@yahyahammad/venastine 0.2.0 → 0.3.8

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 (129) hide show
  1. package/.env.example +19 -14
  2. package/CONFIG_ARCHITECTURE.md +2002 -0
  3. package/NOTICE +1 -1
  4. package/README.md +233 -94
  5. package/agents/builtin/build.md +73 -0
  6. package/agents/builtin/general.md +80 -0
  7. package/agents/builtin/plan.md +23 -5
  8. package/agents/builtin/test.md +65 -0
  9. package/agents/builtin/writer.md +67 -0
  10. package/agents/manager.py +261 -195
  11. package/agents/subagent_tool.py +277 -43
  12. package/agents/tui_commands.py +7 -12
  13. package/bin/venastine.mjs +528 -382
  14. package/bin/venastine.py +71 -0
  15. package/cli_commands.py +835 -0
  16. package/config.py +95 -1198
  17. package/config.yaml +1090 -0
  18. package/config_edit.py +1301 -0
  19. package/config_schema.py +966 -0
  20. package/config_update.py +749 -0
  21. package/core/agent_activity.py +158 -20
  22. package/core/ansi.py +452 -0
  23. package/core/approval.py +27 -5
  24. package/core/call_view.py +364 -0
  25. package/core/cancel.py +121 -0
  26. package/core/chatgpt_auth.py +747 -0
  27. package/core/claims_view.py +107 -0
  28. package/core/client.py +1238 -82
  29. package/core/commands.py +454 -0
  30. package/core/compaction.py +830 -66
  31. package/core/config_loader.py +516 -97
  32. package/core/connect.py +686 -0
  33. package/core/events.py +68 -6
  34. package/core/file_mentions.py +280 -0
  35. package/core/interaction.py +87 -4
  36. package/core/line_pattern.py +128 -0
  37. package/core/loop.py +1516 -440
  38. package/core/memory.py +120 -8
  39. package/core/pipeline_models.py +23 -9
  40. package/core/provider_errors.py +234 -0
  41. package/core/reasoning/confidence_scoring.py +70 -54
  42. package/core/reasoning/json_retry.py +19 -1
  43. package/core/reasoning/orchestrator.py +232 -53
  44. package/core/reasoning/output_writer.py +90 -31
  45. package/core/reasoning/payload_validation.py +133 -7
  46. package/core/reasoning/pipeline_storage.py +427 -361
  47. package/core/reasoning/review.py +659 -615
  48. package/core/reasoning/scholar.py +38 -14
  49. package/core/reasoning/source_corpus.py +110 -4
  50. package/core/replay.py +158 -26
  51. package/core/session.py +36 -12
  52. package/core/session_view.py +259 -0
  53. package/core/session_wake.py +259 -0
  54. package/core/shell_sessions.py +1553 -0
  55. package/core/workspace_trust.py +1 -1
  56. package/credentials.py +308 -180
  57. package/database.py +82 -1
  58. package/{PRIVACY.md → docs/PRIVACY.md} +19 -7
  59. package/docs/SECURITY.md +108 -0
  60. package/docs/THIRD_PARTY_NOTICES.md +148 -0
  61. package/env_secrets.py +118 -77
  62. package/json_store.py +236 -16
  63. package/logging_setup.py +121 -11
  64. package/main.py +1671 -199
  65. package/mcp_client/client.py +257 -23
  66. package/mcp_client/config.py +484 -379
  67. package/mcp_client/registration.py +163 -2
  68. package/memories/manager.py +15 -0
  69. package/memories/tui_commands.py +7 -12
  70. package/package.json +234 -4
  71. package/project_init/config_files.py +4 -0
  72. package/project_init/generator.py +201 -37
  73. package/project_init/manifest.py +418 -406
  74. package/project_init/tui_commands.py +7 -4
  75. package/prompts/revalidate.md +1 -1
  76. package/prompts/revise.md +17 -15
  77. package/prompts/system_prompts.py +64 -10
  78. package/providers.json.example +94 -92
  79. package/pyproject.toml +51 -2
  80. package/requirements.txt +252 -29
  81. package/safety/policy_enforcement.py +353 -17
  82. package/security/askpass.py +419 -0
  83. package/security/capability.py +37 -7
  84. package/security/confine.py +484 -0
  85. package/security/hosted.py +182 -0
  86. package/security/permissions.py +54 -7
  87. package/security/posture.py +124 -12
  88. package/security/protected_paths.py +198 -6
  89. package/security/sandbox.py +3262 -139
  90. package/security/secrets.py +855 -0
  91. package/skills/tui_commands.py +3 -4
  92. package/storage.py +368 -35
  93. package/tools/base.py +465 -305
  94. package/tools/builtin/_math_common.py +306 -16
  95. package/tools/builtin/_net_common.py +267 -61
  96. package/tools/builtin/arxiv.py +681 -223
  97. package/tools/builtin/ask_user.py +231 -213
  98. package/tools/builtin/discrete_math.py +35 -4
  99. package/tools/builtin/fetch_url.py +239 -21
  100. package/tools/builtin/file_ops.py +341 -33
  101. package/tools/builtin/geometry.py +184 -148
  102. package/tools/builtin/get_time.py +2 -2
  103. package/tools/builtin/linear_algebra.py +66 -23
  104. package/tools/builtin/load_skill.py +10 -0
  105. package/tools/builtin/logic.py +171 -12
  106. package/tools/builtin/pin.py +46 -12
  107. package/tools/builtin/probability_stats.py +224 -228
  108. package/tools/builtin/project_docs.py +80 -21
  109. package/tools/builtin/remember.py +21 -4
  110. package/tools/builtin/shell.py +586 -67
  111. package/tools/builtin/shell_sessions.py +613 -0
  112. package/tools/builtin/symbolic_math.py +80 -10
  113. package/tools/builtin/todo.py +5 -3
  114. package/tools/builtin/web_search.py +329 -151
  115. package/tools/context.py +23 -0
  116. package/tools/isolation.py +387 -314
  117. package/tools/registry.py +347 -30
  118. package/tui/app.py +4645 -438
  119. package/tui/app.tcss +185 -2
  120. package/tui/commands.py +10 -154
  121. package/tui/diffs.py +286 -265
  122. package/tui/hosted_driver.py +102 -0
  123. package/tui/markdown.py +60 -10
  124. package/tui/meters.py +103 -7
  125. package/tui/screens.py +346 -92
  126. package/tui/themes.py +59 -19
  127. package/tui/widgets.py +1575 -111
  128. package/SECURITY.md +0 -88
  129. package/THIRD_PARTY_NOTICES.md +0 -100
@@ -25,14 +25,33 @@ and inverting it would create a cycle the moment tools/context.py grows
25
25
  an import of its own.
26
26
  """
27
27
 
28
- from typing import TYPE_CHECKING, Iterable, Optional
28
+ from collections.abc import Iterable
29
+ from typing import TYPE_CHECKING, Optional
29
30
 
30
31
  import config
32
+ from security import hosted
31
33
 
32
34
  if TYPE_CHECKING: # pragma: no cover - typing only, never imported at runtime
33
35
  from tools.context import ToolContext
34
36
 
35
37
 
38
+ # The tools whose approval is decided by `shell_approval_mode` and nothing
39
+ # else, so they have a `tool_permissions` field and no `tool_approvals` one.
40
+ # `shell` since §28 (G3): a field there was the ratchet, and every answer it
41
+ # could give is one the mode already names. The two background-session start
42
+ # tools since ROADMAP_v3 §49 (SS16), because they run a command through the
43
+ # same gate -- a second switch over it would be G3's ratchet again.
44
+ #
45
+ # A set named here rather than `name == "shell"` inline, so the exemption
46
+ # has one spelling that tests can hold against the registry: an import-time
47
+ # check in tools/registry.py requires every name here to be registered with
48
+ # shell's own approval check, so the exemption cannot outlive the gate it is
49
+ # an exemption FOR.
50
+ APPROVAL_BY_SHELL_MODE = frozenset({"shell", "shell_background",
51
+ "shell_monitor",
52
+ "shell_interactive"})
53
+
54
+
36
55
  def _default_for_unknown_tool(tool_name: str) -> bool:
37
56
  """Explicit, named default for tools with no declared field, rather
38
57
  than letting getattr's fallback decide by omission.
@@ -115,6 +134,14 @@ def is_tool_allowed(
115
134
  permissions = config.ToolPermissions()
116
135
  if not getattr(permissions, tool_name, _default_for_unknown_tool(tool_name)):
117
136
  return False
137
+ # ROADMAP_v3 §79 (HD4). A shared hosted deployment refuses the tools
138
+ # that reach the container itself, whatever config.yaml says: every
139
+ # visitor shares that container and one Unix user. Second to the global
140
+ # check and before any context, for the same reason that check is
141
+ # first -- no agent or skill may re-enable what this layer refuses.
142
+ # HD11 adds the math tools, unless the bridge verified their confinement.
143
+ if tool_name in hosted.refused_tools():
144
+ return False
118
145
  if context is not None and context.allowed_tools is not None:
119
146
  return tool_name in context.allowed_tools
120
147
  return True
@@ -153,12 +180,18 @@ def requires_approval(
153
180
  if context is not None
154
181
  else False
155
182
  )
156
- return tool_level or context_level
183
+ # bool(), because the signature says so and an override is only typed
184
+ # where a definition is PARSED (§67 A28): a ToolContext built in code
185
+ # with {"web_search": "false"} returned the string 'false' (batch 220,
186
+ # #250). Truthiness is unchanged -- that string still TIGHTENS, which is
187
+ # the ratchet above working, not a leak -- only the type is.
188
+ return bool(tool_level or context_level)
157
189
 
158
190
 
159
191
  def assert_permissions_declared(tool_names: Iterable[str]) -> None:
160
- """D24: every statically registered tool must have a field in BOTH
161
- config dataclasses. Raises rather than warning, because the failure
192
+ """D24: every statically registered tool must have a declared permission
193
+ field, and -- with one deliberate exception -- a declared approval
194
+ field. Raises rather than warning, because the failure
162
195
  this guards against is invisible at runtime.
163
196
 
164
197
  The precedent is `fetch_url`: registered, documented as working, and
@@ -170,6 +203,13 @@ def assert_permissions_declared(tool_names: Iterable[str]) -> None:
170
203
  _default_for_unknown_tool) is exactly what makes it important that
171
204
  omission is never accidental for built-in ones.
172
205
 
206
+ The exception is APPROVAL_BY_SHELL_MODE in approvals: those tools'
207
+ approval is governed solely by `shell_approval_mode`, and a second
208
+ switch there could only force `always` -- exactly what the mode already
209
+ says. A test pins the set as the SOLE exemption, and tools/registry.py
210
+ requires each name in it to carry shell's own approval check, so a
211
+ future omission cannot ride in on it.
212
+
173
213
  Dynamically-named `mcp__*` tools are exempt by design; they can never
174
214
  have a declared field. Call this after static registration only.
175
215
  """
@@ -179,12 +219,19 @@ def assert_permissions_declared(tool_names: Iterable[str]) -> None:
179
219
  name
180
220
  for name in tool_names
181
221
  if not name.startswith("mcp__")
182
- and not (hasattr(permissions, name) and hasattr(approvals, name))
222
+ and not (hasattr(permissions, name)
223
+ and (name in APPROVAL_BY_SHELL_MODE
224
+ or hasattr(approvals, name)))
183
225
  ]
184
226
  if missing:
185
227
  raise RuntimeError(
186
228
  "Tools are registered with no declared permission/approval "
187
229
  f"field, so they would be silently denied forever: {sorted(missing)}. "
188
- "Add a bool field for each to BOTH config.ToolPermissions and "
189
- "config.ToolApprovals (ROADMAP_v2 §15, D24)."
230
+ "Add a bool field for each to ToolPermissionsModel and -- "
231
+ f"except {sorted(APPROVAL_BY_SHELL_MODE)}, whose approval lives "
232
+ "solely in shell_approval_mode -- to ToolApprovalsModel in "
233
+ "config_schema.py, and a key for each to tool_permissions "
234
+ "(and, with the same exception, tool_approvals) in config.yaml "
235
+ "(ROADMAP_v2 §15, D24). The field names live in Python so this "
236
+ "check stays a build-time one; the values live in the YAML."
190
237
  )
@@ -28,9 +28,11 @@ WHAT THE GUARANTEE IS, AND IS NOT
28
28
  Stated precisely, because a half-understood control is worse than none
29
29
  (UN4). The posture CANNOT be changed by:
30
30
 
31
- * any tool call -- nothing writes `config.py` without an approval
32
- prompt, since it is outside the workspace and `_file_approval_check`
33
- hard-returns True there;
31
+ * any tool call -- nothing writes `config.yaml`, where the values
32
+ live, or `config.py`, which reads them, without an approval prompt:
33
+ both are outside the workspace and `_file_approval_check`
34
+ hard-returns True there. `write` and `edit` are globally denied on
35
+ top of that;
34
36
  * `settings.json` at either tier -- the keys are rejected BY NAME, the
35
37
  way G7 rejects `shell_approval_mode`, because a project's settings
36
38
  beat the user's and arrive with a directory you cloned;
@@ -50,10 +52,29 @@ ORDERING IS SELF-ENFORCING (UN2)
50
52
  --------------------------------
51
53
  `current()` marks the posture bound. `apply_cli()` raises if folding in
52
54
  the command line would now CHANGE it. So a wrong call order in `main.py`
53
- is a LOUD startup failure rather than a silently stale posture -- which
54
- matters because the failure mode of getting it wrong is a run that
55
+ would be a LOUD startup failure rather than a silently stale posture --
56
+ which matters because the failure mode of getting it wrong is a run that
55
57
  believes it is safer than it is, and that is invisible.
56
58
 
59
+ AND THAT GUARD IS DORMANT ON EVERY SHIPPED LAUNCH (batch 142).
60
+ `main.py` calls `apply_cli()` with NO ARGUMENTS -- it is the only
61
+ production call -- so `supplied` is always empty, `candidate` always
62
+ equals `_posture`, and `candidate != _posture` is never true. Measured:
63
+ `current()` then `apply_cli()` does not raise, while the control
64
+ `apply_cli(shell_approval_mode="never")` does. There is no live harm and
65
+ that is the same fact as the finding -- with nothing folded in, a
66
+ premature reader reads the right value anyway -- but a reader creeping
67
+ above that line today produces nothing at all.
68
+
69
+ WHAT ENFORCES UN2 TODAY is
70
+ `test_posture.py::test_no_module_reads_the_posture_at_import`, a source
71
+ scan, which is what AGENTS.md names as the pin. The scan covers IMPORT
72
+ time only, not a reader placed inside `main()` above the `apply_cli()`
73
+ call. The guard below becomes live the day the first posture flag
74
+ exists (`--unsafe` is written about in this module and in
75
+ `test_posture.py` as though it did), and `TestApplyCliOrdering` keeps it
76
+ working for that day -- which is why it is kept rather than removed.
77
+
57
78
  The predicate is "a change after a read", not "a read", and the
58
79
  difference is not a softening. The guard exists to catch *something acted
59
80
  on the config-only value*; if the command line produces the posture that
@@ -68,7 +89,10 @@ Never calling `apply_cli()` at all -- a library import, a bare test --
68
89
  leaves the config-derived default, which is the safe posture.
69
90
 
70
91
  This module imports `config` and the stdlib and nothing else, per
71
- ARCHITECTURE's rule for `security/`.
92
+ ARCHITECTURE's rule for `security/`. Note that `config` is now a shim over
93
+ `config.yaml` and so pulls in `ruamel.yaml` and `pydantic` transitively --
94
+ the rule this sentence states is about what `security/` may NAME, and that
95
+ is unchanged: `config` and the stdlib.
72
96
  """
73
97
 
74
98
  from __future__ import annotations
@@ -76,8 +100,9 @@ from __future__ import annotations
76
100
  import contextlib
77
101
  import dataclasses
78
102
  import os
103
+ from collections.abc import Iterator
79
104
  from dataclasses import dataclass
80
- from typing import Any, Iterator
105
+ from typing import Any
81
106
 
82
107
  import config
83
108
 
@@ -95,7 +120,7 @@ class Posture:
95
120
  legal statement anywhere in the process.
96
121
  """
97
122
 
98
- # "always" / "tiered" / "never" -- validated at bind time, so a typo
123
+ # "always" / "tiered" / "contained" / "never" -- validated at bind time, so a typo
99
124
  # is a startup failure and never a policy that quietly means
100
125
  # something else (the argument capability.validate_mode makes).
101
126
  shell_approval_mode: str
@@ -104,6 +129,19 @@ class Posture:
104
129
  # records that it is not defended against because it cannot be.
105
130
  allow_insecure_fallback: bool
106
131
  auto_approve_fallback: bool
132
+ # ROADMAP_v3 §49 slice 3 (SS35). Whether a command may run in WSL at
133
+ # all. Frozen here rather than read live for the reason the whole
134
+ # module exists, and the reason it belongs here rather than beside
135
+ # `wsl_distro` in plain config: a distro NAME decides where a command
136
+ # lands, and this decides whether it may leave the container.
137
+ allow_wsl_backend: bool
138
+ # ROADMAP_v3 §49 slice 4 (SS46). Whether a command may run on the
139
+ # configured remote host at all. Here rather than beside `ssh_host` in
140
+ # plain config for `allow_wsl_backend`'s reason, with one more: the
141
+ # other SSH keys say WHERE a command would land, and this one says
142
+ # whether this machine may run agent commands on a machine it cannot
143
+ # inspect.
144
+ allow_ssh_backend: bool
107
145
  # Master switch for pattern-based redaction of what leaves a tool.
108
146
  redact_tool_outputs: bool
109
147
  # VENASTINE_REDACT_OFF, the per-run weakening of the line above.
@@ -152,7 +190,24 @@ class Posture:
152
190
  "SHELL_APPROVAL_MODE is 'never' -- no shell command is "
153
191
  "ever asked about, including one that reads outside the "
154
192
  "workspace on the host"))
155
- if self.allow_insecure_fallback and self.auto_approve_fallback:
193
+ elif self.shell_approval_mode == "contained":
194
+ # §48 (CE4). Not a network or host-read weakening -- both still
195
+ # ask -- but anything the container confines runs unasked,
196
+ # including code the agent wrote itself, and in a workspace that
197
+ # holds a `.venastine/` that code can read it.
198
+ reasons.append((
199
+ "shell: contained",
200
+ "SHELL_APPROVAL_MODE is 'contained' -- a command that runs "
201
+ "in the container without network runs without asking, "
202
+ "including code the agent wrote itself; host reads and "
203
+ "network commands still ask"))
204
+ # §48 (CE6). Under "always" the pair is NOT "no ask": the mode check
205
+ # returns before the fallback opt-in is ever read, so every command
206
+ # asks, the host fallback included. Measured, and the badge said
207
+ # otherwise -- so under "always" the pair reports as the fallback
208
+ # alone, which is what it actually is.
209
+ if (self.allow_insecure_fallback and self.auto_approve_fallback
210
+ and self.shell_approval_mode != "always"):
156
211
  reasons.append((
157
212
  "host shell, no ask",
158
213
  "ALLOW_INSECURE_SANDBOX_FALLBACK and "
@@ -162,9 +217,58 @@ class Posture:
162
217
  elif self.allow_insecure_fallback:
163
218
  reasons.append((
164
219
  "host shell fallback",
165
- "ALLOW_INSECURE_SANDBOX_FALLBACK is on -- if Docker is "
166
- "down, commands run on the host with no filesystem or "
167
- "network isolation"))
220
+ "ALLOW_INSECURE_SANDBOX_FALLBACK is on -- if no container "
221
+ "runtime (Docker or Podman) can run the sandbox, commands "
222
+ "run on the host with no filesystem or network isolation"))
223
+ # ROADMAP_v3 §49 slice 3 (SS41). Reported when it is ENABLED rather
224
+ # than when it is used: the badge describes what this process may
225
+ # do, and a user reading it after a WSL command has already run has
226
+ # read it too late. It is its own pair rather than folded into the
227
+ # fallback's, because the two are different weakenings -- the
228
+ # fallback is the host when there is no container, and this is a
229
+ # Linux userland the agent can ask for while one is running.
230
+ if self.allow_wsl_backend:
231
+ # ROADMAP_v3 §49 slice 5b (SS65). The last clause used to read
232
+ # "Only a read-only command inside the workspace runs there
233
+ # without asking" unconditionally, and that is FALSE when the
234
+ # fallback opt-in is on -- measured against the real gate, where
235
+ # `rm -rf /home` on WSL answers asks=False, because WSL is
236
+ # UNCONTAINED and the opt-in sits above the capability rule.
237
+ # A badge that overstates what asks is worse than no badge, and
238
+ # this is §48 (CE6)'s correction applied to a second pair.
239
+ if (self.auto_approve_fallback
240
+ and self.shell_approval_mode not in ("always", "never")):
241
+ asking = ("AUTO_APPROVE_SANDBOX_FALLBACK is on as well, so "
242
+ "NOTHING there asks first -- that opt-in answers "
243
+ "for the whole backend, not only for the host "
244
+ "fallback it is named after. Running as root is "
245
+ "the one exception and always asks")
246
+ else:
247
+ asking = ("Only a read-only command inside the workspace "
248
+ "runs there without asking, and a command that "
249
+ "runs as root always asks")
250
+ reasons.append((
251
+ "wsl backend",
252
+ "ALLOW_WSL_BACKEND is on -- the agent can ask for a command "
253
+ "to run in WSL, which is not a sandbox: it reaches this "
254
+ "harness's own config and authority files, has no memory or "
255
+ "process limits, and can start Windows programs with your "
256
+ "full authority. " + asking))
257
+ # SS55. A THIRD pair, not a widening of the WSL one, because the
258
+ # three weakenings are different: the fallback is the host when
259
+ # there is no container, WSL is a Linux userland beside a running
260
+ # one, and this is a machine that is not this machine at all. A
261
+ # user who reads "uncontained" has to be able to tell which.
262
+ if self.allow_ssh_backend:
263
+ reasons.append((
264
+ "ssh backend",
265
+ "ALLOW_SSH_BACKEND is on -- the agent can ask for a command "
266
+ "to run on the configured remote host, which this harness "
267
+ "cannot inspect: no isolation, no memory or process limits, "
268
+ "and credentials it did not issue. Nothing runs there "
269
+ "without asking you first -- running as root included "
270
+ "(slice 5b) -- and `write` still saves locally while the "
271
+ "shell runs remotely"))
168
272
  if not self.redact_tool_outputs:
169
273
  reasons.append((
170
274
  "no redaction",
@@ -197,6 +301,8 @@ def _from_config() -> Posture:
197
301
  shell_approval_mode=config.SHELL_APPROVAL_MODE,
198
302
  allow_insecure_fallback=bool(config.ALLOW_INSECURE_SANDBOX_FALLBACK),
199
303
  auto_approve_fallback=bool(config.AUTO_APPROVE_SANDBOX_FALLBACK),
304
+ allow_wsl_backend=bool(config.ALLOW_WSL_BACKEND),
305
+ allow_ssh_backend=bool(config.ALLOW_SSH_BACKEND),
200
306
  redact_tool_outputs=bool(config.REDACT_TOOL_OUTPUTS),
201
307
  redact_off_env=_redact_off_from_env(),
202
308
  )
@@ -261,6 +367,12 @@ def apply_cli(**overrides: Any) -> Posture:
261
367
  # the read alone would also make `main()` un-callable twice in one
262
368
  # interpreter, which the suite does constantly -- and a guard that
263
369
  # forces the tests to work around it stops being read as a guard.
370
+ #
371
+ # DORMANT UNTIL THE FIRST POSTURE FLAG EXISTS (batch 142). No shipped
372
+ # launch passes an override, so `supplied` is {} and this cannot
373
+ # fire. The predicate is NOT the thing to revisit -- raising on the
374
+ # read alone was tried and was wrong twice, and the paragraph above
375
+ # is why. See the module docstring for what enforces UN2 meanwhile.
264
376
  if _read and candidate != _posture:
265
377
  raise RuntimeError(
266
378
  "security.posture.apply_cli() would CHANGE the posture after it "
@@ -66,6 +66,7 @@ from __future__ import annotations
66
66
  import logging
67
67
  import os
68
68
  import re
69
+ import stat
69
70
 
70
71
  import config
71
72
  import credentials
@@ -196,7 +197,29 @@ def check_workspace(workspace_dir: str) -> str | None:
196
197
  The message is written for the person who has to fix it, and names
197
198
  both what overlaps and what to do instead: a workspace refusal is
198
199
  always a configuration error, never something the model did.
200
+
201
+ AN EMPTY NAME IS REFUSED BEFORE IT IS RESOLVED (TECHNICAL_DEBT 93,
202
+ batch 160), because `os.path.realpath("")` is the LAUNCH DIRECTORY.
203
+ `config_schema._overlay_env` honours a variable that is set at all --
204
+ `if override is not None`, deliberately generic over its table -- so
205
+ `AGENT_WORKSPACE=` made the launch directory the auto-approved
206
+ workspace, and `WORKSPACE_DIR_EXPLICIT`, which reads the variable's
207
+ presence and never its value, made it the project as well. Measured
208
+ in batch 146: `_file_approval_check('write', ...)` answered no-prompt
209
+ for a file in the launch directory. Here rather than in the merge, so
210
+ the merge stays generic and the one place a bad workspace is refused
211
+ stays the one place; `main()` calls this before anything reads the
212
+ workspace or the project.
199
213
  """
214
+ if not workspace_dir or not workspace_dir.strip():
215
+ return (
216
+ "The workspace is an empty path -- AGENT_WORKSPACE is set to an "
217
+ "empty value, or workspace_dir in config.yaml is blank -- and an "
218
+ "empty path resolves to the directory this was launched from, "
219
+ "which would make it the auto-approved workspace and the "
220
+ "project. Unset AGENT_WORKSPACE, or name a directory; the "
221
+ "default is ./workspace."
222
+ )
200
223
  workspace_real = os.path.realpath(workspace_dir)
201
224
 
202
225
  for protected, label in critical_paths():
@@ -219,6 +242,68 @@ def check_workspace(workspace_dir: str) -> str | None:
219
242
  return None
220
243
 
221
244
 
245
+ def check_project(project_path: str) -> str | None:
246
+ """Why *project_path* may not be the PROJECT, or None (debt item 26).
247
+
248
+ The project is what `config_loader.get_project_path()` resolves, and
249
+ everything downstream follows it: D17 workspace trust, the
250
+ `.venastine/` config tier, `/init`'s destination and `UserMemory`'s
251
+ project scope (WS7, D25/M12). It may not be the harness's own install
252
+ tree.
253
+
254
+ THE DISPARITY THIS CLOSES, and it is why this is a guard rather than a
255
+ convention. Naming that directory was ALREADY refused --
256
+ `check_workspace` above returns a reason for `AGENT_WORKSPACE=<harness
257
+ root>`, and for every non-exempt subfolder of it. But `main()` falls
258
+ back to `os.getcwd()` when nothing names a workspace as the project, so
259
+ launching from a clone reached the same state by saying nothing:
260
+ measured in batch 96 from an owner report and reproduced in batch 107,
261
+ the trust prompt listed the harness's own `AGENTS.md`. The rule was
262
+ enforced when you said it and skipped when you did not.
263
+
264
+ THE MESSAGE NAMES THREE WAYS OUT since batch 111, and used to name two
265
+ and a trap. It said `workspace_dir` in `config.yaml` "is not enough on
266
+ its own" -- true, and no help at all to the person reading it, who had
267
+ just written the directory they meant into the file this refusal is
268
+ about. `workspace_is_project` is the third way (TECHNICAL_DEBT 28), and
269
+ the sentence is now about what to do rather than about what does not
270
+ work.
271
+
272
+ EQUALITY, NOT CONTAINMENT, which is the one place this deliberately
273
+ differs from `check_workspace`. That one is a WRITE boundary and
274
+ refuses anything inside the tree except `workspace/` and `output/`;
275
+ this one answers "is the project the harness", and `<harness>/workspace`
276
+ is a perfectly good project. It is the shipped layout, and it is the
277
+ answer for someone working on the harness itself.
278
+
279
+ A REASON STRING rather than an exception, for `check_workspace`'s
280
+ reason: the caller prints it and exits. REALPATH'D HERE rather than at
281
+ the caller, because `main()`'s implicit branch hands over a bare
282
+ `os.getcwd()` -- and because the realpath is what makes this a question
283
+ about a directory instead of about a spelling. `_relation` compares
284
+ strings, so `<root>/workspace/..` and a Windows path differing only in
285
+ case both read as `disjoint` to it; `os.path.realpath` resolves the
286
+ first on every platform and canonicalises the second on Windows
287
+ (measured, batch 107).
288
+ """
289
+ root = harness_root()
290
+ if _relation(os.path.realpath(project_path), root) != "equal":
291
+ return None
292
+ return (
293
+ f"The project would be the harness's own install tree ({root}), "
294
+ f"because nothing named a workspace as the project and this was "
295
+ f"launched from there. Workspace trust, .venastine/, /init's "
296
+ f"destination and project-scoped memories all follow the project, "
297
+ f"so the harness would be asked to trust its own AGENTS.md and "
298
+ f"/init would scaffold documentation into its own source. Three "
299
+ f"ways out, any one of them enough: launch from the directory you "
300
+ f"want to work in; set AGENT_WORKSPACE to it; or, if you have "
301
+ f"already written it as workspace_dir in config.yaml, set "
302
+ f"workspace_is_project: true beside it. To work on the harness "
303
+ f"itself, AGENT_WORKSPACE=./workspace."
304
+ )
305
+
306
+
222
307
  def readonly_mounts(workspace_real: str) -> list[str]:
223
308
  """Docker `-v` arguments binding every protected path NESTED inside
224
309
  *workspace_real* read-only, in the order docker expects them.
@@ -280,13 +365,24 @@ def readonly_mounts(workspace_real: str) -> list[str]:
280
365
  # project_docs.py unions this set into its own wider denylist rather than
281
366
  # re-spelling the segment.
282
367
  #
368
+ # AND ONE COMPARISON, since batch 146 -- `folded_segments` below, because
369
+ # the fact is not the spelling on its own but the spelling together with
370
+ # how a filesystem will match it. Two of the three consumers had their own
371
+ # comparison and only one of them folded case, which is not a thing either
372
+ # file said; it was a thing you could only find by writing `.VENASTINE/`.
373
+ #
283
374
  # KNOWN LIMITS, in the house style. The file tools DENY; the shell only
284
375
  # ever ASKS, because a token check that learned shell syntax would be a
285
376
  # parser whose bugs auto-approve (G2), and an approval prompt leaves the
286
377
  # whole command text in front of a human instead. The auto-approved INERT
287
378
  # tier executes argv with no shell (§46 EP5), so its tokens ARE its
288
- # arguments and a token check is sound there; every tier that could
289
- # expand a glob or a variable into the segment is asked about anyway.
379
+ # arguments and a token check is sound there; under `tiered` every tier
380
+ # that could expand a glob or a variable into the segment is asked about
381
+ # anyway. That sentence was FALSE until §48 (CE1): a contained command was
382
+ # auto-approved, and `cat .ven*/settings.json; true` read the file unasked.
383
+ # Under `contained` and `never` it is still not true -- such a spelling,
384
+ # or code that opens the file, runs unasked; the ro mount stops a write,
385
+ # and the read is recorded in SECURITY.md as documented risk.
290
386
  # MCP servers exposing their own file tools are a separate surface this
291
387
  # list does not reach, and the user-tier `~/.config/venastine` stays
292
388
  # where it always was -- outside a default workspace it is prompt-gated
@@ -294,16 +390,112 @@ def readonly_mounts(workspace_real: str) -> list[str]:
294
390
  PROTECTED_SEGMENTS = frozenset({".venastine"})
295
391
 
296
392
 
393
+ def hardlinked_segment(resolved_path: str, roots) -> str | None:
394
+ """The protected segment a HARD LINK at *resolved_path* is a second
395
+ name for, or None (TECHNICAL_DEBT 90, batch 160).
396
+
397
+ `protected_segment` reads a resolved PATH, and a hardlink is not a link
398
+ to anything: `realpath` returns the name it was given and `islink` is
399
+ False, so a second name for `.venastine/settings.json` sitting in the
400
+ workspace read as an ordinary file. A `read` through one served the
401
+ protected file -- measured in batch 146, along with the write half that
402
+ `write_text_atomic` already closed by replacing the name.
403
+
404
+ SAME FILE, NOT SAME NAME: a file with more than one name is compared by
405
+ `(st_dev, st_ino)` against every file under the `.venastine/` of each of
406
+ *roots* -- the project and the workspace, which are the directories that
407
+ tree is read from. Both fields are real on Windows as well as POSIX
408
+ (measured: an NTFS hardlink pair reports `st_nlink == 2` and one file
409
+ id). The two rejected alternatives are why it is this one: enumerating a
410
+ file's other names is per-platform, and refusing every `st_nlink > 1`
411
+ denies ordinary files to protect against a layout the write path
412
+ already handled. A file with one name pays one `stat`.
413
+
414
+ A `.venastine/` elsewhere on the disk is not looked for, so a hardlink
415
+ into some other project's tree is not caught; the deny exists for the
416
+ one this harness reads.
417
+ """
418
+ try:
419
+ target = os.stat(resolved_path)
420
+ except OSError:
421
+ return None
422
+ if not stat.S_ISREG(target.st_mode) or target.st_nlink < 2:
423
+ return None
424
+ identity = (target.st_dev, target.st_ino)
425
+ for root in roots:
426
+ try:
427
+ entries = os.listdir(root)
428
+ except OSError:
429
+ continue
430
+ for entry in entries:
431
+ if entry.casefold() not in PROTECTED_SEGMENTS_FOLDED:
432
+ continue
433
+ for dirpath, _dirnames, filenames in os.walk(
434
+ os.path.join(root, entry)):
435
+ for name in filenames:
436
+ try:
437
+ other = os.stat(os.path.join(dirpath, name))
438
+ except OSError:
439
+ continue
440
+ if (other.st_dev, other.st_ino) == identity:
441
+ return entry
442
+ return None
443
+
444
+
445
+ def folded_segments(names) -> frozenset:
446
+ """`names` prepared for the comparison a filesystem will actually make.
447
+
448
+ THE ONE SPELLING OF THE FOLD (batch 146). Both consumers that compare a
449
+ path's segments against a denylist call this -- `protected_segment`
450
+ below and `project_docs._is_readable_doc` -- because the alternative is
451
+ what was there before: two comparisons, one of them case-sensitive, and
452
+ nothing in either file able to say they disagreed.
453
+
454
+ `casefold` rather than `lower`, which is the comparison Python offers
455
+ for "are these the same name to a human"; for ASCII segment names they
456
+ are identical and the difference costs nothing.
457
+ """
458
+ return frozenset(name.casefold() for name in names)
459
+
460
+
461
+ #: PUBLIC, because `shell.py`'s raw-token branch compares against it too
462
+ #: and a second `folded_segments(PROTECTED_SEGMENTS)` at that module's top
463
+ #: would be one derived value computed twice.
464
+ PROTECTED_SEGMENTS_FOLDED = folded_segments(PROTECTED_SEGMENTS)
465
+
466
+
297
467
  def protected_segment(resolved_path: str) -> str | None:
298
468
  """The protected segment *resolved_path* carries, or None.
299
469
 
300
470
  Callers pass a realpath'd path (file_ops' ``resolve_path``, the
301
471
  sandbox's ``_within``), so `..` travel and symlinks are already
302
- collapsed: a link pointing at `.venastine/settings.json` arrives
303
- here as the real thing, and on Windows so does a case-spelled one.
304
- Split on both separators because a raw shell token may carry either.
472
+ collapsed: a link pointing at `.venastine/settings.json` arrives here
473
+ as the real thing. Split on both separators because a raw shell token
474
+ may carry either.
475
+
476
+ CASE-FOLDED ON EVERY PLATFORM, and the word "realpath" above is exactly
477
+ why it has to be (batch 146). This used to say that on Windows a
478
+ case-spelled path "arrives here as the real thing" too, and that is
479
+ true only where the path ALREADY EXISTS for realpath to canonicalise --
480
+ which is not the state a `write` is in. With no `.venastine/` yet,
481
+ `.VENASTINE/mcp.json` resolved to itself, carried no known segment, was
482
+ auto-approved, and created the protected directory under the spelling
483
+ the guard did not know. After that realpath canonicalised every later
484
+ path to the ON-DISK name, so BOTH spellings were admitted in that
485
+ project from then on, and the file `config_loader` opens at startup was
486
+ one of them. Measured, including the control: a project whose directory
487
+ was created lowercase refuses both.
488
+
489
+ Folding unconditionally rather than per platform is the fail-closed
490
+ direction and costs what it costs: on a case-SENSITIVE filesystem a
491
+ directory genuinely named `.VENASTINE` is now denied to the file tools
492
+ with no prompt and no override, because for them this list is the hard
493
+ deny rather than a reason to ask. The shell consumer only ever ASKS, so
494
+ there the same fold costs one prompt.
305
495
  """
306
496
  for part in re.split(r"[\\/]", resolved_path):
307
- if part in PROTECTED_SEGMENTS:
497
+ if part.casefold() in PROTECTED_SEGMENTS_FOLDED:
498
+ # The spelling from the PATH, not from the list, so the refusal
499
+ # names what the caller actually wrote.
308
500
  return part
309
501
  return None