@yahyahammad/venastine 0.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.env.example +14 -0
- package/LICENSE +201 -0
- package/NOTICE +14 -0
- package/PRIVACY.md +66 -0
- package/README.md +878 -0
- package/SECURITY.md +88 -0
- package/THIRD_PARTY_NOTICES.md +100 -0
- package/agents/builtin/compactor.md +75 -0
- package/agents/builtin/explore.md +77 -0
- package/agents/builtin/grill-me.md +27 -0
- package/agents/builtin/initializer.md +94 -0
- package/agents/builtin/pipeline-reviewer.md +100 -0
- package/agents/builtin/plan.md +88 -0
- package/agents/builtin/review.md +74 -0
- package/agents/manager.py +195 -0
- package/agents/subagent_tool.py +215 -0
- package/agents/tui_commands.py +136 -0
- package/bin/venastine.mjs +382 -0
- package/config.py +1198 -0
- package/core/__init__.py +0 -0
- package/core/approval.py +115 -0
- package/core/client.py +1548 -0
- package/core/compaction.py +959 -0
- package/core/config_loader.py +1563 -0
- package/core/events.py +61 -0
- package/core/interaction.py +328 -0
- package/core/loop.py +1353 -0
- package/core/memory.py +370 -0
- package/core/model_windows.py +175 -0
- package/core/pipeline_models.py +176 -0
- package/core/reasoning/__init__.py +0 -0
- package/core/reasoning/authorization.py +130 -0
- package/core/reasoning/base.py +219 -0
- package/core/reasoning/confidence_scoring.py +468 -0
- package/core/reasoning/events.py +149 -0
- package/core/reasoning/json_retry.py +219 -0
- package/core/reasoning/orchestrator.py +1382 -0
- package/core/reasoning/output_writer.py +286 -0
- package/core/reasoning/payload_validation.py +398 -0
- package/core/reasoning/pipeline_storage.py +361 -0
- package/core/reasoning/review.py +615 -0
- package/core/reasoning/scholar.py +434 -0
- package/core/reasoning/source_corpus.py +327 -0
- package/core/reasoning/source_scoring.py +855 -0
- package/core/replay.py +167 -0
- package/core/session.py +153 -0
- package/core/workspace_trust.py +207 -0
- package/credentials.py +180 -0
- package/database.py +252 -0
- package/env_secrets.py +77 -0
- package/json_store.py +123 -0
- package/logging_setup.py +185 -0
- package/main.py +1959 -0
- package/mcp_client/__init__.py +15 -0
- package/mcp_client/client.py +499 -0
- package/mcp_client/config.py +379 -0
- package/mcp_client/registration.py +182 -0
- package/memories/manager.py +132 -0
- package/memories/tui_commands.py +232 -0
- package/package.json +73 -0
- package/project_init/config_files.py +278 -0
- package/project_init/doc_sets.py +330 -0
- package/project_init/generator.py +564 -0
- package/project_init/manifest.py +406 -0
- package/project_init/tui_commands.py +144 -0
- package/prompts/__init__.py +0 -0
- package/prompts/annotate.md +14 -0
- package/prompts/assumption_audit.md +27 -0
- package/prompts/claim_extraction.md +34 -0
- package/prompts/completeness.md +18 -0
- package/prompts/confidence_tiers.md +23 -0
- package/prompts/critic_pass.md +24 -0
- package/prompts/final_synthesis.md +13 -0
- package/prompts/initial_generation.md +14 -0
- package/prompts/preliminary_plan.md +23 -0
- package/prompts/revalidate.md +25 -0
- package/prompts/revise.md +15 -0
- package/prompts/source_grounding.md +66 -0
- package/prompts/system_prompts.py +262 -0
- package/prompts/universal_system_prompt +7 -0
- package/prompts/untrusted_content +1 -0
- package/providers.json.example +92 -0
- package/pyproject.toml +100 -0
- package/requirements.txt +72 -0
- package/safety/__init__.py +0 -0
- package/safety/policy_enforcement.py +682 -0
- package/security/__init__.py +0 -0
- package/security/capability.py +177 -0
- package/security/permissions.py +190 -0
- package/security/posture.py +302 -0
- package/security/protected_paths.py +309 -0
- package/security/sandbox.py +985 -0
- package/skills/builtin/crypto/cryptography-verification.md +66 -0
- package/skills/builtin/math/formal-specification.md +67 -0
- package/skills/builtin/math/numerical-methods.md +66 -0
- package/skills/builtin/math/proof-writing.md +61 -0
- package/skills/builtin/research/experiment-design.md +62 -0
- package/skills/builtin/research/literature-review.md +65 -0
- package/skills/builtin/research/reproducibility.md +55 -0
- package/skills/builtin/research/source-evaluation.md +70 -0
- package/skills/builtin/research/statistical-inference.md +63 -0
- package/skills/builtin/research/technical-writing.md +59 -0
- package/skills/builtin/security/cybersecurity-research.md +63 -0
- package/skills/builtin/software/code-review.md +61 -0
- package/skills/builtin/software/debugging.md +61 -0
- package/skills/builtin/software/test-design.md +67 -0
- package/skills/manager.py +138 -0
- package/skills/tui_commands.py +111 -0
- package/storage.py +1005 -0
- package/tools/base.py +305 -0
- package/tools/builtin/__init__.py +0 -0
- package/tools/builtin/_math_common.py +377 -0
- package/tools/builtin/_net_common.py +61 -0
- package/tools/builtin/arxiv.py +223 -0
- package/tools/builtin/ask_user.py +213 -0
- package/tools/builtin/discrete_math.py +134 -0
- package/tools/builtin/fetch_url.py +153 -0
- package/tools/builtin/file_ops.py +412 -0
- package/tools/builtin/geometry.py +148 -0
- package/tools/builtin/get_time.py +20 -0
- package/tools/builtin/linear_algebra.py +155 -0
- package/tools/builtin/load_skill.py +66 -0
- package/tools/builtin/logic.py +96 -0
- package/tools/builtin/pin.py +193 -0
- package/tools/builtin/probability_stats.py +228 -0
- package/tools/builtin/project_docs.py +405 -0
- package/tools/builtin/remember.py +158 -0
- package/tools/builtin/shell.py +461 -0
- package/tools/builtin/symbolic_math.py +119 -0
- package/tools/builtin/todo.py +147 -0
- package/tools/builtin/web_search.py +151 -0
- package/tools/context.py +136 -0
- package/tools/isolation.py +314 -0
- package/tools/registry.py +897 -0
- package/tui/app.py +3423 -0
- package/tui/app.tcss +458 -0
- package/tui/commands.py +61 -0
- package/tui/diffs.py +265 -0
- package/tui/preferences.py +182 -0
- package/tui/ravens.py +146 -0
- package/tui/screens.py +943 -0
- package/tui/themes.py +366 -0
- package/tui/widgets.py +1239 -0
package/core/replay.py
ADDED
|
@@ -0,0 +1,167 @@
|
|
|
1
|
+
"""
|
|
2
|
+
core/replay.py
|
|
3
|
+
|
|
4
|
+
Turns a stored thread back into something a shell can display
|
|
5
|
+
(ROADMAP_v2 §27, bug 1).
|
|
6
|
+
|
|
7
|
+
WHY THIS EXISTS AT ALL. Resuming a thread loaded its history correctly and
|
|
8
|
+
showed NOTHING: `ConversationMemory(thread_id=...)` reads the full history
|
|
9
|
+
at construction, but the TUI's picker callback swapped the object and wrote
|
|
10
|
+
one line, and `main.py --thread` printed "Resuming thread: <uuid>" and
|
|
11
|
+
dropped straight into the prompt. In the TUI it was worse than blank -- the
|
|
12
|
+
PREVIOUS thread's transcript stayed on screen under the new thread's id, so
|
|
13
|
+
the screen actively lied about which conversation was loaded.
|
|
14
|
+
|
|
15
|
+
WHY IT IS ONE FUNCTION AND NOT TWO RENDERERS. D12 makes the CLI a permanent
|
|
16
|
+
fallback, so both shells replay (§27 T5, §26's L6 again). What may be shown,
|
|
17
|
+
and in what form, is a policy decision that must not differ between them --
|
|
18
|
+
so the policy lives here and each shell only decides how to paint the
|
|
19
|
+
entries it is handed. §26 spent a decision (L2, "one translation site") on
|
|
20
|
+
exactly this shape.
|
|
21
|
+
|
|
22
|
+
THE ARCHIVE, NEVER THE DERIVED VIEW (T3). `storage.archive_history()`
|
|
23
|
+
returns every message ever written, compaction or not.
|
|
24
|
+
`ConversationMemory.messages` on a compacted thread BEGINS with the
|
|
25
|
+
synthesized summary that §21's M8 says must never be mistaken for something
|
|
26
|
+
a person said -- and replaying it would render harness-generated text under
|
|
27
|
+
a `you ›` label, which is precisely the mistake M8 exists to prevent. It
|
|
28
|
+
also means a replayed thread shows what was originally said rather than a
|
|
29
|
+
condensation of it, which is what a human resuming a conversation is looking
|
|
30
|
+
for.
|
|
31
|
+
|
|
32
|
+
THE LIVE VIEW SHOWS MORE THAN THIS ONE, DELIBERATELY (§41). Since batch 41
|
|
33
|
+
the TUI renders a `write` or an `edit` as an inline diff, and a REPLAYED one
|
|
34
|
+
here still shows `param_digest`. That is not a gap to close: a replay reads
|
|
35
|
+
the archive, which records what was called and never what the file held at
|
|
36
|
+
the time, so the digest is genuinely all this function has. The live view is
|
|
37
|
+
better informed because it was there.
|
|
38
|
+
|
|
39
|
+
REASONING IS REPLAYED, SINCE §44. §38 captured thinking for display and
|
|
40
|
+
persisted none of it, so reopening a thread showed the answers with the
|
|
41
|
+
reasoning silently missing -- the same conversation rendered two different
|
|
42
|
+
ways depending on whether the session had been closed. The blocks are
|
|
43
|
+
stored now (storage.MessageLog.thinking) and _reasoning_text pulls the
|
|
44
|
+
prose back out of them. What a shell does with the entry is still the
|
|
45
|
+
shell's: the TUI honours tui.show_thinking, because that is a display
|
|
46
|
+
setting rather than a fact about the thread.
|
|
47
|
+
|
|
48
|
+
TOOL RESULTS ARE SKIPPED, TOOL CALLS ARE ONE LINE (T4). A grounding-heavy
|
|
49
|
+
thread carries hundreds of kilobytes of fetched page text in its
|
|
50
|
+
`tool_result` rows; replaying that would bury the conversation in the
|
|
51
|
+
material the conversation was about. The call is what says what happened,
|
|
52
|
+
and `param_digest` says WHICH url or query -- redacted, because a tool
|
|
53
|
+
argument can carry a credential and this is now the third place one could
|
|
54
|
+
surface.
|
|
55
|
+
"""
|
|
56
|
+
|
|
57
|
+
from __future__ import annotations
|
|
58
|
+
|
|
59
|
+
from typing import List, Optional, Tuple
|
|
60
|
+
from uuid import UUID
|
|
61
|
+
|
|
62
|
+
from tools.registry import registry
|
|
63
|
+
from storage import archive_history
|
|
64
|
+
|
|
65
|
+
#: One replayed entry: (role, text). The roles are transcript palette roles
|
|
66
|
+
#: (tui/themes.role_styles), so the TUI can paint them with what it already
|
|
67
|
+
#: has and the CLI can label them. Deliberately NOT the neutral message
|
|
68
|
+
#: shape -- a caller of this function is rendering, not reasoning about
|
|
69
|
+
#: history, and handing it messages would invite a second policy decision
|
|
70
|
+
#: about tool results at the call site.
|
|
71
|
+
ReplayEntry = Tuple[str, str]
|
|
72
|
+
|
|
73
|
+
|
|
74
|
+
def replay_entries(thread_id: UUID) -> List[ReplayEntry]:
|
|
75
|
+
"""A stored thread as display entries, oldest first (§27 T3/T4).
|
|
76
|
+
|
|
77
|
+
Empty for a thread with no messages -- a `/new` thread that was never
|
|
78
|
+
used is a real state, and the shells write their own "nothing here yet"
|
|
79
|
+
line rather than this function inventing one.
|
|
80
|
+
"""
|
|
81
|
+
entries: List[ReplayEntry] = []
|
|
82
|
+
for message in archive_history(thread_id):
|
|
83
|
+
role = message.get("role")
|
|
84
|
+
if role == "user":
|
|
85
|
+
text = _as_text(message.get("content"))
|
|
86
|
+
if text:
|
|
87
|
+
entries.append(("user", text))
|
|
88
|
+
elif role == "assistant":
|
|
89
|
+
# BEFORE the answer, because that is the order it happened in
|
|
90
|
+
# and the order the live transcript drew it in (§43 RM1 puts
|
|
91
|
+
# one `venastine ›` above whichever came first). A replay that
|
|
92
|
+
# reordered the turn would be the same defect as one that
|
|
93
|
+
# reflowed it.
|
|
94
|
+
reasoning = _reasoning_text(message.get("thinking"))
|
|
95
|
+
if reasoning:
|
|
96
|
+
entries.append(("thinking", reasoning))
|
|
97
|
+
text = _as_text(message.get("text"))
|
|
98
|
+
if text:
|
|
99
|
+
entries.append(("assistant", text))
|
|
100
|
+
for call in message.get("tool_calls") or []:
|
|
101
|
+
entries.append(("tool", _tool_marker(call)))
|
|
102
|
+
# role == "tool": skipped by T4. Not a gap -- see the module
|
|
103
|
+
# docstring. The CALL above is the record that it happened.
|
|
104
|
+
return entries
|
|
105
|
+
|
|
106
|
+
|
|
107
|
+
def _reasoning_text(record) -> str:
|
|
108
|
+
"""The displayable prose out of a stored thinking record (§44).
|
|
109
|
+
|
|
110
|
+
The record holds the PROVIDER'S blocks, because that is what goes back
|
|
111
|
+
on the wire; this is the other half of the same data, for a human. Two
|
|
112
|
+
shapes reach here -- Anthropic's `thinking` blocks and the
|
|
113
|
+
v1-compatible `reasoning_content` one -- and both are read by field
|
|
114
|
+
name rather than by provider, so a third costs a name and not a
|
|
115
|
+
branch.
|
|
116
|
+
|
|
117
|
+
`redacted_thinking` contributes NOTHING and that is correct: it is
|
|
118
|
+
opaque ciphertext meaningful only to the model, so rendering it would
|
|
119
|
+
put a wall of base64 under a `venastine ›` label. It still travels on
|
|
120
|
+
the wire; this function is about what a person sees.
|
|
121
|
+
"""
|
|
122
|
+
if not isinstance(record, dict):
|
|
123
|
+
return ""
|
|
124
|
+
parts = []
|
|
125
|
+
for block in record.get("blocks") or []:
|
|
126
|
+
if not isinstance(block, dict):
|
|
127
|
+
continue
|
|
128
|
+
value = block.get("thinking") or block.get("text")
|
|
129
|
+
if isinstance(value, str) and value.strip():
|
|
130
|
+
parts.append(value.strip())
|
|
131
|
+
return "\n\n".join(parts)
|
|
132
|
+
|
|
133
|
+
|
|
134
|
+
def _tool_marker(call: dict) -> str:
|
|
135
|
+
"""One line for one tool call: name, then a redacted param digest.
|
|
136
|
+
|
|
137
|
+
The digest comes from `registry.call_digest` rather than from a
|
|
138
|
+
local copy -- it redacts BEFORE truncating, and a second
|
|
139
|
+
implementation is how that ordering regresses in one place and
|
|
140
|
+
not the other (§27's note on _param_digest, and the project's
|
|
141
|
+
canonical producer/consumer bug).
|
|
142
|
+
|
|
143
|
+
Through the REGISTRY since §42 (RA3), not `param_digest` direct,
|
|
144
|
+
because dropping a tool's declared rationale param is part of the
|
|
145
|
+
same rule and belongs at the same producer. Importing the registry
|
|
146
|
+
here costs nothing measurable: `main.py` is this module's only
|
|
147
|
+
caller and has already imported it, so it is in `sys.modules`
|
|
148
|
+
(1.9s cold at process start, 0.0s here).
|
|
149
|
+
"""
|
|
150
|
+
name = call.get("name") or "tool"
|
|
151
|
+
digest = registry.call_digest(name, call.get("input"))
|
|
152
|
+
# Standardised to ▸ like the live transcript (tui/app.py both modes).
|
|
153
|
+
# Replay stays redacted via param_digest, same producer as live.
|
|
154
|
+
return f"▸ {name} {digest}".rstrip() if digest else f"▸ {name}"
|
|
155
|
+
|
|
156
|
+
|
|
157
|
+
def _as_text(value: Optional[object]) -> str:
|
|
158
|
+
"""Message content as a displayable string.
|
|
159
|
+
|
|
160
|
+
`content` is whatever was persisted through json.dumps, so it is
|
|
161
|
+
normally a string but need not be; str() rather than a type assertion,
|
|
162
|
+
because a replay that raises on one odd row would take the whole resumed
|
|
163
|
+
conversation with it.
|
|
164
|
+
"""
|
|
165
|
+
if value is None:
|
|
166
|
+
return ""
|
|
167
|
+
return value.strip() if isinstance(value, str) else str(value)
|
package/core/session.py
ADDED
|
@@ -0,0 +1,153 @@
|
|
|
1
|
+
"""
|
|
2
|
+
core/session.py
|
|
3
|
+
|
|
4
|
+
Ephemeral, per-session override of ONE compaction number, and nothing
|
|
5
|
+
else. ROADMAP_v2 §21's revisit, batch 31; halved in batch 44.
|
|
6
|
+
|
|
7
|
+
WHAT THIS MODULE OWNS
|
|
8
|
+
set_trigger record an override against a (provider, model)
|
|
9
|
+
trigger_for read it back, but only for that same pair
|
|
10
|
+
trigger_any the binding-free read, for #89's pin cap
|
|
11
|
+
clear drop it, on a /model switch
|
|
12
|
+
|
|
13
|
+
WHAT IT DOES NOT OWN
|
|
14
|
+
Resolving the number -- core/compaction.py:thresholds() does that, and
|
|
15
|
+
treats what is here as the most local source rather than as the answer.
|
|
16
|
+
Persisting anything: NOTHING here reaches disk, which is the whole point
|
|
17
|
+
(see NEVER PERSISTED below).
|
|
18
|
+
|
|
19
|
+
ONE NUMBER NOW, AND THE OTHER ONE LEFT FOR A GOOD REASON. Batch 31 held
|
|
20
|
+
two, `window` and `trigger`, and was careful to say they were not the same
|
|
21
|
+
number. They still are not -- but they no longer have the same LIFETIME,
|
|
22
|
+
which is what this module is really about:
|
|
23
|
+
|
|
24
|
+
trigger the working-set ceiling. Ephemeral, because it is a knob you
|
|
25
|
+
reach for while a particular thread is behaving badly, and
|
|
26
|
+
restarting should give the derived number back. Still here.
|
|
27
|
+
window what context_limit() returns. It is not a preference at all --
|
|
28
|
+
it is a FACT about a deployment, and a fact does not want to be
|
|
29
|
+
re-entered every launch. Batch 44 moved it to
|
|
30
|
+
core/model_windows.py, a user-tier store keyed on every
|
|
31
|
+
(provider, model) the user has answered for. Keeping an
|
|
32
|
+
ephemeral tier ABOVE that store was the rejected alternative:
|
|
33
|
+
/model would clear a value the store immediately re-supplied,
|
|
34
|
+
so "override cleared" would be true and change nothing.
|
|
35
|
+
|
|
36
|
+
BOUND TO A (provider, model) PAIR, AND THAT IS THE INTERESTING RULE.
|
|
37
|
+
An override applies only when the model that was named is the model
|
|
38
|
+
actually running. A read from any other pair returns None WITHOUT
|
|
39
|
+
clearing, so:
|
|
40
|
+
|
|
41
|
+
* an agent that pins its own model (§18) uses that model's derived
|
|
42
|
+
value, rather than a number tuned for a different one;
|
|
43
|
+
* a critic-routed research pass (§11) does the same -- and, because a
|
|
44
|
+
mismatch does not clear, one routine critic pass cannot silently
|
|
45
|
+
destroy an override set for the chat;
|
|
46
|
+
* a research pass or subagent on the SESSION's model matches, so it
|
|
47
|
+
inherits with no plumbing at all. Propagation is a consequence of the
|
|
48
|
+
binding, not a separate mechanism.
|
|
49
|
+
|
|
50
|
+
Clearing is a separate act, and only /model does it. That is what makes
|
|
51
|
+
"switch away and come back" return the DEFAULT rather than the old
|
|
52
|
+
override: the pair-binding alone would restore it, so both rules are
|
|
53
|
+
needed and they do different jobs.
|
|
54
|
+
|
|
55
|
+
NEVER PERSISTED. Not settings.json (project tier beats user tier there,
|
|
56
|
+
D29 -- and a value that survives a restart is not what was asked for),
|
|
57
|
+
not extra_data, not anywhere. Restarting the harness is the guaranteed
|
|
58
|
+
way back to the derived number.
|
|
59
|
+
|
|
60
|
+
MODULE STATE, written from the UI thread and read from worker threads.
|
|
61
|
+
Same posture as core/compaction.py's `_compacting` flag and for the same
|
|
62
|
+
reason: these are single assignments of an int or a tuple, so there is no
|
|
63
|
+
half-set state to observe. Tests clear it in an autouse fixture, which is
|
|
64
|
+
the rule every other module-level cache in this project follows.
|
|
65
|
+
"""
|
|
66
|
+
|
|
67
|
+
import logging
|
|
68
|
+
from typing import Optional
|
|
69
|
+
|
|
70
|
+
logger = logging.getLogger(__name__)
|
|
71
|
+
|
|
72
|
+
# key -> (provider_name, model, value). One slot each; setting replaces.
|
|
73
|
+
_overrides: dict = {}
|
|
74
|
+
|
|
75
|
+
TRIGGER = "trigger"
|
|
76
|
+
|
|
77
|
+
|
|
78
|
+
def _set(key: str, provider_name: str, model: str, value: int) -> None:
|
|
79
|
+
_overrides[key] = (provider_name, model, int(value))
|
|
80
|
+
logger.info("Session %s override: %s for %s/%s (this session only).",
|
|
81
|
+
key, value, provider_name, model)
|
|
82
|
+
|
|
83
|
+
|
|
84
|
+
def _get(key: str, provider_name: Optional[str],
|
|
85
|
+
model: Optional[str]) -> Optional[int]:
|
|
86
|
+
"""The override for exactly this pair, or None.
|
|
87
|
+
|
|
88
|
+
A pair mismatch is not an error and does not clear -- see the module
|
|
89
|
+
docstring. `provider_name` is Optional because compaction.context_limit
|
|
90
|
+
still accepts callers that do not know it; without a provider no
|
|
91
|
+
override can match, which is the same answer as having none.
|
|
92
|
+
"""
|
|
93
|
+
entry = _overrides.get(key)
|
|
94
|
+
if entry is None or not provider_name:
|
|
95
|
+
return None
|
|
96
|
+
stored_provider, stored_model, value = entry
|
|
97
|
+
if stored_provider == provider_name and stored_model == model:
|
|
98
|
+
return value
|
|
99
|
+
return None
|
|
100
|
+
|
|
101
|
+
|
|
102
|
+
def set_trigger(provider_name: str, model: str, tokens: int) -> None:
|
|
103
|
+
"""Override the working-set compaction ceiling for this pair."""
|
|
104
|
+
_set(TRIGGER, provider_name, model, tokens)
|
|
105
|
+
|
|
106
|
+
|
|
107
|
+
def trigger_for(provider_name: Optional[str],
|
|
108
|
+
model: Optional[str]) -> Optional[int]:
|
|
109
|
+
return _get(TRIGGER, provider_name, model)
|
|
110
|
+
|
|
111
|
+
|
|
112
|
+
def trigger_any() -> Optional[int]:
|
|
113
|
+
"""The trigger override IGNORING the pair binding.
|
|
114
|
+
|
|
115
|
+
Exists for exactly one caller: compaction.pin_measurements, which
|
|
116
|
+
computes #89's cap as PIN_MAX_TRIGGER_FRACTION x the trigger and has
|
|
117
|
+
no model in scope to match against. That caller takes the MINIMUM of
|
|
118
|
+
this and the configured trigger, which is safe under every binding --
|
|
119
|
+
a raised session trigger leaves the cap stricter than it needs to be,
|
|
120
|
+
and a lowered one pulls the cap down with it. Reading the configured
|
|
121
|
+
trigger alone would let `/trigger 12k` leave pins capped at 20k, able
|
|
122
|
+
to floor a thread permanently above its own trigger, which is the
|
|
123
|
+
defect #89 exists to prevent.
|
|
124
|
+
|
|
125
|
+
Do not reach for this anywhere the pair IS available.
|
|
126
|
+
"""
|
|
127
|
+
entry = _overrides.get(TRIGGER)
|
|
128
|
+
return entry[2] if entry else None
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
def clear(key: Optional[str] = None) -> list:
|
|
132
|
+
"""Drop overrides and report which were actually dropped.
|
|
133
|
+
|
|
134
|
+
Called with no key by /model. `/trigger off` clears the same one by
|
|
135
|
+
name; the list form survives a second key being added rather than
|
|
136
|
+
being collapsed to the one that exists today. The return value is what
|
|
137
|
+
the shell says out loud -- a switch that silently discarded a number
|
|
138
|
+
the user set would be exactly the invisible state this feature is
|
|
139
|
+
careful not to introduce.
|
|
140
|
+
|
|
141
|
+
/window is NOT here since batch 44: a remembered window survives the
|
|
142
|
+
switch, and clearing it is `/window off`.
|
|
143
|
+
"""
|
|
144
|
+
keys = [key] if key else [TRIGGER]
|
|
145
|
+
dropped = [name for name in keys if _overrides.pop(name, None) is not None]
|
|
146
|
+
if dropped:
|
|
147
|
+
logger.info("Session overrides cleared: %s.", ", ".join(dropped))
|
|
148
|
+
return dropped
|
|
149
|
+
|
|
150
|
+
|
|
151
|
+
def state() -> dict:
|
|
152
|
+
"""{key: (provider, model, value)} for the bare-command report."""
|
|
153
|
+
return dict(_overrides)
|
|
@@ -0,0 +1,207 @@
|
|
|
1
|
+
"""
|
|
2
|
+
core/workspace_trust.py
|
|
3
|
+
|
|
4
|
+
Workspace trust gate (ROADMAP_v2 §14, D17). Project-level `.venastine/`
|
|
5
|
+
content (agents, skills, settings.json, AGENTS.md, mcp.json) only loads
|
|
6
|
+
after a one-time explicit user confirmation, keyed by resolved project
|
|
7
|
+
path + a content hash of everything under `.venastine/`. Changed content
|
|
8
|
+
after trust was granted produces a different hash and re-triggers the
|
|
9
|
+
prompt rather than silently inheriting stale trust.
|
|
10
|
+
|
|
11
|
+
The store lives at user level (~/.config/venastine/trusted_projects.json).
|
|
12
|
+
Until trust is granted, project-level content does not load AT ALL --
|
|
13
|
+
absent, not loaded-but-disabled. No partial trust.
|
|
14
|
+
"""
|
|
15
|
+
|
|
16
|
+
import hashlib
|
|
17
|
+
import json
|
|
18
|
+
import os
|
|
19
|
+
|
|
20
|
+
from json_store import write_json_atomic
|
|
21
|
+
|
|
22
|
+
|
|
23
|
+
def _trust_store_path() -> str:
|
|
24
|
+
# Resolved at call time, not import time, so tests can redirect the
|
|
25
|
+
# store via HOME/USERPROFILE and a moved config dir is picked up
|
|
26
|
+
# without a restart.
|
|
27
|
+
return os.path.expanduser("~/.config/venastine/trusted_projects.json")
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def venastine_dir(project_path: str) -> str:
|
|
31
|
+
return os.path.join(os.path.realpath(project_path), ".venastine")
|
|
32
|
+
|
|
33
|
+
|
|
34
|
+
#: The project's context document, and the ONE file outside .venastine/
|
|
35
|
+
#: that the loader injects into a system prompt (§44, replacing
|
|
36
|
+
#: .venastine/CONTEXT.md). Defined HERE rather than beside /init, because
|
|
37
|
+
#: the interesting fact about it is that it is inside the trust boundary:
|
|
38
|
+
#: config_loader reads it, tools/builtin/project_docs writes it, and
|
|
39
|
+
#: project_init/doc_sets names it, all from this one constant.
|
|
40
|
+
PROJECT_CONTEXT_FILENAME = "AGENTS.md"
|
|
41
|
+
|
|
42
|
+
|
|
43
|
+
def content_files(project_path: str) -> list[str]:
|
|
44
|
+
"""Sorted PROJECT-RELATIVE paths of everything a grant covers.
|
|
45
|
+
|
|
46
|
+
Everything under .venastine/, plus the root AGENTS.md when it exists.
|
|
47
|
+
|
|
48
|
+
THE ROOT FILE IS IN HERE BECAUSE IT REACHES THE SYSTEM PROMPT (§44).
|
|
49
|
+
D17 gates .venastine/ for one reason: its content becomes instructions,
|
|
50
|
+
and "project" means "arrived with a repo you cloned". Moving the hub
|
|
51
|
+
out of that directory without moving the boundary would have left a
|
|
52
|
+
cloned AGENTS.md entering an opted-in agent's system prompt with no
|
|
53
|
+
prompt at all -- D17's own stated threat, through the door §44 opened.
|
|
54
|
+
This module's invariant is the one that says so: the loader must read
|
|
55
|
+
no file the trust listing omits.
|
|
56
|
+
|
|
57
|
+
PATHS ARE RELATIVE TO THE PROJECT, not to .venastine/, which they were
|
|
58
|
+
before. They have to be: "AGENTS.md" and ".venastine/AGENTS.md" are
|
|
59
|
+
different files and a listing that could not tell them apart would be
|
|
60
|
+
a listing you cannot trust. The strings feed the hash, so every
|
|
61
|
+
existing grant re-prompts exactly once -- correct, because the set
|
|
62
|
+
being consented to has genuinely changed.
|
|
63
|
+
"""
|
|
64
|
+
resolved = os.path.realpath(project_path)
|
|
65
|
+
out = []
|
|
66
|
+
if os.path.isfile(os.path.join(resolved, PROJECT_CONTEXT_FILENAME)):
|
|
67
|
+
out.append(PROJECT_CONTEXT_FILENAME)
|
|
68
|
+
root = venastine_dir(project_path)
|
|
69
|
+
if os.path.isdir(root):
|
|
70
|
+
for dirpath, dirs, files in os.walk(root):
|
|
71
|
+
dirs.sort()
|
|
72
|
+
for fname in sorted(files):
|
|
73
|
+
# posix separators so the hash (and the trust-prompt
|
|
74
|
+
# listing) are identical regardless of platform path
|
|
75
|
+
# conventions
|
|
76
|
+
rel = os.path.relpath(os.path.join(dirpath, fname), resolved)
|
|
77
|
+
out.append(rel.replace(os.sep, "/"))
|
|
78
|
+
return out
|
|
79
|
+
|
|
80
|
+
|
|
81
|
+
def _content_hash(project_path: str) -> str:
|
|
82
|
+
"""Hashes relative path + content of everything content_files() lists.
|
|
83
|
+
|
|
84
|
+
ONE TRAVERSAL, NOT TWO THAT MUST AGREE (§44). This walked the tree a
|
|
85
|
+
second time and had to arrive at the same set as content_files() by
|
|
86
|
+
construction rather than by sharing code -- which #18 already showed
|
|
87
|
+
is not a safe way to keep two traversals aligned, and which AGENTS.md
|
|
88
|
+
recorded as the deeper fix still open. It iterates the listing now, so
|
|
89
|
+
the trust PROMPT and the trust DIGEST are one set of files by
|
|
90
|
+
definition: a file the listing omits cannot enter the hash, and this
|
|
91
|
+
module's invariant is that the loader may read no such file.
|
|
92
|
+
|
|
93
|
+
Two corrections survive from §14, both of which matter because this
|
|
94
|
+
hash is a security control rather than a cache key (Rev. 3):
|
|
95
|
+
|
|
96
|
+
1. `os.walk` yields subdirectories in filesystem order, not sorted
|
|
97
|
+
order. Sorting `files` alone is not enough -- `dirs` must be
|
|
98
|
+
sorted IN PLACE so os.walk itself descends deterministically, or
|
|
99
|
+
the same unchanged project can hash differently between runs and
|
|
100
|
+
re-prompt for trust that was already granted. That sort now lives
|
|
101
|
+
in content_files(), which is the traversal.
|
|
102
|
+
2. Hashing file CONTENT only makes the path invisible to the hash.
|
|
103
|
+
Two files swapping names produces an identical digest -- and the
|
|
104
|
+
directory a definition lives in is what determines how it's
|
|
105
|
+
loaded. The relative path is fed in too, with \\0 separators, so
|
|
106
|
+
concatenation boundaries can't be forged.
|
|
107
|
+
"""
|
|
108
|
+
resolved = os.path.realpath(project_path)
|
|
109
|
+
hasher = hashlib.sha256()
|
|
110
|
+
for rel in content_files(project_path):
|
|
111
|
+
path = os.path.join(resolved, rel.replace("/", os.sep))
|
|
112
|
+
hasher.update(rel.encode("utf-8"))
|
|
113
|
+
hasher.update(b"\0")
|
|
114
|
+
# Bounded reads, and regular files only. This hash runs BEFORE the
|
|
115
|
+
# user has consented to anything, over content that arrived with a
|
|
116
|
+
# directory they cloned: f.read() on a multi-GB blob is one
|
|
117
|
+
# unbounded allocation at startup, and open() on a FIFO (os.walk
|
|
118
|
+
# lists it, and a symlink to one resolves) blocks forever. Neither
|
|
119
|
+
# is content worth reading to decide whether to trust it -- its
|
|
120
|
+
# presence and name are enough.
|
|
121
|
+
if os.path.isfile(path):
|
|
122
|
+
with open(path, "rb") as f:
|
|
123
|
+
for chunk in iter(lambda: f.read(1 << 16), b""):
|
|
124
|
+
hasher.update(chunk)
|
|
125
|
+
else:
|
|
126
|
+
hasher.update(b"<non-regular file>")
|
|
127
|
+
hasher.update(b"\0")
|
|
128
|
+
return hasher.hexdigest()
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
def is_trusted(project_path: str) -> bool:
|
|
132
|
+
"""Whether this project's content may load.
|
|
133
|
+
|
|
134
|
+
"Nothing to trust" is now asked of the LISTING rather than of the
|
|
135
|
+
.venastine/ directory alone (§44). A project can have a root AGENTS.md
|
|
136
|
+
and no .venastine/ at all -- that is the ordinary shape of a repo
|
|
137
|
+
someone shares without their configuration, and it is exactly the
|
|
138
|
+
shape that must not walk into a system prompt unasked. The shortcut
|
|
139
|
+
narrows accordingly: a cloned repo shipping an AGENTS.md now prompts,
|
|
140
|
+
where before it was silent because the directory was absent.
|
|
141
|
+
|
|
142
|
+
An EMPTY .venastine/ also reads as nothing to trust now, where it
|
|
143
|
+
previously hashed to a constant and compared. Nothing loads from an
|
|
144
|
+
empty directory, so there was never a question to ask there.
|
|
145
|
+
"""
|
|
146
|
+
resolved = os.path.realpath(project_path)
|
|
147
|
+
if not content_files(project_path):
|
|
148
|
+
return True # nothing to trust -- no project-level content at all
|
|
149
|
+
current_hash = _content_hash(project_path)
|
|
150
|
+
store = _load_trust_store()
|
|
151
|
+
return store.get(resolved) == current_hash
|
|
152
|
+
|
|
153
|
+
|
|
154
|
+
def grant_trust(project_path: str) -> None:
|
|
155
|
+
resolved = os.path.realpath(project_path)
|
|
156
|
+
store = _load_trust_store()
|
|
157
|
+
store[resolved] = _content_hash(project_path)
|
|
158
|
+
_save_trust_store(store)
|
|
159
|
+
|
|
160
|
+
|
|
161
|
+
def _load_trust_store() -> dict:
|
|
162
|
+
"""The store, or {} for anything unreadable.
|
|
163
|
+
|
|
164
|
+
Fails CLOSED, in both senses. An empty store means nothing is
|
|
165
|
+
trusted, so a damaged file re-triggers the prompt -- which is this
|
|
166
|
+
module's stated posture for unknown state. Previously a partial write
|
|
167
|
+
(a kill, a Ctrl+C, a full disk) raised JSONDecodeError out of
|
|
168
|
+
is_trusted() at EVERY startup for EVERY project until the user found
|
|
169
|
+
and deleted the file by hand, and a valid-JSON-but-non-object store
|
|
170
|
+
escaped as an AttributeError on .get instead.
|
|
171
|
+
"""
|
|
172
|
+
path = _trust_store_path()
|
|
173
|
+
if not os.path.exists(path):
|
|
174
|
+
return {}
|
|
175
|
+
try:
|
|
176
|
+
with open(path, "r", encoding="utf-8-sig") as f:
|
|
177
|
+
data = json.load(f)
|
|
178
|
+
except (OSError, ValueError):
|
|
179
|
+
return {}
|
|
180
|
+
return data if isinstance(data, dict) else {}
|
|
181
|
+
|
|
182
|
+
|
|
183
|
+
def _save_trust_store(store: dict) -> None:
|
|
184
|
+
"""Write atomically, via json_store, which the three sibling user-tier
|
|
185
|
+
stores share. A truncate-then-write leaves the store empty for the
|
|
186
|
+
length of the write, and permanently damaged if the process dies
|
|
187
|
+
inside it -- for a file whose whole job is remembering a security
|
|
188
|
+
decision.
|
|
189
|
+
|
|
190
|
+
RAISES, unlike model_windows and preferences, which warn and return
|
|
191
|
+
False. That policy stays here rather than in the shared helper: a
|
|
192
|
+
trust grant that silently did not persist is worse than a crash, and
|
|
193
|
+
a remembered theme that did not is not.
|
|
194
|
+
|
|
195
|
+
0600 on the TEMP FILE, not on the destination (#19). os.replace
|
|
196
|
+
carries the source file's mode across, so setting it afterwards would
|
|
197
|
+
be both a no-op and a lie; setting it on the destination beforehand
|
|
198
|
+
is worse, since replace overwrites that too.
|
|
199
|
+
|
|
200
|
+
This store holds no secret, so the exposure it closes is only "which
|
|
201
|
+
projects this user has trusted, and their content hashes" --
|
|
202
|
+
information disclosure, not a breach. It was never group- or
|
|
203
|
+
other-WRITABLE, so no one could forge a trust entry. Done for
|
|
204
|
+
consistency with credentials.py rather than because this file is the
|
|
205
|
+
dangerous one.
|
|
206
|
+
"""
|
|
207
|
+
write_json_atomic(_trust_store_path(), store, mode=0o600)
|