Merge #5927: teach ephemeral prompt that redacted credentials are intentional masking

api/streaming.py

@@ -636,6 +636,7 @@ WebUI progress guidance:
 - Each update should say what you are about to check, what you just confirmed, or why the next tool call is needed.
 - Keep updates concise, factual, and in the user's language. One or two short sentences are enough.
 - Do not reveal hidden reasoning, chain-of-thought, private scratchpads, secrets, raw logs, or long tool output.
+- Password, API-key, token, and secret fields are automatically redacted by the system. Treat masked values as intentional redaction, not placeholder text or user input errors, and do not tell the user a stored credential is wrong based on a masked value alone.
 - Final visible assistant replies must be clear, user-facing, and in the user's language, not private planning notes.
 - Do not include terse planning fragments or scratchpad shorthand in visible assistant text. Avoid fragments like "Need script", "Need check logs", "Need inspect email", or "maybe invite"; either omit them or rewrite them as clear user-facing progress.
 - For direct answers or very short tasks, skip progress updates and answer normally.

tests/test_issue5871_redaction_awareness_prompt.py

@@ -0,0 +1,47 @@
+"""Product-semantics coverage for WebUI credential-masking guidance (#5871)."""
+
+import re
+
+from api.streaming import _WEBUI_PROGRESS_PROMPT, _webui_ephemeral_system_prompt
+
+
+def _redaction_guidance() -> str:
+    return next(
+        line for line in _WEBUI_PROGRESS_PROMPT.splitlines()
+        if "automatically redacted" in line
+    )
+
+
+def test_progress_prompt_explains_intentional_credential_redaction():
+    guidance = _redaction_guidance()
+
+    assert "Password, API-key, token, and secret fields" in guidance
+    assert "intentional redaction" in guidance
+    assert "not placeholder text or user input errors" in guidance
+    assert "do not tell the user a stored credential is wrong" in guidance
+    assert "masked value alone" in guidance
+
+
+def test_redaction_guidance_flows_into_ephemeral_system_prompt():
+    prompt = _webui_ephemeral_system_prompt(
+        "Use a concise tone.",
+        surface_context={"source": "webui"},
+    )
+
+    assert _redaction_guidance() in prompt
+
+
+def test_progress_prompt_retains_secret_leak_prevention_guidance():
+    assert (
+        "Do not reveal hidden reasoning, chain-of-thought, private scratchpads, "
+        "secrets, raw logs, or long tool output."
+    ) in _WEBUI_PROGRESS_PROMPT
+
+
+def test_redaction_guidance_does_not_teach_concrete_key_or_mask_formats():
+    guidance = _redaction_guidance()
+
+    assert "***" not in guidance
+    assert "..." not in guidance
+    assert not re.search(r"\b(?:sk-|ghp_|github_pat_|xox[baprs]-|AKIA)\S+", guidance)
+    assert not re.search(r"\b[A-Za-z0-9_-]{24,}\b", guidance)