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
|
@@ -0,0 +1,53 @@
|
|
|
1
|
+
"""One question to the model, with or without a lookup budget. Shared by every agent that
|
|
2
|
+
asks: specialists reviewing the change, and the refuter checking their claims."""
|
|
3
|
+
|
|
4
|
+
from __future__ import annotations
|
|
5
|
+
|
|
6
|
+
import logging
|
|
7
|
+
from typing import Any
|
|
8
|
+
|
|
9
|
+
from mlai_review.application import messages, prompts
|
|
10
|
+
from mlai_review.application.ports import LanguageModel, Message
|
|
11
|
+
from mlai_review.application.tools import LookupBudget, WorkspaceTools
|
|
12
|
+
|
|
13
|
+
_log = logging.getLogger("mlai.orchestrator.ask")
|
|
14
|
+
|
|
15
|
+
|
|
16
|
+
async def ask(
|
|
17
|
+
model: LanguageModel,
|
|
18
|
+
*,
|
|
19
|
+
system: str,
|
|
20
|
+
user: str,
|
|
21
|
+
tools: WorkspaceTools | None,
|
|
22
|
+
budget: LookupBudget | None,
|
|
23
|
+
who: str,
|
|
24
|
+
) -> dict[str, Any]:
|
|
25
|
+
"""The model's JSON answer. With tools and a budget, the model may look around first; the
|
|
26
|
+
budget bounds how far, after which it must answer without tools."""
|
|
27
|
+
if tools is None or budget is None:
|
|
28
|
+
return await model.complete_json(system=system, user=user)
|
|
29
|
+
conversation = [Message("system", system + prompts.LOOKUPS), Message("user", user)]
|
|
30
|
+
calls = spent = 0
|
|
31
|
+
nudge = prompts.LOOKUPS_EXHAUSTED
|
|
32
|
+
while calls < budget.max_calls and spent < budget.max_result_chars:
|
|
33
|
+
turn = await model.next_turn(tuple(conversation), tools.specs)
|
|
34
|
+
if turn.answer is not None:
|
|
35
|
+
_log.info(messages.LOOKUPS_USED, who, calls, spent)
|
|
36
|
+
return turn.answer
|
|
37
|
+
if not turn.tool_calls:
|
|
38
|
+
# It answered in prose instead of JSON: keep its words, ask for the object.
|
|
39
|
+
conversation.append(Message("assistant", turn.text))
|
|
40
|
+
nudge = prompts.ANSWER_AS_JSON
|
|
41
|
+
break
|
|
42
|
+
conversation.append(Message("assistant", tool_calls=turn.tool_calls))
|
|
43
|
+
for call in turn.tool_calls:
|
|
44
|
+
result = await tools.call(call)
|
|
45
|
+
calls += 1
|
|
46
|
+
spent += len(result)
|
|
47
|
+
conversation.append(Message("tool", result, tool_call_id=call.id))
|
|
48
|
+
# No more tools on offer, so the model must answer now, in JSON mode.
|
|
49
|
+
conversation.append(Message("user", nudge))
|
|
50
|
+
turn = await model.next_turn(tuple(conversation), ())
|
|
51
|
+
assert turn.answer is not None
|
|
52
|
+
_log.info(messages.LOOKUPS_USED, who, calls, spent)
|
|
53
|
+
return turn.answer
|
|
@@ -0,0 +1,142 @@
|
|
|
1
|
+
"""The refuter: a second opinion that tries to disprove each candidate finding before the
|
|
2
|
+
publication policy sees it. A claim the refuter breaks is not shown; one it cannot break is
|
|
3
|
+
shown at a lower confidence bar than an unchecked one."""
|
|
4
|
+
|
|
5
|
+
from __future__ import annotations
|
|
6
|
+
|
|
7
|
+
import asyncio
|
|
8
|
+
import logging
|
|
9
|
+
from abc import ABC, abstractmethod
|
|
10
|
+
from dataclasses import replace
|
|
11
|
+
from typing import Any
|
|
12
|
+
|
|
13
|
+
from mlai_review.application import messages, prompts
|
|
14
|
+
from mlai_review.application.asking import ask
|
|
15
|
+
from mlai_review.application.errors import ReviewError
|
|
16
|
+
from mlai_review.application.ports import LanguageModel, Workspace
|
|
17
|
+
from mlai_review.application.tools import LookupBudget, WorkspaceTools
|
|
18
|
+
from mlai_review.domain.context import ReviewContext
|
|
19
|
+
from mlai_review.domain.findings import (
|
|
20
|
+
VALID_GROUNDS,
|
|
21
|
+
Challenge,
|
|
22
|
+
ChallengeOutcome,
|
|
23
|
+
Finding,
|
|
24
|
+
RefutationGround,
|
|
25
|
+
)
|
|
26
|
+
|
|
27
|
+
_log = logging.getLogger("mlai.orchestrator.refuter")
|
|
28
|
+
|
|
29
|
+
# A refutation the refuter itself doubts counts as "unsure", not as a refutation.
|
|
30
|
+
MIN_REFUTATION_CONFIDENCE = 0.5
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
class Challenger(ABC):
|
|
34
|
+
@abstractmethod
|
|
35
|
+
async def one(
|
|
36
|
+
self, context: ReviewContext, finding: Finding, workspace: Workspace | None
|
|
37
|
+
) -> Finding:
|
|
38
|
+
"""The finding, carrying the challenge's outcome where one was made."""
|
|
39
|
+
|
|
40
|
+
async def challenge(
|
|
41
|
+
self, context: ReviewContext, findings: list[Finding], workspace: Workspace | None
|
|
42
|
+
) -> list[Finding]:
|
|
43
|
+
return list(await asyncio.gather(*(self.one(context, f, workspace) for f in findings)))
|
|
44
|
+
|
|
45
|
+
|
|
46
|
+
class NoChallenger(Challenger):
|
|
47
|
+
async def one(
|
|
48
|
+
self, context: ReviewContext, finding: Finding, workspace: Workspace | None
|
|
49
|
+
) -> Finding:
|
|
50
|
+
return finding
|
|
51
|
+
|
|
52
|
+
|
|
53
|
+
class LlmChallenger(Challenger):
|
|
54
|
+
"""One model call per finding, all in parallel. A claim is about one place, so the refuter
|
|
55
|
+
sees the files it is about (its own, and any other changed file it names), not the whole
|
|
56
|
+
change: re-sending a 40,000-token change for every claim made the refuter most of a large
|
|
57
|
+
review's input. Claims about the same file share a prompt prefix, which providers cache. A
|
|
58
|
+
call that fails leaves its finding unchallenged: the policy then treats it as it would
|
|
59
|
+
without a refuter."""
|
|
60
|
+
|
|
61
|
+
def __init__(self, model: LanguageModel, lookups: LookupBudget | None = None) -> None:
|
|
62
|
+
self._model = model
|
|
63
|
+
self._lookups = lookups
|
|
64
|
+
|
|
65
|
+
async def one(
|
|
66
|
+
self, context: ReviewContext, finding: Finding, workspace: Workspace | None
|
|
67
|
+
) -> Finding:
|
|
68
|
+
tools = (
|
|
69
|
+
WorkspaceTools(workspace, self._lookups)
|
|
70
|
+
if workspace is not None and self._lookups is not None
|
|
71
|
+
else None
|
|
72
|
+
)
|
|
73
|
+
try:
|
|
74
|
+
scope = evidence_scope(context, finding)
|
|
75
|
+
answer = await self._challenge_one(prompts.user_message(scope), finding, tools)
|
|
76
|
+
except ReviewError as exc:
|
|
77
|
+
_log.warning(messages.CHALLENGE_FAILED, finding.title, exc)
|
|
78
|
+
return finding
|
|
79
|
+
challenge = _parse(answer)
|
|
80
|
+
_log.info(
|
|
81
|
+
messages.CHALLENGE_OUTCOME, challenge.outcome.value, finding.title, challenge.reason
|
|
82
|
+
)
|
|
83
|
+
return replace(finding, challenge=challenge)
|
|
84
|
+
|
|
85
|
+
async def _challenge_one(
|
|
86
|
+
self, user: str, finding: Finding, tools: WorkspaceTools | None
|
|
87
|
+
) -> dict[str, Any]:
|
|
88
|
+
return await ask(
|
|
89
|
+
self._model,
|
|
90
|
+
system=prompts.SHARED,
|
|
91
|
+
user=user + prompts.refute_task(finding),
|
|
92
|
+
tools=tools,
|
|
93
|
+
budget=self._lookups,
|
|
94
|
+
who="refuter",
|
|
95
|
+
)
|
|
96
|
+
|
|
97
|
+
|
|
98
|
+
# A change this small is sent whole: narrowing it saves little and loses context.
|
|
99
|
+
WHOLE_CHANGE_FILES = 3
|
|
100
|
+
MAX_SCOPE_FILES = 4
|
|
101
|
+
|
|
102
|
+
|
|
103
|
+
def evidence_scope(context: ReviewContext, finding: Finding) -> ReviewContext:
|
|
104
|
+
"""The part of the change a claim can be judged on: the file it points at, plus changed
|
|
105
|
+
files its words name (by file name), up to a few."""
|
|
106
|
+
files = context.changes.files
|
|
107
|
+
if len(files) <= WHOLE_CHANGE_FILES:
|
|
108
|
+
return context
|
|
109
|
+
told = f"{finding.title} {finding.explanation}".lower()
|
|
110
|
+
named = [
|
|
111
|
+
f.path for f in files if f.path != finding.path and _stem(f.path) and _stem(f.path) in told
|
|
112
|
+
]
|
|
113
|
+
paths = (finding.path, *named[: MAX_SCOPE_FILES - 1])
|
|
114
|
+
return context.part(paths)
|
|
115
|
+
|
|
116
|
+
|
|
117
|
+
def _stem(path: str) -> str:
|
|
118
|
+
""" "src/menu/menu.service.ts" -> "menu.service"; short stems ("a", "index") name nothing."""
|
|
119
|
+
name = path.rsplit("/", 1)[-1]
|
|
120
|
+
stem = name.rsplit(".", 1)[0].lower() if "." in name else name.lower()
|
|
121
|
+
return stem if len(stem) >= 6 and stem not in {"index", "module", "service", "utils"} else ""
|
|
122
|
+
|
|
123
|
+
|
|
124
|
+
def _parse(answer: dict[str, Any]) -> Challenge:
|
|
125
|
+
"""The refuter's verdict as typed values; anything unusable is 'unsure', never a
|
|
126
|
+
refutation, so a garbled answer cannot hide a real finding. A refutation stands only on
|
|
127
|
+
a valid ground: "the bug was already there" does not disprove a finding."""
|
|
128
|
+
try:
|
|
129
|
+
outcome = ChallengeOutcome(str(answer.get("verdict", "")).lower())
|
|
130
|
+
confidence = min(max(float(answer.get("confidence", 0.0)), 0.0), 1.0)
|
|
131
|
+
except (TypeError, ValueError):
|
|
132
|
+
return Challenge(ChallengeOutcome.UNSURE, "", 0.0)
|
|
133
|
+
reason = str(answer.get("reason", "")).strip()
|
|
134
|
+
ground: RefutationGround | None = None
|
|
135
|
+
if outcome is ChallengeOutcome.REFUTED:
|
|
136
|
+
try:
|
|
137
|
+
ground = RefutationGround(str(answer.get("ground", "other")).lower())
|
|
138
|
+
except ValueError:
|
|
139
|
+
ground = RefutationGround.OTHER
|
|
140
|
+
if confidence < MIN_REFUTATION_CONFIDENCE or ground not in VALID_GROUNDS:
|
|
141
|
+
outcome = ChallengeOutcome.UNSURE
|
|
142
|
+
return Challenge(outcome, reason, confidence, ground)
|
|
@@ -0,0 +1,194 @@
|
|
|
1
|
+
"""Gathers what a reviewer sees, within a size budget."""
|
|
2
|
+
|
|
3
|
+
from __future__ import annotations
|
|
4
|
+
|
|
5
|
+
import logging
|
|
6
|
+
import re
|
|
7
|
+
|
|
8
|
+
from mlai_review.contracts.review import ReviewPullRequestInput
|
|
9
|
+
from mlai_review.application import messages
|
|
10
|
+
from mlai_review.application.ports import GrepMatch, Outliner, PullRequestReader, Workspace
|
|
11
|
+
from mlai_review.domain.changes import changed_lines, parse_patch
|
|
12
|
+
from mlai_review.domain.context import FileView, ReviewContext
|
|
13
|
+
from mlai_review.domain.diagnostics import Diagnostic, on_changed_lines
|
|
14
|
+
from mlai_review.domain.relations import (
|
|
15
|
+
Relation,
|
|
16
|
+
RelationKind,
|
|
17
|
+
called_names,
|
|
18
|
+
changed_definitions,
|
|
19
|
+
definition_pattern,
|
|
20
|
+
)
|
|
21
|
+
from mlai_review.domain.relevance import Triage, is_source, is_test, triage
|
|
22
|
+
|
|
23
|
+
_log = logging.getLogger("mlai.orchestrator.context")
|
|
24
|
+
|
|
25
|
+
# Budgets for what the reviewer is shown. A file whose view is over the per-file limit is
|
|
26
|
+
# reviewed from its diff only; once the total is spent, remaining files are too.
|
|
27
|
+
MAX_FILE_CHARS = 60_000
|
|
28
|
+
MAX_TOTAL_CHARS = 150_000
|
|
29
|
+
# Files larger than this are not read at all.
|
|
30
|
+
MAX_SOURCE_BYTES = 2_000_000
|
|
31
|
+
# Related code: how many names to look up, matches to scan, and lines to show.
|
|
32
|
+
MAX_LOOKUP_NAMES = 24
|
|
33
|
+
MAX_MATCHES = 4_000
|
|
34
|
+
# A name with more uses than this is too common to be informative; methods (get, handle,
|
|
35
|
+
# serialize...) share names across classes, so their bar is lower.
|
|
36
|
+
MAX_GENERIC_MATCHES = 60
|
|
37
|
+
MAX_GENERIC_METHOD_MATCHES = 15
|
|
38
|
+
MAX_DEFINITIONS_PER_NAME = 2
|
|
39
|
+
MAX_REFERENCES_PER_NAME = 4
|
|
40
|
+
MAX_RELATED_CHARS = 8_000
|
|
41
|
+
|
|
42
|
+
|
|
43
|
+
class ContextBuilder:
|
|
44
|
+
def __init__(
|
|
45
|
+
self,
|
|
46
|
+
reader: PullRequestReader,
|
|
47
|
+
outliner: Outliner,
|
|
48
|
+
*,
|
|
49
|
+
max_file_chars: int = MAX_FILE_CHARS,
|
|
50
|
+
max_total_chars: int = MAX_TOTAL_CHARS,
|
|
51
|
+
) -> None:
|
|
52
|
+
self._reader = reader
|
|
53
|
+
self._outliner = outliner
|
|
54
|
+
self._max_file_chars = max_file_chars
|
|
55
|
+
self._max_total_chars = max_total_chars
|
|
56
|
+
|
|
57
|
+
async def head(self, target: ReviewPullRequestInput) -> tuple[str, str]:
|
|
58
|
+
return await self._reader.head(target)
|
|
59
|
+
|
|
60
|
+
async def changes(self, target: ReviewPullRequestInput, *, since: str | None = None) -> Triage:
|
|
61
|
+
"""What the pull request changed (or only what changed since a commit), with noise
|
|
62
|
+
files set aside."""
|
|
63
|
+
result = triage(await self._reader.changed_files(target, since=since))
|
|
64
|
+
if result.skipped:
|
|
65
|
+
_log.info(messages.FILES_SKIPPED, len(result.skipped), ", ".join(result.skipped))
|
|
66
|
+
return result
|
|
67
|
+
|
|
68
|
+
async def build(
|
|
69
|
+
self,
|
|
70
|
+
target: ReviewPullRequestInput,
|
|
71
|
+
changes: Triage,
|
|
72
|
+
workspace: Workspace | None,
|
|
73
|
+
diagnostics: tuple[Diagnostic, ...] = (),
|
|
74
|
+
) -> ReviewContext:
|
|
75
|
+
"""The reviewer's view of the change. Without a workspace, it is the diffs alone."""
|
|
76
|
+
info = await self._reader.pull_request(target)
|
|
77
|
+
views: dict[str, FileView] = {}
|
|
78
|
+
if workspace is not None:
|
|
79
|
+
budget = self._max_total_chars
|
|
80
|
+
for change in changes.reviewable.files:
|
|
81
|
+
if change.status == "removed":
|
|
82
|
+
continue
|
|
83
|
+
text = await workspace.read(change.path, max_bytes=MAX_SOURCE_BYTES)
|
|
84
|
+
if text is None:
|
|
85
|
+
continue
|
|
86
|
+
view = FileView.focus(text, change.patch, self._outliner.outline(change.path, text))
|
|
87
|
+
if view.chars > self._max_file_chars or view.chars > budget:
|
|
88
|
+
continue
|
|
89
|
+
views[change.path] = view
|
|
90
|
+
budget -= view.chars
|
|
91
|
+
related = await self._related(workspace, changes, views) if workspace else ()
|
|
92
|
+
return ReviewContext(
|
|
93
|
+
pull_request=info,
|
|
94
|
+
changes=changes.reviewable,
|
|
95
|
+
views=views,
|
|
96
|
+
contents_available=workspace is not None,
|
|
97
|
+
skipped=changes.skipped,
|
|
98
|
+
related=related,
|
|
99
|
+
diagnostics=on_changed_lines(diagnostics, changes.reviewable),
|
|
100
|
+
)
|
|
101
|
+
|
|
102
|
+
async def _related(
|
|
103
|
+
self, workspace: Workspace, changes: Triage, views: dict[str, FileView]
|
|
104
|
+
) -> tuple[Relation, ...]:
|
|
105
|
+
"""Definitions of what the change calls, and uses of what it changes, from one search
|
|
106
|
+
over the checkout. Test files are reviewed on their own and not expanded."""
|
|
107
|
+
callees: list[str] = []
|
|
108
|
+
touched: dict[str, str] = {} # changed definition -> file it lives in
|
|
109
|
+
methods: set[str] = set()
|
|
110
|
+
for change in changes.reviewable.files:
|
|
111
|
+
view = views.get(change.path)
|
|
112
|
+
if view is None or view.outline is None or is_test(change.path):
|
|
113
|
+
continue
|
|
114
|
+
defined_here = {d.name for d in view.outline.definitions}
|
|
115
|
+
methods |= {d.name for d in view.outline.definitions if d.depth > 0}
|
|
116
|
+
for name in changed_definitions(view.outline, changed_lines(change.patch)):
|
|
117
|
+
touched.setdefault(name, change.path)
|
|
118
|
+
added = [d.text for d in parse_patch(change.patch) if d.kind == "+"]
|
|
119
|
+
for name in called_names(added, defined_here=defined_here):
|
|
120
|
+
if name not in callees and name not in touched:
|
|
121
|
+
callees.append(name)
|
|
122
|
+
names = [*touched, *callees][:MAX_LOOKUP_NAMES]
|
|
123
|
+
if not names:
|
|
124
|
+
return ()
|
|
125
|
+
pattern = r"\b(?:" + "|".join(re.escape(n) for n in names) + r")\b"
|
|
126
|
+
matches = await workspace.grep(pattern, limit=MAX_MATCHES)
|
|
127
|
+
return _select_related(matches, touched, callees, set(views), methods)
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
def _select_related(
|
|
131
|
+
matches: tuple[GrepMatch, ...],
|
|
132
|
+
touched: dict[str, str],
|
|
133
|
+
callees: list[str],
|
|
134
|
+
changed_paths: set[str],
|
|
135
|
+
methods: frozenset[str] | set[str] = frozenset(),
|
|
136
|
+
) -> tuple[Relation, ...]:
|
|
137
|
+
by_name: dict[str, list[GrepMatch]] = {}
|
|
138
|
+
for match in matches:
|
|
139
|
+
if not is_source(match.path) or match.text.lstrip()[:1] in ("#", "/", "*"):
|
|
140
|
+
continue # not code, or a comment
|
|
141
|
+
for name in (*touched, *callees):
|
|
142
|
+
if re.search(rf"\b{re.escape(name)}\b", match.text):
|
|
143
|
+
by_name.setdefault(name, []).append(match)
|
|
144
|
+
|
|
145
|
+
def too_common(name: str) -> bool:
|
|
146
|
+
limit = MAX_GENERIC_METHOD_MATCHES if name in methods else MAX_GENERIC_MATCHES
|
|
147
|
+
return len(by_name.get(name, [])) > limit
|
|
148
|
+
|
|
149
|
+
related: list[Relation] = []
|
|
150
|
+
for name in callees:
|
|
151
|
+
if too_common(name):
|
|
152
|
+
continue
|
|
153
|
+
definitions = [
|
|
154
|
+
m
|
|
155
|
+
for m in by_name.get(name, [])
|
|
156
|
+
if m.path not in changed_paths and definition_pattern(name).search(m.text)
|
|
157
|
+
]
|
|
158
|
+
for match in definitions[:MAX_DEFINITIONS_PER_NAME]:
|
|
159
|
+
related.append(
|
|
160
|
+
Relation(RelationKind.DEFINITION, name, match.path, match.line, match.text)
|
|
161
|
+
)
|
|
162
|
+
for name, origin in touched.items():
|
|
163
|
+
if too_common(name):
|
|
164
|
+
continue
|
|
165
|
+
# Uses only, in the same language: another class's method of the same name, or a
|
|
166
|
+
# same-named type in the frontend, is not a use of ours.
|
|
167
|
+
uses = [
|
|
168
|
+
m
|
|
169
|
+
for m in by_name.get(name, [])
|
|
170
|
+
if m.path != origin
|
|
171
|
+
and _suffix(m.path) == _suffix(origin)
|
|
172
|
+
and not definition_pattern(name).search(m.text)
|
|
173
|
+
]
|
|
174
|
+
uses.sort(key=lambda m: (is_test(m.path), m.path, m.line)) # code first, then tests
|
|
175
|
+
for match in uses[:MAX_REFERENCES_PER_NAME]:
|
|
176
|
+
related.append(
|
|
177
|
+
Relation(RelationKind.REFERENCE, name, match.path, match.line, match.text)
|
|
178
|
+
)
|
|
179
|
+
kept: list[Relation] = []
|
|
180
|
+
seen: set[tuple[str, int]] = set()
|
|
181
|
+
spent = 0
|
|
182
|
+
for relation in related:
|
|
183
|
+
if (relation.path, relation.line) in seen:
|
|
184
|
+
continue
|
|
185
|
+
spent += len(relation.text) + len(relation.path) + 8
|
|
186
|
+
if spent > MAX_RELATED_CHARS:
|
|
187
|
+
break
|
|
188
|
+
seen.add((relation.path, relation.line))
|
|
189
|
+
kept.append(relation)
|
|
190
|
+
return tuple(kept)
|
|
191
|
+
|
|
192
|
+
|
|
193
|
+
def _suffix(path: str) -> str:
|
|
194
|
+
return path[path.rfind(".") :].lower() if "." in path else ""
|
|
@@ -0,0 +1,17 @@
|
|
|
1
|
+
"""Errors that tell the workflow whether retrying could help."""
|
|
2
|
+
|
|
3
|
+
|
|
4
|
+
class ReviewError(Exception):
|
|
5
|
+
"""A step failed in a way that may succeed on retry (network, rate limit, 5xx)."""
|
|
6
|
+
|
|
7
|
+
|
|
8
|
+
class PermanentReviewError(ReviewError):
|
|
9
|
+
"""A step failed in a way retrying cannot fix (missing permission, PR or repo gone)."""
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
class AccessDenied(PermanentReviewError):
|
|
13
|
+
"""GitHub refused the token (401, 403 or 404): it lacks a permission or cannot see the repository."""
|
|
14
|
+
|
|
15
|
+
|
|
16
|
+
class ForbiddenPath(Exception):
|
|
17
|
+
"""A read asked for a path outside the checkout. Refused, never retried."""
|
|
@@ -0,0 +1,52 @@
|
|
|
1
|
+
"""Feedback on ML AI's comments: a report for the people running a trial, and the team's
|
|
2
|
+
dismissals, which every later review of the repository takes into account."""
|
|
3
|
+
|
|
4
|
+
from __future__ import annotations
|
|
5
|
+
|
|
6
|
+
import logging
|
|
7
|
+
from datetime import datetime, timedelta
|
|
8
|
+
|
|
9
|
+
from mlai_review.application import messages
|
|
10
|
+
from mlai_review.application.errors import ReviewError
|
|
11
|
+
from mlai_review.application.ports import FeedbackSource, Repository
|
|
12
|
+
from mlai_review.domain.feedback import (
|
|
13
|
+
Dismissal,
|
|
14
|
+
FeedbackSummary,
|
|
15
|
+
dismissals,
|
|
16
|
+
judge,
|
|
17
|
+
)
|
|
18
|
+
|
|
19
|
+
_log = logging.getLogger("mlai.orchestrator.feedback")
|
|
20
|
+
|
|
21
|
+
# How far back a review looks for findings the team dismissed, and how many it keeps.
|
|
22
|
+
DISMISSALS_LOOKBACK = timedelta(days=90)
|
|
23
|
+
MAX_DISMISSALS = 20
|
|
24
|
+
|
|
25
|
+
|
|
26
|
+
class FeedbackReporter:
|
|
27
|
+
def __init__(self, source: FeedbackSource) -> None:
|
|
28
|
+
self._source = source
|
|
29
|
+
|
|
30
|
+
async def report(self, full_name: str, since: datetime) -> FeedbackSummary:
|
|
31
|
+
repository = await self._source.locate(full_name)
|
|
32
|
+
comments = await self._source.posted_comments(repository, since)
|
|
33
|
+
runs = await self._source.review_runs(repository, since)
|
|
34
|
+
return FeedbackSummary(comments=tuple(judge(c) for c in comments), runs=runs)
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
class TeamFeedback:
|
|
38
|
+
"""Findings the team dismissed in this repository, for a review about to run. Best
|
|
39
|
+
effort: if feedback cannot be read, the review goes ahead without it."""
|
|
40
|
+
|
|
41
|
+
def __init__(self, source: FeedbackSource, clock: type[datetime] = datetime) -> None:
|
|
42
|
+
self._source = source
|
|
43
|
+
self._clock = clock
|
|
44
|
+
|
|
45
|
+
async def dismissed(self, repository: Repository) -> tuple[Dismissal, ...]:
|
|
46
|
+
since = self._clock.now().astimezone() - DISMISSALS_LOOKBACK
|
|
47
|
+
try:
|
|
48
|
+
comments = await self._source.posted_comments(repository, since, with_threads=False)
|
|
49
|
+
except ReviewError as exc:
|
|
50
|
+
_log.warning(messages.FEEDBACK_UNAVAILABLE, exc)
|
|
51
|
+
return ()
|
|
52
|
+
return dismissals(comments)[:MAX_DISMISSALS]
|
|
@@ -0,0 +1,46 @@
|
|
|
1
|
+
"""Whether two findings are one issue told in different words. Independent readings of a
|
|
2
|
+
change word the same defect differently, and nearby claims share as few words with each other
|
|
3
|
+
as distinct defects do, so words alone cannot tell; a short question to the model can."""
|
|
4
|
+
|
|
5
|
+
from __future__ import annotations
|
|
6
|
+
|
|
7
|
+
import logging
|
|
8
|
+
from abc import ABC, abstractmethod
|
|
9
|
+
|
|
10
|
+
from mlai_review.application import messages, prompts
|
|
11
|
+
from mlai_review.application.errors import ReviewError
|
|
12
|
+
from mlai_review.application.ports import LanguageModel
|
|
13
|
+
from mlai_review.domain.findings import Finding
|
|
14
|
+
|
|
15
|
+
_log = logging.getLogger("mlai.orchestrator.matcher")
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
class IssueMatcher(ABC):
|
|
19
|
+
@abstractmethod
|
|
20
|
+
async def same(self, a: Finding, b: Finding) -> bool:
|
|
21
|
+
"""True when both describe one defect: one fix resolves both."""
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
class NoMatcher(IssueMatcher):
|
|
25
|
+
"""Words alone decide (see domain.findings.same_issue)."""
|
|
26
|
+
|
|
27
|
+
async def same(self, a: Finding, b: Finding) -> bool:
|
|
28
|
+
return False
|
|
29
|
+
|
|
30
|
+
|
|
31
|
+
class LlmIssueMatcher(IssueMatcher):
|
|
32
|
+
"""One small model call per pair. A call that fails keeps the two apart: a repeated
|
|
33
|
+
comment is noise, a lost finding is worse."""
|
|
34
|
+
|
|
35
|
+
def __init__(self, model: LanguageModel) -> None:
|
|
36
|
+
self._model = model
|
|
37
|
+
|
|
38
|
+
async def same(self, a: Finding, b: Finding) -> bool:
|
|
39
|
+
try:
|
|
40
|
+
answer = await self._model.complete_json(
|
|
41
|
+
system=prompts.SAME_ISSUE_SYSTEM, user=prompts.same_issue_question(a, b)
|
|
42
|
+
)
|
|
43
|
+
except ReviewError as exc:
|
|
44
|
+
_log.warning(messages.MATCH_FAILED, a.title, b.title, exc)
|
|
45
|
+
return False
|
|
46
|
+
return answer.get("same") is True
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
"""Wording for warnings the application layer logs, and for what the reviewer's tools answer.
|
|
2
|
+
Adapters own their own wording (see mlai_review.infrastructure.messages)."""
|
|
3
|
+
|
|
4
|
+
# Logged (%-style, for the logging module)
|
|
5
|
+
CONTENTS_UNAVAILABLE = "repository contents unavailable, reviewing diffs only: %s"
|
|
6
|
+
CHANGE_TOO_LARGE = "change too large to review: %d files, %d changed lines"
|
|
7
|
+
FILES_SKIPPED = "left out %d noise file(s): %s"
|
|
8
|
+
INCREMENTAL_REVIEW = "reviewing only the changes since %s, which was reviewed earlier"
|
|
9
|
+
REVIEW_OUTCOME = "review candidates=%d rejected=%d merged=%d refuted=%d published=%d suppressed=%d"
|
|
10
|
+
MATCH_FAILED = "could not compare %r with %r, keeping both: %s"
|
|
11
|
+
MATCHED_LOG = "issues compared=%d joined=%d"
|
|
12
|
+
PARTS_LOG = "reviewing in %d parts of %s files"
|
|
13
|
+
PART_FAILED = "a part of the change could not be reviewed, publishing the others: %s"
|
|
14
|
+
AGREEMENT_LOG = "readings=%d issues by votes %s"
|
|
15
|
+
PASSES_AT_LEAST_ONE = "a review needs at least one pass, got {passes}"
|
|
16
|
+
CHALLENGE_FAILED = "refuter failed on %r, leaving it unchallenged: %s"
|
|
17
|
+
CHALLENGE_OUTCOME = "refuter %s %r: %s"
|
|
18
|
+
SPECIALIST_FAILED = "specialist %s failed, publishing the others: %s"
|
|
19
|
+
WRITE_CHECK_FAILED = "write check failed, publishing the readings: %s"
|
|
20
|
+
RECONCILED = "findings new=%d repeated=%d resolved=%d"
|
|
21
|
+
FEEDBACK_UNAVAILABLE = "team feedback unavailable, reviewing without it: %s"
|
|
22
|
+
DISMISSED_BY_FEEDBACK = "left out %d finding(s) the team dismissed before"
|
|
23
|
+
ANALYSIS_TIMED_OUT = "static analysis not back after %.0fs, reviewing without it"
|
|
24
|
+
ANALYSIS_FAILED = "static analysis unavailable, reviewing without it: %s"
|
|
25
|
+
LOOKUPS_USED = "specialist %s used %d lookup(s), %d chars of results"
|
|
26
|
+
REVIEW_USAGE = (
|
|
27
|
+
"review usage repository=%s pull=%d head=%s model_calls=%d prompt_tokens=%d "
|
|
28
|
+
"cached_tokens=%d completion_tokens=%d"
|
|
29
|
+
)
|
|
30
|
+
|
|
31
|
+
# Answered to the model by its tools (str.format)
|
|
32
|
+
TOOL_UNKNOWN = "error: no tool named {name}"
|
|
33
|
+
TOOL_BAD_ARGUMENTS = "error: bad arguments for {name}: {reason}"
|
|
34
|
+
TOOL_BAD_PATTERN = "error: invalid regular expression: {reason}"
|
|
35
|
+
TOOL_REFUSED = "error: {reason}"
|
|
36
|
+
TOOL_NO_MATCHES = "no matches"
|
|
37
|
+
TOOL_MATCHES_CAPPED = "(more matches not shown; at most {limit} are returned, narrow the search)"
|
|
38
|
+
TOOL_FILE_UNREADABLE = "error: {path} does not exist or is not a text file"
|
|
39
|
+
TOOL_PAST_END = "error: {path} has only {total} lines"
|
|
40
|
+
TOOL_READ_RANGE = "(lines {start}-{end} of {total})"
|
|
41
|
+
TOOL_EMPTY_DIR = "{path} is empty or not a directory"
|