@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.
- package/.env.example +19 -14
- package/CONFIG_ARCHITECTURE.md +2002 -0
- package/NOTICE +1 -1
- package/README.md +233 -94
- package/agents/builtin/build.md +73 -0
- package/agents/builtin/general.md +80 -0
- package/agents/builtin/plan.md +23 -5
- package/agents/builtin/test.md +65 -0
- package/agents/builtin/writer.md +67 -0
- package/agents/manager.py +261 -195
- package/agents/subagent_tool.py +277 -43
- package/agents/tui_commands.py +7 -12
- package/bin/venastine.mjs +528 -382
- package/bin/venastine.py +71 -0
- package/cli_commands.py +835 -0
- package/config.py +95 -1198
- package/config.yaml +1090 -0
- package/config_edit.py +1301 -0
- package/config_schema.py +966 -0
- package/config_update.py +749 -0
- package/core/agent_activity.py +158 -20
- package/core/ansi.py +452 -0
- package/core/approval.py +27 -5
- package/core/call_view.py +364 -0
- package/core/cancel.py +121 -0
- package/core/chatgpt_auth.py +747 -0
- package/core/claims_view.py +107 -0
- package/core/client.py +1238 -82
- package/core/commands.py +454 -0
- package/core/compaction.py +830 -66
- package/core/config_loader.py +516 -97
- package/core/connect.py +686 -0
- package/core/events.py +68 -6
- package/core/file_mentions.py +280 -0
- package/core/interaction.py +87 -4
- package/core/line_pattern.py +128 -0
- package/core/loop.py +1516 -440
- package/core/memory.py +120 -8
- package/core/pipeline_models.py +23 -9
- package/core/provider_errors.py +234 -0
- package/core/reasoning/confidence_scoring.py +70 -54
- package/core/reasoning/json_retry.py +19 -1
- package/core/reasoning/orchestrator.py +232 -53
- package/core/reasoning/output_writer.py +90 -31
- package/core/reasoning/payload_validation.py +133 -7
- package/core/reasoning/pipeline_storage.py +427 -361
- package/core/reasoning/review.py +659 -615
- package/core/reasoning/scholar.py +38 -14
- package/core/reasoning/source_corpus.py +110 -4
- package/core/replay.py +158 -26
- package/core/session.py +36 -12
- package/core/session_view.py +259 -0
- package/core/session_wake.py +259 -0
- package/core/shell_sessions.py +1553 -0
- package/core/workspace_trust.py +1 -1
- package/credentials.py +308 -180
- package/database.py +82 -1
- package/{PRIVACY.md → docs/PRIVACY.md} +19 -7
- package/docs/SECURITY.md +108 -0
- package/docs/THIRD_PARTY_NOTICES.md +148 -0
- package/env_secrets.py +118 -77
- package/json_store.py +236 -16
- package/logging_setup.py +121 -11
- package/main.py +1671 -199
- package/mcp_client/client.py +257 -23
- package/mcp_client/config.py +484 -379
- package/mcp_client/registration.py +163 -2
- package/memories/manager.py +15 -0
- package/memories/tui_commands.py +7 -12
- package/package.json +234 -4
- package/project_init/config_files.py +4 -0
- package/project_init/generator.py +201 -37
- package/project_init/manifest.py +418 -406
- package/project_init/tui_commands.py +7 -4
- package/prompts/revalidate.md +1 -1
- package/prompts/revise.md +17 -15
- package/prompts/system_prompts.py +64 -10
- package/providers.json.example +94 -92
- package/pyproject.toml +51 -2
- package/requirements.txt +252 -29
- package/safety/policy_enforcement.py +353 -17
- package/security/askpass.py +419 -0
- package/security/capability.py +37 -7
- package/security/confine.py +484 -0
- package/security/hosted.py +182 -0
- package/security/permissions.py +54 -7
- package/security/posture.py +124 -12
- package/security/protected_paths.py +198 -6
- package/security/sandbox.py +3262 -139
- package/security/secrets.py +855 -0
- package/skills/tui_commands.py +3 -4
- package/storage.py +368 -35
- package/tools/base.py +465 -305
- package/tools/builtin/_math_common.py +306 -16
- package/tools/builtin/_net_common.py +267 -61
- package/tools/builtin/arxiv.py +681 -223
- package/tools/builtin/ask_user.py +231 -213
- package/tools/builtin/discrete_math.py +35 -4
- package/tools/builtin/fetch_url.py +239 -21
- package/tools/builtin/file_ops.py +341 -33
- package/tools/builtin/geometry.py +184 -148
- package/tools/builtin/get_time.py +2 -2
- package/tools/builtin/linear_algebra.py +66 -23
- package/tools/builtin/load_skill.py +10 -0
- package/tools/builtin/logic.py +171 -12
- package/tools/builtin/pin.py +46 -12
- package/tools/builtin/probability_stats.py +224 -228
- package/tools/builtin/project_docs.py +80 -21
- package/tools/builtin/remember.py +21 -4
- package/tools/builtin/shell.py +586 -67
- package/tools/builtin/shell_sessions.py +613 -0
- package/tools/builtin/symbolic_math.py +80 -10
- package/tools/builtin/todo.py +5 -3
- package/tools/builtin/web_search.py +329 -151
- package/tools/context.py +23 -0
- package/tools/isolation.py +387 -314
- package/tools/registry.py +347 -30
- package/tui/app.py +4645 -438
- package/tui/app.tcss +185 -2
- package/tui/commands.py +10 -154
- package/tui/diffs.py +286 -265
- package/tui/hosted_driver.py +102 -0
- package/tui/markdown.py +60 -10
- package/tui/meters.py +103 -7
- package/tui/screens.py +346 -92
- package/tui/themes.py +59 -19
- package/tui/widgets.py +1575 -111
- package/SECURITY.md +0 -88
- package/THIRD_PARTY_NOTICES.md +0 -100
package/config.yaml
ADDED
|
@@ -0,0 +1,1090 @@
|
|
|
1
|
+
# Venastine harness configuration.
|
|
2
|
+
#
|
|
3
|
+
# Every tunable value the harness has. `config.py` reads this file once at
|
|
4
|
+
# import through `config_schema.py` and publishes the values as module
|
|
5
|
+
# attributes, so the names here are the lower-cased form of the constants the
|
|
6
|
+
# code uses: `max_tokens` is `config.MAX_TOKENS`.
|
|
7
|
+
#
|
|
8
|
+
# WHY THE HISTORY IS NOT IN HERE. Each of these numbers was argued, several
|
|
9
|
+
# were reversed once, and the arguments run to hundreds of lines -- they live
|
|
10
|
+
# in CONFIG_ARCHITECTURE.md, section by section, under the name used here.
|
|
11
|
+
# Read it before changing a value. The comments below say what a value DOES;
|
|
12
|
+
# that document says why it is what it is.
|
|
13
|
+
#
|
|
14
|
+
# THREE THINGS TO KNOW BEFORE EDITING.
|
|
15
|
+
#
|
|
16
|
+
# 1. This file ships with the harness and has exactly ONE location, beside
|
|
17
|
+
# config.py. There is no user-level or project-level copy and no
|
|
18
|
+
# environment variable that redirects it. That absence is deliberate and
|
|
19
|
+
# is a security property rather than an oversight: settings.json refuses
|
|
20
|
+
# shell_approval_mode, ensemble_models, critic_model and embedder_model by
|
|
21
|
+
# name precisely because a project's settings.json beats the user's and
|
|
22
|
+
# arrives with a directory you cloned.
|
|
23
|
+
#
|
|
24
|
+
# 2. Keys marked AUTHORITY decide what this harness may do to its host, or
|
|
25
|
+
# name a provider it sends your research content to. Nothing UNATTENDED
|
|
26
|
+
# can change them -- not a tool call, not a settings file, not an
|
|
27
|
+
# environment variable. This file is the only route and it needs a
|
|
28
|
+
# human: `/config` writes one, but only after saying what the key
|
|
29
|
+
# permits and asking (batch 84).
|
|
30
|
+
#
|
|
31
|
+
# 3. Values are read ONCE, at startup. Editing this file while the harness is
|
|
32
|
+
# running changes nothing until the next launch.
|
|
33
|
+
#
|
|
34
|
+
# A missing key, an unknown key, a duplicated key or a value of the wrong
|
|
35
|
+
# type is a startup error naming the key. There are no defaults behind this
|
|
36
|
+
# file.
|
|
37
|
+
|
|
38
|
+
# ======================================================================
|
|
39
|
+
# Model and loop
|
|
40
|
+
# ======================================================================
|
|
41
|
+
|
|
42
|
+
# The model every path uses when nothing more specific speaks.
|
|
43
|
+
# --model > settings.json default_model > this. AGENT_MODEL overrides.
|
|
44
|
+
model_name: claude-sonnet-5
|
|
45
|
+
|
|
46
|
+
# Per-call OUTPUT ceiling, sent as max_tokens / max_output_tokens /
|
|
47
|
+
# max_completion_tokens. On current Anthropic models it bounds thinking
|
|
48
|
+
# and response text together, so effort spends from the same budget.
|
|
49
|
+
max_tokens: 16000
|
|
50
|
+
|
|
51
|
+
# The default step ceiling: every chat turn, every research pass, and any
|
|
52
|
+
# agent whose frontmatter names no max_steps or names an invalid one.
|
|
53
|
+
max_iterations: 50
|
|
54
|
+
|
|
55
|
+
# How many more times ONE model call is attempted when the provider fails in
|
|
56
|
+
# a way that can pass: a dropped connection, a 408/409/429/5xx, or an error
|
|
57
|
+
# sent inside a stream that had already opened. Total attempts per call is
|
|
58
|
+
# this plus one. The openai and anthropic SDKs still retry a request that
|
|
59
|
+
# fails before its stream opens, inside each attempt. 0 turns this off.
|
|
60
|
+
model_call_max_retries: 2
|
|
61
|
+
|
|
62
|
+
# Seconds before the first of those retries. Each later retry waits twice as
|
|
63
|
+
# long, with up to 25% jitter either way; a server's retry-after can make a
|
|
64
|
+
# wait longer, never shorter, capped at 60s.
|
|
65
|
+
model_call_retry_base_delay_s: 3.0
|
|
66
|
+
|
|
67
|
+
# ======================================================================
|
|
68
|
+
# Subagents
|
|
69
|
+
# ======================================================================
|
|
70
|
+
|
|
71
|
+
# How deep spawn_subagent may nest, checked against
|
|
72
|
+
# ToolContext.subagent_depth.
|
|
73
|
+
subagent_max_depth: 2
|
|
74
|
+
|
|
75
|
+
# How many calls of one response may run at once, for tools whose ToolSpec
|
|
76
|
+
# declares `parallel`. A ceiling, not a target: more calls run in waves of
|
|
77
|
+
# this size. 1 makes the parallel path behave exactly like the sequential
|
|
78
|
+
# one, which is the escape hatch if concurrency is ever suspect.
|
|
79
|
+
subagent_max_parallel: 3
|
|
80
|
+
|
|
81
|
+
# ======================================================================
|
|
82
|
+
# Deep research pipeline
|
|
83
|
+
# ======================================================================
|
|
84
|
+
|
|
85
|
+
# Revise/re-validate loop iterations per claim before the pipeline falls
|
|
86
|
+
# back.
|
|
87
|
+
max_pipeline_retries: 2
|
|
88
|
+
|
|
89
|
+
# Corrective follow-up attempts when a pass returns malformed JSON.
|
|
90
|
+
# Total attempts per pass is this plus one.
|
|
91
|
+
max_json_retries: 2
|
|
92
|
+
|
|
93
|
+
# ======================================================================
|
|
94
|
+
# Ensemble mode
|
|
95
|
+
# ======================================================================
|
|
96
|
+
|
|
97
|
+
# Run Pass 1 once per `ensemble_models` entry. A settings.json
|
|
98
|
+
# `ensemble_mode` outranks this; the roster below has no settings key.
|
|
99
|
+
ensemble_mode: false
|
|
100
|
+
|
|
101
|
+
# AUTHORITY -- no settings.json key exists at any tier, because a roster chooses
|
|
102
|
+
# which providers N research passes call. N is its length and is never
|
|
103
|
+
# configured separately. Fewer than two DISTINCT (provider_name, model)
|
|
104
|
+
# pairs is refused: one model cannot disagree with itself.
|
|
105
|
+
# null, with ensemble_mode false, is the shipped configuration. Example:
|
|
106
|
+
# ensemble_models:
|
|
107
|
+
# - {provider_name: ANTHROPIC, model: claude-opus-5}
|
|
108
|
+
# - {provider_name: OPENAI, model: gpt-5.1}
|
|
109
|
+
ensemble_models: null
|
|
110
|
+
|
|
111
|
+
# ======================================================================
|
|
112
|
+
# Sampling-parameter support
|
|
113
|
+
# ======================================================================
|
|
114
|
+
|
|
115
|
+
# Models that reject temperature / top_p / top_k outright, or ignore them.
|
|
116
|
+
# core/client.py omits those parameters entirely for anything listed, with a
|
|
117
|
+
# WARNING. Steering on these happens through prompting and
|
|
118
|
+
# model_effort_levels instead.
|
|
119
|
+
#
|
|
120
|
+
# The two gemini names IGNORE the parameters (ROADMAP_v3 §87): no error and
|
|
121
|
+
# no effect. Google says every later Gemini will do the same, then refuse
|
|
122
|
+
# them with a 400, so a new Gemini model belongs here when it ships. OpenAI's
|
|
123
|
+
# reasoning models are not listed: core/client.py drops the parameters on
|
|
124
|
+
# OPENAI whenever a reasoning effort other than "none" is sent, which is
|
|
125
|
+
# OpenAI's own rule and needs no model names.
|
|
126
|
+
#
|
|
127
|
+
# On ANTHROPIC every model drops them since batch 234 (API version
|
|
128
|
+
# 2026-10-21 refuses them on all of them), so the claude names below matter
|
|
129
|
+
# for Claude reached through an OpenAI-compatible gateway, by its bare name.
|
|
130
|
+
models_rejecting_sampling_params:
|
|
131
|
+
- claude-fable-5
|
|
132
|
+
- claude-fable-5-1
|
|
133
|
+
- claude-haiku-5-5
|
|
134
|
+
- claude-mythos-5
|
|
135
|
+
- claude-mythos-5-1
|
|
136
|
+
- claude-opus-4-7
|
|
137
|
+
- claude-opus-4-8
|
|
138
|
+
- claude-opus-5
|
|
139
|
+
- claude-opus-5-5
|
|
140
|
+
- claude-sonnet-5
|
|
141
|
+
- claude-sonnet-5-5
|
|
142
|
+
- gemini-3.5-flash-lite
|
|
143
|
+
- gemini-3.6-flash
|
|
144
|
+
|
|
145
|
+
# ======================================================================
|
|
146
|
+
# Critic and embedder routing
|
|
147
|
+
# ======================================================================
|
|
148
|
+
|
|
149
|
+
# AUTHORITY -- names a provider every claim is sent to. `/critic` writes a
|
|
150
|
+
# user-tier store that OUTRANKS this, and settings.json rejects the key by
|
|
151
|
+
# name at every tier. null means no special routing.
|
|
152
|
+
critic_model: null
|
|
153
|
+
|
|
154
|
+
# AUTHORITY -- names a provider that receives claim text and the text of every
|
|
155
|
+
# page a run fetched. `/embedder` outranks this. null means no special
|
|
156
|
+
# routing.
|
|
157
|
+
embedder_model: null
|
|
158
|
+
|
|
159
|
+
# Per-embedder similarity floor and ceiling, matched by model-name PREFIX
|
|
160
|
+
# (the longest matching prefix wins).
|
|
161
|
+
similarity_calibration:
|
|
162
|
+
text-embedding-3:
|
|
163
|
+
floor: 0.1
|
|
164
|
+
ceiling: 0.65
|
|
165
|
+
text-embedding-ada:
|
|
166
|
+
floor: 0.7
|
|
167
|
+
ceiling: 0.92
|
|
168
|
+
e5-:
|
|
169
|
+
floor: 0.7
|
|
170
|
+
ceiling: 0.92
|
|
171
|
+
multilingual-e5:
|
|
172
|
+
floor: 0.7
|
|
173
|
+
ceiling: 0.92
|
|
174
|
+
bge-:
|
|
175
|
+
floor: 0.55
|
|
176
|
+
ceiling: 0.85
|
|
177
|
+
gte-:
|
|
178
|
+
floor: 0.55
|
|
179
|
+
ceiling: 0.88
|
|
180
|
+
nomic-embed:
|
|
181
|
+
floor: 0.4
|
|
182
|
+
ceiling: 0.8
|
|
183
|
+
embed-english:
|
|
184
|
+
floor: 0.25
|
|
185
|
+
ceiling: 0.8
|
|
186
|
+
voyage:
|
|
187
|
+
floor: 0.35
|
|
188
|
+
ceiling: 0.85
|
|
189
|
+
gemini-embedding:
|
|
190
|
+
floor: 0.35
|
|
191
|
+
ceiling: 0.85
|
|
192
|
+
|
|
193
|
+
# For an embedder that matches no prefix above.
|
|
194
|
+
default_similarity_calibration:
|
|
195
|
+
floor: 0.25
|
|
196
|
+
ceiling: 0.85
|
|
197
|
+
|
|
198
|
+
# The query/passage prefixes asymmetric embedders need, matched by
|
|
199
|
+
# model-name prefix. An embedder absent from here is sent bare text.
|
|
200
|
+
embedder_prefixes:
|
|
201
|
+
e5-:
|
|
202
|
+
query: 'query: '
|
|
203
|
+
passage: 'passage: '
|
|
204
|
+
multilingual-e5:
|
|
205
|
+
query: 'query: '
|
|
206
|
+
passage: 'passage: '
|
|
207
|
+
bge-:
|
|
208
|
+
query: 'Represent this sentence for searching relevant passages: '
|
|
209
|
+
passage: ''
|
|
210
|
+
gte-:
|
|
211
|
+
query: ''
|
|
212
|
+
passage: ''
|
|
213
|
+
nomic-embed:
|
|
214
|
+
query: 'search_query: '
|
|
215
|
+
passage: 'search_document: '
|
|
216
|
+
gemini-embedding:
|
|
217
|
+
query_task_type: RETRIEVAL_QUERY
|
|
218
|
+
passage_task_type: RETRIEVAL_DOCUMENT
|
|
219
|
+
text-embedding-004:
|
|
220
|
+
query_task_type: RETRIEVAL_QUERY
|
|
221
|
+
passage_task_type: RETRIEVAL_DOCUMENT
|
|
222
|
+
|
|
223
|
+
# ======================================================================
|
|
224
|
+
# Source authority: domain classes
|
|
225
|
+
# ======================================================================
|
|
226
|
+
|
|
227
|
+
# Class name -> authority score. The second table, below, maps hosts onto
|
|
228
|
+
# these names, so a score is changed in one place.
|
|
229
|
+
domain_authority_classes:
|
|
230
|
+
restricted_registry: 0.9 # .gov .mil .int .edu, gov.uk, ac.uk...
|
|
231
|
+
intergovernmental: 0.9 # WHO, UN, World Bank, OECD, EU
|
|
232
|
+
standards_body: 0.88 # IETF, W3C, ISO, NIST, IEC
|
|
233
|
+
peer_reviewed_publisher: 0.85
|
|
234
|
+
established_news: 0.65
|
|
235
|
+
preprint_server: 0.6 # not peer reviewed; §45 SQ5 refines this
|
|
236
|
+
reference_tertiary: 0.55 # encyclopaedias -- a pointer to a source
|
|
237
|
+
generic_org: 0.55 # an unrestricted registry, small bump
|
|
238
|
+
generic_commercial: 0.45
|
|
239
|
+
blog_platform: 0.3
|
|
240
|
+
forum_qa: 0.25
|
|
241
|
+
social_media: 0.15
|
|
242
|
+
content_farm: 0.1
|
|
243
|
+
|
|
244
|
+
# The score for a host that matches no suffix below; its class reads
|
|
245
|
+
# "unknown".
|
|
246
|
+
default_domain_authority: 0.45
|
|
247
|
+
|
|
248
|
+
# Host suffix -> a class name from the table above. The `<kind>.<cc>`
|
|
249
|
+
# second-level forms are derived in source_scoring.py, not listed here.
|
|
250
|
+
domain_authority_suffixes:
|
|
251
|
+
|
|
252
|
+
# Restricted registries. The bare TLDs; the `<kind>.<cc>`
|
|
253
|
+
# second-level forms are derived in source_scoring.py rather than
|
|
254
|
+
# listed.
|
|
255
|
+
gov: restricted_registry
|
|
256
|
+
mil: restricted_registry
|
|
257
|
+
int: restricted_registry
|
|
258
|
+
edu: restricted_registry
|
|
259
|
+
|
|
260
|
+
# Intergovernmental bodies, whose hosts are ordinary .org/.int
|
|
261
|
+
# names.
|
|
262
|
+
who.int: intergovernmental
|
|
263
|
+
un.org: intergovernmental
|
|
264
|
+
europa.eu: intergovernmental
|
|
265
|
+
oecd.org: intergovernmental
|
|
266
|
+
imf.org: intergovernmental
|
|
267
|
+
worldbank.org: intergovernmental
|
|
268
|
+
iaea.org: intergovernmental
|
|
269
|
+
|
|
270
|
+
# Standards bodies.
|
|
271
|
+
ietf.org: standards_body
|
|
272
|
+
rfc-editor.org: standards_body
|
|
273
|
+
w3.org: standards_body
|
|
274
|
+
iso.org: standards_body
|
|
275
|
+
iec.ch: standards_body
|
|
276
|
+
ieee.org: standards_body
|
|
277
|
+
unicode.org: standards_body
|
|
278
|
+
ecma-international.org: standards_body
|
|
279
|
+
|
|
280
|
+
# Peer-reviewed publishers and indexes.
|
|
281
|
+
nature.com: peer_reviewed_publisher
|
|
282
|
+
science.org: peer_reviewed_publisher
|
|
283
|
+
sciencemag.org: peer_reviewed_publisher
|
|
284
|
+
nejm.org: peer_reviewed_publisher
|
|
285
|
+
thelancet.com: peer_reviewed_publisher
|
|
286
|
+
bmj.com: peer_reviewed_publisher
|
|
287
|
+
cell.com: peer_reviewed_publisher
|
|
288
|
+
pnas.org: peer_reviewed_publisher
|
|
289
|
+
sciencedirect.com: peer_reviewed_publisher
|
|
290
|
+
springer.com: peer_reviewed_publisher
|
|
291
|
+
link.springer.com: peer_reviewed_publisher
|
|
292
|
+
onlinelibrary.wiley.com: peer_reviewed_publisher
|
|
293
|
+
wiley.com: peer_reviewed_publisher
|
|
294
|
+
tandfonline.com: peer_reviewed_publisher
|
|
295
|
+
sagepub.com: peer_reviewed_publisher
|
|
296
|
+
acm.org: peer_reviewed_publisher
|
|
297
|
+
aps.org: peer_reviewed_publisher
|
|
298
|
+
acs.org: peer_reviewed_publisher
|
|
299
|
+
plos.org: peer_reviewed_publisher
|
|
300
|
+
frontiersin.org: peer_reviewed_publisher
|
|
301
|
+
mdpi.com: peer_reviewed_publisher
|
|
302
|
+
jstor.org: peer_reviewed_publisher
|
|
303
|
+
pubmed.ncbi.nlm.nih.gov: peer_reviewed_publisher
|
|
304
|
+
|
|
305
|
+
# Preprint servers and repositories -- not peer reviewed.
|
|
306
|
+
arxiv.org: preprint_server
|
|
307
|
+
biorxiv.org: preprint_server
|
|
308
|
+
medrxiv.org: preprint_server
|
|
309
|
+
chemrxiv.org: preprint_server
|
|
310
|
+
ssrn.com: preprint_server
|
|
311
|
+
papers.ssrn.com: preprint_server
|
|
312
|
+
osf.io: preprint_server
|
|
313
|
+
researchsquare.com: preprint_server
|
|
314
|
+
semanticscholar.org: preprint_server
|
|
315
|
+
|
|
316
|
+
# News organisations with a correction policy and a masthead.
|
|
317
|
+
# An ALLOWLIST, because "news site" is not a property of a
|
|
318
|
+
# domain suffix and the alternative is scoring every .com that
|
|
319
|
+
# publishes articles.
|
|
320
|
+
reuters.com: established_news
|
|
321
|
+
apnews.com: established_news
|
|
322
|
+
bbc.com: established_news
|
|
323
|
+
bbc.co.uk: established_news
|
|
324
|
+
ft.com: established_news
|
|
325
|
+
economist.com: established_news
|
|
326
|
+
nytimes.com: established_news
|
|
327
|
+
washingtonpost.com: established_news
|
|
328
|
+
wsj.com: established_news
|
|
329
|
+
theguardian.com: established_news
|
|
330
|
+
npr.org: established_news
|
|
331
|
+
pbs.org: established_news
|
|
332
|
+
bloomberg.com: established_news
|
|
333
|
+
nikkei.com: established_news
|
|
334
|
+
afp.com: established_news
|
|
335
|
+
dw.com: established_news
|
|
336
|
+
|
|
337
|
+
# Tertiary references -- they summarise sources rather than
|
|
338
|
+
# being one.
|
|
339
|
+
wikipedia.org: reference_tertiary
|
|
340
|
+
wikimedia.org: reference_tertiary
|
|
341
|
+
wiktionary.org: reference_tertiary
|
|
342
|
+
britannica.com: reference_tertiary
|
|
343
|
+
org: generic_org
|
|
344
|
+
|
|
345
|
+
# Blog and newsletter platforms: the host says nothing about
|
|
346
|
+
# the author.
|
|
347
|
+
medium.com: blog_platform
|
|
348
|
+
substack.com: blog_platform
|
|
349
|
+
wordpress.com: blog_platform
|
|
350
|
+
blogspot.com: blog_platform
|
|
351
|
+
blogger.com: blog_platform
|
|
352
|
+
tumblr.com: blog_platform
|
|
353
|
+
wixsite.com: blog_platform
|
|
354
|
+
ghost.io: blog_platform
|
|
355
|
+
reddit.com: forum_qa
|
|
356
|
+
quora.com: forum_qa
|
|
357
|
+
|
|
358
|
+
# Forums and Q&A. Often correct, never attributable.
|
|
359
|
+
stackexchange.com: forum_qa
|
|
360
|
+
stackoverflow.com: forum_qa
|
|
361
|
+
news.ycombinator.com: forum_qa
|
|
362
|
+
4chan.org: content_farm
|
|
363
|
+
|
|
364
|
+
# Social media. Speculation and unattributed claims travel here
|
|
365
|
+
# fastest, which is the whole reason a claim needs grounding
|
|
366
|
+
# elsewhere.
|
|
367
|
+
x.com: social_media
|
|
368
|
+
twitter.com: social_media
|
|
369
|
+
facebook.com: social_media
|
|
370
|
+
instagram.com: social_media
|
|
371
|
+
tiktok.com: social_media
|
|
372
|
+
threads.net: social_media
|
|
373
|
+
linkedin.com: social_media
|
|
374
|
+
mastodon.social: social_media
|
|
375
|
+
bsky.app: social_media
|
|
376
|
+
youtube.com: social_media
|
|
377
|
+
vk.com: social_media
|
|
378
|
+
weibo.com: social_media
|
|
379
|
+
t.me: social_media
|
|
380
|
+
|
|
381
|
+
# Aggregators that republish without attribution, and image
|
|
382
|
+
# boards with no text to ground anything in. The measured
|
|
383
|
+
# case: a Pinterest pin scored similarity 1.0 on a Nobel
|
|
384
|
+
# Prize claim.
|
|
385
|
+
pinterest.com: content_farm
|
|
386
|
+
answers.com: content_farm
|
|
387
|
+
ehow.com: content_farm
|
|
388
|
+
scribd.com: content_farm
|
|
389
|
+
coursehero.com: content_farm
|
|
390
|
+
chegg.com: content_farm
|
|
391
|
+
|
|
392
|
+
# How far a stated reason may move a source's authority score, in either
|
|
393
|
+
# direction.
|
|
394
|
+
authority_adjustment_cap: 0.15
|
|
395
|
+
|
|
396
|
+
# ======================================================================
|
|
397
|
+
# The sources/ artifact
|
|
398
|
+
# ======================================================================
|
|
399
|
+
|
|
400
|
+
# Whether the sources/ artifact keeps the fetched page text alongside the
|
|
401
|
+
# scores, or the scores alone.
|
|
402
|
+
persist_source_text: true
|
|
403
|
+
|
|
404
|
+
# ======================================================================
|
|
405
|
+
# Scholarly authority (OpenAlex)
|
|
406
|
+
# ======================================================================
|
|
407
|
+
|
|
408
|
+
# Whether the OpenAlex lookup runs at all. Off ships no outbound
|
|
409
|
+
# scholarly requests.
|
|
410
|
+
scholar_lookup: false
|
|
411
|
+
|
|
412
|
+
# The OpenAlex base URL.
|
|
413
|
+
scholar_api_url: https://api.openalex.org
|
|
414
|
+
|
|
415
|
+
# Per-request timeout, in seconds.
|
|
416
|
+
scholar_timeout_s: 8.0
|
|
417
|
+
|
|
418
|
+
# Retries per request; total attempts is this plus one.
|
|
419
|
+
scholar_max_retries: 2
|
|
420
|
+
|
|
421
|
+
# How long a lookup is cached in-process, in seconds.
|
|
422
|
+
scholar_cache_ttl_s: 900
|
|
423
|
+
|
|
424
|
+
# Sent as OpenAlex's `mailto` parameter when non-empty, which is their
|
|
425
|
+
# polite-pool convention. Blank sends no address and is the shipped value:
|
|
426
|
+
# an address here leaves the machine on every lookup.
|
|
427
|
+
scholar_mailto: ''
|
|
428
|
+
|
|
429
|
+
# The three weights below split one score and a settings.json
|
|
430
|
+
# `source_scoring` block can override each. They are expected to sum to 1.
|
|
431
|
+
scholar_venue_weight: 0.4
|
|
432
|
+
scholar_citation_weight: 0.4
|
|
433
|
+
scholar_author_weight: 0.2
|
|
434
|
+
|
|
435
|
+
# OpenAlex location type -> venue credit. `unknown` is the fallback for an
|
|
436
|
+
# unrecognised type, and a work with no location at all takes
|
|
437
|
+
# `repository`.
|
|
438
|
+
scholar_venue_credit:
|
|
439
|
+
journal: 1.0
|
|
440
|
+
conference: 1.0
|
|
441
|
+
book series: 0.85
|
|
442
|
+
repository: 0.55 # arXiv, bioRxiv, institutional repositories
|
|
443
|
+
unknown: 0.6
|
|
444
|
+
|
|
445
|
+
# The h-index at which author credit saturates.
|
|
446
|
+
scholar_h_saturation: 60
|
|
447
|
+
|
|
448
|
+
# Below this many works in a field-year cohort, the citation percentile is
|
|
449
|
+
# not trusted.
|
|
450
|
+
scholar_min_cohort_size: 200
|
|
451
|
+
|
|
452
|
+
# A work younger than this has not had time to accrue citations, so the
|
|
453
|
+
# citation term is not held against it.
|
|
454
|
+
scholar_min_citation_age_days: 60
|
|
455
|
+
|
|
456
|
+
# The authority a retracted work is given outright, whatever else it
|
|
457
|
+
# scores.
|
|
458
|
+
scholar_retracted_authority: 0.05
|
|
459
|
+
|
|
460
|
+
# ======================================================================
|
|
461
|
+
# Reasoning effort
|
|
462
|
+
# ======================================================================
|
|
463
|
+
|
|
464
|
+
# Reasoning effort when nothing more specific speaks: --effort >
|
|
465
|
+
# settings.json `effort` > this, with settings.json `tui.effort` winning
|
|
466
|
+
# inside the TUI. null sends no effort at all.
|
|
467
|
+
default_effort: high
|
|
468
|
+
|
|
469
|
+
# Known effort levels per model, for providers whose API exposes no
|
|
470
|
+
# capability query. An EMPTY list means the model accepts none, which is
|
|
471
|
+
# different from being absent -- absent falls through to the list below.
|
|
472
|
+
model_effort_levels:
|
|
473
|
+
gpt-4o: []
|
|
474
|
+
gpt-4o-mini: []
|
|
475
|
+
gpt-4.1: []
|
|
476
|
+
gpt-4.1-mini: []
|
|
477
|
+
gpt-4.1-nano: []
|
|
478
|
+
gpt-3.5-turbo: []
|
|
479
|
+
|
|
480
|
+
# Used for a model with no entry above.
|
|
481
|
+
default_effort_levels:
|
|
482
|
+
- low
|
|
483
|
+
- medium
|
|
484
|
+
- high
|
|
485
|
+
- xhigh
|
|
486
|
+
- max
|
|
487
|
+
- ultra
|
|
488
|
+
|
|
489
|
+
# Effort level -> Google thinking_budget in tokens. -1 hands the choice to
|
|
490
|
+
# the model.
|
|
491
|
+
google_thinking_budgets:
|
|
492
|
+
low: 2048
|
|
493
|
+
medium: 8192
|
|
494
|
+
high: 16384
|
|
495
|
+
xhigh: 24576
|
|
496
|
+
ultra: 32768
|
|
497
|
+
max: -1
|
|
498
|
+
|
|
499
|
+
# ======================================================================
|
|
500
|
+
# Database
|
|
501
|
+
# ======================================================================
|
|
502
|
+
|
|
503
|
+
# The SQLite file behind the conversation database. APP_DB_PATH overrides;
|
|
504
|
+
# the npm launcher points it at ~/.config/venastine so state is global
|
|
505
|
+
# while artifacts stay local.
|
|
506
|
+
db_path: app.db
|
|
507
|
+
|
|
508
|
+
# ======================================================================
|
|
509
|
+
# File-ops workspace
|
|
510
|
+
# ======================================================================
|
|
511
|
+
|
|
512
|
+
# The file tools' read-write root, and a PERMISSION BOUNDARY: file_ops
|
|
513
|
+
# auto-approves writes inside it. AGENT_WORKSPACE overrides, and naming it
|
|
514
|
+
# also makes the resolved workspace the project path -- so pointing it at a
|
|
515
|
+
# home directory would auto-approve writes there for every project. Writing
|
|
516
|
+
# the path HERE moves the workspace alone; the key below moves the project
|
|
517
|
+
# with it.
|
|
518
|
+
workspace_dir: ./workspace
|
|
519
|
+
|
|
520
|
+
# Whether that workspace is also the PROJECT: what workspace trust asks about,
|
|
521
|
+
# which .venastine/ supplies settings and agents, where /init scaffolds, and
|
|
522
|
+
# what project-scoped memories belong to. Naming AGENT_WORKSPACE has always
|
|
523
|
+
# done this; the key is the same statement made in the file, for the reader
|
|
524
|
+
# who has no variable to set. Either route suffices, and neither is needed to
|
|
525
|
+
# work inside a project you launched from -- that is already the project.
|
|
526
|
+
workspace_is_project: false
|
|
527
|
+
|
|
528
|
+
# Hard reject before opening, in both directions (25 MB).
|
|
529
|
+
max_file_size_bytes: 25000000
|
|
530
|
+
|
|
531
|
+
# Per-read line ceiling. A larger `count` is clamped, with a note in the
|
|
532
|
+
# result rather than silently.
|
|
533
|
+
max_read_lines: 500
|
|
534
|
+
|
|
535
|
+
# Per-read character ceiling, the truncation point for sandboxed stdout and
|
|
536
|
+
# stderr, and the bound on an MCP tool's whole result.
|
|
537
|
+
max_read_chars: 50000
|
|
538
|
+
|
|
539
|
+
# ROADMAP_v3 §81 (IM7). The largest image `view_image` attaches, in bytes of
|
|
540
|
+
# the file (3.75 MiB). Base64 inflates it by 4/3, so the encoded payload stays
|
|
541
|
+
# under 5 MiB -- the smallest per-image limit of the wires that carry it --
|
|
542
|
+
# and the image is resent with every request for as long as it stays in view.
|
|
543
|
+
max_image_bytes: 3932160
|
|
544
|
+
|
|
545
|
+
# ======================================================================
|
|
546
|
+
# /init
|
|
547
|
+
# ======================================================================
|
|
548
|
+
|
|
549
|
+
# The agent /init runs.
|
|
550
|
+
initializer_agent: initializer
|
|
551
|
+
|
|
552
|
+
# Most characters read_project_doc returns per file.
|
|
553
|
+
init_read_chars: 20000
|
|
554
|
+
|
|
555
|
+
# Step ceiling for the /init run when its agent names none.
|
|
556
|
+
init_max_steps: 12
|
|
557
|
+
|
|
558
|
+
# Largest existing AGENTS.md /init will revise, in characters. Above it
|
|
559
|
+
# /init stops before any model call: the reply replaces the whole file.
|
|
560
|
+
init_max_hub_chars: 1000000
|
|
561
|
+
|
|
562
|
+
# Most lines of the proposed diff /init prints before asking.
|
|
563
|
+
init_diff_shown_lines: 2000
|
|
564
|
+
|
|
565
|
+
# ======================================================================
|
|
566
|
+
# MCP teardown
|
|
567
|
+
# ======================================================================
|
|
568
|
+
|
|
569
|
+
# Seconds allowed for MCP session teardown at shutdown before the harness
|
|
570
|
+
# stops waiting.
|
|
571
|
+
teardown_budget_s: 10.0
|
|
572
|
+
|
|
573
|
+
# ======================================================================
|
|
574
|
+
# MCP results
|
|
575
|
+
# ======================================================================
|
|
576
|
+
|
|
577
|
+
# Largest single server-sent event, in bytes, a Streamable HTTP ("http")
|
|
578
|
+
# MCP server may stream back. A bigger one fails that tool call with an
|
|
579
|
+
# error naming the limit; the session survives. An answer sent as plain
|
|
580
|
+
# JSON is not an event and is not bounded by this. Legacy "sse" servers
|
|
581
|
+
# keep the MCP SDK's own 1 MiB, which it gives no way to change. 16 MiB.
|
|
582
|
+
# What reaches the model is bounded separately, by max_read_chars.
|
|
583
|
+
mcp_max_event_bytes: 16777216
|
|
584
|
+
|
|
585
|
+
# ======================================================================
|
|
586
|
+
# Shell and sandbox
|
|
587
|
+
# ======================================================================
|
|
588
|
+
|
|
589
|
+
# The shell the sandbox runs commands through. EMPTY means auto-detect,
|
|
590
|
+
# which is the shipped value. AGENT_SHELL overrides.
|
|
591
|
+
shell_binary: ''
|
|
592
|
+
|
|
593
|
+
# AUTHORITY -- part of the frozen security posture. Enables the weak
|
|
594
|
+
# host-subprocess backend, which is unbounded code on the host. Bound once
|
|
595
|
+
# at first read and frozen, so no tool call, settings file, environment
|
|
596
|
+
# variable or slash command can move it.
|
|
597
|
+
allow_insecure_sandbox_fallback: false
|
|
598
|
+
|
|
599
|
+
# AUTHORITY -- runs that fallback with no per-run prompt. The pair above and
|
|
600
|
+
# here reads as worse than either alone. Honoured under `tiered` and
|
|
601
|
+
# `contained`; under `always` every command still asks, this one included.
|
|
602
|
+
auto_approve_sandbox_fallback: false
|
|
603
|
+
|
|
604
|
+
# AUTHORITY -- part of the frozen security posture. Lets the agent ask
|
|
605
|
+
# for a command to run in WSL (ROADMAP_v3 §49, slice 3). It is NOT a
|
|
606
|
+
# weaker container: a command there reaches this harness's own
|
|
607
|
+
# config.yaml and ~/.config, gets no memory, CPU or process limit, and
|
|
608
|
+
# can start Windows programs with your full authority -- Windows
|
|
609
|
+
# interop is a kernel feature there and survives an emptied
|
|
610
|
+
# environment. Only a read-only command whose every argument is inside
|
|
611
|
+
# the workspace runs there without asking; everything else meets a
|
|
612
|
+
# prompt under `tiered` and `contained` alike. Shown on the startup
|
|
613
|
+
# banner and the TUI badge whenever it is on.
|
|
614
|
+
allow_wsl_backend: false
|
|
615
|
+
|
|
616
|
+
# Which WSL distribution the backend runs in. EMPTY means the first one
|
|
617
|
+
# `wsl.exe -l -q` lists, which is WSL's own default. A name that is not
|
|
618
|
+
# installed is refused with the installed names, rather than passed to
|
|
619
|
+
# `wsl.exe` -- which answers an unknown distro by exiting 4294967295
|
|
620
|
+
# and writing its complaint to stdout in UTF-16, so the failure would
|
|
621
|
+
# read as the command's rather than the configuration's.
|
|
622
|
+
wsl_distro: ''
|
|
623
|
+
|
|
624
|
+
# AUTHORITY -- part of the frozen security posture. Lets the agent ask
|
|
625
|
+
# for a command to run on the remote host configured below (ROADMAP_v3
|
|
626
|
+
# §49, slice 4). A remote host is not a sandbox and not a weaker one: it
|
|
627
|
+
# is a machine this harness cannot look at. It gets no memory, CPU or
|
|
628
|
+
# process limit -- those belong to the remote -- and the credentials it
|
|
629
|
+
# uses are ones you issued, not ones this harness controls.
|
|
630
|
+
# NOTHING runs there without asking you first, not even a read-only
|
|
631
|
+
# command: "inside the workspace" is a question about a filesystem this
|
|
632
|
+
# process cannot resolve, so it is not asked. Shown on the startup
|
|
633
|
+
# banner and the TUI badge whenever it is on.
|
|
634
|
+
allow_ssh_backend: false
|
|
635
|
+
|
|
636
|
+
# The one remote host the SSH backend reaches. The agent asks for
|
|
637
|
+
# `backend: "ssh"` and gets THIS host; it never names a host itself,
|
|
638
|
+
# for the reason `wsl_distro` is not the agent's either -- which machine
|
|
639
|
+
# this harness may reach is your question, and which command needs it is
|
|
640
|
+
# the agent's.
|
|
641
|
+
ssh_host: ''
|
|
642
|
+
ssh_user: ''
|
|
643
|
+
ssh_port: 22
|
|
644
|
+
|
|
645
|
+
# The private key to authenticate with. EMPTY offers whatever a running
|
|
646
|
+
# ssh-agent holds. This harness NEVER asks for, holds or stores a
|
|
647
|
+
# passphrase: if the key is encrypted, load it into an agent with
|
|
648
|
+
# `ssh-add` first. Measured, and the reason the backend passes
|
|
649
|
+
# BatchMode=yes: without it, an encrypted key with no agent makes ssh
|
|
650
|
+
# wait forever for a passphrase on a terminal this process does not own.
|
|
651
|
+
ssh_identity_file: ''
|
|
652
|
+
|
|
653
|
+
# Where a command runs on the remote. REQUIRED when the backend is on,
|
|
654
|
+
# and deliberately not defaulted to the login home: a command that runs
|
|
655
|
+
# somewhere nobody named is the failure this backend is most able to
|
|
656
|
+
# cause, and nothing here can see where it landed.
|
|
657
|
+
ssh_remote_workspace: ''
|
|
658
|
+
|
|
659
|
+
# The remote's host key, PINNED. Run `ssh-keyscan -p <port> <host>` and
|
|
660
|
+
# paste its output here (the `#` banner line is ignored). A host whose
|
|
661
|
+
# key does not match this is refused -- not prompted about, because a
|
|
662
|
+
# prompt on first use is a question a headless run has nobody to ask,
|
|
663
|
+
# and because a key that CHANGES is exactly the case worth refusing.
|
|
664
|
+
ssh_host_key: ''
|
|
665
|
+
|
|
666
|
+
# Which `ssh` program to run. EMPTY prefers the operating system's own
|
|
667
|
+
# (%SYSTEMROOT%\System32\OpenSSH\ssh.exe on Windows) and falls back to
|
|
668
|
+
# PATH. Not PATH first: on the machine this was built, PATH resolves to
|
|
669
|
+
# a git-for-Windows build while the OS ships its own, and the two report
|
|
670
|
+
# different text for the same failure.
|
|
671
|
+
ssh_binary: ''
|
|
672
|
+
|
|
673
|
+
# AUTHORITY -- THE gate for shell approval, and part of the frozen posture.
|
|
674
|
+
# always -- every command asks
|
|
675
|
+
# tiered -- read-only commands from inert_commands, every argument
|
|
676
|
+
# inside the workspace, run unprompted; every other command
|
|
677
|
+
# asks, including anything that can run code
|
|
678
|
+
# contained -- anything that runs in the container without network runs
|
|
679
|
+
# unprompted, code included; reads outside the workspace and
|
|
680
|
+
# network commands still ask
|
|
681
|
+
# never -- nothing asks
|
|
682
|
+
# An unknown value raises at startup rather than falling back, because one
|
|
683
|
+
# direction of a fallback asks about everything and the other about
|
|
684
|
+
# nothing, and a typo cannot pick. Rejected by name from settings.json at
|
|
685
|
+
# every tier: a cloned repo could otherwise set `never`.
|
|
686
|
+
# Shell approval is governed solely here: `tool_approvals` deliberately has
|
|
687
|
+
# no `shell` key, so there is no second switch that can disagree with this.
|
|
688
|
+
shell_approval_mode: tiered
|
|
689
|
+
|
|
690
|
+
# The image sandboxed commands run in, under Docker or Podman (whichever
|
|
691
|
+
# runs the sandbox -- ROADMAP_v3 §49, SS22). Fully qualified so Podman
|
|
692
|
+
# resolves it without a short-name alias (SS23). AGENT_SANDBOX_IMAGE
|
|
693
|
+
# overrides.
|
|
694
|
+
sandbox_docker_image: docker.io/library/python:3.13-slim
|
|
695
|
+
|
|
696
|
+
# Wall-clock ceiling for one sandboxed command.
|
|
697
|
+
sandbox_timeout_seconds: 120
|
|
698
|
+
|
|
699
|
+
# Memory ceiling for the sandbox, in MB.
|
|
700
|
+
sandbox_memory_mb: 2048
|
|
701
|
+
|
|
702
|
+
# CPU-time ceiling for the sandbox, in seconds.
|
|
703
|
+
sandbox_cpu_seconds: 30
|
|
704
|
+
|
|
705
|
+
# Process ceiling for the sandbox.
|
|
706
|
+
sandbox_max_pids: 200
|
|
707
|
+
|
|
708
|
+
# Background and monitor shell sessions (ROADMAP_v3 §49, SS12-SS13).
|
|
709
|
+
#
|
|
710
|
+
# The longest a session may run. An agent asking for more is started with
|
|
711
|
+
# this and told so -- never refused.
|
|
712
|
+
shell_session_timeout_cap_s: 3600
|
|
713
|
+
|
|
714
|
+
# How long an INTERACTIVE shell may sit with nothing typed into it
|
|
715
|
+
# before it is closed (ROADMAP_v3 §49, SS25). Unlike a background
|
|
716
|
+
# command, an interactive session does not block your prompt, so this is
|
|
717
|
+
# the only thing that stops a forgotten shell holding a container open
|
|
718
|
+
# until its timeout cap. Measured from the last input, not the last
|
|
719
|
+
# output: a program printing to itself is still one nobody is using.
|
|
720
|
+
shell_session_idle_timeout_s: 600
|
|
721
|
+
|
|
722
|
+
# How many sessions of any kind may be live at once, counted across the
|
|
723
|
+
# main agent and every subagent. Each is its own container under the
|
|
724
|
+
# memory ceiling above. An interactive shell counts here -- it holds a
|
|
725
|
+
# container like any other -- even though it does not block your prompt.
|
|
726
|
+
shell_session_max_live: 4
|
|
727
|
+
|
|
728
|
+
# How many times in a row finished sessions may wake the agent with no
|
|
729
|
+
# message from you. Past it, waking stops, the prompt unblocks, and the
|
|
730
|
+
# results wait for your next message.
|
|
731
|
+
shell_session_max_consecutive_wakes: 10
|
|
732
|
+
|
|
733
|
+
# How much of a session's output is kept, in memory only and never on
|
|
734
|
+
# disk: its first characters and its most recent ones. The middle is
|
|
735
|
+
# dropped, and the agent is told where.
|
|
736
|
+
shell_session_output_head_chars: 10000
|
|
737
|
+
shell_session_output_tail_chars: 190000
|
|
738
|
+
|
|
739
|
+
# ======================================================================
|
|
740
|
+
# Output redaction and tool isolation
|
|
741
|
+
# ======================================================================
|
|
742
|
+
|
|
743
|
+
# AUTHORITY -- the master switch for pattern-based redaction of everything
|
|
744
|
+
# leaving a tool, and part of the frozen posture. There is deliberately no
|
|
745
|
+
# settings.json key. VENASTINE_REDACT_OFF can only ever turn redaction
|
|
746
|
+
# OFF for a run, never on, and is bound at the same moment as this.
|
|
747
|
+
redact_tool_outputs: true
|
|
748
|
+
|
|
749
|
+
# Wall clock for one BUDGET_COMPUTE tool call, and the CPU ceiling handed
|
|
750
|
+
# to the isolated child process that runs it.
|
|
751
|
+
tool_compute_timeout_s: 20
|
|
752
|
+
|
|
753
|
+
# First words that make a sandboxed command need the network. Anything not
|
|
754
|
+
# listed runs with networking off.
|
|
755
|
+
network_allowed_commands:
|
|
756
|
+
- pip
|
|
757
|
+
- pip3
|
|
758
|
+
- curl
|
|
759
|
+
- wget
|
|
760
|
+
- git
|
|
761
|
+
- npm
|
|
762
|
+
- yarn
|
|
763
|
+
- pnpm
|
|
764
|
+
- apt
|
|
765
|
+
- apt-get
|
|
766
|
+
- yum
|
|
767
|
+
- dnf
|
|
768
|
+
- brew
|
|
769
|
+
- cargo
|
|
770
|
+
|
|
771
|
+
# Read-only inspection commands eligible for the auto-approved inert tier
|
|
772
|
+
# under `shell_approval_mode: tiered`. Sound as a bare token list only
|
|
773
|
+
# because that tier executes argv with no shell, so the tokens ARE the
|
|
774
|
+
# arguments.
|
|
775
|
+
#
|
|
776
|
+
# WHAT BELONGS HERE, after batch 117 found two things that did not. A
|
|
777
|
+
# command qualifies only if it cannot write, and "cannot write" has to hold
|
|
778
|
+
# for its ARGUMENTS as well as its name:
|
|
779
|
+
#
|
|
780
|
+
# * A writing FLAG is guarded by `_DANGEROUS_FLAGS` in security/sandbox.py
|
|
781
|
+
# -- `tree -o FILE` and `date -s TIME` were inert with no entry there
|
|
782
|
+
# until batch 117, so the tier reported `writes=False` about a write.
|
|
783
|
+
# * A writing POSITIONAL cannot be guarded at all. `uniq IN OUT` writes to
|
|
784
|
+
# its second argument with no flag to catch, so `uniq` was REMOVED from
|
|
785
|
+
# this list rather than given an entry that could not work. Teaching the
|
|
786
|
+
# classifier which positions are outputs would make it a shell parser,
|
|
787
|
+
# which security/sandbox.py's own G2 argument forbids.
|
|
788
|
+
#
|
|
789
|
+
# The bound that still held throughout: `_escapes_workspace` reads every
|
|
790
|
+
# token after the first as a path, so a write aimed outside the workspace
|
|
791
|
+
# was refused regardless. What was wrong was the profile's claim, not the
|
|
792
|
+
# containment.
|
|
793
|
+
inert_commands:
|
|
794
|
+
- ls
|
|
795
|
+
- cat
|
|
796
|
+
- pwd
|
|
797
|
+
- echo
|
|
798
|
+
- grep
|
|
799
|
+
- wc
|
|
800
|
+
- head
|
|
801
|
+
- tail
|
|
802
|
+
- date
|
|
803
|
+
- whoami
|
|
804
|
+
- uname
|
|
805
|
+
- which
|
|
806
|
+
- where
|
|
807
|
+
- type
|
|
808
|
+
- file
|
|
809
|
+
- stat
|
|
810
|
+
- du
|
|
811
|
+
- df
|
|
812
|
+
- tree
|
|
813
|
+
- diff
|
|
814
|
+
- md5sum
|
|
815
|
+
- sha256sum
|
|
816
|
+
|
|
817
|
+
# ======================================================================
|
|
818
|
+
# Research grants and review
|
|
819
|
+
# ======================================================================
|
|
820
|
+
|
|
821
|
+
# Ceiling on granted tool calls in one run, across every grant.
|
|
822
|
+
max_granted_tool_calls: 150
|
|
823
|
+
|
|
824
|
+
# How long an attended approval waits for an answer before giving up on
|
|
825
|
+
# it.
|
|
826
|
+
attended_approval_timeout_s: 600
|
|
827
|
+
|
|
828
|
+
# Whether a finished research run is reviewed. A settings.json
|
|
829
|
+
# `research.subagent_review` outranks this; --no-review escapes it for one
|
|
830
|
+
# run.
|
|
831
|
+
subagent_review: false
|
|
832
|
+
|
|
833
|
+
# Findings kept from one review; the rest are dropped with a warning.
|
|
834
|
+
max_review_findings: 25
|
|
835
|
+
|
|
836
|
+
# Rounds a single finding may be refined before it is recorded as
|
|
837
|
+
# rejected.
|
|
838
|
+
max_review_refinements: 3
|
|
839
|
+
|
|
840
|
+
# ======================================================================
|
|
841
|
+
# Compaction
|
|
842
|
+
# ======================================================================
|
|
843
|
+
|
|
844
|
+
# How much of the model's context window a thread may fill before
|
|
845
|
+
# compaction folds it. What the fraction really names is the HEADROOM --
|
|
846
|
+
# the remainder, which the step after the trigger fires has to fit inside,
|
|
847
|
+
# because nothing in the harness recovers from a prompt that overflows the
|
|
848
|
+
# window. A bare share gets that wrong at both ends of the roster: 0.10 of
|
|
849
|
+
# 1M is 100k and 0.10 of 64k is 6.4k, while ONE capped tool read is 12,500
|
|
850
|
+
# proxy tokens. Hence the two clamps below.
|
|
851
|
+
compaction_trigger_fraction: 0.90
|
|
852
|
+
|
|
853
|
+
# The headroom never shrinks below this, however small the window. Sized
|
|
854
|
+
# from max_read_chars (50,000 chars = 12,500 proxy tokens, the largest a
|
|
855
|
+
# single bounded read can return) plus the assistant turn that asked for
|
|
856
|
+
# it. Binds under a ~160k window and nowhere above one, so a 262k or 1M
|
|
857
|
+
# thread gets exactly the fraction above and a 64k thread keeps a usable
|
|
858
|
+
# margin instead of 6.4k.
|
|
859
|
+
compaction_min_headroom_tokens: 16000
|
|
860
|
+
|
|
861
|
+
# ...and never grows past this share of the window, which is what stops
|
|
862
|
+
# the floor above from eating a small window whole -- 16k of a 32k window
|
|
863
|
+
# is half of it. This is the clamp that keeps the trigger positive and
|
|
864
|
+
# above keep_recent_tokens at the bottom of the range.
|
|
865
|
+
compaction_max_headroom_fraction: 0.25
|
|
866
|
+
|
|
867
|
+
# The trigger used when NO MODEL IS IN SCOPE. A settings.json
|
|
868
|
+
# `compaction.trigger_tokens` overrides it.
|
|
869
|
+
compaction_trigger_tokens: 40000
|
|
870
|
+
|
|
871
|
+
# The early warning fires this many tokens BEFORE the trigger. Must stay
|
|
872
|
+
# strictly below the trigger, or the warning arrives after the thing it
|
|
873
|
+
# warns about; the loader refuses a pair that cannot work.
|
|
874
|
+
compaction_warning_margin_tokens: 8000
|
|
875
|
+
|
|
876
|
+
# Recent conversation held out of a fold verbatim.
|
|
877
|
+
compaction_keep_recent_tokens: 4000
|
|
878
|
+
|
|
879
|
+
# Recent turns held out of a fold verbatim, whichever bound binds first.
|
|
880
|
+
compaction_keep_recent_turns: 3
|
|
881
|
+
|
|
882
|
+
# The share of the trigger that pinned content may occupy, so a pin cannot
|
|
883
|
+
# force permanent compaction.
|
|
884
|
+
pin_max_trigger_fraction: 0.5
|
|
885
|
+
|
|
886
|
+
# Default fold strength: one of the keys of `compaction_target_ratios`.
|
|
887
|
+
compaction_strength: 3
|
|
888
|
+
|
|
889
|
+
# Attempts at one fold before it is abandoned.
|
|
890
|
+
compaction_max_retries: 2
|
|
891
|
+
|
|
892
|
+
# Strength -> the fraction of the original character count a fold aims
|
|
893
|
+
# for. INTEGER keys, and they are the closed set a settings.json
|
|
894
|
+
# `compaction.strength` is checked against.
|
|
895
|
+
compaction_target_ratios:
|
|
896
|
+
1: 0.4
|
|
897
|
+
2: 0.25
|
|
898
|
+
3: 0.15
|
|
899
|
+
4: 0.1
|
|
900
|
+
5: 0.05
|
|
901
|
+
|
|
902
|
+
# How far over target a fold may land before it is retried.
|
|
903
|
+
compaction_ratio_tolerance: 0.5
|
|
904
|
+
|
|
905
|
+
# Default strategy.
|
|
906
|
+
# chain -- fold the previous summary together with the new span
|
|
907
|
+
# rederive -- re-summarise the whole thread from the archive
|
|
908
|
+
compaction_strategy: chain
|
|
909
|
+
|
|
910
|
+
# The closed vocabulary `compaction_strategy` and any settings.json
|
|
911
|
+
# `compaction.strategy` are checked against.
|
|
912
|
+
compaction_strategies:
|
|
913
|
+
- rederive
|
|
914
|
+
- chain
|
|
915
|
+
|
|
916
|
+
# Memories injected into a prompt at once. 0 disables the `remember` tool
|
|
917
|
+
# rather than injecting nothing.
|
|
918
|
+
max_injected_memories: 50
|
|
919
|
+
|
|
920
|
+
# ======================================================================
|
|
921
|
+
# Catalog text
|
|
922
|
+
# ======================================================================
|
|
923
|
+
|
|
924
|
+
# Cap on a skill's or agent's `name` and `description` -- the two strings
|
|
925
|
+
# that reach every run's system prompt from a .md file with no tool call
|
|
926
|
+
# and no further consent. The newline collapse that comes with it matters
|
|
927
|
+
# more than the number.
|
|
928
|
+
max_catalog_text_chars: 300
|
|
929
|
+
|
|
930
|
+
# ======================================================================
|
|
931
|
+
# Summaries, references and the checklist
|
|
932
|
+
# ======================================================================
|
|
933
|
+
|
|
934
|
+
# The agent that does the summarising, for both folds and /summary.
|
|
935
|
+
compactor_agent: compactor
|
|
936
|
+
|
|
937
|
+
# How long a whole-thread summary may be, in characters. ABSOLUTE rather
|
|
938
|
+
# than a ratio: this summary is injected into every turn of a REFERENCING
|
|
939
|
+
# thread, where nothing it replaces bounds it.
|
|
940
|
+
summary_target_chars: 2000
|
|
941
|
+
|
|
942
|
+
# Referenced threads attachable to one thread at once. REFUSES rather than
|
|
943
|
+
# dropping the oldest, because a reference is something a person named.
|
|
944
|
+
max_injected_refs: 3
|
|
945
|
+
|
|
946
|
+
# The checklist vocabulary, and the enum in todo_write's own schema. A
|
|
947
|
+
# plain list rather than an enum so a list stored under an older
|
|
948
|
+
# vocabulary reads back as data instead of raising.
|
|
949
|
+
todo_statuses:
|
|
950
|
+
- pending
|
|
951
|
+
- in_progress
|
|
952
|
+
- completed
|
|
953
|
+
|
|
954
|
+
# Checklist ceiling. REFUSES rather than truncating: a list silently cut
|
|
955
|
+
# would leave the model believing it had recorded work it has forgotten.
|
|
956
|
+
max_todo_items: 50
|
|
957
|
+
|
|
958
|
+
# Research passes compact only this far from the model's actual context
|
|
959
|
+
# window -- a safety net against a hard provider error mid-pipeline, not a
|
|
960
|
+
# working-set policy.
|
|
961
|
+
compaction_pipeline_backstop_tokens: 20000
|
|
962
|
+
|
|
963
|
+
# ======================================================================
|
|
964
|
+
# Context windows
|
|
965
|
+
# ======================================================================
|
|
966
|
+
|
|
967
|
+
# Model -> input window, in tokens. A FALLBACK now, not the only source:
|
|
968
|
+
# core/client.py queries the provider where the SDK exposes it, and
|
|
969
|
+
# core/model_windows.py holds the user's own `/window` answer, which
|
|
970
|
+
# outranks both. A miss here warns once per model.
|
|
971
|
+
model_context_windows:
|
|
972
|
+
claude-opus-5: 1000000
|
|
973
|
+
claude-sonnet-5: 1000000
|
|
974
|
+
claude-fable-5: 1000000
|
|
975
|
+
claude-opus-4-8: 1000000
|
|
976
|
+
claude-opus-4-7: 1000000
|
|
977
|
+
claude-opus-4-6: 1000000
|
|
978
|
+
claude-sonnet-4-6: 1000000
|
|
979
|
+
claude-haiku-4-5: 200000
|
|
980
|
+
gpt-5.1: 400000
|
|
981
|
+
gemini-2.5-pro: 1000000
|
|
982
|
+
qwen3.8-max-preview: 983616
|
|
983
|
+
|
|
984
|
+
# For a model absent from the table above and unqueryable. Logged at
|
|
985
|
+
# WARNING once per model, because a wrong window now decides when every
|
|
986
|
+
# thread folds.
|
|
987
|
+
default_context_window: 256000
|
|
988
|
+
|
|
989
|
+
# ======================================================================
|
|
990
|
+
# Connecting a model
|
|
991
|
+
# ======================================================================
|
|
992
|
+
|
|
993
|
+
# The Hugging Face OAuth app /connect huggingface signs in through, by its
|
|
994
|
+
# client id -- a PUBLIC value, not a secret. This one is the project's own
|
|
995
|
+
# app, registered without a secret, with the device code flow and the scopes
|
|
996
|
+
# `openid profile inference-api` (ROADMAP_v3 §79, HD5). Another app's id
|
|
997
|
+
# signs in through that app instead; empty hides the sign-in and leaves a
|
|
998
|
+
# pasted token working.
|
|
999
|
+
hf_oauth_client_id: "3a823316-767d-4abe-8e55-6c0354ce2c27"
|
|
1000
|
+
|
|
1001
|
+
# ======================================================================
|
|
1002
|
+
# Tool permissions and approvals
|
|
1003
|
+
# ======================================================================
|
|
1004
|
+
|
|
1005
|
+
# AUTHORITY -- whether each registered tool may be called AT ALL.
|
|
1006
|
+
# is_tool_allowed() reads this before any context or approval logic and a
|
|
1007
|
+
# tool context can never widen it, so a false here cannot be re-enabled at
|
|
1008
|
+
# runtime by an agent, a skill or a settings file -- only by editing this
|
|
1009
|
+
# file. read / view_image / write / edit / shell ship false for that reason,
|
|
1010
|
+
# and so do the five background-session tools after shell: a session is a
|
|
1011
|
+
# shell command that keeps running, and enabling one without the others is
|
|
1012
|
+
# allowed but rarely useful. view_image follows read because it reads a file
|
|
1013
|
+
# off disk the same way; it also needs a provider marked `accepts_images` in
|
|
1014
|
+
# providers.json before it is offered at all (ROADMAP_v3 §81, IM6).
|
|
1015
|
+
# Every registered tool must appear in BOTH tables or the harness refuses
|
|
1016
|
+
# to start (D24).
|
|
1017
|
+
tool_permissions:
|
|
1018
|
+
web_search: true
|
|
1019
|
+
fetch_url: true
|
|
1020
|
+
get_time: true
|
|
1021
|
+
arxiv_search: true
|
|
1022
|
+
symbolic_math: true
|
|
1023
|
+
linear_algebra: true
|
|
1024
|
+
probability_stats: true
|
|
1025
|
+
discrete_math: true
|
|
1026
|
+
logic: true
|
|
1027
|
+
geometry: true
|
|
1028
|
+
read: false
|
|
1029
|
+
view_image: false
|
|
1030
|
+
write: false
|
|
1031
|
+
edit: false
|
|
1032
|
+
shell: false
|
|
1033
|
+
shell_background: false
|
|
1034
|
+
shell_monitor: false
|
|
1035
|
+
shell_interactive: false
|
|
1036
|
+
shell_input: false
|
|
1037
|
+
shell_sessions: false
|
|
1038
|
+
shell_output: false
|
|
1039
|
+
shell_kill: false
|
|
1040
|
+
load_skill: true
|
|
1041
|
+
spawn_subagent: true
|
|
1042
|
+
pin: true
|
|
1043
|
+
unpin: true
|
|
1044
|
+
remember: true
|
|
1045
|
+
read_project_doc: true
|
|
1046
|
+
write_project_doc: true
|
|
1047
|
+
ask_user: true
|
|
1048
|
+
todo_write: true
|
|
1049
|
+
|
|
1050
|
+
# AUTHORITY -- whether each tool needs a human yes before it runs. Approval ORs
|
|
1051
|
+
# across every layer, so a true here is a one-way ratchet that can only
|
|
1052
|
+
# add prompts. `shell`, `shell_background`, `shell_monitor` and
|
|
1053
|
+
# `shell_interactive` are deliberately absent: their approval is governed
|
|
1054
|
+
# solely by `shell_approval_mode` above -- there is no second switch, so
|
|
1055
|
+
# there is nothing here that can disagree with it. Listing, reading and
|
|
1056
|
+
# stopping a session ask nothing: they act only on this conversation's own
|
|
1057
|
+
# sessions. `shell_input` is here and ships false because approving the
|
|
1058
|
+
# shell approved what you type into it (SS29) -- set it true to be asked
|
|
1059
|
+
# about every line instead.
|
|
1060
|
+
# A tool gated here is not merely denied where nothing can ask -- it stops
|
|
1061
|
+
# being advertised at all, which is why ask_user, pin and todo_write are
|
|
1062
|
+
# ungated.
|
|
1063
|
+
tool_approvals:
|
|
1064
|
+
web_search: false
|
|
1065
|
+
fetch_url: false
|
|
1066
|
+
get_time: false
|
|
1067
|
+
arxiv_search: false
|
|
1068
|
+
symbolic_math: false
|
|
1069
|
+
linear_algebra: false
|
|
1070
|
+
probability_stats: false
|
|
1071
|
+
discrete_math: false
|
|
1072
|
+
logic: false
|
|
1073
|
+
geometry: false
|
|
1074
|
+
read: false
|
|
1075
|
+
view_image: false
|
|
1076
|
+
write: false
|
|
1077
|
+
edit: false
|
|
1078
|
+
shell_input: false
|
|
1079
|
+
shell_sessions: false
|
|
1080
|
+
shell_output: false
|
|
1081
|
+
shell_kill: false
|
|
1082
|
+
load_skill: false
|
|
1083
|
+
spawn_subagent: true
|
|
1084
|
+
pin: false
|
|
1085
|
+
unpin: false
|
|
1086
|
+
remember: true
|
|
1087
|
+
read_project_doc: false
|
|
1088
|
+
write_project_doc: true
|
|
1089
|
+
ask_user: false
|
|
1090
|
+
todo_write: false
|