sbxloop 0.7.4__tar.gz → 0.7.6__tar.gz

{sbxloop-0.7.4 → sbxloop-0.7.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: sbxloop
-Version: 0.7.4
+Version: 0.7.6
 Summary: Agentic loop orchestration on Docker Sandboxes (sbx) with isolated credential domains
 Project-URL: Homepage, https://github.com/brettbergin/sbxloop
 Project-URL: Repository, https://github.com/brettbergin/sbxloop

{sbxloop-0.7.4 → sbxloop-0.7.6}/src/sbxloop/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.7.4'
-__version_tuple__ = version_tuple = (0, 7, 4)
+__version__ = version = '0.7.6'
+__version_tuple__ = version_tuple = (0, 7, 6)
 
 __commit_id__ = commit_id = None

{sbxloop-0.7.4 → sbxloop-0.7.6}/src/sbxloop/daemon/backlog.py

@@ -47,7 +47,12 @@ AUDIT_INSTRUCTIONS = (
     "the tool that ran you: its planner, prompts, verify lint, delivery, daemon — "
     "go under `.sbxloop/backlog/tool/` instead; they are filed to the tool's own "
     "tracker when the operator configured one, otherwise only noted in the closing "
-    "comment, never filed as issues of this project."
+    "comment, never filed as issues of this project. VERIFICATION: an audit changes "
+    "no code, so its tasks need NO verify_commands (leave the list empty) — or at "
+    "most a check on the findings you wrote (e.g. `ls .sbxloop/backlog/*.md`). Never "
+    "the project's test suite, build or lint: an audit whose charter is to find a "
+    "failing test cannot also promise the suite is green (field failure rakvqn6fr, "
+    "where exactly that verify command sank the audit and its findings)."
 )
 
 TOOL_SUBDIR = BACKLOG_SUBDIR / "tool"

{sbxloop-0.7.4 → sbxloop-0.7.6}/src/sbxloop/daemon/loop.py

@@ -665,7 +665,12 @@ class DaemonLoop:
             self._notify(
                 f"✅ {item.item_id} done ({report.task_summary})"
                 + (f" · PR {report.delivery[1]}" if report.delivery else "")
-                + (f" · filed {len(report.filed)} finding(s)" if item.kind == "audit" else "")
+                + (
+                    f" · filed {len(report.filed) + len(report.tool_filed)} finding(s)"
+                    + (f" ({len(report.tool_filed)} upstream)" if report.tool_filed else "")
+                    if item.kind == "audit"
+                    else ""
+                )
             )
             self._file_review(item, run_id, report)
             return "done"
@@ -682,6 +687,17 @@ class DaemonLoop:
             self._file_postmortem(item, run_id, f"delivery failed: {report.delivery_error}")
             return "delivery_failed"
         reason = str(error) if error is not None else f"run ended {report.state}"
+        if item.kind == "audit" and result is not None:
+            # An audit that failed on the harness (a verify command it never
+            # needed) still wrote its findings; they are evidence, not code,
+            # so file them rather than lose them (field: rakvqn6fr).
+            filed = self._collect_backlog(run_id, source)
+            tool = self._collect_tool_findings(run_id, source)
+            if filed or tool.filed:
+                self._notify(
+                    f"🔎 {item.item_id} failed but its findings were filed: "
+                    f"{len(filed) + len(tool.filed)} issue(s)"
+                )
         attempts_left = self.config.daemon.max_attempts_per_item - item.attempts
         self._set_breaker(self._breaker_opened_at, self._consecutive_failures + 1)
         self.dstore.finish_ledger(run_id, "failed", now)

{sbxloop-0.7.4 → sbxloop-0.7.6}/src/sbxloop/daemon/sources.py

@@ -412,6 +412,7 @@ class GitHubIssueSource:
             }
             if trigger not in names:
                 return False
+            stale_delivered = self.labels.delivered in names
             epoch = self._trigger_epoch(ops, number, trigger)
             token = uuid.uuid4().hex
             self._comment(
@@ -449,10 +450,14 @@ class GitHubIssueSource:
         # mode was configured *then*, and a daemon restarted with
         # close_on_success back on would otherwise leave it beside
         # in-progress. Best-effort — a leftover label must not un-claim.
-        self._guard(
-            "clear delivered label",
-            lambda ops: self._remove_label(ops, number, self.labels.delivered),
-        )
+        # Only when the re-GET actually showed it: a blind DELETE 404s on
+        # every fresh issue and the event stream renders each as an error
+        # panel (field noise on every audit claim).
+        if stale_delivered:
+            self._guard(
+                "clear delivered label",
+                lambda ops: self._remove_label(ops, number, self.labels.delivered),
+            )
         return True
 
     def _trigger_epoch(self, ops: GithubOps, number: str, trigger: str | None = None) -> str: