mlai-code 0.3.3 → 0.3.5
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +1 -1
- package/package.json +1 -1
- package/src/checkpoints.mjs +217 -0
- package/src/cli.mjs +153 -26
- package/src/cloud.mjs +2 -0
- package/src/engine.mjs +1 -1
- package/src/headless.mjs +39 -11
- package/src/home.mjs +3 -2
- package/src/host.mjs +141 -13
- package/src/ide-quick.mjs +147 -0
- package/src/ide.mjs +182 -0
- package/src/netbox.mjs +1 -1
- package/src/parity.mjs +17 -0
- package/src/relay.mjs +1 -1
- package/src/render.mjs +16 -5
- package/src/runtime.mjs +2 -0
- package/src/server.mjs +46 -7
- package/src/session.mjs +31 -8
- package/src/tui/acp.mjs +121 -19
- package/src/tui/app.mjs +2 -1
- package/src/tui/chat.mjs +35 -10
- package/src/tui/cmd-config.mjs +3 -1
- package/src/tui/cmd-core.mjs +9 -5
- package/src/tui/cmd-extensibility.mjs +5 -0
- package/src/tui/extensions-local.mjs +2 -0
- package/src/tui/mcp-config.mjs +22 -4
- package/src/tui/review-cli.mjs +15 -4
- package/src/tui/snapshots.mjs +57 -16
- package/src/tui/transcript.mjs +13 -2
- package/src/two-words.mjs +14 -8
- package/src/view-sync.mjs +48 -0
- package/src/winenv.mjs +1 -1
- package/src/workspace-api.mjs +21 -4
- package/vendor/client-sdk/index.cjs +2 -1
- package/vendor/cloud-cli/mlai.mjs +23 -11
- package/worker/architecture_worker.py +33 -4
- package/worker/browser_tool.py +19 -1
- package/worker/chat_fork.py +4 -1
- package/worker/checkpoints.py +71 -8
- package/worker/command_safety.py +56 -3
- package/worker/connected_turn.py +30 -0
- package/worker/conversation.py +125 -13
- package/worker/git_sync.py +4 -4
- package/worker/image_generation.py +2 -2
- package/worker/kimi_bridge.py +251 -31
- package/worker/kimi_gates.py +25 -6
- package/worker/kimi_hook.py +154 -13
- package/worker/media_tools.py +4 -1
- package/worker/mlai_hooks.py +222 -3
- package/worker/mlai_mcp.py +26 -6
- package/worker/mlai_review/README.md +35 -0
- package/worker/mlai_review/__init__.py +1 -0
- package/worker/mlai_review/__main__.py +261 -0
- package/worker/mlai_review/_compat.py +29 -0
- package/worker/mlai_review/analysis/__init__.py +3 -0
- package/worker/mlai_review/analysis/detect.py +50 -0
- package/worker/mlai_review/analysis/main.py +128 -0
- package/worker/mlai_review/analysis/messages.py +8 -0
- package/worker/mlai_review/analysis/tools.py +194 -0
- package/worker/mlai_review/application/__init__.py +1 -0
- package/worker/mlai_review/application/asking.py +53 -0
- package/worker/mlai_review/application/challengers.py +142 -0
- package/worker/mlai_review/application/context_builder.py +194 -0
- package/worker/mlai_review/application/errors.py +17 -0
- package/worker/mlai_review/application/feedback.py +52 -0
- package/worker/mlai_review/application/matching.py +46 -0
- package/worker/mlai_review/application/messages.py +41 -0
- package/worker/mlai_review/application/ports.py +222 -0
- package/worker/mlai_review/application/prompts.py +342 -0
- package/worker/mlai_review/application/review_service.py +200 -0
- package/worker/mlai_review/application/reviewers.py +438 -0
- package/worker/mlai_review/application/streaming.py +91 -0
- package/worker/mlai_review/application/tools.py +138 -0
- package/worker/mlai_review/application/usage.py +145 -0
- package/worker/mlai_review/application/write_check.py +46 -0
- package/worker/mlai_review/contracts/__init__.py +5 -0
- package/worker/mlai_review/contracts/analysis.py +57 -0
- package/worker/mlai_review/contracts/review.py +42 -0
- package/worker/mlai_review/domain/__init__.py +4 -0
- package/worker/mlai_review/domain/changes.py +77 -0
- package/worker/mlai_review/domain/context.py +147 -0
- package/worker/mlai_review/domain/diagnostics.py +239 -0
- package/worker/mlai_review/domain/feedback.py +216 -0
- package/worker/mlai_review/domain/findings.py +505 -0
- package/worker/mlai_review/domain/lifecycle.py +85 -0
- package/worker/mlai_review/domain/partition.py +42 -0
- package/worker/mlai_review/domain/relations.py +166 -0
- package/worker/mlai_review/domain/relevance.py +202 -0
- package/worker/mlai_review/domain/report.py +99 -0
- package/worker/mlai_review/domain/structure.py +56 -0
- package/worker/mlai_review/domain/writes.py +86 -0
- package/worker/mlai_review/eval.py +221 -0
- package/worker/mlai_review/eval_cases/clean-rename.json +11 -0
- package/worker/mlai_review/eval_cases/missing-ownership-check.json +22 -0
- package/worker/mlai_review/eval_cases/off-by-one.json +22 -0
- package/worker/mlai_review/eval_cases/sql-injection.json +20 -0
- package/worker/mlai_review/feedback_report.py +60 -0
- package/worker/mlai_review/infrastructure/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/analysis/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/analysis/none.py +12 -0
- package/worker/mlai_review/infrastructure/analysis/translate.py +17 -0
- package/worker/mlai_review/infrastructure/github/__init__.py +0 -0
- package/worker/mlai_review/infrastructure/github/client.py +445 -0
- package/worker/mlai_review/infrastructure/github/feedback.py +212 -0
- package/worker/mlai_review/infrastructure/github/http.py +31 -0
- package/worker/mlai_review/infrastructure/github/rendering.py +398 -0
- package/worker/mlai_review/infrastructure/github/token.py +20 -0
- package/worker/mlai_review/infrastructure/local/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/local/analyzer.py +48 -0
- package/worker/mlai_review/infrastructure/local/events.py +100 -0
- package/worker/mlai_review/infrastructure/local/git_reader.py +233 -0
- package/worker/mlai_review/infrastructure/local/publisher.py +138 -0
- package/worker/mlai_review/infrastructure/messages.py +54 -0
- package/worker/mlai_review/infrastructure/models/__init__.py +0 -0
- package/worker/mlai_review/infrastructure/models/openai_compatible.py +269 -0
- package/worker/mlai_review/infrastructure/structure/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/structure/heuristic.py +186 -0
- package/worker/mlai_review/infrastructure/structure/tree_sitter.py +186 -0
- package/worker/mlai_review/infrastructure/workspace/__init__.py +1 -0
- package/worker/mlai_review/infrastructure/workspace/local.py +105 -0
- package/worker/mlai_review/infrastructure/workspace/memory.py +77 -0
- package/worker/mlai_review/infrastructure/workspace/paths.py +17 -0
- package/worker/mlai_review/pipeline.py +89 -0
- package/worker/mlai_review/runtime.py +194 -0
- package/worker/mlai_review/settings.py +139 -0
- package/worker/preview.py +15 -0
- package/worker/product_facts.py +3 -1
- package/worker/repo_policy.py +5 -0
- package/worker/roles.py +29 -2
- package/worker/runner.py +164 -23
- package/worker/runtime_control.py +95 -9
- package/worker/secret_scan.py +3 -2
- package/worker/server_memory.py +122 -0
- package/worker/subagents.py +6 -2
- package/worker/tool_catalog.py +5 -2
- package/worker/verification.py +113 -8
package/worker/kimi_bridge.py
CHANGED
|
@@ -18,7 +18,7 @@ import handover
|
|
|
18
18
|
STALL_SECONDS = int(os.getenv('MINI_KIMI_STALL_SECONDS', '60'))
|
|
19
19
|
from runner import ROOT, REPO, event, write, unvault, VAULT, vault_guidance, connect_guidance, resolve_role, inference_progress, CREDITS_OUT_NOTE
|
|
20
20
|
from lifecycle import boot_id, identity, load
|
|
21
|
-
from conversation import enabled as conversation_enabled, readonly as conversation_readonly, route as route_conversation, recent_history, honor_no_change, NO_CHANGE_NOTE
|
|
21
|
+
from conversation import enabled as conversation_enabled, readonly as conversation_readonly, route as route_conversation, recent_history, honor_no_change, NO_CHANGE_NOTE, supersede
|
|
22
22
|
from design_director import restore_choice, prepare as prepare_design, browser_review, DESIGN_BUILD_GUIDANCE, DESIGN_BRIEF_GUIDANCE
|
|
23
23
|
import ask_first
|
|
24
24
|
from runtime_control import completion_report, check_key, clip
|
|
@@ -172,6 +172,12 @@ class ModelProxy:
|
|
|
172
172
|
|
|
173
173
|
def forward(self, handler, body):
|
|
174
174
|
body = unvault(body) # Vault secret values never reach the model
|
|
175
|
+
try: # BUG-0527: search/list/command output lines from hard-denied files never reach the model
|
|
176
|
+
import mlai_hooks
|
|
177
|
+
for m in body.get('messages') or []:
|
|
178
|
+
if isinstance(m, dict) and m.get('role') == 'tool' and isinstance(m.get('content'), str):
|
|
179
|
+
m['content'] = mlai_hooks.filter_denied_lines(m['content'], repo=REPO)[0]
|
|
180
|
+
except Exception: pass
|
|
175
181
|
body['model'] = self.model
|
|
176
182
|
if getattr(self, 'images_off', False): strip_images(body)
|
|
177
183
|
stream = bool(body.get('stream'))
|
|
@@ -217,6 +223,8 @@ class ModelProxy:
|
|
|
217
223
|
# The endpoint refused image input: carry on text-only for the rest of the run instead of failing the step.
|
|
218
224
|
self.images_off = True
|
|
219
225
|
event('note', content='Images could not be sent to the model, so this run continues without them.')
|
|
226
|
+
# BUG-0230: the app offers a one-click resend with a model that can see images.
|
|
227
|
+
event('vision_unavailable', model=os.getenv('MINI_MODEL_ID', ''), content=(os.getenv('MINI_MODEL_LABEL') or 'This model') + ' could not see the attached image, so its answer is not based on what the image shows.')
|
|
220
228
|
req = urllib.request.Request(self.endpoint, data=json.dumps(body).encode(), headers=dict(req.header_items()))
|
|
221
229
|
res = urllib.request.urlopen(req, timeout=STALL_SECONDS)
|
|
222
230
|
elif not (effort and exc.code == 400 and b'reasoning' in exc.detail.lower()): raise
|
|
@@ -308,9 +316,13 @@ Working rules:
|
|
|
308
316
|
- The project is already a git repository. When a scaffolder runs inside it, skip its own git init (forge init --no-git, cargo new --vcs none, create-* --no-git); a nested repository breaks checkpoints and GitHub sync.
|
|
309
317
|
- Read before you edit. Prefer Grep/Glob and focused Read ranges over dumping whole files. Use mcp__mlai__repo_map once to orient in an unfamiliar repository and mcp__mlai__symbols to find declarations.
|
|
310
318
|
- Change source files with Edit (exact text replacement) or Write, never with sed -i, perl -pi or a python/node script that rewrites the file: those break silently on JSX/TSX quotes and braces. After a batch of edits, run the type check or build before you report anything as working.
|
|
311
|
-
-
|
|
319
|
+
- You have live web access in every turn: use mcp__mlai__web_search and mcp__mlai__web_fetch whenever the answer depends on current public facts you cannot see in the repository (documentation and versions, but also weather, news, prices, scores or local time). Never say you lack web, real-time or weather access. For research answers, record the material claims with mcp__mlai__evidence and link the source URLs. Never put credentials or private code in queries.
|
|
312
320
|
- After changing code, run the relevant tests, build or type check with Bash and read the real exit status. A passing unrelated suite does not prove the requested behavior. Never claim something works without running it.
|
|
313
321
|
- Do not fabricate data, mocks or placeholder implementations in shipped code unless the user asked for them.
|
|
322
|
+
- Do what the user asked, and only that (BUG-0226). A request to install a package, run a command or answer a question is done when that is done: do not add features, tools, commands or files nobody asked for, do not rename the package, remove "private" from package.json, publish, install globally or push unless the user asked for it. When extra work looks useful, finish the request, then suggest it in one sentence and ask; do not start it.
|
|
323
|
+
- Once the user approves a plan or task list, build exactly what it says (BUG-0245). If an approved item turns out to need a different approach, stop and ask before changing it; if you had to change it, the final message must name the item, what you did instead and why, in plain words.
|
|
324
|
+
- Copy commit SHAs, PR numbers, branch names, IDs and version numbers into your answer from the tool output of this task, character for character; never type them from memory or guess them (BUG-0458). A PR's head commit is the head SHA the tool returned, not the base branch's commit.
|
|
325
|
+
- To look at the local app in a browser, use mcp__mlai__browser: it is the built-in Playwright browser. Do not search for, install or script Playwright or Chromium yourself; if the tool reports it is not set up, say so (BUG-0202).
|
|
314
326
|
- For web apps, run the app with mcp__mlai__preview {"action":"start"} (port 3000, dev server with hot reload; it installs missing packages first and keeps running after the task, when it is swapped for the production build the user shares). Vite needs server.host true and server.allowedHosts true; Next.js and static index.html work as is. Servers you start with Bash in the background stop when the task ends. Check UI work with mcp__mlai__browser at a phone width (390) and a desktop width (1280). A hydration error that names data-myki, jcpwm, data-gr-*, data-lastpass or data-1p-* comes from the user's browser extension, not the app: do not rewrite components for it (the production preview does not show it); in Next.js root layouts keep suppressHydrationWarning on <html> and <body>.
|
|
315
327
|
- Commands that run longer than a minute should use run_in_background, then WaitFor/TaskOutput. Never sleep in a loop. Work that must outlive this turn (training, benchmark sweeps, pipelines, extra servers) goes in mcp__mlai__job: it logs, survives the turn, and is restarted from its resume command if it stops unexpectedly. Make such work checkpoint its progress.
|
|
316
328
|
- If a command cannot reach a host because the workspace network blocks it, call mcp__mlai__network {domain, reason} once, then retry after the user allows it.
|
|
@@ -318,13 +330,15 @@ Working rules:
|
|
|
318
330
|
- Computer memory (mcp__mlai__memory) persists across tasks on this chat's computer only, not in the sidebar project's Memory card: when you learn how this project is run or tested, a convention, a user preference or a correction, save it with mcp__mlai__memory (never secrets). Read a listed fact before re-discovering it.
|
|
319
331
|
- Only claim that tests, builds or type checks pass when you ran them in this task and saw the result; the final message is checked against the commands you actually ran. Do not edit, skip or delete existing tests to make them pass unless the task asks for it, and say so when you do.
|
|
320
332
|
- Your final message is what the user reads on their phone: say what changed, how you verified it (commands and results) and anything left open, briefly. When the user asked to see output (for example "run it and show me the output"), paste that output in a code block (up to about 60 lines) instead of only saying it ran.
|
|
321
|
-
- Refer to files by their path inside the project (src/app.js), never by the computer's absolute path.
|
|
333
|
+
- Refer to files by their path inside the project (src/app.js), never by the computer's absolute path.
|
|
334
|
+
- Earlier messages of this chat come with the request: answer questions about what the user already told you from them. Do not state a conclusion (such as "I don't know") before tool calls that could change it; say what you are checking instead.'''
|
|
322
335
|
|
|
323
336
|
SHELL_ALWAYS = os.getenv('MINI_KIMI_SHELL', '1') != '0'
|
|
324
337
|
TALK_RULES = '''This turn is a {route} request, and you have the full workspace shell. Run any command that helps: status checks, logs, processes, tests, network calls, CLIs that use the Vault keys in your environment. When the question is about work the user already started or approved (a job, sweep, deploy, server or earlier task), act on it in this turn: check it, and resume or restart it if it stalled, rather than asking the user to say go. Do not begin new implementation or edit project files the user has not asked for; recommend those changes instead. Answer directly and concisely.'''
|
|
325
338
|
TALK_RESEARCH = '''Research with mcp__mlai__web_search and mcp__mlai__web_fetch, record material claims with mcp__mlai__evidence, and link the source URLs in the answer.'''
|
|
326
339
|
INSPECT_RULES = '''The user asked for no changes in this turn (an assessment, review or "ask before fixing"). Do not edit, create or delete project files, install packages, start services or commit; those calls are refused. You may read the code and run existing tests and checks to reproduce defects. Report each finding with its evidence (file and line, failing command and output) and the fix you recommend, then ask whether to apply the fixes. Results of the checks you run are investigation evidence, not proof that the project is broken.'''
|
|
327
|
-
|
|
340
|
+
# BUG-0235: a build or deploy request in Plan mode gets a pointer to the Mode switch, never a missing-tools story.
|
|
341
|
+
PLAN_RULES = '''The user picked Plan mode: they review a plan before anything is built. Do not try to create or change files and do not ask the user to run commands themselves. If the message is only a question, just answer it. For work, write the plan as a short numbered list (files to add or change, what each step does and how it will be checked), using project paths such as index.html. End with one sentence saying they can tap Build this plan to start. If they ask you to build, create files, deploy or go ahead while Plan mode is still on, do not try: say in one or two sentences that Plan mode only plans, and that they can tap Build this plan or switch Mode to Approve or Auto, then send the request again. Never describe this mode as a tool restriction, never say a tool is missing or not found, and never list tool names.'''
|
|
328
342
|
READONLY_RULES = '''This request is read-only: answer, research or plan. Do not modify files, run builds or start servers. Inspect the repository with Read, Grep and Glob, research with mcp__mlai__web_search and mcp__mlai__web_fetch when needed (record claims with mcp__mlai__evidence and link sources), then answer directly and concisely. If you propose a plan, make it concrete and numbered.'''
|
|
329
343
|
|
|
330
344
|
GLM_RULES = '''Model notes (GLM):
|
|
@@ -361,7 +375,7 @@ def context_size(config):
|
|
|
361
375
|
def end_reason_for(run_state, result, state):
|
|
362
376
|
"""Why a run ended, stored with its usage so token spend can be split by outcome (e.g. verification_incomplete)."""
|
|
363
377
|
if result.get('exit_status') == 'VerificationIncomplete' or state.get('_soft_unverified'): return 'verification_incomplete' # usage split keeps counting them
|
|
364
|
-
if run_state == 'limited': return 'limited_' + ('tokens'
|
|
378
|
+
if run_state == 'limited': return 'limited_' + ({'tokens': 'tokens', 'steps': 'steps', 'sandbox': 'sandbox'}.get(state.get('limit_kind') or 'time') or 'time')
|
|
365
379
|
if (ROOT / 'budget-declined').exists(): return 'budget_declined'
|
|
366
380
|
return run_state if run_state in ('completed', 'stopped', 'error', 'interrupted') else 'completed'
|
|
367
381
|
|
|
@@ -424,6 +438,10 @@ def attached_image_note(config):
|
|
|
424
438
|
'sentence and suggest switching to the default model or describing the image; do not guess or run OCR as a substitute.]'
|
|
425
439
|
% (len(images), '' if len(images) == 1 else 's', listed))
|
|
426
440
|
|
|
441
|
+
# BUG-0230: the model is told it cannot see the image, so it says so instead of guessing from pixel values.
|
|
442
|
+
IMAGE_OMITTED = ('[image omitted: this model cannot see images. Do not guess what the image shows from pixel values, file data or '
|
|
443
|
+
'its name. Tell the person in one sentence that you could not see it and that they can switch to a model that can see images.]')
|
|
444
|
+
|
|
427
445
|
def strip_images(body):
|
|
428
446
|
"""Replace image parts with a short text note. Returns how many were removed."""
|
|
429
447
|
n = 0
|
|
@@ -432,9 +450,65 @@ def strip_images(body):
|
|
|
432
450
|
if not isinstance(c, list): continue
|
|
433
451
|
for i, part in enumerate(c):
|
|
434
452
|
if isinstance(part, dict) and part.get('type') in ('image_url', 'input_image', 'image'):
|
|
435
|
-
c[i] = {'type': 'text', 'text':
|
|
453
|
+
c[i] = {'type': 'text', 'text': IMAGE_OMITTED}; n += 1
|
|
436
454
|
return n
|
|
437
455
|
|
|
456
|
+
# Never in the engine's environment, so never in the environment of the commands it runs (BUG-0517). The MCP server
|
|
457
|
+
# gets the image/research/hub values it needs through its own env in mcp.json; the engine reads the proxy token from config.toml.
|
|
458
|
+
ENGINE_ENV_HIDDEN = ('MODEL_API_KEY', 'SIDE_MODEL_API_KEY', 'VERCEL_OIDC_TOKEN', 'EXA_API_KEY', 'FIRECRAWL_API_KEY', 'OPENROUTER_API_KEY',
|
|
459
|
+
'PIXABAY_API_KEY', 'BRANDFETCH_API_KEY', 'MINI_VAULT', 'MINI_KIMI_PROXY_TOKEN', 'MINI_IMAGE_URL', 'MINI_IMAGE_TOKEN',
|
|
460
|
+
'MINI_RESEARCH_URL', 'MINI_RESEARCH_TOKEN', 'MLAI_HUB_TOKEN')
|
|
461
|
+
|
|
462
|
+
def engine_env(environ):
|
|
463
|
+
return {k: v for k, v in environ.items() if k not in ENGINE_ENV_HIDDEN}
|
|
464
|
+
|
|
465
|
+
LAST_OUTPUT = '' # the newest successful shell output in this run
|
|
466
|
+
SANDBOX_DENIALS = 0 # tool calls the OS sandbox refused in this run
|
|
467
|
+
SANDBOX_DENIED = re.compile(r'Operation not permitted|Read-only file system|sandbox-exec: .*denied|deny\(\d+\) file-write', re.I)
|
|
468
|
+
SANDBOX_STUCK_AFTER = 3
|
|
469
|
+
SANDBOX_STOP_NOTE = ('Stopped: the OS sandbox blocked writing outside this project %d times, and a run started with -p cannot ask you to allow it. '
|
|
470
|
+
'Nothing outside the project was changed. Run it interactively (mlai) to approve it, or work inside the project folder.' % SANDBOX_STUCK_AFTER)
|
|
471
|
+
|
|
472
|
+
def sandbox_stuck(config):
|
|
473
|
+
return bool(config.get('unattended')) and config.get('mode') == 'auto' and SANDBOX_DENIALS >= SANDBOX_STUCK_AFTER
|
|
474
|
+
|
|
475
|
+
def fill_dangling_output(answer, output, max_lines=40):
|
|
476
|
+
"""BUG-0505: "Created squares.py and ran it - output:" followed by nothing (the model expects the UI to show the
|
|
477
|
+
command). The newest command output goes under such a line; an answer that already shows output is unchanged."""
|
|
478
|
+
text = str(output or '').strip('\n')
|
|
479
|
+
if not text.strip() or not answer: return answer
|
|
480
|
+
lines = answer.split('\n')
|
|
481
|
+
for i, line in enumerate(lines):
|
|
482
|
+
if not re.search(r'(?i)\boutput:\s*$', line): continue
|
|
483
|
+
rest = [l for l in lines[i + 1:] if l.strip()]
|
|
484
|
+
if rest and (rest[0].lstrip().startswith('```') or rest[0].startswith((' ', '\t'))): continue
|
|
485
|
+
if rest and not re.search(r'(?i)\boutput:\s*$', rest[0]) and i + 1 < len(lines) and lines[i + 1].strip(): continue
|
|
486
|
+
shown = text.split('\n')
|
|
487
|
+
if len(shown) > max_lines: shown = shown[:max_lines] + ['...']
|
|
488
|
+
block = '```\n' + '\n'.join(shown).replace('```', "'''") + '\n```'
|
|
489
|
+
return '\n'.join(lines[:i + 1] + [block] + lines[i + 1:])
|
|
490
|
+
return answer
|
|
491
|
+
|
|
492
|
+
# CLI tool names (--tools / --disallowedTools / permission rules) -> the engine tools that stay out of the offer.
|
|
493
|
+
DENY_TO_ENGINE = {'Bash': ('Bash',), 'Edit': ('Write', 'Edit'), 'WebSearch': ('WebSearch',), 'Agent': ('Agent',)}
|
|
494
|
+
|
|
495
|
+
def denied_tools(config):
|
|
496
|
+
"""BUG-0510: a whole-tool deny rule (`--tools Read` denies Bash, Edit, ...) also removes the tool from what the agent
|
|
497
|
+
is offered, so it says at once that it cannot do the work instead of trying every tool and being blocked.
|
|
498
|
+
Rules with a pattern (Bash(rm:*)) still go through the permission check only."""
|
|
499
|
+
ext = config.get('extensions') if isinstance(config.get('extensions'), dict) else {}
|
|
500
|
+
perms = ext.get('permissions') if isinstance(ext.get('permissions'), dict) else {}
|
|
501
|
+
if perms.get('bypass') is True: return []
|
|
502
|
+
out = []
|
|
503
|
+
for rule in perms.get('deny') if isinstance(perms.get('deny'), list) else []:
|
|
504
|
+
if not isinstance(rule, str) or '(' in rule: continue
|
|
505
|
+
out += [t for t in DENY_TO_ENGINE.get({'write': 'Edit', 'edit': 'Edit', 'bash': 'Bash', 'shell': 'Bash', 'websearch': 'WebSearch', 'agent': 'Agent'}.get(rule.strip().lower(), ''), ()) if t not in out]
|
|
506
|
+
return out
|
|
507
|
+
|
|
508
|
+
def step_cap(config, steps):
|
|
509
|
+
"""Agent steps per turn for the engine. CLI --max-turns (hardSteps) is an exact cap; otherwise at least 10 (BUG-0492)."""
|
|
510
|
+
return max(1, steps) if config.get('hardSteps') else max(10, steps)
|
|
511
|
+
|
|
438
512
|
def write_home(config, proxy, mode, readonly, steps):
|
|
439
513
|
HOME.mkdir(parents=True, exist_ok=True)
|
|
440
514
|
hook = sys.executable + ' ' + str(Path(__file__).resolve().parent / 'kimi_hook.py')
|
|
@@ -444,13 +518,13 @@ def write_home(config, proxy, mode, readonly, steps):
|
|
|
444
518
|
'telemetry = false',
|
|
445
519
|
'extra_skill_dirs = [' + toml_str(HOME / 'skills') + ']',
|
|
446
520
|
'',
|
|
447
|
-
'[providers.mlai]', 'type = "openai"', 'base_url = ' + toml_str('http://127.0.0.1:%d/v1' % proxy.port), '
|
|
521
|
+
'[providers.mlai]', 'type = "openai"', 'base_url = ' + toml_str('http://127.0.0.1:%d/v1' % proxy.port), 'api_key = ' + toml_str(proxy.token), '',
|
|
448
522
|
'[models."mlai/main"]', 'provider = "mlai"', 'model = ' + toml_str(config.get('model') or 'glm'), 'max_context_size = %d' % context_size(config), 'capabilities = ' + json.dumps(model_capabilities(config)), '',
|
|
449
523
|
'[image]', 'max_edge_px = 1568', '',
|
|
450
|
-
'[loop_control]', 'max_steps_per_turn = %d' %
|
|
524
|
+
'[loop_control]', 'max_steps_per_turn = %d' % step_cap(config, steps), 'max_attempts_per_step = 4', 'reserved_context_size = %d' % RESERVED_CONTEXT, '',
|
|
451
525
|
'[task]', 'bash_auto_background_on_timeout = true', 'print_background_mode = "exit"', '',
|
|
452
526
|
'[mcp]', 'startup_timeout_ms = 20000', 'tool_timeout_ms = 600000', '',
|
|
453
|
-
'[tools]', 'disabled = ["CronCreate", "CronDelete", "CronList", "AskUserQuestion", "ExitPlanMode", "FetchURL", "AgentSwarm", "CreateGoal", "GetGoal", "SetGoalBudget", "UpdateGoal"' + (', "Write", "Edit", "Bash", "Agent"' if readonly else '') + ']', '',
|
|
527
|
+
'[tools]', 'disabled = ["CronCreate", "CronDelete", "CronList", "AskUserQuestion", "ExitPlanMode", "FetchURL", "AgentSwarm", "CreateGoal", "GetGoal", "SetGoalBudget", "UpdateGoal"' + (', "Write", "Edit", "Bash", "Agent"' if readonly else ''.join(', ' + toml_str(t) for t in denied_tools(config))) + ']', '',
|
|
454
528
|
]
|
|
455
529
|
events = [('PreToolUse', '', 600), ('Stop', '', 120)]
|
|
456
530
|
try:
|
|
@@ -459,8 +533,11 @@ def write_home(config, proxy, mode, readonly, steps):
|
|
|
459
533
|
except Exception: pass
|
|
460
534
|
for event_name, matcher, timeout in events:
|
|
461
535
|
lines += ['[[hooks]]', 'event = ' + toml_str(event_name)] + (['matcher = ' + toml_str(matcher)] if matcher else []) + ['command = ' + toml_str(hook + ' ' + event_name), 'timeout = %d' % timeout, '']
|
|
536
|
+
# BUG-0517: the per-run proxy token lives in this private file, not in the environment every agent command inherits.
|
|
462
537
|
(HOME / 'config.toml').write_text('\n'.join(lines))
|
|
463
|
-
|
|
538
|
+
try: os.chmod(HOME / 'config.toml', 0o600)
|
|
539
|
+
except OSError: pass
|
|
540
|
+
server = {'command': sys.executable, 'args': [str(Path(__file__).resolve().parent / 'mlai_mcp.py')], 'toolTimeoutMs': 600000, 'env': {'MINI_ROOT': str(ROOT), 'MINI_REPO': REPO, 'MINI_READONLY': '1' if readonly or config.get('inspect') or 'Write' in denied_tools(config) else '0', 'MINI_SUBAGENT': '1' if config.get('subagent') else '0', **({'FIRECRAWL_API_KEY': os.environ['FIRECRAWL_API_KEY']} if os.getenv('FIRECRAWL_API_KEY') else {}), **({'MINI_RESEARCH_URL': os.environ['MINI_RESEARCH_URL'], 'MINI_RESEARCH_TOKEN': os.environ['MINI_RESEARCH_TOKEN']} if os.getenv('MINI_RESEARCH_URL') and os.getenv('MINI_RESEARCH_TOKEN') else {}), **({'EXA_API_KEY': os.environ['EXA_API_KEY'], 'MINI_RESEARCH_FALLBACK': 'exa'} if os.getenv('EXA_API_KEY') and os.getenv('MINI_RESEARCH_FALLBACK') == 'exa' else {}), **({} if readonly or config.get('inspect') else computer_use.env()), **hub_env(), **{k: os.environ[k] for k in ('MINI_IMAGE_URL', 'MINI_IMAGE_TOKEN') if os.environ.get(k)}}}
|
|
464
541
|
servers = {'mlai': server}
|
|
465
542
|
try:
|
|
466
543
|
import mcp_local # the user's own MCP servers from the CLI (/mcp): .mlai/mcp.json, .mcp.json, .cursor/mcp.json, settings
|
|
@@ -509,12 +586,27 @@ def agent_file(config, readonly, extra_system):
|
|
|
509
586
|
# ---------------------------------------------------------------- stream mapping
|
|
510
587
|
TOOL_EVENT = {'Read': ('read_file', 'path'), 'Write': ('write_file', 'path'), 'Edit': ('edit_file', 'path'), 'Grep': ('search_text', 'pattern'), 'Glob': ('list_files', 'pattern'),
|
|
511
588
|
'FetchURL': ('web_fetch', 'url'), 'Skill': ('skill_load', 'skill'), 'mcp__mlai__web_search': ('web_search', 'query'), 'mcp__mlai__repo_map': ('repo_map', ''), 'mcp__mlai__symbols': ('symbols', 'query'), 'mcp__mlai__integration_search': ('tool_search', 'query'), 'mcp__mlai__integration_call': ('integration', 'name'), 'mcp__mlai__memory': ('memory', 'action'), 'mcp__mlai__scope': ('scope', ''), 'mcp__mlai__deploy': ('deploy', 'action')}
|
|
512
|
-
|
|
589
|
+
# A write the user or a permission rule refused never happened (BUG-0496: it was counted as an edit).
|
|
590
|
+
FAILED = re.compile(r'Command failed with exit code: (\d+)|^(Error|error):|<system>ERROR|Tool execution failed|Tool call was blocked|was not run|The user rejected this action|The user did not approve|Blocked by your permission rule', re.M)
|
|
513
591
|
# A command the hook refused (e.g. a test masked by `| tail` or `; git status`) never ran, so it is not a failed check.
|
|
514
592
|
HOOK_BLOCKED = re.compile(r'Tool call was blocked|This call was not run\.')
|
|
515
593
|
|
|
516
594
|
ENGINE_NAMES = re.compile(r'\bKimi(?:[ -]Code)?[ -]CLI\b|\bKimi Code\b(?= (?:agent|engine|harness))', re.I)
|
|
517
595
|
|
|
596
|
+
REFERS_BACK = re.compile(r'\b(?:above|previous (?:message|reply|answer)|earlier (?:message|reply|answer)|already (?:listed|given|provided|written))\b', re.I)
|
|
597
|
+
|
|
598
|
+
def final_answer(last, texts):
|
|
599
|
+
"""BUG-0192: the agent wrote the full answer, then a short closing message pointing back to it ("the tips above are
|
|
600
|
+
the complete deliverable"). Only the closing message was kept as the answer, so the list vanished from the chat after
|
|
601
|
+
the run (and after reload). When the last message is short and refers back, keep the long answer before it too."""
|
|
602
|
+
last = str(last or '').strip()
|
|
603
|
+
if not last or len(last) > 1500 or not REFERS_BACK.search(last): return last
|
|
604
|
+
earlier = [t for t in (texts or [])[:-1] if t and t.strip() != last]
|
|
605
|
+
if not earlier: return last
|
|
606
|
+
body = max(earlier[-4:], key=len)
|
|
607
|
+
if len(body) < max(600, 2 * len(last)) or body in last: return last
|
|
608
|
+
return body.rstrip() + '\n\n' + last
|
|
609
|
+
|
|
518
610
|
def public_text(text):
|
|
519
611
|
"""What the user reads: project paths instead of the computer's absolute paths, and no internal engine name
|
|
520
612
|
(BUG-0099, BUG-0185). The model's own words are otherwise unchanged."""
|
|
@@ -522,7 +614,9 @@ def public_text(text):
|
|
|
522
614
|
for root in sorted({str(REPO).rstrip('/'), '/vercel/sandbox/repo'}, key=len, reverse=True):
|
|
523
615
|
if root and len(root) > 1:
|
|
524
616
|
out = re.sub(re.escape(root) + r'/(?=\S)', '', out)
|
|
525
|
-
|
|
617
|
+
# BUG-0537: on the user's own computer the folder has a real name; show it instead of a placeholder.
|
|
618
|
+
label = Path(root).name if os.getenv('MINI_LAPTOP') == '1' and root == str(REPO).rstrip('/') and Path(root).name else 'the project folder'
|
|
619
|
+
out = re.sub(re.escape(root) + r'(?![\w/.-])', lambda _m: label, out)
|
|
526
620
|
out = re.sub(r'/vercel/sandbox(?:/[\w.-]+)*', 'the workspace', out)
|
|
527
621
|
return ENGINE_NAMES.sub('ML AI Code', out)
|
|
528
622
|
|
|
@@ -562,6 +656,8 @@ class StreamMapper:
|
|
|
562
656
|
if isinstance(content, str) and content.strip():
|
|
563
657
|
content = public_text(content)
|
|
564
658
|
self.last_text = content.strip(); event('assistant', content=content)
|
|
659
|
+
if not hasattr(self, 'texts'): self.texts = []
|
|
660
|
+
self.texts = (self.texts + [self.last_text])[-12:] # BUG-0192
|
|
565
661
|
try: write('kimi-last-text.json', {'text': self.last_text[:20000], 'time': time.time()})
|
|
566
662
|
except OSError: pass
|
|
567
663
|
calls = msg.get('tool_calls') or []
|
|
@@ -594,12 +690,20 @@ class StreamMapper:
|
|
|
594
690
|
# EXEC-001: `npm test | tail` exits with tail's status. Reconcile with the runner's own summary.
|
|
595
691
|
check_code, integrity = verification.check_exit(str(args.get('command', '')), code, content, is_check)
|
|
596
692
|
event('output', output=content[:12000], returncode=code if check_code is None else check_code, **({'integrity': integrity} if integrity else {}))
|
|
693
|
+
if name == 'Bash' and args.get('run_in_background') and not code:
|
|
694
|
+
# BUG-0203: remember the start command so a new computer can offer Restart (never restarted by itself).
|
|
695
|
+
try:
|
|
696
|
+
import server_memory; server_memory.remember(ROOT, str(args.get('command', '')), REPO)
|
|
697
|
+
except Exception: pass
|
|
597
698
|
st = self.state
|
|
598
699
|
if name in ('Write', 'Edit') and not code:
|
|
599
700
|
st['edits'] = st.get('edits', 0) + 1
|
|
600
701
|
# Editing prose does not make test evidence stale; editing anything else does.
|
|
601
702
|
if not PROSE_PATH.search(str(args.get('path', ''))): st['last_edit'] = time.time()
|
|
602
703
|
if not code: self.checkpoint(name, args)
|
|
704
|
+
global LAST_OUTPUT, SANDBOX_DENIALS
|
|
705
|
+
if name == 'Bash' and not code: LAST_OUTPUT = content
|
|
706
|
+
if code and name in ('Bash', 'Write', 'Edit') and SANDBOX_DENIED.search(content): SANDBOX_DENIALS += 1
|
|
603
707
|
if name == 'Bash':
|
|
604
708
|
command = str(args.get('command', ''))
|
|
605
709
|
try:
|
|
@@ -616,6 +720,16 @@ class StreamMapper:
|
|
|
616
720
|
write('checks.json', checks[-20:]); event('verification', **record)
|
|
617
721
|
if check_code == 0: st['last_check'] = time.time()
|
|
618
722
|
elif check_code is not None: st['last_failed_check'] = time.time()
|
|
723
|
+
elif code == 0 and not args.get('run_in_background') and not st.get('inspect') and not HOOK_BLOCKED.search(content):
|
|
724
|
+
# BUG-0542: no test suite here; running the changed program and seeing it work is the check.
|
|
725
|
+
try:
|
|
726
|
+
ran = verification.program_run_check(command, REPO)
|
|
727
|
+
if ran and not verification.has_test_suite(REPO):
|
|
728
|
+
record = {'label': clip('Ran ' + command, 160), 'command': command[:2000], 'returncode': 0, 'executed': True, 'scope': 'repository', 'kind': 'program-run', 'output': content[-4000:], 'time': time.time()}
|
|
729
|
+
checks = [c for c in load(ROOT, 'checks.json', []) if check_key(c.get('command')) != check_key(command)] + [record]
|
|
730
|
+
write('checks.json', checks[-20:]); event('verification', **record)
|
|
731
|
+
st['last_check'] = time.time()
|
|
732
|
+
except Exception: pass
|
|
619
733
|
remote = remote_check(command, code, content)
|
|
620
734
|
if remote:
|
|
621
735
|
checks = [c for c in load(ROOT, 'checks.json', []) if c.get('command') != remote['command']] + [remote]
|
|
@@ -635,10 +749,39 @@ def remote_check(command, code, output):
|
|
|
635
749
|
if pending and code in (0, 8):
|
|
636
750
|
return {'label': 'Remote CI', 'command': REMOTE_CI, 'returncode': None, 'executed': False, 'scope': 'ci', 'output': 'Remote checks are still running.\n' + (output or '')[-1500:], 'time': time.time()}
|
|
637
751
|
return {'label': 'Remote CI', 'command': REMOTE_CI, 'returncode': code, 'executed': True, 'scope': 'ci', 'output': (output or '')[-4000:], 'time': time.time()}
|
|
638
|
-
if PUSHED.search(command) and not code:
|
|
752
|
+
if PUSHED.search(command) and not code and push_succeeded(command, output):
|
|
753
|
+
if not has_ci(REPO):
|
|
754
|
+
# BUG-0456: nothing will ever report on this push, so a pending row would never clear.
|
|
755
|
+
event('note', content='Changes were pushed. This repository has no CI workflow, so there are no remote checks to wait for.')
|
|
756
|
+
return None
|
|
639
757
|
return {'label': 'Remote CI', 'command': REMOTE_CI, 'returncode': None, 'executed': False, 'scope': 'ci', 'output': 'Changes were pushed. The remote checks for that push were not observed in this run.', 'time': time.time()}
|
|
640
758
|
return None
|
|
641
759
|
|
|
760
|
+
# BUG-0456: `git push ... 2>&1 | tail` or `git push; echo done` exits 0 even when the push failed
|
|
761
|
+
# ("fatal: could not read Username for 'https://github.com'"). Only a push whose output shows no failure counts.
|
|
762
|
+
PUSH_FAILED = re.compile(r'^\s*(?:fatal|error):|\[rejected\]|\[remote rejected\]|failed to push|could not read (?:Username|Password)|Authentication failed|Permission denied|Repository not found|could not resolve host|HTTP 4\d\d|GraphQL: |no commits between|pull request create failed|not mergeable', re.I | re.M)
|
|
763
|
+
PUSH_REF = re.compile(r'^\s*[-+*=!]?\s*(?:\[new branch\]|\[new tag\]|[0-9a-f]{7,}\.\.\.?[0-9a-f]{7,})?\s*\S+\s+->\s+\S+|Everything up-to-date|https://\S+/pull/\d+|\b(?:Merged|merged) pull request\b|\bwas already merged\b', re.M)
|
|
764
|
+
|
|
765
|
+
def push_succeeded(command, output):
|
|
766
|
+
"""No failure in the output, and either a pushed ref / PR URL in it or a quiet command that printed nothing."""
|
|
767
|
+
text = str(output or '')
|
|
768
|
+
if PUSH_FAILED.search(text): return False
|
|
769
|
+
if PUSH_REF.search(text): return True
|
|
770
|
+
# `git push -q` prints nothing on success; anything else that printed but shows no ref did not confirm the push.
|
|
771
|
+
return not re.sub(r'Command (?:executed|ran) successfully\.?', '', text).strip()
|
|
772
|
+
|
|
773
|
+
CI_FILES = ('.github/workflows', '.gitlab-ci.yml', '.circleci', 'azure-pipelines.yml', 'bitbucket-pipelines.yml', '.buildkite', '.travis.yml', 'Jenkinsfile', '.drone.yml', 'vercel.json', 'netlify.toml')
|
|
774
|
+
|
|
775
|
+
def has_ci(repo):
|
|
776
|
+
"""True when the repository configures a CI or deploy check that would report on a push."""
|
|
777
|
+
root = Path(repo or '.')
|
|
778
|
+
for name in CI_FILES:
|
|
779
|
+
p = root / name
|
|
780
|
+
try:
|
|
781
|
+
if p.is_file() or (p.is_dir() and any(p.iterdir())): return True
|
|
782
|
+
except OSError: continue
|
|
783
|
+
return False
|
|
784
|
+
|
|
642
785
|
PROGRESS_KEYS = ('last_edit', 'last_check', 'last_failed_check', 'edits')
|
|
643
786
|
|
|
644
787
|
# Commands the harness never replays on its own: builds that share the live preview's output folder (a Next.js build under a
|
|
@@ -692,6 +835,16 @@ def recheck_stale(state, env, seconds_left, limit=3):
|
|
|
692
835
|
|
|
693
836
|
SOFT_ISSUE = re.compile(r'^(Code changed \(|You changed code \(|Your todo list still has open items)', re.I)
|
|
694
837
|
|
|
838
|
+
def plan_turn_ok(report):
|
|
839
|
+
"""A read-only turn with no changed files and no failed check (BUG-0241)."""
|
|
840
|
+
return not (report.get('changedFiles') or []) and not any(int(report.get(k) or 0) for k in ('checksFailed', 'checksInconclusive', 'externalChecksFailed'))
|
|
841
|
+
|
|
842
|
+
def plan_turn_report(report):
|
|
843
|
+
out = dict(report, unfinishedTasks=[])
|
|
844
|
+
out['verification'] = re.sub(r' \d+ plan items? remains? unfinished\.', '', str(report.get('verification') or ''))
|
|
845
|
+
if out.get('verificationStatus') == 'needs_attention': out['verificationStatus'] = 'passed' if int(report.get('checksPassed') or 0) else 'unverified'
|
|
846
|
+
return out
|
|
847
|
+
|
|
695
848
|
def soft_unverified(report, unresolved):
|
|
696
849
|
"""Nothing failed: the only gaps are checks that predate the last edit (and could not be replayed) or open task-list
|
|
697
850
|
items. The work is shown as done with a note, not as an interrupted run with a Resume button."""
|
|
@@ -734,7 +887,13 @@ def workspace_note(config, event):
|
|
|
734
887
|
_, rows = dev_jobs.reconcile(ROOT, REPO, event, env)
|
|
735
888
|
note += dev_jobs.context_note(rows)
|
|
736
889
|
except Exception as exc: note += 'Job supervisor unavailable: ' + str(exc)[:200] + '\n'
|
|
737
|
-
|
|
890
|
+
laptop = bool(config.get('laptop')) or os.getenv('MINI_LAPTOP') == '1'
|
|
891
|
+
if laptop and config.get('conversation_route', 'build') == 'build':
|
|
892
|
+
# BUG-0550: on the user's own computer a repository's .mlai/setup.sh, devcontainer command or install scripts
|
|
893
|
+
# (npm postinstall) never run automatically: they come from the repo, not the user. Installs go through the
|
|
894
|
+
# agent's normal commands, which the user approves.
|
|
895
|
+
note += 'Project dependencies were not installed automatically on this computer. If the task needs them, run the install command as a normal step (it asks for approval when needed).\n'
|
|
896
|
+
elif config.get('conversation_route', 'build') == 'build' and os.getenv('MINI_ENV_SETUP', '1') != '0':
|
|
738
897
|
try:
|
|
739
898
|
import env_setup
|
|
740
899
|
note += env_setup.ensure(ROOT, REPO, event, env)
|
|
@@ -794,6 +953,14 @@ def build_context(config):
|
|
|
794
953
|
if config.get('conversation_history'):
|
|
795
954
|
task = 'Recent conversation in this workspace (context only; the current request takes priority):\n' + json.dumps(config['conversation_history']) + '\n\n' + task
|
|
796
955
|
if config.get('_handover'): task = config['_handover'] + task # model switch: brief sits right above the recent conversation
|
|
956
|
+
restored = load(ROOT, 'restored.json', None) # BUG-0598: the user restored a checkpoint since the last run
|
|
957
|
+
if isinstance(restored, dict) and restored.get('at'):
|
|
958
|
+
label = str(restored.get('label') or 'an earlier checkpoint')[:160]
|
|
959
|
+
task = ('Files restored: after the last reply, the user restored the project to ' + label + ' (' + str(restored['at'])[:40] + '). '
|
|
960
|
+
'Changes made after that point were undone. Earlier replies in this chat that describe later changes no longer match the files; '
|
|
961
|
+
'read the files before you say what the project contains.\n\n' + task)
|
|
962
|
+
try: (ROOT / 'restored.json').unlink()
|
|
963
|
+
except OSError: pass
|
|
797
964
|
if not readonly and not config.get('resuming') and config.get('conversation_route', 'build') == 'build':
|
|
798
965
|
try:
|
|
799
966
|
guidance = planning_guidance(config['task'], REPO, config.get('mode', 'approve'), phase1=role_engine.phase1(config.get('_role')))
|
|
@@ -811,6 +978,7 @@ def build_context(config):
|
|
|
811
978
|
try:
|
|
812
979
|
import mlai_hooks
|
|
813
980
|
if mlai_hooks.active(): task = mlai_hooks.guidance() + '\n' + task # project hooks and hard-deny rules
|
|
981
|
+
if config.get('laptop') or os.getenv('MINI_LAPTOP') == '1': task = mlai_hooks.LAPTOP_RULES.replace('call cannot_complete', 'call mcp__mlai__cannot_complete') + '\n' + task # BUG-0523/0529/0531
|
|
814
982
|
except Exception: pass
|
|
815
983
|
try: # CLI custom subagents and opt-in project memory
|
|
816
984
|
from subagents import agents_guidance
|
|
@@ -869,7 +1037,9 @@ def mode_rules(mode, ask_first=False):
|
|
|
869
1037
|
'(naming, file layout, colours) or that the request already answers.\nOnce they agree, continue in the user\'s mode:\n' + after)
|
|
870
1038
|
if mode == 'auto':
|
|
871
1039
|
return ('Mode: Auto. Edits and ordinary commands run without approval. Still refused or asked: recursive deletes, history rewrites, force pushes, data drops, '
|
|
872
|
-
'sudo beyond package installs and services, and anything touching the harness state. Pick contained alternatives instead of asking.'
|
|
1040
|
+
'sudo beyond package installs and services, and anything touching the harness state. Pick contained alternatives instead of asking. '
|
|
1041
|
+
'Auto runs without approvals, so keep to the request: do only what the user asked, the smallest change that does it. Ask before any extra work '
|
|
1042
|
+
'(new features, tools or files, package.json name or "private" changes, publishing, global installs, pushes or PRs the user did not ask for).')
|
|
873
1043
|
if mode == 'brief':
|
|
874
1044
|
return ('Mode: Brief. Nothing is built until the user approves your task list. First read the repository (read-only commands, file reads, searches). Then clarify: record one to three short questions about what the request leaves open '
|
|
875
1045
|
'(style, content, images, scope) with mcp__mlai__brief_questions, each with your recommended answer. Then call TodoList with 3 to 10 concrete, '
|
|
@@ -968,6 +1138,8 @@ def auto_continue(config, state, unresolved, reason, readonly, goal, plan_on_hol
|
|
|
968
1138
|
import wakeups
|
|
969
1139
|
if config.get('subagent') or readonly or state.get('inspect') or goal or reason == 'stopped': return None
|
|
970
1140
|
if os.getenv('MINI_VERIFY_RESUMES', '2') == '0': return None
|
|
1141
|
+
# BUG-0557: a one-shot run (mlai -p) exits after this turn; nothing would continue it, so never promise it.
|
|
1142
|
+
if os.getenv('MINI_ONE_SHOT') == '1': return None
|
|
971
1143
|
try:
|
|
972
1144
|
if plan_on_hold(): return None
|
|
973
1145
|
except Exception: pass
|
|
@@ -1011,6 +1183,7 @@ def main():
|
|
|
1011
1183
|
if config.get('wake'): config['task'] = goals.wake_prompt(config['wake'], goals.read(ROOT))
|
|
1012
1184
|
goal = goals.start(ROOT, config, event)
|
|
1013
1185
|
global KIMI_VERSION
|
|
1186
|
+
os.environ['MINI_MODEL_LABEL'] = str(config.get('modelLabel') or ''); os.environ['MINI_MODEL_ID'] = str(config.get('modelId') or '')
|
|
1014
1187
|
stage('Preparing the coding engine')
|
|
1015
1188
|
try: ensure_runtime()
|
|
1016
1189
|
except Exception as exc:
|
|
@@ -1045,11 +1218,20 @@ def main():
|
|
|
1045
1218
|
event('finished', state='completed')
|
|
1046
1219
|
return
|
|
1047
1220
|
start_handover(config)
|
|
1221
|
+
servers_note = ''
|
|
1222
|
+
if not config.get('subagent'):
|
|
1223
|
+
try:
|
|
1224
|
+
import server_memory # BUG-0203: this chat moved to a new computer; say which servers stopped
|
|
1225
|
+
gone = server_memory.stopped(ROOT)
|
|
1226
|
+
if gone: event('servers_stopped', servers=gone, content='This chat is on a new cloud computer. Servers that were running before have stopped.')
|
|
1227
|
+
servers_note = server_memory.agent_note(ROOT)
|
|
1228
|
+
except Exception: pass
|
|
1048
1229
|
if honor_no_change(config, shell=SHELL_ALWAYS): event('note', content=NO_CHANGE_NOTE)
|
|
1049
1230
|
if conversation_enabled(config):
|
|
1050
1231
|
stage('Reading your request')
|
|
1051
1232
|
selected = restore_choice(config, ROOT)
|
|
1052
1233
|
ask_first.apply(config, REPO) # the harness decides whether this run asks first, for every client
|
|
1234
|
+
if config.get('replaces') and supersede(ROOT, config['replaces']): event('note', content='The earlier version of this message was replaced; the agent continues from the new one.') # BUG-0460
|
|
1053
1235
|
history = recent_history(ROOT) or load(ROOT, 'imported-context.json', {}).get('messages', [])
|
|
1054
1236
|
config['conversation_history'] = history
|
|
1055
1237
|
if selected:
|
|
@@ -1068,12 +1250,14 @@ def main():
|
|
|
1068
1250
|
write_home(config, proxy, config.get('mode', 'approve'), readonly, steps)
|
|
1069
1251
|
extra = ''
|
|
1070
1252
|
talk = conversation_enabled(config) and not readonly and config.get('conversation_route') in ('chat', 'research', 'plan')
|
|
1253
|
+
if servers_note: extra += servers_note + '\n'
|
|
1071
1254
|
if config.get('inspect') and not readonly: extra += INSPECT_RULES + '\n'
|
|
1072
1255
|
if talk: extra += TALK_RULES.replace('{route}', config['conversation_route']) + ('\n' + TALK_RESEARCH if config['conversation_route'] == 'research' else '') + '\n'
|
|
1073
1256
|
if config.get('require_changes') and not readonly: extra += 'This task explicitly requires a repository patch. An unchanged repository is not completion. If you cannot fix it, report the blocker honestly.\n'
|
|
1074
1257
|
if config.get('selected_design') and not readonly: extra += DESIGN_BUILD_GUIDANCE + '\n'
|
|
1075
1258
|
if config.get('design_in_brief') and brief_first(config) and not readonly: extra += DESIGN_BRIEF_GUIDANCE + '\n'
|
|
1076
1259
|
from connector_search import compact_catalog, read_only as read_only_tools
|
|
1260
|
+
import connected_turn
|
|
1077
1261
|
offered = read_only_tools(config.get('tools', [])) if readonly else config.get('tools', [])
|
|
1078
1262
|
if offered:
|
|
1079
1263
|
catalog, compacted = compact_catalog(offered)
|
|
@@ -1093,35 +1277,49 @@ def main():
|
|
|
1093
1277
|
from reply_language import turn_note
|
|
1094
1278
|
language_note = turn_note(config)
|
|
1095
1279
|
image_note = attached_image_note(config)
|
|
1280
|
+
if image_note and 'image_in' not in model_capabilities(config) and not config.get('subagent'):
|
|
1281
|
+
event('vision_unavailable', model=str(config.get('modelId') or ''), content=str(config.get('modelLabel') or 'This model') + ' reads text only and cannot see the attached image.') # BUG-0230
|
|
1096
1282
|
if image_note: task += '\n\n' + image_note
|
|
1097
1283
|
if language_note: task += '\n\n' + language_note
|
|
1098
1284
|
base = subprocess.run(['git', 'rev-parse', 'HEAD'], cwd=REPO, capture_output=True, text=True).stdout.strip() or None
|
|
1099
1285
|
start_dirty = None
|
|
1100
|
-
|
|
1101
|
-
|
|
1102
|
-
|
|
1103
|
-
|
|
1104
|
-
|
|
1105
|
-
|
|
1286
|
+
(ROOT / 'start-dirty.json').unlink(missing_ok=True)
|
|
1287
|
+
# Files already changed or untracked before the turn are the user's own work, not something this turn did
|
|
1288
|
+
# (no-changes turns; BUG-0495/0497: the checker and the changed-files summary; BUG-0558: changedFiles lists only what this turn changed).
|
|
1289
|
+
# Read-only turns too: a question in a folder with uncommitted work must not report that work as this turn's changes.
|
|
1290
|
+
try:
|
|
1291
|
+
from runtime_control import save_start_dirty
|
|
1292
|
+
start_dirty = save_start_dirty(ROOT, REPO, base)
|
|
1293
|
+
except Exception: start_dirty = None
|
|
1106
1294
|
from conversation import research_required
|
|
1107
1295
|
evidence_key = str(config.get('runId') or config.get('run_id') or hashlib.sha256((config['task'] + str(time.time())).encode()).hexdigest()[:16])
|
|
1108
1296
|
role_id = (config.get('_role') or {}).get('id') if isinstance(config.get('_role'), dict) else config.get('_role')
|
|
1109
1297
|
state = {'mode': config.get('mode', 'approve'), 'brief': config.get('mode') == 'brief' or brief_first(config), 'brief_return': config.get('mode', 'approve') if brief_first(config) else None, 'readonly': readonly, 'inspect': bool(config.get('inspect')) and not readonly, 'deadline': time.time() + WALL_SECONDS, 'land_at': time.time() + WALL_SECONDS - LAND_SECONDS,
|
|
1110
|
-
'
|
|
1298
|
+
'step_limit': steps, # BUG-0589: enforced per command/edit in kimi_hook.step_limit_gate
|
|
1299
|
+
'require_changes': bool(config.get('require_changes')), 'task': config['task'][:4000], 'role': config.get('_role'), 'base': base, 'edits': 0, 'started': time.time(),
|
|
1111
1300
|
'plan_approval': config.get('planApproval') if config.get('planApproval') in ('always', 'guided', 'off') else 'guided',
|
|
1112
1301
|
'large': large, 'selected_design': bool(config.get('selected_design')) and not readonly, 'look': bool(config.get('design_in_brief')) and brief_first(config) and not readonly,
|
|
1113
|
-
'research': bool(research_required(config)), 'evidence_key': evidence_key,
|
|
1302
|
+
'research': bool(research_required(config)), 'evidence_key': evidence_key, 'quick': bool(config.get('quick_question')),
|
|
1114
1303
|
'account_left': account_left(config),
|
|
1115
1304
|
# BUG-0127: Guided promises "approve the plan": a build turn shows its task list before the first edit.
|
|
1116
|
-
|
|
1117
|
-
|
|
1305
|
+
# BUG-0590: "Approve the task list first: Always" waits for the plan in Auto too.
|
|
1306
|
+
'plan_first': ((config.get('mode', 'approve') == 'approve' and config.get('planApproval', 'guided') in ('guided', 'always'))
|
|
1307
|
+
or (config.get('mode', 'approve') == 'auto' and config.get('planApproval') == 'always')) and not readonly
|
|
1308
|
+
and not config.get('resuming') and not config.get('go') and not config.get('subagent') and config.get('conversation_route', 'build') in ('build', 'chat', 'research'),
|
|
1309
|
+
# BUG-0345: a pure question (chat/research route) skips the plan review; an edit in that turn still waits for it.
|
|
1310
|
+
'question_turn': config.get('conversation_route', 'build') in ('chat', 'research') and not brief_first(config)}
|
|
1118
1311
|
save_state(state, static=True)
|
|
1119
1312
|
for stale in ('scope.json', 'kimi-commands.json', 'verification.json', 'kimi-last-text.json'): (ROOT / stale).unlink(missing_ok=True)
|
|
1120
|
-
env =
|
|
1313
|
+
env = engine_env(os.environ)
|
|
1121
1314
|
env.update(VAULT)
|
|
1122
|
-
env.update({'KIMI_CODE_HOME': str(HOME), 'KIMI_DISABLE_TELEMETRY': '1', '
|
|
1123
|
-
|
|
1315
|
+
env.update({'KIMI_CODE_HOME': str(HOME), 'KIMI_DISABLE_TELEMETRY': '1', 'MINI_ROOT': str(ROOT), 'MINI_REPO': REPO, 'NO_COLOR': '1', 'CI': '1'})
|
|
1316
|
+
# BUG-0460: after an Edit or Retry the old session still holds the replaced turn, so start fresh from the history.
|
|
1317
|
+
prior = None if config.get('_handover') or config.get('replaces') else resumable_session(readonly, role_id, handover.model_key(config))
|
|
1124
1318
|
if prior:
|
|
1319
|
+
try:
|
|
1320
|
+
sync = connected_turn.resume_note((load(ROOT, 'kimi-session.json', {}) or {}).get('connected'), offered, config.get('connected'))
|
|
1321
|
+
if sync: task = sync + '\n\n' + task; event('note', content='Connected apps changed since the last message; the agent has the current list')
|
|
1322
|
+
except Exception: pass # never blocks a run
|
|
1125
1323
|
cmd = [*kimi_argv(), '-S', prior, '-p', task, '--output-format', 'stream-json']
|
|
1126
1324
|
event('note', content='Continuing the previous engine session')
|
|
1127
1325
|
else: cmd = [*kimi_argv(), '-p', task, '--output-format', 'stream-json', '--agent-file', str(agent)]
|
|
@@ -1150,7 +1348,14 @@ def main():
|
|
|
1150
1348
|
except Exception: pass
|
|
1151
1349
|
if stopped(): killed['reason'] = 'stopped'
|
|
1152
1350
|
elif time.time() > state['deadline']: killed['reason'] = 'limited'; killed['limit'] = 'time'
|
|
1153
|
-
|
|
1351
|
+
# CLI --max-turns: the whole run (repair rounds included) stops once the agent asked for more model
|
|
1352
|
+
# steps than allowed (BUG-0492).
|
|
1353
|
+
elif config.get('hardSteps') and proxy.calls > steps: killed['reason'] = 'limited'; killed['limit'] = 'steps'
|
|
1354
|
+
# BUG-0519: a -p run in auto mode has nobody to approve a write the OS sandbox refused; it used to
|
|
1355
|
+
# keep calling the model until the time limit.
|
|
1356
|
+
elif sandbox_stuck(config): killed['reason'] = 'limited'; killed['limit'] = 'sandbox'
|
|
1357
|
+
# CLI --max-budget-credits (hardBudget): the caller's own cap gets no grace past it (BUG-0509).
|
|
1358
|
+
elif state.get('account_left') is not None and (totals()['tokens'] > state['account_left'] + (0 if config.get('hardBudget') else ACCOUNT_GRACE) or time.time() - float((load(ROOT, 'kimi-hook.json', {}) or {}).get('account_stop_at') or 1e18) > ACCOUNT_WRAP_SECONDS):
|
|
1154
1359
|
# The hook already asked the agent to wrap up at the allowance; this is the hard stop.
|
|
1155
1360
|
event('account_limit', phase='stopped', content=CREDITS_OUT_NOTE); killed['reason'] = 'limited'; killed['limit'] = 'tokens'
|
|
1156
1361
|
elif time.time() - (load(ROOT, 'kimi-stopped', {}) or {}).get('time', 1e18) > FINISH_GRACE: killed['reason'] = 'finished'
|
|
@@ -1176,7 +1381,7 @@ def main():
|
|
|
1176
1381
|
code = child.wait(); stderr.close()
|
|
1177
1382
|
(ROOT / 'child.json').unlink(missing_ok=True)
|
|
1178
1383
|
if not mapper.session: mapper.session = newest_session(launched)
|
|
1179
|
-
if mapper.session: write('kimi-session.json', {'session': mapper.session, 'time': time.time(), 'readonly': readonly, 'role': role_id, 'model': handover.model_key(config)})
|
|
1384
|
+
if mapper.session: write('kimi-session.json', {'session': mapper.session, 'time': time.time(), 'readonly': readonly, 'role': role_id, 'model': handover.model_key(config), 'connected': connected_turn.fingerprint(offered)})
|
|
1180
1385
|
if killed.get('limit'): state['limit_kind'] = killed['limit']
|
|
1181
1386
|
return code, killed['reason']
|
|
1182
1387
|
def resume(prompt): return [*kimi_argv(), '-S', mapper.session, '-p', prompt + ('\n\n' + language_note if language_note else ''), '--output-format', 'stream-json']
|
|
@@ -1245,7 +1450,7 @@ def main():
|
|
|
1245
1450
|
if checkpoints:
|
|
1246
1451
|
try: checkpoints.flush()
|
|
1247
1452
|
except Exception: pass
|
|
1248
|
-
answer = mapper.last_text
|
|
1453
|
+
answer = final_answer(mapper.last_text, getattr(mapper, 'texts', []))
|
|
1249
1454
|
if answer and state.get('_acceptance_note') and not unresolved: answer += '\n\n' + state['_acceptance_note']
|
|
1250
1455
|
if reason == 'finished': reason = None
|
|
1251
1456
|
continuing = auto_continue(config, state, unresolved, reason, readonly, goal, plan_on_hold)
|
|
@@ -1262,6 +1467,8 @@ def main():
|
|
|
1262
1467
|
if killed['reason'] == 'limited':
|
|
1263
1468
|
limit = state.get('limit_kind') or 'time'
|
|
1264
1469
|
limit_text = (CREDITS_OUT_NOTE if limit == 'tokens'
|
|
1470
|
+
else 'Stopped after %d agent step%s (the --max-turns limit); your files are saved.' % (steps, '' if steps == 1 else 's') if limit == 'steps'
|
|
1471
|
+
else SANDBOX_STOP_NOTE if limit == 'sandbox'
|
|
1265
1472
|
else 'Each run can work for up to %d minutes. This one reached that limit and stopped; your files are saved and %s.' % (max(1, WALL_SECONDS // 60), 'the work continues automatically in a minute' if continuing else 'Resume continues from them'))
|
|
1266
1473
|
result = {'exit_status': 'LimitsExceeded', 'limit': limit, 'limitNote': limit_text, 'submission': answer or limit_text}
|
|
1267
1474
|
run_state = 'limited'
|
|
@@ -1295,10 +1502,20 @@ def main():
|
|
|
1295
1502
|
if tidied: event('note', content=tidied)
|
|
1296
1503
|
except Exception: pass
|
|
1297
1504
|
report = completion_report(ROOT, REPO, state.get('base'))
|
|
1505
|
+
if start_dirty is not None and report.get('changedFiles'):
|
|
1506
|
+
# BUG-0558: the user's own uncommitted edits from before the turn are not this turn's changes.
|
|
1507
|
+
try:
|
|
1508
|
+
from runtime_control import changed_during_turn
|
|
1509
|
+
report['changedFiles'] = changed_during_turn(REPO, report['changedFiles'], start_dirty)
|
|
1510
|
+
except Exception: pass
|
|
1298
1511
|
try:
|
|
1299
1512
|
import work_ledger
|
|
1300
1513
|
work_ledger.record(ROOT, config.get('task', ''), locals().get('stated') or answer or '', load(ROOT, 'checks.json', []), load(ROOT, 'design-browser-evidence.json', []), run_state)
|
|
1301
1514
|
except Exception: pass
|
|
1515
|
+
if run_state == 'completed' and readonly and plan_turn_ok(report):
|
|
1516
|
+
# BUG-0241: a Plan (read-only) turn that changed nothing is a plan, not unfinished work. Its task list is
|
|
1517
|
+
# what the build will do later, so it never ends as Verification incomplete with a Resume that cannot help.
|
|
1518
|
+
unresolved = []; report = plan_turn_report(report)
|
|
1302
1519
|
if run_state == 'completed' and (unresolved or report.get('verificationStatus') == 'needs_attention') and soft_unverified(report, unresolved):
|
|
1303
1520
|
# Done, with an honest note; the review card still lists the open steps and out-of-date checks.
|
|
1304
1521
|
report['verificationStatus'] = 'needs_attention'; state['_soft_unverified'] = True
|
|
@@ -1332,6 +1549,9 @@ def main():
|
|
|
1332
1549
|
ticked = sync_plan_progress(REPO, load(ROOT, 'todos.json', []), report.get('verificationStatus') != 'needs_attention')
|
|
1333
1550
|
if ticked: event('note', content='Marked %d finished deliverable%s in PLAN.md' % (ticked, '' if ticked == 1 else 's'))
|
|
1334
1551
|
except Exception: pass
|
|
1552
|
+
if answer:
|
|
1553
|
+
filled = fill_dangling_output(answer, LAST_OUTPUT)
|
|
1554
|
+
if filled != answer: answer = filled; result['submission'] = answer
|
|
1335
1555
|
if answer: event('answer', content=answer)
|
|
1336
1556
|
open_questions = unshown_questions()
|
|
1337
1557
|
if open_questions: event('questions', questions=open_questions)
|