hexcli 2.14.0__tar.gz → 2.15.0__tar.gz

{hexcli-2.14.0 → hexcli-2.15.0}/CHANGELOG.md

@@ -4,6 +4,36 @@ Full evidence for every claim below — including the experiments that failed
 lives in `docs/V2_PLAN.md` §14. Numbers are pass^k over repeated live runs on
 the Hexagon NPU, not single-run anecdotes.
 
+## 2.15.0 — 2026-09-18
+
+A minor release: one tool error the model reads is new (`edit_file` on
+identical strings).
+
+> **Gate: PASS** on the pinned 24-case set after one recheck (`agentic-5`
+> 4/5, a counting answer with no edit in it, then 6/6). Own arm, fresh
+> server, seed 20260917, 9 invalid of 245 runs. Against the v2.11.1 12-run
+> baseline: run-level 402/514 vs 172/221, −0.4 %, Fisher p = 0.92; pass^k
+> 26 → 28 of 46 shared cases, McNemar p = 0.63. The error fired in the arm
+> on 3 `tests-claim-1` runs and nowhere else; that case went 4/5 against
+> 12/24 (p = 0.34, ceiling). `self-correct-1` read 2/5 in the arm on a
+> server 75 minutes into its traffic (the arm's invalid runs were climbing),
+> and the error had not fired in any of its runs; a paired control on a
+> fresh server, ten runs each on the parent tree and this one, came back
+> **10/10 vs 9/10, p = 1.0**. In the one control run where the error did
+> fire the model had sent a hallucinated `main()` as both strings, was told
+> nothing changed, read the file, and then made the real fix. Smoke miss
+> on `agentic-2` re-measured 9/10 against its recorded 75 %.
+
+- `edit_file` refuses an edit whose `old_string` and `new_string` are the
+  same text, instead of answering "Edited". 37 of the 840 `edit_file` calls
+  on record sent identical strings and were told the edit landed; the model
+  then ran or read the file believing it had changed it (`tests-claim-1`:
+  "The median function was correctly fixed", after an edit of `mean` to
+  itself; `runit-1` before 2.14.0: the broken line edited to itself, twice,
+  to the step limit). The tool now says nothing changed and what to send
+  instead, and leaves the file byte-identical. A missing file still reports
+  the path first.
+
 ## 2.14.0 — 2026-09-17
 
 A minor release: one tool behaves differently for the model (a `write_file`

{hexcli-2.14.0 → hexcli-2.15.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: hexcli
-Version: 2.14.0
+Version: 2.15.0
 Summary: Local Hexagon NPU terminal agent for Snapdragon X Elite Windows ARM64
 Project-URL: Homepage, https://github.com/NathanL15/Hex-CLI
 Project-URL: Repository, https://github.com/NathanL15/Hex-CLI

{hexcli-2.14.0 → hexcli-2.15.0}/hexcli/__init__.py

@@ -3,4 +3,4 @@
 # The one place the version is written. pyproject.toml reads it (hatch
 # dynamic version), agent.VERSION re-exports it, and CI refuses a release
 # tag that does not match it.
-__version__ = "2.14.0"
+__version__ = "2.15.0"

{hexcli-2.14.0 → hexcli-2.15.0}/hexcli/tools.py

@@ -423,6 +423,17 @@ def edit_file_tool(path_text: str, old_string: str, new_string: str) -> str:
         raise RuntimeError("edit_file requires a non-empty 'old_string'. Use write_file to overwrite the whole file.")
     if not path.exists():
         raise RuntimeError(f"File not found: {path}.{missing_path_hint(path)}")
+    if old_string == new_string:
+        # 37 of the 840 edit_file calls on record (2026-09-17) sent the same
+        # text as old and new and were told "Edited"; the model then ran or
+        # read the file believing it had changed it (tests-claim-1: "The
+        # median function was correctly fixed", after an edit of mean to
+        # itself). A no-op is not an edit, and saying so is the only way the
+        # next step can be the change that was meant.
+        raise RuntimeError(
+            f"old_string and new_string are identical, so nothing changed in {path}. "
+            "Put the changed text in new_string, or read_file to see what is there now."
+        )
     content = path.read_text(encoding="utf-8")
     tier = "exact"
     if content.count(old_string) == 1: