AgentX-Dev 3.8.0__tar.gz → 3.9.0__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-3.9.0}/CHANGELOG.md +150 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/PKG-INFO +20 -4
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/README.md +18 -3
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/ChatModel.py +7 -1
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/DefaultTools.py +2382 -2321
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Session.py +93 -15
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Skills/Install.py +24 -5
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/WebTools.py +128 -28
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/app.py +35 -11
- agentx_dev-3.9.0/agentx_dev/cli/app_keys.py +118 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/app_mcp.py +25 -8
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/approval.py +128 -19
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/context.py +8 -2
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/fullscreen.py +66 -17
- agentx_dev-3.9.0/agentx_dev/cli/keystore.py +99 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/mcp.py +171 -22
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/preview.py +16 -2
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/sessions.py +11 -2
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/team_record.py +82 -4
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/trust.py +8 -4
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/typeahead.py +41 -4
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/ui.py +75 -13
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/webui.py +521 -47
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/webui/app.js +496 -63
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/webui/index.html +165 -135
- agentx_dev-3.9.0/agentx_dev/resources/webui/orb.js +272 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/webui/style.css +677 -515
- agentx_dev-3.9.0/agentx_dev/safe_exe.py +75 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev.egg-info/PKG-INFO +20 -4
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev.egg-info/SOURCES.txt +4 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev.egg-info/requires.txt +1 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/pyproject.toml +2 -1
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/AGENTX.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/CONTRIBUTING.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/LICENSE +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/MANIFEST.in +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Agents/Agent.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Agents/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/AsyncTools.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/AutoSetup.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Cache.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Compiler.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Config.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Embeddings.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Evals.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Handoffs.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Loader.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/MCP.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Media.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Memory.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Observability.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Operator.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Pictures.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Planner.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/RunLog.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Runner/AgentRun.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Runner/AsyncAgentRun.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Runner/Persistence.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Runner/Trim.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Runner/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Runner/promptTemplate.yaml +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Skills/Attach.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Skills/Defaults.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Skills/Store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Skills/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Splitters.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Streaming.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/SubAgents.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Supervisor.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/SupervisorMemory.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/Tools.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/VectorStores/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/VectorStores/chroma_store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/VectorStores/pg_store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/VectorStores/qdrant_store.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/__main__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_optional.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_supervisor/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_supervisor/core.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_supervisor/plans.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_supervisor/prompts.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_supervisor/results.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/_supervisor/steps.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/app_models.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/app_team.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/app_usage.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/app_voice.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/commands.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/compact.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/memory.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/models.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/recall.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/remote.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/safepaths.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/settings.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/team.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/tools.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/ui_input.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/ui_text.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/cli/voice.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/__init__.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/promptTemplate.yaml +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/skills/auditing-a-project/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/skills/changing-files/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/skills/delegating-work/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/skills/exploring-a-project/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/skills/running-code/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/skills/web-research/SKILL.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev/resources/webui/icon.svg +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev.egg-info/dependency_links.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev.egg-info/entry_points.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/agentx_dev.egg-info/top_level.txt +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/README.md +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/agentic_rag_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/async_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/async_quickstart.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/auto_features_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/caching_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/chatbot_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/complete_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/concurrent_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/concurrent_tool_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/file_agent_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/function_calling_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/mcp_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/mcp_github_triage_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/observability_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/orchestration_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/planner_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/robust_link_scraper.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/subagents_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/supervisor_codebase_analysis_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/supervisor_example.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/sync_quickstart.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/v3_1_1_features_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/v3_1_comprehensive_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/examples/v3_1_features_demo.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/host/build_data.py +0 -0
- {agentx_dev-3.8.0 → agentx_dev-3.9.0}/setup.cfg +0 -0
|
@@ -6,6 +6,156 @@ follows [Keep a Changelog](https://keepachangelog.com/); versioning is
|
|
|
6
6
|
|
|
7
7
|
## [Unreleased]
|
|
8
8
|
|
|
9
|
+
## [3.9.0] - 2026-11-16
|
|
10
|
+
|
|
11
|
+
### Added
|
|
12
|
+
|
|
13
|
+
- **Two conversations side by side in the app window.** Drag a conversation from **Recent** onto the chat (or use
|
|
14
|
+
the button that shows on the row) and it opens on the right half of the window, beside the one you are in; the
|
|
15
|
+
folder menu's **Open a folder beside…** opens a new conversation in another project there. Each side is an
|
|
16
|
+
agentx of its own, with its own history, settings, approvals, orb and memory, so both can work at once; they use
|
|
17
|
+
the same API keys. Closing a side (its ✕) saves its conversation and ends its agentx. On a phone-sized window the
|
|
18
|
+
side covers the chat until you close it. Drag the line between the sides to give either one more room (double-click it
|
|
19
|
+
for half and half; the split is remembered). Each side's **Details** belongs to that side: the left one opens
|
|
20
|
+
beside the left conversation, not past the other side.
|
|
21
|
+
- **Details can be resized.** Drag its left edge to make the panel wider or narrower (from 260 pixels to 60% of
|
|
22
|
+
the window); the width is remembered, and a double-click on the edge puts it back. It works as a column, when it
|
|
23
|
+
slides over the chat, and on either side of the split view.
|
|
24
|
+
|
|
25
|
+
- **API keys in the app window.** Settings has an **API keys** tab: paste your Claude or OpenAI key to save it,
|
|
26
|
+
replace it later, or remove it. Keys are kept encrypted by your system's credential store (Windows Credential
|
|
27
|
+
Manager, the macOS Keychain, the Secret Service on Linux), never in a file, and the page shows only their last
|
|
28
|
+
four characters. A saved key wins over the environment variable; agentx switches to it at your next message and
|
|
29
|
+
keeps the conversation. `keyring` is now installed with agentx.
|
|
30
|
+
|
|
31
|
+
### Changed
|
|
32
|
+
|
|
33
|
+
- **The agentx orb.** A sphere of lime dots now shows what agentx is doing in the app window (the terminal keeps
|
|
34
|
+
its spinner).
|
|
35
|
+
While it thinks, its surface bubbles, dots pushing out and settling back; while a step runs, rings of dots light
|
|
36
|
+
up and turn on their own axes in opposite directions; when a turn finishes it rests in a slow turn with a soft
|
|
37
|
+
glow ("Done" shows for a moment); while it waits for your answer it folds into a cube that turns slowly and breathes, a band of light
|
|
38
|
+
sweeping down it. In the window, a new
|
|
39
|
+
conversation shows it large under "What should we work on?"; once you write it moves to the top, large, with a
|
|
40
|
+
small pill under it naming what it shows (Thinking, the step it runs, Waiting for your answer, Done and how long
|
|
41
|
+
it took, then Ready), and stays there for the rest of the conversation. The large orb can be grabbed with the mouse and spun; let go and it keeps the flick's
|
|
42
|
+
momentum, then settles into its slow turn in the direction you threw it.
|
|
43
|
+
- **A tidier conversation in the app window.** The mode note ("Auto mode is on.") is now a small chip: hover it
|
|
44
|
+
to read what the mode allows, click it to open Details. MCP servers that need signing in are listed in Details
|
|
45
|
+
only, with a yellow dot on its button, instead of a notice in the conversation. A new conversation no longer
|
|
46
|
+
repeats the model, its maker and the mode under the question (the message box already shows the model and the
|
|
47
|
+
mode), which leaves room for a larger orb. Once you write, the agentx header box (version, model, folder) and the mode
|
|
48
|
+
chip are gone, and the line over the message box no longer repeats what the orb's pill says. A new conversation
|
|
49
|
+
shows the mode chip again.
|
|
50
|
+
|
|
51
|
+
### Fixed
|
|
52
|
+
|
|
53
|
+
- **A conversation opened again shows what it did.** In the app window, steps that a model wrapped in its response
|
|
54
|
+
format (`React_`) showed as "React" rows with the tool's name, or its note, as text; they now read as the tool
|
|
55
|
+
they ran (Run, Edit, Read...), with its note and the final answer. Details counts the tokens of the whole
|
|
56
|
+
conversation, across every start: a conversation's saved total was replaced by the last start's count each
|
|
57
|
+
turn, so an old one showed zeros or too little; it now adds up (tokens from the cache too). And Details lists the
|
|
58
|
+
team jobs a conversation ran before: each job now names its conversation in `job.md`, and an older one is matched
|
|
59
|
+
by its task (or, when the history has been summarised since, by when it ran).
|
|
60
|
+
- **An MCP sign-in in the app window.** `/mcp login` showed the whole sign-in address as text while the orb said
|
|
61
|
+
"Ready". It is now a small card with an **Open the sign-in page** button, and the orb folds into its cube with
|
|
62
|
+
"Waiting for you to sign in to <server>" until the sign-in ends (Esc cancels it).
|
|
63
|
+
- **A wider conversation in the app window.** The conversation and the message box use up to 1000 pixels instead
|
|
64
|
+
of 768, so a wide window has less empty space on each side.
|
|
65
|
+
- **Esc stops agentx in the app window while you type.** "Esc to stop" did nothing while the cursor was in the
|
|
66
|
+
message box; now it stops the turn from there too (what you typed stays), and the orb's pill reads "Stopping…"
|
|
67
|
+
until the turn ends. A stop during the pause before a model request is sent again (a rate limit asks for seconds)
|
|
68
|
+
now lands at once, instead of after the pause.
|
|
69
|
+
|
|
70
|
+
## [3.8.1] - 2026-11-13
|
|
71
|
+
|
|
72
|
+
### Changed
|
|
73
|
+
|
|
74
|
+
- **The app window opens like the docs site.** The mark rises with a soft glow, `agentx` types itself in letter by
|
|
75
|
+
letter and a lime cursor keeps blinking after it; "Starting..." (dots lighting up in turn) shows until agentx is
|
|
76
|
+
ready, and the welcome grows out of it. While it works, that blinking cursor replaces the round spinner.
|
|
77
|
+
- **The folder trust question highlights "No, exit".** Enter alone now exits; trust a folder with `1`, `y`, or Up
|
|
78
|
+
then Enter. The app window's trust card does the same.
|
|
79
|
+
- **"Don't ask again" is `2` in approval menus.** The `a` shortcut is gone from the arrow-key menu (it still works in
|
|
80
|
+
the typed `[y/a/n]` prompt), so a word typed for the chat cannot pick it.
|
|
81
|
+
- **Menus wait a moment before they take keys.** Keys typed before a menu appeared are thrown away, and keys pressed
|
|
82
|
+
in its first 0.4 seconds are ignored (Ctrl+C still cancels). In full-screen mode a menu waits while the input line
|
|
83
|
+
holds text (`Clear or send your message to answer`); keys keep editing the message and Enter sends it.
|
|
84
|
+
- **App window cards do not take the focus while you type.** When the message box holds text, or another field has
|
|
85
|
+
the focus, a new card leaves the focus where it is; from an empty message box `1`, `2`, `3` answer it but Enter and
|
|
86
|
+
Space do not.
|
|
87
|
+
- **Some granted commands ask again.** A "don't ask again" command grant no longer covers a call that uses an option
|
|
88
|
+
through which the program runs another program (`git fetch --upload-pack`, `make -f`, `rg --pre` and similar), and
|
|
89
|
+
`find`, `sed`, `tar`, `rsync`, `vim`, `git submodule`, `git rebase` and similar are never remembered.
|
|
90
|
+
- **An MCP server launcher that exists only as a `.ps1` script needs its full path.** A bare `"command"` is now looked
|
|
91
|
+
up on `PATH` only; one not found reads `not connected: <name> not found on PATH`.
|
|
92
|
+
- **A parent folder's `AGENTX.md` needs that folder trusted.** To use a repository's top-level file when you start
|
|
93
|
+
in a subfolder, trust the top folder once.
|
|
94
|
+
|
|
95
|
+
### Fixed
|
|
96
|
+
|
|
97
|
+
- **`--remote` on a VPS prints the ssh command that works.** It assumed SSH on port 22 and guessed the address, so
|
|
98
|
+
a server on another port (2222, say) timed out. In an SSH session agentx now uses the address and port that
|
|
99
|
+
session came in on (`ssh -p 2222 ...`).
|
|
100
|
+
|
|
101
|
+
### Security
|
|
102
|
+
|
|
103
|
+
- **Keys typed for the chat never answer an approval.** Keys still waiting when a prompt opens go to your message
|
|
104
|
+
line, never to the prompt; menus throw away keys typed before they appeared and ignore the first 0.4 seconds.
|
|
105
|
+
- **A stray Enter never trusts a folder.** The trust question opens on "No, exit", and keys typed while agentx
|
|
106
|
+
started are thrown away.
|
|
107
|
+
- **A command grant cannot approve another program.** "Don't ask again" on a command no longer approves a later call
|
|
108
|
+
that runs another program through its options (`git fetch --upload-pack`, `git grep -O`, `git -c`, `make -f` or
|
|
109
|
+
`VAR=value`, `go test -exec`, `cargo --config`, `rg --pre`, `npm --script-shell`) or holds `$ * ? [ {` for such a
|
|
110
|
+
program. Programs that run others through their options or scripts (`find`, `sed`, `tar`, `rsync`, `scp`, `zip`,
|
|
111
|
+
`vim`, `parallel`, `crontab`, `at`...) and git sub-commands such as `submodule`, `rebase`, `difftool`, `config` and
|
|
112
|
+
`hook` are approved once and never remembered.
|
|
113
|
+
- **An edit grant covers edits only.** Answering "don't ask again" at an edit prompt no longer lets web searches and
|
|
114
|
+
page fetches run unasked; they still ask until you choose `accept-edits` yourself or allow the web at a web prompt.
|
|
115
|
+
- **`/clear` and `/resume` also end a web "don't ask again" grant.**
|
|
116
|
+
- **The approval diff shows every hidden line break.** Lines split at LF only: a line break swapped for a form feed,
|
|
117
|
+
a vertical tab, U+2028 or a lone CR shows as a changed line with an escape such as `\x0c`, not as "only line
|
|
118
|
+
endings change".
|
|
119
|
+
- **Backslash paths on Linux and macOS.** A backslash is part of a file name there, so the deny rules (and the
|
|
120
|
+
prompt) also judge the file the name really reaches: a link named `x\y` cannot lead past the `.git` protection.
|
|
121
|
+
- **The window's key is never on a command line.** The window agentx opens gets a one-time start link (`?c=`) that
|
|
122
|
+
works once, for 2 minutes; the page trades it for the key. A second use is refused, the key is changed and the
|
|
123
|
+
terminal prints a warning with the new link.
|
|
124
|
+
- **An approval too long to show in full is declined.** A card whose action, target or preview would be cut is
|
|
125
|
+
declined without asking, with a notice saying why; "don't ask again" is not offered for a rule too long to show.
|
|
126
|
+
- **Links in answers cannot add attributes.** A link's address ends at a space, `(`, `)` or a quote, and each link is
|
|
127
|
+
built once, so model text cannot break out of it.
|
|
128
|
+
- **Saving Settings keeps declined project rules off.** Project allow rules you turned down at start, and rules added
|
|
129
|
+
to the project file since, stay off for the session after the window saves the settings.
|
|
130
|
+
- **Open another folder imports nothing from that folder** before the new agentx asks whether you trust it.
|
|
131
|
+
- **Pictures are never saved through a planted link.** When the project's `.agentx` or `.agentx/uploads` is a link,
|
|
132
|
+
a picture goes to `~/.agentx/uploads/`; the message names where the file really is.
|
|
133
|
+
- **Turning remote control off closes the windows opened through Tailscale** (their conversations are saved), and
|
|
134
|
+
the 30-second re-share check can no longer share the window again afterwards. agentx also changes its key, so a
|
|
135
|
+
link shared with another device stops working; windows on this computer get the new key and keep going.
|
|
136
|
+
- **Nothing is started from the project folder on Windows.** `run_shell` takes git-bash from Program Files, a
|
|
137
|
+
per-user Git install or `PATH` without the current folder, else System32's PowerShell; stopping a command runs
|
|
138
|
+
System32's `taskkill`; Tailscale, Edge, Chrome and PowerShell for the app window are found the same way.
|
|
139
|
+
- **MCP stdio servers' bare commands are found on `PATH` only**, never in the project folder (the server still
|
|
140
|
+
starts in the current folder).
|
|
141
|
+
- **Read-only auto-allows are re-checked on every connect.** Tools allowed because the server marked them read-only
|
|
142
|
+
are kept in `"auto"` in `mcp.json`, apart from your own `/mcp allow` choices; one the server stops marking read-only
|
|
143
|
+
asks again, with a warning. On the first connect after upgrading, saved `"allow"` entries the server marks
|
|
144
|
+
read-only move to `"auto"`.
|
|
145
|
+
- **A sign-in is tied to the server's address.** When a server's URL changes, the old sign-in is deleted instead of
|
|
146
|
+
being sent there; a server taken out of `mcp.json` is disconnected and its sign-in deleted; `/mcp add` under a
|
|
147
|
+
name used before starts signed out.
|
|
148
|
+
- **A parent folder's `AGENTX.md` is read only when you trust that folder.** The folder you start in and your home
|
|
149
|
+
folder's file are always read.
|
|
150
|
+
- **`grep` and `find_files` do not follow links out of the project.** A linked file that leads outside is skipped and
|
|
151
|
+
linked folders (Windows junctions included) are never entered.
|
|
152
|
+
- **`web_fetch` connects only to the address it checked**, on every redirect, so DNS rebinding cannot reach a local
|
|
153
|
+
address; a redirect to a private address reads `ERROR: refused`.
|
|
154
|
+
- **Saved conversations and their titles are readable by you only** (`0600`, new folders `0700`), and `~/.agentx` is
|
|
155
|
+
created yours only.
|
|
156
|
+
- **`/skills install` refuses a repository whose skill files have `:` in their paths** (another drive, or an NTFS
|
|
157
|
+
stream, on Windows) and never writes outside its download folder.
|
|
158
|
+
|
|
9
159
|
## [3.8.0] - 2026-11-08
|
|
10
160
|
|
|
11
161
|
### Added
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: agentx-dev
|
|
3
|
-
Version: 3.
|
|
3
|
+
Version: 3.9.0
|
|
4
4
|
Summary: Build LLM agents in Python, and get agentx: a coding assistant for your terminal. GPT and Claude, permission-gated tools, skills, MCP, team jobs with helpers that run in parallel, RAG, memory and evals.
|
|
5
5
|
Author-email: Bruce-Arhin Shadrach <brucearhin098@gmail.com>
|
|
6
6
|
License-Expression: MIT
|
|
@@ -16,6 +16,7 @@ Requires-Dist: pydantic>=2.0
|
|
|
16
16
|
Requires-Dist: PyYAML>=5.3.1
|
|
17
17
|
Requires-Dist: rich>=10.0
|
|
18
18
|
Requires-Dist: prompt_toolkit>=3.0
|
|
19
|
+
Requires-Dist: keyring>=24
|
|
19
20
|
Provides-Extra: anthropic
|
|
20
21
|
Requires-Dist: anthropic>=0.36; extra == "anthropic"
|
|
21
22
|
Provides-Extra: mcp
|
|
@@ -100,6 +101,21 @@ Every code block below is something you can paste and run.
|
|
|
100
101
|
|
|
101
102
|
---
|
|
102
103
|
|
|
104
|
+
## What's new in 3.9 — the orb and two conversations at once
|
|
105
|
+
|
|
106
|
+
The `agentx` app window has an **orb**: a sphere of lime dots that shows what agentx is doing. It bubbles while it
|
|
107
|
+
thinks, turns rings of dots while a step runs, and folds into a turning cube while it waits for you. Drag a
|
|
108
|
+
conversation from **Recent** onto the chat to run **two conversations side by side**, each with its own history
|
|
109
|
+
and approvals. **Settings > API keys** keeps your Claude and OpenAI keys encrypted by the system. See
|
|
110
|
+
[What's new](docs/whats-new.md) and [Upgrading](docs/guides/upgrading.md).
|
|
111
|
+
|
|
112
|
+
## What's new in 3.8.1 — a security release
|
|
113
|
+
|
|
114
|
+
3.8.1 closes gaps in the `agentx` app where a cloned project, a web page or keys typed at the wrong moment could get
|
|
115
|
+
something past you: the trust question now defaults to "No", "don't ask again" covers only what it says, nothing is
|
|
116
|
+
started from the project folder on Windows, and the app window no longer puts its key on a command line. See
|
|
117
|
+
[What's new](docs/whats-new.md) and [Upgrading](docs/guides/upgrading.md).
|
|
118
|
+
|
|
103
119
|
## What's new in 3.8 — skills, the app window and remote use
|
|
104
120
|
|
|
105
121
|
An agent can now read short **skills**: guides for a kind of work ("how to explore a project", "how to change
|
|
@@ -167,9 +183,9 @@ agentx # or: python -m agentx_dev
|
|
|
167
183
|
|
|
168
184
|
| Feature | What you get |
|
|
169
185
|
|---|---|
|
|
170
|
-
| **Approvals and modes** | Reads run; edits, commands, delete and move ask first, with a diff or the full command shown. Modes `default`, `accept-edits`, `plan` and `auto` (set on purpose only, never from Shift+Tab or a project file).
|
|
171
|
-
| **Trust a folder** | The first start in a folder asks "Accessing workspace ... Yes / No". Nothing from the folder (`AGENTX.md`, its settings) is read before you say yes. `.git`, `.agentx` and shell/git config files are never written, in any mode. |
|
|
172
|
-
| **`AGENTX.md`** | Instructions for the model (how to run the tests, the layout, conventions), read from your home folder and from the working folder up. `/init` drafts one. |
|
|
186
|
+
| **Approvals and modes** | Reads run; edits, commands, delete and move ask first, with a diff or the full command shown. Modes `default`, `accept-edits`, `plan` and `auto` (set on purpose only, never from Shift+Tab or a project file). "Don't ask again" (`2` in the menu) remembers all edits, or a plain command prefix, for the conversation. |
|
|
187
|
+
| **Trust a folder** | The first start in a folder asks "Accessing workspace ... Yes / No", with No highlighted, so Enter alone exits. Nothing from the folder (`AGENTX.md`, its settings) is read before you say yes. `.git`, `.agentx` and shell/git config files are never written, in any mode. |
|
|
188
|
+
| **`AGENTX.md`** | Instructions for the model (how to run the tests, the layout, conventions), read from your home folder and from the working folder up (a parent folder's file only when you trust that folder). `/init` drafts one. |
|
|
173
189
|
| **Sessions** | Conversations autosave. `--continue`, `--resume [ID]` and `/resume` pick one up again. |
|
|
174
190
|
| **Slash commands** | `/help`, `/clear`, `/model`, `/mode`, `/settings`, `/resume`, `/sessions`, `/init`, `/team`, `/team-models`, `/exit`. Typing `/` shows the list. |
|
|
175
191
|
| **Team jobs** | For a big job the agent calls `start_team_job` (or you type `/team <task>`): a supervisor plans up to six steps and three helpers (explorer reads, coder edits, runner runs commands) carry them out through the same approval gate. One message may run up to 5 team jobs, so a big task continues batch after batch. A helper that runs out of its 100 steps hands on its partial work. |
|
|
@@ -36,6 +36,21 @@ Every code block below is something you can paste and run.
|
|
|
36
36
|
|
|
37
37
|
---
|
|
38
38
|
|
|
39
|
+
## What's new in 3.9 — the orb and two conversations at once
|
|
40
|
+
|
|
41
|
+
The `agentx` app window has an **orb**: a sphere of lime dots that shows what agentx is doing. It bubbles while it
|
|
42
|
+
thinks, turns rings of dots while a step runs, and folds into a turning cube while it waits for you. Drag a
|
|
43
|
+
conversation from **Recent** onto the chat to run **two conversations side by side**, each with its own history
|
|
44
|
+
and approvals. **Settings > API keys** keeps your Claude and OpenAI keys encrypted by the system. See
|
|
45
|
+
[What's new](docs/whats-new.md) and [Upgrading](docs/guides/upgrading.md).
|
|
46
|
+
|
|
47
|
+
## What's new in 3.8.1 — a security release
|
|
48
|
+
|
|
49
|
+
3.8.1 closes gaps in the `agentx` app where a cloned project, a web page or keys typed at the wrong moment could get
|
|
50
|
+
something past you: the trust question now defaults to "No", "don't ask again" covers only what it says, nothing is
|
|
51
|
+
started from the project folder on Windows, and the app window no longer puts its key on a command line. See
|
|
52
|
+
[What's new](docs/whats-new.md) and [Upgrading](docs/guides/upgrading.md).
|
|
53
|
+
|
|
39
54
|
## What's new in 3.8 — skills, the app window and remote use
|
|
40
55
|
|
|
41
56
|
An agent can now read short **skills**: guides for a kind of work ("how to explore a project", "how to change
|
|
@@ -103,9 +118,9 @@ agentx # or: python -m agentx_dev
|
|
|
103
118
|
|
|
104
119
|
| Feature | What you get |
|
|
105
120
|
|---|---|
|
|
106
|
-
| **Approvals and modes** | Reads run; edits, commands, delete and move ask first, with a diff or the full command shown. Modes `default`, `accept-edits`, `plan` and `auto` (set on purpose only, never from Shift+Tab or a project file).
|
|
107
|
-
| **Trust a folder** | The first start in a folder asks "Accessing workspace ... Yes / No". Nothing from the folder (`AGENTX.md`, its settings) is read before you say yes. `.git`, `.agentx` and shell/git config files are never written, in any mode. |
|
|
108
|
-
| **`AGENTX.md`** | Instructions for the model (how to run the tests, the layout, conventions), read from your home folder and from the working folder up. `/init` drafts one. |
|
|
121
|
+
| **Approvals and modes** | Reads run; edits, commands, delete and move ask first, with a diff or the full command shown. Modes `default`, `accept-edits`, `plan` and `auto` (set on purpose only, never from Shift+Tab or a project file). "Don't ask again" (`2` in the menu) remembers all edits, or a plain command prefix, for the conversation. |
|
|
122
|
+
| **Trust a folder** | The first start in a folder asks "Accessing workspace ... Yes / No", with No highlighted, so Enter alone exits. Nothing from the folder (`AGENTX.md`, its settings) is read before you say yes. `.git`, `.agentx` and shell/git config files are never written, in any mode. |
|
|
123
|
+
| **`AGENTX.md`** | Instructions for the model (how to run the tests, the layout, conventions), read from your home folder and from the working folder up (a parent folder's file only when you trust that folder). `/init` drafts one. |
|
|
109
124
|
| **Sessions** | Conversations autosave. `--continue`, `--resume [ID]` and `/resume` pick one up again. |
|
|
110
125
|
| **Slash commands** | `/help`, `/clear`, `/model`, `/mode`, `/settings`, `/resume`, `/sessions`, `/init`, `/team`, `/team-models`, `/exit`. Typing `/` shows the list. |
|
|
111
126
|
| **Team jobs** | For a big job the agent calls `start_team_job` (or you type `/team <task>`): a supervisor plans up to six steps and three helpers (explorer reads, coder edits, runner runs commands) carry them out through the same approval gate. One message may run up to 5 team jobs, so a big task continues batch after batch. A helper that runs out of its 100 steps hands on its partial work. |
|
|
@@ -1515,7 +1515,13 @@ class BaseChatModel(ABC):
|
|
|
1515
1515
|
hook(attempt + 1, e, delay)
|
|
1516
1516
|
except Exception: # showing a retry must never change it
|
|
1517
1517
|
pass
|
|
1518
|
-
|
|
1518
|
+
# in short naps: a stop (Esc in the app window) is raised in this thread and lands only
|
|
1519
|
+
# between Python instructions, so one long sleep would hold it for the whole pause
|
|
1520
|
+
left = delay
|
|
1521
|
+
while left > 0:
|
|
1522
|
+
nap = min(left, 0.1)
|
|
1523
|
+
time.sleep(nap)
|
|
1524
|
+
left -= nap
|
|
1519
1525
|
raise last_exc
|
|
1520
1526
|
|
|
1521
1527
|
|