Skip to content

The agent that grows with you

fix(skills): clip index descriptions at any whitespace, not only spaces

Tekniumb66ab2bb7b47
diff --git a/agent/skill_utils.py b/agent/skill_utils.py
index c12ed03c174..f24faec3d4e 100644
--- a/agent/skill_utils.py
+++ b/agent/skill_utils.py
@@ -840,6 +840,8 @@ def resolve_skill_config_values(config_vars: list[dict[str, Any]]) -> dict[str,
# description over it and tests/skills/test_authoring_standards.py holds shipped skills to it, so only legacy or
# hand-edited skills ever reach clip_skill_description().
SKILL_PROMPT_DESC_LIMIT = 60
+_LAST_WHITESPACE_RE = re.compile(r"\s\S*$")
+_CLIP_TRAIL_RE = re.compile(r"[\s.,;:\-\u2013\u2014(/]+$")
def _normalize_skill_description(frontmatter: dict[str, Any]) -> str:
@@ -852,11 +854,11 @@ def clip_skill_description(desc: str) -> str:
"""*desc* fitted to SKILL_PROMPT_DESC_LIMIT, cut at a word boundary so the index never shows a half word."""
if len(desc) <= SKILL_PROMPT_DESC_LIMIT:
return desc
- head = desc[:SKILL_PROMPT_DESC_LIMIT - 2] # one char past the budget: a space there means the last word is whole
- cut = head.rfind(" ")
- if cut <= 0: # a single word longer than the budget has no boundary to cut at
+ head = desc[:SKILL_PROMPT_DESC_LIMIT - 2] # one char past the budget: whitespace there means the last word is whole
+ boundary = _LAST_WHITESPACE_RE.search(head) # any whitespace (tab, NBSP, ...) separates words, not just " "
+ if not boundary or boundary.start() == 0: # a single word longer than the budget has no boundary to cut at
return desc[:SKILL_PROMPT_DESC_LIMIT - 3] + "..."
- return head[:cut].rstrip(" .,;:-\u2013\u2014(/") + "..."
+ return _CLIP_TRAIL_RE.sub("", head[:boundary.start()]) + "..."
def extract_skill_description(frontmatter: dict[str, Any]) -> str:
diff --git a/tests/agent/test_skills_index_format.py b/tests/agent/test_skills_index_format.py
index 9e35e8d5014..3d0748ec579 100644
--- a/tests/agent/test_skills_index_format.py
+++ b/tests/agent/test_skills_index_format.py
@@ -4,7 +4,7 @@ import re
import pytest
from agent.prompt_builder import build_skills_system_prompt, clear_skills_system_prompt_cache
-from agent.skill_utils import SKILL_PROMPT_DESC_LIMIT
+from agent.skill_utils import SKILL_PROMPT_DESC_LIMIT, clip_skill_description
from agent.skills_index_delta import index_entries
LONG = "Use when cutting, publishing, or abandoning a hermes-agent release and its notes."
@@ -47,3 +47,11 @@ def test_index_rows_carry_each_name_once_and_never_a_half_word(monkeypatch, tmp_
assert LONG.startswith(kept) and LONG[len(kept)] in " ," # cut lands between words
# The turn-delta parser reads flattened rows like nested ones.
assert set(index_entries(prompt)) == {"ascii-art", "pixel-art", "computer-use", "hermes-release"}
+
+
[email protected]("sep", ["\t", "\u00a0"], ids=["tab", "nbsp"])
+def test_clip_cuts_at_any_whitespace_never_inside_a_word(sep):
+ desc = f"alpha{sep}" + f"bravo{sep}" * 12
+ shown = clip_skill_description(desc)
+ assert len(shown) <= SKILL_PROMPT_DESC_LIMIT and shown.endswith("...")
+ assert set(shown[:-3].split(sep)) == {"alpha", "bravo"} # whole words only, no "bra..."