mlai-code 0.3.2 → 0.3.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (157) hide show
  1. package/README.md +1 -1
  2. package/package.json +1 -1
  3. package/src/checkpoints.mjs +217 -0
  4. package/src/cli.mjs +178 -31
  5. package/src/cloud.mjs +2 -0
  6. package/src/engine.mjs +1 -1
  7. package/src/headless.mjs +25 -9
  8. package/src/home.mjs +3 -2
  9. package/src/host.mjs +141 -13
  10. package/src/ide-quick.mjs +147 -0
  11. package/src/ide.mjs +182 -0
  12. package/src/netbox.mjs +1 -1
  13. package/src/parity.mjs +20 -1
  14. package/src/relay.mjs +1 -1
  15. package/src/render.mjs +12 -3
  16. package/src/runtime.mjs +23 -1
  17. package/src/sandbox.mjs +19 -3
  18. package/src/server.mjs +46 -7
  19. package/src/session.mjs +53 -10
  20. package/src/shared-data.mjs +1 -1
  21. package/src/tui/acp.mjs +121 -19
  22. package/src/tui/app.mjs +24 -5
  23. package/src/tui/chat.mjs +43 -16
  24. package/src/tui/cmd-config.mjs +3 -1
  25. package/src/tui/cmd-core.mjs +11 -7
  26. package/src/tui/cmd-extensibility.mjs +13 -2
  27. package/src/tui/commands.mjs +2 -1
  28. package/src/tui/extensions-local.mjs +2 -0
  29. package/src/tui/mcp-config.mjs +22 -4
  30. package/src/tui/review-cli.mjs +15 -4
  31. package/src/tui/snapshots.mjs +57 -16
  32. package/src/tui/statusline.mjs +2 -1
  33. package/src/two-words.mjs +14 -8
  34. package/src/view-sync.mjs +48 -0
  35. package/src/wakeups.mjs +2 -2
  36. package/src/winenv.mjs +1 -1
  37. package/src/workspace-api.mjs +21 -4
  38. package/vendor/client-sdk/index.cjs +5 -4
  39. package/vendor/cloud-cli/mlai.mjs +184 -28
  40. package/worker/architecture_worker.py +33 -4
  41. package/worker/browser_tool.py +19 -1
  42. package/worker/build_planner.py +75 -7
  43. package/worker/chat_fork.py +4 -1
  44. package/worker/checkpoints.py +71 -8
  45. package/worker/cloud_pty.py +91 -2
  46. package/worker/command_safety.py +80 -3
  47. package/worker/compaction_brief.py +186 -0
  48. package/worker/connected_turn.py +30 -0
  49. package/worker/context_manager.py +6 -0
  50. package/worker/conversation.py +139 -12
  51. package/worker/deck_render.js +11 -0
  52. package/worker/deck_tools.py +825 -0
  53. package/worker/git_sync.py +42 -17
  54. package/worker/goals.py +4 -0
  55. package/worker/image_generation.py +102 -23
  56. package/worker/kimi_bridge.py +580 -52
  57. package/worker/kimi_gates.py +74 -6
  58. package/worker/kimi_hook.py +213 -19
  59. package/worker/media_tools.py +182 -0
  60. package/worker/memory.py +2 -2
  61. package/worker/mlai_hooks.py +314 -5
  62. package/worker/mlai_mcp.py +49 -11
  63. package/worker/mlai_review/README.md +35 -0
  64. package/worker/mlai_review/__init__.py +1 -0
  65. package/worker/mlai_review/__main__.py +261 -0
  66. package/worker/mlai_review/_compat.py +29 -0
  67. package/worker/mlai_review/analysis/__init__.py +3 -0
  68. package/worker/mlai_review/analysis/detect.py +50 -0
  69. package/worker/mlai_review/analysis/main.py +128 -0
  70. package/worker/mlai_review/analysis/messages.py +8 -0
  71. package/worker/mlai_review/analysis/tools.py +194 -0
  72. package/worker/mlai_review/application/__init__.py +1 -0
  73. package/worker/mlai_review/application/asking.py +53 -0
  74. package/worker/mlai_review/application/challengers.py +142 -0
  75. package/worker/mlai_review/application/context_builder.py +194 -0
  76. package/worker/mlai_review/application/errors.py +17 -0
  77. package/worker/mlai_review/application/feedback.py +52 -0
  78. package/worker/mlai_review/application/matching.py +46 -0
  79. package/worker/mlai_review/application/messages.py +41 -0
  80. package/worker/mlai_review/application/ports.py +222 -0
  81. package/worker/mlai_review/application/prompts.py +342 -0
  82. package/worker/mlai_review/application/review_service.py +200 -0
  83. package/worker/mlai_review/application/reviewers.py +438 -0
  84. package/worker/mlai_review/application/streaming.py +91 -0
  85. package/worker/mlai_review/application/tools.py +138 -0
  86. package/worker/mlai_review/application/usage.py +145 -0
  87. package/worker/mlai_review/application/write_check.py +46 -0
  88. package/worker/mlai_review/contracts/__init__.py +5 -0
  89. package/worker/mlai_review/contracts/analysis.py +57 -0
  90. package/worker/mlai_review/contracts/review.py +42 -0
  91. package/worker/mlai_review/domain/__init__.py +4 -0
  92. package/worker/mlai_review/domain/changes.py +77 -0
  93. package/worker/mlai_review/domain/context.py +147 -0
  94. package/worker/mlai_review/domain/diagnostics.py +239 -0
  95. package/worker/mlai_review/domain/feedback.py +216 -0
  96. package/worker/mlai_review/domain/findings.py +505 -0
  97. package/worker/mlai_review/domain/lifecycle.py +85 -0
  98. package/worker/mlai_review/domain/partition.py +42 -0
  99. package/worker/mlai_review/domain/relations.py +166 -0
  100. package/worker/mlai_review/domain/relevance.py +202 -0
  101. package/worker/mlai_review/domain/report.py +99 -0
  102. package/worker/mlai_review/domain/structure.py +56 -0
  103. package/worker/mlai_review/domain/writes.py +86 -0
  104. package/worker/mlai_review/eval.py +221 -0
  105. package/worker/mlai_review/eval_cases/clean-rename.json +11 -0
  106. package/worker/mlai_review/eval_cases/missing-ownership-check.json +22 -0
  107. package/worker/mlai_review/eval_cases/off-by-one.json +22 -0
  108. package/worker/mlai_review/eval_cases/sql-injection.json +20 -0
  109. package/worker/mlai_review/feedback_report.py +60 -0
  110. package/worker/mlai_review/infrastructure/__init__.py +1 -0
  111. package/worker/mlai_review/infrastructure/analysis/__init__.py +1 -0
  112. package/worker/mlai_review/infrastructure/analysis/none.py +12 -0
  113. package/worker/mlai_review/infrastructure/analysis/translate.py +17 -0
  114. package/worker/mlai_review/infrastructure/github/__init__.py +0 -0
  115. package/worker/mlai_review/infrastructure/github/client.py +445 -0
  116. package/worker/mlai_review/infrastructure/github/feedback.py +212 -0
  117. package/worker/mlai_review/infrastructure/github/http.py +31 -0
  118. package/worker/mlai_review/infrastructure/github/rendering.py +398 -0
  119. package/worker/mlai_review/infrastructure/github/token.py +20 -0
  120. package/worker/mlai_review/infrastructure/local/__init__.py +1 -0
  121. package/worker/mlai_review/infrastructure/local/analyzer.py +48 -0
  122. package/worker/mlai_review/infrastructure/local/events.py +100 -0
  123. package/worker/mlai_review/infrastructure/local/git_reader.py +233 -0
  124. package/worker/mlai_review/infrastructure/local/publisher.py +138 -0
  125. package/worker/mlai_review/infrastructure/messages.py +54 -0
  126. package/worker/mlai_review/infrastructure/models/__init__.py +0 -0
  127. package/worker/mlai_review/infrastructure/models/openai_compatible.py +269 -0
  128. package/worker/mlai_review/infrastructure/structure/__init__.py +1 -0
  129. package/worker/mlai_review/infrastructure/structure/heuristic.py +186 -0
  130. package/worker/mlai_review/infrastructure/structure/tree_sitter.py +186 -0
  131. package/worker/mlai_review/infrastructure/workspace/__init__.py +1 -0
  132. package/worker/mlai_review/infrastructure/workspace/local.py +105 -0
  133. package/worker/mlai_review/infrastructure/workspace/memory.py +77 -0
  134. package/worker/mlai_review/infrastructure/workspace/paths.py +17 -0
  135. package/worker/mlai_review/pipeline.py +89 -0
  136. package/worker/mlai_review/runtime.py +194 -0
  137. package/worker/mlai_review/settings.py +139 -0
  138. package/worker/model_stream.py +21 -7
  139. package/worker/offline_guard.py +1 -1
  140. package/worker/preview.py +62 -4
  141. package/worker/product_facts.py +20 -0
  142. package/worker/project_context.py +33 -7
  143. package/worker/reply_language.py +92 -10
  144. package/worker/repo_policy.py +5 -0
  145. package/worker/roles.json +199 -37
  146. package/worker/roles.py +271 -9
  147. package/worker/runner.py +366 -53
  148. package/worker/runtime_control.py +64 -3
  149. package/worker/secret_scan.py +3 -2
  150. package/worker/server_memory.py +122 -0
  151. package/worker/site_check.py +279 -0
  152. package/worker/skill_library.py +15 -3
  153. package/worker/skill_pack.py +170 -16
  154. package/worker/subagents.py +6 -2
  155. package/worker/tool_catalog.py +4 -3
  156. package/worker/verification.py +156 -19
  157. package/worker/wakeups.py +56 -4
@@ -0,0 +1,128 @@
1
+ """Entry point of the sandbox job.
2
+
3
+ Environment:
4
+ MLAI_REPOSITORY owner/name
5
+ MLAI_HEAD_SHA the commit to analyse
6
+ MLAI_GITHUB_TOKEN a short-lived installation token with contents:read on that one repository
7
+ MLAI_RESULT_URL where to PUT the result (a write-only blob URL); omitted: print to stdout
8
+ MLAI_CLONE_BASE defaults to https://github.com
9
+
10
+ Local use: `mlai-analyse --root /path/to/checkout` skips the clone and prints the result.
11
+ """
12
+
13
+ from __future__ import annotations
14
+
15
+ import argparse
16
+ import base64
17
+ import json
18
+ import os
19
+ import shutil
20
+ import subprocess
21
+ import sys
22
+ import tempfile
23
+ import urllib.request
24
+ from pathlib import Path
25
+
26
+ from mlai_review.analysis.detect import Tool, detect
27
+ from mlai_review.analysis.tools import ToolFailed, install_node_dependencies, run_tool
28
+ from mlai_review.contracts.analysis import AnalysisResult, Diagnostic
29
+
30
+ CLONE_TIMEOUT_SECONDS = 600
31
+
32
+
33
+ def analyse(
34
+ root: Path, *, allowed: tuple[Tool, ...] | None = None, install: bool = True
35
+ ) -> AnalysisResult:
36
+ """`allowed` limits the tools (eslint runs the repository's own config, which is code);
37
+ `install` False uses the dependencies already there, never touching a developer's
38
+ node_modules."""
39
+ tools = tuple(t for t in detect(root) if allowed is None or t in allowed)
40
+ diagnostics: list[Diagnostic] = []
41
+ errors: dict[str, str] = {}
42
+ ran: list[str] = []
43
+ node = Tool.TSC in tools or Tool.ESLINT in tools
44
+ if install and node and not (root / "node_modules").is_dir():
45
+ try:
46
+ install_node_dependencies(root)
47
+ except ToolFailed as exc:
48
+ errors["npm"] = str(exc)
49
+ for tool in tools:
50
+ try:
51
+ diagnostics += run_tool(tool, root)
52
+ ran.append(tool.value)
53
+ except ToolFailed as exc:
54
+ errors[tool.value] = str(exc)
55
+ return AnalysisResult(tools_run=tuple(ran), diagnostics=tuple(diagnostics), errors=errors)
56
+
57
+
58
+ def clone(repository: str, sha: str, token: str, base: str, into: Path) -> None:
59
+ credentials = base64.b64encode(f"x-access-token:{token}".encode()).decode()
60
+ env = {
61
+ **os.environ,
62
+ "GIT_TERMINAL_PROMPT": "0",
63
+ "GIT_CONFIG_COUNT": "1",
64
+ "GIT_CONFIG_KEY_0": "http.extraheader",
65
+ "GIT_CONFIG_VALUE_0": f"Authorization: Basic {credentials}",
66
+ }
67
+ url = f"{base.rstrip('/')}/{repository}.git"
68
+ for args in (
69
+ ["init", "-q"],
70
+ ["fetch", "-q", "--depth", "1", "--no-tags", url, sha],
71
+ ["checkout", "-q", "--detach", "FETCH_HEAD"],
72
+ ):
73
+ subprocess.run(
74
+ ["git", "-C", str(into), *args], check=True, env=env, timeout=CLONE_TIMEOUT_SECONDS
75
+ )
76
+
77
+
78
+ def publish(result: AnalysisResult, url: str | None) -> None:
79
+ body = json.dumps(result.as_dict()).encode()
80
+ if not url:
81
+ sys.stdout.write(body.decode() + "\n")
82
+ return
83
+ request = urllib.request.Request(
84
+ url,
85
+ data=body,
86
+ method="PUT",
87
+ headers={"x-ms-blob-type": "BlockBlob", "Content-Type": "application/json"},
88
+ )
89
+ with urllib.request.urlopen(request, timeout=60):
90
+ pass
91
+
92
+
93
+ def main() -> int:
94
+ parser = argparse.ArgumentParser(description=__doc__)
95
+ parser.add_argument("--root", type=Path, help="analyse an existing checkout instead of cloning")
96
+ parser.add_argument("--out", type=Path, help="write the result here instead of stdout")
97
+ parser.add_argument("--no-eslint", action="store_true", help="skip eslint (runs repo code)")
98
+ args = parser.parse_args()
99
+ if args.root:
100
+ allowed = tuple(t for t in Tool if t is not Tool.ESLINT) if args.no_eslint else None
101
+ result = analyse(args.root, allowed=allowed)
102
+ if args.out:
103
+ args.out.write_text(json.dumps(result.as_dict()))
104
+ else:
105
+ publish(result, None)
106
+ return 0
107
+ repository = os.environ["MLAI_REPOSITORY"]
108
+ sha = os.environ["MLAI_HEAD_SHA"]
109
+ token = os.environ.pop("MLAI_GITHUB_TOKEN") # out of the environment before any tool runs
110
+ result_url = os.environ.get("MLAI_RESULT_URL")
111
+ base = os.environ.get("MLAI_CLONE_BASE", "https://github.com")
112
+ workdir = Path(tempfile.mkdtemp(prefix="mlai-analysis-"))
113
+ root = workdir / "repo"
114
+ root.mkdir()
115
+ try:
116
+ clone(repository, sha, token, base, root)
117
+ del token
118
+ result = analyse(root)
119
+ except subprocess.CalledProcessError as exc:
120
+ result = AnalysisResult(errors={"clone": f"git exited with {exc.returncode}"})
121
+ finally:
122
+ shutil.rmtree(workdir, ignore_errors=True)
123
+ publish(result, result_url)
124
+ return 0
125
+
126
+
127
+ if __name__ == "__main__":
128
+ sys.exit(main())
@@ -0,0 +1,8 @@
1
+ """Wording for why a tool could not run. Reported to operators in the result, never to the
2
+ model."""
3
+
4
+ TOOL_NOT_AVAILABLE = "{name}: not available: {error}"
5
+ TOOL_TIMED_OUT = "{name}: took longer than {seconds:.0f}s"
6
+ TOOL_EXITED = "{name}: exited with {code}: {detail}"
7
+ TOOL_UNREADABLE_OUTPUT = "{name}: unreadable output: {error}"
8
+ ESLINT_MISSING = "eslint: not installed in the repository (node_modules/.bin/eslint)"
@@ -0,0 +1,194 @@
1
+ """Running each tool and turning what it prints into diagnostics. Every tool is run with a
2
+ timeout and a clean environment; its exit code is not an error (tools exit non-zero when they
3
+ find something), only its output matters."""
4
+
5
+ from __future__ import annotations
6
+
7
+ import json
8
+ import os
9
+ import re
10
+ import subprocess
11
+ from collections.abc import Callable
12
+ from pathlib import Path
13
+
14
+ from mlai_review.analysis import messages
15
+ from mlai_review.analysis.detect import Tool
16
+ from mlai_review.contracts.analysis import Diagnostic
17
+
18
+ TOOL_TIMEOUT_SECONDS = 600
19
+ INSTALL_TIMEOUT_SECONDS = 900
20
+ MAX_DIAGNOSTICS = 2_000
21
+
22
+ _TSC_LINE = re.compile(
23
+ r"^(?P<path>[^(]+)\((?P<line>\d+),\d+\): "
24
+ r"(?P<level>error|warning) (?P<rule>TS\d+): (?P<message>.*)$"
25
+ )
26
+ # Ruff rules that mean the code is wrong, not merely untidy: undefined and misused names,
27
+ # syntax errors, invalid constructs. Everything else is reported as a warning.
28
+ _RUFF_ERRORS = frozenset(
29
+ {
30
+ "F821",
31
+ "F822",
32
+ "F823",
33
+ "F811",
34
+ "F632",
35
+ "F601",
36
+ "F602",
37
+ "F631",
38
+ "F633",
39
+ "F701",
40
+ "F702",
41
+ "F704",
42
+ "F706",
43
+ "F707",
44
+ }
45
+ )
46
+
47
+
48
+ class ToolFailed(Exception):
49
+ """The tool could not run at all (missing, crashed, timed out)."""
50
+
51
+
52
+ def run_tool(tool: Tool, root: Path) -> list[Diagnostic]:
53
+ runner: Callable[[Path], list[Diagnostic]] = {
54
+ Tool.RUFF: _ruff,
55
+ Tool.TSC: _tsc,
56
+ Tool.ESLINT: _eslint,
57
+ }[tool]
58
+ return runner(root)[:MAX_DIAGNOSTICS]
59
+
60
+
61
+ def install_node_dependencies(root: Path) -> None:
62
+ """Dependencies for tsc and eslint. Lifecycle scripts never run: they are arbitrary code
63
+ from the pull request, and the tools do not need them."""
64
+ if not (root / "package.json").is_file():
65
+ return
66
+ lockfile = (root / "package-lock.json").is_file() or (root / "npm-shrinkwrap.json").is_file()
67
+ command = [
68
+ "npm",
69
+ "ci" if lockfile else "install",
70
+ "--ignore-scripts",
71
+ "--no-audit",
72
+ "--no-fund",
73
+ ]
74
+ _execute(command, root, INSTALL_TIMEOUT_SECONDS, name="npm")
75
+
76
+
77
+ def _ruff(root: Path) -> list[Diagnostic]:
78
+ output = _execute(
79
+ ["ruff", "check", "--output-format", "json", "--exit-zero", "."],
80
+ root,
81
+ TOOL_TIMEOUT_SECONDS,
82
+ name="ruff",
83
+ )
84
+ try:
85
+ items = json.loads(output or "[]")
86
+ except json.JSONDecodeError as exc:
87
+ raise ToolFailed(messages.TOOL_UNREADABLE_OUTPUT.format(name="ruff", error=exc)) from exc
88
+ return [
89
+ Diagnostic(
90
+ tool="ruff",
91
+ path=_relative(root, item["filename"]),
92
+ line=int(item["location"]["row"]),
93
+ rule=str(item.get("code") or ""),
94
+ message=str(item.get("message", "")),
95
+ level="error" if _is_ruff_error(str(item.get("code") or "")) else "warning",
96
+ )
97
+ for item in items
98
+ ]
99
+
100
+
101
+ def _tsc(root: Path) -> list[Diagnostic]:
102
+ binary = root / "node_modules" / ".bin" / "tsc"
103
+ command = [str(binary) if binary.is_file() else "tsc", "--noEmit", "--pretty", "false"]
104
+ output = _execute(command, root, TOOL_TIMEOUT_SECONDS, name="tsc", ok_codes=(0, 1, 2))
105
+ diagnostics = []
106
+ for line in output.splitlines():
107
+ match = _TSC_LINE.match(line.strip())
108
+ if match:
109
+ diagnostics.append(
110
+ Diagnostic(
111
+ tool="tsc",
112
+ path=_relative(root, match["path"]),
113
+ line=int(match["line"]),
114
+ rule=match["rule"],
115
+ message=match["message"],
116
+ level=match["level"],
117
+ )
118
+ )
119
+ return diagnostics
120
+
121
+
122
+ def _eslint(root: Path) -> list[Diagnostic]:
123
+ binary = root / "node_modules" / ".bin" / "eslint"
124
+ if not binary.is_file():
125
+ raise ToolFailed(messages.ESLINT_MISSING)
126
+ output = _execute(
127
+ [str(binary), ".", "--format", "json", "--no-error-on-unmatched-pattern"],
128
+ root,
129
+ TOOL_TIMEOUT_SECONDS,
130
+ name="eslint",
131
+ ok_codes=(0, 1),
132
+ )
133
+ try:
134
+ files = json.loads(output or "[]")
135
+ except json.JSONDecodeError as exc:
136
+ raise ToolFailed(messages.TOOL_UNREADABLE_OUTPUT.format(name="eslint", error=exc)) from exc
137
+ return [
138
+ Diagnostic(
139
+ tool="eslint",
140
+ path=_relative(root, file["filePath"]),
141
+ line=int(message.get("line") or 1),
142
+ rule=str(message.get("ruleId") or ""),
143
+ message=str(message.get("message", "")),
144
+ level="error" if message.get("severity") == 2 else "warning",
145
+ )
146
+ for file in files
147
+ for message in file.get("messages", [])
148
+ ]
149
+
150
+
151
+ def _execute(
152
+ command: list[str],
153
+ root: Path,
154
+ timeout: float,
155
+ *,
156
+ name: str,
157
+ ok_codes: tuple[int, ...] = (0,),
158
+ ) -> str:
159
+ env = {
160
+ "PATH": os.environ.get("PATH", "/usr/local/bin:/usr/bin:/bin"),
161
+ "HOME": str(root.parent),
162
+ "LANG": "C.UTF-8",
163
+ "CI": "1",
164
+ }
165
+ try:
166
+ done = subprocess.run(
167
+ command, cwd=root, env=env, capture_output=True, text=True, timeout=timeout
168
+ )
169
+ except FileNotFoundError as exc:
170
+ raise ToolFailed(messages.TOOL_NOT_AVAILABLE.format(name=name, error=exc)) from exc
171
+ except subprocess.TimeoutExpired as exc:
172
+ raise ToolFailed(messages.TOOL_TIMED_OUT.format(name=name, seconds=timeout)) from exc
173
+ if done.returncode not in ok_codes:
174
+ detail = (done.stderr or done.stdout).strip().splitlines()
175
+ raise ToolFailed(
176
+ messages.TOOL_EXITED.format(
177
+ name=name, code=done.returncode, detail=detail[-1] if detail else ""
178
+ )
179
+ )
180
+ return done.stdout
181
+
182
+
183
+ def _is_ruff_error(code: str) -> bool:
184
+ return code in _RUFF_ERRORS or code.startswith("E9")
185
+
186
+
187
+ def _relative(root: Path, path: str) -> str:
188
+ candidate = Path(path)
189
+ if candidate.is_absolute():
190
+ try:
191
+ return candidate.resolve().relative_to(root.resolve()).as_posix()
192
+ except ValueError:
193
+ return candidate.as_posix()
194
+ return candidate.as_posix()
@@ -0,0 +1 @@
1
+ """Use cases and the ports they need. Depends on the domain and mlai_contracts only."""
@@ -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)