makker 0.3.0__py3-none-any.whl
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.
- makker/__init__.py +33 -0
- makker/__main__.py +23 -0
- makker/_version.py +24 -0
- makker/agent/__init__.py +0 -0
- makker/agent/added_dirs.py +117 -0
- makker/agent/approval.py +278 -0
- makker/agent/backend.py +160 -0
- makker/agent/builder.py +530 -0
- makker/agent/capabilities.py +355 -0
- makker/agent/context_editing.py +74 -0
- makker/agent/memory.py +267 -0
- makker/agent/openrouter.py +34 -0
- makker/agent/plot_julia.py +786 -0
- makker/agent/plot_julia_src.py +806 -0
- makker/agent/prompts.py +244 -0
- makker/agent/recovery.py +127 -0
- makker/agent/summarization.py +163 -0
- makker/agent/titling.py +60 -0
- makker/agent/tool_output.py +86 -0
- makker/agent/tools.py +358 -0
- makker/agent/turns.py +431 -0
- makker/agent/windows_paths.py +133 -0
- makker/approval_preview.py +270 -0
- makker/context_panel.py +239 -0
- makker/credentials.py +163 -0
- makker/display.py +188 -0
- makker/eval/__init__.py +22 -0
- makker/eval/_gemini_compat.py +65 -0
- makker/eval/report.py +475 -0
- makker/eval/runconfig.py +91 -0
- makker/eval/scorers.py +712 -0
- makker/eval/solver.py +295 -0
- makker/eval/tasks/__init__.py +1 -0
- makker/eval/tasks/_corpus/MiniRes/Project.toml +3 -0
- makker/eval/tasks/_corpus/MiniRes/README.md +4 -0
- makker/eval/tasks/_corpus/MiniRes/docs/api.md +7 -0
- makker/eval/tasks/_corpus/MiniRes/examples/advanced/sweep.jl +8 -0
- makker/eval/tasks/_corpus/MiniRes/examples/quickstart.jl +6 -0
- makker/eval/tasks/_corpus/MiniRes/src/MiniRes.jl +13 -0
- makker/eval/tasks/_corpus/MiniRes/src/grid.jl +8 -0
- makker/eval/tasks/_corpus/MiniRes/src/physics/darcy.jl +13 -0
- makker/eval/tasks/_corpus/MiniRes/src/physics/wells.jl +15 -0
- makker/eval/tasks/_corpus/MiniRes/src/solver/newton.jl +15 -0
- makker/eval/tasks/_corpus/__init__.py +54 -0
- makker/eval/tasks/api_discovery.py +122 -0
- makker/eval/tasks/battmo.py +88 -0
- makker/eval/tasks/calibration.py +82 -0
- makker/eval/tasks/canary.py +41 -0
- makker/eval/tasks/ensembles.py +209 -0
- makker/eval/tasks/filesystem.py +290 -0
- makker/eval/tasks/filesystem_source.py +80 -0
- makker/eval/tasks/fimbul.py +51 -0
- makker/eval/tasks/guardrails.py +47 -0
- makker/eval/tasks/jutuldarcy.py +139 -0
- makker/eval/tasks/mocca.py +90 -0
- makker/eval/tasks/plotting.py +111 -0
- makker/eval/tasks/search.py +211 -0
- makker/eval/tasks/search_source.py +71 -0
- makker/eval/tasks/usage.py +138 -0
- makker/interfaces/__init__.py +0 -0
- makker/interfaces/cli/__init__.py +10 -0
- makker/interfaces/cli/_helpers.py +197 -0
- makker/interfaces/cli/doctor.py +443 -0
- makker/interfaces/cli/eval.py +198 -0
- makker/interfaces/cli/init.py +274 -0
- makker/interfaces/cli/key.py +117 -0
- makker/interfaces/cli/main.py +207 -0
- makker/interfaces/cli/review.py +538 -0
- makker/interfaces/cli/run.py +412 -0
- makker/interfaces/cli/sessions.py +45 -0
- makker/interfaces/cli/sysimage.py +223 -0
- makker/interfaces/cli/transcript.py +102 -0
- makker/interfaces/cli/upgrade.py +369 -0
- makker/interfaces/cli/web.py +314 -0
- makker/interfaces/server/__init__.py +13 -0
- makker/interfaces/server/access.py +148 -0
- makker/interfaces/server/app.py +2443 -0
- makker/interfaces/server/manager.py +292 -0
- makker/interfaces/server/protocol.py +321 -0
- makker/interfaces/server/update_helper.py +199 -0
- makker/interfaces/server/updates.py +446 -0
- makker/interfaces/server/web_dist/assets/index-BcWXsrUW.js +54 -0
- makker/interfaces/server/web_dist/assets/index-Cbigi-bs.css +1 -0
- makker/interfaces/server/web_dist/index.html +17 -0
- makker/interfaces/tui/__init__.py +5 -0
- makker/interfaces/tui/app.py +1779 -0
- makker/interfaces/tui/approval_menu.py +227 -0
- makker/interfaces/tui/commands.py +143 -0
- makker/interfaces/tui/model_menu.py +416 -0
- makker/interfaces/tui/prompt.py +210 -0
- makker/interfaces/tui/tool_display.py +440 -0
- makker/interfaces/tui/widgets.py +991 -0
- makker/julia/__init__.py +0 -0
- makker/julia/precompile.py +120 -0
- makker/julia/requirements.py +101 -0
- makker/julia/session.py +49 -0
- makker/julia/threads.py +126 -0
- makker/julia_runtime/Makker/Project.toml +21 -0
- makker/julia_runtime/Makker/src/Makker.jl +99 -0
- makker/julia_runtime/Makker/src/ensemble.jl +104 -0
- makker/julia_runtime/Makker/src/plots.jl +292 -0
- makker/julia_runtime/Makker/src/web_views.jl +77 -0
- makker/juliakernel/__init__.py +19 -0
- makker/juliakernel/config.py +35 -0
- makker/juliakernel/connection.py +261 -0
- makker/juliakernel/kernel.py +509 -0
- makker/juliakernel/result.py +37 -0
- makker/juliakernel/server.jl +263 -0
- makker/juliakernel/text.py +211 -0
- makker/lab/__init__.py +19 -0
- makker/lab/fakes.py +744 -0
- makker/lab/live.py +225 -0
- makker/lab/profile_startup.py +144 -0
- makker/lab/profile_turn.py +70 -0
- makker/lab/rasterize.py +73 -0
- makker/lab/scenarios.py +255 -0
- makker/lab/tui.py +194 -0
- makker/legacy.py +298 -0
- makker/models.py +441 -0
- makker/ollama_client.py +177 -0
- makker/open_file.py +54 -0
- makker/paths.py +205 -0
- makker/preview.py +70 -0
- makker/py.typed +0 -0
- makker/recent_models.py +47 -0
- makker/review/__init__.py +40 -0
- makker/review/curate.py +135 -0
- makker/review/dashboard.py +216 -0
- makker/review/dashboard_template.html +384 -0
- makker/review/discovery.py +159 -0
- makker/review/eval_link.py +81 -0
- makker/review/export.py +74 -0
- makker/review/findings.py +176 -0
- makker/review/issues.py +190 -0
- makker/review/prompt.py +195 -0
- makker/review/reviewer.py +195 -0
- makker/review/server.py +109 -0
- makker/review/settings.py +28 -0
- makker/session.py +520 -0
- makker/session_host.py +706 -0
- makker/simulators/README.md +50 -0
- makker/simulators/__init__.py +0 -0
- makker/simulators/base.py +63 -0
- makker/simulators/battmo/__init__.py +3 -0
- makker/simulators/battmo/adapter.py +39 -0
- makker/simulators/battmo/julia_env/MakkerBattMo/Project.toml +21 -0
- makker/simulators/battmo/julia_env/MakkerBattMo/src/MakkerBattMo.jl +43 -0
- makker/simulators/battmo/julia_env/Project.toml +34 -0
- makker/simulators/battmo/skills/battmo-cycling/SKILL.md +33 -0
- makker/simulators/battmo/skills/battmo-overview/SKILL.md +55 -0
- makker/simulators/battmo/skills/battmo-plotting/SKILL.md +46 -0
- makker/simulators/env_setup.py +1070 -0
- makker/simulators/fimbul/__init__.py +3 -0
- makker/simulators/fimbul/adapter.py +43 -0
- makker/simulators/fimbul/julia_env/MakkerFimbul/Project.toml +23 -0
- makker/simulators/fimbul/julia_env/MakkerFimbul/src/MakkerFimbul.jl +89 -0
- makker/simulators/fimbul/julia_env/Project.toml +40 -0
- makker/simulators/fimbul/skills/fimbul-overview/SKILL.md +72 -0
- makker/simulators/fimbul/skills/fimbul-plotting/SKILL.md +69 -0
- makker/simulators/jutuldarcy/__init__.py +3 -0
- makker/simulators/jutuldarcy/adapter.py +41 -0
- makker/simulators/jutuldarcy/julia_env/MakkerJutulDarcy/Project.toml +24 -0
- makker/simulators/jutuldarcy/julia_env/MakkerJutulDarcy/src/MakkerJutulDarcy.jl +149 -0
- makker/simulators/jutuldarcy/julia_env/Project.toml +37 -0
- makker/simulators/jutuldarcy/skills/jutuldarcy-overview/SKILL.md +82 -0
- makker/simulators/jutuldarcy/skills/jutuldarcy-plotting/SKILL.md +74 -0
- makker/simulators/jutuldarcy/skills/jutuldarcy-wells/SKILL.md +67 -0
- makker/simulators/mocca/__init__.py +3 -0
- makker/simulators/mocca/adapter.py +40 -0
- makker/simulators/mocca/julia_env/MakkerMocca/Project.toml +22 -0
- makker/simulators/mocca/julia_env/MakkerMocca/src/MakkerMocca.jl +55 -0
- makker/simulators/mocca/julia_env/Project.toml +35 -0
- makker/simulators/mocca/skills/mocca-overview/SKILL.md +125 -0
- makker/simulators/registry.py +147 -0
- makker/simulators/shared_skills/ensembles/SKILL.md +94 -0
- makker/simulators/shared_skills/investigation-loop/SKILL.md +101 -0
- makker/simulators/shared_skills/julia-and-repl/SKILL.md +93 -0
- makker/simulators/shared_skills/plotting-basics/SKILL.md +138 -0
- makker/simulators/shared_skills/workspace-and-source/SKILL.md +110 -0
- makker/simulators/warmup.py +267 -0
- makker/sysimage.py +548 -0
- makker/sysimage_build.py +881 -0
- makker/tool_labels.py +35 -0
- makker/trace/__init__.py +10 -0
- makker/trace/log.py +155 -0
- makker/trace/messages.py +46 -0
- makker/trace/recorder.py +141 -0
- makker/trace/schema.py +129 -0
- makker/transcript/__init__.py +8 -0
- makker/transcript/attempts.py +67 -0
- makker/transcript/bundle.py +18 -0
- makker/transcript/embed.py +55 -0
- makker/transcript/events.py +77 -0
- makker/transcript/highlight.py +143 -0
- makker/transcript/html.py +1031 -0
- makker/transcript/markdown.py +176 -0
- makker/transcript/markdown_html.py +65 -0
- makker/transcript/report.py +1076 -0
- makker/update_check.py +329 -0
- makker/user_config.py +51 -0
- makker/workspace.py +1170 -0
- makker-0.3.0.dist-info/METADATA +157 -0
- makker-0.3.0.dist-info/RECORD +206 -0
- makker-0.3.0.dist-info/WHEEL +4 -0
- makker-0.3.0.dist-info/entry_points.txt +2 -0
- makker-0.3.0.dist-info/licenses/LICENSE +21 -0
makker/__init__.py
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
1
|
+
"""Makker: an AI agent for computational science."""
|
|
2
|
+
|
|
3
|
+
from __future__ import annotations
|
|
4
|
+
|
|
5
|
+
from importlib.metadata import PackageNotFoundError, version
|
|
6
|
+
|
|
7
|
+
|
|
8
|
+
def _resolve_version() -> str:
|
|
9
|
+
"""Best-effort running version, robust across install shapes.
|
|
10
|
+
|
|
11
|
+
Preference order:
|
|
12
|
+
1. Installed distribution metadata (wheel, ``uv tool install``, editable) —
|
|
13
|
+
the canonical source, derived from git tags by hatch-vcs at build time.
|
|
14
|
+
2. The ``_version.py`` file hatch-vcs bakes into the wheel, in case metadata
|
|
15
|
+
is unavailable (e.g. running from a source tree that was built but not
|
|
16
|
+
installed).
|
|
17
|
+
3. ``"0.0.0+unknown"`` when neither is present (a bare source checkout with no
|
|
18
|
+
build), so callers always get a string.
|
|
19
|
+
"""
|
|
20
|
+
|
|
21
|
+
try:
|
|
22
|
+
return version("makker")
|
|
23
|
+
except PackageNotFoundError:
|
|
24
|
+
pass
|
|
25
|
+
try:
|
|
26
|
+
from makker._version import __version__ as baked # type: ignore[import-not-found]
|
|
27
|
+
|
|
28
|
+
return baked
|
|
29
|
+
except Exception:
|
|
30
|
+
return "0.0.0+unknown"
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
__version__ = _resolve_version()
|
makker/__main__.py
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
"""Entry point for the Makker CLI.
|
|
2
|
+
|
|
3
|
+
Delegates to ``interfaces.cli.main`` so the CLI plumbing has one home.
|
|
4
|
+
"""
|
|
5
|
+
|
|
6
|
+
from __future__ import annotations
|
|
7
|
+
|
|
8
|
+
import sys
|
|
9
|
+
import warnings
|
|
10
|
+
|
|
11
|
+
# The LangChain packages import pydantic.v1 so they can accept V1 models, and on
|
|
12
|
+
# Python 3.14+ that import alone warns at the start of every command. Makker
|
|
13
|
+
# never hands them a V1 model, so the warning says nothing about this program.
|
|
14
|
+
warnings.filterwarnings(
|
|
15
|
+
"ignore",
|
|
16
|
+
message="Core Pydantic V1 functionality isn't compatible with Python 3.14",
|
|
17
|
+
category=UserWarning,
|
|
18
|
+
)
|
|
19
|
+
|
|
20
|
+
from makker.interfaces.cli import main # noqa: E402
|
|
21
|
+
|
|
22
|
+
if __name__ == "__main__":
|
|
23
|
+
sys.exit(main())
|
makker/_version.py
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
1
|
+
# file generated by vcs-versioning
|
|
2
|
+
# don't change, don't track in version control
|
|
3
|
+
from __future__ import annotations
|
|
4
|
+
|
|
5
|
+
__all__ = [
|
|
6
|
+
"__version__",
|
|
7
|
+
"__version_tuple__",
|
|
8
|
+
"version",
|
|
9
|
+
"version_tuple",
|
|
10
|
+
"__commit_id__",
|
|
11
|
+
"commit_id",
|
|
12
|
+
]
|
|
13
|
+
|
|
14
|
+
version: str
|
|
15
|
+
__version__: str
|
|
16
|
+
__version_tuple__: tuple[int | str, ...]
|
|
17
|
+
version_tuple: tuple[int | str, ...]
|
|
18
|
+
commit_id: str | None
|
|
19
|
+
__commit_id__: str | None
|
|
20
|
+
|
|
21
|
+
__version__ = version = '0.3.0'
|
|
22
|
+
__version_tuple__ = version_tuple = (0, 3, 0)
|
|
23
|
+
|
|
24
|
+
__commit_id__ = commit_id = None
|
makker/agent/__init__.py
ADDED
|
File without changes
|
|
@@ -0,0 +1,117 @@
|
|
|
1
|
+
"""Track additional working directories the user adds to the session.
|
|
2
|
+
|
|
3
|
+
The agent's filesystem uses real paths (see ``agent.builder.build_backend``), so
|
|
4
|
+
an added folder is already reachable at its absolute path by every tool:
|
|
5
|
+
``read_file``, ``grep``, ``write_file``, ``execute``, and ``run_julia`` alike.
|
|
6
|
+
Adding a folder just validates the path and records it on the live backend so
|
|
7
|
+
the session can list what was added (via the ``--add-dir`` flag at launch or the
|
|
8
|
+
TUI ``/add-dir`` command). Records are session-scoped: they live until the
|
|
9
|
+
process exits and are not persisted.
|
|
10
|
+
"""
|
|
11
|
+
|
|
12
|
+
from __future__ import annotations
|
|
13
|
+
|
|
14
|
+
import re
|
|
15
|
+
from dataclasses import dataclass
|
|
16
|
+
from pathlib import Path
|
|
17
|
+
|
|
18
|
+
from makker.paths import workspace_root
|
|
19
|
+
|
|
20
|
+
# Collapse anything outside this set to a single dash for a readable short name.
|
|
21
|
+
_UNSAFE_NAME_CHARS = re.compile(r"[^A-Za-z0-9._-]+")
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
class AddDirError(ValueError):
|
|
25
|
+
"""A folder could not be added (missing path, not a directory, ...)."""
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
@dataclass(frozen=True)
|
|
29
|
+
class AddedDir:
|
|
30
|
+
"""A folder the user added to the session, used at its real ``path``."""
|
|
31
|
+
|
|
32
|
+
name: str
|
|
33
|
+
path: Path
|
|
34
|
+
|
|
35
|
+
|
|
36
|
+
def _slugify(path: Path) -> str:
|
|
37
|
+
"""A safe, human-readable short name derived from the folder's basename."""
|
|
38
|
+
|
|
39
|
+
base = _UNSAFE_NAME_CHARS.sub("-", path.name).strip("-._")
|
|
40
|
+
return base or "dir"
|
|
41
|
+
|
|
42
|
+
|
|
43
|
+
def _unique_name(taken: set[str], desired: str) -> str:
|
|
44
|
+
"""``desired`` if free, else ``desired-2``, ``desired-3``, ... so two
|
|
45
|
+
folders that share a basename get distinct names."""
|
|
46
|
+
|
|
47
|
+
if desired not in taken:
|
|
48
|
+
return desired
|
|
49
|
+
index = 2
|
|
50
|
+
while f"{desired}-{index}" in taken:
|
|
51
|
+
index += 1
|
|
52
|
+
return f"{desired}-{index}"
|
|
53
|
+
|
|
54
|
+
|
|
55
|
+
def resolve_dir(raw: str | Path, *, workspace: Path | None = None) -> Path:
|
|
56
|
+
"""Resolve a user-supplied folder to an absolute, existing directory.
|
|
57
|
+
|
|
58
|
+
``~`` and relative paths are resolved against the workspace root (the
|
|
59
|
+
user's launch directory), matching how they'd read the path at the prompt.
|
|
60
|
+
Raises ``AddDirError`` if the path is missing or not a directory.
|
|
61
|
+
"""
|
|
62
|
+
|
|
63
|
+
base = (workspace or workspace_root()).resolve()
|
|
64
|
+
candidate = Path(raw).expanduser()
|
|
65
|
+
if not candidate.is_absolute():
|
|
66
|
+
candidate = base / candidate
|
|
67
|
+
candidate = candidate.resolve()
|
|
68
|
+
if not candidate.exists():
|
|
69
|
+
raise AddDirError(f"no such directory: {candidate}")
|
|
70
|
+
if not candidate.is_dir():
|
|
71
|
+
raise AddDirError(f"not a directory: {candidate}")
|
|
72
|
+
return candidate
|
|
73
|
+
|
|
74
|
+
|
|
75
|
+
def _registry(backend) -> list[AddedDir]:
|
|
76
|
+
"""The mutable list of added folders, stored on the live backend."""
|
|
77
|
+
dirs = getattr(backend, "_added_dirs", None)
|
|
78
|
+
if dirs is None:
|
|
79
|
+
dirs = []
|
|
80
|
+
backend._added_dirs = dirs
|
|
81
|
+
return dirs
|
|
82
|
+
|
|
83
|
+
|
|
84
|
+
def added_dirs(backend) -> list[AddedDir]:
|
|
85
|
+
"""The folders added to this session, sorted by name."""
|
|
86
|
+
return sorted(_registry(backend), key=lambda entry: entry.name)
|
|
87
|
+
|
|
88
|
+
|
|
89
|
+
def add_dir(
|
|
90
|
+
backend,
|
|
91
|
+
raw: str | Path,
|
|
92
|
+
*,
|
|
93
|
+
workspace: Path | None = None,
|
|
94
|
+
) -> AddedDir:
|
|
95
|
+
"""Validate ``raw`` and record it as an added folder; return its ``AddedDir``.
|
|
96
|
+
|
|
97
|
+
The real-path backend already reads and writes any real path, so the agent
|
|
98
|
+
uses the folder's absolute path directly; recording it just lets the session
|
|
99
|
+
list and track what was added. Idempotent (an already-added folder returns
|
|
100
|
+
its existing record). The workspace itself is rejected because it is already
|
|
101
|
+
the working directory. Raises ``AddDirError`` for a missing or non-directory path.
|
|
102
|
+
"""
|
|
103
|
+
|
|
104
|
+
ws = (workspace or workspace_root()).resolve()
|
|
105
|
+
path = resolve_dir(raw, workspace=ws)
|
|
106
|
+
if path == ws:
|
|
107
|
+
raise AddDirError("the workspace is already the working directory, no need to add it")
|
|
108
|
+
|
|
109
|
+
registry = _registry(backend)
|
|
110
|
+
for entry in registry:
|
|
111
|
+
if entry.path == path:
|
|
112
|
+
return entry
|
|
113
|
+
|
|
114
|
+
name = _unique_name({entry.name for entry in registry}, _slugify(path))
|
|
115
|
+
entry = AddedDir(name=name, path=path)
|
|
116
|
+
registry.append(entry)
|
|
117
|
+
return entry
|
makker/agent/approval.py
ADDED
|
@@ -0,0 +1,278 @@
|
|
|
1
|
+
"""Human-in-the-loop approval modes for workspace file and shell tools."""
|
|
2
|
+
|
|
3
|
+
from __future__ import annotations
|
|
4
|
+
|
|
5
|
+
from collections.abc import Iterable
|
|
6
|
+
from copy import deepcopy
|
|
7
|
+
from enum import StrEnum
|
|
8
|
+
from typing import Any, Protocol
|
|
9
|
+
|
|
10
|
+
from langchain.agents.middleware import InterruptOnConfig
|
|
11
|
+
|
|
12
|
+
# Tools that require human approval. Action name → human-readable description
|
|
13
|
+
# shown in the approval card. All entries use approve/reject only.
|
|
14
|
+
_APPROVAL_TOOLS: dict[str, str] = {
|
|
15
|
+
"execute": "Run a shell command in the workspace.",
|
|
16
|
+
"write_file": "Write a file in the workspace.",
|
|
17
|
+
"edit_file": "Edit a file in the workspace.",
|
|
18
|
+
"delete": "Delete a file or directory (directories go recursively).",
|
|
19
|
+
}
|
|
20
|
+
|
|
21
|
+
# Deliberately excludes `delete`: these are the tools "accept edits" mode and the
|
|
22
|
+
# session allowlist wave through in bulk, and a recursive delete is not something
|
|
23
|
+
# a blanket "always allow file edits" should cover. It stays one prompt per call.
|
|
24
|
+
_WORKSPACE_EDIT_TOOLS = frozenset({"write_file", "edit_file"})
|
|
25
|
+
_SHELL_TOOLS = frozenset({"execute"})
|
|
26
|
+
|
|
27
|
+
# Session allowlist categories.
|
|
28
|
+
ALLOWLIST_FILE_EDITS = "file_edits"
|
|
29
|
+
ALLOWLIST_SHELL = "shell"
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
def interrupt_on_config() -> dict[str, InterruptOnConfig]:
|
|
33
|
+
"""``interrupt_on`` map keyed by tool name (used by `create_deep_agent`)."""
|
|
34
|
+
|
|
35
|
+
return {
|
|
36
|
+
name: {"allowed_decisions": ["approve", "reject"], "description": description}
|
|
37
|
+
for name, description in _APPROVAL_TOOLS.items()
|
|
38
|
+
}
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
class ApprovalMode(StrEnum):
|
|
42
|
+
"""How side-effecting Deep Agents tools are gated before execution."""
|
|
43
|
+
|
|
44
|
+
ASK = "ask"
|
|
45
|
+
"""Prompt before every gated tool (default)."""
|
|
46
|
+
|
|
47
|
+
WORKSPACE = "workspace"
|
|
48
|
+
"""Auto-allow workspace write_file and edit_file; still prompt for the rest."""
|
|
49
|
+
|
|
50
|
+
AUTO = "auto"
|
|
51
|
+
"""Auto-allow all configured approval tools (non-interactive trust mode)."""
|
|
52
|
+
|
|
53
|
+
def display_label(self) -> str:
|
|
54
|
+
return {
|
|
55
|
+
ApprovalMode.ASK: "default",
|
|
56
|
+
ApprovalMode.WORKSPACE: "accept edits",
|
|
57
|
+
ApprovalMode.AUTO: "auto",
|
|
58
|
+
}[self]
|
|
59
|
+
|
|
60
|
+
def cycle_next(self) -> ApprovalMode:
|
|
61
|
+
order = (ApprovalMode.ASK, ApprovalMode.WORKSPACE, ApprovalMode.AUTO)
|
|
62
|
+
index = order.index(self)
|
|
63
|
+
return order[(index + 1) % len(order)]
|
|
64
|
+
|
|
65
|
+
|
|
66
|
+
class ToolAllowlist:
|
|
67
|
+
"""Per-session categories the user chose to always allow."""
|
|
68
|
+
|
|
69
|
+
def __init__(self) -> None:
|
|
70
|
+
self._categories: set[str] = set()
|
|
71
|
+
|
|
72
|
+
def add(self, category: str) -> None:
|
|
73
|
+
self._categories.add(category)
|
|
74
|
+
|
|
75
|
+
def contains(self, category: str) -> bool:
|
|
76
|
+
return category in self._categories
|
|
77
|
+
|
|
78
|
+
def __bool__(self) -> bool:
|
|
79
|
+
return bool(self._categories)
|
|
80
|
+
|
|
81
|
+
def categories(self) -> frozenset[str]:
|
|
82
|
+
return frozenset(self._categories)
|
|
83
|
+
|
|
84
|
+
|
|
85
|
+
def tool_allowlist_category(tool_name: str) -> str | None:
|
|
86
|
+
if tool_name in _WORKSPACE_EDIT_TOOLS:
|
|
87
|
+
return ALLOWLIST_FILE_EDITS
|
|
88
|
+
if tool_name in _SHELL_TOOLS:
|
|
89
|
+
return ALLOWLIST_SHELL
|
|
90
|
+
return None
|
|
91
|
+
|
|
92
|
+
|
|
93
|
+
def categories_for_interrupt(value: Any) -> frozenset[str]:
|
|
94
|
+
if not isinstance(value, dict):
|
|
95
|
+
return frozenset()
|
|
96
|
+
action_requests = value.get("action_requests")
|
|
97
|
+
if not isinstance(action_requests, list):
|
|
98
|
+
return frozenset()
|
|
99
|
+
categories: set[str] = set()
|
|
100
|
+
for action in action_requests:
|
|
101
|
+
if not isinstance(action, dict):
|
|
102
|
+
continue
|
|
103
|
+
category = tool_allowlist_category(str(action.get("name") or ""))
|
|
104
|
+
if category:
|
|
105
|
+
categories.add(category)
|
|
106
|
+
return frozenset(categories)
|
|
107
|
+
|
|
108
|
+
|
|
109
|
+
def interrupt_matches_allowlist(value: Any, allowlist: ToolAllowlist) -> bool:
|
|
110
|
+
categories = categories_for_interrupt(value)
|
|
111
|
+
if not categories:
|
|
112
|
+
return False
|
|
113
|
+
return all(allowlist.contains(category) for category in categories)
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
def always_allow_categories(value: Any) -> frozenset[str]:
|
|
117
|
+
"""Categories a front end may offer to 'always allow' for this interrupt.
|
|
118
|
+
|
|
119
|
+
Mirrors the TUI's policy so every interface behaves the same: shell approvals
|
|
120
|
+
are deliberately one-off (a blanket 'always allow shell' is too broad), so a
|
|
121
|
+
shell-only interrupt offers nothing; file edits and other categories do.
|
|
122
|
+
"""
|
|
123
|
+
categories = categories_for_interrupt(value)
|
|
124
|
+
if not categories or categories == {ALLOWLIST_SHELL}:
|
|
125
|
+
return frozenset()
|
|
126
|
+
return categories
|
|
127
|
+
|
|
128
|
+
|
|
129
|
+
def parse_approval_mode(value: str | None) -> ApprovalMode:
|
|
130
|
+
if not value:
|
|
131
|
+
return ApprovalMode.ASK
|
|
132
|
+
normalized = value.strip().lower()
|
|
133
|
+
for mode in ApprovalMode:
|
|
134
|
+
if mode == normalized:
|
|
135
|
+
return mode
|
|
136
|
+
raise ValueError(
|
|
137
|
+
f"unknown approval mode {value!r}; expected one of: "
|
|
138
|
+
+ ", ".join(m.value for m in ApprovalMode)
|
|
139
|
+
)
|
|
140
|
+
|
|
141
|
+
|
|
142
|
+
def interrupt_on_for_mode(mode: ApprovalMode) -> dict[str, InterruptOnConfig]:
|
|
143
|
+
"""Map an approval mode to Deep Agents ``interrupt_on`` configuration."""
|
|
144
|
+
|
|
145
|
+
if mode == ApprovalMode.AUTO:
|
|
146
|
+
return {}
|
|
147
|
+
if mode == ApprovalMode.WORKSPACE:
|
|
148
|
+
# "Accept edits" auto-allows exactly the workspace edit tools; anything
|
|
149
|
+
# else that is gated at all stays gated, so a newly gated tool is not
|
|
150
|
+
# silently waved through by this mode.
|
|
151
|
+
full = interrupt_on_config()
|
|
152
|
+
return {name: cfg for name, cfg in full.items() if name not in _WORKSPACE_EDIT_TOOLS}
|
|
153
|
+
return interrupt_on_config()
|
|
154
|
+
|
|
155
|
+
|
|
156
|
+
def interrupt_is_workspace_edits_only(value: Any) -> bool:
|
|
157
|
+
"""True when every action in an interrupt is write_file or edit_file."""
|
|
158
|
+
|
|
159
|
+
if not isinstance(value, dict):
|
|
160
|
+
return False
|
|
161
|
+
action_requests = value.get("action_requests")
|
|
162
|
+
if not isinstance(action_requests, list) or not action_requests:
|
|
163
|
+
return False
|
|
164
|
+
for action in action_requests:
|
|
165
|
+
if not isinstance(action, dict):
|
|
166
|
+
return False
|
|
167
|
+
if action.get("name") not in _WORKSPACE_EDIT_TOOLS:
|
|
168
|
+
return False
|
|
169
|
+
return True
|
|
170
|
+
|
|
171
|
+
|
|
172
|
+
def should_auto_approve_interrupt(
|
|
173
|
+
value: Any,
|
|
174
|
+
mode: ApprovalMode,
|
|
175
|
+
*,
|
|
176
|
+
allowlist: ToolAllowlist | None = None,
|
|
177
|
+
) -> bool:
|
|
178
|
+
if allowlist is not None and interrupt_matches_allowlist(value, allowlist):
|
|
179
|
+
return True
|
|
180
|
+
if mode == ApprovalMode.AUTO:
|
|
181
|
+
return True
|
|
182
|
+
if mode == ApprovalMode.WORKSPACE:
|
|
183
|
+
return interrupt_is_workspace_edits_only(value)
|
|
184
|
+
return False
|
|
185
|
+
|
|
186
|
+
|
|
187
|
+
# ---------------------------------------------------------------------------
|
|
188
|
+
# Interrupt decisions and resume payloads.
|
|
189
|
+
#
|
|
190
|
+
# Shared across interfaces (TUI, server) so every front end resolves an
|
|
191
|
+
# approval interrupt the same way. An interrupt payload follows deepagents'
|
|
192
|
+
# contract: a dict with an ``action_requests`` list and a parallel
|
|
193
|
+
# ``review_configs`` list of allowed-decision policies.
|
|
194
|
+
|
|
195
|
+
# The decisions a front end can resolve an interrupt with.
|
|
196
|
+
SUPPORTED_APPROVAL_DECISIONS = frozenset({"approve", "reject", "respond"})
|
|
197
|
+
|
|
198
|
+
|
|
199
|
+
class SupportsInterrupt(Protocol):
|
|
200
|
+
"""A pending interrupt: an id and its raw payload value."""
|
|
201
|
+
|
|
202
|
+
interrupt_id: str
|
|
203
|
+
value: Any
|
|
204
|
+
|
|
205
|
+
|
|
206
|
+
def review_config_map(review_configs: Any) -> dict[str, frozenset[str]]:
|
|
207
|
+
"""Map each action name to its allowed decisions from a ``review_configs`` list."""
|
|
208
|
+
|
|
209
|
+
config_map: dict[str, frozenset[str]] = {}
|
|
210
|
+
if not isinstance(review_configs, list):
|
|
211
|
+
return config_map
|
|
212
|
+
for review in review_configs:
|
|
213
|
+
if not isinstance(review, dict):
|
|
214
|
+
continue
|
|
215
|
+
action_name = review.get("action_name")
|
|
216
|
+
allowed = review.get("allowed_decisions")
|
|
217
|
+
if isinstance(action_name, str) and isinstance(allowed, list):
|
|
218
|
+
config_map[action_name] = frozenset(str(item) for item in allowed)
|
|
219
|
+
return config_map
|
|
220
|
+
|
|
221
|
+
|
|
222
|
+
def allowed_decisions_for_interrupt(value: Any) -> frozenset[str]:
|
|
223
|
+
"""Return the intersection of decisions allowed across an interrupt's actions.
|
|
224
|
+
|
|
225
|
+
An interrupt may bundle multiple ``action_requests``; each can declare its
|
|
226
|
+
own ``allowed_decisions`` in a sibling ``review_configs`` block. A front end
|
|
227
|
+
can only resume an interrupt with a decision every action accepts, so we
|
|
228
|
+
intersect. Empty / malformed payloads fall back to all supported decisions,
|
|
229
|
+
matching deepagents' default.
|
|
230
|
+
"""
|
|
231
|
+
|
|
232
|
+
if not isinstance(value, dict):
|
|
233
|
+
return SUPPORTED_APPROVAL_DECISIONS
|
|
234
|
+
|
|
235
|
+
config_map = review_config_map(value.get("review_configs"))
|
|
236
|
+
action_requests = value.get("action_requests")
|
|
237
|
+
if not isinstance(action_requests, list):
|
|
238
|
+
return SUPPORTED_APPROVAL_DECISIONS
|
|
239
|
+
|
|
240
|
+
shared: frozenset[str] | None = None
|
|
241
|
+
for action in action_requests:
|
|
242
|
+
if not isinstance(action, dict):
|
|
243
|
+
continue
|
|
244
|
+
action_name = str(action.get("name") or "")
|
|
245
|
+
allowed = config_map.get(action_name, SUPPORTED_APPROVAL_DECISIONS)
|
|
246
|
+
shared = allowed if shared is None else shared & allowed
|
|
247
|
+
return shared if shared is not None else SUPPORTED_APPROVAL_DECISIONS
|
|
248
|
+
|
|
249
|
+
|
|
250
|
+
def pending_allowed_decisions(interrupts: Iterable[SupportsInterrupt]) -> frozenset[str]:
|
|
251
|
+
"""Decisions every pending interrupt accepts, scoped to the supported set."""
|
|
252
|
+
|
|
253
|
+
shared: frozenset[str] | None = None
|
|
254
|
+
for interrupt in interrupts:
|
|
255
|
+
allowed = allowed_decisions_for_interrupt(interrupt.value)
|
|
256
|
+
shared = allowed if shared is None else shared & allowed
|
|
257
|
+
if shared is None:
|
|
258
|
+
return frozenset()
|
|
259
|
+
return shared & SUPPORTED_APPROVAL_DECISIONS
|
|
260
|
+
|
|
261
|
+
|
|
262
|
+
def build_resume_payload(
|
|
263
|
+
interrupts: Iterable[SupportsInterrupt],
|
|
264
|
+
decision: dict[str, str],
|
|
265
|
+
) -> dict[str, dict[str, list[dict[str, str]]]]:
|
|
266
|
+
"""Apply one decision to every action of every pending interrupt, for ``resume``.
|
|
267
|
+
|
|
268
|
+
The result is the ``Command(resume=...)`` payload deepagents expects: keyed
|
|
269
|
+
by interrupt id, each carrying one decision per ``action_request``.
|
|
270
|
+
"""
|
|
271
|
+
|
|
272
|
+
payload: dict[str, dict[str, list[dict[str, str]]]] = {}
|
|
273
|
+
for interrupt in interrupts:
|
|
274
|
+
value = interrupt.value if isinstance(interrupt.value, dict) else {}
|
|
275
|
+
action_requests = value.get("action_requests")
|
|
276
|
+
count = len(action_requests) if isinstance(action_requests, list) else 1
|
|
277
|
+
payload[interrupt.interrupt_id] = {"decisions": [deepcopy(decision) for _ in range(count)]}
|
|
278
|
+
return payload
|
makker/agent/backend.py
ADDED
|
@@ -0,0 +1,160 @@
|
|
|
1
|
+
"""Filesystem backend for the agent's workspace.
|
|
2
|
+
|
|
3
|
+
``WorkspaceShellBackend`` is the real-path default: it reads and writes the real
|
|
4
|
+
filesystem from the workspace, refuses writes into read-only roots (installed
|
|
5
|
+
package source in the shared Julia depot), and blocks shell ``julia``.
|
|
6
|
+
``WorkspaceCompositeBackend`` is the top-level composite, handling Windows
|
|
7
|
+
drive-absolute glob patterns.
|
|
8
|
+
"""
|
|
9
|
+
|
|
10
|
+
from __future__ import annotations
|
|
11
|
+
|
|
12
|
+
from pathlib import Path
|
|
13
|
+
|
|
14
|
+
from deepagents.backends import CompositeBackend, LocalShellBackend
|
|
15
|
+
from deepagents.backends.protocol import (
|
|
16
|
+
DeleteResult,
|
|
17
|
+
EditResult,
|
|
18
|
+
ExecuteResponse,
|
|
19
|
+
GlobResult,
|
|
20
|
+
WriteResult,
|
|
21
|
+
)
|
|
22
|
+
|
|
23
|
+
from makker.agent.windows_paths import split_windows_glob
|
|
24
|
+
|
|
25
|
+
_BLOCKED_INTERPRETERS = frozenset({"julia"})
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
def interpreter_invocation(command: str, names: frozenset[str] | tuple[str, ...]) -> str | None:
|
|
29
|
+
"""Name of the interpreter a shell command launches, or ``None``.
|
|
30
|
+
|
|
31
|
+
Looks at the executable position of each pipeline/sequence segment, so
|
|
32
|
+
``julia -e ...`` and ``echo x | julia`` match while an interpreter name
|
|
33
|
+
appearing as data (``ls ~/.julia/...``, ``grep julia src/``) does not.
|
|
34
|
+
Shared by the workspace backend's execute guard (which blocks only
|
|
35
|
+
``julia``, the session kernel exists precisely for it) and the bench
|
|
36
|
+
scorers that audit recorded execute calls.
|
|
37
|
+
"""
|
|
38
|
+
import re
|
|
39
|
+
|
|
40
|
+
for segment in re.split(r"[;|&]+|\$\(", command):
|
|
41
|
+
head = segment.strip().split()
|
|
42
|
+
if head and Path(head[0]).name in names:
|
|
43
|
+
return Path(head[0]).name
|
|
44
|
+
return None
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
_NO_JULIA_SHELL_MSG = (
|
|
48
|
+
"Error: spawning `julia` through the shell is disabled. Julia code runs in "
|
|
49
|
+
"the persistent session via `run_julia`: shared state, streamed output, no "
|
|
50
|
+
'cold start. Use `run_julia` for code and `include("file.jl")` for scripts.'
|
|
51
|
+
)
|
|
52
|
+
|
|
53
|
+
_READ_ONLY_MSG = (
|
|
54
|
+
"Error: '{path}' is read-only installed package source in the shared Julia "
|
|
55
|
+
"depot. Read and grep it freely (it's reference material), but don't edit "
|
|
56
|
+
"it: the depot is shared across projects. Write your own code in the "
|
|
57
|
+
"workspace, or to change the package itself, `Pkg.develop` it "
|
|
58
|
+
"(makker init --source-path ...) and edit that checkout."
|
|
59
|
+
)
|
|
60
|
+
|
|
61
|
+
|
|
62
|
+
class WorkspaceCompositeBackend(CompositeBackend):
|
|
63
|
+
"""Top-level composite that rewrites a Windows drive-absolute glob into a relative
|
|
64
|
+
pattern plus a base path (see :func:`split_windows_glob`): ``pathlib``'s
|
|
65
|
+
``rglob`` rejects absolute patterns, so the absolute ``glob`` form the skills
|
|
66
|
+
use silently matches nothing on Windows otherwise. A no-op off Windows.
|
|
67
|
+
"""
|
|
68
|
+
|
|
69
|
+
def glob(self, pattern: str, path: str | None = None) -> GlobResult:
|
|
70
|
+
pattern, path = split_windows_glob(pattern, path)
|
|
71
|
+
return super().glob(pattern, path=path)
|
|
72
|
+
|
|
73
|
+
async def aglob(self, pattern: str, path: str | None = None) -> GlobResult:
|
|
74
|
+
pattern, path = split_windows_glob(pattern, path)
|
|
75
|
+
return await super().aglob(pattern, path=path)
|
|
76
|
+
|
|
77
|
+
|
|
78
|
+
class WorkspaceShellBackend(LocalShellBackend):
|
|
79
|
+
"""Workspace default backend: real paths, with a guard against shell ``julia``.
|
|
80
|
+
|
|
81
|
+
Constructed with ``virtual_mode=False`` so the file tools speak the real
|
|
82
|
+
filesystem: a relative path resolves against the workspace (the backend's
|
|
83
|
+
``cwd``) and an absolute path as itself, the same way ``run_julia`` and
|
|
84
|
+
``execute`` resolve paths from that working directory. One path string
|
|
85
|
+
therefore names one file in the file tools, the shell, and the REPL. Package
|
|
86
|
+
source, skills, memory, and added folders are all read and written at their
|
|
87
|
+
real paths through this one backend.
|
|
88
|
+
|
|
89
|
+
``readonly_roots`` are real directories writes are refused under: the
|
|
90
|
+
shared Julia depot's installed package source. Reads and greps there are
|
|
91
|
+
fine (that's how the agent studies a package); ``write``, ``edit`` and
|
|
92
|
+
``delete`` are blocked, so the agent can't corrupt the depot for other
|
|
93
|
+
projects. A ``Pkg.develop`` checkout lives outside the depot and stays
|
|
94
|
+
writable, so the registry-vs-dev distinction falls out of the path, with no
|
|
95
|
+
special-casing.
|
|
96
|
+
|
|
97
|
+
``execute`` refuses to launch ``julia``: Julia belongs in the session
|
|
98
|
+
kernel, where state persists and output streams; a shell julia is a cold
|
|
99
|
+
process that shares nothing with the session. A prompt rule alone does not
|
|
100
|
+
reliably stop models from shelling out; the tool guaranteeing it does.
|
|
101
|
+
Other interpreters are deliberately not blocked: shell python and friends
|
|
102
|
+
are part of general competence, and scientific results are kept honest by
|
|
103
|
+
the bench's trace checks, not by a blanket ban.
|
|
104
|
+
"""
|
|
105
|
+
|
|
106
|
+
def __init__(self, *args, readonly_roots: tuple[Path, ...] = (), **kwargs) -> None:
|
|
107
|
+
super().__init__(*args, **kwargs)
|
|
108
|
+
self._readonly_roots = tuple(Path(root).resolve() for root in readonly_roots)
|
|
109
|
+
|
|
110
|
+
def _readonly_reason(self, file_path: str) -> str | None:
|
|
111
|
+
"""Corrective message if ``file_path`` lands inside a read-only root, else ``None``."""
|
|
112
|
+
if not self._readonly_roots:
|
|
113
|
+
return None
|
|
114
|
+
path = Path(file_path)
|
|
115
|
+
target = path if path.is_absolute() else self.cwd / path
|
|
116
|
+
try:
|
|
117
|
+
target = target.resolve()
|
|
118
|
+
except OSError:
|
|
119
|
+
return None
|
|
120
|
+
for root in self._readonly_roots:
|
|
121
|
+
if target == root or root in target.parents:
|
|
122
|
+
return _READ_ONLY_MSG.format(path=file_path)
|
|
123
|
+
return None
|
|
124
|
+
|
|
125
|
+
def write(self, file_path: str, content: str) -> WriteResult:
|
|
126
|
+
reason = self._readonly_reason(file_path)
|
|
127
|
+
if reason is not None:
|
|
128
|
+
return WriteResult(error=reason)
|
|
129
|
+
return super().write(file_path, content)
|
|
130
|
+
|
|
131
|
+
def edit(
|
|
132
|
+
self,
|
|
133
|
+
file_path: str,
|
|
134
|
+
old_string: str,
|
|
135
|
+
new_string: str,
|
|
136
|
+
replace_all: bool = False,
|
|
137
|
+
) -> EditResult:
|
|
138
|
+
reason = self._readonly_reason(file_path)
|
|
139
|
+
if reason is not None:
|
|
140
|
+
return EditResult(error=reason)
|
|
141
|
+
return super().edit(file_path, old_string, new_string, replace_all=replace_all)
|
|
142
|
+
|
|
143
|
+
def delete(self, file_path: str) -> DeleteResult:
|
|
144
|
+
# Deleting a directory removes it recursively, so the depot guard matters
|
|
145
|
+
# here more than for write/edit: one call could take out a package's
|
|
146
|
+
# whole source tree.
|
|
147
|
+
reason = self._readonly_reason(file_path)
|
|
148
|
+
if reason is not None:
|
|
149
|
+
return DeleteResult(error=reason)
|
|
150
|
+
return super().delete(file_path)
|
|
151
|
+
|
|
152
|
+
def execute(self, command: str, *, timeout: int | None = None) -> ExecuteResponse:
|
|
153
|
+
if interpreter_invocation(command, _BLOCKED_INTERPRETERS) is not None:
|
|
154
|
+
return ExecuteResponse(output=_NO_JULIA_SHELL_MSG, exit_code=2)
|
|
155
|
+
return super().execute(command, timeout=timeout)
|
|
156
|
+
|
|
157
|
+
async def aexecute(self, command: str, *, timeout: int | None = None) -> ExecuteResponse:
|
|
158
|
+
if interpreter_invocation(command, _BLOCKED_INTERPRETERS) is not None:
|
|
159
|
+
return ExecuteResponse(output=_NO_JULIA_SHELL_MSG, exit_code=2)
|
|
160
|
+
return await super().aexecute(command, timeout=timeout)
|