@yahyahammad/venastine 0.2.0 → 0.3.9

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 (152) hide show
  1. package/.env.example +9 -4
  2. package/CONFIG_ARCHITECTURE.md +2030 -0
  3. package/LICENSE +201 -201
  4. package/NOTICE +1 -1
  5. package/README.md +1034 -895
  6. package/agents/builtin/build.md +73 -0
  7. package/agents/builtin/compactor.md +75 -75
  8. package/agents/builtin/explore.md +77 -77
  9. package/agents/builtin/general.md +80 -0
  10. package/agents/builtin/pipeline-reviewer.md +100 -100
  11. package/agents/builtin/plan.md +106 -88
  12. package/agents/builtin/review.md +74 -74
  13. package/agents/builtin/test.md +65 -0
  14. package/agents/builtin/writer.md +67 -0
  15. package/agents/manager.py +68 -2
  16. package/agents/subagent_tool.py +472 -238
  17. package/agents/tui_commands.py +238 -151
  18. package/bin/venastine.mjs +150 -4
  19. package/bin/venastine.py +71 -0
  20. package/cli_commands.py +838 -0
  21. package/config.py +95 -1198
  22. package/config.yaml +1103 -0
  23. package/config_edit.py +1301 -0
  24. package/config_schema.py +971 -0
  25. package/config_update.py +749 -0
  26. package/core/agent_activity.py +256 -118
  27. package/core/ansi.py +452 -0
  28. package/core/approval.py +137 -115
  29. package/core/call_view.py +364 -0
  30. package/core/cancel.py +121 -0
  31. package/core/chatgpt_auth.py +747 -0
  32. package/core/claims_view.py +107 -0
  33. package/core/client.py +2715 -1559
  34. package/core/commands.py +460 -0
  35. package/core/compaction.py +1735 -971
  36. package/core/config_loader.py +1982 -1563
  37. package/core/connect.py +686 -0
  38. package/core/events.py +123 -61
  39. package/core/file_mentions.py +280 -0
  40. package/core/interaction.py +411 -328
  41. package/core/line_pattern.py +128 -0
  42. package/core/loop.py +2566 -1490
  43. package/core/memory.py +482 -370
  44. package/core/pipeline_models.py +190 -176
  45. package/core/provider_errors.py +234 -0
  46. package/core/reasoning/authorization.py +130 -130
  47. package/core/reasoning/base.py +219 -219
  48. package/core/reasoning/confidence_scoring.py +484 -468
  49. package/core/reasoning/events.py +149 -149
  50. package/core/reasoning/json_retry.py +19 -1
  51. package/core/reasoning/orchestrator.py +1561 -1382
  52. package/core/reasoning/output_writer.py +345 -286
  53. package/core/reasoning/payload_validation.py +524 -398
  54. package/core/reasoning/pipeline_storage.py +88 -22
  55. package/core/reasoning/review.py +659 -615
  56. package/core/reasoning/scholar.py +458 -434
  57. package/core/reasoning/source_corpus.py +433 -327
  58. package/core/reasoning/source_scoring.py +855 -855
  59. package/core/replay.py +315 -183
  60. package/core/session.py +36 -12
  61. package/core/session_view.py +259 -0
  62. package/core/session_wake.py +259 -0
  63. package/core/shell_sessions.py +1553 -0
  64. package/core/workspace_trust.py +1 -1
  65. package/credentials.py +148 -20
  66. package/database.py +333 -252
  67. package/{PRIVACY.md → docs/PRIVACY.md} +78 -66
  68. package/docs/SECURITY.md +108 -0
  69. package/docs/THIRD_PARTY_NOTICES.md +148 -0
  70. package/env_secrets.py +50 -9
  71. package/json_store.py +343 -123
  72. package/logging_setup.py +295 -185
  73. package/main.py +3433 -1961
  74. package/mcp_client/client.py +733 -499
  75. package/mcp_client/config.py +126 -21
  76. package/mcp_client/registration.py +343 -182
  77. package/memories/manager.py +147 -132
  78. package/memories/tui_commands.py +319 -244
  79. package/package.json +303 -73
  80. package/project_init/config_files.py +282 -278
  81. package/project_init/generator.py +728 -564
  82. package/project_init/manifest.py +18 -6
  83. package/project_init/tui_commands.py +147 -144
  84. package/prompts/critic_pass.md +24 -24
  85. package/prompts/revalidate.md +25 -25
  86. package/prompts/revise.md +2 -0
  87. package/prompts/source_grounding.md +66 -66
  88. package/prompts/system_prompts.py +316 -262
  89. package/providers.json.example +4 -2
  90. package/pyproject.toml +149 -100
  91. package/requirements.txt +295 -72
  92. package/safety/policy_enforcement.py +1082 -746
  93. package/security/askpass.py +419 -0
  94. package/security/capability.py +207 -177
  95. package/security/confine.py +484 -0
  96. package/security/hosted.py +182 -0
  97. package/security/permissions.py +237 -190
  98. package/security/posture.py +124 -12
  99. package/security/protected_paths.py +501 -309
  100. package/security/sandbox.py +4108 -985
  101. package/security/secrets.py +855 -0
  102. package/skills/builtin/math/formal-specification.md +67 -67
  103. package/skills/builtin/math/numerical-methods.md +66 -66
  104. package/skills/builtin/research/experiment-design.md +62 -62
  105. package/skills/builtin/research/reproducibility.md +55 -55
  106. package/skills/builtin/research/source-evaluation.md +70 -70
  107. package/skills/builtin/research/statistical-inference.md +63 -63
  108. package/skills/builtin/research/technical-writing.md +59 -59
  109. package/skills/builtin/software/code-review.md +61 -61
  110. package/skills/builtin/software/debugging.md +61 -61
  111. package/skills/builtin/software/test-design.md +67 -67
  112. package/skills/tui_commands.py +110 -111
  113. package/storage.py +1338 -1005
  114. package/tools/base.py +161 -1
  115. package/tools/builtin/_math_common.py +667 -377
  116. package/tools/builtin/_net_common.py +225 -19
  117. package/tools/builtin/arxiv.py +497 -39
  118. package/tools/builtin/ask_user.py +231 -213
  119. package/tools/builtin/discrete_math.py +35 -4
  120. package/tools/builtin/fetch_url.py +371 -153
  121. package/tools/builtin/file_ops.py +720 -412
  122. package/tools/builtin/geometry.py +48 -12
  123. package/tools/builtin/get_time.py +2 -2
  124. package/tools/builtin/linear_algebra.py +66 -23
  125. package/tools/builtin/load_skill.py +76 -66
  126. package/tools/builtin/logic.py +171 -12
  127. package/tools/builtin/pin.py +227 -193
  128. package/tools/builtin/probability_stats.py +12 -16
  129. package/tools/builtin/project_docs.py +464 -405
  130. package/tools/builtin/remember.py +175 -158
  131. package/tools/builtin/shell.py +980 -461
  132. package/tools/builtin/shell_sessions.py +613 -0
  133. package/tools/builtin/symbolic_math.py +80 -10
  134. package/tools/builtin/todo.py +149 -147
  135. package/tools/builtin/web_search.py +191 -13
  136. package/tools/context.py +159 -136
  137. package/tools/isolation.py +107 -34
  138. package/tools/registry.py +1242 -925
  139. package/tui/app.py +8318 -4100
  140. package/tui/app.tcss +747 -556
  141. package/tui/commands.py +26 -170
  142. package/tui/diffs.py +34 -13
  143. package/tui/history.py +128 -128
  144. package/tui/hosted_driver.py +102 -0
  145. package/tui/markdown.py +1329 -1279
  146. package/tui/meters.py +364 -268
  147. package/tui/ravens.py +146 -146
  148. package/tui/screens.py +1255 -943
  149. package/tui/themes.py +599 -559
  150. package/tui/widgets.py +3813 -2332
  151. package/SECURITY.md +0 -88
  152. package/THIRD_PARTY_NOTICES.md +0 -100
package/pyproject.toml CHANGED
@@ -1,100 +1,149 @@
1
- # Issue #144 -- requirements.txt was the whole of what this project said
2
- # about how to install it, and it is the one packaging format that cannot
3
- # express a Python floor or an optional group.
4
- #
5
- # Python 3.11 is a hard requirement, not a preference: mcp_client/client.py
6
- # uses asyncio.timeout (3.11+), and 27 sites use runtime `X | Y` annotations
7
- # (3.10+). On 3.10 a user got `AttributeError: module 'asyncio' has no
8
- # attribute 'timeout'` from inside the MCP connect path, AFTER
9
- # `pip install -r requirements.txt` reported success. requires-python turns
10
- # that into a refusal at install time.
11
-
12
- # setuptools>=77 is the floor for PEP 639: the SPDX `license = "Apache-2.0"` string below is
13
- # read as a classifier expression only from that release on. An older backend treats it as a
14
- # path and fails the build rather than warning, so this floor and that line move together.
15
- [build-system]
16
- requires = ["setuptools>=77"]
17
- build-backend = "setuptools.build_meta"
18
-
19
- [project]
20
- name = "venastine-research-harness"
21
- version = "0.2.0"
22
- description = "A from-scratch agentic harness: tool-using chat plus a ten-pass self-critiquing research pipeline."
23
- readme = "README.md"
24
- requires-python = ">=3.11"
25
- # PEP 639: a bare SPDX expression, not the deprecated `{ text = ... }` table, and no
26
- # `License :: OSI Approved ::` classifier -- PEP 639 supersedes those and setuptools warns on
27
- # both. `license-files` is declared rather than left to the backend's default glob so the file
28
- # that ships is the one named here.
29
- license = "Apache-2.0"
30
- # NOTICE ships WITH the licence, not beside it in the repo only: a NOTICE that
31
- # is not listed here exists in a checkout and is absent from every sdist and
32
- # wheel, which defeats the reason for having one. Apache-2.0 4(d) is what makes
33
- # the pairing meaningful downstream.
34
- license-files = ["LICENSE", "NOTICE"]
35
- authors = [{ name = "Yahya Hammad" }]
36
- maintainers = [{ name = "Yahya Hammad" }]
37
- classifiers = [
38
- "Programming Language :: Python :: 3",
39
- "Programming Language :: Python :: 3.11",
40
- "Programming Language :: Python :: 3.12",
41
- "Programming Language :: Python :: 3.13",
42
- "Intended Audience :: Science/Research",
43
- "Topic :: Scientific/Engineering :: Artificial Intelligence",
44
- ]
45
-
46
- # ONE dependency list, and it stays in requirements.txt -- the file that
47
- # already carries the D22 rationale comments explaining why `textual` and
48
- # `mcp` are upper-bounded. Inlining the list here would create a second
49
- # hand-maintained copy of the same fact, which is the shape #144 was filed
50
- # about in the first place ("a hand-maintained list is only as complete as
51
- # the file it lives in"). setuptools strips comments and blank lines.
52
- #
53
- # THIS KEY MUST STAY INSIDE [project], above the sub-tables. TOML scopes a
54
- # bare key to the most recent table header, so when this sat below
55
- # [project.urls] it parsed as `project.urls.dynamic` and setuptools refused
56
- # the whole file: "`project.urls.dynamic` must be string". `pip install -e .`
57
- # failed outright, which also meant requires-python -- the entire point of
58
- # #144 -- was never enforced, because the install never got far enough to
59
- # check it. Moving a key past a table header here is not cosmetic.
60
- dynamic = ["dependencies"]
61
-
62
- [project.urls]
63
- Homepage = "https://github.com/YahyaHammad/Venastine-Research-Harness"
64
- Repository = "https://github.com/YahyaHammad/Venastine-Research-Harness"
65
- Issues = "https://github.com/YahyaHammad/Venastine-Research-Harness/issues"
66
-
67
- [project.optional-dependencies]
68
- # markitdown's only consumer is tools/builtin/file_ops.py, which implements
69
- # read/write/edit -- all three globally denied, and per S24 they cannot be
70
- # re-enabled at runtime at all (is_tool_allowed reads config.ToolPermissions
71
- # directly, settings.json has no permissions section, and D14 forbids a
72
- # ToolContext widening anything). The import is lazy, so a default install
73
- # never reaches it. It brought 11 transitive distributions, 5 of them for
74
- # markitdown alone.
75
- #
76
- # protobuf moves with it rather than staying in the default set: nothing in
77
- # the repo imports it, and the only installed distribution requiring it is
78
- # onnxruntime, which arrives through markitdown. Left behind it would pin a
79
- # transitive dependency of a package no longer installed by default.
80
- documents = ["markitdown==0.1.6", "protobuf==7.35.1"]
81
-
82
- # Optional in the code too, and documented as such at
83
- # core/reasoning/output_writer.py:185 -- the PDF path is guarded by
84
- # `except ImportError: return`. Declared here so the group has a name;
85
- # neither was ever in requirements.txt.
86
- pdf = ["markdown", "weasyprint"]
87
-
88
- [tool.setuptools.dynamic]
89
- dependencies = { file = ["requirements.txt"] }
90
-
91
- [tool.setuptools]
92
- # DECLARATIVE ONLY -- this project is run from a checkout, not installed.
93
- # `pip install .` would otherwise put `config`, `main`, `storage` and
94
- # `conftest` into site-packages as top-level names, where they collide with
95
- # anything. Empty lists mean `pip install -e .` resolves the dependencies
96
- # and enforces the Python floor without installing a single module, which
97
- # is exactly what is wanted here. README's `pip install -r requirements.txt`
98
- # remains the documented install path.
99
- packages = []
100
- py-modules = []
1
+ # Issue #144 -- requirements.txt was the whole of what this project said
2
+ # about how to install it, and it is the one packaging format that cannot
3
+ # express a Python floor or an optional group.
4
+ #
5
+ # Python 3.11 is a hard requirement, not a preference: mcp_client/client.py
6
+ # uses asyncio.timeout (3.11+), and 27 sites use runtime `X | Y` annotations
7
+ # (3.10+). On 3.10 a user got `AttributeError: module 'asyncio' has no
8
+ # attribute 'timeout'` from inside the MCP connect path, AFTER
9
+ # `pip install -r requirements.txt` reported success. requires-python turns
10
+ # that into a refusal at install time.
11
+
12
+ # setuptools>=77 is the floor for PEP 639: the SPDX `license = "Apache-2.0"` string below is
13
+ # read as a classifier expression only from that release on. An older backend treats it as a
14
+ # path and fails the build rather than warning, so this floor and that line move together.
15
+ [build-system]
16
+ requires = ["setuptools>=77"]
17
+ build-backend = "setuptools.build_meta"
18
+
19
+ [project]
20
+ name = "venastine-research-harness"
21
+ version = "0.3.9"
22
+ description = "A from-scratch agentic harness: tool-using chat plus a ten-pass self-critiquing research pipeline."
23
+ readme = "README.md"
24
+ requires-python = ">=3.11"
25
+ # PEP 639: a bare SPDX expression, not the deprecated `{ text = ... }` table, and no
26
+ # `License :: OSI Approved ::` classifier -- PEP 639 supersedes those and setuptools warns on
27
+ # both. `license-files` is declared rather than left to the backend's default glob so the file
28
+ # that ships is the one named here.
29
+ license = "Apache-2.0"
30
+ # NOTICE ships WITH the licence, not beside it in the repo only: a NOTICE that
31
+ # is not listed here exists in a checkout and is absent from every sdist and
32
+ # wheel, which defeats the reason for having one. Apache-2.0 4(d) is what makes
33
+ # the pairing meaningful downstream.
34
+ license-files = ["LICENSE", "NOTICE"]
35
+ authors = [{ name = "Yahya Hammad" }]
36
+ maintainers = [{ name = "Yahya Hammad" }]
37
+ classifiers = [
38
+ "Programming Language :: Python :: 3",
39
+ "Programming Language :: Python :: 3.11",
40
+ "Programming Language :: Python :: 3.12",
41
+ "Programming Language :: Python :: 3.13",
42
+ "Intended Audience :: Science/Research",
43
+ "Topic :: Scientific/Engineering :: Artificial Intelligence",
44
+ ]
45
+
46
+ # ONE dependency list, and it stays in requirements.txt -- the file that
47
+ # already carries the D22 rationale comments explaining why `textual` and
48
+ # `mcp` are upper-bounded. Inlining the list here would create a second
49
+ # hand-maintained copy of the same fact, which is the shape #144 was filed
50
+ # about in the first place ("a hand-maintained list is only as complete as
51
+ # the file it lives in"). setuptools strips comments and blank lines.
52
+ #
53
+ # THIS KEY MUST STAY INSIDE [project], above the sub-tables. TOML scopes a
54
+ # bare key to the most recent table header, so when this sat below
55
+ # [project.urls] it parsed as `project.urls.dynamic` and setuptools refused
56
+ # the whole file: "`project.urls.dynamic` must be string". `pip install -e .`
57
+ # failed outright, which also meant requires-python -- the entire point of
58
+ # #144 -- was never enforced, because the install never got far enough to
59
+ # check it. Moving a key past a table header here is not cosmetic.
60
+ dynamic = ["dependencies"]
61
+
62
+ [project.urls]
63
+ Homepage = "https://github.com/YahyaHammad/Venastine-Research-Harness"
64
+ Repository = "https://github.com/YahyaHammad/Venastine-Research-Harness"
65
+ Issues = "https://github.com/YahyaHammad/Venastine-Research-Harness/issues"
66
+
67
+ [project.optional-dependencies]
68
+ # markitdown's only consumer is tools/builtin/file_ops.py, which implements
69
+ # read/write/edit -- all three globally denied, and per S24 they cannot be
70
+ # re-enabled at runtime at all (is_tool_allowed reads config.ToolPermissions
71
+ # directly, settings.json has no permissions section, and D14 forbids a
72
+ # ToolContext widening anything). The import is lazy, so a default install
73
+ # never reaches it. It brought 11 transitive distributions, 5 of them for
74
+ # markitdown alone.
75
+ #
76
+ # protobuf moves with it rather than staying in the default set: nothing in
77
+ # the repo imports it, and the only installed distribution requiring it is
78
+ # onnxruntime, which arrives through markitdown. Left behind it would pin a
79
+ # transitive dependency of a package no longer installed by default.
80
+ documents = ["markitdown==0.1.9", "protobuf==7.36.2"]
81
+
82
+ # Optional in the code too, and documented as such at
83
+ # core/reasoning/output_writer.py:185 -- the PDF path is guarded by
84
+ # `except ImportError: return`. Declared here so the group has a name;
85
+ # neither was ever in requirements.txt.
86
+ pdf = ["markdown", "weasyprint"]
87
+
88
+ # S18-07, batch 136. These four were at the bottom of requirements.txt, and
89
+ # `dynamic = ["dependencies"]` above reads that file, so they were hard
90
+ # requirements of the DISTRIBUTION: `pip install venastine-research-harness`
91
+ # resolved a test runner and its scheduler, and bin/venastine.mjs installed
92
+ # the same set into every end user's venv.
93
+ #
94
+ # NAMED HERE RATHER THAN READ FROM THE FILE, because setuptools'
95
+ # `optional-dependencies` is all-or-nothing: making this one group dynamic
96
+ # would force `documents` and `pdf` into files of their own, which is churn
97
+ # to two groups that work in order to avoid typing four names. The names are
98
+ # therefore in two places, and `test_the_test_extra_matches_requirements_test`
99
+ # is what stops them drifting -- #144's rule is that a hand-maintained
100
+ # duplicate is only as good as the gate over it.
101
+ #
102
+ # Floors, not pins, matching requirements-test.txt: D22's argument for
103
+ # pinning is about packages whose behaviour CI's RESULT depends on.
104
+ test = [
105
+ "pytest>=8.0",
106
+ "pytest-mock>=3.14",
107
+ "pytest-asyncio>=1.0",
108
+ "pytest-xdist>=3.6",
109
+ "starlette>=0.40",
110
+ ]
111
+
112
+ [tool.setuptools.dynamic]
113
+ dependencies = { file = ["requirements.txt"] }
114
+
115
+ [tool.ruff]
116
+ # Lint gate (lint.yml), judged against the 3.11 floor even though
117
+ # developers may run 3.12/3.13 locally -- UP modernisations must be valid
118
+ # where CI runs, not just where they were written.
119
+ target-version = "py311"
120
+
121
+ [tool.ruff.lint]
122
+ select = ["E", "F", "I", "UP"]
123
+ ignore = [
124
+ # Long docstring/comment lines are the record (decision ids, trace
125
+ # excerpts, URLs). Reflowing them for a ruler would bury the content.
126
+ "E501",
127
+ # `Optional[X]` is this codebase's idiom (300+ sites); `X | None`
128
+ # rewrites are churn with zero behaviour change.
129
+ "UP045",
130
+ # Deferred imports are architectural: tools/registry.py <-> agents/
131
+ # subagent_tool.py would cycle at import time, and several tests import
132
+ # only after arranging fixtures/patches.
133
+ "E402",
134
+ # printf-style `%` is the logging convention (AGENTS.md: interpolate via
135
+ # logger args, never f-strings at the call site) and the user-facing
136
+ # string idiom; mass conversion is churn for no behaviour change.
137
+ "UP031",
138
+ ]
139
+
140
+ [tool.setuptools]
141
+ # DECLARATIVE ONLY -- this project is run from a checkout, not installed.
142
+ # `pip install .` would otherwise put `config`, `main`, `storage` and
143
+ # `conftest` into site-packages as top-level names, where they collide with
144
+ # anything. Empty lists mean `pip install -e .` resolves the dependencies
145
+ # and enforces the Python floor without installing a single module, which
146
+ # is exactly what is wanted here. README's `pip install -r requirements.txt`
147
+ # remains the documented install path.
148
+ packages = []
149
+ py-modules = []
package/requirements.txt CHANGED
@@ -1,72 +1,295 @@
1
- # The Python floor is 3.11 (mcp_client/client.py uses asyncio.timeout).
2
- # requirements.txt cannot express that; pyproject.toml's requires-python
3
- # can, and reads THIS file for the list below via dynamic dependencies --
4
- # so there is one dependency list, not two. See issue #144.
5
- #
6
- # markitdown and protobuf moved to pyproject's `documents` extra: their
7
- # only reachable consumer is behind read/write/edit, which are globally
8
- # denied. `pip install -e ".[documents]"` brings them back.
9
-
10
- anthropic==0.116.0
11
- ddgs==9.14.4
12
- httpx==0.28.1
13
- openai==2.45.0
14
- # NOTE (ROADMAP_v2 §16): this pin has no ThinkingConfig.thinking_budget --
15
- # that field arrived in a later release. core/client.py checks for it at
16
- # runtime and reports no effort levels on GOOGLE rather than sending a
17
- # field the SDK cannot carry. Moving this pin enables Google effort; it
18
- # also puts §9's verified Google implementation back in scope for retest.
19
- google-genai==1.0.0
20
- pydantic==2.13.4
21
- sympy==1.13.3
22
- sqlmodel==0.0.39
23
- python-dotenv==1.2.2
24
- matplotlib==3.10.3
25
- pyyaml==6.0.3
26
- # ROADMAP_v2 §16 TUI. Upper-bounded per D22 -- the mcp lesson was that an
27
- # unpinned dependency mid-major-version-transition migrates you by accident.
28
- textual>=1.0,<2.0
29
-
30
- # Imported DIRECTLY at tui/widgets.py:8-9 (`rich.syntax.Syntax`,
31
- # `rich.text.Text`) and previously undeclared -- present only because
32
- # textual, mcp and httpx all happen to require it, so its major version was
33
- # chosen by whatever textual's own range resolved to. Upper-bounded per D22
34
- # for D22's own reason, and this was the one package the project imported
35
- # without declaring, i.e. the one place such a migration would be silent.
36
- rich>=13,<16
37
-
38
- # ROADMAP_v2 §17 MCP client. Upper-bounded per D22 -- and D22 exists
39
- # BECAUSE of this package: it was mid-major-version transition when §17
40
- # was written, and 2.0.0 landed 2026-07-28 before implementation started.
41
- #
42
- # v2 is a real break for consumers, verified in hand rather than from the
43
- # docs: result fields are snake_case (.content/.structured_content/
44
- # .is_error, was camelCase), timeouts are float seconds (was timedelta),
45
- # the exception is MCPError (was McpError), streamablehttp_client is
46
- # removed, and httpx/httpx-sse are replaced by httpx2 -- which is why
47
- # there are now two HTTP stacks installed. httpx stays; anthropic, openai
48
- # and ddgs all still need it.
49
- # EXACT, not the `>=2.0,<3` range this used to carry (batch 32, CI run
50
- # 32778904550). D22's upper bound did its job -- no major migration -- but a
51
- # MINOR still changed observable behaviour underneath it: 2.1.0 stopped
52
- # passing a crashed tool's exception text back to the client ("the
53
- # exception's own text stays on the server"), so a test asserting the
54
- # message went red on main while every local run stayed green. The range
55
- # meant CI's result depended on WHEN it ran, and CI was not testing what
56
- # anyone runs locally -- which is the more serious half of that.
57
- #
58
- # So this is pinned the way anthropic, openai and google-genai already are.
59
- # Moving it is now a deliberate step: read 2.1.0's split first -- a tool
60
- # raising ToolError on purpose keeps its message, a tool that CRASHES has
61
- # its text withheld -- and re-run tests/test_mcp_client.py, which is written
62
- # to hold on both versions.
63
- mcp==2.0.0
64
-
65
- # --- Test dependencies -------------------------------------------------
66
- # Previously absent, which the docs recorded as a known open item since
67
- # §14: the suite has always needed pytest and pytest-mock, and §16's TUI
68
- # pilot tests add pytest-asyncio. Listed now so `pip install -r` produces
69
- # a working suite instead of one that fails at import.
70
- pytest>=8.0
71
- pytest-mock>=3.14
72
- pytest-asyncio>=1.0
1
+ # The Python floor is 3.11 (mcp_client/client.py uses asyncio.timeout).
2
+ # requirements.txt cannot express that; pyproject.toml's requires-python
3
+ # can, and reads THIS file for the list below via dynamic dependencies --
4
+ # so there is one dependency list, not two. See issue #144.
5
+ #
6
+ # markitdown and protobuf moved to pyproject's `documents` extra: their
7
+ # only reachable consumer is behind read/write/edit, which are globally
8
+ # denied. `pip install -e ".[documents]"` brings them back.
9
+
10
+ # Batch 234, from Dependabot's group (PR #271): anthropic 0.116.0 -> 2.1.0
11
+ # and openai 2.45.0 -> 3.36.0, each the newest release at least two weeks
12
+ # old, the owner's rule from batches 172 and 208 (the PR proposed 2.3.0 and
13
+ # 3.41.0). Both move their HTTP stack to httpx2. anthropic 2.0 also sends
14
+ # `anthropic-version: 2026-10-21` and the key as `Authorization: Bearer`;
15
+ # that API version refuses a temperature on every model, and 1.0 took the
16
+ # parameter out of `messages.stream()`, so core/client.py drops one for
17
+ # ANTHROPIC (ROADMAP_v3 A82). An empty key, which 0.116.0 refused as a
18
+ # TypeError, became a retried "Connection error." (httpx2 refuses the
19
+ # header `Bearer `), so api_initialization refuses it first. Measured with
20
+ # core/client.py's own calls against a local mock on both pin sets;
21
+ # test_sdk_conformance.py re-checks the shapes. What to re-measure when
22
+ # these move: that the stream, the effort query and the error types still
23
+ # match, since the suite fakes both packages everywhere else.
24
+ anthropic==2.1.0
25
+ ddgs==9.16.0
26
+ httpx==0.28.1
27
+ openai==3.36.0
28
+ # ROADMAP_v3 §82 (GS1), batch 172: 1.0.0 -> 2.29.0, the owner's choice.
29
+ # Gemini 3 refuses a request whose current turn holds a function call
30
+ # without the thought signature it came with; 1.0.0's Part had no field
31
+ # to hold one, and every model a new Google account can call is Gemini 3,
32
+ # so every GOOGLE tool loop failed on its second request. 2.29.0 was the
33
+ # newest release at least two weeks old when chosen. It also carries
34
+ # ThinkingConfig.thinking_budget, which core/client.py checks for at
35
+ # runtime, so Google reasoning effort is ON now (ROADMAP_v2 §16), and
36
+ # §9's Google implementation was retested live against it. Its
37
+ # transport is httpx now, not requests. 2.32.0 since batch 234, the newest
38
+ # release at least two weeks old; test_sdk_conformance.py is the evidence,
39
+ # with no live Gemini call (no key on this machine).
40
+ google-genai==2.32.0
41
+ # Batch 208, from Dependabot's group, the owner's choice of which: the
42
+ # packages below are ones the suite runs for real rather than faking, so
43
+ # a green suite is evidence about them, and each release was at least two
44
+ # weeks old. pydantic 2.14.0 fixed GHSA-gj4f-r7mf-jjxh (an unbounded
45
+ # Fraction exponent) and GHSA-955g-x97g-8mcp (quadratic multi-host URL
46
+ # validation). sympy 1.14.0 gives test_math_tools.py's measured expansion
47
+ # the same 74,723 characters 1.13.3 did.
48
+ #
49
+ # NO matplotlib SINCE BATCH 212, the owner's choice. It drew one bar chart
50
+ # per research run, and with numpy, fonttools, pillow and seven smaller
51
+ # packages it was 126 MB of a 367 MB install, measured in a fresh venv;
52
+ # core/reasoning/output_writer.py draws the chart as SVG with the standard
53
+ # library now. pillow's floor (batch 201) went with it: nothing else here
54
+ # reaches pillow.
55
+ pydantic==2.14.1
56
+ sympy==1.14.0
57
+ # 0.0.48 since batch 234: test_storage_e2e.py runs it for real.
58
+ sqlmodel==0.0.48
59
+ python-dotenv==1.2.4
60
+ pyyaml==6.0.3
61
+ # config.yaml, via config_schema.py. TWO YAML PARSERS IN ONE TREE IS
62
+ # DELIBERATE -- do not "unify" them. pyyaml above parses agent/skill
63
+ # frontmatter, where nothing is ever written back and ~70 tests pin the
64
+ # behaviour; ruamel is here for the half pyyaml cannot do, which is
65
+ # preserving a hand-authored file's comments across a WRITE. Since batch 84
66
+ # two modules do exactly that -- config_edit.py behind /config, and
67
+ # config_update.py putting a user's values back after an npm update replaces
68
+ # the file -- while the IMPORT path still uses typ="safe". (This said
69
+ # "nothing writes config.yaml yet ... this pin is bought for the config
70
+ # editor that will", which those batches made false.)
71
+ #
72
+ # EXACT, for the reason `mcp` and `textual` are exact: an upper bound lets a
73
+ # MINOR change behaviour underneath CI and makes CI's result depend on when
74
+ # it ran. Two facts to re-measure when this pin moves: 0.18 REMOVED the
75
+ # module-level `load()` / `safe_load()` functions, so the `YAML()` object API
76
+ # is the only one, and `typ="safe"` is what refuses `!!python/...` tags in a
77
+ # file that decides the tool permissions.
78
+ ruamel.yaml==0.19.1
79
+ # ROADMAP_v2 §16 TUI. EXACT, and this comment is textual's OWN record
80
+ # rather than a line borrowed from D22's, which is about `mcp`. The
81
+ # `>=1.0,<2.0` it replaces was written BEFORE 2.0 existed -- 1.0.0 shipped
82
+ # 2024-12-12 and 2.0.0 on 2025-02-16 -- so it was a
83
+ # don't-migrate-me-by-accident ceiling and never a judgement about 2.x.
84
+ # Nothing in the repo had evaluated 2.x, and the ceiling then sat there
85
+ # for twenty months while upstream reached 8.2.8.
86
+ #
87
+ # Exact for the reason batch 32 made `mcp` exact: an upper bound stops a
88
+ # MAJOR and lets a MINOR change behaviour underneath CI, which makes CI's
89
+ # result depend on when it ran.
90
+ #
91
+ # Moving it again is a deliberate step. Each of these was verified in hand
92
+ # against the installed package rather than read from a changelog, and
93
+ # each is what to re-measure next time:
94
+ # * `Static.renderable` is `.content` since 6.0.0 (`.visual` is the
95
+ # rendered form). 43 sites across 8 test files read it.
96
+ # * `Static.update` sends a Rich `Text` through `visualize()`, which
97
+ # converts it with `console=widget.app.console` -- so updating an
98
+ # UNMOUNTED widget raises NoActiveAppError, and the suite builds these
99
+ # panels bare everywhere. `tui.widgets.as_content` is the answer.
100
+ # The conversion also drops `no_wrap` / `overflow`, silently; those
101
+ # are `text-wrap` / `text-overflow` in app.tcss now.
102
+ # * Textual markup replaced Rich markup at 2.0.0, and `markup=False` IS
103
+ # honoured now -- on 1.0.0 it was stored and never read, which is what
104
+ # tui/screens.py's `Text(...)` rule was written against (RA1).
105
+ # * `textual-ansi` is GONE. `ansi-dark` / `ansi-light` replaced it at
106
+ # 8.2.5, `Theme` grew an `ansi: bool`, and the built-in roster is 21
107
+ # rather than 12 -- so SELECTABLE_THEMES in tests/test_themes.py, and
108
+ # every documented test count, moves with the pin.
109
+ # * `dim` is resolved into the drawn COLOUR now instead of reaching the
110
+ # cell as an attribute, which is what a test reading a drawn style
111
+ # off the compositor sees.
112
+ # * The kitty keyboard protocol reaches the WINDOWS driver since 6.6.0.
113
+ # Batch 54's `shift+enter` note was written against its absence.
114
+ # Then re-run the TUI pilots dependabot.yml names -- ~410 run_test sites.
115
+ textual==8.2.8
116
+
117
+ # Imported DIRECTLY at tui/widgets.py:8-9 (`rich.syntax.Syntax`,
118
+ # `rich.text.Text`) and previously undeclared -- present only because
119
+ # textual, mcp and httpx all happen to require it, so its major version was
120
+ # chosen by whatever textual's own range resolved to. Upper-bounded per D22
121
+ # for D22's own reason, and this was the one package the project imported
122
+ # without declaring, i.e. the one place such a migration would be silent.
123
+ # The FLOOR is textual's, not a guess: 8.2.8 requires `rich>=14.2.0`, so
124
+ # `>=13` stated a version this project can no longer be installed with.
125
+ rich>=14.2,<16
126
+
127
+ # ROADMAP_v2 §17 MCP client. Upper-bounded per D22 -- and D22 exists
128
+ # BECAUSE of this package: it was mid-major-version transition when §17
129
+ # was written, and 2.0.0 landed 2026-07-28 before implementation started.
130
+ #
131
+ # v2 is a real break for consumers, verified in hand rather than from the
132
+ # docs: result fields are snake_case (.content/.structured_content/
133
+ # .is_error, was camelCase), timeouts are float seconds (was timedelta),
134
+ # the exception is MCPError (was McpError), streamablehttp_client is
135
+ # removed, and httpx/httpx-sse are replaced by httpx2 -- which is why
136
+ # there are now two HTTP stacks installed. httpx stays: the harness imports
137
+ # it itself (fetch_url, arxiv, scholar, /connect, the hosted sign-in) and
138
+ # google-genai needs it. Since batch 234 anthropic and openai are on httpx2
139
+ # too, and ddgs needs neither (measured with `pip show` in a fresh venv).
140
+ # EXACT, not the `>=2.0,<3` range this used to carry (batch 32, CI run
141
+ # 32778904550). D22's upper bound did its job -- no major migration -- but a
142
+ # MINOR still changed observable behaviour underneath it: 2.1.0 stopped
143
+ # passing a crashed tool's exception text back to the client ("the
144
+ # exception's own text stays on the server"), so a test asserting the
145
+ # message went red on main while every local run stayed green. The range
146
+ # meant CI's result depended on WHEN it ran, and CI was not testing what
147
+ # anyone runs locally -- which is the more serious half of that.
148
+ #
149
+ # So this is pinned the way anthropic, openai and google-genai already are.
150
+ # Moving it is now a deliberate step: read 2.1.0's split first -- a tool
151
+ # raising ToolError on purpose keeps its message, a tool that CRASHES has
152
+ # its text withheld -- and re-run tests/test_mcp_client.py, which is written
153
+ # to hold on both versions.
154
+ #
155
+ # 2.2.0 SINCE BATCH 165's follow-up, for CVE-2026-59951: a Streamable HTTP
156
+ # SERVER kept every session's state with no expiry and no cap, so a client
157
+ # could grow its memory without bound. This harness is an MCP client and
158
+ # serves nothing, so it was never exposed -- the bump is because pip-audit
159
+ # fails on any advisory, and a fix existed. Taken the deliberate way, the
160
+ # split read first: 2.1.0 withholds a crashed tool's text and drops
161
+ # outputSchema/structuredContent for tools that return content blocks;
162
+ # 2.2.0 follows redirects only within the same origin (or http -> https on
163
+ # one host) and no longer honours a passed-in client's follow_redirects.
164
+ # Neither touches what this client does. test_mcp_client.py,
165
+ # test_mcp_config.py and the integration servers all pass on it, and
166
+ # mcp-types moves with it in lockstep.
167
+ #
168
+ # 2.3.0 SINCE BATCH 208, the deliberate way again. Its release notes'
169
+ # behaviour changes, against this client: httpx2 >=2.10.0 is required
170
+ # (it was 2.5.0); a 2025-11-25-or-earlier connection no longer sends an
171
+ # empty `_meta` or `params`, and `initialize` leaves `experimental` out,
172
+ # neither of which mcp_client/ sends or reads; a call the server rejects
173
+ # with HeaderMismatch is re-listed and retried once, and a rejected call
174
+ # never ran. ONE MORE CHANGE DOES TOUCH IT, which the notes' "the default
175
+ # stays 1 MiB" hides: on 2.2.0's httpx2 a streamed answer had no size limit,
176
+ # and 2.3.0 fails a call whose one server-sent event passes
177
+ # `max_sse_event_size`, 1 MiB by default, which a streamed screenshot does.
178
+ # The owner chose 16 MiB, as config.yaml's `mcp_max_event_bytes`;
179
+ # CONFIG_ARCHITECTURE.md has the measurement and what it does not cover.
180
+ # The registration fact in mcp_client/registration.py was re-measured on
181
+ # it.
182
+ mcp==2.3.0
183
+
184
+ # ROADMAP_v3 §49 (SS8). The engine for a monitor session's pattern -- a
185
+ # regex the MODEL wrote, matched against program output nobody controls,
186
+ # inside the harness's own process. Imported by core/line_pattern.py and
187
+ # nowhere else.
188
+ #
189
+ # NOT the stdlib `re`, and the reason was measured: `re` holds the GIL for
190
+ # the whole match, so `(a+)+$` against 25 a's and a `b` froze EVERY thread
191
+ # for 1.9 s -- the TUI's included -- and each extra character doubles it.
192
+ # RE2 matches in linear time, releases the GIL, and refuses backreferences
193
+ # and lookaround at compile time, so the agent is told instead of the
194
+ # harness hanging.
195
+ #
196
+ # EXACT, for the reason textual and mcp are exact. BSD-3-Clause; wheels
197
+ # for cp310-cp314 on Windows (x64, x86, arm64), macOS 13-15 (arm64,
198
+ # x86_64) and manylinux_2_28 (x86_64, aarch64). There is NO wheel for musl
199
+ # (Alpine) or glibc older than 2.28, where pip builds the sdist and needs a
200
+ # C++ toolchain and Abseil. Another package ships a module named `re2`
201
+ # that falls back to `re`, which is why core/line_pattern.py refuses a
202
+ # `re2` from any other distribution. What to re-measure when this moves:
203
+ # `re2.Options` still has `log_errors` and `max_mem`; a compile error is
204
+ # still `re2.error` and still silent with `log_errors = False` (it writes
205
+ # to stderr through Abseil otherwise, which paints over Textual); and the
206
+ # wheel list above.
207
+ google-re2==1.1.20251105
208
+
209
+ # ROADMAP_v3 §49 slice 5a (SS57). AES-256-GCM for the harness-held secret
210
+ # store, and scrypt is the stdlib's -- `hashlib` has the KDF and no cipher,
211
+ # so this is the dependency or a hand-rolled AEAD, and this repository
212
+ # ships a `cryptography-verification` skill whose whole subject is not
213
+ # doing the second one.
214
+ #
215
+ # DECLARING IT ADDS NOTHING TO AN INSTALL, measured rather than assumed:
216
+ # google-genai==1.0.0 -> google-auth>=2.14.1 -> cryptography>=38.0.3, with
217
+ # no extra marker, so it is already in the hard closure of the line above.
218
+ # Re-measured on the 2.29.0 pin (batch 172): google-auth>=2.56.0 ->
219
+ # cryptography>=38.0.3 (>=41.0.5 on Python 3.14), still no extra marker.
220
+ # What the declaration buys is that a dependency this project's SECURITY
221
+ # rests on is pinned by this project rather than by whatever google-auth
222
+ # happens to resolve to.
223
+ #
224
+ # Apache-2.0 OR BSD-3-Clause -- permissive, so THIRD_PARTY_NOTICES.md's
225
+ # no-copyleft claim still holds. It is the SECOND dependency shipping
226
+ # compiled native code (google-re2 was the first): the wheels carry a Rust
227
+ # extension and OpenSSL, and that file's paragraph names both now.
228
+ #
229
+ # A FLOOR AND A CEILING rather than an exact pin, like `rich`: an exact pin
230
+ # here would fight google-auth's own floor on a resolver that has to
231
+ # satisfy both. The ceiling is D22's rule, so a MAJOR cannot change
232
+ # behaviour underneath CI. The floor was the AEAD API this uses, stable
233
+ # since 3.x, until batch 208 raised it to 50.0.0 for the reason batch
234
+ # 201 floored pillow (a floor batch 212 removed with matplotlib): pip leaves an installed package alone while it still satisfies
235
+ # the range, and 42.0 through 49.x carry between 15 and 2 known
236
+ # advisories (OSV, measured 2026-10-31), so a venv built earlier kept
237
+ # one. 50.0.0 is the first release with none. What to re-measure when it
238
+ # moves: `AESGCM` still raises `cryptography.exceptions.InvalidTag` with NO
239
+ # message (the store relies on the three failure causes being
240
+ # indistinguishable), and the wheel list still covers this project's
241
+ # platforms.
242
+ cryptography>=50.0.0,<52
243
+
244
+ # ROADMAP_v3 §83 (CG6, CG16). Checks the ID token a ChatGPT-plan sign-in
245
+ # returns against OpenAI's published keys: RS256, issuer, audience, expiry
246
+ # and nonce. The owner chose declaring it over a hand-written RS256 check
247
+ # on `cryptography` -- signature verification is code to get exactly right,
248
+ # and this library is the one everybody else gets it right with.
249
+ #
250
+ # DECLARING IT ADDS NOTHING TO AN INSTALL, measured the way `cryptography`'s
251
+ # entry was: mcp==2.3.0 -> pyjwt[crypto]>=2.10.1, with no marker, so it is
252
+ # already in the hard closure of the mcp line above. What the declaration
253
+ # buys is that a check this project's sign-in rests on no longer depends on
254
+ # mcp happening to pull it in. MIT, so THIRD_PARTY_NOTICES.md's no-copyleft
255
+ # claim holds; pure Python.
256
+ #
257
+ # A FLOOR AND A CEILING, for `cryptography`'s reason: an exact pin here
258
+ # would fight mcp's own floor, `>=2.10.1`. The ceiling is D22's rule. The
259
+ # floor was mcp's until batch 208 raised it to 2.15.0, for `cryptography`'s
260
+ # reason above: the owner's venv held 2.13.0, which the old floor allowed and
261
+ # which carries 29 known advisories (OSV, measured 2026-10-31), two of
262
+ # them in exactly the calls core/chatgpt_auth.py makes -- a RecursionError
263
+ # out of `jwt.decode` on a deeply nested header (GHSA-8wjv-2p76-3863), and
264
+ # one malformed RSA key aborting a whole `PyJWKSet` (GHSA-w6j9-cwv2-h6wq).
265
+ # 2.15.0 is the first release with none. What to re-measure when it
266
+ # moves: `jwt.decode` still refuses an algorithm outside the list it is
267
+ # given (the `none` and HS256 tests in tests/test_chatgpt_auth.py), and
268
+ # `PyJWKSet` still reads OpenAI's JWKS.
269
+ PyJWT>=2.15.0,<3
270
+
271
+ # TECHNICAL_DEBT 18, batch 116. A PRODUCTION dependency: `arxiv.py` imports
272
+ # it at module load, and pyproject.toml reads this file for its dependency
273
+ # list via `dynamic = ["dependencies"]`.
274
+ #
275
+ # THE TIDY THIS COMMENT WARNED ABOUT HAS NOW HAPPENED, and the warning was
276
+ # right about the risk and wrong about the mechanism. It said a tidy of the
277
+ # test block "would drop it from the shipped set", as though the shipped set
278
+ # were something other than this file -- it was not, which is why the test
279
+ # block itself shipped. Batch 136 (S18-07) moved that block to
280
+ # requirements-test.txt, so the distinction is real now and defusedxml is
281
+ # unambiguously on the shipped side of it. It stayed because this paragraph
282
+ # said why; that is the comment earning its place rather than the claim it
283
+ # happened to get wrong.
284
+ #
285
+ # stdlib ElementTree expands declared entities on a feed this project does
286
+ # not write (Bandit B405/B314). Pinned exactly rather than floored, the way
287
+ # textual is: item 18 asks for "version-pinned re-verification", and a
288
+ # parser's refusals are exactly what a silent minor bump could change.
289
+ defusedxml==0.7.1
290
+
291
+ # Test dependencies live in requirements-test.txt, NOT here (S18-07,
292
+ # batch 136). pyproject.toml reads THIS file for `[project].dependencies`
293
+ # via `dynamic`, so anything below this line is a hard requirement of the
294
+ # published package and of every end user's venv. Install the suite's own
295
+ # packages with `pip install -r requirements-test.txt`.