Summary
Several source audit queries rank comments, safe wrappers, sanitizer vocabulary, or parser identifiers instead of executable risk sites: raw-diagnostic-echo, unbounded-json-parse, file-read-all-bytes, environment-secret-source, credential-term, process-argument-list, watcher-boundary, channel-boundary, and StringBuilder materialization. Add code/call-site origin constraints and positive-evidence exclusions.
Evidence source
Dogfooded with the locally built cdidx 1.37.2 on origin/main 098db58. Full tests passed before audit.
Duplicate check
No open issues existed at filing time; closed issues were intentionally not used as blockers per filing instructions.
Additional origin-classification gap
MCP tool schema descriptions embedded in production string literals are classified as ordinary string_literal origins. Risk recipes can therefore match example code such as new HttpClient inside tool descriptions. Add a help-like or schema_description origin so implementation audits can exclude these examples without dropping real runtime strings. Evidence: src/CodeIndex/Mcp/McpToolDefinitions.cs.
Additional exhaustive-audit evidence
A second full recipe pass produced 95 draft candidates that collapse into the same origin/positive-evidence problem rather than 95 independent defects. The affected families include process/subprocess policy, plugin and hook lifecycle, regex timeout policy, JSON/stream materialization, SQLite command policy, cancellation and broad-catch review, secret/auth vocabulary, path-case probes, recursive deletion, and resource ownership.
Representative safe or policy-bearing paths include BoundedJson.cs, BoundedRegex.cs, SqliteCommandPolicy.cs, ProcessLaunchPolicy.cs, SubprocessEnvironmentPolicy.cs, DiagnosticRedactor.cs, SensitiveNameClassifier.cs, and ExtensionAssemblyLoadContext.cs. Recipe issue-drafts should distinguish an unsafe call site from the implementation of the safety wrapper or policy itself, and should avoid producing both a family-level draft and redundant per-file drafts for the same evidence.
Summary
Several source audit queries rank comments, safe wrappers, sanitizer vocabulary, or parser identifiers instead of executable risk sites: raw-diagnostic-echo, unbounded-json-parse, file-read-all-bytes, environment-secret-source, credential-term, process-argument-list, watcher-boundary, channel-boundary, and StringBuilder materialization. Add code/call-site origin constraints and positive-evidence exclusions.
Evidence source
Dogfooded with the locally built cdidx 1.37.2 on origin/main 098db58. Full tests passed before audit.
Duplicate check
No open issues existed at filing time; closed issues were intentionally not used as blockers per filing instructions.
Additional origin-classification gap
MCP tool schema descriptions embedded in production string literals are classified as ordinary string_literal origins. Risk recipes can therefore match example code such as new HttpClient inside tool descriptions. Add a help-like or schema_description origin so implementation audits can exclude these examples without dropping real runtime strings. Evidence: src/CodeIndex/Mcp/McpToolDefinitions.cs.
Additional exhaustive-audit evidence
A second full recipe pass produced 95 draft candidates that collapse into the same origin/positive-evidence problem rather than 95 independent defects. The affected families include process/subprocess policy, plugin and hook lifecycle, regex timeout policy, JSON/stream materialization, SQLite command policy, cancellation and broad-catch review, secret/auth vocabulary, path-case probes, recursive deletion, and resource ownership.
Representative safe or policy-bearing paths include
BoundedJson.cs,BoundedRegex.cs,SqliteCommandPolicy.cs,ProcessLaunchPolicy.cs,SubprocessEnvironmentPolicy.cs,DiagnosticRedactor.cs,SensitiveNameClassifier.cs, andExtensionAssemblyLoadContext.cs. Recipe issue-drafts should distinguish an unsafe call site from the implementation of the safety wrapper or policy itself, and should avoid producing both a family-level draft and redundant per-file drafts for the same evidence.