AgentX-Dev 3.7.0__tar.gz → 3.8.0__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.
Files changed (145) hide show
  1. agentx_dev-3.8.0/AGENTX.md +28 -0
  2. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/CHANGELOG.md +416 -0
  3. {agentx_dev-3.7.0/agentx_dev.egg-info → agentx_dev-3.8.0}/PKG-INFO +148 -43
  4. agentx_dev-3.7.0/PKG-INFO → agentx_dev-3.8.0/README.md +131 -90
  5. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/ChatModel.py +355 -21
  6. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/DefaultTools.py +195 -37
  7. agentx_dev-3.8.0/agentx_dev/Evals.py +821 -0
  8. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Media.py +2 -2
  9. agentx_dev-3.8.0/agentx_dev/Pictures.py +227 -0
  10. agentx_dev-3.8.0/agentx_dev/RunLog.py +306 -0
  11. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Runner/AgentRun.py +154 -13
  12. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Runner/AsyncAgentRun.py +79 -17
  13. agentx_dev-3.8.0/agentx_dev/Runner/Trim.py +286 -0
  14. agentx_dev-3.8.0/agentx_dev/Skills/Attach.py +170 -0
  15. agentx_dev-3.8.0/agentx_dev/Skills/Defaults.py +132 -0
  16. agentx_dev-3.8.0/agentx_dev/Skills/Install.py +449 -0
  17. agentx_dev-3.8.0/agentx_dev/Skills/Store.py +320 -0
  18. agentx_dev-3.8.0/agentx_dev/Skills/__init__.py +8 -0
  19. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/SubAgents.py +6 -0
  20. agentx_dev-3.8.0/agentx_dev/Supervisor.py +1065 -0
  21. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/VectorStores/chroma_store.py +1 -1
  22. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/VectorStores/pg_store.py +1 -1
  23. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/VectorStores/qdrant_store.py +1 -1
  24. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/__init__.py +13 -1
  25. agentx_dev-3.8.0/agentx_dev/_supervisor/__init__.py +1 -0
  26. agentx_dev-3.8.0/agentx_dev/_supervisor/core.py +718 -0
  27. agentx_dev-3.8.0/agentx_dev/_supervisor/plans.py +347 -0
  28. agentx_dev-3.8.0/agentx_dev/_supervisor/prompts.py +69 -0
  29. agentx_dev-3.8.0/agentx_dev/_supervisor/results.py +146 -0
  30. agentx_dev-3.8.0/agentx_dev/_supervisor/steps.py +426 -0
  31. agentx_dev-3.8.0/agentx_dev/cli/app.py +1695 -0
  32. agentx_dev-3.8.0/agentx_dev/cli/app_mcp.py +295 -0
  33. agentx_dev-3.8.0/agentx_dev/cli/app_models.py +277 -0
  34. agentx_dev-3.8.0/agentx_dev/cli/app_team.py +191 -0
  35. agentx_dev-3.8.0/agentx_dev/cli/app_usage.py +174 -0
  36. agentx_dev-3.8.0/agentx_dev/cli/app_voice.py +215 -0
  37. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/approval.py +24 -3
  38. agentx_dev-3.8.0/agentx_dev/cli/commands.py +434 -0
  39. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/compact.py +4 -2
  40. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/context.py +13 -5
  41. agentx_dev-3.8.0/agentx_dev/cli/fullscreen.py +638 -0
  42. agentx_dev-3.8.0/agentx_dev/cli/mcp.py +880 -0
  43. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/memory.py +236 -227
  44. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/models.py +67 -12
  45. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/preview.py +8 -0
  46. agentx_dev-3.8.0/agentx_dev/cli/recall.py +406 -0
  47. agentx_dev-3.8.0/agentx_dev/cli/remote.py +298 -0
  48. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/sessions.py +204 -137
  49. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/settings.py +107 -2
  50. agentx_dev-3.8.0/agentx_dev/cli/team.py +1299 -0
  51. agentx_dev-3.8.0/agentx_dev/cli/team_record.py +162 -0
  52. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/tools.py +85 -4
  53. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/trust.py +123 -122
  54. agentx_dev-3.8.0/agentx_dev/cli/typeahead.py +237 -0
  55. agentx_dev-3.8.0/agentx_dev/cli/ui.py +1104 -0
  56. agentx_dev-3.8.0/agentx_dev/cli/ui_input.py +151 -0
  57. agentx_dev-3.8.0/agentx_dev/cli/ui_text.py +294 -0
  58. agentx_dev-3.8.0/agentx_dev/cli/voice.py +154 -0
  59. agentx_dev-3.8.0/agentx_dev/cli/webui.py +1840 -0
  60. agentx_dev-3.8.0/agentx_dev/resources/skills/auditing-a-project/SKILL.md +44 -0
  61. agentx_dev-3.8.0/agentx_dev/resources/skills/changing-files/SKILL.md +23 -0
  62. agentx_dev-3.8.0/agentx_dev/resources/skills/delegating-work/SKILL.md +20 -0
  63. agentx_dev-3.8.0/agentx_dev/resources/skills/exploring-a-project/SKILL.md +30 -0
  64. agentx_dev-3.8.0/agentx_dev/resources/skills/running-code/SKILL.md +21 -0
  65. agentx_dev-3.8.0/agentx_dev/resources/skills/web-research/SKILL.md +20 -0
  66. agentx_dev-3.8.0/agentx_dev/resources/webui/app.js +1308 -0
  67. agentx_dev-3.8.0/agentx_dev/resources/webui/icon.svg +1 -0
  68. agentx_dev-3.8.0/agentx_dev/resources/webui/index.html +135 -0
  69. agentx_dev-3.8.0/agentx_dev/resources/webui/style.css +515 -0
  70. agentx_dev-3.7.0/README.md → agentx_dev-3.8.0/agentx_dev.egg-info/PKG-INFO +195 -12
  71. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev.egg-info/SOURCES.txt +39 -0
  72. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev.egg-info/requires.txt +15 -7
  73. agentx_dev-3.8.0/agentx_dev.egg-info/top_level.txt +1 -0
  74. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/host/build_data.py +8 -2
  75. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/pyproject.toml +32 -22
  76. agentx_dev-3.7.0/AGENTX.md +0 -308
  77. agentx_dev-3.7.0/agentx_dev/Evals.py +0 -514
  78. agentx_dev-3.7.0/agentx_dev/Supervisor.py +0 -2833
  79. agentx_dev-3.7.0/agentx_dev/cli/app.py +0 -1081
  80. agentx_dev-3.7.0/agentx_dev/cli/commands.py +0 -229
  81. agentx_dev-3.7.0/agentx_dev/cli/team.py +0 -317
  82. agentx_dev-3.7.0/agentx_dev/cli/ui.py +0 -646
  83. agentx_dev-3.7.0/agentx_dev.egg-info/top_level.txt +0 -8
  84. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/CONTRIBUTING.md +0 -0
  85. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/LICENSE +0 -0
  86. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/MANIFEST.in +0 -0
  87. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Agents/Agent.py +0 -0
  88. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Agents/__init__.py +0 -0
  89. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/AsyncTools.py +0 -0
  90. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/AutoSetup.py +0 -0
  91. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Cache.py +0 -0
  92. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Compiler.py +0 -0
  93. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Config.py +0 -0
  94. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Embeddings.py +0 -0
  95. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Handoffs.py +0 -0
  96. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Loader.py +0 -0
  97. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/MCP.py +0 -0
  98. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Memory.py +0 -0
  99. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Observability.py +0 -0
  100. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Operator.py +0 -0
  101. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Planner.py +0 -0
  102. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Runner/Persistence.py +0 -0
  103. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Runner/__init__.py +0 -0
  104. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Runner/promptTemplate.yaml +0 -0
  105. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Session.py +0 -0
  106. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Splitters.py +0 -0
  107. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Streaming.py +0 -0
  108. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/SupervisorMemory.py +0 -0
  109. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/Tools.py +0 -0
  110. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/VectorStores/__init__.py +0 -0
  111. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/WebTools.py +0 -0
  112. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/__main__.py +0 -0
  113. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/_optional.py +0 -0
  114. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/__init__.py +0 -0
  115. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/cli/safepaths.py +0 -0
  116. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/resources/__init__.py +0 -0
  117. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev/resources/promptTemplate.yaml +0 -0
  118. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev.egg-info/dependency_links.txt +0 -0
  119. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/agentx_dev.egg-info/entry_points.txt +0 -0
  120. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/README.md +0 -0
  121. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/agentic_rag_demo.py +0 -0
  122. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/async_example.py +0 -0
  123. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/async_quickstart.py +0 -0
  124. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/auto_features_example.py +0 -0
  125. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/caching_example.py +0 -0
  126. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/chatbot_example.py +0 -0
  127. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/complete_example.py +0 -0
  128. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/concurrent_example.py +0 -0
  129. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/concurrent_tool_example.py +0 -0
  130. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/file_agent_demo.py +0 -0
  131. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/function_calling_demo.py +0 -0
  132. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/mcp_demo.py +0 -0
  133. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/mcp_github_triage_demo.py +0 -0
  134. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/observability_example.py +0 -0
  135. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/orchestration_demo.py +0 -0
  136. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/planner_example.py +0 -0
  137. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/robust_link_scraper.py +0 -0
  138. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/subagents_demo.py +0 -0
  139. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/supervisor_codebase_analysis_demo.py +0 -0
  140. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/supervisor_example.py +0 -0
  141. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/sync_quickstart.py +0 -0
  142. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/v3_1_1_features_demo.py +0 -0
  143. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/v3_1_comprehensive_demo.py +0 -0
  144. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/examples/v3_1_features_demo.py +0 -0
  145. {agentx_dev-3.7.0 → agentx_dev-3.8.0}/setup.cfg +0 -0
@@ -0,0 +1,28 @@
1
+ # AGENTX.md
2
+
3
+ This repo is agentx_dev, a Python framework for tool-using agents, and `agentx`, its terminal app. Check the
4
+ Skills list before choosing tools and follow the skills that fit. This file is sent with every request: keep it short.
5
+
6
+ ## Layout
7
+ - `agentx_dev/Runner/`: the agent loop (`AgentRun.py`, `AsyncAgentRun.py`), trimming (`Trim.py`).
8
+ - `agentx_dev/cli/`: the terminal app (`app.py`, `commands.py`, `team.py`, `tools.py`, `context.py`, `ui.py`).
9
+ - `agentx_dev/Skills/` and `agentx_dev/resources/skills/`: the skill store and the built-in skills.
10
+ - `agentx_dev/ChatModel.py`: providers and token usage. `DefaultTools.py`: the file and command tools.
11
+ - `tests/`: pytest; `tests/conftest.py` has `MockModel` (scripted replies, no API calls).
12
+ - `docs/`: user docs. The full agent guide (guardrails, tools, Supervisor rules) is `docs/reference/agent-guide.md`.
13
+
14
+ ## Checks
15
+ - Run `python -m pytest -q` (about 3,300 tests, about three minutes) after a change and report the result.
16
+ - New behaviour gets a test with `MockModel`; never call a real API in tests.
17
+ - After editing `docs/`, run `python host/build_data.py`.
18
+ - Add a line to `CHANGELOG.md` under `[Unreleased]` for anything a user would notice.
19
+
20
+ ## Conventions
21
+ - Plain, short English in messages, docs and comments. Keep the existing style.
22
+ - Read a file before editing it; read only the lines you need from big files (offset and limit).
23
+ - Print only what you need from `run_python` and `run_shell`; output is cut at 30,000 characters.
24
+ - For code structure (functions, classes, counts) use Python's `ast`, not string matching.
25
+ - If a tool warns about a repeated call, change approach instead of repeating it.
26
+ - Your final reply is what the user sees: give the facts found, with file names and line numbers.
27
+ - Always commit your changes after modifying files or adding new ones. Run the relevant checks first; commit only your own changes.
28
+ - Commits: no `Co-Authored-By` lines.
@@ -4,6 +4,422 @@ All notable changes to `agentx-dev` are documented here. Format loosely
4
4
  follows [Keep a Changelog](https://keepachangelog.com/); versioning is
5
5
  [Semver](https://semver.org/).
6
6
 
7
+ ## [Unreleased]
8
+
9
+ ## [3.8.0] - 2026-11-08
10
+
11
+ ### Added
12
+
13
+ - **One conversation per device in remote use.** Each device that opens the window (an iPad, a laptop, the
14
+ computer's own window) gets its own conversation and agent, with its own approvals, Stop and status, working at
15
+ the same time as the others in the same folder; before, every window shared one conversation, so a second
16
+ device's message slipped into the running task. The computer's conversation (the terminal's, or the first
17
+ window's) is listed on the others as **Live** and can be opened to share it, as before. `/exit` on a device ends
18
+ only its own; one unused for 15 minutes is put away and reopened when the device comes back; settings saved on
19
+ one device reach all; a window says when another conversation is working in the folder. Inside: each
20
+ conversation has its own team lane (a team job's questions, notes and `/skip` never reach another's), the
21
+ conversation titles, work log and team records take several writers, and one MCP sign-in runs at a time.
22
+ - **`agentx --ui`: the app window.** The same agentx in a desktop-style window instead of the terminal: the
23
+ conversations of the folder on the left (open one, or start a new one), your messages on the right, the agent's
24
+ steps as short rows (Read, Edit, Run, with look-around steps folded into one Explored row), answers rendered as
25
+ markdown, approvals as cards with the diff and Yes / don't ask again / No (with a reason) and keys 1-3, the other
26
+ questions as cards, the team's running steps, typed messages waiting with a button to take each back, and a
27
+ message box with the mode and model menus, `/` for commands and skills, a picture button (and paste or drop;
28
+ saved in `.agentx/uploads/`), a microphone button (`/voice`) and a send button that is Stop while it works.
29
+ It opens as an Edge or Chrome app window when one is installed, else in the browser. The app itself runs
30
+ unchanged; the window is another front end, served on 127.0.0.1 with a random token per start and a Host check.
31
+ `--port` and `--no-open`. No new packages.
32
+ - **The app window's side panel, Settings and remote use.** **Details** shows how full the conversation is, the
33
+ tokens used, the last team job step by step, the MCP servers by name with their own Sign in / Reconnect / Tools /
34
+ Sign out buttons, and the project memory. **Settings** saves the default model and mode, the allow and deny rules,
35
+ the steps and the token limit, for all folders or this folder (rules and limits apply at once). The **+** menu adds
36
+ a picture, mentions a file (search) or opens another folder in a new window. `agentx --ui --remote` keeps the window
37
+ 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
38
+ the system. A notice that MCP servers need sign-in has a button per server.
39
+ - **Remote control.** `/remote-control` in the terminal makes the running session usable from an iPad, a phone or
40
+ another computer: the terminal keeps showing everything, the window mirrors it, and messages, answers, menu choices
41
+ and approvals from the window go into the terminal's own prompt (the first answer wins). In the app window it adds
42
+ the remote ways to the window already served. With Tailscale on, `--remote` and `/remote-control` also listen on
43
+ the computer's Tailscale address (Tailscale devices only) and print a link for Safari; otherwise the SSH way. They
44
+ say when no SSH server runs, or when Windows Firewall blocks SSH on a Public network.
45
+ - **The model's thinking, shown.** On Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 agentx asks for a readable summary of
46
+ the reasoning they do on every step anyway (`thinking.display: "summarized"`, same cost; nothing is sent back on
47
+ later requests), and on GPT reasoning models on the Responses API for OpenAI's reasoning summary (asked once more
48
+ only if the account allows it). The app window shows it, and the model's notes between steps, as Thinking blocks
49
+ as they happen, with a live Thinking… line; the terminal shows it before an edit or a command. In the library:
50
+ `Claude(show_reasoning=True)`, `GPT(show_reasoning=True)`, `"reasoning"` in a `call_with_tools` result, and a
51
+ `{"type": "reasoning"}` runner event.
52
+ - **MCP sign-in from a remote window.** The sign-in page is a card in the window, and pasting the address of the
53
+ page that could not load afterwards finishes the sign-in on the agentx computer.
54
+ - **Conversation titles and closing the window.** After its first answer, a conversation gets a short title from the
55
+ model (one small request, on a thread of its own), shown in the app window's list and in `/resume`; older ones
56
+ show the first thing you typed, no longer agentx's summary note. Closing the app window quits agentx (a running turn
57
+ stops and is kept, the conversation is saved); a reload does not. Not with `--no-open` or `--remote`.
58
+ - **`agentx`: commands while it works, and `/unsend`.** Commands that only show something (`/usage`, `/settings`,
59
+ `/help`, `/team log`, `/skills`, `/mcp`, `/memory`, `/sessions`, and `/model`, `/team-models`, `/limit`,
60
+ `/fullscreen` with nothing after them) run at once while the agent works, and `/mode` changes the mode at once;
61
+ they used to wait (`runs when the agent is done`). `/voice` starts recording at once: Enter stops it and the words
62
+ go into the line being typed, to check and send. `/unsend` takes back a waiting message (or a team note) before it
63
+ is read: the newest, `/unsend 2`, or `/unsend all`; the waiting lines are numbered. Commands that change the
64
+ conversation or the setup still wait. The agent and the team's runner are told not to wait inside a command
65
+ (`sleep`, `--watch`): they check CI or a deploy once and report. The zero-to-hero guide has `/init` as step 4.
66
+ - **`agentx`: `/voice`.** Records from the default microphone until Enter (Ctrl+C cancels), writes out what was said
67
+ with OpenAI's speech-to-text (`gpt-4o-transcribe`, falling back to `whisper-1`; `AGENTX_VOICE_MODEL` picks another)
68
+ and puts the words in the next input line to check before sending, in the normal and the full-screen screen.
69
+ Claude models take no audio, so it needs `OPENAI_API_KEY` whatever the chat model. The microphone is read with
70
+ `sounddevice`: new extra `agentx-dev[voice]`. A recording stops by itself at 10 minutes; a microphone that cannot
71
+ record at 16 kHz records at its own rate.
72
+ - **Agents can look at pictures.** `read_path` on a PNG, JPEG, GIF or WebP file hands the model the picture itself
73
+ (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
74
+ works on Claude (inside the `tool_result`) and GPT (Chat Completions and Responses: the tool message gets the text,
75
+ and a message right after the tool results carries the picture). The type comes from the file's contents; a
76
+ picture over 5 MB or 8,000 pixels is made smaller with Pillow when it is installed, and refused otherwise. Each
77
+ request carries the newest three pictures from tool results; older ones become a note naming the file (the history
78
+ keeps them). Logs, the screen and team summaries show one line (`Picture shot.png: PNG, 1440x900, 245 KB.`); in
79
+ `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
80
+ module `agentx_dev.Pictures` (`picture_result`, `ToolContent`) for tools of your own.
81
+ - **`agentx`: type while the agent works.** On the normal screen what you type while the spinner runs shows under it,
82
+ and Enter sends it to the agent at its next step, so it can change course without waiting. During a team job a
83
+ message goes to the team: running helpers read it at their next step, later steps start with it, and the main
84
+ agent reads it in the job's result (to add work). `/skip <step number or helper>` skips a team step not started
85
+ yet. Another command typed then runs when the agent is done. A sent message waits under the spinner (`◌ ...`) and,
86
+ once read, leaves one line: `❯ text ✓ read by <the agent or the helper>`. The spinner says `type to add to the
87
+ task`. Full-screen mode does the same (a message typed
88
+ there used to wait for the end of the turn). Approval prompts still get every key.
89
+ - **`AgentRunner.interjections` and `AsyncAgentRunner.interjections`.** Set it to a function that returns the messages
90
+ a user sent while the run was going; each is added to the run before the next step, and `stream` yields
91
+ `{"type": "interjection", "content": ...}` for it.
92
+
93
+ - **`log=` on `AgentRunner` and `AsyncAgentRunner`: watch a run as it goes.** `log=True` prints one short line per
94
+ step to stderr; a file path appends JSON lines (`.jsonl`) or readable lines; a `logging.Logger` sends them to your
95
+ logging setup; a function receives each event. Lines: the start, each model call (time, tokens in, cached and
96
+ out), each tool result (ok or error, arguments and result cut short), and the end (outcome and totals) or the
97
+ error. Secrets are redacted, and a log that cannot be written never stops the run. New `RunLog` class
98
+ (`RunLog.chain` writes to several). See "Watching one runner" in the observability guide.
99
+ - **The eval harness catches up.** `EvalRunner` runs an `AsyncAgentRunner`, a `Supervisor` or an `AsyncSupervisor`
100
+ as well as an `AgentRunner` (async ones from a notebook too); new checks `outcome_is`, `output_field`,
101
+ `used_skill`, `used_agent`, `subtasks_ok` and `EvalCase(max_tool_errors=)`, with the same fields in JSON cases;
102
+ `repeat=` runs each case several times (the summary shows `[2/3]`), `tags=` picks cases; results carry
103
+ `counted_tokens` (cached input as a tenth), `model_calls` and the failed tools; `compare_reports` and
104
+ `python -m agentx_dev.Evals compare a.json b.json` put two reports side by side; the CLI takes `--repeat`,
105
+ `--tag` and `--log`.
106
+
107
+ - **Public extension points on `Supervisor` and `AsyncSupervisor`:** `prepare_plan(plan)` (change a plan before it
108
+ runs), `run_step(agent_name, query, run)` (wrap or skip every step), `spawn_specialist(spec)` (build the helpers a
109
+ plan defines your own way) and `planner_note` (text added to the planner's list of specialists). A planned helper's
110
+ `new_agent` may name a `model`, kept by the plan checker and passed as `AgentSpec.model`. The `agentx` team runner
111
+ now uses only these, no underscore methods. See "Changing how a Supervisor runs" in the sub-agents guide.
112
+
113
+ ### Changed
114
+
115
+ - **`Claude()` defaults to `claude-opus-5-5` with `max_tokens=16000`** (was `claude-sonnet-4-6`, 4096). Opus 5.5
116
+ always thinks and the thinking counts toward the limit, so 4096 could cut answers short. `agentx`'s help text and
117
+ hints name `claude-opus-5-5` and `gpt-6.1-sol`. See Upgrading.
118
+
119
+ - **`Supervisor` and `AsyncSupervisor` share one core.** Planning, replanning, synthesis, a step's retries and the
120
+ rounds of a session are written once (they were two copies that had to be fixed twice). `AsyncSupervisor` awaits
121
+ real work; `Supervisor` runs the same code with sync calls, still in your thread, with no event loop, so it keeps
122
+ working in a notebook. Each class keeps its own way of running a plan: one step at a time, or side by side. No
123
+ change in behaviour: the same events and log lines.
124
+
125
+ - **Smaller source files.** `Supervisor.py` (2,900 lines) keeps the two classes; its helpers moved to the private
126
+ package `agentx_dev/_supervisor/`. The `agentx` app's `ChatApp` gets its MCP, model, usage and team commands from
127
+ four files (`cli/app_mcp.py`, `app_models.py`, `app_usage.py`, `app_team.py`), and `cli/ui.py` its text and input
128
+ helpers from `ui_text.py` and `ui_input.py`. Every name is still importable from where it was.
129
+
130
+ - **The public API is written down** in `docs/reference/public-api.md`: every name exported from `agentx_dev`, the
131
+ stable import paths below it, what is not public (underscore names, the `agentx` command's Python modules), and
132
+ how patch, minor and major versions may change it. A test fails when the package and the page disagree.
133
+
134
+ - **A smaller base install.** `chromadb`, `qdrant-client`, `psycopg`, `pandas` and `openpyxl` moved from the base
135
+ dependencies to extras: `[rag]` (the three vector stores; `[chroma]`, `[qdrant]`, `[pgvector]` one at a time),
136
+ `[excel]` (spreadsheets in `Media`), `[all]`. A missing package's error names the extra. See Upgrading.
137
+
138
+ - **The package says what it is.** The PyPI summary, keywords, README opening, docs index and docs site name both
139
+ halves: the framework and the `agentx` terminal assistant, with helpers that run in parallel. New cookbook pattern
140
+ 40, "Let helpers work in parallel", for the terminal app and for `AsyncSupervisor` in your own code.
141
+
142
+ ### Fixed
143
+
144
+ - **The iPad's link keeps working after the computer sleeps.** Tailscale could stop sharing agentx's port (after a
145
+ sleep or a Tailscale restart) while agentx still ran, so the link was dead. agentx now checks every 30 seconds
146
+ and shares it again (the same link), saying so in the terminal, and at start clears entries earlier runs left
147
+ shared with nothing behind them. The tests no longer reach the Tailscale of the machine they run on.
148
+ - **The ssh command for another computer uses the Wi-Fi address again.** With Tailscale on, it used this computer's
149
+ Tailscale address (100.x), which times out from a laptop without Tailscale; the Tailscale address is now offered
150
+ as the one to use from a computer with Tailscale on.
151
+ - **The app window shows agentx at work, as the terminal does.** In `/remote-control` from the terminal it said
152
+ "Waiting for agentx…" with nothing moving: now the spinner line (what it does, the time and the tokens this
153
+ message has used) shows there too, starts the moment a message is sent, and a soft light runs along the message
154
+ box while it works. A conversation starts with the agentx header (version, model, folder), like the terminal's
155
+ box, and the link and ssh steps of `/remote-control` stay in the terminal instead of filling the window.
156
+ - **"Open in Safari" opens the app window, not a box saying to start agentx.** The page took its key out of
157
+ the address, so opening it in another browser, a new tab or from a bookmark had none, and the box could not be
158
+ closed. The key now stays in the address after a `#` (never sent to the server). Without a key, or with one from
159
+ an earlier start, the box says which, and takes the link agentx printed pasted in.
160
+ - **`agentx --remote` opens the terminal with remote control on.** Before, `--remote` needed `--ui` and gave the app
161
+ window only; now `agentx --remote` is `agentx` followed by `/remote-control` (`--ui --remote` still serves the
162
+ window only).
163
+ - **When Tailscale gives no link for the iPad, agentx says so first.** HTTPS or Serve turned off in the Tailscale
164
+ account (also after it worked) left only the 127.0.0.1 address, with the reason under the SSH steps; now the reason
165
+ and where to turn it on come first, Tailscale's own link to allow Serve is passed on, and agentx no longer waits
166
+ 20 seconds for Tailscale there.
167
+ - **The app window's Details panel fits the screen.** A long MCP server name or address stretched it, so on an iPad
168
+ it scrolled sideways; long names and addresses now wrap inside it. Below a wide desktop window (an iPad, a
169
+ half-screen window) it slides over the chat instead of squeezing it, and tapping beside it or Escape closes it.
170
+ - **`agentx`: a turn that stops before it finishes is kept.** Out of credit, any other error, Ctrl+C or the token
171
+ limit used to drop the whole turn, so after switching model (say Claude ran out of credit) the new model did not
172
+ know what the stopped turn had changed. Now your message and a short record (files changed, commands run, the last
173
+ thing it said, why it stopped) go into the conversation and the work log, and `continue` picks up from there, on
174
+ the same model or after `/model <name>`. An out-of-credit error says so plainly and what to do instead of
175
+ printing the raw `BadRequestError`. Leaving prints how to come back: `agentx --continue` (with `--model <name>`).
176
+ - **`run_shell` and `run_python` stop at their time limit.** On Windows the limit killed only the shell and then
177
+ waited, with no limit, for programs it had started (`ssh`, `sleep`, a dev server) to close the output, so a 5-minute
178
+ limit could run for as long as such a program did (`Running run_shell... 7m 28s`). The whole process tree is now
179
+ stopped at the limit, and the timeout message keeps what the command printed before it. A command that ends but
180
+ leaves a program running in the background without redirecting its output (`cmd &`) returns at once with a note
181
+ on how to start it (`nohup cmd > log.txt 2>&1 &`) instead of holding the tool until the limit. Commands get an
182
+ empty stdin, so none can read the terminal. Ctrl+C while a command runs now stops it and everything it started
183
+ (a hung `ssh` used to keep the turn waiting after "Turn interrupted").
184
+ - **A helper that calls a tool it does not have is told so and carries on.** A "final checks" step that also had to
185
+ fix what failed went to the runner (it runs commands but cannot edit); its call to `edit_file` ended the step with
186
+ "(agent emitted an unrecognized action and no answer text)", shown as Done. agentx's agent and helpers now get the
187
+ framework's `strict_tool_dispatch` reply (the tool is not theirs, here are the ones they have), a step that still
188
+ ends with no report is marked failed, and the planner is told to give a step only to a helper with every tool it
189
+ needs.
190
+ - **A team job uses the GPT models in `teamModels` too.** With a Claude chat model and `gpt-6-luna` in the list
191
+ (its key found), every step of every job ran on a Claude model: the planner was told "no model is preferred" and
192
+ picked its own provider. It is now told the user listed these models so the work is shared among them, and, when
193
+ the list has both Claude and GPT models, to give at least one step to each provider.
194
+ - **Team steps start in plan order.** When a step finished and freed two steps at once, the one that had been waiting
195
+ longer went first: step 3 ran before step 2. A step now gives way to an earlier step that is ready at the same
196
+ time (for at most 3 seconds, so a step that never starts cannot hold up the rest).
197
+ - **The spinner shows minutes**: `Team is working... 11m 23s`, not `683s` (and `1h 02m` past an hour).
198
+ - **`agentx --model claude-opus-5.5` works.** A Claude name written with a dot got a 404 ("model: claude-opus-5.5
199
+ was not found") on the first message; it is now read as the dashed id (`claude-opus-5-5`). GPT names keep their dot.
200
+ - **A `teamModels` note may be 200 characters** (was 80, and a longer one was dropped at every start). A note past
201
+ 200 is now shortened, with a warning, instead of dropped.
202
+ - **A step with no action and no answer no longer ends the run.** A reply cut off by the output limit (seen with a
203
+ Claude Sonnet 5.5 helper after 20 tool calls) came back empty and the run ended with "(agent emitted an
204
+ unrecognized action and no answer text)". The model is now asked again, up to twice per run.
205
+ - **`agentx` finds an API key saved in Windows after the terminal was opened.** A team model on the other provider
206
+ showed `Not used: gpt-... (no OPENAI_API_KEY)` when the key was set with `setx` or the Environment Variables dialog
207
+ after the terminal started. agentx now reads a missing `ANTHROPIC_API_KEY` or `OPENAI_API_KEY` from the saved user
208
+ (then system) variables at start, before each team job and on `/model`, and says which it used (never the value).
209
+ - **A team job's task may be 12,000 characters** (was 4,000): the main agent's detailed tasks were refused twice in
210
+ one job.
211
+ - **Claude: every agent request was declined on claude-opus-5 and claude-opus-5-5.** The agent's step tool asked
212
+ for a `Thought` (the model's reasoning, written out), and these models decline a request that asks for their
213
+ reasoning (`stop_reason: "refusal"`, category `reasoning_extraction`) before writing anything: `agentx` showed
214
+ nothing at all, or "unrecognized action". The step tool is now sent to Claude without its `Thought` field
215
+ (GPT keeps it). And a decline, or `max_tokens` used up with nothing written, now raises `ModelDeclined` with the
216
+ provider's reason (also in `agentx.log`) instead of returning an empty answer.
217
+ - **Evals: failed tool calls were never counted.** The count looked for results starting with "ToolError" in
218
+ `completion.tool_calls`, which lists only the calls that worked, so it was always 0. It now counts them from the
219
+ run log, recovered ones included.
220
+ - **`agentx`: an installed skill is ready at once.** `/skills install` and `/skills remove` used to wait for a new
221
+ conversation (`/clear`, or a restart); now the conversation you are in gets the change, with its history kept.
222
+ `/skills reload` does the same for a skill folder copied in by hand.
223
+ - **`agentx`: Claude models use prompt caching.** `agentx` built Claude models without it, so every step of a long
224
+ run sent the whole conversation again at full price and counted it in full against the token limit (one helper
225
+ used 229k tokens on 8 tool calls and stopped part-way). The system prompt, the tools and the conversation so far
226
+ are now read from Anthropic's cache on the next step, billed and counted at about a tenth.
227
+ - **`agentx`: a run that hits the step limit shows what it did.** It ended with nothing on the screen; it now shows
228
+ the summary and says to type `continue`. The limit is 300 steps a message (was 50): `"maxSteps"` in your settings
229
+ or `--max-steps` changes it. A short "continue" (or "go on", "carry on", "keep going") right after that stop
230
+ tells the agent to pick up from its last step instead of re-running the checks already in the conversation.
231
+ - **`agentx`: a team job saves every helper's work as it goes**, in `.agentx/team/<date>-<task>/` (`job.md` and one
232
+ file per step with the helper's whole report), and the finished reports join the project's `recall` index: a
233
+ closed terminal no longer loses the research.
234
+ - **`agentx`: team steps are shown by role.** The planner gives every step a name for what its helper does
235
+ (`git-checker`, `payments-fixer`, `migration-verifier`), shown in the plan, the live block, the done lines and
236
+ `/team log` in place of `runner` / `coder` (`git-checker@claude-sonnet-5-5`). The step still runs on the
237
+ built-in helper; a step without a name shows the helper's as before.
238
+ - **`agentx`: `/skills install` fetches only the skill files.** It asks GitHub for the repository's file list and
239
+ downloads just the skill folders' files, so a big repository costs what its skills weigh (`morluto/rea`, 6,677
240
+ files and 130 MB, was refused as "too many files"; its skill is 8 files). Skills in hidden folders such as
241
+ `.agents/skills` are found (they were skipped), a `/blob/.../SKILL.md` link installs that skill, and the archive
242
+ fallback counts only the files it unpacks.
243
+ - **`agentx`: `/skills install` handles repositories with big demo files.** Only the folders holding a `SKILL.md`
244
+ (or the folder the link names) are unpacked, so a repository's videos and docs no longer count: the download may
245
+ be up to 100 MB, and a skill's own files up to 10 MB each (music, fonts), 50 MB in all.
246
+ `latent-spaces/brag` (65 MB, mostly demo videos) was refused as "bigger than the limit". Names before the link
247
+ install only those skills (`/skills install brag <link>`), a misspelt subcommand gets "Did you mean /skills
248
+ install?", and an over-limit download says what the limit is.
249
+ - **Structured output on Opus 5.5, Sonnet 5.5 and Fable 5.1.** These models answer a forced tool choice with a
250
+ 400, which `output_schema` and the answer parser used. The call now goes with `tool_choice: auto` and a line
251
+ asking for the tool; another model that refuses it is found from the error and remembered.
252
+ - **`agentx`: a saved model whose key is gone no longer stops the start.** When the other provider's key is set
253
+ (you moved from GPT to Claude, say), `agentx` lists that provider's models (Claude: `claude-opus-5-5`,
254
+ `claude-sonnet-5-5`, `claude-haiku-5-5`, `claude-fable-5-1`) or asks for a name, uses the one picked, and offers
255
+ to save it in the settings file the old name came from.
256
+ - **An OpenAI model that needs the Responses API for tool calls is found out once per program, not once per model
257
+ object.** Each new `agentx` session, team helper and benchmark task used to send one chat-completions request that
258
+ the provider rejected before switching. Now the switch is remembered for that model at that address until the
259
+ program exits.
260
+ - **Benchmark:** a result file keeps the whole model name (`gpt-6.1-sol` was saved as `gpt-6.json`), and an answer
261
+ check says which words it found instead of "the answer covers it". The first baseline, 10/10 on `gpt-6.1-sol`,
262
+ is in `bench/results/`.
263
+ - **`agentx`: cached input counts as a tenth toward the token limit and each team step's share.** A helper's
264
+ requests are mostly input the provider serves from its cache, billed at about a tenth, but they were counted in
265
+ full: in one audit each explorer was cut off after about 15 calls at "220k tokens", reported partial, and the agent
266
+ started two more team jobs. The limit question also says how many tokens were sent in full, and the team block and
267
+ done lines show the cached share (`230k tokens, 80% cached`).
268
+ - **`agentx`: helpers can open `auditing-a-project`** (version 2; it was listed only for the main agent, so helpers
269
+ told to follow it searched the disk for it). A step's report line drops markdown marks (`##`, `**`).
270
+
271
+ - **`agentx`: a no at the token-limit question now ends a team job cleanly.** New steps still started after the no
272
+ (three explorers at once, in one audit), and skill reads were not refused, so wrapping up ran past its allowance and
273
+ the whole turn was thrown away. Now no step starts after a no (it is listed as `not started`), `use_skill` is refused
274
+ like every other tool, and the team reports with what it has.
275
+ - **`agentx`: files are read with the file tools.** The base prompt and the runner helper say to read files with
276
+ `read_path` and search with `grep`/`find_files`, not print them with `run_python` or `run_shell` (one audit's runner
277
+ spent its whole token share printing files with Python). Helpers open a skill only when their step names it or it
278
+ covers exactly that step.
279
+
280
+ - **`agentx` team helpers no longer read the same files over and over.** Helpers kept 12,000 to 24,000 characters of
281
+ tool results whole, lost the files they had read and read them again (eight to ten times each on one audit, about
282
+ 2.5 million tokens and twelve limit questions). They now keep 20,000 to 40,000 like the chat, and a third whole read
283
+ of a file that has not changed in the step gets a note instead (an edit or any command resets it; a read with
284
+ `offset`/`limit` always works).
285
+ - **A team's report reaches the chat whole.** The result text is cut at 20,000 characters (was 6,000, which cut an
286
+ audit's findings and made the agent start a second job).
287
+
288
+ - **The wheel installs only `agentx_dev`.** `examples` and `host` (the docs site) were installed as top-level
289
+ packages too; they stay in the source download. The license is now an SPDX expression (`MIT`, with
290
+ `license-files`), which needs setuptools 77 or later to build.
291
+ - **An MCP sign-in survives a restart.** The SDK loaded a saved token as never expiring, sent it after it had run
292
+ out, and asked for a new sign-in instead of renewing it; after a restart it also guessed the renewal address
293
+ (`<server>/token`, a 404 for Supabase). agentx now keeps when the token runs out and the server's OAuth details from
294
+ the sign-in, and hands both back, so the token is renewed at the right address.
295
+ - **MCP sign-in to servers that issue a client secret (Supabase).** A secret given without a method is now sent
296
+ in the token request; the sign-in failed with "Required parameter: client_secret".
297
+ - **One stuck request no longer waits about 12 minutes.** A timed-out request is retried once at most (the client
298
+ has usually retried it already; a timed-out request may still be billed), and a 429 waits for its `Retry-After`.
299
+ - **A spent budget is not retried.** `CostBudgetExceeded` (and anything `model.on_usage` raises) is raised once; on
300
+ the plain chat path the call used to be retried up to three times after the budget was spent, paying each time.
301
+
302
+ ### Added
303
+
304
+ - **A benchmark for `agentx`** (`python -m bench`): ten fixed tasks on two sample repos (find, explain, fix, fix and
305
+ test, validate, add a feature, rename, audit, a team job, run and report), each checked automatically (the tests,
306
+ hidden tests, the files), with pass/fail, time, model requests, tokens, the cached share and tool calls recorded;
307
+ `compare` sets two runs side by side. It asks before it uses your API key. See `bench/README.md`.
308
+ - **`agentx`: `/mcp reload [name]`** (also `/mcp refresh`): reads `~/.agentx/mcp.json` again and reconnects every
309
+ server (or one) with the saved sign-ins, never opening the browser; servers taken out of the file are dropped and
310
+ the conversation gets the new tools. On Windows it also reads the user's saved environment variables, so a key set
311
+ with `setx` after the start is found. `MCPManager.reload(name=None, env=None)`.
312
+ - **`agentx`: full-screen mode** (`agentx --fullscreen`, `/fullscreen on|off`, `"fullscreen": true` in
313
+ `~/.agentx/settings.json`; off by default). The welcome box, the input and the status line stay put while only the
314
+ conversation scrolls (mouse wheel, PgUp / PgDn, Ctrl+End). Questions are asked above the input; a message typed
315
+ while the agent works is kept for later; Ctrl+C and Ctrl+D work as before; on exit the conversation is written to
316
+ the terminal. The app runs unchanged on a worker thread: its console writes into the screen, and the spinner is
317
+ drawn by the screen (`Progress.drawn_by`).
318
+ - **`agentx`: the agent looking around is one tree.** Skill reads, `recall`, reads, searches, listings and web calls
319
+ in a row are drawn as one flow (`● Skills using-superpowers` / `├─ recall ...` / `│ ⎿ Found 2 notes` /
320
+ `└─ read_path ...`), redrawn in place above the spinner while it runs (a blinking yellow dot, the running step
321
+ 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
322
+ an edit or a command still is.
323
+
324
+ - **`agentx`: the supervisor can make helpers.** A plan step may define a helper (`new_agent`: a name, instructions,
325
+ tools). agentx builds it the way it builds explorer, coder and runner (gated tools, step
326
+ wrapping, redaction, 100 steps, trimming, skills); its tools come only from the gated set, never `delete_path`,
327
+ `move_file`, `start_team_job` or `delegate`, and it gets only the MCP tools it names. A helper with only reading
328
+ tools runs beside other reading steps. At most 6 per job; the plan shows each one and its tools. With helper
329
+ models set (`/team-models`), it may run on one of them (`"model"` in `new_agent`, read before the plan checker drops
330
+ it): the planner is told the list, a model not on it is not used, and the name shows it (`reviewer@gpt-6-luna`).
331
+ - **`Skills(guard=fn)`**: `fn()` is asked before each `use_skill` read; a text it returns is the answer instead.
332
+ - **`agentx`: reading steps of a team job run side by side.** Explorer steps that do not depend on each other run
333
+ up to 3 at once (`AsyncSupervisor` with `max_parallel=3` and a lock: a coder or runner step runs alone). A plan
334
+ without `depends_on` keeps its order, with neighbouring explorer steps grouped. Helpers run in worker threads; every
335
+ approval, team question, limit question and screen line is handed to the main thread, so prompts never overlap,
336
+ and Ctrl+C stops the job at once (a helper stops at its next tool call or model reply).
337
+ - **`agentx`: each team step has a token share,** one token limit's worth (at least 100,000); past it the step's
338
+ tools tell the helper to report what it has.
339
+ - **`agentx`: `a` at the token-limit question** goes on and does not ask again in this message.
340
+ - **`agentx`: team helpers work in the background.** Their tool calls are no longer printed. A block above the
341
+ spinner, redrawn in place, shows one line per running step (helper · step name, what it has done, time, tokens);
342
+ a finished step leaves one line (`✓ explorer · security audit Done (20 tool uses · 41k tokens · 1m 30s)`) and the
343
+ start of its report. `/team log` lists every call of the last team job. Steps are named by their plan id.
344
+ - **The `auditing-a-project` skill** (bundled): map, run the project's own checks, read by area with explorers side
345
+ by side, verify, and report Fix first / Medium / Low / Done well / Checks / Test gaps. `delegating-work` (version
346
+ 2) says which steps run side by side.
347
+
348
+ - **Smaller requests.** `AgentRunner(..., trim_results=True)` (and `AsyncAgentRunner`, or a `TrimPolicy` from
349
+ `agentx_dev.Runner.Trim`) sends the model a one-line stub in place of each big tool result outside a window of the newest
350
+ 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
351
+ arguments of old calls; the saved history is unchanged and `use_skill` results are never
352
+ trimmed. `Permissions.read_max_chars` makes `read_path` return a long file in numbered parts with a note on how
353
+ to continue. `Skills(remember_reads=True)` answers a second read of a skill with a pointer back.
354
+ - **`agentx`:** all of the above are on (reads in parts of 30,000 characters); summarising starts at 60,000
355
+ characters (was 150,000) and is measured on what the model is sent. On a measured 15-file turn the input sent
356
+ fell from about 231,000 tokens to about 136,000. `/usage` shows the tokens used this session, per model, and by
357
+ the last turn.
358
+ - **Fewer whole-file re-reads.** A trimmed `read_path` result now names the file and its line count (or the line
359
+ range of a slice) and says to grep it or read only the lines needed; `agentx`'s base prompt and the
360
+ `exploring-a-project` skill (version 3) tell agents to write down the facts they will need as they read.
361
+ - **Cache-friendly trimming.** `TrimPolicy(step_chars=)`: once the whole results pass `keep_chars`, the oldest are
362
+ stubbed in one go down to `keep_chars - step_chars` (default half), so the start of the request stays the same,
363
+ and cacheable, until the next step; `keep_last` is now 1. Team helpers keep 12,000-24,000 characters. On a replayed
364
+ 40-step team brainstorm the input cost fell by about 28-41%, depending on the cached price.
365
+ - **A short `AGENTX.md`.** This repo's `AGENTX.md` is now about 1,800 characters (was 13,300), since `agentx` sends
366
+ it with every request; the full guide moved to `docs/reference/agent-guide.md`.
367
+ - **Older skill texts are stubbed in a long conversation.** `TrimPolicy(shrink_exempt_at=60_000)`: past that many
368
+ characters of conversation, every `use_skill` result but the newest is sent as a note that the skill still applies
369
+ and how to reopen it. In `agentx`, `run_python` and `run_shell` output is cut at 30,000 characters (was 100,000 and
370
+ 40,000), with a note to print less. Replaying a 21-step message that read five skills, the input sent fell by
371
+ about 15%.
372
+ - **OpenAI cached tokens are counted** (`TokenUsage.total_cache_read_tokens`, from either API), and `/usage` shows how
373
+ much of the input came from the provider's cache.
374
+ - **A token limit per message in `agentx`.** Past 200,000 tokens (`/limit`, `"tokenLimit"` in settings; a project
375
+ may only lower it) agentx asks before it goes on; a no wraps up (tool calls are refused, the agent answers with
376
+ what it has, the turn is kept), and only a wrap-up that runs long too ends the turn like Ctrl+C.
377
+ `BaseChatModel.on_usage`, a hook called after every recorded reply, is what it uses.
378
+ - **Web tools in `agentx`.** `web_search` and `web_fetch` (readable text, at most 30,000 characters, marked as data
379
+ from the internet); the gate asks before each in `default` and `plan` mode (`a`: the rest of the session). The
380
+ `web-research` skill is listed, and the team's explorer gets them.
381
+ - **MCP servers in `agentx`.** `/mcp add <name> <url>` (`--sse`), `/mcp add <name> -- <command> [args]`, `/mcp remove`,
382
+ `/mcp login`, `/mcp logout`; kept in `~/.agentx/mcp.json`, `${VAR}` for keys. A browser sign-in (OAuth) runs only on
383
+ `/mcp login`; tokens go to the system's credential store (`keyring`, split across entries when long) and renew
384
+ by themselves. Tools appear as `<server>__<tool>` for the chat and every team helper, ask before running even in
385
+ auto mode, and their results are marked as outside data and cut at 30,000 characters. The `mcp` extra now needs
386
+ `mcp>=1.10,<2` (2.x changed the client API and is not supported yet) and `keyring>=24`.
387
+ - **MCP tools sorted by what they do.** The first time agentx sees a server's tools, the ones the server marks
388
+ read-only run without asking (saved per server in `mcp.json` as `allow`/`seen`); the rest ask. `/mcp tools <name>`
389
+ lists them, `/mcp allow|ask <name> <tool...|read-only|all>` changes them; allowing a destructive tool asks first.
390
+ - **Model requests in `agentx` wait up to 5 minutes and are retried visibly.** `model.on_retry(attempt, error,
391
+ delay)` is called before each retry; `agentx` prints a line.
392
+ - **Tool results are one line in `agentx`.** Under each call: `⎿ Read 51 lines`, `⎿ Found 2 files`, `⎿ 3 matches in
393
+ 12 files searched`, `⎿ Found 1 note`, a command's first line of output and how many more, or its exit code and
394
+ last line in red when it failed. Refusals and errors are still shown in full.
395
+ - **Skill reads are one line in `agentx`.** Their text is never printed; a run of them shows on the spinner as it
396
+ grows and then stays as `● Skills a → b → c (+ file)`, with the notes written meanwhile after it. A failed read
397
+ is still shown.
398
+ - **A new `agentx` start screen.** A box with the agentx mark, the model and folder, tips for this folder and
399
+ what's new; MCP sign-in problems as one warning line; notes with a bar down the left; a rule above the input
400
+ and the mode under it (`⏵⏵ auto mode on (shift+tab to cycle) · model · folder`).
401
+ - **`agentx` input colours:** the `/name` at the start of the input line is coloured as you type (command cyan, skill
402
+ magenta, unknown red) and the selected row of the `/` pop-up is highlighted.
403
+ - **Skills first.** The skill list tells the model to check it before choosing tools and to prefer a skill the user
404
+ installed when it fits; the user's own skills are listed ahead of the built-in ones, and the terminal app's base
405
+ prompt says the same.
406
+ - **`agentx` project notes and recall:** a readable note per worthwhile turn in `.agentx/notes/` (asked, result,
407
+ files changed, commands), indexed in `.agentx/memory/` with OpenAI embeddings when `OPENAI_API_KEY` is set
408
+ (local similarity search otherwise, keywords if both fail). The agent gets a `recall` tool, a fresh
409
+ conversation's first message gets up to 3 close notes, and `/memory [words]` shows or searches them.
410
+
411
+ - **Skills.** `AgentRunner(..., skills=True)` (and `AsyncAgentRunner`) gives an agent a `use_skill` tool and a short
412
+ list of skills at the end of its system prompt; the model reads a skill's full text only when it needs it. The
413
+ list is built the same way every time, so the prompt stays identical for the provider's prompt cache. Five
414
+ skills ship with agentx (`exploring-a-project`, `changing-files`, `running-code`, `web-research`,
415
+ `delegating-work`) and are created in `~/.agentx/skills/` on first use; edits and deletions are respected. The
416
+ agent's own tools are described by a generated `custom-tools` skill. `SKILL.md` folders written for Claude work
417
+ as they are. See the new [guide](docs/guides/skills.md).
418
+ - **`agentx` terminal app:** skills are on by default; `/skills` lists them, `/skills install <github link or
419
+ folder>` installs after showing what it found (the exact commit is downloaded and recorded; archives are
420
+ 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
421
+ agent has no tool for it, and `.agentx/*` and `~/.agentx` stay protected from its file tools.
422
+
7
423
  ## [3.7.0] - 2026-10-15
8
424
 
9
425
  ### Added