autoforge-agent 0.4.1__tar.gz → 0.4.3__tar.gz

autoforge_agent-0.4.3/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Yangyi Zhang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

autoforge_agent-0.4.3/MANIFEST.in

@@ -0,0 +1,10 @@
+# The sdist ships the test suite (setuptools picks up tests/test*.py on its own),
+# so it has to ship everything that suite imports or reads:
+#   - conftest.py and the shared fakes the terminal and MCP tests drive
+#   - probes/: the captured generator replies test_json_repair reads by path
+# Without this the tarball carried 39 test modules that could not be collected,
+# plus two that failed on a missing fixture. Found by unpacking the release and
+# running it -- which is now a job in ci.yml, and was briefly a job in
+# tool-market's reproduce workflow before it moved to the repo that can fix it.
+recursive-include tests *.py
+recursive-include probes *.txt *.md *.py

{autoforge_agent-0.4.1/autoforge_agent.egg-info → autoforge_agent-0.4.3}/PKG-INFO

@@ -1,16 +1,19 @@
 Metadata-Version: 2.4
 Name: autoforge-agent
-Version: 0.4.1
+Version: 0.4.3
 Summary: Agent framework with verifiable self-made tools (the forge pipeline).
 Author: Zhang Yangyi
 License: MIT
 Keywords: agents,llm,tools,self-improvement,verification
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
+License-File: LICENSE
 Requires-Dist: requests>=2.28
 Provides-Extra: dev
 Requires-Dist: pytest>=7; extra == "dev"
 Requires-Dist: pytest-xdist; extra == "dev"
+Requires-Dist: pyyaml; extra == "dev"
+Dynamic: license-file
 
 # autoforge
 
@@ -97,7 +100,7 @@ auto config
   base_url   https://aiping.cn/api/v1  config C:\Users\you\.autoforge\config.json
   model      DeepSeek-V4.1-Flash       config C:\Users\you\.autoforge\config.json
   api_key    QC-5...e262 (len 68)      config C:\Users\you\.autoforge\config.json
-  max_tokens 3000                      config C:\Users\you\.autoforge\config.json
+  max_tokens 32768                     default
   proxy      True                      config C:\Users\you\.autoforge\config.json
 ```
 
@@ -112,6 +115,19 @@ auto --policy supervised                    # keep the harness, drop the latitud
 AUTOFORGE_POLICY=supervised auto            # same, this shell only
 ```
 
+`max_tokens` is a budget rather than a limit on the answer: the generator's model
+reasons before it writes, and the trace is spent out of the same budget. At the old
+3000 the trace took the whole of it and the envelope never began, so the default is
+32768. A provider with a lower ceiling is not a problem -- a 400 that names
+`max_tokens` is answered at the number it names. Two things to know before tuning it
+down: an envelope cut off mid-JSON fails to parse, and a truncated one that *does*
+parse arrives with empty `probes`, which the pipeline accepts as a tool that has been
+verified. If a gateway answers `503 暂无可用服务商` for a large cap on a reasoning
+model no matter how long you wait, that is its routing rather than the number you
+sent: the combination measured to work through the aiping gateway is
+`--model Qwen3.5-Flash` with `AUTOFORGE_MAX_TOKENS=4000`, while a direct provider
+endpoint takes the full 32768. Numbers and controls in `probes/FINDINGS.md`.
+
 Why bother with a file when environment variables exist: a variable exported
 *after* a terminal was opened is invisible to that terminal — Windows and POSIX
 alike inherit the environment at process start. That turns configuration into
@@ -168,6 +184,35 @@ with `/` as text. Paste a block and it collapses to a one-line
 `[Pasted text #1: 40 lines → …]` placeholder — the text is kept on disk and
 expanded again before it reaches the model.
 
+### Selecting part of a line
+
+```
+you> check the empty file case
+        └──────┘  Shift+arrows select, Ctrl+Insert copies, Ctrl+Delete cuts
+```
+
+`Shift+Left`/`Shift+Right` extend a selection from where the cursor is,
+`Shift+Home`/`Shift+End` take it to either end, and the span is drawn in
+reverse video. `Ctrl+Insert` copies it, `Ctrl+Delete` cuts it, `Shift+Insert`
+and `^V`/`^Y` paste, and `Backspace`, `Delete`, `^K`, `^U` and `^W` remove the
+selection when there is one. Typing over a selection replaces it. `^C` is left
+alone: it still stops the run, because a terminal that can copy but cannot be
+interrupted cannot be left.
+
+**The mouse stays the console's.** The editor can take it — `AUTOFORGE_MOUSE=1`
+— and keep the selection itself, which is the only way it ever hears about a
+drag on Windows: the console paints its own highlight, keeps it, and reports
+nothing, so a drag over the middle of the line followed by `Delete` deletes the
+last character instead. But taking the mouse takes it for the whole session:
+quick-edit off means the console stops selecting *and stops scrolling on the
+wheel*, so reading back through what the agent printed stops working. That is a
+worse trade than the bug, and it was reported as such, so the default is to
+leave the mouse alone and select on the keys instead.
+
+The keys an X terminal has always had keep working: `Ctrl+A`/`Ctrl+E` for the
+ends of the line, `Ctrl+W` for the previous word, `Ctrl+U`/`Ctrl+K` to kill to
+either end.
+
 The behaviour degrades honestly: a pipe, a redirect, or a test gets the plain
 cooked-mode reader and no heartbeat, because there is no terminal to own.
 
@@ -310,9 +355,15 @@ context but `force=True` still runs it. `rehab()` puts it back on trial.
 
 ```
 score = w_text·similarity + w_success·success_rate + w_trust·state_trust
-        − w_cost·cost_penalty − w_over·over_trigger_penalty
 ```
 
+`cost` and `over_trigger` were terms here until 2026-09-28, when the library
+was measured and both scored a constant for all 120 tools — `cost_hint` is
+"cheap" everywhere and nothing re-probes a forged tool against its
+`TriggerProbe`, so `trigger_misses` stays 0. They were removed rather than
+weighted at zero: a tunable weight on a constant is a knob that can only make
+retrieval worse, and `amend_self` would have had no signal to notice it by.
+
 **Every freedom is declared, and each declaration is classified.** Two presets
 ship — `full` (the default: nothing denied) and `supervised` — selected with
 `--policy` or `AUTOFORGE_POLICY`, and `autoforge config` prints which is live.
@@ -475,3 +526,16 @@ similarity rather than embeddings — swap `text_similarity` for a vector index
 when the library is big enough to need it. `RoleType.FORGE` is declarative only:
 a topology names a forge role, it does not yet change which tools that child can
 reach.
+
+## Related work by the same author
+
+The same claim -- *a number is meaningless until it is shown to survive its own
+verification* -- is made and measured in other domains:
+
+| repo | the number it refuses to trust |
+| ---- | ---- |
+| [autoforge](https://github.com/ZhangYangyi03/autoforge) | a tool's fitness, until an oracle outside the tool agrees |
+| [agentic-eda](https://github.com/ZhangYangyi03/agentic-eda) | a circuit's area, until equivalence to the reference netlist is proven |
+| [debt-verify](https://github.com/ZhangYangyi03/debt-verify) | a debt clause decision, until it survives the published revision record |
+| [tool-market](https://github.com/ZhangYangyi03/tool-market) | a tool's liveness, until the hash chain says which revision is live |
+| [agent-safety-bench](https://github.com/ZhangYangyi03/agent-safety-bench) | a model's safety compliance, measured rather than assumed |

autoforge_agent-0.4.1/PKG-INFO → autoforge_agent-0.4.3/README.md

@@ -1,17 +1,3 @@
-Metadata-Version: 2.4
-Name: autoforge-agent
-Version: 0.4.1
-Summary: Agent framework with verifiable self-made tools (the forge pipeline).
-Author: Zhang Yangyi
-License: MIT
-Keywords: agents,llm,tools,self-improvement,verification
-Requires-Python: >=3.10
-Description-Content-Type: text/markdown
-Requires-Dist: requests>=2.28
-Provides-Extra: dev
-Requires-Dist: pytest>=7; extra == "dev"
-Requires-Dist: pytest-xdist; extra == "dev"
-
 # autoforge
 
 An agent framework that **builds its own tools** — and makes them earn their keep.
@@ -97,7 +83,7 @@ auto config
   base_url   https://aiping.cn/api/v1  config C:\Users\you\.autoforge\config.json
   model      DeepSeek-V4.1-Flash       config C:\Users\you\.autoforge\config.json
   api_key    QC-5...e262 (len 68)      config C:\Users\you\.autoforge\config.json
-  max_tokens 3000                      config C:\Users\you\.autoforge\config.json
+  max_tokens 32768                     default
   proxy      True                      config C:\Users\you\.autoforge\config.json
 ```
 
@@ -112,6 +98,19 @@ auto --policy supervised                    # keep the harness, drop the latitud
 AUTOFORGE_POLICY=supervised auto            # same, this shell only
 ```
 
+`max_tokens` is a budget rather than a limit on the answer: the generator's model
+reasons before it writes, and the trace is spent out of the same budget. At the old
+3000 the trace took the whole of it and the envelope never began, so the default is
+32768. A provider with a lower ceiling is not a problem -- a 400 that names
+`max_tokens` is answered at the number it names. Two things to know before tuning it
+down: an envelope cut off mid-JSON fails to parse, and a truncated one that *does*
+parse arrives with empty `probes`, which the pipeline accepts as a tool that has been
+verified. If a gateway answers `503 暂无可用服务商` for a large cap on a reasoning
+model no matter how long you wait, that is its routing rather than the number you
+sent: the combination measured to work through the aiping gateway is
+`--model Qwen3.5-Flash` with `AUTOFORGE_MAX_TOKENS=4000`, while a direct provider
+endpoint takes the full 32768. Numbers and controls in `probes/FINDINGS.md`.
+
 Why bother with a file when environment variables exist: a variable exported
 *after* a terminal was opened is invisible to that terminal — Windows and POSIX
 alike inherit the environment at process start. That turns configuration into
@@ -168,6 +167,35 @@ with `/` as text. Paste a block and it collapses to a one-line
 `[Pasted text #1: 40 lines → …]` placeholder — the text is kept on disk and
 expanded again before it reaches the model.
 
+### Selecting part of a line
+
+```
+you> check the empty file case
+        └──────┘  Shift+arrows select, Ctrl+Insert copies, Ctrl+Delete cuts
+```
+
+`Shift+Left`/`Shift+Right` extend a selection from where the cursor is,
+`Shift+Home`/`Shift+End` take it to either end, and the span is drawn in
+reverse video. `Ctrl+Insert` copies it, `Ctrl+Delete` cuts it, `Shift+Insert`
+and `^V`/`^Y` paste, and `Backspace`, `Delete`, `^K`, `^U` and `^W` remove the
+selection when there is one. Typing over a selection replaces it. `^C` is left
+alone: it still stops the run, because a terminal that can copy but cannot be
+interrupted cannot be left.
+
+**The mouse stays the console's.** The editor can take it — `AUTOFORGE_MOUSE=1`
+— and keep the selection itself, which is the only way it ever hears about a
+drag on Windows: the console paints its own highlight, keeps it, and reports
+nothing, so a drag over the middle of the line followed by `Delete` deletes the
+last character instead. But taking the mouse takes it for the whole session:
+quick-edit off means the console stops selecting *and stops scrolling on the
+wheel*, so reading back through what the agent printed stops working. That is a
+worse trade than the bug, and it was reported as such, so the default is to
+leave the mouse alone and select on the keys instead.
+
+The keys an X terminal has always had keep working: `Ctrl+A`/`Ctrl+E` for the
+ends of the line, `Ctrl+W` for the previous word, `Ctrl+U`/`Ctrl+K` to kill to
+either end.
+
 The behaviour degrades honestly: a pipe, a redirect, or a test gets the plain
 cooked-mode reader and no heartbeat, because there is no terminal to own.
 
@@ -310,9 +338,15 @@ context but `force=True` still runs it. `rehab()` puts it back on trial.
 
 ```
 score = w_text·similarity + w_success·success_rate + w_trust·state_trust
-        − w_cost·cost_penalty − w_over·over_trigger_penalty
 ```
 
+`cost` and `over_trigger` were terms here until 2026-09-28, when the library
+was measured and both scored a constant for all 120 tools — `cost_hint` is
+"cheap" everywhere and nothing re-probes a forged tool against its
+`TriggerProbe`, so `trigger_misses` stays 0. They were removed rather than
+weighted at zero: a tunable weight on a constant is a knob that can only make
+retrieval worse, and `amend_self` would have had no signal to notice it by.
+
 **Every freedom is declared, and each declaration is classified.** Two presets
 ship — `full` (the default: nothing denied) and `supervised` — selected with
 `--policy` or `AUTOFORGE_POLICY`, and `autoforge config` prints which is live.
@@ -475,3 +509,16 @@ similarity rather than embeddings — swap `text_similarity` for a vector index
 when the library is big enough to need it. `RoleType.FORGE` is declarative only:
 a topology names a forge role, it does not yet change which tools that child can
 reach.
+
+## Related work by the same author
+
+The same claim -- *a number is meaningless until it is shown to survive its own
+verification* -- is made and measured in other domains:
+
+| repo | the number it refuses to trust |
+| ---- | ---- |
+| [autoforge](https://github.com/ZhangYangyi03/autoforge) | a tool's fitness, until an oracle outside the tool agrees |
+| [agentic-eda](https://github.com/ZhangYangyi03/agentic-eda) | a circuit's area, until equivalence to the reference netlist is proven |
+| [debt-verify](https://github.com/ZhangYangyi03/debt-verify) | a debt clause decision, until it survives the published revision record |
+| [tool-market](https://github.com/ZhangYangyi03/tool-market) | a tool's liveness, until the hash chain says which revision is live |
+| [agent-safety-bench](https://github.com/ZhangYangyi03/agent-safety-bench) | a model's safety compliance, measured rather than assumed |

{autoforge_agent-0.4.1 → autoforge_agent-0.4.3}/autoforge/__init__.py

@@ -15,6 +15,75 @@ Design pillars
 Freedom is the default; verification lets you *trust* what the freedom produced.
 """
 
+import os
+
+# No console window for any child this *process* starts.
+#
+# The agent is started by a scheduled task through pythonw, which owns no
+# console. Windows gives a console-subsystem child of a console-less parent a
+# *fresh* console -- a visible window on the operator's desktop -- so every
+# subprocess that starts without a flag flashes one. Measured on 2026-09-16 by
+# enumerating visible top-level windows by pid: flags=0 yields one visible
+# PseudoConsoleWindow, DETACHED_PROCESS none. `GetConsoleWindow()` is the wrong
+# probe for this: it reports non-zero even under DETACHED_PROCESS, where no
+# window is shown. The flag is CREATE_NO_WINDOW, not DETACHED_PROCESS: see the
+# long note at `_NO_WINDOW` below for why the difference is load-bearing.
+#
+# Done here, at import, rather than at each call site: the spawns live in
+# webtools (the search worker), market, schedule, cpu/safety and mcp, and a rule
+# that has to be remembered at five places is a rule that will be missing from
+# the sixth. Nothing in this package does `from subprocess import Popen`, so the
+# class is always looked up on the module and one assignment covers all of them.
+#
+# AUTOFORGE_KEEP_CONSOLE=1 opts out, for the case where a visible window is what
+# is wanted -- debugging a child that dies before it can log.
+if os.name == "nt" and not os.environ.get("AUTOFORGE_KEEP_CONSOLE"):
+    import subprocess as _subprocess
+
+    #: CREATE_NO_WINDOW: the child gets a console, but it is hidden -- and every
+    #: console-subsystem process it starts inherits that hidden console instead
+    #: of being handed a fresh one of its own.
+    #:
+    #: The previous value was DETACHED_PROCESS (0x00000008), chosen because it
+    #: was measured to produce no visible window for the *direct* child. It is
+    #: the wrong flag, for two reasons that only show up one level deeper or one
+    #: pipe further:
+    #:
+    #:   * DETACHED_PROCESS creates no console at all, so the child has none to
+    #:     pass down. Every console-subsystem process *it* starts is therefore a
+    #:     console-less parent's child -- exactly the condition this block exists
+    #:     to prevent -- and Windows gives it a fresh, visible console. One level
+    #:     of detachment only moves the flash from the child to the grandchild.
+    #:     `shell=True` always adds that level: cmd.exe is the child, the real
+    #:     command is the flashed grandchild.
+    #:   * a detached child has no console for `wsl.exe` to write to, and it
+    #:     then exits 0 with empty stdout and empty stderr on a captured pipe --
+    #:     measured on this host. That silent-nothing is what `wsl_isolation`
+    #:     used to read as "the distro is not answering".
+    #:
+    #: CREATE_NO_WINDOW has neither problem. Verified on 2026-09-16 by
+    #: enumerating visible top-level windows by pid under the default Windows
+    #: Terminal host: DETACHED_PROCESS -> one CASCADIA_HOSTING_WINDOW_CLASS
+    #: window, which also took focus; CREATE_NO_WINDOW -> none, with the
+    #: command's output intact. The wsl carve-out that used to live here is gone
+    #: with it: CREATE_NO_WINDOW needs no per-executable exception.
+    _NO_WINDOW = 0x08000000        # CREATE_NO_WINDOW
+
+    class _QuietPopen(_subprocess.Popen):
+        """subprocess.Popen, but the child never gets a console window."""
+
+        def __init__(self, *args, **kwargs):
+            flags = int(kwargs.get("creationflags") or 0)
+            # OR'd, not assigned: a caller that asks for CREATE_NEW_PROCESS_GROUP
+            # or a detached browser keeps its flag.
+            kwargs["creationflags"] = flags | _NO_WINDOW
+            super().__init__(*args, **kwargs)
+
+    if _subprocess.Popen.__name__ != "_QuietPopen":
+        _subprocess.Popen = _QuietPopen
+        globals()["_QuietPopen"] = _QuietPopen
+
+
 # One source of truth: pyproject.toml. Hardcoding the version here meant the
 # package reported 0.2.0 while pyproject said 0.4.0 -- a number nobody would
 # notice was wrong until they quoted it. Read it from the installed metadata