TruffleHog backstop: soft-warn (flag for review) instead of hard-reject; drop FP Twilio regex

app.py

@@ -197,19 +197,15 @@ async def donate(request: Request):
             status_code=422,
         )
 
-    # --- authoritative backstop: TruffleHog over the scrubbed trace ---------
-    # Catches what the regex pass cannot (vendor tokens with no fixed prefix,
-    # newly-rotated formats, etc.). Any flagged detector rejects the donation.
+    # --- TruffleHog soft-warn pass over the scrubbed trace ------------------
+    # Catches what the regex pass cannot (vendor tokens with no fixed prefix).
+    # Run WITHOUT verification so it never transmits candidate secrets — which
+    # means occasional false positives (e.g. a 32-char hash read as a "Box"
+    # token). It therefore does NOT auto-reject: findings are surfaced to the
+    # contributor in the response and recorded on the PR for the maintainer to
+    # review, on top of the human review every donation already gets. The regex
+    # pass above stays the hard block for crisp, high-confidence secret formats.
     th_detectors = trufflehog_findings(cleaned)
-    if th_detectors:
-        return JSONResponse(
-            {
-                "error": "secrets_found",
-                "detail": "TruffleHog (server backstop) flagged likely secrets that survived the client scrub. Donation rejected — clean the flagged values and retry.",
-                "detectors": th_detectors,
-            },
-            status_code=422,
-        )
 
     if not HF_TOKEN or not DATASET_REPO:
         # Not yet configured — accept-validate but don't pretend to publish.
@@ -224,14 +220,22 @@ async def donate(request: Request):
 
     # --- open the PR on the contributor's behalf ---------------------------
     try:
-        pr_url = _open_pr(cleaned, harness, filename)
+        pr_url = _open_pr(cleaned, harness, filename, warnings=th_detectors)
     except Exception as e:  # noqa: BLE001 — surface a clean message to the skill
         return JSONResponse({"error": "publish_failed", "detail": str(e)}, status_code=502)
 
-    return {"status": "submitted", "pr_url": pr_url, "path": f"sessions/{harness}/{filename}"}
+    resp = {"status": "submitted", "pr_url": pr_url, "path": f"sessions/{harness}/{filename}"}
+    if th_detectors:
+        resp["warnings"] = {
+            "trufflehog_unverified": th_detectors,
+            "note": "TruffleHog flagged these without verification — often false positives "
+                    "on high-entropy strings, so the donation was NOT blocked. Please confirm "
+                    "none is a real secret; a maintainer will also review before merging.",
+        }
+    return resp
 
 
-def _open_pr(cleaned_text, harness, filename):
+def _open_pr(cleaned_text, harness, filename, warnings=None):
     """Open a PR to the dataset with the cleaned trace, under the project token."""
     from huggingface_hub import HfApi, CommitOperationAdd
 
@@ -240,12 +244,20 @@ def _open_pr(cleaned_text, harness, filename):
         path_in_repo=f"sessions/{harness}/{filename}",
         path_or_fileobj=io.BytesIO(cleaned_text.encode("utf-8")),
     )
+    description = "Anonymous donation via Trace Commons ingestion server."
+    if warnings:
+        description += (
+            "\n\n⚠️ **Maintainer review needed.** TruffleHog (unverified) flagged the "
+            "following detector(s). These are frequently false positives on high-entropy "
+            "strings (hashes, IDs, base64), but confirm none is a real secret before merging:\n- "
+            + "\n- ".join(warnings)
+        )
     commit = api.create_commit(
         repo_id=DATASET_REPO,
         repo_type="dataset",
         operations=[op],
         commit_message=f"Donate {harness} trace ({filename})",
-        commit_description="Anonymous donation via Trace Commons ingestion server.",
+        commit_description=description,
         create_pr=True,
     )
     # create_commit returns an object whose pr_url is set when create_pr=True