 'logo', l['format'] != 'svg'))
152
+ seen = set()
153
+ for l in order:
154
+ if l['type'] not in ('logo', 'icon', 'symbol'): continue
155
+ tag = l['type'] + {'light': '-for-dark-bg', 'dark': '-for-light-bg'}.get(l.get('theme'), '')
156
+ if tag in seen or len(saved) >= 4: continue
157
+ try: data = _get(l['src'], limit=MAX_LOGO_BYTES)
158
+ except Exception: continue
159
+ ext = {'svg': '.svg', 'png': '.png', 'jpeg': '.jpg', 'webp': '.webp'}[l['format']]
160
+ if ext == '.svg' and b'<svg' not in data[:2000].lower(): continue
161
+ path = d / ('%s-%s%s' % (slug, tag, ext)); path.write_bytes(data); seen.add(tag)
162
+ saved.append({'path': str(path.relative_to(self.repo)), 'type': l['type'], 'use_on': {'light': 'dark backgrounds', 'dark': 'light backgrounds'}.get(l.get('theme'), 'any background'), 'format': l['format']})
163
+ self.emit('brand_lookup', content='Brand for ' + domain + (': ' + info.get('name') if info.get('name') else ''))
164
+ return {'name': info.get('name'), 'domain': domain, 'description': info.get('description'), 'colors': info.get('colors'), 'fonts': info.get('fonts'), 'logos': saved,
165
+ 'note': 'Pick the logo whose use_on matches the deck background (theme.logo). Use the brand colours as accent/accent2 only if text stays readable (4.5:1).'}
166
+
167
+ def call(name, args):
168
+ """Tool side (runner or the Kimi MCP server): forward to the loopback gateway, which holds the keys."""
169
+ import os
170
+ base, token = os.getenv('MINI_IMAGE_URL', ''), os.getenv('MINI_IMAGE_TOKEN', '')
171
+ if not base or not token: raise ValueError('Stock photos and brand lookup are not configured on this server.')
172
+ url = base.rsplit('/', 1)[0] + ('/stock' if name == 'stock_photo' else '/brand')
173
+ req = urllib.request.Request(url, data=json.dumps(args).encode(), headers={'Authorization': 'Bearer ' + token, 'Content-Type': 'application/json'})
174
+ try:
175
+ with urllib.request.build_opener(urllib.request.ProxyHandler({})).open(req, timeout=120) as r: return json.load(r)['result']
176
+ except urllib.error.HTTPError as exc:
177
+ try: message = json.load(exc).get('error', 'Request failed')
178
+ except ValueError: message = 'Request failed'
179
+ raise ValueError(message) from None
package/worker/memory.py CHANGED
@@ -81,10 +81,10 @@ class MemoryStore:
81
81
  def index_text(self, limit=6000):
82
82
  facts = self.list()
83
83
  if not facts:
84
- return ('Project memory: empty. This workspace keeps durable facts across tasks in mcp__mlai__memory. When you learn something that will matter next time '
84
+ return ('Computer memory (this chat\'s computer only; the project Memory card is separate): empty. This workspace keeps durable facts across tasks in mcp__mlai__memory. When you learn something that will matter next time '
85
85
  '(how to run or test this project, a convention, a preference the user stated, a correction they gave, where something lives), save it with '
86
86
  'mcp__mlai__memory {"action":"save","name":"short-name","type":"project|user|feedback|reference","description":"one line","content":"the fact"}. Never store secrets.')
87
- text = 'Project memory (facts saved in earlier tasks; read one with mcp__mlai__memory {"action":"read","name":"<name>"}; save new durable facts with action "save", correct wrong ones by saving the same name, drop them with "forget"):\n'
87
+ text = 'Computer memory (this chat\'s computer only; the project Memory card is separate; facts saved in earlier tasks; read one with mcp__mlai__memory {"action":"read","name":"<name>"}; save new durable facts with action "save", correct wrong ones by saving the same name, drop them with "forget"):\n'
88
88
  body = self.index.read_text() if self.index.is_file() else ''
89
89
  body = '\n'.join(l for l in body.splitlines() if l.startswith('- ') or l.startswith('## '))
90
90
  return (text + body)[:limit]