-
Made command help position-independent without stealing flag values, scrubbed credentials and URL userinfo from report and model-prose surfaces, and kept oversized score caches deterministically bounded even when one run touches more than 5,000 entries. Existing cache entries are scrubbed and rewritten when they contain pre-hardening model secrets. Report generation now also closes model discovery resources after every configured scoring run.
-
Rebuilt calibration from one fixed-time Git-aware cohort, bound reviewed cases to content identities, rejected archive-only provenance, excluded the derived fixture from its own cohort, and kept the existing qualitative scoring contract after screening all 2,002 mappings.
-
Confined manifest enrichment to regular in-repository files, corrected unknown-command help guidance, and completed
slither helpshell completion. -
Added canonical command help, output discovery (
slither outputs), and dependency-free bash/zsh completion from shared command metadata. -
Added strict
version --json, read-onlydoctor, compact report summaries, all canonical review-lane inventories, and privacy-minimal eval Markdown. -
Added the dependency-free
slither agentJSONL bridge with bounded, repository-bound report, query, context, and optional feedback operations. -
Added opt-in
slither.outcome/v1ledger persistence with owner-only file permissions, replacement checks, and privacy-minimal derived records. -
Added deterministic multi-report
slither evaloutput with input-integrity validation, partial-tail warnings, and calibration metrics. -
Documented offline behavior, protocol limits, snapshot invalidation, and evaluation formulas.