chore: public OSS release 2.9.2

CHANGELOG.md

@@ -1,3 +1,7 @@
+### 2.9.2 (Monday, August 10, 2026)
+### Features/Bug Fixes
+* fix(llm): retry malformed structured responses
+---
 ### 2.9.1 (Monday, August 10, 2026)
 ### Features/Bug Fixes
 * fix(llm): add bounded connection retries

docs/release/skillspector-2.9.2.md

@@ -0,0 +1,48 @@
+# SkillSpector v2.9.2
+
+Released: 2026-08-10
+
+## Summary
+
+This patch release makes structured LLM response handling more resilient to transient malformed payloads. It retries validation and structured-output parse failures with bounded backoff while retaining fail-closed batch isolation after the retry budget is exhausted.
+
+## Highlights
+
+- Structured LLM response failures now receive bounded retries before a batch is isolated.
+
+## Added
+
+- None.
+
+## Changed
+
+- Applied the structured-response retry policy centrally to semantic analyzers, the meta-analyzer, TP4, and gap-fill paths.
+
+## Fixed
+
+- Prevented transient malformed structured responses from immediately exhausting required LLM batches.
+
+## Security
+
+- None.
+
+## Breaking Changes and Migration
+
+- None.
+
+## Deprecations
+
+- None.
+
+## Validation
+
+- `git diff --check release/2.9.1..origin/main` — passed.
+- Automated CI — lint, unit, integration, and Docker smoke checks passed; Sonar analysis succeeded.
+
+## Known Limitations
+
+- Requests still fail closed after the bounded retry budget is exhausted.
+
+## References
+
+- `CHANGELOG.md`

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "skillspector"
-version = "2.9.1"
+version = "2.9.2"
 description = "SkillSpector: Security scanner for AI agent skills (Claude Code, Cursor, and similar). Scans skills for vulnerabilities, malicious patterns, and security risks before installation. Supports Git repos, URLs, zips, and local directories; runs static pattern checks and optional LLM semantic analysis; outputs terminal, JSON, and Markdown reports with risk scoring."
 readme = "README.md"
 license = "Apache-2.0"

src/skillspector/inspection_ledger.py

@@ -78,7 +78,7 @@ class LedgerReason(StrEnum):
     LedgerReason.SYNTAX_ERROR: "Python source could not be parsed.",
     LedgerReason.LLM_BATCH_FAILED: "LLM analysis failed for this file range.",
     LedgerReason.LLM_STRUCTURED_RESPONSE_INVALID: (
-        "LLM returned a malformed structured response after retry."
+        "LLM returned a malformed structured response after bounded retries."
     ),
     LedgerReason.LLM_CONNECTION_RETRIES_EXHAUSTED: ("LLM connection failed after bounded retries."),
     LedgerReason.ANALYZER_RUNTIME_ERROR: ("Analyzer failed after beginning applicable work."),

src/skillspector/llm_analyzer_base.py

@@ -62,9 +62,11 @@
 logger = get_logger(__name__)
 
 DEFAULT_MAX_LLM_CONCURRENCY = 10
-STRUCTURED_RESPONSE_MAX_ATTEMPTS = 2
 API_CONNECTION_MAX_RETRIES = 3
 API_CONNECTION_RETRY_DELAYS_SECONDS = (0.5, 1.0, 2.0)
+STRUCTURED_RESPONSE_MAX_RETRIES = 3
+STRUCTURED_RESPONSE_MAX_ATTEMPTS = STRUCTURED_RESPONSE_MAX_RETRIES + 1
+STRUCTURED_RESPONSE_RETRY_DELAYS_SECONDS = API_CONNECTION_RETRY_DELAYS_SECONDS
 LLM_BATCH_MAX_ATTEMPTS = STRUCTURED_RESPONSE_MAX_ATTEMPTS + API_CONNECTION_MAX_RETRIES
 
 
@@ -624,11 +626,16 @@ def _invoke_batch_with_retries(self, batch: Batch, prompt: str) -> tuple[Batch,
                     or attempt == LLM_BATCH_MAX_ATTEMPTS
                 ):
                     raise
+                delay = STRUCTURED_RESPONSE_RETRY_DELAYS_SECONDS[structured_retries]
                 structured_retries += 1
                 logger.warning(
-                    "LLM structured response validation failed for %s; retrying once",
+                    "LLM structured response validation failed for %s; retrying in %.2fs (%d/%d)",
                     batch.file_label,
+                    delay,
+                    structured_retries,
+                    STRUCTURED_RESPONSE_MAX_RETRIES,
                 )
+                time.sleep(delay)
             except Exception as exc:
                 if (
                     not _is_retryable_api_connection_error(exc)
@@ -685,11 +692,16 @@ async def _ainvoke_batch_with_retries(self, batch: Batch, prompt: str) -> tuple[
                     or attempt == LLM_BATCH_MAX_ATTEMPTS
                 ):
                     raise
+                delay = STRUCTURED_RESPONSE_RETRY_DELAYS_SECONDS[structured_retries]
                 structured_retries += 1
                 logger.warning(
-                    "LLM structured response validation failed for %s; retrying once",
+                    "LLM structured response validation failed for %s; retrying in %.2fs (%d/%d)",
                     batch.file_label,
+                    delay,
+                    structured_retries,
+                    STRUCTURED_RESPONSE_MAX_RETRIES,
                 )
+                await asyncio.sleep(delay)
             except Exception as exc:
                 if (
                     not _is_retryable_api_connection_error(exc)
@@ -793,8 +805,9 @@ async def arun_batches(
         native retry timing remains provider-managed. Unrecovered errors cost
         only their own batch and are omitted from the result.
         Malformed structured responses (Pydantic ``ValidationError`` or CLI
-        JSON parse failures) are retried once and then isolated to their batch.
-        A batch makes at most five outer chat-model invocations even when both
+        JSON parse failures) receive three bounded exponential-backoff retries
+        and are then isolated to their batch. A batch makes at most seven outer
+        chat-model invocations even when both
         retry policies apply; native provider retries can make additional HTTP
         requests within one invocation.
         Callers can detect partial results by comparing the returned batches