AgentX-Dev 3.8.0__tar.gz → 3.8.1__tar.gz
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.
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/CHANGELOG.md +1588 -1499
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/PKG-INFO +11 -4
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/README.md +1428 -1421
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/DefaultTools.py +2382 -2321
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Session.py +39 -4
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Skills/Install.py +24 -5
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/WebTools.py +128 -28
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/app.py +17 -6
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/app_mcp.py +25 -8
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/approval.py +128 -19
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/context.py +8 -2
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/fullscreen.py +66 -17
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/mcp.py +171 -22
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/preview.py +16 -2
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/sessions.py +11 -2
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/trust.py +8 -4
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/typeahead.py +41 -4
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/ui.py +75 -13
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/webui.py +278 -42
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/webui/app.js +1386 -1308
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/webui/style.css +24 -4
- agentx_dev-3.8.1/agentx_dev/safe_exe.py +75 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev.egg-info/PKG-INFO +11 -4
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev.egg-info/SOURCES.txt +1 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/pyproject.toml +100 -100
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/AGENTX.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/CONTRIBUTING.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/LICENSE +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/MANIFEST.in +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Agents/Agent.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Agents/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/AsyncTools.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/AutoSetup.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Cache.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/ChatModel.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Compiler.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Config.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Embeddings.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Evals.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Handoffs.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Loader.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/MCP.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Media.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Memory.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Observability.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Operator.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Pictures.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Planner.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/RunLog.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Runner/AgentRun.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Runner/AsyncAgentRun.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Runner/Persistence.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Runner/Trim.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Runner/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Runner/promptTemplate.yaml +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Skills/Attach.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Skills/Defaults.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Skills/Store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Skills/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Splitters.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Streaming.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/SubAgents.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Supervisor.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/SupervisorMemory.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/Tools.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/VectorStores/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/VectorStores/chroma_store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/VectorStores/pg_store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/VectorStores/qdrant_store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/__main__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_optional.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_supervisor/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_supervisor/core.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_supervisor/plans.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_supervisor/prompts.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_supervisor/results.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/_supervisor/steps.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/app_models.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/app_team.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/app_usage.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/app_voice.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/commands.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/compact.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/memory.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/models.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/recall.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/remote.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/safepaths.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/settings.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/team.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/team_record.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/tools.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/ui_input.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/ui_text.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/cli/voice.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/promptTemplate.yaml +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/skills/auditing-a-project/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/skills/changing-files/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/skills/delegating-work/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/skills/exploring-a-project/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/skills/running-code/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/skills/web-research/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/webui/icon.svg +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev/resources/webui/index.html +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev.egg-info/dependency_links.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev.egg-info/entry_points.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev.egg-info/requires.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/agentx_dev.egg-info/top_level.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/README.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/agentic_rag_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/async_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/async_quickstart.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/auto_features_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/caching_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/chatbot_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/complete_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/concurrent_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/concurrent_tool_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/file_agent_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/function_calling_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/mcp_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/mcp_github_triage_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/observability_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/orchestration_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/planner_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/robust_link_scraper.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/subagents_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/supervisor_codebase_analysis_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/supervisor_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/sync_quickstart.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/v3_1_1_features_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/v3_1_comprehensive_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/examples/v3_1_features_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/host/build_data.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.8.1}/setup.cfg +0 -0
|
@@ -1,1499 +1,1588 @@
|
|
|
1
|
-
# Changelog
|
|
2
|
-
|
|
3
|
-
All notable changes to `agentx-dev` are documented here. Format loosely
|
|
4
|
-
follows [Keep a Changelog](https://keepachangelog.com/); versioning is
|
|
5
|
-
[Semver](https://semver.org/).
|
|
6
|
-
|
|
7
|
-
## [Unreleased]
|
|
8
|
-
|
|
9
|
-
## [3.8.
|
|
10
|
-
|
|
11
|
-
###
|
|
12
|
-
|
|
13
|
-
- **
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
window's
|
|
18
|
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
`
|
|
51
|
-
`
|
|
52
|
-
- **
|
|
53
|
-
page
|
|
54
|
-
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
- **
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
- **`
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
- **
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
|
|
144
|
-
|
|
145
|
-
|
|
146
|
-
and
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
|
|
151
|
-
|
|
152
|
-
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
|
|
159
|
-
|
|
160
|
-
|
|
161
|
-
|
|
162
|
-
|
|
163
|
-
|
|
164
|
-
|
|
165
|
-
|
|
166
|
-
|
|
167
|
-
|
|
168
|
-
|
|
169
|
-
|
|
170
|
-
- **`agentx`:
|
|
171
|
-
|
|
172
|
-
|
|
173
|
-
|
|
174
|
-
|
|
175
|
-
|
|
176
|
-
-
|
|
177
|
-
|
|
178
|
-
|
|
179
|
-
|
|
180
|
-
|
|
181
|
-
|
|
182
|
-
|
|
183
|
-
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
|
|
188
|
-
|
|
189
|
-
|
|
190
|
-
|
|
191
|
-
|
|
192
|
-
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
|
|
196
|
-
|
|
197
|
-
|
|
198
|
-
|
|
199
|
-
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
203
|
-
|
|
204
|
-
|
|
205
|
-
|
|
206
|
-
|
|
207
|
-
|
|
208
|
-
|
|
209
|
-
|
|
210
|
-
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
214
|
-
|
|
215
|
-
|
|
216
|
-
|
|
217
|
-
|
|
218
|
-
|
|
219
|
-
|
|
220
|
-
|
|
221
|
-
|
|
222
|
-
|
|
223
|
-
-
|
|
224
|
-
|
|
225
|
-
|
|
226
|
-
|
|
227
|
-
-
|
|
228
|
-
the
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
(
|
|
236
|
-
|
|
237
|
-
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
|
|
242
|
-
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
|
|
246
|
-
|
|
247
|
-
|
|
248
|
-
|
|
249
|
-
-
|
|
250
|
-
|
|
251
|
-
|
|
252
|
-
-
|
|
253
|
-
(
|
|
254
|
-
|
|
255
|
-
|
|
256
|
-
- **
|
|
257
|
-
|
|
258
|
-
|
|
259
|
-
|
|
260
|
-
|
|
261
|
-
|
|
262
|
-
|
|
263
|
-
|
|
264
|
-
|
|
265
|
-
|
|
266
|
-
|
|
267
|
-
|
|
268
|
-
|
|
269
|
-
|
|
270
|
-
|
|
271
|
-
|
|
272
|
-
(
|
|
273
|
-
|
|
274
|
-
|
|
275
|
-
|
|
276
|
-
`
|
|
277
|
-
|
|
278
|
-
|
|
279
|
-
|
|
280
|
-
|
|
281
|
-
|
|
282
|
-
|
|
283
|
-
|
|
284
|
-
|
|
285
|
-
|
|
286
|
-
|
|
287
|
-
|
|
288
|
-
|
|
289
|
-
|
|
290
|
-
|
|
291
|
-
- **
|
|
292
|
-
|
|
293
|
-
|
|
294
|
-
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
|
|
298
|
-
|
|
299
|
-
|
|
300
|
-
|
|
301
|
-
|
|
302
|
-
|
|
303
|
-
|
|
304
|
-
|
|
305
|
-
|
|
306
|
-
|
|
307
|
-
`
|
|
308
|
-
|
|
309
|
-
|
|
310
|
-
|
|
311
|
-
|
|
312
|
-
- **`agentx`:
|
|
313
|
-
|
|
314
|
-
|
|
315
|
-
|
|
316
|
-
|
|
317
|
-
|
|
318
|
-
-
|
|
319
|
-
|
|
320
|
-
|
|
321
|
-
|
|
322
|
-
|
|
323
|
-
|
|
324
|
-
-
|
|
325
|
-
|
|
326
|
-
|
|
327
|
-
|
|
328
|
-
|
|
329
|
-
|
|
330
|
-
|
|
331
|
-
|
|
332
|
-
- **`agentx`:
|
|
333
|
-
|
|
334
|
-
|
|
335
|
-
|
|
336
|
-
|
|
337
|
-
|
|
338
|
-
|
|
339
|
-
|
|
340
|
-
|
|
341
|
-
|
|
342
|
-
|
|
343
|
-
|
|
344
|
-
|
|
345
|
-
|
|
346
|
-
|
|
347
|
-
|
|
348
|
-
|
|
349
|
-
|
|
350
|
-
|
|
351
|
-
|
|
352
|
-
|
|
353
|
-
|
|
354
|
-
|
|
355
|
-
|
|
356
|
-
|
|
357
|
-
|
|
358
|
-
|
|
359
|
-
|
|
360
|
-
|
|
361
|
-
|
|
362
|
-
|
|
363
|
-
|
|
364
|
-
|
|
365
|
-
|
|
366
|
-
|
|
367
|
-
|
|
368
|
-
|
|
369
|
-
|
|
370
|
-
|
|
371
|
-
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
- **A
|
|
375
|
-
|
|
376
|
-
|
|
377
|
-
|
|
378
|
-
|
|
379
|
-
|
|
380
|
-
|
|
381
|
-
|
|
382
|
-
|
|
383
|
-
|
|
384
|
-
|
|
385
|
-
|
|
386
|
-
|
|
387
|
-
|
|
388
|
-
|
|
389
|
-
|
|
390
|
-
|
|
391
|
-
|
|
392
|
-
|
|
393
|
-
|
|
394
|
-
|
|
395
|
-
|
|
396
|
-
|
|
397
|
-
|
|
398
|
-
|
|
399
|
-
|
|
400
|
-
|
|
401
|
-
- **`agentx
|
|
402
|
-
|
|
403
|
-
|
|
404
|
-
|
|
405
|
-
|
|
406
|
-
|
|
407
|
-
|
|
408
|
-
|
|
409
|
-
|
|
410
|
-
|
|
411
|
-
|
|
412
|
-
|
|
413
|
-
|
|
414
|
-
|
|
415
|
-
|
|
416
|
-
|
|
417
|
-
|
|
418
|
-
-
|
|
419
|
-
|
|
420
|
-
|
|
421
|
-
|
|
422
|
-
|
|
423
|
-
|
|
424
|
-
|
|
425
|
-
|
|
426
|
-
|
|
427
|
-
|
|
428
|
-
|
|
429
|
-
|
|
430
|
-
|
|
431
|
-
|
|
432
|
-
|
|
433
|
-
|
|
434
|
-
|
|
435
|
-
|
|
436
|
-
|
|
437
|
-
|
|
438
|
-
|
|
439
|
-
|
|
440
|
-
|
|
441
|
-
|
|
442
|
-
|
|
443
|
-
|
|
444
|
-
|
|
445
|
-
|
|
446
|
-
|
|
447
|
-
|
|
448
|
-
|
|
449
|
-
|
|
450
|
-
|
|
451
|
-
|
|
452
|
-
|
|
453
|
-
|
|
454
|
-
|
|
455
|
-
|
|
456
|
-
|
|
457
|
-
|
|
458
|
-
|
|
459
|
-
|
|
460
|
-
|
|
461
|
-
|
|
462
|
-
|
|
463
|
-
|
|
464
|
-
|
|
465
|
-
|
|
466
|
-
|
|
467
|
-
|
|
468
|
-
|
|
469
|
-
|
|
470
|
-
|
|
471
|
-
|
|
472
|
-
|
|
473
|
-
|
|
474
|
-
|
|
475
|
-
|
|
476
|
-
|
|
477
|
-
|
|
478
|
-
|
|
479
|
-
|
|
480
|
-
|
|
481
|
-
|
|
482
|
-
|
|
483
|
-
|
|
484
|
-
|
|
485
|
-
|
|
486
|
-
|
|
487
|
-
|
|
488
|
-
|
|
489
|
-
|
|
490
|
-
- **`
|
|
491
|
-
|
|
492
|
-
- **
|
|
493
|
-
|
|
494
|
-
|
|
495
|
-
|
|
496
|
-
|
|
497
|
-
|
|
498
|
-
|
|
499
|
-
|
|
500
|
-
|
|
501
|
-
|
|
502
|
-
|
|
503
|
-
|
|
504
|
-
|
|
505
|
-
|
|
506
|
-
|
|
507
|
-
|
|
508
|
-
|
|
509
|
-
|
|
510
|
-
|
|
511
|
-
|
|
512
|
-
|
|
513
|
-
|
|
514
|
-
|
|
515
|
-
|
|
516
|
-
|
|
517
|
-
|
|
518
|
-
`
|
|
519
|
-
|
|
520
|
-
|
|
521
|
-
|
|
522
|
-
|
|
523
|
-
|
|
524
|
-
|
|
525
|
-
|
|
526
|
-
|
|
527
|
-
|
|
528
|
-
|
|
529
|
-
|
|
530
|
-
|
|
531
|
-
|
|
532
|
-
|
|
533
|
-
|
|
534
|
-
|
|
535
|
-
|
|
536
|
-
|
|
537
|
-
|
|
538
|
-
|
|
539
|
-
|
|
540
|
-
|
|
541
|
-
|
|
542
|
-
|
|
543
|
-
|
|
544
|
-
|
|
545
|
-
- **
|
|
546
|
-
|
|
547
|
-
|
|
548
|
-
|
|
549
|
-
|
|
550
|
-
|
|
551
|
-
|
|
552
|
-
|
|
553
|
-
|
|
554
|
-
|
|
555
|
-
|
|
556
|
-
|
|
557
|
-
|
|
558
|
-
|
|
559
|
-
|
|
560
|
-
|
|
561
|
-
|
|
562
|
-
|
|
563
|
-
|
|
564
|
-
|
|
565
|
-
-
|
|
566
|
-
|
|
567
|
-
|
|
568
|
-
|
|
569
|
-
|
|
570
|
-
|
|
571
|
-
|
|
572
|
-
|
|
573
|
-
|
|
574
|
-
|
|
575
|
-
|
|
576
|
-
|
|
577
|
-
|
|
578
|
-
|
|
579
|
-
|
|
580
|
-
|
|
581
|
-
|
|
582
|
-
|
|
583
|
-
|
|
584
|
-
|
|
585
|
-
|
|
586
|
-
|
|
587
|
-
|
|
588
|
-
|
|
589
|
-
|
|
590
|
-
|
|
591
|
-
|
|
592
|
-
|
|
593
|
-
|
|
594
|
-
|
|
595
|
-
|
|
596
|
-
|
|
597
|
-
|
|
598
|
-
|
|
599
|
-
|
|
600
|
-
|
|
601
|
-
|
|
602
|
-
-
|
|
603
|
-
|
|
604
|
-
|
|
605
|
-
|
|
606
|
-
|
|
607
|
-
`
|
|
608
|
-
`
|
|
609
|
-
|
|
610
|
-
|
|
611
|
-
|
|
612
|
-
|
|
613
|
-
|
|
614
|
-
|
|
615
|
-
|
|
616
|
-
|
|
617
|
-
|
|
618
|
-
|
|
619
|
-
|
|
620
|
-
|
|
621
|
-
|
|
622
|
-
|
|
623
|
-
|
|
624
|
-
|
|
625
|
-
|
|
626
|
-
so
|
|
627
|
-
|
|
628
|
-
|
|
629
|
-
- The
|
|
630
|
-
|
|
631
|
-
|
|
632
|
-
|
|
633
|
-
|
|
634
|
-
|
|
635
|
-
|
|
636
|
-
|
|
637
|
-
|
|
638
|
-
|
|
639
|
-
|
|
640
|
-
|
|
641
|
-
|
|
642
|
-
|
|
643
|
-
|
|
644
|
-
|
|
645
|
-
|
|
646
|
-
|
|
647
|
-
-
|
|
648
|
-
`
|
|
649
|
-
|
|
650
|
-
|
|
651
|
-
|
|
652
|
-
|
|
653
|
-
|
|
654
|
-
|
|
655
|
-
|
|
656
|
-
|
|
657
|
-
|
|
658
|
-
|
|
659
|
-
|
|
660
|
-
|
|
661
|
-
|
|
662
|
-
|
|
663
|
-
|
|
664
|
-
|
|
665
|
-
|
|
666
|
-
|
|
667
|
-
|
|
668
|
-
|
|
669
|
-
|
|
670
|
-
|
|
671
|
-
|
|
672
|
-
|
|
673
|
-
|
|
674
|
-
- **
|
|
675
|
-
|
|
676
|
-
|
|
677
|
-
|
|
678
|
-
|
|
679
|
-
|
|
680
|
-
|
|
681
|
-
|
|
682
|
-
|
|
683
|
-
|
|
684
|
-
|
|
685
|
-
|
|
686
|
-
-
|
|
687
|
-
|
|
688
|
-
|
|
689
|
-
|
|
690
|
-
|
|
691
|
-
|
|
692
|
-
|
|
693
|
-
|
|
694
|
-
|
|
695
|
-
|
|
696
|
-
|
|
697
|
-
|
|
698
|
-
|
|
699
|
-
|
|
700
|
-
|
|
701
|
-
|
|
702
|
-
|
|
703
|
-
|
|
704
|
-
|
|
705
|
-
|
|
706
|
-
|
|
707
|
-
`
|
|
708
|
-
|
|
709
|
-
|
|
710
|
-
|
|
711
|
-
|
|
712
|
-
|
|
713
|
-
-
|
|
714
|
-
|
|
715
|
-
|
|
716
|
-
|
|
717
|
-
|
|
718
|
-
|
|
719
|
-
|
|
720
|
-
|
|
721
|
-
|
|
722
|
-
|
|
723
|
-
|
|
724
|
-
|
|
725
|
-
|
|
726
|
-
-
|
|
727
|
-
|
|
728
|
-
`
|
|
729
|
-
|
|
730
|
-
|
|
731
|
-
|
|
732
|
-
|
|
733
|
-
|
|
734
|
-
|
|
735
|
-
|
|
736
|
-
|
|
737
|
-
|
|
738
|
-
|
|
739
|
-
|
|
740
|
-
|
|
741
|
-
|
|
742
|
-
|
|
743
|
-
|
|
744
|
-
|
|
745
|
-
|
|
746
|
-
|
|
747
|
-
|
|
748
|
-
|
|
749
|
-
|
|
750
|
-
|
|
751
|
-
`
|
|
752
|
-
|
|
753
|
-
|
|
754
|
-
|
|
755
|
-
|
|
756
|
-
|
|
757
|
-
|
|
758
|
-
|
|
759
|
-
|
|
760
|
-
|
|
761
|
-
|
|
762
|
-
|
|
763
|
-
|
|
764
|
-
|
|
765
|
-
|
|
766
|
-
|
|
767
|
-
|
|
768
|
-
|
|
769
|
-
|
|
770
|
-
|
|
771
|
-
|
|
772
|
-
|
|
773
|
-
|
|
774
|
-
|
|
775
|
-
|
|
776
|
-
|
|
777
|
-
|
|
778
|
-
|
|
779
|
-
|
|
780
|
-
|
|
781
|
-
|
|
782
|
-
|
|
783
|
-
|
|
784
|
-
|
|
785
|
-
|
|
786
|
-
|
|
787
|
-
|
|
788
|
-
|
|
789
|
-
-
|
|
790
|
-
|
|
791
|
-
|
|
792
|
-
|
|
793
|
-
|
|
794
|
-
|
|
795
|
-
|
|
796
|
-
|
|
797
|
-
|
|
798
|
-
|
|
799
|
-
|
|
800
|
-
|
|
801
|
-
|
|
802
|
-
|
|
803
|
-
|
|
804
|
-
|
|
805
|
-
|
|
806
|
-
|
|
807
|
-
|
|
808
|
-
|
|
809
|
-
|
|
810
|
-
|
|
811
|
-
|
|
812
|
-
|
|
813
|
-
|
|
814
|
-
|
|
815
|
-
|
|
816
|
-
|
|
817
|
-
|
|
818
|
-
|
|
819
|
-
|
|
820
|
-
|
|
821
|
-
|
|
822
|
-
|
|
823
|
-
|
|
824
|
-
|
|
825
|
-
|
|
826
|
-
|
|
827
|
-
-
|
|
828
|
-
|
|
829
|
-
|
|
830
|
-
|
|
831
|
-
-
|
|
832
|
-
|
|
833
|
-
|
|
834
|
-
|
|
835
|
-
|
|
836
|
-
|
|
837
|
-
|
|
838
|
-
|
|
839
|
-
|
|
840
|
-
|
|
841
|
-
|
|
842
|
-
|
|
843
|
-
-
|
|
844
|
-
|
|
845
|
-
|
|
846
|
-
|
|
847
|
-
|
|
848
|
-
|
|
849
|
-
|
|
850
|
-
|
|
851
|
-
|
|
852
|
-
|
|
853
|
-
|
|
854
|
-
|
|
855
|
-
|
|
856
|
-
|
|
857
|
-
|
|
858
|
-
|
|
859
|
-
|
|
860
|
-
|
|
861
|
-
|
|
862
|
-
|
|
863
|
-
|
|
864
|
-
|
|
865
|
-
|
|
866
|
-
|
|
867
|
-
|
|
868
|
-
|
|
869
|
-
|
|
870
|
-
|
|
871
|
-
|
|
872
|
-
|
|
873
|
-
|
|
874
|
-
|
|
875
|
-
|
|
876
|
-
|
|
877
|
-
|
|
878
|
-
|
|
879
|
-
|
|
880
|
-
|
|
881
|
-
|
|
882
|
-
|
|
883
|
-
|
|
884
|
-
|
|
885
|
-
|
|
886
|
-
|
|
887
|
-
|
|
888
|
-
|
|
889
|
-
|
|
890
|
-
|
|
891
|
-
|
|
892
|
-
|
|
893
|
-
|
|
894
|
-
|
|
895
|
-
|
|
896
|
-
|
|
897
|
-
|
|
898
|
-
|
|
899
|
-
|
|
900
|
-
|
|
901
|
-
|
|
902
|
-
|
|
903
|
-
|
|
904
|
-
|
|
905
|
-
|
|
906
|
-
- **
|
|
907
|
-
`
|
|
908
|
-
|
|
909
|
-
|
|
910
|
-
|
|
911
|
-
|
|
912
|
-
|
|
913
|
-
|
|
914
|
-
|
|
915
|
-
|
|
916
|
-
|
|
917
|
-
|
|
918
|
-
|
|
919
|
-
|
|
920
|
-
|
|
921
|
-
|
|
922
|
-
|
|
923
|
-
|
|
924
|
-
|
|
925
|
-
|
|
926
|
-
|
|
927
|
-
|
|
928
|
-
|
|
929
|
-
|
|
930
|
-
|
|
931
|
-
|
|
932
|
-
|
|
933
|
-
|
|
934
|
-
|
|
935
|
-
|
|
936
|
-
|
|
937
|
-
|
|
938
|
-
|
|
939
|
-
|
|
940
|
-
|
|
941
|
-
|
|
942
|
-
|
|
943
|
-
|
|
944
|
-
|
|
945
|
-
|
|
946
|
-
|
|
947
|
-
|
|
948
|
-
|
|
949
|
-
|
|
950
|
-
|
|
951
|
-
|
|
952
|
-
|
|
953
|
-
|
|
954
|
-
|
|
955
|
-
|
|
956
|
-
|
|
957
|
-
|
|
958
|
-
|
|
959
|
-
|
|
960
|
-
|
|
961
|
-
|
|
962
|
-
|
|
963
|
-
|
|
964
|
-
|
|
965
|
-
|
|
966
|
-
|
|
967
|
-
|
|
968
|
-
-
|
|
969
|
-
|
|
970
|
-
|
|
971
|
-
|
|
972
|
-
|
|
973
|
-
|
|
974
|
-
|
|
975
|
-
-
|
|
976
|
-
`
|
|
977
|
-
|
|
978
|
-
|
|
979
|
-
|
|
980
|
-
|
|
981
|
-
|
|
982
|
-
|
|
983
|
-
-
|
|
984
|
-
`
|
|
985
|
-
|
|
986
|
-
|
|
987
|
-
|
|
988
|
-
|
|
989
|
-
|
|
990
|
-
|
|
991
|
-
|
|
992
|
-
|
|
993
|
-
|
|
994
|
-
|
|
995
|
-
|
|
996
|
-
`
|
|
997
|
-
|
|
998
|
-
|
|
999
|
-
|
|
1000
|
-
|
|
1001
|
-
|
|
1002
|
-
|
|
1003
|
-
|
|
1004
|
-
|
|
1005
|
-
|
|
1006
|
-
|
|
1007
|
-
|
|
1008
|
-
|
|
1009
|
-
|
|
1010
|
-
|
|
1011
|
-
|
|
1012
|
-
|
|
1013
|
-
|
|
1014
|
-
|
|
1015
|
-
|
|
1016
|
-
|
|
1017
|
-
|
|
1018
|
-
|
|
1019
|
-
|
|
1020
|
-
|
|
1021
|
-
|
|
1022
|
-
|
|
1023
|
-
-
|
|
1024
|
-
|
|
1025
|
-
|
|
1026
|
-
|
|
1027
|
-
|
|
1028
|
-
|
|
1029
|
-
|
|
1030
|
-
|
|
1031
|
-
|
|
1032
|
-
|
|
1033
|
-
|
|
1034
|
-
|
|
1035
|
-
|
|
1036
|
-
|
|
1037
|
-
|
|
1038
|
-
|
|
1039
|
-
|
|
1040
|
-
|
|
1041
|
-
|
|
1042
|
-
|
|
1043
|
-
|
|
1044
|
-
|
|
1045
|
-
|
|
1046
|
-
|
|
1047
|
-
|
|
1048
|
-
|
|
1049
|
-
|
|
1050
|
-
|
|
1051
|
-
|
|
1052
|
-
|
|
1053
|
-
|
|
1054
|
-
|
|
1055
|
-
|
|
1056
|
-
|
|
1057
|
-
|
|
1058
|
-
|
|
1059
|
-
|
|
1060
|
-
|
|
1061
|
-
|
|
1062
|
-
|
|
1063
|
-
|
|
1064
|
-
|
|
1065
|
-
|
|
1066
|
-
|
|
1067
|
-
|
|
1068
|
-
|
|
1069
|
-
|
|
1070
|
-
|
|
1071
|
-
|
|
1072
|
-
|
|
1073
|
-
`
|
|
1074
|
-
|
|
1075
|
-
|
|
1076
|
-
|
|
1077
|
-
|
|
1078
|
-
|
|
1079
|
-
|
|
1080
|
-
|
|
1081
|
-
|
|
1082
|
-
|
|
1083
|
-
|
|
1084
|
-
|
|
1085
|
-
|
|
1086
|
-
|
|
1087
|
-
|
|
1088
|
-
|
|
1089
|
-
|
|
1090
|
-
|
|
1091
|
-
|
|
1092
|
-
|
|
1093
|
-
|
|
1094
|
-
|
|
1095
|
-
|
|
1096
|
-
|
|
1097
|
-
|
|
1098
|
-
|
|
1099
|
-
|
|
1100
|
-
|
|
1101
|
-
|
|
1102
|
-
|
|
1103
|
-
|
|
1104
|
-
|
|
1105
|
-
|
|
1106
|
-
|
|
1107
|
-
|
|
1108
|
-
|
|
1109
|
-
|
|
1110
|
-
|
|
1111
|
-
|
|
1112
|
-
|
|
1113
|
-
|
|
1114
|
-
|
|
1115
|
-
|
|
1116
|
-
|
|
1117
|
-
|
|
1118
|
-
|
|
1119
|
-
|
|
1120
|
-
|
|
1121
|
-
|
|
1122
|
-
|
|
1123
|
-
|
|
1124
|
-
|
|
1125
|
-
|
|
1126
|
-
|
|
1127
|
-
|
|
1128
|
-
|
|
1129
|
-
|
|
1130
|
-
|
|
1131
|
-
|
|
1132
|
-
|
|
1133
|
-
|
|
1134
|
-
|
|
1135
|
-
|
|
1136
|
-
|
|
1137
|
-
|
|
1138
|
-
|
|
1139
|
-
|
|
1140
|
-
|
|
1141
|
-
|
|
1142
|
-
|
|
1143
|
-
|
|
1144
|
-
|
|
1145
|
-
|
|
1146
|
-
|
|
1147
|
-
|
|
1148
|
-
|
|
1149
|
-
|
|
1150
|
-
|
|
1151
|
-
|
|
1152
|
-
|
|
1153
|
-
|
|
1154
|
-
|
|
1155
|
-
|
|
1156
|
-
|
|
1157
|
-
|
|
1158
|
-
|
|
1159
|
-
|
|
1160
|
-
|
|
1161
|
-
|
|
1162
|
-
|
|
1163
|
-
|
|
1164
|
-
|
|
1165
|
-
|
|
1166
|
-
|
|
1167
|
-
|
|
1168
|
-
|
|
1169
|
-
`
|
|
1170
|
-
|
|
1171
|
-
|
|
1172
|
-
|
|
1173
|
-
|
|
1174
|
-
|
|
1175
|
-
|
|
1176
|
-
|
|
1177
|
-
|
|
1178
|
-
|
|
1179
|
-
|
|
1180
|
-
|
|
1181
|
-
|
|
1182
|
-
|
|
1183
|
-
|
|
1184
|
-
|
|
1185
|
-
|
|
1186
|
-
|
|
1187
|
-
|
|
1188
|
-
|
|
1189
|
-
|
|
1190
|
-
|
|
1191
|
-
|
|
1192
|
-
|
|
1193
|
-
|
|
1194
|
-
|
|
1195
|
-
|
|
1196
|
-
|
|
1197
|
-
|
|
1198
|
-
|
|
1199
|
-
|
|
1200
|
-
|
|
1201
|
-
|
|
1202
|
-
|
|
1203
|
-
|
|
1204
|
-
|
|
1205
|
-
|
|
1206
|
-
|
|
1207
|
-
|
|
1208
|
-
|
|
1209
|
-
|
|
1210
|
-
|
|
1211
|
-
|
|
1212
|
-
|
|
1213
|
-
|
|
1214
|
-
|
|
1215
|
-
|
|
1216
|
-
|
|
1217
|
-
|
|
1218
|
-
|
|
1219
|
-
|
|
1220
|
-
|
|
1221
|
-
|
|
1222
|
-
|
|
1223
|
-
|
|
1224
|
-
|
|
1225
|
-
|
|
1226
|
-
|
|
1227
|
-
|
|
1228
|
-
|
|
1229
|
-
|
|
1230
|
-
|
|
1231
|
-
|
|
1232
|
-
|
|
1233
|
-
|
|
1234
|
-
|
|
1235
|
-
|
|
1236
|
-
|
|
1237
|
-
|
|
1238
|
-
|
|
1239
|
-
|
|
1240
|
-
|
|
1241
|
-
|
|
1242
|
-
|
|
1243
|
-
|
|
1244
|
-
|
|
1245
|
-
|
|
1246
|
-
|
|
1247
|
-
|
|
1248
|
-
|
|
1249
|
-
|
|
1250
|
-
|
|
1251
|
-
|
|
1252
|
-
|
|
1253
|
-
|
|
1254
|
-
|
|
1255
|
-
|
|
1256
|
-
|
|
1257
|
-
|
|
1258
|
-
|
|
1259
|
-
|
|
1260
|
-
|
|
1261
|
-
|
|
1262
|
-
|
|
1263
|
-
|
|
1264
|
-
`
|
|
1265
|
-
|
|
1266
|
-
|
|
1267
|
-
|
|
1268
|
-
|
|
1269
|
-
`
|
|
1270
|
-
|
|
1271
|
-
|
|
1272
|
-
|
|
1273
|
-
|
|
1274
|
-
|
|
1275
|
-
|
|
1276
|
-
|
|
1277
|
-
|
|
1278
|
-
|
|
1279
|
-
|
|
1280
|
-
|
|
1281
|
-
|
|
1282
|
-
|
|
1283
|
-
|
|
1284
|
-
|
|
1285
|
-
|
|
1286
|
-
|
|
1287
|
-
|
|
1288
|
-
|
|
1289
|
-
|
|
1290
|
-
|
|
1291
|
-
|
|
1292
|
-
|
|
1293
|
-
|
|
1294
|
-
|
|
1295
|
-
|
|
1296
|
-
|
|
1297
|
-
|
|
1298
|
-
|
|
1299
|
-
|
|
1300
|
-
|
|
1301
|
-
|
|
1302
|
-
|
|
1303
|
-
|
|
1304
|
-
|
|
1305
|
-
|
|
1306
|
-
|
|
1307
|
-
|
|
1308
|
-
|
|
1309
|
-
|
|
1310
|
-
|
|
1311
|
-
|
|
1312
|
-
|
|
1313
|
-
|
|
1314
|
-
|
|
1315
|
-
|
|
1316
|
-
|
|
1317
|
-
|
|
1318
|
-
|
|
1319
|
-
|
|
1320
|
-
are
|
|
1321
|
-
|
|
1322
|
-
|
|
1323
|
-
|
|
1324
|
-
|
|
1325
|
-
|
|
1326
|
-
|
|
1327
|
-
|
|
1328
|
-
|
|
1329
|
-
-
|
|
1330
|
-
|
|
1331
|
-
|
|
1332
|
-
|
|
1333
|
-
|
|
1334
|
-
|
|
1335
|
-
|
|
1336
|
-
|
|
1337
|
-
|
|
1338
|
-
-
|
|
1339
|
-
|
|
1340
|
-
|
|
1341
|
-
|
|
1342
|
-
|
|
1343
|
-
|
|
1344
|
-
|
|
1345
|
-
|
|
1346
|
-
|
|
1347
|
-
|
|
1348
|
-
|
|
1349
|
-
|
|
1350
|
-
|
|
1351
|
-
|
|
1352
|
-
|
|
1353
|
-
|
|
1354
|
-
|
|
1355
|
-
|
|
1356
|
-
|
|
1357
|
-
|
|
1358
|
-
|
|
1359
|
-
|
|
1360
|
-
|
|
1361
|
-
|
|
1362
|
-
|
|
1363
|
-
|
|
1364
|
-
|
|
1365
|
-
|
|
1366
|
-
|
|
1367
|
-
|
|
1368
|
-
|
|
1369
|
-
-
|
|
1370
|
-
|
|
1371
|
-
|
|
1372
|
-
|
|
1373
|
-
|
|
1374
|
-
|
|
1375
|
-
|
|
1376
|
-
|
|
1377
|
-
|
|
1378
|
-
|
|
1379
|
-
- `
|
|
1380
|
-
|
|
1381
|
-
|
|
1382
|
-
|
|
1383
|
-
|
|
1384
|
-
|
|
1385
|
-
|
|
1386
|
-
|
|
1387
|
-
|
|
1388
|
-
|
|
1389
|
-
|
|
1390
|
-
|
|
1391
|
-
|
|
1392
|
-
|
|
1393
|
-
|
|
1394
|
-
-
|
|
1395
|
-
is
|
|
1396
|
-
|
|
1397
|
-
|
|
1398
|
-
|
|
1399
|
-
|
|
1400
|
-
-
|
|
1401
|
-
|
|
1402
|
-
|
|
1403
|
-
-
|
|
1404
|
-
|
|
1405
|
-
-
|
|
1406
|
-
-
|
|
1407
|
-
-
|
|
1408
|
-
|
|
1409
|
-
|
|
1410
|
-
|
|
1411
|
-
|
|
1412
|
-
|
|
1413
|
-
|
|
1414
|
-
|
|
1415
|
-
|
|
1416
|
-
|
|
1417
|
-
|
|
1418
|
-
|
|
1419
|
-
|
|
1420
|
-
|
|
1421
|
-
|
|
1422
|
-
|
|
1423
|
-
|
|
1424
|
-
|
|
1425
|
-
|
|
1426
|
-
|
|
1427
|
-
|
|
1428
|
-
|
|
1429
|
-
|
|
1430
|
-
|
|
1431
|
-
|
|
1432
|
-
|
|
1433
|
-
|
|
1434
|
-
|
|
1435
|
-
|
|
1436
|
-
|
|
1437
|
-
|
|
1438
|
-
|
|
1439
|
-
|
|
1440
|
-
|
|
1441
|
-
|
|
1442
|
-
|
|
1443
|
-
|
|
1444
|
-
|
|
1445
|
-
|
|
1446
|
-
|
|
1447
|
-
|
|
1448
|
-
|
|
1449
|
-
|
|
1450
|
-
|
|
1451
|
-
|
|
1452
|
-
|
|
1453
|
-
|
|
1454
|
-
|
|
1455
|
-
|
|
1456
|
-
|
|
1457
|
-
|
|
1458
|
-
|
|
1459
|
-
|
|
1460
|
-
|
|
1461
|
-
|
|
1462
|
-
|
|
1463
|
-
-
|
|
1464
|
-
|
|
1465
|
-
|
|
1466
|
-
|
|
1467
|
-
|
|
1468
|
-
|
|
1469
|
-
|
|
1470
|
-
|
|
1471
|
-
|
|
1472
|
-
|
|
1473
|
-
|
|
1474
|
-
|
|
1475
|
-
|
|
1476
|
-
|
|
1477
|
-
|
|
1478
|
-
|
|
1479
|
-
|
|
1480
|
-
|
|
1481
|
-
|
|
1482
|
-
-
|
|
1483
|
-
-
|
|
1484
|
-
|
|
1485
|
-
|
|
1486
|
-
|
|
1487
|
-
|
|
1488
|
-
|
|
1489
|
-
|
|
1490
|
-
|
|
1491
|
-
|
|
1492
|
-
-
|
|
1493
|
-
|
|
1494
|
-
|
|
1495
|
-
|
|
1496
|
-
|
|
1497
|
-
|
|
1498
|
-
|
|
1499
|
-
|
|
1
|
+
# Changelog
|
|
2
|
+
|
|
3
|
+
All notable changes to `agentx-dev` are documented here. Format loosely
|
|
4
|
+
follows [Keep a Changelog](https://keepachangelog.com/); versioning is
|
|
5
|
+
[Semver](https://semver.org/).
|
|
6
|
+
|
|
7
|
+
## [Unreleased]
|
|
8
|
+
|
|
9
|
+
## [3.8.1] - 2026-11-13
|
|
10
|
+
|
|
11
|
+
### Changed
|
|
12
|
+
|
|
13
|
+
- **The app window opens like the docs site.** The mark rises with a soft glow, `agentx` types itself in letter by
|
|
14
|
+
letter and a lime cursor keeps blinking after it; "Starting..." (dots lighting up in turn) shows until agentx is
|
|
15
|
+
ready, and the welcome grows out of it. While it works, that blinking cursor replaces the round spinner.
|
|
16
|
+
- **The folder trust question highlights "No, exit".** Enter alone now exits; trust a folder with `1`, `y`, or Up
|
|
17
|
+
then Enter. The app window's trust card does the same.
|
|
18
|
+
- **"Don't ask again" is `2` in approval menus.** The `a` shortcut is gone from the arrow-key menu (it still works in
|
|
19
|
+
the typed `[y/a/n]` prompt), so a word typed for the chat cannot pick it.
|
|
20
|
+
- **Menus wait a moment before they take keys.** Keys typed before a menu appeared are thrown away, and keys pressed
|
|
21
|
+
in its first 0.4 seconds are ignored (Ctrl+C still cancels). In full-screen mode a menu waits while the input line
|
|
22
|
+
holds text (`Clear or send your message to answer`); keys keep editing the message and Enter sends it.
|
|
23
|
+
- **App window cards do not take the focus while you type.** When the message box holds text, or another field has
|
|
24
|
+
the focus, a new card leaves the focus where it is; from an empty message box `1`, `2`, `3` answer it but Enter and
|
|
25
|
+
Space do not.
|
|
26
|
+
- **Some granted commands ask again.** A "don't ask again" command grant no longer covers a call that uses an option
|
|
27
|
+
through which the program runs another program (`git fetch --upload-pack`, `make -f`, `rg --pre` and similar), and
|
|
28
|
+
`find`, `sed`, `tar`, `rsync`, `vim`, `git submodule`, `git rebase` and similar are never remembered.
|
|
29
|
+
- **An MCP server launcher that exists only as a `.ps1` script needs its full path.** A bare `"command"` is now looked
|
|
30
|
+
up on `PATH` only; one not found reads `not connected: <name> not found on PATH`.
|
|
31
|
+
- **A parent folder's `AGENTX.md` needs that folder trusted.** To use a repository's top-level file when you start
|
|
32
|
+
in a subfolder, trust the top folder once.
|
|
33
|
+
|
|
34
|
+
### Fixed
|
|
35
|
+
|
|
36
|
+
- **`--remote` on a VPS prints the ssh command that works.** It assumed SSH on port 22 and guessed the address, so
|
|
37
|
+
a server on another port (2222, say) timed out. In an SSH session agentx now uses the address and port that
|
|
38
|
+
session came in on (`ssh -p 2222 ...`).
|
|
39
|
+
|
|
40
|
+
### Security
|
|
41
|
+
|
|
42
|
+
- **Keys typed for the chat never answer an approval.** Keys still waiting when a prompt opens go to your message
|
|
43
|
+
line, never to the prompt; menus throw away keys typed before they appeared and ignore the first 0.4 seconds.
|
|
44
|
+
- **A stray Enter never trusts a folder.** The trust question opens on "No, exit", and keys typed while agentx
|
|
45
|
+
started are thrown away.
|
|
46
|
+
- **A command grant cannot approve another program.** "Don't ask again" on a command no longer approves a later call
|
|
47
|
+
that runs another program through its options (`git fetch --upload-pack`, `git grep -O`, `git -c`, `make -f` or
|
|
48
|
+
`VAR=value`, `go test -exec`, `cargo --config`, `rg --pre`, `npm --script-shell`) or holds `$ * ? [ {` for such a
|
|
49
|
+
program. Programs that run others through their options or scripts (`find`, `sed`, `tar`, `rsync`, `scp`, `zip`,
|
|
50
|
+
`vim`, `parallel`, `crontab`, `at`...) and git sub-commands such as `submodule`, `rebase`, `difftool`, `config` and
|
|
51
|
+
`hook` are approved once and never remembered.
|
|
52
|
+
- **An edit grant covers edits only.** Answering "don't ask again" at an edit prompt no longer lets web searches and
|
|
53
|
+
page fetches run unasked; they still ask until you choose `accept-edits` yourself or allow the web at a web prompt.
|
|
54
|
+
- **`/clear` and `/resume` also end a web "don't ask again" grant.**
|
|
55
|
+
- **The approval diff shows every hidden line break.** Lines split at LF only: a line break swapped for a form feed,
|
|
56
|
+
a vertical tab, U+2028 or a lone CR shows as a changed line with an escape such as `\x0c`, not as "only line
|
|
57
|
+
endings change".
|
|
58
|
+
- **Backslash paths on Linux and macOS.** A backslash is part of a file name there, so the deny rules (and the
|
|
59
|
+
prompt) also judge the file the name really reaches: a link named `x\y` cannot lead past the `.git` protection.
|
|
60
|
+
- **The window's key is never on a command line.** The window agentx opens gets a one-time start link (`?c=`) that
|
|
61
|
+
works once, for 2 minutes; the page trades it for the key. A second use is refused, the key is changed and the
|
|
62
|
+
terminal prints a warning with the new link.
|
|
63
|
+
- **An approval too long to show in full is declined.** A card whose action, target or preview would be cut is
|
|
64
|
+
declined without asking, with a notice saying why; "don't ask again" is not offered for a rule too long to show.
|
|
65
|
+
- **Links in answers cannot add attributes.** A link's address ends at a space, `(`, `)` or a quote, and each link is
|
|
66
|
+
built once, so model text cannot break out of it.
|
|
67
|
+
- **Saving Settings keeps declined project rules off.** Project allow rules you turned down at start, and rules added
|
|
68
|
+
to the project file since, stay off for the session after the window saves the settings.
|
|
69
|
+
- **Open another folder imports nothing from that folder** before the new agentx asks whether you trust it.
|
|
70
|
+
- **Pictures are never saved through a planted link.** When the project's `.agentx` or `.agentx/uploads` is a link,
|
|
71
|
+
a picture goes to `~/.agentx/uploads/`; the message names where the file really is.
|
|
72
|
+
- **Turning remote control off closes the windows opened through Tailscale** (their conversations are saved), and
|
|
73
|
+
the 30-second re-share check can no longer share the window again afterwards. agentx also changes its key, so a
|
|
74
|
+
link shared with another device stops working; windows on this computer get the new key and keep going.
|
|
75
|
+
- **Nothing is started from the project folder on Windows.** `run_shell` takes git-bash from Program Files, a
|
|
76
|
+
per-user Git install or `PATH` without the current folder, else System32's PowerShell; stopping a command runs
|
|
77
|
+
System32's `taskkill`; Tailscale, Edge, Chrome and PowerShell for the app window are found the same way.
|
|
78
|
+
- **MCP stdio servers' bare commands are found on `PATH` only**, never in the project folder (the server still
|
|
79
|
+
starts in the current folder).
|
|
80
|
+
- **Read-only auto-allows are re-checked on every connect.** Tools allowed because the server marked them read-only
|
|
81
|
+
are kept in `"auto"` in `mcp.json`, apart from your own `/mcp allow` choices; one the server stops marking read-only
|
|
82
|
+
asks again, with a warning. On the first connect after upgrading, saved `"allow"` entries the server marks
|
|
83
|
+
read-only move to `"auto"`.
|
|
84
|
+
- **A sign-in is tied to the server's address.** When a server's URL changes, the old sign-in is deleted instead of
|
|
85
|
+
being sent there; a server taken out of `mcp.json` is disconnected and its sign-in deleted; `/mcp add` under a
|
|
86
|
+
name used before starts signed out.
|
|
87
|
+
- **A parent folder's `AGENTX.md` is read only when you trust that folder.** The folder you start in and your home
|
|
88
|
+
folder's file are always read.
|
|
89
|
+
- **`grep` and `find_files` do not follow links out of the project.** A linked file that leads outside is skipped and
|
|
90
|
+
linked folders (Windows junctions included) are never entered.
|
|
91
|
+
- **`web_fetch` connects only to the address it checked**, on every redirect, so DNS rebinding cannot reach a local
|
|
92
|
+
address; a redirect to a private address reads `ERROR: refused`.
|
|
93
|
+
- **Saved conversations and their titles are readable by you only** (`0600`, new folders `0700`), and `~/.agentx` is
|
|
94
|
+
created yours only.
|
|
95
|
+
- **`/skills install` refuses a repository whose skill files have `:` in their paths** (another drive, or an NTFS
|
|
96
|
+
stream, on Windows) and never writes outside its download folder.
|
|
97
|
+
|
|
98
|
+
## [3.8.0] - 2026-11-08
|
|
99
|
+
|
|
100
|
+
### Added
|
|
101
|
+
|
|
102
|
+
- **One conversation per device in remote use.** Each device that opens the window (an iPad, a laptop, the
|
|
103
|
+
computer's own window) gets its own conversation and agent, with its own approvals, Stop and status, working at
|
|
104
|
+
the same time as the others in the same folder; before, every window shared one conversation, so a second
|
|
105
|
+
device's message slipped into the running task. The computer's conversation (the terminal's, or the first
|
|
106
|
+
window's) is listed on the others as **Live** and can be opened to share it, as before. `/exit` on a device ends
|
|
107
|
+
only its own; one unused for 15 minutes is put away and reopened when the device comes back; settings saved on
|
|
108
|
+
one device reach all; a window says when another conversation is working in the folder. Inside: each
|
|
109
|
+
conversation has its own team lane (a team job's questions, notes and `/skip` never reach another's), the
|
|
110
|
+
conversation titles, work log and team records take several writers, and one MCP sign-in runs at a time.
|
|
111
|
+
- **`agentx --ui`: the app window.** The same agentx in a desktop-style window instead of the terminal: the
|
|
112
|
+
conversations of the folder on the left (open one, or start a new one), your messages on the right, the agent's
|
|
113
|
+
steps as short rows (Read, Edit, Run, with look-around steps folded into one Explored row), answers rendered as
|
|
114
|
+
markdown, approvals as cards with the diff and Yes / don't ask again / No (with a reason) and keys 1-3, the other
|
|
115
|
+
questions as cards, the team's running steps, typed messages waiting with a button to take each back, and a
|
|
116
|
+
message box with the mode and model menus, `/` for commands and skills, a picture button (and paste or drop;
|
|
117
|
+
saved in `.agentx/uploads/`), a microphone button (`/voice`) and a send button that is Stop while it works.
|
|
118
|
+
It opens as an Edge or Chrome app window when one is installed, else in the browser. The app itself runs
|
|
119
|
+
unchanged; the window is another front end, served on 127.0.0.1 with a random token per start and a Host check.
|
|
120
|
+
`--port` and `--no-open`. No new packages.
|
|
121
|
+
- **The app window's side panel, Settings and remote use.** **Details** shows how full the conversation is, the
|
|
122
|
+
tokens used, the last team job step by step, the MCP servers by name with their own Sign in / Reconnect / Tools /
|
|
123
|
+
Sign out buttons, and the project memory. **Settings** saves the default model and mode, the allow and deny rules,
|
|
124
|
+
the steps and the token limit, for all folders or this folder (rules and limits apply at once). The **+** menu adds
|
|
125
|
+
a picture, mentions a file (search) or opens another folder in a new window. `agentx --ui --remote` keeps the window
|
|
126
|
+
on the server's 127.0.0.1 and prints the `ssh -N -L` command to use it from another computer. Light or dark follows
|
|
127
|
+
the system. A notice that MCP servers need sign-in has a button per server.
|
|
128
|
+
- **Remote control.** `/remote-control` in the terminal makes the running session usable from an iPad, a phone or
|
|
129
|
+
another computer: the terminal keeps showing everything, the window mirrors it, and messages, answers, menu choices
|
|
130
|
+
and approvals from the window go into the terminal's own prompt (the first answer wins). In the app window it adds
|
|
131
|
+
the remote ways to the window already served. With Tailscale on, `--remote` and `/remote-control` also listen on
|
|
132
|
+
the computer's Tailscale address (Tailscale devices only) and print a link for Safari; otherwise the SSH way. They
|
|
133
|
+
say when no SSH server runs, or when Windows Firewall blocks SSH on a Public network.
|
|
134
|
+
- **The model's thinking, shown.** On Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 agentx asks for a readable summary of
|
|
135
|
+
the reasoning they do on every step anyway (`thinking.display: "summarized"`, same cost; nothing is sent back on
|
|
136
|
+
later requests), and on GPT reasoning models on the Responses API for OpenAI's reasoning summary (asked once more
|
|
137
|
+
only if the account allows it). The app window shows it, and the model's notes between steps, as Thinking blocks
|
|
138
|
+
as they happen, with a live Thinking… line; the terminal shows it before an edit or a command. In the library:
|
|
139
|
+
`Claude(show_reasoning=True)`, `GPT(show_reasoning=True)`, `"reasoning"` in a `call_with_tools` result, and a
|
|
140
|
+
`{"type": "reasoning"}` runner event.
|
|
141
|
+
- **MCP sign-in from a remote window.** The sign-in page is a card in the window, and pasting the address of the
|
|
142
|
+
page that could not load afterwards finishes the sign-in on the agentx computer.
|
|
143
|
+
- **Conversation titles and closing the window.** After its first answer, a conversation gets a short title from the
|
|
144
|
+
model (one small request, on a thread of its own), shown in the app window's list and in `/resume`; older ones
|
|
145
|
+
show the first thing you typed, no longer agentx's summary note. Closing the app window quits agentx (a running turn
|
|
146
|
+
stops and is kept, the conversation is saved); a reload does not. Not with `--no-open` or `--remote`.
|
|
147
|
+
- **`agentx`: commands while it works, and `/unsend`.** Commands that only show something (`/usage`, `/settings`,
|
|
148
|
+
`/help`, `/team log`, `/skills`, `/mcp`, `/memory`, `/sessions`, and `/model`, `/team-models`, `/limit`,
|
|
149
|
+
`/fullscreen` with nothing after them) run at once while the agent works, and `/mode` changes the mode at once;
|
|
150
|
+
they used to wait (`runs when the agent is done`). `/voice` starts recording at once: Enter stops it and the words
|
|
151
|
+
go into the line being typed, to check and send. `/unsend` takes back a waiting message (or a team note) before it
|
|
152
|
+
is read: the newest, `/unsend 2`, or `/unsend all`; the waiting lines are numbered. Commands that change the
|
|
153
|
+
conversation or the setup still wait. The agent and the team's runner are told not to wait inside a command
|
|
154
|
+
(`sleep`, `--watch`): they check CI or a deploy once and report. The zero-to-hero guide has `/init` as step 4.
|
|
155
|
+
- **`agentx`: `/voice`.** Records from the default microphone until Enter (Ctrl+C cancels), writes out what was said
|
|
156
|
+
with OpenAI's speech-to-text (`gpt-4o-transcribe`, falling back to `whisper-1`; `AGENTX_VOICE_MODEL` picks another)
|
|
157
|
+
and puts the words in the next input line to check before sending, in the normal and the full-screen screen.
|
|
158
|
+
Claude models take no audio, so it needs `OPENAI_API_KEY` whatever the chat model. The microphone is read with
|
|
159
|
+
`sounddevice`: new extra `agentx-dev[voice]`. A recording stops by itself at 10 minutes; a microphone that cannot
|
|
160
|
+
record at 16 kHz records at its own rate.
|
|
161
|
+
- **Agents can look at pictures.** `read_path` on a PNG, JPEG, GIF or WebP file hands the model the picture itself
|
|
162
|
+
(it used to answer `file is not utf-8 text`), so an agent or a team helper that takes a screenshot can see it. It
|
|
163
|
+
works on Claude (inside the `tool_result`) and GPT (Chat Completions and Responses: the tool message gets the text,
|
|
164
|
+
and a message right after the tool results carries the picture). The type comes from the file's contents; a
|
|
165
|
+
picture over 5 MB or 8,000 pixels is made smaller with Pillow when it is installed, and refused otherwise. Each
|
|
166
|
+
request carries the newest three pictures from tool results; older ones become a note naming the file (the history
|
|
167
|
+
keeps them). Logs, the screen and team summaries show one line (`Picture shot.png: PNG, 1440x900, 245 KB.`); in
|
|
168
|
+
`agentx` the tool line reads `Picture, 1440x900`. In `agentx` a picture outside the project folder (a screenshot in the user's Pictures folder) is opened after an approval prompt instead of being refused; other files outside stay refused. The agent and the team planner are told they can do this. New
|
|
169
|
+
module `agentx_dev.Pictures` (`picture_result`, `ToolContent`) for tools of your own.
|
|
170
|
+
- **`agentx`: type while the agent works.** On the normal screen what you type while the spinner runs shows under it,
|
|
171
|
+
and Enter sends it to the agent at its next step, so it can change course without waiting. During a team job a
|
|
172
|
+
message goes to the team: running helpers read it at their next step, later steps start with it, and the main
|
|
173
|
+
agent reads it in the job's result (to add work). `/skip <step number or helper>` skips a team step not started
|
|
174
|
+
yet. Another command typed then runs when the agent is done. A sent message waits under the spinner (`◌ ...`) and,
|
|
175
|
+
once read, leaves one line: `❯ text ✓ read by <the agent or the helper>`. The spinner says `type to add to the
|
|
176
|
+
task`. Full-screen mode does the same (a message typed
|
|
177
|
+
there used to wait for the end of the turn). Approval prompts still get every key.
|
|
178
|
+
- **`AgentRunner.interjections` and `AsyncAgentRunner.interjections`.** Set it to a function that returns the messages
|
|
179
|
+
a user sent while the run was going; each is added to the run before the next step, and `stream` yields
|
|
180
|
+
`{"type": "interjection", "content": ...}` for it.
|
|
181
|
+
|
|
182
|
+
- **`log=` on `AgentRunner` and `AsyncAgentRunner`: watch a run as it goes.** `log=True` prints one short line per
|
|
183
|
+
step to stderr; a file path appends JSON lines (`.jsonl`) or readable lines; a `logging.Logger` sends them to your
|
|
184
|
+
logging setup; a function receives each event. Lines: the start, each model call (time, tokens in, cached and
|
|
185
|
+
out), each tool result (ok or error, arguments and result cut short), and the end (outcome and totals) or the
|
|
186
|
+
error. Secrets are redacted, and a log that cannot be written never stops the run. New `RunLog` class
|
|
187
|
+
(`RunLog.chain` writes to several). See "Watching one runner" in the observability guide.
|
|
188
|
+
- **The eval harness catches up.** `EvalRunner` runs an `AsyncAgentRunner`, a `Supervisor` or an `AsyncSupervisor`
|
|
189
|
+
as well as an `AgentRunner` (async ones from a notebook too); new checks `outcome_is`, `output_field`,
|
|
190
|
+
`used_skill`, `used_agent`, `subtasks_ok` and `EvalCase(max_tool_errors=)`, with the same fields in JSON cases;
|
|
191
|
+
`repeat=` runs each case several times (the summary shows `[2/3]`), `tags=` picks cases; results carry
|
|
192
|
+
`counted_tokens` (cached input as a tenth), `model_calls` and the failed tools; `compare_reports` and
|
|
193
|
+
`python -m agentx_dev.Evals compare a.json b.json` put two reports side by side; the CLI takes `--repeat`,
|
|
194
|
+
`--tag` and `--log`.
|
|
195
|
+
|
|
196
|
+
- **Public extension points on `Supervisor` and `AsyncSupervisor`:** `prepare_plan(plan)` (change a plan before it
|
|
197
|
+
runs), `run_step(agent_name, query, run)` (wrap or skip every step), `spawn_specialist(spec)` (build the helpers a
|
|
198
|
+
plan defines your own way) and `planner_note` (text added to the planner's list of specialists). A planned helper's
|
|
199
|
+
`new_agent` may name a `model`, kept by the plan checker and passed as `AgentSpec.model`. The `agentx` team runner
|
|
200
|
+
now uses only these, no underscore methods. See "Changing how a Supervisor runs" in the sub-agents guide.
|
|
201
|
+
|
|
202
|
+
### Changed
|
|
203
|
+
|
|
204
|
+
- **`Claude()` defaults to `claude-opus-5-5` with `max_tokens=16000`** (was `claude-sonnet-4-6`, 4096). Opus 5.5
|
|
205
|
+
always thinks and the thinking counts toward the limit, so 4096 could cut answers short. `agentx`'s help text and
|
|
206
|
+
hints name `claude-opus-5-5` and `gpt-6.1-sol`. See Upgrading.
|
|
207
|
+
|
|
208
|
+
- **`Supervisor` and `AsyncSupervisor` share one core.** Planning, replanning, synthesis, a step's retries and the
|
|
209
|
+
rounds of a session are written once (they were two copies that had to be fixed twice). `AsyncSupervisor` awaits
|
|
210
|
+
real work; `Supervisor` runs the same code with sync calls, still in your thread, with no event loop, so it keeps
|
|
211
|
+
working in a notebook. Each class keeps its own way of running a plan: one step at a time, or side by side. No
|
|
212
|
+
change in behaviour: the same events and log lines.
|
|
213
|
+
|
|
214
|
+
- **Smaller source files.** `Supervisor.py` (2,900 lines) keeps the two classes; its helpers moved to the private
|
|
215
|
+
package `agentx_dev/_supervisor/`. The `agentx` app's `ChatApp` gets its MCP, model, usage and team commands from
|
|
216
|
+
four files (`cli/app_mcp.py`, `app_models.py`, `app_usage.py`, `app_team.py`), and `cli/ui.py` its text and input
|
|
217
|
+
helpers from `ui_text.py` and `ui_input.py`. Every name is still importable from where it was.
|
|
218
|
+
|
|
219
|
+
- **The public API is written down** in `docs/reference/public-api.md`: every name exported from `agentx_dev`, the
|
|
220
|
+
stable import paths below it, what is not public (underscore names, the `agentx` command's Python modules), and
|
|
221
|
+
how patch, minor and major versions may change it. A test fails when the package and the page disagree.
|
|
222
|
+
|
|
223
|
+
- **A smaller base install.** `chromadb`, `qdrant-client`, `psycopg`, `pandas` and `openpyxl` moved from the base
|
|
224
|
+
dependencies to extras: `[rag]` (the three vector stores; `[chroma]`, `[qdrant]`, `[pgvector]` one at a time),
|
|
225
|
+
`[excel]` (spreadsheets in `Media`), `[all]`. A missing package's error names the extra. See Upgrading.
|
|
226
|
+
|
|
227
|
+
- **The package says what it is.** The PyPI summary, keywords, README opening, docs index and docs site name both
|
|
228
|
+
halves: the framework and the `agentx` terminal assistant, with helpers that run in parallel. New cookbook pattern
|
|
229
|
+
40, "Let helpers work in parallel", for the terminal app and for `AsyncSupervisor` in your own code.
|
|
230
|
+
|
|
231
|
+
### Fixed
|
|
232
|
+
|
|
233
|
+
- **The iPad's link keeps working after the computer sleeps.** Tailscale could stop sharing agentx's port (after a
|
|
234
|
+
sleep or a Tailscale restart) while agentx still ran, so the link was dead. agentx now checks every 30 seconds
|
|
235
|
+
and shares it again (the same link), saying so in the terminal, and at start clears entries earlier runs left
|
|
236
|
+
shared with nothing behind them. The tests no longer reach the Tailscale of the machine they run on.
|
|
237
|
+
- **The ssh command for another computer uses the Wi-Fi address again.** With Tailscale on, it used this computer's
|
|
238
|
+
Tailscale address (100.x), which times out from a laptop without Tailscale; the Tailscale address is now offered
|
|
239
|
+
as the one to use from a computer with Tailscale on.
|
|
240
|
+
- **The app window shows agentx at work, as the terminal does.** In `/remote-control` from the terminal it said
|
|
241
|
+
"Waiting for agentx…" with nothing moving: now the spinner line (what it does, the time and the tokens this
|
|
242
|
+
message has used) shows there too, starts the moment a message is sent, and a soft light runs along the message
|
|
243
|
+
box while it works. A conversation starts with the agentx header (version, model, folder), like the terminal's
|
|
244
|
+
box, and the link and ssh steps of `/remote-control` stay in the terminal instead of filling the window.
|
|
245
|
+
- **"Open in Safari" opens the app window, not a box saying to start agentx.** The page took its key out of
|
|
246
|
+
the address, so opening it in another browser, a new tab or from a bookmark had none, and the box could not be
|
|
247
|
+
closed. The key now stays in the address after a `#` (never sent to the server). Without a key, or with one from
|
|
248
|
+
an earlier start, the box says which, and takes the link agentx printed pasted in.
|
|
249
|
+
- **`agentx --remote` opens the terminal with remote control on.** Before, `--remote` needed `--ui` and gave the app
|
|
250
|
+
window only; now `agentx --remote` is `agentx` followed by `/remote-control` (`--ui --remote` still serves the
|
|
251
|
+
window only).
|
|
252
|
+
- **When Tailscale gives no link for the iPad, agentx says so first.** HTTPS or Serve turned off in the Tailscale
|
|
253
|
+
account (also after it worked) left only the 127.0.0.1 address, with the reason under the SSH steps; now the reason
|
|
254
|
+
and where to turn it on come first, Tailscale's own link to allow Serve is passed on, and agentx no longer waits
|
|
255
|
+
20 seconds for Tailscale there.
|
|
256
|
+
- **The app window's Details panel fits the screen.** A long MCP server name or address stretched it, so on an iPad
|
|
257
|
+
it scrolled sideways; long names and addresses now wrap inside it. Below a wide desktop window (an iPad, a
|
|
258
|
+
half-screen window) it slides over the chat instead of squeezing it, and tapping beside it or Escape closes it.
|
|
259
|
+
- **`agentx`: a turn that stops before it finishes is kept.** Out of credit, any other error, Ctrl+C or the token
|
|
260
|
+
limit used to drop the whole turn, so after switching model (say Claude ran out of credit) the new model did not
|
|
261
|
+
know what the stopped turn had changed. Now your message and a short record (files changed, commands run, the last
|
|
262
|
+
thing it said, why it stopped) go into the conversation and the work log, and `continue` picks up from there, on
|
|
263
|
+
the same model or after `/model <name>`. An out-of-credit error says so plainly and what to do instead of
|
|
264
|
+
printing the raw `BadRequestError`. Leaving prints how to come back: `agentx --continue` (with `--model <name>`).
|
|
265
|
+
- **`run_shell` and `run_python` stop at their time limit.** On Windows the limit killed only the shell and then
|
|
266
|
+
waited, with no limit, for programs it had started (`ssh`, `sleep`, a dev server) to close the output, so a 5-minute
|
|
267
|
+
limit could run for as long as such a program did (`Running run_shell... 7m 28s`). The whole process tree is now
|
|
268
|
+
stopped at the limit, and the timeout message keeps what the command printed before it. A command that ends but
|
|
269
|
+
leaves a program running in the background without redirecting its output (`cmd &`) returns at once with a note
|
|
270
|
+
on how to start it (`nohup cmd > log.txt 2>&1 &`) instead of holding the tool until the limit. Commands get an
|
|
271
|
+
empty stdin, so none can read the terminal. Ctrl+C while a command runs now stops it and everything it started
|
|
272
|
+
(a hung `ssh` used to keep the turn waiting after "Turn interrupted").
|
|
273
|
+
- **A helper that calls a tool it does not have is told so and carries on.** A "final checks" step that also had to
|
|
274
|
+
fix what failed went to the runner (it runs commands but cannot edit); its call to `edit_file` ended the step with
|
|
275
|
+
"(agent emitted an unrecognized action and no answer text)", shown as Done. agentx's agent and helpers now get the
|
|
276
|
+
framework's `strict_tool_dispatch` reply (the tool is not theirs, here are the ones they have), a step that still
|
|
277
|
+
ends with no report is marked failed, and the planner is told to give a step only to a helper with every tool it
|
|
278
|
+
needs.
|
|
279
|
+
- **A team job uses the GPT models in `teamModels` too.** With a Claude chat model and `gpt-6-luna` in the list
|
|
280
|
+
(its key found), every step of every job ran on a Claude model: the planner was told "no model is preferred" and
|
|
281
|
+
picked its own provider. It is now told the user listed these models so the work is shared among them, and, when
|
|
282
|
+
the list has both Claude and GPT models, to give at least one step to each provider.
|
|
283
|
+
- **Team steps start in plan order.** When a step finished and freed two steps at once, the one that had been waiting
|
|
284
|
+
longer went first: step 3 ran before step 2. A step now gives way to an earlier step that is ready at the same
|
|
285
|
+
time (for at most 3 seconds, so a step that never starts cannot hold up the rest).
|
|
286
|
+
- **The spinner shows minutes**: `Team is working... 11m 23s`, not `683s` (and `1h 02m` past an hour).
|
|
287
|
+
- **`agentx --model claude-opus-5.5` works.** A Claude name written with a dot got a 404 ("model: claude-opus-5.5
|
|
288
|
+
was not found") on the first message; it is now read as the dashed id (`claude-opus-5-5`). GPT names keep their dot.
|
|
289
|
+
- **A `teamModels` note may be 200 characters** (was 80, and a longer one was dropped at every start). A note past
|
|
290
|
+
200 is now shortened, with a warning, instead of dropped.
|
|
291
|
+
- **A step with no action and no answer no longer ends the run.** A reply cut off by the output limit (seen with a
|
|
292
|
+
Claude Sonnet 5.5 helper after 20 tool calls) came back empty and the run ended with "(agent emitted an
|
|
293
|
+
unrecognized action and no answer text)". The model is now asked again, up to twice per run.
|
|
294
|
+
- **`agentx` finds an API key saved in Windows after the terminal was opened.** A team model on the other provider
|
|
295
|
+
showed `Not used: gpt-... (no OPENAI_API_KEY)` when the key was set with `setx` or the Environment Variables dialog
|
|
296
|
+
after the terminal started. agentx now reads a missing `ANTHROPIC_API_KEY` or `OPENAI_API_KEY` from the saved user
|
|
297
|
+
(then system) variables at start, before each team job and on `/model`, and says which it used (never the value).
|
|
298
|
+
- **A team job's task may be 12,000 characters** (was 4,000): the main agent's detailed tasks were refused twice in
|
|
299
|
+
one job.
|
|
300
|
+
- **Claude: every agent request was declined on claude-opus-5 and claude-opus-5-5.** The agent's step tool asked
|
|
301
|
+
for a `Thought` (the model's reasoning, written out), and these models decline a request that asks for their
|
|
302
|
+
reasoning (`stop_reason: "refusal"`, category `reasoning_extraction`) before writing anything: `agentx` showed
|
|
303
|
+
nothing at all, or "unrecognized action". The step tool is now sent to Claude without its `Thought` field
|
|
304
|
+
(GPT keeps it). And a decline, or `max_tokens` used up with nothing written, now raises `ModelDeclined` with the
|
|
305
|
+
provider's reason (also in `agentx.log`) instead of returning an empty answer.
|
|
306
|
+
- **Evals: failed tool calls were never counted.** The count looked for results starting with "ToolError" in
|
|
307
|
+
`completion.tool_calls`, which lists only the calls that worked, so it was always 0. It now counts them from the
|
|
308
|
+
run log, recovered ones included.
|
|
309
|
+
- **`agentx`: an installed skill is ready at once.** `/skills install` and `/skills remove` used to wait for a new
|
|
310
|
+
conversation (`/clear`, or a restart); now the conversation you are in gets the change, with its history kept.
|
|
311
|
+
`/skills reload` does the same for a skill folder copied in by hand.
|
|
312
|
+
- **`agentx`: Claude models use prompt caching.** `agentx` built Claude models without it, so every step of a long
|
|
313
|
+
run sent the whole conversation again at full price and counted it in full against the token limit (one helper
|
|
314
|
+
used 229k tokens on 8 tool calls and stopped part-way). The system prompt, the tools and the conversation so far
|
|
315
|
+
are now read from Anthropic's cache on the next step, billed and counted at about a tenth.
|
|
316
|
+
- **`agentx`: a run that hits the step limit shows what it did.** It ended with nothing on the screen; it now shows
|
|
317
|
+
the summary and says to type `continue`. The limit is 300 steps a message (was 50): `"maxSteps"` in your settings
|
|
318
|
+
or `--max-steps` changes it. A short "continue" (or "go on", "carry on", "keep going") right after that stop
|
|
319
|
+
tells the agent to pick up from its last step instead of re-running the checks already in the conversation.
|
|
320
|
+
- **`agentx`: a team job saves every helper's work as it goes**, in `.agentx/team/<date>-<task>/` (`job.md` and one
|
|
321
|
+
file per step with the helper's whole report), and the finished reports join the project's `recall` index: a
|
|
322
|
+
closed terminal no longer loses the research.
|
|
323
|
+
- **`agentx`: team steps are shown by role.** The planner gives every step a name for what its helper does
|
|
324
|
+
(`git-checker`, `payments-fixer`, `migration-verifier`), shown in the plan, the live block, the done lines and
|
|
325
|
+
`/team log` in place of `runner` / `coder` (`git-checker@claude-sonnet-5-5`). The step still runs on the
|
|
326
|
+
built-in helper; a step without a name shows the helper's as before.
|
|
327
|
+
- **`agentx`: `/skills install` fetches only the skill files.** It asks GitHub for the repository's file list and
|
|
328
|
+
downloads just the skill folders' files, so a big repository costs what its skills weigh (`morluto/rea`, 6,677
|
|
329
|
+
files and 130 MB, was refused as "too many files"; its skill is 8 files). Skills in hidden folders such as
|
|
330
|
+
`.agents/skills` are found (they were skipped), a `/blob/.../SKILL.md` link installs that skill, and the archive
|
|
331
|
+
fallback counts only the files it unpacks.
|
|
332
|
+
- **`agentx`: `/skills install` handles repositories with big demo files.** Only the folders holding a `SKILL.md`
|
|
333
|
+
(or the folder the link names) are unpacked, so a repository's videos and docs no longer count: the download may
|
|
334
|
+
be up to 100 MB, and a skill's own files up to 10 MB each (music, fonts), 50 MB in all.
|
|
335
|
+
`latent-spaces/brag` (65 MB, mostly demo videos) was refused as "bigger than the limit". Names before the link
|
|
336
|
+
install only those skills (`/skills install brag <link>`), a misspelt subcommand gets "Did you mean /skills
|
|
337
|
+
install?", and an over-limit download says what the limit is.
|
|
338
|
+
- **Structured output on Opus 5.5, Sonnet 5.5 and Fable 5.1.** These models answer a forced tool choice with a
|
|
339
|
+
400, which `output_schema` and the answer parser used. The call now goes with `tool_choice: auto` and a line
|
|
340
|
+
asking for the tool; another model that refuses it is found from the error and remembered.
|
|
341
|
+
- **`agentx`: a saved model whose key is gone no longer stops the start.** When the other provider's key is set
|
|
342
|
+
(you moved from GPT to Claude, say), `agentx` lists that provider's models (Claude: `claude-opus-5-5`,
|
|
343
|
+
`claude-sonnet-5-5`, `claude-haiku-5-5`, `claude-fable-5-1`) or asks for a name, uses the one picked, and offers
|
|
344
|
+
to save it in the settings file the old name came from.
|
|
345
|
+
- **An OpenAI model that needs the Responses API for tool calls is found out once per program, not once per model
|
|
346
|
+
object.** Each new `agentx` session, team helper and benchmark task used to send one chat-completions request that
|
|
347
|
+
the provider rejected before switching. Now the switch is remembered for that model at that address until the
|
|
348
|
+
program exits.
|
|
349
|
+
- **Benchmark:** a result file keeps the whole model name (`gpt-6.1-sol` was saved as `gpt-6.json`), and an answer
|
|
350
|
+
check says which words it found instead of "the answer covers it". The first baseline, 10/10 on `gpt-6.1-sol`,
|
|
351
|
+
is in `bench/results/`.
|
|
352
|
+
- **`agentx`: cached input counts as a tenth toward the token limit and each team step's share.** A helper's
|
|
353
|
+
requests are mostly input the provider serves from its cache, billed at about a tenth, but they were counted in
|
|
354
|
+
full: in one audit each explorer was cut off after about 15 calls at "220k tokens", reported partial, and the agent
|
|
355
|
+
started two more team jobs. The limit question also says how many tokens were sent in full, and the team block and
|
|
356
|
+
done lines show the cached share (`230k tokens, 80% cached`).
|
|
357
|
+
- **`agentx`: helpers can open `auditing-a-project`** (version 2; it was listed only for the main agent, so helpers
|
|
358
|
+
told to follow it searched the disk for it). A step's report line drops markdown marks (`##`, `**`).
|
|
359
|
+
|
|
360
|
+
- **`agentx`: a no at the token-limit question now ends a team job cleanly.** New steps still started after the no
|
|
361
|
+
(three explorers at once, in one audit), and skill reads were not refused, so wrapping up ran past its allowance and
|
|
362
|
+
the whole turn was thrown away. Now no step starts after a no (it is listed as `not started`), `use_skill` is refused
|
|
363
|
+
like every other tool, and the team reports with what it has.
|
|
364
|
+
- **`agentx`: files are read with the file tools.** The base prompt and the runner helper say to read files with
|
|
365
|
+
`read_path` and search with `grep`/`find_files`, not print them with `run_python` or `run_shell` (one audit's runner
|
|
366
|
+
spent its whole token share printing files with Python). Helpers open a skill only when their step names it or it
|
|
367
|
+
covers exactly that step.
|
|
368
|
+
|
|
369
|
+
- **`agentx` team helpers no longer read the same files over and over.** Helpers kept 12,000 to 24,000 characters of
|
|
370
|
+
tool results whole, lost the files they had read and read them again (eight to ten times each on one audit, about
|
|
371
|
+
2.5 million tokens and twelve limit questions). They now keep 20,000 to 40,000 like the chat, and a third whole read
|
|
372
|
+
of a file that has not changed in the step gets a note instead (an edit or any command resets it; a read with
|
|
373
|
+
`offset`/`limit` always works).
|
|
374
|
+
- **A team's report reaches the chat whole.** The result text is cut at 20,000 characters (was 6,000, which cut an
|
|
375
|
+
audit's findings and made the agent start a second job).
|
|
376
|
+
|
|
377
|
+
- **The wheel installs only `agentx_dev`.** `examples` and `host` (the docs site) were installed as top-level
|
|
378
|
+
packages too; they stay in the source download. The license is now an SPDX expression (`MIT`, with
|
|
379
|
+
`license-files`), which needs setuptools 77 or later to build.
|
|
380
|
+
- **An MCP sign-in survives a restart.** The SDK loaded a saved token as never expiring, sent it after it had run
|
|
381
|
+
out, and asked for a new sign-in instead of renewing it; after a restart it also guessed the renewal address
|
|
382
|
+
(`<server>/token`, a 404 for Supabase). agentx now keeps when the token runs out and the server's OAuth details from
|
|
383
|
+
the sign-in, and hands both back, so the token is renewed at the right address.
|
|
384
|
+
- **MCP sign-in to servers that issue a client secret (Supabase).** A secret given without a method is now sent
|
|
385
|
+
in the token request; the sign-in failed with "Required parameter: client_secret".
|
|
386
|
+
- **One stuck request no longer waits about 12 minutes.** A timed-out request is retried once at most (the client
|
|
387
|
+
has usually retried it already; a timed-out request may still be billed), and a 429 waits for its `Retry-After`.
|
|
388
|
+
- **A spent budget is not retried.** `CostBudgetExceeded` (and anything `model.on_usage` raises) is raised once; on
|
|
389
|
+
the plain chat path the call used to be retried up to three times after the budget was spent, paying each time.
|
|
390
|
+
|
|
391
|
+
### Added
|
|
392
|
+
|
|
393
|
+
- **A benchmark for `agentx`** (`python -m bench`): ten fixed tasks on two sample repos (find, explain, fix, fix and
|
|
394
|
+
test, validate, add a feature, rename, audit, a team job, run and report), each checked automatically (the tests,
|
|
395
|
+
hidden tests, the files), with pass/fail, time, model requests, tokens, the cached share and tool calls recorded;
|
|
396
|
+
`compare` sets two runs side by side. It asks before it uses your API key. See `bench/README.md`.
|
|
397
|
+
- **`agentx`: `/mcp reload [name]`** (also `/mcp refresh`): reads `~/.agentx/mcp.json` again and reconnects every
|
|
398
|
+
server (or one) with the saved sign-ins, never opening the browser; servers taken out of the file are dropped and
|
|
399
|
+
the conversation gets the new tools. On Windows it also reads the user's saved environment variables, so a key set
|
|
400
|
+
with `setx` after the start is found. `MCPManager.reload(name=None, env=None)`.
|
|
401
|
+
- **`agentx`: full-screen mode** (`agentx --fullscreen`, `/fullscreen on|off`, `"fullscreen": true` in
|
|
402
|
+
`~/.agentx/settings.json`; off by default). The welcome box, the input and the status line stay put while only the
|
|
403
|
+
conversation scrolls (mouse wheel, PgUp / PgDn, Ctrl+End). Questions are asked above the input; a message typed
|
|
404
|
+
while the agent works is kept for later; Ctrl+C and Ctrl+D work as before; on exit the conversation is written to
|
|
405
|
+
the terminal. The app runs unchanged on a worker thread: its console writes into the screen, and the spinner is
|
|
406
|
+
drawn by the screen (`Progress.drawn_by`).
|
|
407
|
+
- **`agentx`: the agent looking around is one tree.** Skill reads, `recall`, reads, searches, listings and web calls
|
|
408
|
+
in a row are drawn as one flow (`● Skills using-superpowers` / `├─ recall ...` / `│ ⎿ Found 2 notes` /
|
|
409
|
+
`└─ read_path ...`), redrawn in place above the spinner while it runs (a blinking yellow dot, the running step
|
|
410
|
+
blinking white) and printed when the agent moves on, with a green dot (red when the last step failed). The agent's narration leading into those steps or into its answer is no longer printed; a note before
|
|
411
|
+
an edit or a command still is.
|
|
412
|
+
|
|
413
|
+
- **`agentx`: the supervisor can make helpers.** A plan step may define a helper (`new_agent`: a name, instructions,
|
|
414
|
+
tools). agentx builds it the way it builds explorer, coder and runner (gated tools, step
|
|
415
|
+
wrapping, redaction, 100 steps, trimming, skills); its tools come only from the gated set, never `delete_path`,
|
|
416
|
+
`move_file`, `start_team_job` or `delegate`, and it gets only the MCP tools it names. A helper with only reading
|
|
417
|
+
tools runs beside other reading steps. At most 6 per job; the plan shows each one and its tools. With helper
|
|
418
|
+
models set (`/team-models`), it may run on one of them (`"model"` in `new_agent`, read before the plan checker drops
|
|
419
|
+
it): the planner is told the list, a model not on it is not used, and the name shows it (`reviewer@gpt-6-luna`).
|
|
420
|
+
- **`Skills(guard=fn)`**: `fn()` is asked before each `use_skill` read; a text it returns is the answer instead.
|
|
421
|
+
- **`agentx`: reading steps of a team job run side by side.** Explorer steps that do not depend on each other run
|
|
422
|
+
up to 3 at once (`AsyncSupervisor` with `max_parallel=3` and a lock: a coder or runner step runs alone). A plan
|
|
423
|
+
without `depends_on` keeps its order, with neighbouring explorer steps grouped. Helpers run in worker threads; every
|
|
424
|
+
approval, team question, limit question and screen line is handed to the main thread, so prompts never overlap,
|
|
425
|
+
and Ctrl+C stops the job at once (a helper stops at its next tool call or model reply).
|
|
426
|
+
- **`agentx`: each team step has a token share,** one token limit's worth (at least 100,000); past it the step's
|
|
427
|
+
tools tell the helper to report what it has.
|
|
428
|
+
- **`agentx`: `a` at the token-limit question** goes on and does not ask again in this message.
|
|
429
|
+
- **`agentx`: team helpers work in the background.** Their tool calls are no longer printed. A block above the
|
|
430
|
+
spinner, redrawn in place, shows one line per running step (helper · step name, what it has done, time, tokens);
|
|
431
|
+
a finished step leaves one line (`✓ explorer · security audit Done (20 tool uses · 41k tokens · 1m 30s)`) and the
|
|
432
|
+
start of its report. `/team log` lists every call of the last team job. Steps are named by their plan id.
|
|
433
|
+
- **The `auditing-a-project` skill** (bundled): map, run the project's own checks, read by area with explorers side
|
|
434
|
+
by side, verify, and report Fix first / Medium / Low / Done well / Checks / Test gaps. `delegating-work` (version
|
|
435
|
+
2) says which steps run side by side.
|
|
436
|
+
|
|
437
|
+
- **Smaller requests.** `AgentRunner(..., trim_results=True)` (and `AsyncAgentRunner`, or a `TrimPolicy` from
|
|
438
|
+
`agentx_dev.Runner.Trim`) sends the model a one-line stub in place of each big tool result outside a window of the newest
|
|
439
|
+
ones (20,000 to 40,000 characters, moved in steps, at least the newest one), stubs an older word-for-word copy of a later result, and shortens long
|
|
440
|
+
arguments of old calls; the saved history is unchanged and `use_skill` results are never
|
|
441
|
+
trimmed. `Permissions.read_max_chars` makes `read_path` return a long file in numbered parts with a note on how
|
|
442
|
+
to continue. `Skills(remember_reads=True)` answers a second read of a skill with a pointer back.
|
|
443
|
+
- **`agentx`:** all of the above are on (reads in parts of 30,000 characters); summarising starts at 60,000
|
|
444
|
+
characters (was 150,000) and is measured on what the model is sent. On a measured 15-file turn the input sent
|
|
445
|
+
fell from about 231,000 tokens to about 136,000. `/usage` shows the tokens used this session, per model, and by
|
|
446
|
+
the last turn.
|
|
447
|
+
- **Fewer whole-file re-reads.** A trimmed `read_path` result now names the file and its line count (or the line
|
|
448
|
+
range of a slice) and says to grep it or read only the lines needed; `agentx`'s base prompt and the
|
|
449
|
+
`exploring-a-project` skill (version 3) tell agents to write down the facts they will need as they read.
|
|
450
|
+
- **Cache-friendly trimming.** `TrimPolicy(step_chars=)`: once the whole results pass `keep_chars`, the oldest are
|
|
451
|
+
stubbed in one go down to `keep_chars - step_chars` (default half), so the start of the request stays the same,
|
|
452
|
+
and cacheable, until the next step; `keep_last` is now 1. Team helpers keep 12,000-24,000 characters. On a replayed
|
|
453
|
+
40-step team brainstorm the input cost fell by about 28-41%, depending on the cached price.
|
|
454
|
+
- **A short `AGENTX.md`.** This repo's `AGENTX.md` is now about 1,800 characters (was 13,300), since `agentx` sends
|
|
455
|
+
it with every request; the full guide moved to `docs/reference/agent-guide.md`.
|
|
456
|
+
- **Older skill texts are stubbed in a long conversation.** `TrimPolicy(shrink_exempt_at=60_000)`: past that many
|
|
457
|
+
characters of conversation, every `use_skill` result but the newest is sent as a note that the skill still applies
|
|
458
|
+
and how to reopen it. In `agentx`, `run_python` and `run_shell` output is cut at 30,000 characters (was 100,000 and
|
|
459
|
+
40,000), with a note to print less. Replaying a 21-step message that read five skills, the input sent fell by
|
|
460
|
+
about 15%.
|
|
461
|
+
- **OpenAI cached tokens are counted** (`TokenUsage.total_cache_read_tokens`, from either API), and `/usage` shows how
|
|
462
|
+
much of the input came from the provider's cache.
|
|
463
|
+
- **A token limit per message in `agentx`.** Past 200,000 tokens (`/limit`, `"tokenLimit"` in settings; a project
|
|
464
|
+
may only lower it) agentx asks before it goes on; a no wraps up (tool calls are refused, the agent answers with
|
|
465
|
+
what it has, the turn is kept), and only a wrap-up that runs long too ends the turn like Ctrl+C.
|
|
466
|
+
`BaseChatModel.on_usage`, a hook called after every recorded reply, is what it uses.
|
|
467
|
+
- **Web tools in `agentx`.** `web_search` and `web_fetch` (readable text, at most 30,000 characters, marked as data
|
|
468
|
+
from the internet); the gate asks before each in `default` and `plan` mode (`a`: the rest of the session). The
|
|
469
|
+
`web-research` skill is listed, and the team's explorer gets them.
|
|
470
|
+
- **MCP servers in `agentx`.** `/mcp add <name> <url>` (`--sse`), `/mcp add <name> -- <command> [args]`, `/mcp remove`,
|
|
471
|
+
`/mcp login`, `/mcp logout`; kept in `~/.agentx/mcp.json`, `${VAR}` for keys. A browser sign-in (OAuth) runs only on
|
|
472
|
+
`/mcp login`; tokens go to the system's credential store (`keyring`, split across entries when long) and renew
|
|
473
|
+
by themselves. Tools appear as `<server>__<tool>` for the chat and every team helper, ask before running even in
|
|
474
|
+
auto mode, and their results are marked as outside data and cut at 30,000 characters. The `mcp` extra now needs
|
|
475
|
+
`mcp>=1.10,<2` (2.x changed the client API and is not supported yet) and `keyring>=24`.
|
|
476
|
+
- **MCP tools sorted by what they do.** The first time agentx sees a server's tools, the ones the server marks
|
|
477
|
+
read-only run without asking (saved per server in `mcp.json` as `allow`/`seen`); the rest ask. `/mcp tools <name>`
|
|
478
|
+
lists them, `/mcp allow|ask <name> <tool...|read-only|all>` changes them; allowing a destructive tool asks first.
|
|
479
|
+
- **Model requests in `agentx` wait up to 5 minutes and are retried visibly.** `model.on_retry(attempt, error,
|
|
480
|
+
delay)` is called before each retry; `agentx` prints a line.
|
|
481
|
+
- **Tool results are one line in `agentx`.** Under each call: `⎿ Read 51 lines`, `⎿ Found 2 files`, `⎿ 3 matches in
|
|
482
|
+
12 files searched`, `⎿ Found 1 note`, a command's first line of output and how many more, or its exit code and
|
|
483
|
+
last line in red when it failed. Refusals and errors are still shown in full.
|
|
484
|
+
- **Skill reads are one line in `agentx`.** Their text is never printed; a run of them shows on the spinner as it
|
|
485
|
+
grows and then stays as `● Skills a → b → c (+ file)`, with the notes written meanwhile after it. A failed read
|
|
486
|
+
is still shown.
|
|
487
|
+
- **A new `agentx` start screen.** A box with the agentx mark, the model and folder, tips for this folder and
|
|
488
|
+
what's new; MCP sign-in problems as one warning line; notes with a bar down the left; a rule above the input
|
|
489
|
+
and the mode under it (`⏵⏵ auto mode on (shift+tab to cycle) · model · folder`).
|
|
490
|
+
- **`agentx` input colours:** the `/name` at the start of the input line is coloured as you type (command cyan, skill
|
|
491
|
+
magenta, unknown red) and the selected row of the `/` pop-up is highlighted.
|
|
492
|
+
- **Skills first.** The skill list tells the model to check it before choosing tools and to prefer a skill the user
|
|
493
|
+
installed when it fits; the user's own skills are listed ahead of the built-in ones, and the terminal app's base
|
|
494
|
+
prompt says the same.
|
|
495
|
+
- **`agentx` project notes and recall:** a readable note per worthwhile turn in `.agentx/notes/` (asked, result,
|
|
496
|
+
files changed, commands), indexed in `.agentx/memory/` with OpenAI embeddings when `OPENAI_API_KEY` is set
|
|
497
|
+
(local similarity search otherwise, keywords if both fail). The agent gets a `recall` tool, a fresh
|
|
498
|
+
conversation's first message gets up to 3 close notes, and `/memory [words]` shows or searches them.
|
|
499
|
+
|
|
500
|
+
- **Skills.** `AgentRunner(..., skills=True)` (and `AsyncAgentRunner`) gives an agent a `use_skill` tool and a short
|
|
501
|
+
list of skills at the end of its system prompt; the model reads a skill's full text only when it needs it. The
|
|
502
|
+
list is built the same way every time, so the prompt stays identical for the provider's prompt cache. Five
|
|
503
|
+
skills ship with agentx (`exploring-a-project`, `changing-files`, `running-code`, `web-research`,
|
|
504
|
+
`delegating-work`) and are created in `~/.agentx/skills/` on first use; edits and deletions are respected. The
|
|
505
|
+
agent's own tools are described by a generated `custom-tools` skill. `SKILL.md` folders written for Claude work
|
|
506
|
+
as they are. See the new [guide](docs/guides/skills.md).
|
|
507
|
+
- **`agentx` terminal app:** skills are on by default; `/skills` lists them, `/skills install <github link or
|
|
508
|
+
folder>` installs after showing what it found (the exact commit is downloaded and recorded; archives are
|
|
509
|
+
unpacked with path, link and size limits), `/skills remove <name>` deletes one. Every skill the agent has is also a slash command: typing `/` lists the commands and then the skills, and `/<skill> [what you want]` (or `/skills use <name> ...` for a skill named like a command) makes the agent read that skill and do the task. Team helpers get `use_skill` and the skills that fit their role's tools, so a team step can use a skill; a helper's skill reads show on screen like its other tool calls (`● [explorer] use_skill brainstorming`); `/skill` is a short form of `/skills`. Only the user can install: the
|
|
510
|
+
agent has no tool for it, and `.agentx/*` and `~/.agentx` stay protected from its file tools.
|
|
511
|
+
|
|
512
|
+
## [3.7.0] - 2026-10-15
|
|
513
|
+
|
|
514
|
+
### Added
|
|
515
|
+
|
|
516
|
+
- **The `agentx` command**: a terminal coding assistant (`agentx`, or `python -m agentx_dev`), built on
|
|
517
|
+
`AgentRunner`, `DefaultTools` and `Session`. Run it in a project folder; it works with GPT and Claude from
|
|
518
|
+
`OPENAI_API_KEY` / `ANTHROPIC_API_KEY`. Flags: `--model`, `--mode`, `--add-dir`, `--continue`,
|
|
519
|
+
`--resume [ID]`, `--version`. See the new [guide](docs/guides/cli.md).
|
|
520
|
+
- **Approvals and modes.** Reads run; edits, commands and delete / move ask first, with a diff or the full
|
|
521
|
+
command shown. Modes `default`, `accept-edits`, `plan` and `auto` (edits, shell and Python run without asking; delete, move and unknown tools still ask; entered only on purpose, never from Shift+Tab, an `a` answer or a project settings file). Answering `a` ("don't ask again") remembers
|
|
522
|
+
all edits or a plain command prefix for the conversation, never for interpreters, shells, compound
|
|
523
|
+
commands, `run_python`, delete or move. A preview that cannot be shown, or a command over 20,000
|
|
524
|
+
characters, is declined with the reason.
|
|
525
|
+
- **Default protections.** Writes, edits, deletes and moves of the whole `.git`, of `.agentx/*` in the
|
|
526
|
+
project, and (when the home folder is reachable) of `~/.agentx`, git config files and common shell
|
|
527
|
+
profiles are always refused, in every mode. Deny and allow rules are a convenience layer, not a secrets
|
|
528
|
+
vault: `grep`, `find_files` and `list_directory` can still read a file a read rule denies, and
|
|
529
|
+
`run_shell` / `run_python` are not covered by path rules.
|
|
530
|
+
- **Settings and rules.** `~/.agentx/settings.json` and `.agentx/settings.json` (`model`, `mode`,
|
|
531
|
+
`addDirs`, `permissions.allow` / `deny` as `tool:glob`). Project settings are untrusted: an
|
|
532
|
+
`accept-edits` mode and `addDirs` outside the project are ignored with a warning, and the project's
|
|
533
|
+
allow rules are listed at startup and honoured only if you answer yes to "Honour these project rules?"
|
|
534
|
+
(deny rules from the project are always kept).
|
|
535
|
+
- **`AGENTX.md` context.** `~/.agentx/AGENTX.md` and the files from the working folder up through its
|
|
536
|
+
parents are passed to the model as instructions (nearest last, 20,000 characters); `/init` drafts one.
|
|
537
|
+
- **Sessions.** Conversations autosave to `~/.agentx/sessions/` and resume with `--continue`,
|
|
538
|
+
`--resume` and `/resume`. The files hold tool results (file contents) and are not encrypted.
|
|
539
|
+
- **Team jobs.** For a big job the chat agent can call `start_team_job` (the user approves each hand-off,
|
|
540
|
+
unless you run in auto mode or have allowed it with your own rule; `/team <task>` does it by hand): a
|
|
541
|
+
`Supervisor` plans up to six steps and three helpers (explorer reads,
|
|
542
|
+
coder edits, runner runs commands) carry them out one at a time. Every helper call goes through the same
|
|
543
|
+
gate, modes, deny rules and prompts as the chat; the result joins the conversation. A deny rule
|
|
544
|
+
`start_team_job:*` turns it off.
|
|
545
|
+
- **Helper models for team jobs.** Allow extra models with `teamModels` (in either settings file; a folder's
|
|
546
|
+
list replaces yours) or `/team-models`, and the supervisor picks per step which model a helper runs on
|
|
547
|
+
(Claude and GPT can be mixed when both keys are set). Every helper then names its model, the chat's own
|
|
548
|
+
included (`coder@claude-opus-5-5`, `runner@claude-sonnet-5-5`), with the same wording for every model so
|
|
549
|
+
none is favoured, and each model can carry a note on what it is for (`gpt-6.1-sol="hard coding"`), which
|
|
550
|
+
the supervisor reads; without one, the chat model is asked once what each model is good for and its
|
|
551
|
+
answer is kept as an `auto:` note (`/team-models describe` asks again). The approval box lists the models. A list that comes from a project's settings file
|
|
552
|
+
is announced at startup. With no list nothing changes. Team helpers now get 100 steps (was 25); a helper
|
|
553
|
+
that runs out of steps hands on its partial work instead of failing the step, and one message may run up to
|
|
554
|
+
5 team jobs, so a big task continues batch after batch without you typing "continue". See the guide.
|
|
555
|
+
- **Project memory.** Each folder can have its own model: the first time you trust a folder, `agentx`
|
|
556
|
+
writes `.agentx/settings.json` with the model in use, and `/model <name> --folder` changes it (plain
|
|
557
|
+
`/model <name>` stays for the session). The app keeps a work log in `.agentx/log.md`, written by the app
|
|
558
|
+
(not the model) after each finished turn with the request, the files changed, the commands run and the
|
|
559
|
+
start of the answer; keys are hidden and the file is readable by its owner only where the platform
|
|
560
|
+
allows. A fresh conversation (a start without `--continue` / `--resume`, and `/clear`) starts from the
|
|
561
|
+
newest entries of that log, and a very long conversation is summarised before it overflows (the older
|
|
562
|
+
messages become one summary; the last six of your messages stay as they were). Not kept for the home
|
|
563
|
+
folder or a drive root. See the guide.
|
|
564
|
+
- **Slash commands.** `/help`, `/clear`, `/model`, `/mode`, `/settings`, `/resume`, `/sessions`, `/init`,
|
|
565
|
+
`/team`, `/team-models`, `/exit`. Typing `/` pops up the list with a one-line description of each. The
|
|
566
|
+
first run writes `~/.agentx/settings.json` with the model in use, once it is built, so there is a file to
|
|
567
|
+
edit; `/settings` shows what is in force and where the files are. A `"model": "(no model)"` line that an
|
|
568
|
+
earlier build of this release could write there is ignored with a warning to delete it.
|
|
569
|
+
- **Trust and menus.** The first start in a folder asks whether you trust it ("Accessing workspace", Yes /
|
|
570
|
+
No); nothing from the folder (`AGENTX.md`, project settings) is read before you answer, and the answer is
|
|
571
|
+
kept in `~/.agentx/trusted.json`. Choices are numbered menus you move through with the arrow keys and
|
|
572
|
+
confirm with Enter (or a number, or `y` / `a` / `n`): the approval prompt, the folder question and the
|
|
573
|
+
`/resume` list. Without a terminal they fall back to typed answers.
|
|
574
|
+
- **A prompt that looks alive.** The input prompt is coloured and its cursor blinks while it waits for you.
|
|
575
|
+
- **Progress.** While the agent works, a spinner with the seconds elapsed shows what it is doing
|
|
576
|
+
("Thinking...", "Running run_python... 12s"). It is paused while a prompt waits and drawn only on a
|
|
577
|
+
terminal. Python code may run for 2 minutes and shell commands for 5 (the framework's defaults, 10 s and
|
|
578
|
+
30 s, stopped web fetches and test runs); time spent waiting for approval does not count.
|
|
579
|
+
- **`Session.stream(user_input, **kw)`**: like `Session.invoke`, but yields the runner's events and absorbs the
|
|
580
|
+
result on the final `completion` event. A turn abandoned before then is not recorded.
|
|
581
|
+
- **New base dependency: `prompt_toolkit>=3.0`** (the input line: history, multi-line, completion, status
|
|
582
|
+
bar). `rich` was already a dependency.
|
|
583
|
+
|
|
584
|
+
### Fixed
|
|
585
|
+
|
|
586
|
+
- **A conversation can continue after a run that ended on a final answer given as a function call.** With
|
|
587
|
+
function calling the model delivers its answer as a call (`React_` or `respond`) and the run ended
|
|
588
|
+
without a result for it, so the next turn was rejected by OpenAI ("An assistant message with
|
|
589
|
+
'tool_calls' must be followed by tool messages" / "No tool output found for function call"). The history
|
|
590
|
+
a run returns now closes that call, and a history saved by an older version is repaired when resumed.
|
|
591
|
+
This affects `Session` and `chat_history` users as well as the CLI.
|
|
592
|
+
- **`edit_file` no longer rewrites a file's line endings.** It read and wrote through the platform's newline
|
|
593
|
+
translation, so on Windows an LF file came back CRLF and on Linux and macOS a CRLF file came back LF, with
|
|
594
|
+
every line changed. It now keeps every byte outside the edit. A model's bare newlines in `find` and
|
|
595
|
+
`replace` are put in the file's own ending (CRLF if the file mostly uses it), so CRLF files can be edited
|
|
596
|
+
across lines.
|
|
597
|
+
|
|
598
|
+
## [3.6.2] - 2026-10-09
|
|
599
|
+
|
|
600
|
+
### Added
|
|
601
|
+
|
|
602
|
+
- **`Supervisor(ask_when_done=True)` / `follow_up=True`** (and the same on `AsyncSupervisor`): when a
|
|
603
|
+
round's answer is ready the framework itself asks the operator "What should I do next?" through the
|
|
604
|
+
`ask_user` asker (the answer, the outcome and the question in one prompt) and returns the reply in
|
|
605
|
+
`SupervisorResult.next_step`. `follow_up=True` runs another round on each reply, with the earlier
|
|
606
|
+
results as context, until the operator replies blank or with a closing word (any of `done`, `stop`,
|
|
607
|
+
`quit`, `exit`, `no`, `nothing`, `that's all`, `that's it`, `all done`, `nope`, `bye`, `goodbye`,
|
|
608
|
+
`thanks`, `thank you`, `no thanks`, `no thank you`, `done thanks`, `done thank you`, `nothing thanks`,
|
|
609
|
+
`stop thanks`; case-insensitive; apostrophes (`thats all` works), commas, extra spaces and trailing `.`, `!`, `?` ignored); `max_follow_ups` (default
|
|
610
|
+
5) bounds it and `SupervisorResult.rounds` lists the rounds. Asked whatever the outcome. The
|
|
611
|
+
checkpoint uses no `max_questions` slot and is never saved to memory; only round 1's result is.
|
|
612
|
+
New `checkpoint` and `round` stream events. Either setting needs `ask_user`. With `ask_when_done`
|
|
613
|
+
on, the Supervisor asks itself, so the planner and the helpers are told not to ask "what next?"
|
|
614
|
+
(no double question); the built-in terminal prompt puts the question first.
|
|
615
|
+
|
|
616
|
+
### Fixed
|
|
617
|
+
|
|
618
|
+
- **"When done, ask me" now reaches `ask_user`.** With `ask_user=True` and a task that said to ask the
|
|
619
|
+
operator what to do when finished, the helper wrote the question into its answer and the run ended
|
|
620
|
+
with nobody to answer. A helper's instruction covered only a missing fact; it now also covers a task
|
|
621
|
+
that tells it to check with the user, and says to call `ask_user` at that point and act on the reply.
|
|
622
|
+
The planner is told to put that instruction in the step's query rather than plan a step that only
|
|
623
|
+
asks.
|
|
624
|
+
|
|
625
|
+
- A planner-defined helper built with no tools (a research step the planner forgot to give `web`) is
|
|
626
|
+
now called out in the verbose `[spawn]` line ("no tools, so it can only reason"). The planner is
|
|
627
|
+
reminded, as the last rule of the helper instructions, to give every `new_agent` the tools its job
|
|
628
|
+
needs, and `capabilities` is accepted as an alias for `tools` in a `new_agent` (the 3.5 spelling).
|
|
629
|
+
- The verbose line "has the same tools as the registered specialist" no longer fires because both
|
|
630
|
+
agents were given `ask_user`.
|
|
631
|
+
|
|
632
|
+
### Changed
|
|
633
|
+
|
|
634
|
+
- **The vector-store clients are core dependencies.** `chromadb`, `qdrant-client` and `psycopg[binary]`
|
|
635
|
+
now install with `pip install agentx-dev`, so `ChromaVectorStore`, `QdrantVectorStore` and
|
|
636
|
+
`PgVectorStore` work without an extra. `[chroma]`, `[qdrant]` and `[pgvector]` remain as aliases.
|
|
637
|
+
The base install is larger (chromadb brings onnxruntime and others); `anthropic`, `mcp` and
|
|
638
|
+
`opentelemetry` are still extras.
|
|
639
|
+
- The planner's catalog lists each registered agent's tools under its description (framework tools
|
|
640
|
+
`ask_user` and `delegate` left out, long lists cut at 12), so a description no longer has to say
|
|
641
|
+
that an agent can write files. Registered agents whose runner has no tool registry are unchanged.
|
|
642
|
+
- The verbose `[spawn]` line ends with `+ ask_user` for a helper that can ask. `result.spawned` is
|
|
643
|
+
unchanged.
|
|
644
|
+
|
|
645
|
+
### Docs
|
|
646
|
+
|
|
647
|
+
- Sub-agents guide: how to let helpers write files (`capabilities={"web", "files"}` and
|
|
648
|
+
`allowed_paths`), the capability words, and asking at a checkpoint; FAQ entry.
|
|
649
|
+
|
|
650
|
+
## [3.6.1] - 2026-10-09
|
|
651
|
+
|
|
652
|
+
### Fixed
|
|
653
|
+
|
|
654
|
+
- A package that was installed but failed to import was reported as not installed. A user whose project
|
|
655
|
+
folder held a `google.py` saw "ChromaVectorStore requires the chromadb package" although chromadb was
|
|
656
|
+
installed: the file hid the real `google` package that chromadb's dependencies use, and the catch-all
|
|
657
|
+
hid that. The error now says which case it is. Not installed: it names the extra
|
|
658
|
+
(`python -m pip install "agentx-dev[chroma]"`) and the Python that is running. Installed but failing:
|
|
659
|
+
it shows the real error and common causes (a file named like a library, a version conflict) and no
|
|
660
|
+
install command. Applies to `ChromaVectorStore`, `QdrantVectorStore`, `PgVectorStore`, `MCPClient`,
|
|
661
|
+
`OTelHook` and `Claude`.
|
|
662
|
+
|
|
663
|
+
### Docs
|
|
664
|
+
|
|
665
|
+
- Install pages list the `chroma`, `qdrant` and `pgvector` extras and say why a plain install leaves
|
|
666
|
+
them out; the troubleshooting page covers both messages.
|
|
667
|
+
|
|
668
|
+
## [3.6.0] - 2026-10-09
|
|
669
|
+
|
|
670
|
+
Sub-agents: the Supervisor can create its own helpers. Opt-in, with the behavior changes listed below.
|
|
671
|
+
|
|
672
|
+
### Added
|
|
673
|
+
|
|
674
|
+
- **Inline helpers**: a plan step can carry `new_agent` (name, free-form
|
|
675
|
+
instructions, tools). The helper runs the step, is reused by name, and is
|
|
676
|
+
discarded when the run ends. Recovery plans can define a replacement for a
|
|
677
|
+
stuck step.
|
|
678
|
+
- **`delegate` tool**: specialists hand a side job to a fresh sub-agent (clean
|
|
679
|
+
context) and get a short summary back; refusals are plain text and a helper
|
|
680
|
+
that fails comes back as a tool error the caller's stuck logic sees.
|
|
681
|
+
`AgentRunner(..., delegation=SpawnConfig(...))` gives a standalone runner the
|
|
682
|
+
same tool; `runner.spawned` lists what it created.
|
|
683
|
+
- **One ceiling**: `SpawnConfig(capabilities=..., tools=..., max_spawns=...,
|
|
684
|
+
max_depth=...)` bounds every helper; spawns inside it need no approval and
|
|
685
|
+
over-asks are clipped, not fatal. New read-only `files_read` preset.
|
|
686
|
+
- **`AsyncSupervisor(spawn_config=...)`**: async parity; spawned steps run in
|
|
687
|
+
parallel. `AsyncAgentRunner` gained `system_addendum=`.
|
|
688
|
+
- `SupervisorResult.spawned`, `spawn` and `delegate_result` stream events,
|
|
689
|
+
`AgentSpec`, `ToolRegistry.unregister`, `AgentRunner.add_tool` /
|
|
690
|
+
`remove_tool`.
|
|
691
|
+
- New guide: Sub-agents.
|
|
692
|
+
- **Ask the operator**: `Supervisor(ask_user=True | function)` lets the planner ask for facts the task
|
|
693
|
+
leaves out and gives every agent an `ask_user` tool. `True` uses a built-in asker that works in
|
|
694
|
+
notebooks (the input box), terminals and IDE consoles and never blocks a headless process; a
|
|
695
|
+
function routes questions through your own channel (chatbot, backend). `max_questions`,
|
|
696
|
+
`ask_timeout`, `SupervisorResult.asked`, `question`/`answer` events,
|
|
697
|
+
`agentx_dev.ask_human_tool()`; waiting on the operator pauses the persistent deadline. The built-in prompt names who is asking (`[planner]` or the agent's name) with a raised-hand marker. Each agent
|
|
698
|
+
may ask once per run (the planner may ask several questions in its one round), and after the
|
|
699
|
+
operator answers the planner is told to do the work instead of planning another question. The
|
|
700
|
+
helper-creation instructions now say a helper can only look things up with the tools it is given
|
|
701
|
+
(this applies to every Supervisor that enables spawning).
|
|
702
|
+
- **Supervisor long-term memory**: `Supervisor(memory=store)` (any vector store: `VectorStore`, Chroma,
|
|
703
|
+
Qdrant, pgvector) gives the planner and every dispatched step the few relevant stored items as a
|
|
704
|
+
FROM MEMORY block, saves each operator answer and each completed run's final answer, and answers an
|
|
705
|
+
exact repeat of an answered question from memory (a run whose synthesis was cut off by the budget
|
|
706
|
+
saves nothing). `memory_top_k`, `memory_min_score`,
|
|
707
|
+
`memory_write`, `SupervisorResult.memory`, `memory` stream events.
|
|
708
|
+
- `web_fetch_tool(text_only=True)`: return an HTML page as readable text (scripts, styles and tags
|
|
709
|
+
removed) instead of the first 50,000 characters of raw markup, which on a modern page can hold no
|
|
710
|
+
content at all. Default off (unchanged for existing agents); the helpers a Supervisor builds use it.
|
|
711
|
+
Helpers are also told not to answer with a script unless they have a code tool, and to say what
|
|
712
|
+
they tried and what was missing.
|
|
713
|
+
- `SpawnConfig(max_iterations=15)`: the step limit for the helpers a Supervisor creates (it was fixed at 15).
|
|
714
|
+
Registered specialists keep the limit set on their own `AgentRunner`. Helpers are told their step limit
|
|
715
|
+
so they can finish with a Final_Answer in time.
|
|
716
|
+
- A helper that runs out of steps (`iteration_limit`) is retried with a note that it ran out, a short
|
|
717
|
+
recap of what it did, and an instruction to do fewer lookups; other failures keep the generic retry note.
|
|
718
|
+
- The unknown-tool error now lists the valid tools and says that a response-format name such as `React_`
|
|
719
|
+
is not a tool, so a model that called one can recover.
|
|
720
|
+
- `RunBudget` now carries a `PausableClock` (it holds a lock), so a `RunBudget` can no longer be
|
|
721
|
+
deep-copied or pickled.
|
|
722
|
+
|
|
723
|
+
### Changed
|
|
724
|
+
|
|
725
|
+
- A spawned agent no longer stays on the supervisor after the run.
|
|
726
|
+
- A spawn is no longer refused (or rerouted) because an existing specialist
|
|
727
|
+
has the same tools. The `spawn` event gained `origin`, `tools`, `dropped`,
|
|
728
|
+
`reused` and `refused`; `rerouted_from` is always `None`.
|
|
729
|
+
- A `Supervisor` / `AsyncSupervisor` with `persistence=` and no `spawn_config`
|
|
730
|
+
now allows helpers with web and read-only file access (`max_spawns=6`) and
|
|
731
|
+
gives its specialists the `delegate` tool. Pass
|
|
732
|
+
`spawn_config=SpawnConfig(enabled=False)` to opt out.
|
|
733
|
+
- A legacy spawn config (`enabled=True` with neither `capabilities` nor
|
|
734
|
+
`tools`) keeps the 3.5 approval flow and gives specialists no `delegate`
|
|
735
|
+
tool; set a ceiling to turn delegation on.
|
|
736
|
+
- The planner prompt teaches `new_agent` instead of the `__spawn__` step; old
|
|
737
|
+
`__spawn__` steps still work, but their `name` must now match
|
|
738
|
+
`[A-Za-z0-9_-]{1,40}`.
|
|
739
|
+
- Agent names starting with `__` and names of the form `delegate_<N>` are
|
|
740
|
+
reserved: a `new_agent` that uses one is dropped by plan sanitization.
|
|
741
|
+
- `SpawnConfig.max_spawns` is `Optional[int]` (unset: 3, or 6 in ceiling mode).
|
|
742
|
+
`SUPERVISOR_SPAWN_INSTRUCTION` was removed.
|
|
743
|
+
|
|
744
|
+
## [3.5.0] - 2026-10-04
|
|
745
|
+
|
|
746
|
+
Agents that keep working through errors. Opt-in; one default change (below).
|
|
747
|
+
|
|
748
|
+
### Added
|
|
749
|
+
|
|
750
|
+
- **Persistent mode** (`persistence=Persistence(...)`) for `AgentRunner`,
|
|
751
|
+
`AsyncAgentRunner`, `Supervisor` and `AsyncSupervisor`. When a run gets
|
|
752
|
+
stuck (the same call repeated, or tool errors in a row) it is
|
|
753
|
+
asked, in four escalating steps, to find the root cause, try a
|
|
754
|
+
different approach, probe its assumptions, and say what blocks it,
|
|
755
|
+
before it may stop; real progress resets the ladder. Runs stop on
|
|
756
|
+
`max_minutes` or the model's cost cap, not a step count.
|
|
757
|
+
- **Context compaction** for long runs: old turns become notes, the task
|
|
758
|
+
(with its attachments) and recent turns are untouched, and the failed
|
|
759
|
+
attempts are kept word for word.
|
|
760
|
+
- **Patient retries**: transient provider errors (429, 5xx, timeouts,
|
|
761
|
+
connection errors) are retried with backoff until the deadline; auth
|
|
762
|
+
and invalid-request errors still raise at once.
|
|
763
|
+
- **Supervisors replan**: a step that does not finish is replanned
|
|
764
|
+
around under one shared deadline; finished steps are kept and can be
|
|
765
|
+
depended on. A recovery step names the steps it redoes
|
|
766
|
+
(`"replaces": [...]`); a failed step counts as resolved only when its
|
|
767
|
+
replacement finishes. New `replan`, `reflect`, `compact` and `budget`
|
|
768
|
+
stream events.
|
|
769
|
+
- **Outcomes**: every completion has `.outcome` and (persistent runs)
|
|
770
|
+
`.progress`; `SubtaskResult` and `SupervisorResult` have `.outcome`.
|
|
771
|
+
- New guide: Long-running agents.
|
|
772
|
+
|
|
773
|
+
### Changed
|
|
774
|
+
|
|
775
|
+
- A `Supervisor`/`AsyncSupervisor` now treats a specialist that finished
|
|
776
|
+
with an outcome other than `done` (for example `iteration_limit`) as a
|
|
777
|
+
failed attempt: it retries up to `max_subtask_retries` with the reason
|
|
778
|
+
fed back, then flags the step. Previously such a result was accepted
|
|
779
|
+
as a success. This applies without `persistence`. Consequences: each
|
|
780
|
+
give-up costs one extra attempt by default (`max_subtask_retries=1`);
|
|
781
|
+
the flagged step's recap is no longer passed to downstream
|
|
782
|
+
specialists; steps that `depends_on` it are skipped; and
|
|
783
|
+
`SupervisorResult.outcome` is `partial` or `stuck`.
|
|
784
|
+
- While `persistence` is set the tool-result cache is off, and the
|
|
785
|
+
iteration cap is `persistence.max_turns`.
|
|
786
|
+
|
|
787
|
+
## [3.4.3] - 2026-09-26
|
|
788
|
+
|
|
789
|
+
Two path-resolution fixes for Windows sandboxes. Non-breaking.
|
|
790
|
+
|
|
791
|
+
### Fixed
|
|
792
|
+
|
|
793
|
+
- **`full_access(["/workspace"])` now means the project's workspace
|
|
794
|
+
folder, not the drive root.** On Windows a rooted path without a drive
|
|
795
|
+
letter resolves to the current drive, so the sandbox became
|
|
796
|
+
`C:\workspace`; `auto_create_paths` then created it, and an agent
|
|
797
|
+
asked for `workspace/spam.csv` searched an empty directory outside the
|
|
798
|
+
project and reported the file missing. Slash-rooted sandbox roots are
|
|
799
|
+
now re-rooted at the running project, matching what a leading slash
|
|
800
|
+
already meant for tool arguments. POSIX absolute paths (`/workspace`
|
|
801
|
+
as a Docker mount) and drive-qualified paths are untouched.
|
|
802
|
+
- **`workspace/report.md` no longer nests into
|
|
803
|
+
`<workspace>/workspace/report.md`.** The `workspace` field documented
|
|
804
|
+
this pass-through, but only the `./workspace/...` spelling had it.
|
|
805
|
+
Models name the workspace folder constantly, so the missing case hit
|
|
806
|
+
often; new files still resolve into the workspace as before.
|
|
807
|
+
|
|
808
|
+
## [3.4.2] - 2026-09-19
|
|
809
|
+
|
|
810
|
+
`media=` everywhere, CSV / text / Excel attachments, and structured output
|
|
811
|
+
on models that need the Responses API. Non-breaking.
|
|
812
|
+
|
|
813
|
+
### Added
|
|
814
|
+
|
|
815
|
+
- **`media=` on chat models and structured output.** `llm.invoke`,
|
|
816
|
+
`llm.ainvoke`, and `with_structured_output(...).invoke` / `.ainvoke` take
|
|
817
|
+
`media=[...]` (paths, URLs, `Media`, part dicts), appended to the last
|
|
818
|
+
user turn, matching the agent runner. Media inside `content` lists
|
|
819
|
+
already worked; this adds the argument.
|
|
820
|
+
- **CSV, text files, and Excel in `media=`.** Text-like files (CSV, TSV,
|
|
821
|
+
TXT, Markdown, JSON, XML, YAML, HTML) are read as text -- exactly as
|
|
822
|
+
written, UTF-8 or Windows-1252, CRLF normalised -- and spreadsheets
|
|
823
|
+
(.xlsx / .xls / .ods) are converted with pandas (`dtype=object`, so
|
|
824
|
+
cell values aren't re-inferred) to one CSV block per sheet. Both reach
|
|
825
|
+
GPT (chat and Responses) and Claude as a labelled text block, which
|
|
826
|
+
every model accepts. New `Media.text()` / `Media.spreadsheet()`; a
|
|
827
|
+
100,000-character cap (`max_chars=`, `truncate=True`). pandas and
|
|
828
|
+
openpyxl are now core dependencies, so spreadsheets work after a plain
|
|
829
|
+
`pip install agentx-dev` (`[excel]` still works). Date cells without a
|
|
830
|
+
time are written as dates, not `... 00:00:00`.
|
|
831
|
+
- **Media docs section.** Its own nav group: Overview & flow, With chat
|
|
832
|
+
models, With agents, Reference (constructors, per-provider wire
|
|
833
|
+
formats, errors). `docs/guides/media.md` is now a pointer to it.
|
|
834
|
+
|
|
835
|
+
### Fixed
|
|
836
|
+
|
|
837
|
+
- **`with_structured_output` on models that need the Responses API.**
|
|
838
|
+
Structured output forces a single tool; if the Responses API rejects a
|
|
839
|
+
forced tool choice for the model, `GPT` retries with
|
|
840
|
+
`tool_choice="required"` (the same tool when there is only one), then
|
|
841
|
+
`"auto"`, remembering it per model. `StructuredOutputRunnable` also
|
|
842
|
+
accepts a JSON text answer that validates against the schema.
|
|
843
|
+
- **`.txt` media failed on Claude.** 3.4.0 stored text files as base64
|
|
844
|
+
documents, which Anthropic only accepts for PDFs. Text files now go
|
|
845
|
+
out as text; older base64 text parts in saved histories are decoded
|
|
846
|
+
on the way out.
|
|
847
|
+
- **Unsupported files reached the provider.** `Media.document()`
|
|
848
|
+
accepted any file type -- including CSV, which Windows' own type table
|
|
849
|
+
labels `application/vnd.ms-excel` -- and sent it as an attachment
|
|
850
|
+
neither provider accepts. Word, PowerPoint and archives now raise
|
|
851
|
+
`ValueError` with the fix before any request; a text file or
|
|
852
|
+
spreadsheet by URL raises too, since only PDFs can be fetched by URL.
|
|
853
|
+
- **Misleading WARNING on recovered calls.** The retry wrapper logged
|
|
854
|
+
every non-retryable 400 at WARNING before the caller could recover, so a
|
|
855
|
+
call that switched to the Responses API and succeeded still printed the
|
|
856
|
+
original "Function tools with reasoning_effort are not supported" text.
|
|
857
|
+
Now DEBUG; recoveries log their own WARNING and real failures log ERROR.
|
|
858
|
+
|
|
859
|
+
## [3.4.1] - 2026-09-18
|
|
860
|
+
|
|
861
|
+
Tool calls on OpenAI models that require the Responses API.
|
|
862
|
+
|
|
863
|
+
### Fixed
|
|
864
|
+
|
|
865
|
+
- **Tool calls failed on models that need the Responses API.** Some
|
|
866
|
+
OpenAI models (e.g. `gpt-6-astra`) reject function tools combined with
|
|
867
|
+
reasoning on `/v1/chat/completions`, and the error's suggested fix --
|
|
868
|
+
`reasoning_effort='none'` -- is itself rejected by those models, so
|
|
869
|
+
3.4.0's parameter adaptation had nowhere to go. `GPT` now switches a
|
|
870
|
+
model's tool calls to `/v1/responses` when the provider says to,
|
|
871
|
+
retries, and remembers it (WARNING-logged). History is translated only
|
|
872
|
+
at the wire (`function_call` / `function_call_output` items paired by
|
|
873
|
+
`call_id`), so agents, `tool_calls`, and completions are unchanged.
|
|
874
|
+
Learned parameter fixes carry across endpoints (`reasoning.effort` maps
|
|
875
|
+
to `reasoning_effort`). Requests are sent with `store=False`, matching
|
|
876
|
+
chat completions. Plain text calls and streaming stay on chat
|
|
877
|
+
completions.
|
|
878
|
+
- Requests chat completions can't express (a document URL) go to the
|
|
879
|
+
Responses API, which accepts `file_url`, instead of raising.
|
|
880
|
+
|
|
881
|
+
### Added
|
|
882
|
+
|
|
883
|
+
- `GPT(use_responses_api=None | True | False)`: automatic (default),
|
|
884
|
+
always, or never.
|
|
885
|
+
|
|
886
|
+
## [3.4.0] - 2026-09-18
|
|
887
|
+
|
|
888
|
+
Media input for GPT and Claude, models that adapt to each generation's
|
|
889
|
+
parameter support, and workspace-rooted paths. Non-breaking; two
|
|
890
|
+
defaults changed -- see docs/guides/upgrading.md.
|
|
891
|
+
|
|
892
|
+
### Added
|
|
893
|
+
|
|
894
|
+
- **Media input for GPT and Claude.** New `Media` type
|
|
895
|
+
(`Media.image()`, `.document()`, `.audio()`, `.from_path()`,
|
|
896
|
+
`.from_url()`, `.from_bytes()`) and a `media=[...]` argument on
|
|
897
|
+
`AgentRunner.invoke` / `stream` and `AsyncAgentRunner.ainvoke` /
|
|
898
|
+
`astream`. Items can be paths, URLs, `Media`, or content-part dicts
|
|
899
|
+
in either provider's native shape; each model renders them in its
|
|
900
|
+
own wire format (OpenAI `image_url` / `file` / `input_audio`,
|
|
901
|
+
Anthropic `image` / `document`). Unsupported combinations -- audio to
|
|
902
|
+
Claude, a document URL to GPT -- raise `ValueError` before any
|
|
903
|
+
request. Media is stored as plain JSON dicts, so `completion.history`
|
|
904
|
+
and `Session.save()` keep working. Guide: docs/guides/media.md.
|
|
905
|
+
|
|
906
|
+
- **Models adapt to per-generation parameter support.** `GPT` and
|
|
907
|
+
`Claude` read a parameter-compatibility 400 (which names the
|
|
908
|
+
parameter and, for enums, the allowed values), make the smallest
|
|
909
|
+
change it asks for, retry, and remember it per model: an unsupported
|
|
910
|
+
`reasoning_effort` moves to the nearest supported value, `max_tokens`
|
|
911
|
+
becomes `max_completion_tokens` on reasoning models (up front for
|
|
912
|
+
known families), parameters a model lacks are dropped, and Claude's
|
|
913
|
+
`max_tokens` is clamped to the model's cap. Logged at WARNING; other
|
|
914
|
+
400s raise unchanged. `adapt_params=False` opts out.
|
|
915
|
+
|
|
916
|
+
- **`Claude` gains `top_p`, `top_k`, `thinking`, `stop_sequences`.**
|
|
917
|
+
|
|
918
|
+
### Changed
|
|
919
|
+
|
|
920
|
+
- `Claude(temperature=)` defaults to `None` (not sent; the API default is
|
|
921
|
+
the old 1.0). Always sending it conflicted with `top_p` on newer models
|
|
922
|
+
and with extended thinking.
|
|
923
|
+
- `GPT(reasoning_effort=)` accepts any string, not just
|
|
924
|
+
`none | low | medium | high`, so `minimal` and `xhigh` are expressible.
|
|
925
|
+
|
|
926
|
+
### Fixed
|
|
927
|
+
|
|
928
|
+
- **`reasoning_effort="none"` crashed on models that don't take it**
|
|
929
|
+
("Unsupported value: 'reasoning_effort' does not support 'none' with
|
|
930
|
+
this model"). The models docs even recommended `"none"` as a fix for
|
|
931
|
+
tool-calling conflicts; that advice is gone.
|
|
932
|
+
- **Workspace-rooted paths resolved to the drive root.** With a workspace
|
|
933
|
+
set, `/bruce.jpeg` became `C:\bruce.jpeg` on Windows (a leading slash
|
|
934
|
+
with no drive means the current drive's root) and the sandbox rejected
|
|
935
|
+
it. A leading slash now means the workspace root; the re-rooted path
|
|
936
|
+
is still sandbox-checked, so traversal is rejected as before.
|
|
937
|
+
- **`run_python` ran in the host process's directory, not the
|
|
938
|
+
workspace.** It passed no `cwd`, so `open("bruce.jpeg")` looked
|
|
939
|
+
wherever the program was launched from. It now starts in the workspace,
|
|
940
|
+
matching `run_shell`.
|
|
941
|
+
- **Message-list input stringified media into the prompt.** A user turn
|
|
942
|
+
whose content was a list (text + image) went through `str()`, so an
|
|
943
|
+
attached image reached the model as base64 text. Text and media are
|
|
944
|
+
now split, and earlier turns keep their list content.
|
|
945
|
+
- **Claude returned the wrong block with extended thinking.** It read
|
|
946
|
+
`response.content[0].text`; with thinking the first block is a
|
|
947
|
+
`thinking` block. Text blocks are now joined.
|
|
948
|
+
- `GPT.Initialize` / `stream_text` / `astream_text` now run messages
|
|
949
|
+
through the same OpenAI translator as `call_with_tools`.
|
|
950
|
+
|
|
951
|
+
## [3.3.1] - 2026-09-09
|
|
952
|
+
|
|
953
|
+
Correctness fixes in the agent loop, found by driving it with a
|
|
954
|
+
scripted model rather than reading it. No API changes; every fix
|
|
955
|
+
replaces a silent wrong answer or a runaway cost with correct
|
|
956
|
+
behaviour. Nine regression tests ship with them
|
|
957
|
+
(`tests/test_agent_loop_flaws.py`).
|
|
958
|
+
|
|
959
|
+
### Fixed
|
|
960
|
+
|
|
961
|
+
- **Native binding had no loop-level spiral guard.** The repeat breaker
|
|
962
|
+
sat after the native branch's `continue`, so `bind_tools_natively`
|
|
963
|
+
runs never reached it: a model stuck re-issuing one call burned the
|
|
964
|
+
full `max_iterations` (20 LLM turns where text mode stopped at 3) and
|
|
965
|
+
returned the "hit max_iterations" recap quoting the dup-guard's own
|
|
966
|
+
warning text instead of the data.
|
|
967
|
+
|
|
968
|
+
- **`respond` batched with real tool calls dropped them.** A model that
|
|
969
|
+
emitted "do X, and here is my answer" in one turn had X silently
|
|
970
|
+
discarded -- never dispatched, never in `completion.tool_calls`, no
|
|
971
|
+
error. The batch now runs and the answer defers one turn.
|
|
972
|
+
|
|
973
|
+
- **`completion.history` could not be replayed as `chat_history`.** The
|
|
974
|
+
copy filter kept only truthy `{role, content}`, which dropped every
|
|
975
|
+
tool-calling assistant turn (`content=""`), stripped `tool_call_id`
|
|
976
|
+
off `role="tool"` messages, and replayed a stored system prompt on top
|
|
977
|
+
of the fresh one -- so a follow-up call sent two system messages, zero
|
|
978
|
+
assistant turns, and an orphaned tool message providers reject. Fixed
|
|
979
|
+
in both runners.
|
|
980
|
+
|
|
981
|
+
- **A sync `AgentRunner` silently ignored async tools.** `known_tools`
|
|
982
|
+
unioned only the two sync registry tables while the registry accepted,
|
|
983
|
+
listed and prompt-advertised async ones, so calling one fell through
|
|
984
|
+
to implicit-final and returned `action_input` as the answer with an
|
|
985
|
+
empty `tool_calls` list. Now unions all four, matching the async
|
|
986
|
+
runner; the sync dispatcher already returns a clear `ToolError`, so
|
|
987
|
+
the silent wrong answer became an actionable one.
|
|
988
|
+
|
|
989
|
+
- **`AsyncAgentRunner` regressed the function-calling message shape.**
|
|
990
|
+
It appended its FC turn as raw JSON text with no `tool_calls` block
|
|
991
|
+
and never set `_last_function_call_id`, so observations went back as
|
|
992
|
+
`role="user"` -- text-mode shape while in FC mode, losing provider
|
|
993
|
+
correlation and the cached prefix.
|
|
994
|
+
|
|
995
|
+
- **A `BaseException` from a specialist orphaned its siblings.**
|
|
996
|
+
`CancelledError` and `KeyboardInterrupt` are not `Exception`, so
|
|
997
|
+
`_run_subtask`'s handler never saw them; `t.result()` re-raised and
|
|
998
|
+
left sibling tasks running unowned, with in-flight LLM calls still
|
|
999
|
+
billing. The scheduler now owns its tasks in a `try`/`finally`, which
|
|
1000
|
+
also covers a consumer that stops iterating `astream` early.
|
|
1001
|
+
|
|
1002
|
+
- **The tool cache collided on tool name.** It is a process-wide
|
|
1003
|
+
singleton keyed on `(tool_name, args)` with no record of which
|
|
1004
|
+
implementation ran, so two runners whose tools merely share a name --
|
|
1005
|
+
`search`, `fetch`, `query`, routine across Supervisor specialists --
|
|
1006
|
+
served each other's results and the second function never ran. Keys
|
|
1007
|
+
now fold in the callable's `module.qualname`, so same-name /
|
|
1008
|
+
different-implementation misses while genuinely identical tools still
|
|
1009
|
+
share, including across processes for the disk-backed `FileCache`.
|
|
1010
|
+
|
|
1011
|
+
### Docs
|
|
1012
|
+
|
|
1013
|
+
- Streaming documentation described parameters the API rejects.
|
|
1014
|
+
`AsyncAgentRunner.astream()` and `HandoffCoordinator.stream()` take no
|
|
1015
|
+
`stream_tokens` (only the sync `AgentRunner.stream()` does), and every
|
|
1016
|
+
`text_delta` example was built on a default model where
|
|
1017
|
+
`use_function_calling` auto-detects to `True` and yields zero deltas.
|
|
1018
|
+
`simple_stream` was documented with the wrong signature entirely.
|
|
1019
|
+
Added a table naming exactly which stream methods accept what.
|
|
1020
|
+
|
|
1021
|
+
### Tests
|
|
1022
|
+
|
|
1023
|
+
- Repaired 20 stale text-mode tests. `MockModel` defined
|
|
1024
|
+
`call_with_tools` unconditionally, so 3.1.7's auto-detect routed every
|
|
1025
|
+
text script down the function-calling path and the runner returned
|
|
1026
|
+
`""`. The mock now advertises the capability only when scripted for
|
|
1027
|
+
it. Suite: 206 passed, 3 skipped.
|
|
1028
|
+
|
|
1029
|
+
## [3.3.0] - 2026-08-19
|
|
1030
|
+
|
|
1031
|
+
Dependency DAGs for the Supervisor. Plans declare which steps consume
|
|
1032
|
+
which, and the executor derives ordering, parallelism, AND context
|
|
1033
|
+
routing from those edges — unifying the old split where sequential
|
|
1034
|
+
mode had threading but no parallelism and concurrent mode had
|
|
1035
|
+
parallelism but no threading. Design: docs/design/3.3-depends-on-dag.md.
|
|
1036
|
+
|
|
1037
|
+
### Added
|
|
1038
|
+
|
|
1039
|
+
- **`depends_on` plan steps.** Every plan step now carries an `id`;
|
|
1040
|
+
a step that consumes an earlier step's output lists that id in
|
|
1041
|
+
`depends_on`. Sync `Supervisor` executes in stable topological
|
|
1042
|
+
order; `AsyncSupervisor` runs a completion-driven scheduler where a
|
|
1043
|
+
step starts the MOMENT its dependencies finish (not on wave
|
|
1044
|
+
barriers) and independent steps run concurrently. In DAG mode each
|
|
1045
|
+
step is threaded ONLY its direct dependencies' results — explicit
|
|
1046
|
+
routing instead of "everything prior", which also stops the
|
|
1047
|
+
per-entry context budget shrinking as plans grow.
|
|
1048
|
+
|
|
1049
|
+
- **Plan sanitization that never fails a run.** Missing/duplicate ids
|
|
1050
|
+
auto-assigned, unknown dependencies dropped (a planner typo degrades
|
|
1051
|
+
to a root step, not a dead run), self-deps dropped, cycles broken
|
|
1052
|
+
deterministically (back-edge in plan order), spawn steps cannot be
|
|
1053
|
+
depended on. If repairs were needed, the plan is re-requested once
|
|
1054
|
+
(`max_plan_retries`, default 1) with the repair warnings appended;
|
|
1055
|
+
the sanitized original is kept when the retry is no better.
|
|
1056
|
+
|
|
1057
|
+
- **Failure cascade + `skipped` flag.** A step whose dependency failed
|
|
1058
|
+
(after retries / success-check) is skipped, transitively, with
|
|
1059
|
+
`SubtaskResult.skipped=True` and an error naming the failed dep.
|
|
1060
|
+
Independent branches keep running; synthesis runs over what
|
|
1061
|
+
succeeded. Skipped steps never dispatch — no tokens burned
|
|
1062
|
+
downstream of garbage.
|
|
1063
|
+
|
|
1064
|
+
- **`skip_when` conditional execution.** A step may declare
|
|
1065
|
+
`{"step": <direct dep id>, "field": <typed output field>, "is": <value>}`;
|
|
1066
|
+
evaluated in Python (no LLM call) against the dependency's 3.2
|
|
1067
|
+
structured output, dotted paths supported, strictly FAIL-OPEN (any
|
|
1068
|
+
doubt → the step runs). Condition-skips do NOT cascade — dependents
|
|
1069
|
+
treat them as empty successes ("retrieval unnecessary" is not
|
|
1070
|
+
"answering impossible").
|
|
1071
|
+
|
|
1072
|
+
- **`Specialist` registry entries.** `agents={}` now also accepts
|
|
1073
|
+
`Specialist(description, runner, depends_on=[...names...],
|
|
1074
|
+
output_schema=..., when_to_use=...)`. The extras render into the
|
|
1075
|
+
planning catalog (`typically after:` / `returns: Schema(fields)` /
|
|
1076
|
+
`use when:`) so the planner can write real dependency graphs and
|
|
1077
|
+
`skip_when` conditions against actual field names. `depends_on`
|
|
1078
|
+
here is a planner HINT, never an execution constraint (the same
|
|
1079
|
+
specialist can appear twice in one plan; step-ids disambiguate).
|
|
1080
|
+
Classic `(description, runner)` tuples keep working — they are
|
|
1081
|
+
wrapped internally, and `Specialist` tuple-unpacks for older code.
|
|
1082
|
+
|
|
1083
|
+
- **`AsyncSupervisor(max_parallel=N)`.** Caps concurrent sub-tasks for
|
|
1084
|
+
rate-limited deployments. `sequential=True` is now sugar for
|
|
1085
|
+
`max_parallel=1` with all-prior threading.
|
|
1086
|
+
|
|
1087
|
+
- **`SubtaskResult.step_id` / `.depends_on` / `.skipped`** and a
|
|
1088
|
+
`step_id` field on `dispatch` / `subtask_result` stream events.
|
|
1089
|
+
|
|
1090
|
+
### Backward compatibility
|
|
1091
|
+
|
|
1092
|
+
A plan where NO step declares `depends_on` runs with byte-identical
|
|
1093
|
+
legacy semantics: sync + async-sequential thread all prior results in
|
|
1094
|
+
plan order; async-concurrent runs everything at once with no
|
|
1095
|
+
threading. Verified by regression tests against the 3.2 behaviour.
|
|
1096
|
+
|
|
1097
|
+
## [3.2.0] - 2026-08-13
|
|
1098
|
+
|
|
1099
|
+
Typed multi-agent pipelines. Specialists can now declare a Pydantic
|
|
1100
|
+
output schema once and pass validated instances to each other through
|
|
1101
|
+
the Supervisor, instead of downstream agents re-parsing prose.
|
|
1102
|
+
|
|
1103
|
+
### Added
|
|
1104
|
+
|
|
1105
|
+
- **`output_schema` on the `AgentRunner` / `AsyncAgentRunner`
|
|
1106
|
+
constructor.** Declare the runner's output shape once
|
|
1107
|
+
(`AgentRunner(..., output_schema=QueryIntent)`) instead of passing it
|
|
1108
|
+
on every call or describing JSON in the prompt. A per-call
|
|
1109
|
+
`output_schema=` still wins when both are set. `None` keeps the exact
|
|
1110
|
+
pre-3.2 behaviour: no coercion, `completion.output` stays `None`.
|
|
1111
|
+
|
|
1112
|
+
- **Schema coercion via forced native function calling.** When a schema
|
|
1113
|
+
is in play, the final answer is converted by forcing a provider-native
|
|
1114
|
+
tool call against the schema (constrained decoding), not by regexing
|
|
1115
|
+
JSON out of prose. The ReAct loop itself is untouched: tool selection
|
|
1116
|
+
and intermediate reasoning run exactly as before, and the coercion
|
|
1117
|
+
happens once, after the loop finishes. Models without a
|
|
1118
|
+
`call_with_tools` implementation fall back to the previous text-JSON
|
|
1119
|
+
parsing, so custom `BaseChatModel` subclasses keep working.
|
|
1120
|
+
`completion.content` keeps the human-readable answer alongside
|
|
1121
|
+
`completion.output` in every case.
|
|
1122
|
+
|
|
1123
|
+
- **`SubtaskResult.output`.** The Supervisor now preserves each
|
|
1124
|
+
specialist's validated Pydantic instance next to its `content` text.
|
|
1125
|
+
Consumers that only read `content` are unaffected.
|
|
1126
|
+
|
|
1127
|
+
- **Structured specialist-to-specialist handoff.** When an earlier step
|
|
1128
|
+
produced typed output, `_build_augmented_query` serializes it into the
|
|
1129
|
+
next specialist's context as a labelled JSON block
|
|
1130
|
+
(`STRUCTURED OUTPUT (QueryIntent): {...}`) followed by the summary
|
|
1131
|
+
text, so downstream steps parse fields rather than interpreting
|
|
1132
|
+
sentences like `INTENT: ... SEARCH_QUERY: ...`.
|
|
1133
|
+
|
|
1134
|
+
- **`vector_search_tool` pipeline options.** New kwargs:
|
|
1135
|
+
`max_text_chars` (default 500; pass `0` for full untruncated passages,
|
|
1136
|
+
which a reranker judging evidence actually needs) and
|
|
1137
|
+
`structured_output` (default False; when True the tool returns a JSON
|
|
1138
|
+
array of `{id, text, vector_score, metadata}` instead of the
|
|
1139
|
+
human-formatted list). Defaults preserve existing behaviour byte-for-
|
|
1140
|
+
byte.
|
|
1141
|
+
|
|
1142
|
+
### Fixed
|
|
1143
|
+
|
|
1144
|
+
- **Supervisor planning prompt contradicted the execution engine.** The
|
|
1145
|
+
planner rule said sub-agents "do NOT see previous steps' output" and
|
|
1146
|
+
discouraged dependency chains, but the dispatcher has threaded prior
|
|
1147
|
+
findings into every step since `_build_augmented_query` shipped.
|
|
1148
|
+
The rule now tells the planner that sequential steps receive earlier
|
|
1149
|
+
results (structured when available) and that chains like
|
|
1150
|
+
intent -> retrieval -> reranking are a good plan shape, while still
|
|
1151
|
+
requiring same-specialist steps to merge and banning report-only steps.
|
|
1152
|
+
|
|
1153
|
+
## [3.1.7] — 2026-07-27
|
|
1154
|
+
|
|
1155
|
+
### Changed
|
|
1156
|
+
|
|
1157
|
+
- **`use_function_calling` default flipped to auto-detect** on
|
|
1158
|
+
`AgentRunner` / `AsyncAgentRunner`. The parameter's default type is
|
|
1159
|
+
now `Optional[bool] = None`; `None` resolves to `True` when the
|
|
1160
|
+
model class overrides `BaseChatModel.call_with_tools` (both `GPT`
|
|
1161
|
+
and `Claude` do) and to `False` when it doesn't (or when
|
|
1162
|
+
`bind_tools_natively=True`). Callers passing `True`/`False`
|
|
1163
|
+
explicitly are unaffected. Rationale: text-mode ReAct requires the
|
|
1164
|
+
model to emit strict JSON with any long `action_input` string
|
|
1165
|
+
properly escaped — a 1200-word markdown draft with unescaped
|
|
1166
|
+
newlines or quotes reliably breaks `json.loads` and killed the run.
|
|
1167
|
+
Function-calling mode routes the parser through the SDK's typed
|
|
1168
|
+
channel so escaping is handled automatically. The historical
|
|
1169
|
+
default (`False`) was the fragile option; the new default matches
|
|
1170
|
+
what most users actually want.
|
|
1171
|
+
|
|
1172
|
+
### Fixed
|
|
1173
|
+
|
|
1174
|
+
- **Malformed parser JSON no longer crashes the run.** When the
|
|
1175
|
+
text-mode assistant response failed `json.loads` (typically because
|
|
1176
|
+
a long `action_input` string had unescaped `"`, `\n`, or backticks),
|
|
1177
|
+
the framework used to raise `JSONDecodeError` and unwind the whole
|
|
1178
|
+
invocation. The runner now (1) tries a regex-based salvage that
|
|
1179
|
+
extracts `{Thought, action, action_input}` from the raw text
|
|
1180
|
+
covering the common "outer envelope valid, inner string broke
|
|
1181
|
+
escaping" failure, and (2) if salvage fails, feeds a targeted fix
|
|
1182
|
+
hint back to the model (`"your last response was not valid JSON;
|
|
1183
|
+
emit …, escape newlines as \n"`) and continues the loop bounded
|
|
1184
|
+
by `max_iterations`. Exhaustion returns a clear framework message
|
|
1185
|
+
rather than an uncaught exception. Applied to both sync and async
|
|
1186
|
+
runners via a shared `_salvage_react_json` helper.
|
|
1187
|
+
The salvager's action-name regex is intentionally strict
|
|
1188
|
+
(`[A-Za-z_][A-Za-z0-9_.\- ]{0,79}`) so it can't hallucinate an
|
|
1189
|
+
"action" out of an unrelated `"key":"value"` pair inside malformed
|
|
1190
|
+
JSON.
|
|
1191
|
+
|
|
1192
|
+
- **Verbose trace in `bind_tools_natively` mode now prints tool
|
|
1193
|
+
name + args + response.** Previously native runs showed blank
|
|
1194
|
+
`[tool.call.start]` / `[tool.call.complete]` pairs (the
|
|
1195
|
+
observability layer fires them without the trace context), so you
|
|
1196
|
+
couldn't tell which tool the model actually invoked or what came
|
|
1197
|
+
back. The runner now prints `[tool] Invoking '<name>' with args:
|
|
1198
|
+
<input>` and `[tool] Response: <preview>` (or `[tool] Error: ...`
|
|
1199
|
+
when the dispatch raised) in the post-dispatch loop, matching the
|
|
1200
|
+
format text-mode and function-calling mode use. Mirrored to the
|
|
1201
|
+
async runner.
|
|
1202
|
+
|
|
1203
|
+
- **`web_fetch_tool(vector_store=...)` auto-ingests fetched pages into
|
|
1204
|
+
a vector store** instead of dumping raw HTML into the model's
|
|
1205
|
+
context. Fixes the TPM-limit trap: when a research agent fetches
|
|
1206
|
+
four articles in parallel (via ``multi_tool_use.parallel`` or
|
|
1207
|
+
native binding), the combined bodies can easily exceed 40k tokens
|
|
1208
|
+
and blow past a 30k TPM ceiling on the very next model call.
|
|
1209
|
+
New parameters on ``web_fetch_tool``:
|
|
1210
|
+
|
|
1211
|
+
| Kwarg | Default | Effect |
|
|
1212
|
+
|---|---|---|
|
|
1213
|
+
| ``vector_store`` | ``None`` | When set, each fetch is HTML-stripped, chunked with ``TextSplitter``, embedded via the store's embeddings, and added with ``{src: url, chunk_index, total_chunks}`` metadata. The tool response becomes a compact summary (URL, byte count, chunk count, 240-char preview) — NOT the raw body. The model then calls ``vector_search`` / ``Rag`` to pull only the passages it needs. |
|
|
1214
|
+
| ``chunk_size`` | ``1500`` | Characters per chunk when ``vector_store`` is set. Ignored otherwise. |
|
|
1215
|
+
| ``chunk_overlap`` | ``200`` | Overlap between adjacent chunks so a fact spanning a boundary is still retrievable. Ignored otherwise. |
|
|
1216
|
+
|
|
1217
|
+
Backwards-compatible: the positional ``cache_dir`` signature keeps
|
|
1218
|
+
working; `web_fetch_tool()` with no ``vector_store`` returns raw
|
|
1219
|
+
body as before. Ingest and cache_dir compose — enable both and get
|
|
1220
|
+
disk-cached full bodies AND searchable chunks. HTML stripping is
|
|
1221
|
+
minimal and dependency-free (regex-based: script/style blocks
|
|
1222
|
+
dropped whole, then tags stripped, whitespace collapsed) so the
|
|
1223
|
+
ingest path adds no new install dependency. On JSON/plain-text
|
|
1224
|
+
responses the stripper is a near no-op.
|
|
1225
|
+
|
|
1226
|
+
The observation returned to the model shows topical coverage --
|
|
1227
|
+
first, middle, and last chunk previews (up to 3 samples,
|
|
1228
|
+
deduplicated for short pages) -- so the model can tell what
|
|
1229
|
+
topics the page actually covers, not just the intro paragraph.
|
|
1230
|
+
Without this the model would only see the page's opening and
|
|
1231
|
+
wouldn't know to query for topics discussed later in the same
|
|
1232
|
+
page. Explicit instruction in the observation ("query with
|
|
1233
|
+
SPECIFIC keywords from the topics above; do NOT re-fetch; do
|
|
1234
|
+
NOT ask for the full body") steers the model toward the RAG path
|
|
1235
|
+
on follow-up turns.
|
|
1236
|
+
|
|
1237
|
+
- **`multi_tool_use.parallel` now reaches its dispatch path.**
|
|
1238
|
+
When GPT wanted to batch several tool calls into one turn (fetch N
|
|
1239
|
+
URLs concurrently, run M searches at once), it emitted OpenAI's
|
|
1240
|
+
synthetic `multi_tool_use.parallel` meta-tool. The registry's
|
|
1241
|
+
`_dispatch_multi_parallel` / `_adispatch_multi_parallel` handlers
|
|
1242
|
+
already knew how to unpack it, but the runner loop's known-tools
|
|
1243
|
+
guardrail rejected the name FIRST as unregistered — dumping the
|
|
1244
|
+
raw `{"tool_uses": [...]}` payload into the user-facing "final
|
|
1245
|
+
answer" and never invoking any of the nested calls. Added
|
|
1246
|
+
`multi_tool_use.parallel` to the recognized action set in both
|
|
1247
|
+
sync and async runners so the meta-tool flows through to dispatch
|
|
1248
|
+
and the existing unpackers run. Nested calls with the `functions.`
|
|
1249
|
+
prefix are normalized before dispatch (same as top-level FC
|
|
1250
|
+
calls), so the model can emit either shape.
|
|
1251
|
+
|
|
1252
|
+
- **`Permissions.full_access` / `read_only` auto-wrap a bare string.**
|
|
1253
|
+
Passing `full_access("./workspace")` used to iterate the string
|
|
1254
|
+
into 11 single-character "subtrees" (Python's `list("./workspace")`)
|
|
1255
|
+
— every path check silently rejected because no real path could
|
|
1256
|
+
ever match a `"."` or `"/"` "allowed subtree". The classmethod
|
|
1257
|
+
now detects a bare string and treats it as `[allowed_paths]`, so
|
|
1258
|
+
`full_access("./workspace")` does the intuitive thing (equivalent
|
|
1259
|
+
to `full_access(["./workspace"])` and auto-infers the workspace).
|
|
1260
|
+
Same fix on `read_only`. List inputs are unchanged.
|
|
1261
|
+
|
|
1262
|
+
- **`Permissions.full_access` now accepts (and auto-infers)
|
|
1263
|
+
`workspace`.** The classmethod set `allowed_paths` but not
|
|
1264
|
+
`workspace`, so short paths like `write_file(path="report.md")`
|
|
1265
|
+
resolved to CWD (outside the sandbox) and raised
|
|
1266
|
+
`PermissionError: access denied` — a landmine that every caller of
|
|
1267
|
+
`Permissions.full_access(["./workspace"])` hit sooner or later.
|
|
1268
|
+
New signature: `full_access(allowed_paths, *, workspace=None)`.
|
|
1269
|
+
When `workspace` isn't passed AND `allowed_paths` has exactly one
|
|
1270
|
+
entry, that path is auto-set as the workspace (the "project-scoped
|
|
1271
|
+
agent whose one allowed subtree IS its workspace" case, which is
|
|
1272
|
+
99% of use). Two or more paths stay ambiguous and require an
|
|
1273
|
+
explicit `workspace=` if short-path resolution is wanted. Pass an
|
|
1274
|
+
explicit `workspace=` string to override the auto-choice.
|
|
1275
|
+
Backwards-compatible on the positional signature; adds a keyword
|
|
1276
|
+
argument that existing callers didn't use.
|
|
1277
|
+
|
|
1278
|
+
## [3.1.5] — 2026-07-26
|
|
1279
|
+
|
|
1280
|
+
### Fixed
|
|
1281
|
+
|
|
1282
|
+
- **Text-mode tool results no longer use `role: "function"`.** In text
|
|
1283
|
+
mode (the default — no `use_function_calling`) the runner fed each tool
|
|
1284
|
+
observation back to the model as a `role: "function"` message. Newer
|
|
1285
|
+
OpenAI models reject that role outright (`400 … 'messages[N].role' does
|
|
1286
|
+
not support 'function' with this model`, e.g. gpt-5.x), and Anthropic
|
|
1287
|
+
never accepted it — text-mode multi-tool runs on Claude were latently
|
|
1288
|
+
broken too; older GPT models simply still tolerated the legacy role.
|
|
1289
|
+
Tool observations now go back as a plain `role: "user"` turn framed as
|
|
1290
|
+
`Observation: …`, which every provider and model generation accepts and
|
|
1291
|
+
which matches the ReAct template's own few-shot convention.
|
|
1292
|
+
Function-calling mode is unchanged (native `role: "tool"` +
|
|
1293
|
+
`tool_call_id`). The async runner was additionally emitting `function`
|
|
1294
|
+
unconditionally (even in FC mode); it now uses the same shared helper.
|
|
1295
|
+
|
|
1296
|
+
### Added
|
|
1297
|
+
|
|
1298
|
+
- **`subtask_success_check` on `Supervisor` / `AsyncSupervisor`.** Opt-in
|
|
1299
|
+
predicate `(SubtaskResult) -> bool | str` that decides whether a
|
|
1300
|
+
*returned* (non-raised) sub-task result is actually acceptable — the
|
|
1301
|
+
"ran fine but produced nothing useful" case a plain retry can't catch
|
|
1302
|
+
(a scraper that saved 0 links, an extractor that found nothing). Return
|
|
1303
|
+
`True` to accept, or `False`/a `str` reason to reject; a rejected
|
|
1304
|
+
result is retried like a raised error, with the reason fed back into the
|
|
1305
|
+
query, bounded by `max_subtask_retries`. After retries are exhausted the
|
|
1306
|
+
last result is returned with its `error` set (content preserved). A
|
|
1307
|
+
check that itself raises is treated as "accept" so a buggy predicate
|
|
1308
|
+
can't wedge the run. Default `None` keeps the exceptions-only behavior.
|
|
1309
|
+
New example `examples/robust_link_scraper.py` wires it together with a
|
|
1310
|
+
scraping `system_addendum` (parse relative+absolute hrefs, fall back to
|
|
1311
|
+
`sitemap.xml` on JS-rendered sites).
|
|
1312
|
+
|
|
1313
|
+
## [3.1.4] — 2026-07-26
|
|
1314
|
+
|
|
1315
|
+
### Fixed
|
|
1316
|
+
|
|
1317
|
+
- **A malformed-JSON tool argument no longer crashes the whole agent
|
|
1318
|
+
run.** When a model emitted a Python snippet or a Windows path as a
|
|
1319
|
+
tool-call argument — `re.findall(r'\d+')`, `C:\Users` — the `\d` / `\U`
|
|
1320
|
+
are illegal JSON escapes, and the OpenAI adapter's eager
|
|
1321
|
+
`json.loads(call.function.arguments)` raised `JSONDecodeError` and
|
|
1322
|
+
`raise`d it, unwinding the entire ReAct loop before the agent's own
|
|
1323
|
+
retry machinery could act. Under a Supervisor this surfaced as a bare
|
|
1324
|
+
`ERROR: Invalid \escape: line 1 column 598` and the sub-task was
|
|
1325
|
+
abandoned. Now:
|
|
1326
|
+
- `_parse_tool_arguments` repairs the common case (backslashes that
|
|
1327
|
+
don't begin a valid JSON escape are doubled), recovering `\d`,
|
|
1328
|
+
`\w`, `\s`, etc. with zero extra round-trips. A backslash before a
|
|
1329
|
+
valid-escape letter (`\b`, `\n`, …) remains ambiguous and is left
|
|
1330
|
+
as the escape — a documented limit.
|
|
1331
|
+
- When repair fails, `call_with_tools` returns a dedicated
|
|
1332
|
+
`invalid_tool_args` result and the loop feeds the error back as a
|
|
1333
|
+
retryable observation ("your arguments weren't valid JSON — escape
|
|
1334
|
+
backslashes and resend"), bounded by `max_iterations`, in all
|
|
1335
|
+
three modes across both `AgentRunner` and `AsyncAgentRunner`.
|
|
1336
|
+
`Claude` was already immune (its tool inputs arrive pre-parsed).
|
|
1337
|
+
|
|
1338
|
+
- **A tool-call preamble is no longer returned as the final answer.**
|
|
1339
|
+
Models routinely end a turn with an announcement instead of an action
|
|
1340
|
+
— "I'll look up your recent scores to get a clear view of your
|
|
1341
|
+
communication skills. Just a second!" — and every "no tool call
|
|
1342
|
+
found" branch in both runners was coded as *this text is the answer,
|
|
1343
|
+
break*. The loop terminated on iteration 1 and the caller got a
|
|
1344
|
+
promise instead of a result. Three sites per runner were affected:
|
|
1345
|
+
the native-binding path (`type != "tool_use"`), the
|
|
1346
|
+
`use_function_calling` path (parser unresolved), and the JSON-text
|
|
1347
|
+
path (response didn't parse). `max_iterations` never helped, because
|
|
1348
|
+
the break happened before any iteration was spent.
|
|
1349
|
+
|
|
1350
|
+
The runner now feeds the model one corrective nudge — "your last turn
|
|
1351
|
+
had no action, so nothing happened; do it, don't announce it" — and
|
|
1352
|
+
continues the loop. Verified against both `AgentRunner` and
|
|
1353
|
+
`AsyncAgentRunner` in all three modes.
|
|
1354
|
+
|
|
1355
|
+
### Added
|
|
1356
|
+
|
|
1357
|
+
- **Proactive "act, don't announce" system-prompt clause.** The reactive
|
|
1358
|
+
`text_turn_nudges` fix corrects an agent *after* it narrates instead of
|
|
1359
|
+
acting; this clause heads it off. When (and only when) an agent has
|
|
1360
|
+
tools, its system prompt now tells it to call the tool rather than
|
|
1361
|
+
reply "I'll do X / just a second" and stop — and to report what it DID
|
|
1362
|
+
in past tense. Injected in all three modes across both runners; skipped
|
|
1363
|
+
for tool-less chat agents, where prose is the correct answer. Sits
|
|
1364
|
+
before any `system_addendum` so a caller's role instructions still win.
|
|
1365
|
+
|
|
1366
|
+
- **`max_subtask_retries` on `Supervisor` / `AsyncSupervisor`** (default
|
|
1367
|
+
`1`). A sub-task that raised used to be recorded as an error and the
|
|
1368
|
+
Supervisor moved straight to synthesis — no second attempt. Now a
|
|
1369
|
+
failed sub-task is re-dispatched up to this many times, with the prior
|
|
1370
|
+
error appended to the query so the specialist knows what to fix
|
|
1371
|
+
("your previous attempt failed with X — diagnose and try again").
|
|
1372
|
+
Bounded and informed: only raised exceptions trigger a retry (a
|
|
1373
|
+
sub-task that returns content is accepted as-is, since the Supervisor
|
|
1374
|
+
can't tell "terse but correct" from "wrong"), and the error text is
|
|
1375
|
+
fed back rather than blindly re-running. Set to `0` for the old
|
|
1376
|
+
quit-on-first-failure behavior. Applies in both sequential and
|
|
1377
|
+
concurrent async modes.
|
|
1378
|
+
|
|
1379
|
+
- **`text_turn_nudges` on `AgentRunner` / `AsyncAgentRunner`** (default
|
|
1380
|
+
`1`). Caps the re-prompts described above at one extra LLM call per
|
|
1381
|
+
run; after the budget is spent the model's text stands as the answer.
|
|
1382
|
+
Set to `0` for the previous behavior. Automatically skipped when no
|
|
1383
|
+
tools are registered, since a runner with no tools is a plain chat
|
|
1384
|
+
call and prose genuinely is the answer there.
|
|
1385
|
+
|
|
1386
|
+
## [3.1.3] — 2026-07-22
|
|
1387
|
+
|
|
1388
|
+
Docs-only patch. No code changes since 3.1.2. Users on 3.1.2 don't
|
|
1389
|
+
need to upgrade for functionality; upgrade to pick up the improved
|
|
1390
|
+
onboarding docs bundled in the sdist.
|
|
1391
|
+
|
|
1392
|
+
### Documentation
|
|
1393
|
+
|
|
1394
|
+
- **Tools doc rewritten to answer "how do I actually use these?"**
|
|
1395
|
+
Added §0 `How each built-in tool is registered` as the entry
|
|
1396
|
+
section. Two registration paths — auto vs manual — laid out in a
|
|
1397
|
+
table on the first screen. Six runnable subsections covering every
|
|
1398
|
+
combination:
|
|
1399
|
+
- §0.1 DefaultTools via `Permissions(...)` (auto)
|
|
1400
|
+
- §0.2 WebTools via `tools=[web_search_tool(), web_fetch_tool()]`
|
|
1401
|
+
- §0.3 RAG via `TextSplitter` -> `VectorStore.add_documents` ->
|
|
1402
|
+
`vector_search_tool(store)`
|
|
1403
|
+
- §0.4 Handoffs via `handoff_tool` + `HandoffCoordinator`
|
|
1404
|
+
- §0.5 Custom `StructuredTool` from scratch
|
|
1405
|
+
- §0.6 Fully-loaded runner combining all of the above
|
|
1406
|
+
- §0.7 Rules on name collisions, invisible-denied-capabilities,
|
|
1407
|
+
async-tool behavior
|
|
1408
|
+
The existing inventory + wrapper / controls / cheat-sheet sections
|
|
1409
|
+
are unchanged; they now sit after the "how to use them" primer
|
|
1410
|
+
instead of before it.
|
|
1411
|
+
|
|
1412
|
+
## [3.1.2] — 2026-07-22
|
|
1413
|
+
|
|
1414
|
+
Patch release. Two independent fixes.
|
|
1415
|
+
|
|
1416
|
+
### Fixed
|
|
1417
|
+
|
|
1418
|
+
- **`llm_judge` correctly parses YES/NO across providers.** The judge
|
|
1419
|
+
parser was comparing the reply's first word to the literal string
|
|
1420
|
+
`"YES"`. GPT-4o answers `"YES,"` (comma-suffixed), which failed the
|
|
1421
|
+
equality check and marked every genuine PASS as FAIL. Claude replies
|
|
1422
|
+
`"YES"` without punctuation so the bug hid during local development.
|
|
1423
|
+
Fixed by matching `\b(YES|NO)\b` (word-boundary regex, case-
|
|
1424
|
+
insensitive) at the start of the reply. Handles every real shape:
|
|
1425
|
+
`YES`, `YES.`, `YES!`, `YES, exactly right`, `Yes.`, `yes -- reason`.
|
|
1426
|
+
Ambiguous replies (`Maybe`, empty string) still fail closed.
|
|
1427
|
+
- Regression test `test_llm_judge_parses_various_verdict_shapes`
|
|
1428
|
+
covers 8 YES shapes, 5 NO shapes, and 4 ambiguous replies.
|
|
1429
|
+
|
|
1430
|
+
### Added
|
|
1431
|
+
|
|
1432
|
+
- **`agentx_dev.Tools` is a one-stop tools namespace.** Users no
|
|
1433
|
+
longer need to remember which module each tool lives in:
|
|
1434
|
+
|
|
1435
|
+
```python
|
|
1436
|
+
from agentx_dev.Tools import (
|
|
1437
|
+
StandardTool, StructuredTool,
|
|
1438
|
+
AsyncStandardTool, AsyncStructuredTool,
|
|
1439
|
+
web_search_tool, web_fetch_tool,
|
|
1440
|
+
vector_search_tool, handoff_tool,
|
|
1441
|
+
DefaultTools, Permissions,
|
|
1442
|
+
)
|
|
1443
|
+
```
|
|
1444
|
+
|
|
1445
|
+
Both this form and the pre-existing `from agentx_dev import X`
|
|
1446
|
+
form coexist. Implementation uses PEP 562 module `__getattr__`
|
|
1447
|
+
and `__dir__` so re-exports are lazy (no import cost for modules
|
|
1448
|
+
the caller doesn't touch) and show up in IDE autocomplete +
|
|
1449
|
+
`dir(agentx_dev.Tools)`.
|
|
1450
|
+
|
|
1451
|
+
## [3.1.1] — 2026-07-21
|
|
1452
|
+
|
|
1453
|
+
Second batch of 3.1 features + a full docs + brand pass.
|
|
1454
|
+
|
|
1455
|
+
### Added
|
|
1456
|
+
|
|
1457
|
+
**Streaming through orchestration**
|
|
1458
|
+
- `Supervisor.stream()` / `AsyncSupervisor.astream()` emit
|
|
1459
|
+
`plan_start` / `plan` / `dispatch` / `subtask_result` /
|
|
1460
|
+
`synthesize_start` / `final` / `completion` events.
|
|
1461
|
+
- `HandoffCoordinator.stream()` / `.astream()` emit `invoke` /
|
|
1462
|
+
`completion` / `handoff` / `final` / `result` events per hop.
|
|
1463
|
+
- Legacy `.run()` / `.arun()` refactored to consume the streams (no
|
|
1464
|
+
code duplication).
|
|
1465
|
+
|
|
1466
|
+
**Prompt optimization — `Compiled`**
|
|
1467
|
+
- New `agentx_dev.Compiler` module.
|
|
1468
|
+
- `Compiled(runner_factory, trainset, ...)` iteratively refines a
|
|
1469
|
+
runner's `system_addendum` against the eval harness. Half of
|
|
1470
|
+
DSPy's power at a tenth of the surface.
|
|
1471
|
+
|
|
1472
|
+
**Anthropic Batch API**
|
|
1473
|
+
- `Claude.batch(requests)` submits many prompts at Anthropic's 50%-off
|
|
1474
|
+
batch rate, polls to completion, returns results in submission order.
|
|
1475
|
+
- Per-request error dicts on failure; token usage funneled into
|
|
1476
|
+
`TokenUsage` so cost tracking stays a single source of truth.
|
|
1477
|
+
|
|
1478
|
+
**Vector store adapters — `agentx_dev.VectorStores`**
|
|
1479
|
+
- `ChromaVectorStore`, `QdrantVectorStore`, `PgVectorStore` — same
|
|
1480
|
+
public shape as the in-memory `VectorStore` (`add` / `search` /
|
|
1481
|
+
`delete` / `clear` / `__len__` / `embeddings`).
|
|
1482
|
+
- `vector_search_tool()` and `SemanticMemory` accept any of them.
|
|
1483
|
+
- SDK imports lazy; friendly `ImportError` when the underlying SDK
|
|
1484
|
+
is missing.
|
|
1485
|
+
|
|
1486
|
+
**Trace viewer (`viewer/`)**
|
|
1487
|
+
- Self-hosted single-page app that reads `FileHook` JSONL and renders
|
|
1488
|
+
a timeline with type/text filters, summary sidebar, JSON drill-down.
|
|
1489
|
+
- Works from `file://`, no server required.
|
|
1490
|
+
|
|
1491
|
+
**Docs site (`host/`)**
|
|
1492
|
+
- Full editorial dark-first design system (JetBrains Mono headings,
|
|
1493
|
+
Inter body, `#B8FF3E` electric-lime accent).
|
|
1494
|
+
- Command palette (`Cmd+K`) with keyboard navigation and live search.
|
|
1495
|
+
- Hero code snippet with hand-tinted syntax highlighting.
|
|
1496
|
+
- Reading progress bar, breadcrumbs, header anchor links.
|
|
1497
|
+
- Sidebar sliding active marker, collapsible groups.
|
|
1498
|
+
- Code copy buttons, language labels.
|
|
1499
|
+
- Right-rail auto-TOC with `IntersectionObserver` scrollspy.
|
|
1500
|
+
- Dark/light theme toggle, persisted.
|
|
1501
|
+
- Cache-busted assets so edits land on refresh without hard-reload.
|
|
1502
|
+
|
|
1503
|
+
**Brand identity (`brand/`)**
|
|
1504
|
+
- Full brand kit: 5 SVG assets (`mark`, `mono`, `wordmark`, `logo-full`,
|
|
1505
|
+
`app-icon`), `BRAND.md` strategy doc, rendered brand-kit HTML deck.
|
|
1506
|
+
- Copy audit dropped "small" (weak) and "LangChain" references from
|
|
1507
|
+
all marketing surfaces.
|
|
1508
|
+
- Favicon wired into docs + trace viewer.
|
|
1509
|
+
|
|
1510
|
+
**Test suite (`tests/`)**
|
|
1511
|
+
- Restored + expanded pytest suite: 127 tests passing (3 skipped for
|
|
1512
|
+
absent optional SDKs).
|
|
1513
|
+
- Coverage: parser + all `AgentType` variants, `ToolRegistry`
|
|
1514
|
+
(dispatch / dup-guard / circuit-breaker / timeout), Permissions
|
|
1515
|
+
(capability gating + sandbox + traversal), budgets (cost / rate /
|
|
1516
|
+
retry / non-retryable HTTP), runner loop (streaming + output_schema
|
|
1517
|
+
+ chat history), embeddings + `VectorStore` + `SemanticMemory`,
|
|
1518
|
+
handoffs (bounded hops + history sanitization), evals harness
|
|
1519
|
+
(all assertion helpers + JSON case loaders), vector-store adapter
|
|
1520
|
+
shape conformance.
|
|
1521
|
+
|
|
1522
|
+
**Docs (`docs/`)**
|
|
1523
|
+
- Full docs tree (34 pages), including new pages for:
|
|
1524
|
+
vector store adapters, prompt optimization, batch API, trace viewer,
|
|
1525
|
+
and a **use-cases** landing (13 concrete scenarios with runnable code).
|
|
1526
|
+
- Rewrote **Tools** page to enumerate every built-in tool with args,
|
|
1527
|
+
return shape, capability flag, and use-case guidance.
|
|
1528
|
+
- Rewrote **Agents** page to cover all four orchestration
|
|
1529
|
+
architectures (Solo / Supervisor / Handoffs / Compiled) with
|
|
1530
|
+
decision trees, worked examples, and cheat sheet.
|
|
1531
|
+
- **Agentic RAG chatbot** as use case §13 — multi-query decomposition,
|
|
1532
|
+
parallel retrieval, self-critique, citations, user memory.
|
|
1533
|
+
|
|
1534
|
+
**Examples**
|
|
1535
|
+
- `examples/agentic_rag_demo.py` — the runnable version of the
|
|
1536
|
+
agentic RAG use case. Auto-seeds a KB if none exists, `--demo` flag
|
|
1537
|
+
runs a 3-turn scripted session proving user-notes recall works.
|
|
1538
|
+
|
|
1539
|
+
**Package**
|
|
1540
|
+
- `[chroma]`, `[qdrant]`, `[pgvector]`, `[dev]` extras added.
|
|
1541
|
+
- `[anthropic]` bumped to `>=0.36` (Batch API + prompt cache).
|
|
1542
|
+
|
|
1543
|
+
### Fixed
|
|
1544
|
+
|
|
1545
|
+
- `AgentRunner._iter_run` in `bind_tools_natively=True` mode uses a
|
|
1546
|
+
minimal system prompt instead of the AgentType template so the
|
|
1547
|
+
ReAct `action/action_input` scaffold no longer fights the native
|
|
1548
|
+
tool interface. Previously produced JSON-blob answers under GPT.
|
|
1549
|
+
- `HandoffCoordinator._sanitize_history_for_next_agent` strips tool
|
|
1550
|
+
and function role messages between hops so tool_call_ids from a
|
|
1551
|
+
previous agent don't leak into the next model's call (OpenAI 400).
|
|
1552
|
+
- Docs site marker positioning uses double-`requestAnimationFrame` +
|
|
1553
|
+
`document.fonts.ready` so the sidebar accent bar lands on the
|
|
1554
|
+
correct row even on a cold font cache.
|
|
1555
|
+
- Primary hero CTA color uses `#doc .hero-cta a.primary` selector to
|
|
1556
|
+
outrank `#doc a` link styling (previously rendered lime-on-lime
|
|
1557
|
+
and was invisible).
|
|
1558
|
+
|
|
1559
|
+
### Notes
|
|
1560
|
+
|
|
1561
|
+
- Package version bumped from `3.0.6` to `3.1.1`. The 3.1.0 release
|
|
1562
|
+
did not ship publicly — 3.1.1 is the first 3.1-tagged PyPI release
|
|
1563
|
+
and includes both batches of features.
|
|
1564
|
+
|
|
1565
|
+
## [3.1.0] — internal only (commits 52840e7)
|
|
1566
|
+
|
|
1567
|
+
First batch of 3.1 features. Committed but not released to PyPI.
|
|
1568
|
+
Merged into 3.1.1 for the public release.
|
|
1569
|
+
|
|
1570
|
+
### Added
|
|
1571
|
+
- Anthropic prompt caching (`Claude(enable_prompt_cache=True)`).
|
|
1572
|
+
- Parallel per-turn tool dispatch in `AgentRunner`
|
|
1573
|
+
(`bind_tools_natively=True`, `parallel_tool_workers`).
|
|
1574
|
+
- Semantic memory (`SemanticMemory`, embeddings-backed retrieval).
|
|
1575
|
+
- RAG core (`Embeddings`, `HashEmbeddings`, `OpenAIEmbeddings`,
|
|
1576
|
+
`VectorStore`, `VectorHit`, `vector_search_tool()`).
|
|
1577
|
+
- Agent-to-agent handoffs (`HandoffRequest`, `handoff_tool`,
|
|
1578
|
+
`HandoffCoordinator`, `HandoffResult`).
|
|
1579
|
+
- Evals harness (`EvalCase`, `EvalRunner`, `EvalReport`, 7 assertion
|
|
1580
|
+
helpers, JSON case loader, `python -m agentx_dev.Evals run` CLI).
|
|
1581
|
+
- `TokenUsage.cache_hit_ratio` property.
|
|
1582
|
+
|
|
1583
|
+
## [3.0.6] — 2026-03 (baseline)
|
|
1584
|
+
|
|
1585
|
+
Security hardening baseline (SSRF guard on `web_fetch`, HMAC-signed
|
|
1586
|
+
persistent state, scrubbed subprocess env, path sanitizer,
|
|
1587
|
+
`permissions.json` mode 0o600, ReDoS guard on `grep`,
|
|
1588
|
+
`invoke`/`ainvoke` accept bare strings and message lists).
|