assistant 2.2.0b1__tar.gz → 2.2.2b1__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 (122) hide show
  1. {assistant-2.2.0b1/assistant.egg-info → assistant-2.2.2b1}/PKG-INFO +58 -7
  2. {assistant-2.2.0b1 → assistant-2.2.2b1}/README.md +56 -6
  3. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/__init__.py +6 -1
  4. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/agent.py +161 -30
  5. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/builtin_cmds.py +585 -79
  6. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/config.py +46 -0
  7. assistant-2.2.2b1/assistant/display.py +134 -0
  8. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/listen/interface.py +6 -7
  9. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/main.py +65 -163
  10. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/memory.py +87 -3
  11. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/__init__.py +1 -0
  12. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/google.py +2 -1
  13. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/llama_cpp.py +1 -0
  14. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/nim.py +2 -1
  15. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/ollama.py +1 -0
  16. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/openai.py +1 -0
  17. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/providers/openrouter.py +8 -3
  18. assistant-2.2.2b1/assistant/nlp/providers/vllm.py +38 -0
  19. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/shell.py +656 -129
  20. assistant-2.2.2b1/assistant/ptk_shell/toolbox.py +132 -0
  21. assistant-2.2.2b1/assistant/search_objects.py +230 -0
  22. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/setup_wizard.py +108 -15
  23. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/tools.py +259 -33
  24. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/traces/cmd.py +1 -0
  25. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/traces/store.py +127 -40
  26. {assistant-2.2.0b1 → assistant-2.2.2b1/assistant.egg-info}/PKG-INFO +58 -7
  27. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant.egg-info/SOURCES.txt +9 -0
  28. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant.egg-info/requires.txt +1 -0
  29. {assistant-2.2.0b1 → assistant-2.2.2b1}/setup.cfg +0 -1
  30. {assistant-2.2.0b1 → assistant-2.2.2b1}/setup.py +1 -0
  31. assistant-2.2.2b1/tests/test_builtin_cmds.py +398 -0
  32. assistant-2.2.2b1/tests/test_builtin_echo.py +559 -0
  33. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_compaction.py +114 -0
  34. assistant-2.2.2b1/tests/test_display.py +236 -0
  35. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_interrupt.py +17 -25
  36. assistant-2.2.2b1/tests/test_join_live.py +88 -0
  37. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_listen_interrupt.py +8 -8
  38. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_memory.py +20 -0
  39. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_model_compare.py +1 -1
  40. assistant-2.2.2b1/tests/test_reasoning_effort.py +280 -0
  41. assistant-2.2.2b1/tests/test_search_objects.py +139 -0
  42. assistant-2.2.2b1/tests/test_setup_wizard.py +68 -0
  43. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_startup_performance.py +8 -5
  44. assistant-2.2.2b1/tests/test_thinking_loop.py +128 -0
  45. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_tool_handling.py +47 -0
  46. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_tools.py +131 -8
  47. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_traces_agent.py +15 -7
  48. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_tts_streaming.py +203 -35
  49. {assistant-2.2.0b1 → assistant-2.2.2b1}/xontrib/assistant.py +4 -3
  50. assistant-2.2.0b1/assistant/nlp/providers/vllm.py +0 -22
  51. assistant-2.2.0b1/tests/test_builtin_cmds.py +0 -220
  52. assistant-2.2.0b1/tests/test_setup_wizard.py +0 -22
  53. {assistant-2.2.0b1 → assistant-2.2.2b1}/LICENSE +0 -0
  54. {assistant-2.2.0b1 → assistant-2.2.2b1}/MANIFEST.in +0 -0
  55. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/__main__.py +0 -0
  56. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/as_client.py +0 -0
  57. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/as_service.py +0 -0
  58. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/codecache.py +0 -0
  59. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/compress.py +0 -0
  60. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/conditions.py +0 -0
  61. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/entry_points/__init__.py +0 -0
  62. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/entry_points/run_assistant.py +0 -0
  63. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/envindex.py +0 -0
  64. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/execer.py +0 -0
  65. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/history.py +0 -0
  66. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/listen/__init__.py +0 -0
  67. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/listen/__main__.py +0 -0
  68. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/listen/main.py +0 -0
  69. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/__init__.py +0 -0
  70. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/clear.py +0 -0
  71. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/exit.py +0 -0
  72. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/http_client.py +0 -0
  73. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/interface.py +0 -0
  74. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/message.py +0 -0
  75. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/think.py +0 -0
  76. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/nlp/xonsh.py +0 -0
  77. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/parser.py +0 -0
  78. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/projectfile.py +0 -0
  79. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/__init__.py +0 -0
  80. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/bindings.py +0 -0
  81. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/completer.py +0 -0
  82. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/dialogs.py +0 -0
  83. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/floats.py +0 -0
  84. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/layout.py +0 -0
  85. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/operators.py +0 -0
  86. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/strformat_utils.py +0 -0
  87. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/styles.py +0 -0
  88. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/ptk_shell/windows.py +0 -0
  89. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/say/__init__.py +0 -0
  90. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/say/interface.py +0 -0
  91. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/time.py +0 -0
  92. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/tools_worker.py +0 -0
  93. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/traces/__init__.py +0 -0
  94. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/traces/export.py +0 -0
  95. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/traces/schema.py +0 -0
  96. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/utils.py +0 -0
  97. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/version.py +0 -0
  98. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant/vocabulary.py +0 -0
  99. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant.egg-info/dependency_links.txt +0 -0
  100. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant.egg-info/entry_points.txt +0 -0
  101. {assistant-2.2.0b1 → assistant-2.2.2b1}/assistant.egg-info/top_level.txt +0 -0
  102. {assistant-2.2.0b1 → assistant-2.2.2b1}/pyproject.toml +0 -0
  103. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_bindings.py +0 -0
  104. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_compress.py +0 -0
  105. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_config.py +0 -0
  106. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_envindex.py +0 -0
  107. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_envindex_suggestions.py +0 -0
  108. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_git_hook.py +0 -0
  109. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_history_bridge.py +0 -0
  110. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_http_client.py +0 -0
  111. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_init.py +0 -0
  112. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_interrupt_real.py +0 -0
  113. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_memory_sqlite.py +0 -0
  114. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_projectfile.py +0 -0
  115. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_tmux_interactive.py +0 -0
  116. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_traces_cmd.py +0 -0
  117. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_traces_export.py +0 -0
  118. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_traces_schema.py +0 -0
  119. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_traces_store.py +0 -0
  120. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_utils.py +0 -0
  121. {assistant-2.2.0b1 → assistant-2.2.2b1}/tests/test_xontrib.py +0 -0
  122. {assistant-2.2.0b1 → assistant-2.2.2b1}/xontrib/__init__.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: assistant
3
- Version: 2.2.0b1
3
+ Version: 2.2.2b1
4
4
  Summary: Your very own Assistant. Because you deserve it.
5
5
  Home-page: https://gitlab.com/waser-technologies/technologies/assistant
6
6
  Author: Danny Waser
@@ -29,6 +29,7 @@ Requires-Dist: xonsh
29
29
  Requires-Dist: prompt_toolkit>3.0.36
30
30
  Requires-Dist: questionary
31
31
  Requires-Dist: colorama
32
+ Requires-Dist: click
32
33
  Requires-Dist: rich
33
34
  Requires-Dist: halo
34
35
  Requires-Dist: requests
@@ -126,6 +127,8 @@ pacman -S python-assistant
126
127
 
127
128
  </details>
128
129
 
130
+ Read [Use Assistant as Your Default Shell](#use-assistant-as-your-default-shell) for more information.
131
+
129
132
  ### 2. Run the Configuration Wizard
130
133
 
131
134
  On the first launch the wizard starts automatically. You can also re-run it at any time:
@@ -153,7 +156,7 @@ The configuration is saved to `~/.assistant/config.toml`.
153
156
  assistant
154
157
  ```
155
158
 
156
- That's it — you're in. Type commands, ask questions, or mix both.
159
+ That's it: you're in. Type commands, ask questions, or mix both.
157
160
 
158
161
  ---
159
162
 
@@ -434,16 +437,64 @@ Toggle shortcuts (Ctrl+O, Ctrl+Q) update the config automatically when pressed,
434
437
 
435
438
  ## Use Assistant as Your Default Shell
436
439
 
437
- > **This is not recommended in beta!**
440
+ Assistant is a full Xonsh shell, so it reads the same RC files (`~/.config/xonsh/rc.xsh` or `~/.config/xonsh/rc.d/`) which means your existing config carries over as-is.
441
+
442
+ Read more about Xonsh [Installation Guide](https://xon.sh/install.html) and its [Run Control Files](https://xon.sh/xonshrc.html).
443
+
444
+ ### Install Assistant in a Stable, Isolated Environment
445
+
446
+ (so a system Python upgrade can't break your shell)
447
+
448
+ **Example with `uv`**
449
+
450
+ ```bash
451
+ env_root=~/.venvs
452
+ env_path=${env_root}/assistant
453
+ python_path=${env_path}/bin/python
454
+ mkdir -p $env_root
455
+ uv venv $env_path --python 3.13
456
+ uv pip install --python ${python_path} assistant
457
+ ```
458
+
459
+ From there to update Assistant, simply run:
460
+ ```bash
461
+ uv pip install --python ${python_path} -U assistant
462
+ ```
463
+
464
+ ### Create a Dedicated Terminal Profile
465
+
466
+ Since Assistant is a natural language shell and Xonsh is not ready to be fully POSIX compliant,
467
+ it is highly recommended that you create a dedicated profile for Assistant beside Xonsh and your other shells,
468
+ directly inside your terminal application,
469
+ where you can set Assistant as your default shell, instead of registering it for the system.
470
+
471
+ The path to Assistant is defined as shown in this example.
472
+ ```bash
473
+ assistant_bin=${env_path}/bin/assistant
474
+ ```
475
+
476
+ You can also create a symbolic link to Assistant in `~/.local/bin` (or any other location in `$PATH`) so that Assistant is easily accessible.
477
+
478
+ ```bash
479
+ ln -s $assistant_bin ~/.local/bin
480
+ ```
481
+
482
+ Or you could add `${env_path}/bin` to your `$PATH`.
438
483
 
439
- You should be able to add the location of `assistant` at the end of `/etc/shells`. You'll then be able to set `Assistant` as your default shell using `chsh`.
484
+ ### Register Assistant as a Shell for the System
485
+ > [!WARNING]
486
+ > If you read the Xonsh installation guide correctly, you understand that **you shouldn't try this method** unless:
487
+ > 1. you know what you are doing and want to test deeper interaction with the POSIX compliance of the system
488
+ > 2. you are ready to assume the consequences of potentially triggering multiple inference calls every time the system executes a command that exits with a non-zero code.
489
+ >
490
+ > Do the following at your own risk.
440
491
 
441
492
  ```bash
442
- sudo sh -c 'w=$(which assistant); echo $w >> /etc/shells'
443
- chsh -s $(which assistant)
493
+ sudo sh -c "echo ${assistant_bin} >> /etc/shells"
494
+ chsh -s $assistant_bin
444
495
  ```
445
496
 
446
- Log out and when you come back, `Assistant` should be your default shell.
497
+ Log out and back in, Assistant is now your default shell.
447
498
 
448
499
  ---
449
500
 
@@ -75,6 +75,8 @@ pacman -S python-assistant
75
75
 
76
76
  </details>
77
77
 
78
+ Read [Use Assistant as Your Default Shell](#use-assistant-as-your-default-shell) for more information.
79
+
78
80
  ### 2. Run the Configuration Wizard
79
81
 
80
82
  On the first launch the wizard starts automatically. You can also re-run it at any time:
@@ -102,7 +104,7 @@ The configuration is saved to `~/.assistant/config.toml`.
102
104
  assistant
103
105
  ```
104
106
 
105
- That's it — you're in. Type commands, ask questions, or mix both.
107
+ That's it: you're in. Type commands, ask questions, or mix both.
106
108
 
107
109
  ---
108
110
 
@@ -383,16 +385,64 @@ Toggle shortcuts (Ctrl+O, Ctrl+Q) update the config automatically when pressed,
383
385
 
384
386
  ## Use Assistant as Your Default Shell
385
387
 
386
- > **This is not recommended in beta!**
388
+ Assistant is a full Xonsh shell, so it reads the same RC files (`~/.config/xonsh/rc.xsh` or `~/.config/xonsh/rc.d/`) which means your existing config carries over as-is.
389
+
390
+ Read more about Xonsh [Installation Guide](https://xon.sh/install.html) and its [Run Control Files](https://xon.sh/xonshrc.html).
391
+
392
+ ### Install Assistant in a Stable, Isolated Environment
393
+
394
+ (so a system Python upgrade can't break your shell)
395
+
396
+ **Example with `uv`**
397
+
398
+ ```bash
399
+ env_root=~/.venvs
400
+ env_path=${env_root}/assistant
401
+ python_path=${env_path}/bin/python
402
+ mkdir -p $env_root
403
+ uv venv $env_path --python 3.13
404
+ uv pip install --python ${python_path} assistant
405
+ ```
406
+
407
+ From there to update Assistant, simply run:
408
+ ```bash
409
+ uv pip install --python ${python_path} -U assistant
410
+ ```
411
+
412
+ ### Create a Dedicated Terminal Profile
413
+
414
+ Since Assistant is a natural language shell and Xonsh is not ready to be fully POSIX compliant,
415
+ it is highly recommended that you create a dedicated profile for Assistant beside Xonsh and your other shells,
416
+ directly inside your terminal application,
417
+ where you can set Assistant as your default shell, instead of registering it for the system.
418
+
419
+ The path to Assistant is defined as shown in this example.
420
+ ```bash
421
+ assistant_bin=${env_path}/bin/assistant
422
+ ```
423
+
424
+ You can also create a symbolic link to Assistant in `~/.local/bin` (or any other location in `$PATH`) so that Assistant is easily accessible.
425
+
426
+ ```bash
427
+ ln -s $assistant_bin ~/.local/bin
428
+ ```
429
+
430
+ Or you could add `${env_path}/bin` to your `$PATH`.
387
431
 
388
- You should be able to add the location of `assistant` at the end of `/etc/shells`. You'll then be able to set `Assistant` as your default shell using `chsh`.
432
+ ### Register Assistant as a Shell for the System
433
+ > [!WARNING]
434
+ > If you read the Xonsh installation guide correctly, you understand that **you shouldn't try this method** unless:
435
+ > 1. you know what you are doing and want to test deeper interaction with the POSIX compliance of the system
436
+ > 2. you are ready to assume the consequences of potentially triggering multiple inference calls every time the system executes a command that exits with a non-zero code.
437
+ >
438
+ > Do the following at your own risk.
389
439
 
390
440
  ```bash
391
- sudo sh -c 'w=$(which assistant); echo $w >> /etc/shells'
392
- chsh -s $(which assistant)
441
+ sudo sh -c "echo ${assistant_bin} >> /etc/shells"
442
+ chsh -s $assistant_bin
393
443
  ```
394
444
 
395
- Log out and when you come back, `Assistant` should be your default shell.
445
+ Log out and back in, Assistant is now your default shell.
396
446
 
397
447
  ---
398
448
 
@@ -5,7 +5,7 @@ import warnings
5
5
  warnings.filterwarnings("ignore", category=DeprecationWarning, module="webrtcvad")
6
6
  warnings.filterwarnings("ignore", category=UserWarning, module="webrtcvad")
7
7
 
8
- __version__ = "2.2.0b1"
8
+ __version__ = "2.2.2b1"
9
9
 
10
10
  USERNAME = os.environ.get("USER", "root")
11
11
  HOME = os.environ.get("HOME", f"/home/{USERNAME}" if USERNAME != "root" else "/root")
@@ -62,6 +62,7 @@ DEFAULT_CONFIG = {
62
62
  "provider": "ollama",
63
63
  "context_length": 256000,
64
64
  "think": False,
65
+ "reasoning_effort": None,
65
66
  "temperature": 0.1,
66
67
  "max_token": 1,
67
68
  },
@@ -131,7 +132,9 @@ DEFAULT_CONFIG = {
131
132
  "max_turns": 50,
132
133
  "temperature": 0.1,
133
134
  "max_command_output_length": 10000,
135
+ "max_tool_output_lines": 50,
134
136
  "command_timeout": 180,
137
+ "tool_timeouts": {"xonsh": 180, "web_search": 15, "read_webpage": 20},
135
138
  },
136
139
  "memory": {
137
140
  "enabled": False,
@@ -161,6 +164,8 @@ DEFAULT_CONFIG = {
161
164
  # end of the system prompt, after the "Relevant commands" block.
162
165
  "suggestions_enabled": True,
163
166
  "suggestions_budget": 800,
167
+ "thinking_loop_enabled": True,
168
+ "thinking_loop_max_turns": 5,
164
169
  },
165
170
  "voice": {
166
171
  "tts_enabled": True,
@@ -11,7 +11,7 @@ import types
11
11
  from typing import Any, Callable, Dict, List, Optional
12
12
 
13
13
  from assistant.config import DEFAULT_CONFIG_PATH, get_config
14
- from assistant.main import get_random_thinking_sentence
14
+ from assistant.nlp.interface import get_random_thinking_sentence
15
15
  from assistant.memory import memory
16
16
  from assistant.nlp.providers import get_provider
17
17
  from assistant.nlp.xonsh import XONSH_INTRO
@@ -19,6 +19,19 @@ from assistant.tools import TOOL_REGISTRY
19
19
  from assistant.traces.store import store as trace_store
20
20
  from assistant.utils import count_message_tokens, count_messages_tokens
21
21
 
22
+ # Appended to a search tool's output so the model knows the structured result
23
+ # object landed in the session (see _process_tool_calls).
24
+ CONVERSATION_SEARCH_HINT = (
25
+ "\n\nResults assigned to __conversation__ in the session environment. "
26
+ "You can interact with them using the xonsh tool. Use `__conversation__?` "
27
+ "to get more information about them."
28
+ )
29
+ MEMORY_SEARCH_HINT = (
30
+ "\n\nResults assigned to __memory__ in the session environment. "
31
+ "You can interact with them using the xonsh tool. Use `__memory__?` "
32
+ "to get more information about them."
33
+ )
34
+
22
35
 
23
36
  # Handoff exception to break the tool loop and switch agents
24
37
  class AgentHandoff(Exception):
@@ -43,8 +56,12 @@ class Agent:
43
56
  model_cfg = cfg.get_model_config(model_id)
44
57
  self.model_name = model_cfg.get("id", "qwen3.5:latest")
45
58
  self.context_length = model_cfg.get("context_length", 256000)
59
+ # NOTE: config key "max_token" is shared between the completer (wants 1 = single word)
60
+ # and the main agent (wants a full output budget). The clamp below enforces a minimum
61
+ # of 4096 tokens for the agent, so the completer's max_token=1 is safely overridden.
62
+ self.max_tokens = max(4096, int(model_cfg.get("max_token", 4096)))
46
63
  self.think = model_cfg.get("think", False)
47
-
64
+ self.reasoning_effort = model_cfg.get("reasoning_effort", None)
48
65
  # Resolve Provider
49
66
  provider_id = model_cfg.get("provider", "ollama")
50
67
  provider_cfg = cfg.get_provider_config(provider_id)
@@ -144,6 +161,10 @@ class Orchestrator:
144
161
  self._summ_model_id = None # Resolved dedicated model id
145
162
  self._active_agent = None # Agent of the current chat turn
146
163
 
164
+ # Thinking-loop detection (Phase 8): consecutive thinking turns + one-shot nudge flag.
165
+ self._consecutive_thinking_turns = 0
166
+ self._thinking_warned = False
167
+
147
168
  self._ensure_config()
148
169
 
149
170
  # Session setup: trace recording (Phase 1). Controlled by config
@@ -254,7 +275,15 @@ class Orchestrator:
254
275
  try:
255
276
  from xonsh.built_ins import XSH
256
277
 
257
- vital = {k: v for k, v in XSH.ctx.items() if k.startswith("__")}
278
+ # Stale search results across a reset would mislead the next
279
+ # session — drop them (they start with __ and would otherwise
280
+ # survive the vital filter).
281
+ stale_search = ("__conversation__", "__memory__")
282
+ vital = {
283
+ k: v
284
+ for k, v in XSH.ctx.items()
285
+ if k.startswith("__") and k not in stale_search
286
+ }
258
287
  XSH.ctx.clear()
259
288
  XSH.ctx.update(vital)
260
289
  except Exception:
@@ -661,6 +690,10 @@ class Orchestrator:
661
690
  total_turns = 0
662
691
  max_turns = self._cfg.max_turns or 50
663
692
 
693
+ # Reset thinking-loop detection state for this chat() call.
694
+ self._consecutive_thinking_turns = 0
695
+ self._thinking_warned = False
696
+
664
697
  # Sync the static sys prompt once per session if possible
665
698
  # Cache for system prompt tokens to avoid re-calculating identical blocks
666
699
  while total_turns < max_turns:
@@ -767,13 +800,18 @@ class Orchestrator:
767
800
  None if self._is_intro else (agent.tools if agent.tools else None)
768
801
  )
769
802
  use_think = False if self._is_intro else agent.think
803
+ use_reasoning_effort = (
804
+ None if self._is_intro else agent.reasoning_effort
805
+ )
770
806
 
771
807
  kwargs = agent.provider.build_chat_kwargs(
772
808
  model=agent.model_name,
773
809
  messages=api_messages,
774
810
  temperature=self._cfg.temperature,
775
811
  think=use_think,
812
+ reasoning_effort=use_reasoning_effort,
776
813
  tools=use_tools,
814
+ max_tokens=agent.max_tokens,
777
815
  )
778
816
  if _is_debug:
779
817
  sys.stderr.write(f"[DEBUG] Calling API ({agent.model_name})...\n")
@@ -782,7 +820,7 @@ class Orchestrator:
782
820
  self._emit(
783
821
  {
784
822
  "type": "status",
785
- "text": f"{agent.name}: {get_random_thinking_sentence()()}...",
823
+ "text": f"{agent.name}: {get_random_thinking_sentence()}...",
786
824
  }
787
825
  )
788
826
  if _is_debug:
@@ -823,6 +861,10 @@ class Orchestrator:
823
861
  return ctx + [{"role": "assistant", "content": ""}]
824
862
  raise e
825
863
 
864
+ # Thinking-loop detection (Phase 8): nudge the model after N consecutive
865
+ # thinking turns so it stops reasoning and answers directly.
866
+ self._detect_thinking_loop(think_text, context_messages)
867
+
826
868
  assistant_msg = {"role": "assistant", "content": visible}
827
869
  # Count tokens after adding reasoning - reasoning is preserved for trace
828
870
  # fidelity (kept in the message), so it counts toward the token budget.
@@ -1137,6 +1179,42 @@ class Orchestrator:
1137
1179
 
1138
1180
  return visible, think_text, tool_calls, is_streamed, response_meta
1139
1181
 
1182
+ def _detect_thinking_loop(self, think_text, context_messages):
1183
+ """Detect a thinking loop and nudge the model to answer directly.
1184
+
1185
+ After ``thinking_loop_max_turns`` consecutive turns where the model
1186
+ produced only reasoning (no visible content, no tool calls), inject a
1187
+ one-shot user-role nudge into ``context_messages`` and emit a status
1188
+ event. The nudge fires once per ``chat()`` call (``_thinking_warned``);
1189
+ any turn with visible content or tool calls resets the counter.
1190
+
1191
+ Args:
1192
+ think_text: The reasoning text from this turn (empty if none).
1193
+ context_messages: The mutable message list to append the nudge to.
1194
+ """
1195
+ if think_text:
1196
+ self._consecutive_thinking_turns += 1
1197
+ loop_max = self._cfg.thinking_loop_max_turns
1198
+ if (
1199
+ self._cfg.thinking_loop_enabled
1200
+ and loop_max > 0
1201
+ and self._consecutive_thinking_turns >= loop_max
1202
+ and not self._thinking_warned
1203
+ ):
1204
+ self._thinking_warned = True
1205
+ context_messages.append(
1206
+ {"role": "user", "content": "Please stop reasoning and answer directly."}
1207
+ )
1208
+ self._emit(
1209
+ {
1210
+ "type": "status",
1211
+ "text": "⚠️ detected thinking loop — nudging model to answer directly",
1212
+ }
1213
+ )
1214
+ else:
1215
+ self._consecutive_thinking_turns = 0
1216
+ self._thinking_warned = False
1217
+
1140
1218
  def _get_summ_agent(self):
1141
1219
  """Resolve the LLM used by density-layer compaction (couches 1-3).
1142
1220
 
@@ -1189,7 +1267,10 @@ class Orchestrator:
1189
1267
  # Use config limit or a generous 64k default.
1190
1268
  # We no longer hard-cap at 32k to allow modern large-context models.
1191
1269
  effective_limit = int(agent.context_length or 64000)
1192
- context_limit = effective_limit * threshold
1270
+ # Reserve a minimum output budget so the model always has room to respond.
1271
+ # The input must never consume more than (context_length - output_reserve).
1272
+ output_reserve = max(4096, int(agent.max_tokens or 4096))
1273
+ context_limit = min(int(effective_limit * threshold), effective_limit - output_reserve)
1193
1274
 
1194
1275
  current_tokens = count_messages_tokens(run_messages)
1195
1276
  if current_tokens > context_limit:
@@ -1212,6 +1293,14 @@ class Orchestrator:
1212
1293
  run_messages.clear()
1213
1294
  run_messages.extend(system_msgs)
1214
1295
  run_messages.extend([m for m in context_messages if m["role"] != "system"])
1296
+ # Hard safety net: if compaction still left us over the limit,
1297
+ # drop oldest non-system messages until we fit.
1298
+ while count_messages_tokens(run_messages) > context_limit:
1299
+ non_system = [i for i, m in enumerate(run_messages) if m.get("role") != "system"]
1300
+ if not non_system:
1301
+ break
1302
+ oldest_idx = non_system[0]
1303
+ run_messages.pop(oldest_idx)
1215
1304
  # Couche 3 rolling summary: injected as a trailing system block (the
1216
1305
  # lead system prompt is never mutated, keeping the KV-cache warm). It
1217
1306
  # lives outside the compaction branch so a summary produced earlier is
@@ -1331,35 +1420,25 @@ class Orchestrator:
1331
1420
  if "sigint_pipe_w" in _params:
1332
1421
  extra_kwargs["sigint_pipe_w"] = sigint_pipe_w
1333
1422
 
1423
+ _t0 = time.monotonic()
1334
1424
  res = fn(args, **extra_kwargs)
1425
+ _wall = time.monotonic() - _t0
1335
1426
  except KeyboardInterrupt as ke:
1336
1427
  # Re-raise KeyboardInterrupt so parent loop can handle cleanup
1337
1428
  raise ke
1338
1429
 
1339
1430
  # Check for abort after tool execution
1340
1431
  if self._abort_event.is_set():
1341
- self._abort_event.clear()
1342
1432
  return None
1343
1433
 
1344
- self._emit(
1345
- {
1346
- "type": "tool_output",
1347
- "command": cmd_display,
1348
- "output": res["content"],
1349
- "is_error": res["is_error"],
1350
- }
1351
- )
1352
-
1353
- tool_msg = {
1354
- "role": "tool",
1355
- "tool_call_id": t.id,
1356
- "name": fn_name,
1357
- "content": res["content"] or "",
1358
- }
1359
- # Soft bound (rtk-like): if the output exceeds the context budget,
1360
- # compress it for the model context (deterministic, command-aware).
1361
- # The full output is kept on _full_output for trace storage.
1362
- content = tool_msg["content"]
1434
+ # Structured display payload (e.g. search_conversation rows)
1435
+ # rides the event for the shell; it never enters tool_msg, so
1436
+ # the LLM context stays the flat ``content`` string.
1437
+ # Compute the output ONCE, shared by the model context and the
1438
+ # user display (issue #10: they must always be the same).
1439
+ content = res["content"] or ""
1440
+ display_output = content
1441
+ compressed = False
1363
1442
  if content:
1364
1443
  bound = getattr(self._cfg, "max_tool_output_in_context", None)
1365
1444
  if not bound:
@@ -1372,13 +1451,52 @@ class Orchestrator:
1372
1451
  if bound and len(content) > bound:
1373
1452
  from assistant.compress import compress_tool_output
1374
1453
 
1375
- tool_msg["_full_output"] = content
1376
- tool_msg["content"] = compress_tool_output(
1454
+ display_output = compress_tool_output(
1377
1455
  content,
1378
1456
  command=cmd_display if fn_name == "xonsh" else fn_name,
1379
1457
  limit=bound,
1380
1458
  )
1381
- tool_msg["_compressed"] = True
1459
+ compressed = True
1460
+
1461
+ # Search tools inject a structured result object into the
1462
+ # session (__conversation__ / __memory__); surface that to the
1463
+ # model so it can refine/display them via the xonsh tool. The
1464
+ # hint rides both the display output and (when compressed) the
1465
+ # restored full output.
1466
+ result_kind = res.get("result_kind")
1467
+ if result_kind == "conversation":
1468
+ hint = CONVERSATION_SEARCH_HINT
1469
+ elif result_kind == "memory":
1470
+ hint = MEMORY_SEARCH_HINT
1471
+ else:
1472
+ hint = ""
1473
+ if hint:
1474
+ display_output += hint
1475
+
1476
+ event = {
1477
+ "type": "tool_output",
1478
+ "command": cmd_display,
1479
+ "tool": fn_name,
1480
+ "output": display_output,
1481
+ "is_error": res["is_error"],
1482
+ "wall": round(_wall, 3),
1483
+ "artifact": len(display_output or ""),
1484
+ }
1485
+ if res.get("results") is not None:
1486
+ event["results"] = res["results"]
1487
+ if res.get("result_kind"):
1488
+ event["result_kind"] = res["result_kind"]
1489
+ self._emit(event)
1490
+
1491
+ tool_msg = {
1492
+ "role": "tool",
1493
+ "tool_call_id": t.id,
1494
+ "name": fn_name,
1495
+ "content": display_output,
1496
+ }
1497
+ if compressed:
1498
+ tool_msg["_full_output"] = content + hint
1499
+ tool_msg["_compressed"] = True
1382
1500
  # Optimization: populate cache immediately
1383
1501
  tool_msg["_tokens"] = count_message_tokens(tool_msg)
1384
1502
  context_messages.append(tool_msg)
@@ -1394,12 +1512,25 @@ class Orchestrator:
1394
1512
  return handoff_target
1395
1513
 
1396
1514
  def _get_tool_display_name(self, name, args):
1515
+ """Return the user-facing display text for a tool call.
1516
+
1517
+ Always shows the tool's primary argument value (when given) so the
1518
+ input box never repeats the ``tool_name(...)`` wrapper the title
1519
+ already provides. Falls back to the bare tool name when no argument
1520
+ is meaningful.
1521
+ """
1397
1522
  if name == "xonsh":
1398
1523
  return args.get("code", "")
1399
1524
  if name == "web_search":
1400
- return f'web_search("{args.get("query", "")}")'
1525
+ return args.get("query", "") or name
1401
1526
  if name == "memory_search":
1402
- return f'memory_search("{args.get("query", "")}")'
1527
+ return args.get("query", "") or name
1528
+ if name == "memory_add":
1529
+ return args.get("text", "") or name
1530
+ if name == "cmd_search":
1531
+ return args.get("query", "") or name
1532
+ if name == "search_conversation":
1533
+ return args.get("query", "") or name
1403
1534
  return name
1404
1535
 
1405
1536
  def _handle_handoff_return(self, agent_name, response, context_messages, agent_id):