@yahyahammad/venastine 0.3.8 → 0.3.10
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 +19 -19
- package/CONFIG_ARCHITECTURE.md +2139 -2002
- package/LICENSE +201 -201
- package/README.md +1043 -1034
- package/agents/builtin/compactor.md +75 -75
- package/agents/builtin/explore.md +77 -77
- package/agents/builtin/pipeline-reviewer.md +100 -100
- package/agents/builtin/plan.md +106 -106
- package/agents/builtin/review.md +74 -74
- package/agents/manager.py +261 -261
- package/agents/subagent_tool.py +472 -472
- package/agents/tui_commands.py +238 -146
- package/bin/venastine.mjs +584 -528
- package/bin/venastine.py +73 -71
- package/cli_commands.py +848 -835
- package/config.yaml +64 -0
- package/config_edit.py +1301 -1301
- package/config_schema.py +987 -966
- package/config_update.py +749 -749
- package/core/agent_activity.py +256 -256
- package/core/ansi.py +452 -452
- package/core/approval.py +137 -137
- package/core/call_view.py +364 -364
- package/core/cancel.py +121 -121
- package/core/chatgpt_auth.py +747 -747
- package/core/claims_view.py +107 -107
- package/core/client.py +2756 -2715
- package/core/commands.py +464 -454
- package/core/compaction.py +1735 -1735
- package/core/config_loader.py +1982 -1982
- package/core/connect.py +686 -686
- package/core/events.py +123 -123
- package/core/file_mentions.py +280 -280
- package/core/interaction.py +411 -411
- package/core/loop.py +2566 -2566
- package/core/memory.py +482 -482
- package/core/pipeline_models.py +190 -190
- package/core/provider_errors.py +274 -234
- package/core/reasoning/authorization.py +130 -130
- package/core/reasoning/base.py +219 -219
- package/core/reasoning/confidence_scoring.py +484 -484
- package/core/reasoning/events.py +149 -149
- package/core/reasoning/orchestrator.py +1561 -1561
- package/core/reasoning/output_writer.py +345 -345
- package/core/reasoning/payload_validation.py +524 -524
- package/core/reasoning/pipeline_storage.py +427 -427
- package/core/reasoning/scholar.py +467 -458
- package/core/reasoning/source_corpus.py +433 -433
- package/core/reasoning/source_scoring.py +855 -855
- package/core/replay.py +315 -315
- package/core/shell_sessions.py +1553 -1553
- package/credentials.py +307 -307
- package/database.py +333 -333
- package/docs/PRIVACY.md +78 -78
- package/docs/SECURITY.md +114 -108
- package/docs/THIRD_PARTY_NOTICES.md +148 -148
- package/env_secrets.py +118 -118
- package/json_store.py +343 -343
- package/logging_setup.py +295 -295
- package/main.py +3433 -3433
- package/mcp_client/client.py +780 -733
- package/mcp_client/config.py +484 -484
- package/mcp_client/registration.py +343 -343
- package/memories/manager.py +147 -147
- package/memories/tui_commands.py +319 -239
- package/package.json +303 -303
- package/project_init/config_files.py +282 -282
- package/project_init/generator.py +728 -728
- package/project_init/manifest.py +418 -418
- package/project_init/tui_commands.py +147 -147
- package/prompts/critic_pass.md +24 -24
- package/prompts/revalidate.md +25 -25
- package/prompts/revise.md +17 -17
- package/prompts/source_grounding.md +66 -66
- package/prompts/system_prompts.py +316 -316
- package/providers.json.example +94 -94
- package/pyproject.toml +149 -149
- package/requirements.txt +314 -295
- package/safety/policy_enforcement.py +1082 -1082
- package/security/askpass.py +419 -419
- package/security/capability.py +207 -207
- package/security/confine.py +484 -484
- package/security/hosted.py +182 -182
- package/security/permissions.py +237 -237
- package/security/protected_paths.py +501 -501
- package/security/sandbox.py +4108 -4108
- package/skills/builtin/math/formal-specification.md +67 -67
- package/skills/builtin/math/numerical-methods.md +66 -66
- package/skills/builtin/research/experiment-design.md +62 -62
- package/skills/builtin/research/reproducibility.md +55 -55
- package/skills/builtin/research/source-evaluation.md +70 -70
- package/skills/builtin/research/statistical-inference.md +63 -63
- package/skills/builtin/research/technical-writing.md +59 -59
- package/skills/builtin/software/code-review.md +61 -61
- package/skills/builtin/software/debugging.md +61 -61
- package/skills/builtin/software/test-design.md +67 -67
- package/skills/tui_commands.py +110 -110
- package/storage.py +1338 -1338
- package/tools/base.py +465 -465
- package/tools/builtin/_math_common.py +667 -667
- package/tools/builtin/_net_common.py +1145 -267
- package/tools/builtin/_page_text.py +490 -0
- package/tools/builtin/_pdf_text.py +113 -0
- package/tools/builtin/arxiv.py +685 -681
- package/tools/builtin/fetch_url.py +639 -371
- package/tools/builtin/file_ops.py +720 -720
- package/tools/builtin/geometry.py +184 -184
- package/tools/builtin/load_skill.py +76 -76
- package/tools/builtin/pin.py +227 -227
- package/tools/builtin/probability_stats.py +224 -224
- package/tools/builtin/project_docs.py +464 -464
- package/tools/builtin/remember.py +175 -175
- package/tools/builtin/shell.py +980 -980
- package/tools/builtin/todo.py +149 -149
- package/tools/builtin/web_search.py +402 -328
- package/tools/context.py +159 -159
- package/tools/isolation.py +387 -387
- package/tools/registry.py +1252 -1242
- package/tui/app.py +8372 -8307
- package/tui/app.tcss +754 -739
- package/tui/commands.py +26 -26
- package/tui/diffs.py +286 -286
- package/tui/history.py +128 -128
- package/tui/hosted_driver.py +102 -102
- package/tui/markdown.py +1329 -1329
- package/tui/meters.py +364 -364
- package/tui/ravens.py +146 -146
- package/tui/screens.py +1255 -1197
- package/tui/themes.py +599 -599
- package/tui/widgets.py +3861 -3796
package/agents/manager.py
CHANGED
|
@@ -1,261 +1,261 @@
|
|
|
1
|
-
"""
|
|
2
|
-
agents/manager.py
|
|
3
|
-
|
|
4
|
-
ROADMAP_v2 §18. AgentManager discovers agent `.md` files across the three
|
|
5
|
-
tiers and exposes lookup by name plus the §18 scope-composition rules.
|
|
6
|
-
|
|
7
|
-
Thin by design: discovery, frontmatter parsing, and tier precedence
|
|
8
|
-
(D29: harness > user > project) all live in core/config_loader.py, which
|
|
9
|
-
runs at initialize() time. This file adds only the agent-specific policy
|
|
10
|
-
composition that §18 owns and that §20 (subagent review) will call:
|
|
11
|
-
|
|
12
|
-
* active_context() -- the ToolContext for a session-scoped active
|
|
13
|
-
agent (/agent switch): the agent's own restrictions at depth 0.
|
|
14
|
-
* child_context() -- the ToolContext for a spawned subagent: C6's
|
|
15
|
-
intersection-with-parent (never union) plus a depth increment (C3).
|
|
16
|
-
* system_prompt_for() -- base + catalogs + the agent's body + opt-in
|
|
17
|
-
project context, assembled in one place so the TUI worker and
|
|
18
|
-
spawn_subagent cannot drift apart.
|
|
19
|
-
"""
|
|
20
|
-
|
|
21
|
-
from typing import Optional
|
|
22
|
-
|
|
23
|
-
from core import config_loader
|
|
24
|
-
from core.config_loader import AgentDef
|
|
25
|
-
from memories.manager import manager as memory_manager
|
|
26
|
-
from tools.context import ToolContext
|
|
27
|
-
|
|
28
|
-
#: What /grill-me asks, in either shell (§18; shared since ROADMAP_v3 §89).
|
|
29
|
-
GRILL_ME_TASK = "Grill this thread: surface what still needs a decision."
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
class AgentManager:
|
|
33
|
-
"""Lookup + scope composition over the loader's discovered agents."""
|
|
34
|
-
|
|
35
|
-
def get(self, name: str) -> Optional[AgentDef]:
|
|
36
|
-
return config_loader.get_agent(name)
|
|
37
|
-
|
|
38
|
-
def names(self) -> list:
|
|
39
|
-
return sorted(config_loader.get_agents())
|
|
40
|
-
|
|
41
|
-
def all(self) -> dict:
|
|
42
|
-
return config_loader.get_agents()
|
|
43
|
-
|
|
44
|
-
@staticmethod
|
|
45
|
-
def active_context(agent: AgentDef) -> ToolContext:
|
|
46
|
-
"""ToolContext for running AS this agent (session-scoped /agent
|
|
47
|
-
switch). The agent's own restrictions at depth 0 -- no parent to
|
|
48
|
-
intersect with, no nesting."""
|
|
49
|
-
allowed = (
|
|
50
|
-
set(agent.allowed_tools)
|
|
51
|
-
if agent.allowed_tools is not None else None
|
|
52
|
-
)
|
|
53
|
-
targets = (
|
|
54
|
-
set(agent.spawn_targets)
|
|
55
|
-
if agent.spawn_targets is not None else None
|
|
56
|
-
)
|
|
57
|
-
return ToolContext(
|
|
58
|
-
allowed_tools=allowed,
|
|
59
|
-
approval_overrides=dict(agent.approval_overrides),
|
|
60
|
-
subagent_depth=0,
|
|
61
|
-
spawn_targets=targets,
|
|
62
|
-
)
|
|
63
|
-
|
|
64
|
-
@staticmethod
|
|
65
|
-
def child_context(
|
|
66
|
-
agent: AgentDef, parent_context: Optional[ToolContext],
|
|
67
|
-
) -> ToolContext:
|
|
68
|
-
"""ToolContext for a subagent spawned WHILE running under
|
|
69
|
-
parent_context. C6: intersection, never union -- a restricted
|
|
70
|
-
parent must not escape its restriction by spawning a permissive
|
|
71
|
-
child. C3: depth increments so the nesting limit is enforceable.
|
|
72
|
-
|
|
73
|
-
An agent that declares no allowed_tools inherits the parent's
|
|
74
|
-
restriction unchanged (it may not widen it either).
|
|
75
|
-
|
|
76
|
-
APPROVAL COMPOSES THE SAME WAY (§14-§18 review f12, decision S2).
|
|
77
|
-
§18's spec sketch built the child's overrides from the agent
|
|
78
|
-
definition alone, which drops the parent's. C6 reasoned about
|
|
79
|
-
allowed_tools and never considered the approval axis, so the same
|
|
80
|
-
escalation it exists to prevent was open one field over: a parent
|
|
81
|
-
declaring `approval_overrides: {web_search: true}` prompts on
|
|
82
|
-
every search, then spawns a child that computes
|
|
83
|
-
approval_needed(web_search) == False and runs it ungated.
|
|
84
|
-
|
|
85
|
-
Union rather than replace. Only True entries carry -- a False is
|
|
86
|
-
indistinguishable from absent under D14's OR, so unioning can
|
|
87
|
-
only ever tighten, which is what makes it safe to do silently.
|
|
88
|
-
|
|
89
|
-
BOTH SIDES ARE FILTERED, and batch 150 is why. The parent's dict
|
|
90
|
-
was filtered and the agent's was merged over it raw, so a `False`
|
|
91
|
-
in the child's frontmatter OVERWROTE the parent's `True` and the
|
|
92
|
-
sentence above was false of the line below it: a parent that
|
|
93
|
-
prompts on every search spawned a child that did not, which is
|
|
94
|
-
the exact escalation the paragraph before this one describes.
|
|
95
|
-
It did not heal one hop further out either -- the next `if v`
|
|
96
|
-
dropped the False rather than restoring the True, so the
|
|
97
|
-
tightening was gone permanently rather than suspended for one
|
|
98
|
-
run. Three of the four quadrants had tests; the one the code got
|
|
99
|
-
wrong was the one with none.
|
|
100
|
-
|
|
101
|
-
The cost is not only the gate. `candidate_approvals` reads this
|
|
102
|
-
same dict, and it is BOTH the sign-off notice and the grant the
|
|
103
|
-
child receives -- measured, the loosened tool left the candidate
|
|
104
|
-
list entirely, so the user was neither asked at call time nor
|
|
105
|
-
told at sign-off time.
|
|
106
|
-
|
|
107
|
-
`active_context` needs no filter and deliberately has none: at
|
|
108
|
-
depth 0 there is no parent to overwrite, a False is inert under
|
|
109
|
-
D14's OR, and the next hop's filter drops it. Measured both
|
|
110
|
-
ways rather than reasoned about, because "the other composer
|
|
111
|
-
looks similar" is how a fix grows past its finding."""
|
|
112
|
-
parent = parent_context or ToolContext()
|
|
113
|
-
if agent.allowed_tools is not None:
|
|
114
|
-
own = set(agent.allowed_tools)
|
|
115
|
-
child_allowed = own & (
|
|
116
|
-
parent.allowed_tools if parent.allowed_tools is not None
|
|
117
|
-
else own
|
|
118
|
-
)
|
|
119
|
-
else:
|
|
120
|
-
child_allowed = (
|
|
121
|
-
set(parent.allowed_tools)
|
|
122
|
-
if parent.allowed_tools is not None else None
|
|
123
|
-
)
|
|
124
|
-
overrides = {k: v for k, v in parent.approval_overrides.items() if v}
|
|
125
|
-
overrides.update(
|
|
126
|
-
{k: v for k, v in agent.approval_overrides.items() if v})
|
|
127
|
-
# SPAWN TARGETS COMPOSE THE SAME WAY, for C6's own reason one
|
|
128
|
-
# field further out. The refusal that reads this field is checked
|
|
129
|
-
# against the context a run holds, so a field that applied only at
|
|
130
|
-
# depth 0 would be escapable by nesting: an agent forbidden to
|
|
131
|
-
# reach `build` could spawn something that was not, and have that
|
|
132
|
-
# one spawn it. Intersection makes the restriction survive the
|
|
133
|
-
# hop, and an agent declaring none inherits its parent's roster
|
|
134
|
-
# unchanged rather than widening back to everything.
|
|
135
|
-
if agent.spawn_targets is not None:
|
|
136
|
-
own_targets = set(agent.spawn_targets)
|
|
137
|
-
child_targets = own_targets & (
|
|
138
|
-
parent.spawn_targets if parent.spawn_targets is not None
|
|
139
|
-
else own_targets
|
|
140
|
-
)
|
|
141
|
-
else:
|
|
142
|
-
child_targets = (
|
|
143
|
-
set(parent.spawn_targets)
|
|
144
|
-
if parent.spawn_targets is not None else None
|
|
145
|
-
)
|
|
146
|
-
return ToolContext(
|
|
147
|
-
allowed_tools=child_allowed,
|
|
148
|
-
approval_overrides=overrides,
|
|
149
|
-
subagent_depth=parent.subagent_depth + 1,
|
|
150
|
-
spawn_targets=child_targets,
|
|
151
|
-
)
|
|
152
|
-
|
|
153
|
-
@staticmethod
|
|
154
|
-
def candidate_approvals(child_context: ToolContext) -> list:
|
|
155
|
-
"""Tools the subagent could reach that would each need approval.
|
|
156
|
-
|
|
157
|
-
ONE helper, TWO callers -- the sign-off prompt's notice text and
|
|
158
|
-
the grant the child actually receives. Computing the list twice
|
|
159
|
-
would let what the user approved drift from what the child got,
|
|
160
|
-
which is the whole substance of the sign-off.
|
|
161
|
-
|
|
162
|
-
Params are empty because there are none yet: a path-dependent
|
|
163
|
-
approval_check (file_ops outside the workspace, shell on a
|
|
164
|
-
non-inert command) cannot be resolved before the call exists, so
|
|
165
|
-
those still prompt at call time through the inherited channel.
|
|
166
|
-
|
|
167
|
-
§25 (R2) filters those tools OUT of the list rather than merely
|
|
168
|
-
letting them re-prompt. They were never grantable by name, and the
|
|
169
|
-
loop now enforces that -- so leaving them here would make the
|
|
170
|
-
sign-off notice promise authority the grant does not carry, which
|
|
171
|
-
is worse than not offering it. What the user reads and what the
|
|
172
|
-
child receives have to be the same list; that is the whole
|
|
173
|
-
substance of a sign-off, and this helper exists to keep them one.
|
|
174
|
-
|
|
175
|
-
§25 (R13) adds the second filter, and #67 is why it is here rather
|
|
176
|
-
than only in core/reasoning/authorization.py. Both functions
|
|
177
|
-
answer "what may be pre-granted", both were built from
|
|
178
|
-
headless_hidden + grantable, and only the pipeline's one then
|
|
179
|
-
subtracted an exclusion list -- which lived in the pipeline's
|
|
180
|
-
module where this one could not see it. So this list offered
|
|
181
|
-
`spawn_subagent` and `remember`, the two tools R4 and M17 exist to
|
|
182
|
-
keep out of a grant, and R4's argument is ABOUT this sign-off.
|
|
183
|
-
Now both read one per-tool declaration.
|
|
184
|
-
|
|
185
|
-
EMPTY FOR EVERY SHIPPED AGENT, on every install, and that is a
|
|
186
|
-
ruled scope rather than a gap (TECHNICAL_DEBT 83, batch 139).
|
|
187
|
-
The pool this returns from is the gated-with-no-approval_check
|
|
188
|
-
set -- the MCP tools plus `write_project_doc` -- and the
|
|
189
|
-
intersection with `child_context` is what empties it: all six
|
|
190
|
-
spawnable shipped agents narrow `allowed_tools` to built-ins,
|
|
191
|
-
and an agent file cannot name `mcp__server__tool` for a server
|
|
192
|
-
that did not exist when it was written. So the sign-off's
|
|
193
|
-
PICKER branch belongs to an agent that declines to narrow, or
|
|
194
|
-
one naming a gated tool outright, which is a user- or
|
|
195
|
-
project-tier file. A shipped spawn gets the no-candidates
|
|
196
|
-
branch, which `tui/screens.py` builds deliberately.
|
|
197
|
-
|
|
198
|
-
The LOOSE half of the pair: GRANT_SIGNOFF_ONLY passes here and is
|
|
199
|
-
refused by the pipeline. A human answering in the moment, about
|
|
200
|
-
one named agent, for one turn, is the consent write_project_doc's
|
|
201
|
-
notice was written for; ten unattended passes on one launch-time
|
|
202
|
-
tick is not.
|
|
203
|
-
"""
|
|
204
|
-
from tools.base import GRANT_NEVER
|
|
205
|
-
from tools.registry import registry
|
|
206
|
-
|
|
207
|
-
return sorted(
|
|
208
|
-
name for name in registry.headless_hidden(child_context)
|
|
209
|
-
if registry.grantable(name)
|
|
210
|
-
and registry.grant_policy(name) != GRANT_NEVER
|
|
211
|
-
)
|
|
212
|
-
|
|
213
|
-
@staticmethod
|
|
214
|
-
def system_prompt_for(agent: AgentDef, base_prompt: str,
|
|
215
|
-
active_skills=None, context=None,
|
|
216
|
-
callable_only: bool = False,
|
|
217
|
-
granted=None) -> str:
|
|
218
|
-
"""The full system prompt for a run AS this agent: base + skill /
|
|
219
|
-
agent catalogs, then the agent's own body, then the project's
|
|
220
|
-
AGENTS.md iff the agent opts in (use_project_context). One
|
|
221
|
-
assembly point so every caller produces the same prompt.
|
|
222
|
-
|
|
223
|
-
active_skills (§19) reaches only the catalog, where it MARKS the
|
|
224
|
-
skills whose bodies the shell pins separately. Bodies are not
|
|
225
|
-
added here -- see with_catalogs' K6 note.
|
|
226
|
-
|
|
227
|
-
context (§20 review f19, generalised): the ToolContext this run
|
|
228
|
-
will actually use. It reaches with_catalogs, which suppresses a
|
|
229
|
-
catalog whose tool the context excludes -- otherwise a restricted
|
|
230
|
-
agent is told to "call the load_skill tool" it does not have.
|
|
231
|
-
Pass the SAME context the run gets: child_context() for a spawned
|
|
232
|
-
subagent, so a parent's restriction suppresses the invitation too;
|
|
233
|
-
active_context() for a /agent switch. None means the run is under
|
|
234
|
-
global policy only, and both catalogs apply."""
|
|
235
|
-
import prompts.system_prompts as system_prompts
|
|
236
|
-
|
|
237
|
-
# callable_only / granted (#68): forwarded, not decided here.
|
|
238
|
-
# This function knows which agent is running; whether anything
|
|
239
|
-
# can answer an approval question is the RUN's fact, and the
|
|
240
|
-
# caller is the one holding it. Defaulting to the attended
|
|
241
|
-
# answer keeps both TUI call sites correct with no argument.
|
|
242
|
-
prompt = system_prompts.with_catalogs(
|
|
243
|
-
base_prompt, active_skills, context,
|
|
244
|
-
callable_only=callable_only, granted=granted)
|
|
245
|
-
prompt = f"{prompt}\n\n## Agent: {agent.name}\n\n{agent.body}"
|
|
246
|
-
context = config_loader.context_for_agent(agent)
|
|
247
|
-
if context:
|
|
248
|
-
prompt = f"{prompt}\n\n## Project context\n\n{context}"
|
|
249
|
-
# ROADMAP_v2 §21b: durable memories, the third context tier beside
|
|
250
|
-
# the project's AGENTS.md and gated the same way (use_memory).
|
|
251
|
-
# Appended HERE and not inside with_catalogs, which feeds
|
|
252
|
-
# pass_prompt() -- injecting there would put every memory into all
|
|
253
|
-
# ten research passes, which is §19's K6 trap in its second
|
|
254
|
-
# instance.
|
|
255
|
-
memories = memory_manager.prompt_fragment(agent)
|
|
256
|
-
if memories:
|
|
257
|
-
prompt = f"{prompt}\n\n{memories}"
|
|
258
|
-
return prompt
|
|
259
|
-
|
|
260
|
-
|
|
261
|
-
manager = AgentManager()
|
|
1
|
+
"""
|
|
2
|
+
agents/manager.py
|
|
3
|
+
|
|
4
|
+
ROADMAP_v2 §18. AgentManager discovers agent `.md` files across the three
|
|
5
|
+
tiers and exposes lookup by name plus the §18 scope-composition rules.
|
|
6
|
+
|
|
7
|
+
Thin by design: discovery, frontmatter parsing, and tier precedence
|
|
8
|
+
(D29: harness > user > project) all live in core/config_loader.py, which
|
|
9
|
+
runs at initialize() time. This file adds only the agent-specific policy
|
|
10
|
+
composition that §18 owns and that §20 (subagent review) will call:
|
|
11
|
+
|
|
12
|
+
* active_context() -- the ToolContext for a session-scoped active
|
|
13
|
+
agent (/agent switch): the agent's own restrictions at depth 0.
|
|
14
|
+
* child_context() -- the ToolContext for a spawned subagent: C6's
|
|
15
|
+
intersection-with-parent (never union) plus a depth increment (C3).
|
|
16
|
+
* system_prompt_for() -- base + catalogs + the agent's body + opt-in
|
|
17
|
+
project context, assembled in one place so the TUI worker and
|
|
18
|
+
spawn_subagent cannot drift apart.
|
|
19
|
+
"""
|
|
20
|
+
|
|
21
|
+
from typing import Optional
|
|
22
|
+
|
|
23
|
+
from core import config_loader
|
|
24
|
+
from core.config_loader import AgentDef
|
|
25
|
+
from memories.manager import manager as memory_manager
|
|
26
|
+
from tools.context import ToolContext
|
|
27
|
+
|
|
28
|
+
#: What /grill-me asks, in either shell (§18; shared since ROADMAP_v3 §89).
|
|
29
|
+
GRILL_ME_TASK = "Grill this thread: surface what still needs a decision."
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
class AgentManager:
|
|
33
|
+
"""Lookup + scope composition over the loader's discovered agents."""
|
|
34
|
+
|
|
35
|
+
def get(self, name: str) -> Optional[AgentDef]:
|
|
36
|
+
return config_loader.get_agent(name)
|
|
37
|
+
|
|
38
|
+
def names(self) -> list:
|
|
39
|
+
return sorted(config_loader.get_agents())
|
|
40
|
+
|
|
41
|
+
def all(self) -> dict:
|
|
42
|
+
return config_loader.get_agents()
|
|
43
|
+
|
|
44
|
+
@staticmethod
|
|
45
|
+
def active_context(agent: AgentDef) -> ToolContext:
|
|
46
|
+
"""ToolContext for running AS this agent (session-scoped /agent
|
|
47
|
+
switch). The agent's own restrictions at depth 0 -- no parent to
|
|
48
|
+
intersect with, no nesting."""
|
|
49
|
+
allowed = (
|
|
50
|
+
set(agent.allowed_tools)
|
|
51
|
+
if agent.allowed_tools is not None else None
|
|
52
|
+
)
|
|
53
|
+
targets = (
|
|
54
|
+
set(agent.spawn_targets)
|
|
55
|
+
if agent.spawn_targets is not None else None
|
|
56
|
+
)
|
|
57
|
+
return ToolContext(
|
|
58
|
+
allowed_tools=allowed,
|
|
59
|
+
approval_overrides=dict(agent.approval_overrides),
|
|
60
|
+
subagent_depth=0,
|
|
61
|
+
spawn_targets=targets,
|
|
62
|
+
)
|
|
63
|
+
|
|
64
|
+
@staticmethod
|
|
65
|
+
def child_context(
|
|
66
|
+
agent: AgentDef, parent_context: Optional[ToolContext],
|
|
67
|
+
) -> ToolContext:
|
|
68
|
+
"""ToolContext for a subagent spawned WHILE running under
|
|
69
|
+
parent_context. C6: intersection, never union -- a restricted
|
|
70
|
+
parent must not escape its restriction by spawning a permissive
|
|
71
|
+
child. C3: depth increments so the nesting limit is enforceable.
|
|
72
|
+
|
|
73
|
+
An agent that declares no allowed_tools inherits the parent's
|
|
74
|
+
restriction unchanged (it may not widen it either).
|
|
75
|
+
|
|
76
|
+
APPROVAL COMPOSES THE SAME WAY (§14-§18 review f12, decision S2).
|
|
77
|
+
§18's spec sketch built the child's overrides from the agent
|
|
78
|
+
definition alone, which drops the parent's. C6 reasoned about
|
|
79
|
+
allowed_tools and never considered the approval axis, so the same
|
|
80
|
+
escalation it exists to prevent was open one field over: a parent
|
|
81
|
+
declaring `approval_overrides: {web_search: true}` prompts on
|
|
82
|
+
every search, then spawns a child that computes
|
|
83
|
+
approval_needed(web_search) == False and runs it ungated.
|
|
84
|
+
|
|
85
|
+
Union rather than replace. Only True entries carry -- a False is
|
|
86
|
+
indistinguishable from absent under D14's OR, so unioning can
|
|
87
|
+
only ever tighten, which is what makes it safe to do silently.
|
|
88
|
+
|
|
89
|
+
BOTH SIDES ARE FILTERED, and batch 150 is why. The parent's dict
|
|
90
|
+
was filtered and the agent's was merged over it raw, so a `False`
|
|
91
|
+
in the child's frontmatter OVERWROTE the parent's `True` and the
|
|
92
|
+
sentence above was false of the line below it: a parent that
|
|
93
|
+
prompts on every search spawned a child that did not, which is
|
|
94
|
+
the exact escalation the paragraph before this one describes.
|
|
95
|
+
It did not heal one hop further out either -- the next `if v`
|
|
96
|
+
dropped the False rather than restoring the True, so the
|
|
97
|
+
tightening was gone permanently rather than suspended for one
|
|
98
|
+
run. Three of the four quadrants had tests; the one the code got
|
|
99
|
+
wrong was the one with none.
|
|
100
|
+
|
|
101
|
+
The cost is not only the gate. `candidate_approvals` reads this
|
|
102
|
+
same dict, and it is BOTH the sign-off notice and the grant the
|
|
103
|
+
child receives -- measured, the loosened tool left the candidate
|
|
104
|
+
list entirely, so the user was neither asked at call time nor
|
|
105
|
+
told at sign-off time.
|
|
106
|
+
|
|
107
|
+
`active_context` needs no filter and deliberately has none: at
|
|
108
|
+
depth 0 there is no parent to overwrite, a False is inert under
|
|
109
|
+
D14's OR, and the next hop's filter drops it. Measured both
|
|
110
|
+
ways rather than reasoned about, because "the other composer
|
|
111
|
+
looks similar" is how a fix grows past its finding."""
|
|
112
|
+
parent = parent_context or ToolContext()
|
|
113
|
+
if agent.allowed_tools is not None:
|
|
114
|
+
own = set(agent.allowed_tools)
|
|
115
|
+
child_allowed = own & (
|
|
116
|
+
parent.allowed_tools if parent.allowed_tools is not None
|
|
117
|
+
else own
|
|
118
|
+
)
|
|
119
|
+
else:
|
|
120
|
+
child_allowed = (
|
|
121
|
+
set(parent.allowed_tools)
|
|
122
|
+
if parent.allowed_tools is not None else None
|
|
123
|
+
)
|
|
124
|
+
overrides = {k: v for k, v in parent.approval_overrides.items() if v}
|
|
125
|
+
overrides.update(
|
|
126
|
+
{k: v for k, v in agent.approval_overrides.items() if v})
|
|
127
|
+
# SPAWN TARGETS COMPOSE THE SAME WAY, for C6's own reason one
|
|
128
|
+
# field further out. The refusal that reads this field is checked
|
|
129
|
+
# against the context a run holds, so a field that applied only at
|
|
130
|
+
# depth 0 would be escapable by nesting: an agent forbidden to
|
|
131
|
+
# reach `build` could spawn something that was not, and have that
|
|
132
|
+
# one spawn it. Intersection makes the restriction survive the
|
|
133
|
+
# hop, and an agent declaring none inherits its parent's roster
|
|
134
|
+
# unchanged rather than widening back to everything.
|
|
135
|
+
if agent.spawn_targets is not None:
|
|
136
|
+
own_targets = set(agent.spawn_targets)
|
|
137
|
+
child_targets = own_targets & (
|
|
138
|
+
parent.spawn_targets if parent.spawn_targets is not None
|
|
139
|
+
else own_targets
|
|
140
|
+
)
|
|
141
|
+
else:
|
|
142
|
+
child_targets = (
|
|
143
|
+
set(parent.spawn_targets)
|
|
144
|
+
if parent.spawn_targets is not None else None
|
|
145
|
+
)
|
|
146
|
+
return ToolContext(
|
|
147
|
+
allowed_tools=child_allowed,
|
|
148
|
+
approval_overrides=overrides,
|
|
149
|
+
subagent_depth=parent.subagent_depth + 1,
|
|
150
|
+
spawn_targets=child_targets,
|
|
151
|
+
)
|
|
152
|
+
|
|
153
|
+
@staticmethod
|
|
154
|
+
def candidate_approvals(child_context: ToolContext) -> list:
|
|
155
|
+
"""Tools the subagent could reach that would each need approval.
|
|
156
|
+
|
|
157
|
+
ONE helper, TWO callers -- the sign-off prompt's notice text and
|
|
158
|
+
the grant the child actually receives. Computing the list twice
|
|
159
|
+
would let what the user approved drift from what the child got,
|
|
160
|
+
which is the whole substance of the sign-off.
|
|
161
|
+
|
|
162
|
+
Params are empty because there are none yet: a path-dependent
|
|
163
|
+
approval_check (file_ops outside the workspace, shell on a
|
|
164
|
+
non-inert command) cannot be resolved before the call exists, so
|
|
165
|
+
those still prompt at call time through the inherited channel.
|
|
166
|
+
|
|
167
|
+
§25 (R2) filters those tools OUT of the list rather than merely
|
|
168
|
+
letting them re-prompt. They were never grantable by name, and the
|
|
169
|
+
loop now enforces that -- so leaving them here would make the
|
|
170
|
+
sign-off notice promise authority the grant does not carry, which
|
|
171
|
+
is worse than not offering it. What the user reads and what the
|
|
172
|
+
child receives have to be the same list; that is the whole
|
|
173
|
+
substance of a sign-off, and this helper exists to keep them one.
|
|
174
|
+
|
|
175
|
+
§25 (R13) adds the second filter, and #67 is why it is here rather
|
|
176
|
+
than only in core/reasoning/authorization.py. Both functions
|
|
177
|
+
answer "what may be pre-granted", both were built from
|
|
178
|
+
headless_hidden + grantable, and only the pipeline's one then
|
|
179
|
+
subtracted an exclusion list -- which lived in the pipeline's
|
|
180
|
+
module where this one could not see it. So this list offered
|
|
181
|
+
`spawn_subagent` and `remember`, the two tools R4 and M17 exist to
|
|
182
|
+
keep out of a grant, and R4's argument is ABOUT this sign-off.
|
|
183
|
+
Now both read one per-tool declaration.
|
|
184
|
+
|
|
185
|
+
EMPTY FOR EVERY SHIPPED AGENT, on every install, and that is a
|
|
186
|
+
ruled scope rather than a gap (TECHNICAL_DEBT 83, batch 139).
|
|
187
|
+
The pool this returns from is the gated-with-no-approval_check
|
|
188
|
+
set -- the MCP tools plus `write_project_doc` -- and the
|
|
189
|
+
intersection with `child_context` is what empties it: all six
|
|
190
|
+
spawnable shipped agents narrow `allowed_tools` to built-ins,
|
|
191
|
+
and an agent file cannot name `mcp__server__tool` for a server
|
|
192
|
+
that did not exist when it was written. So the sign-off's
|
|
193
|
+
PICKER branch belongs to an agent that declines to narrow, or
|
|
194
|
+
one naming a gated tool outright, which is a user- or
|
|
195
|
+
project-tier file. A shipped spawn gets the no-candidates
|
|
196
|
+
branch, which `tui/screens.py` builds deliberately.
|
|
197
|
+
|
|
198
|
+
The LOOSE half of the pair: GRANT_SIGNOFF_ONLY passes here and is
|
|
199
|
+
refused by the pipeline. A human answering in the moment, about
|
|
200
|
+
one named agent, for one turn, is the consent write_project_doc's
|
|
201
|
+
notice was written for; ten unattended passes on one launch-time
|
|
202
|
+
tick is not.
|
|
203
|
+
"""
|
|
204
|
+
from tools.base import GRANT_NEVER
|
|
205
|
+
from tools.registry import registry
|
|
206
|
+
|
|
207
|
+
return sorted(
|
|
208
|
+
name for name in registry.headless_hidden(child_context)
|
|
209
|
+
if registry.grantable(name)
|
|
210
|
+
and registry.grant_policy(name) != GRANT_NEVER
|
|
211
|
+
)
|
|
212
|
+
|
|
213
|
+
@staticmethod
|
|
214
|
+
def system_prompt_for(agent: AgentDef, base_prompt: str,
|
|
215
|
+
active_skills=None, context=None,
|
|
216
|
+
callable_only: bool = False,
|
|
217
|
+
granted=None) -> str:
|
|
218
|
+
"""The full system prompt for a run AS this agent: base + skill /
|
|
219
|
+
agent catalogs, then the agent's own body, then the project's
|
|
220
|
+
AGENTS.md iff the agent opts in (use_project_context). One
|
|
221
|
+
assembly point so every caller produces the same prompt.
|
|
222
|
+
|
|
223
|
+
active_skills (§19) reaches only the catalog, where it MARKS the
|
|
224
|
+
skills whose bodies the shell pins separately. Bodies are not
|
|
225
|
+
added here -- see with_catalogs' K6 note.
|
|
226
|
+
|
|
227
|
+
context (§20 review f19, generalised): the ToolContext this run
|
|
228
|
+
will actually use. It reaches with_catalogs, which suppresses a
|
|
229
|
+
catalog whose tool the context excludes -- otherwise a restricted
|
|
230
|
+
agent is told to "call the load_skill tool" it does not have.
|
|
231
|
+
Pass the SAME context the run gets: child_context() for a spawned
|
|
232
|
+
subagent, so a parent's restriction suppresses the invitation too;
|
|
233
|
+
active_context() for a /agent switch. None means the run is under
|
|
234
|
+
global policy only, and both catalogs apply."""
|
|
235
|
+
import prompts.system_prompts as system_prompts
|
|
236
|
+
|
|
237
|
+
# callable_only / granted (#68): forwarded, not decided here.
|
|
238
|
+
# This function knows which agent is running; whether anything
|
|
239
|
+
# can answer an approval question is the RUN's fact, and the
|
|
240
|
+
# caller is the one holding it. Defaulting to the attended
|
|
241
|
+
# answer keeps both TUI call sites correct with no argument.
|
|
242
|
+
prompt = system_prompts.with_catalogs(
|
|
243
|
+
base_prompt, active_skills, context,
|
|
244
|
+
callable_only=callable_only, granted=granted)
|
|
245
|
+
prompt = f"{prompt}\n\n## Agent: {agent.name}\n\n{agent.body}"
|
|
246
|
+
context = config_loader.context_for_agent(agent)
|
|
247
|
+
if context:
|
|
248
|
+
prompt = f"{prompt}\n\n## Project context\n\n{context}"
|
|
249
|
+
# ROADMAP_v2 §21b: durable memories, the third context tier beside
|
|
250
|
+
# the project's AGENTS.md and gated the same way (use_memory).
|
|
251
|
+
# Appended HERE and not inside with_catalogs, which feeds
|
|
252
|
+
# pass_prompt() -- injecting there would put every memory into all
|
|
253
|
+
# ten research passes, which is §19's K6 trap in its second
|
|
254
|
+
# instance.
|
|
255
|
+
memories = memory_manager.prompt_fragment(agent)
|
|
256
|
+
if memories:
|
|
257
|
+
prompt = f"{prompt}\n\n{memories}"
|
|
258
|
+
return prompt
|
|
259
|
+
|
|
260
|
+
|
|
261
|
+
manager = AgentManager()
|