fix(background-review): reject unresolved failures as skills

agent/background_review.py

@@ -284,6 +284,15 @@ _SKILL_REVIEW_PROMPT = (
     "  • One-off task narratives. A user asking 'summarize today's "
     "market' or 'analyze this PR' is not a class of work that warrants "
     "a skill.\n\n"
+    "  • Unresolved failures: if the session ended WITHOUT actually "
+    "finding a working method — you tried several things, none worked, "
+    "and told the user to check manually — do NOT write those attempts "
+    "up as a 'reliable workflow' or 'recommended approach'. That presents "
+    "an untested sequence of failures as validated guidance a future "
+    "session will trust and repeat. Either say 'Nothing to save', or, "
+    "only if you are independently confident of a real working alternative "
+    "(not something you are merely guessing might work), capture ONLY that "
+    "alternative — never the dead ends, and never dressed up as best practice.\n\n"
     "If a tool failed because of setup state, capture the FIX (install "
     "command, config step, env var to set) under an existing setup or "
     "troubleshooting skill — never 'this tool does not work' as a "
@@ -377,6 +386,15 @@ _COMBINED_REVIEW_PROMPT = (
     "  • One-off task narratives. A user asking 'summarize today's "
     "market' or 'analyze this PR' is not a class of work that warrants "
     "a skill.\n\n"
+    "  • Unresolved failures: if the session ended WITHOUT actually "
+    "finding a working method — you tried several things, none worked, "
+    "and told the user to check manually — do NOT write those attempts "
+    "up as a 'reliable workflow' or 'recommended approach'. That presents "
+    "an untested sequence of failures as validated guidance a future "
+    "session will trust and repeat. Either say 'Nothing to save', or, "
+    "only if you are independently confident of a real working alternative "
+    "(not something you are merely guessing might work), capture ONLY that "
+    "alternative — never the dead ends, and never dressed up as best practice.\n\n"
     "If a tool failed because of setup state, capture the FIX (install "
     "command, config step, env var to set) under an existing setup or "
     "troubleshooting skill — never 'this tool does not work' as a "

tests/run_agent/test_review_prompt_class_first.py

@@ -88,6 +88,28 @@ def test_combined_review_prompt_has_memory_section():
 
 
 
+def _assert_unresolved_failure_guidance(prompt: str, label: str) -> None:
+    """Unresolved task attempts must not become persistent skill guidance."""
+    lower = prompt.lower()
+    assert "unresolved failures" in lower, f"{label}: must identify unresolved failures"
+    assert "working method" in lower, f"{label}: must require a working method"
+    assert "told the user to check manually" in lower, (
+        f"{label}: must recognize an explicitly unresolved session"
+    )
+    assert "never the dead ends" in lower, f"{label}: must exclude failed attempts"
+    assert "independently confident" in lower, (
+        f"{label}: must limit exceptions to verified alternatives"
+    )
+
+
+def test_skill_review_prompt_rejects_unresolved_failures():
+    _assert_unresolved_failure_guidance(AIAgent._SKILL_REVIEW_PROMPT, "_SKILL_REVIEW_PROMPT")
+
+
+def test_combined_review_prompt_rejects_unresolved_failures():
+    _assert_unresolved_failure_guidance(AIAgent._COMBINED_REVIEW_PROMPT, "_COMBINED_REVIEW_PROMPT")
+
+
 # ---------------------------------------------------------------------------
 # Anti-pattern guidance — see issue #6051. The reviewer was learning transient
 # environment failures (e.g. "browser tools do not work" from a fresh-install