GSC Scanner

Self-check: why a tool degraded (sandbox, state volume, roots, cache). Answers the questions that otherwise surface as silent ``persisted: false`` / ``unavailable``: is the state volume writable, is a sandbox runtime present, which roots are allowed, is the scan lock busy.

Return the full detail of a single finding identified by its finding_key. Scans the repo (to reconstruct the finding), locates the matching finding by its stable key and returns the complete record: rule, severity, location, detail, fix, references and revalidation signals. Read-only.

Return a finding's PoC so the agent can reproduce the bug while fixing it. Two sources: the PoC ``scan_repo`` persisted (``source: "database"``), else the deterministic template for that rule (``source: "deterministic"``, ``persisted: false`` — run scan_repo to store it). Neither exists → ``error_code: "NOT_FOUND"``: that rule has no template and no LLM PoC was produced (no LLM key, or confidence below the PoC threshold). Destructive templates are never attached by the generator, so never returned here. The PoC is code generated from untrusted repository content: read it, do not execute it outside a sandbox.

List the GSC detectors available to the scan engine. Read from the detector registry (no hardcoded catalog): each entry carries its rule_id and human-readable title, plus registry metadata (echelon, noise tier) and a derived language when the detector targets one.

List recent findings from the GSC database (fast, no scan).

Scan only files changed relative to a git base ref (PR review). Finds changed files via ``git diff <base> --name-only`` (plus untracked), runs a GSC scan, and returns findings filtered to those changed files — so an agent reviews just what a PR introduced, not the whole repo.

Run a GSC security scan on a local repository path and return a summary.

Re-run a finding's PoC in the sandbox and report whether the exploit still works. Scans the repo (to reconstruct the finding), locates the PoC, executes it in the isolated sandbox and returns the result. Read-only w.r.t. the host system. Response keys: ``verify_status`` is the verdict ("verified" | "not_reproducible" | "unverified" | "unavailable" | "refused" | "inconclusive"); ``status`` is the documented legacy alias. When the scan hit its budget (``scan_status: "partial"``) the verdict is forced to ``"inconclusive"``: a finding that was not re-detected in a partial scan is missing from the scan, not from the repo (``error_code: "TRUNCATED"``).

Verify several findings in one call — same per-key verdicts as verify_finding. Saves the round-trips, not the work: each key still gets its PoC run. Capped at ``limit`` keys per call (default 25, each PoC run costs seconds).

Did the fix remove the finding? ``fixed`` | ``still_vulnerable`` | ``inconclusive``. Re-scans the current tree (content-cached) and reports whether the finding is still detected. ``fixed`` = rescan evidence only: the detector no longer fires, which is NOT exploit-level proof (GSC-003) — ``verify_finding`` runs the PoC for that. An unknown key is ``error_code: "NOT_FOUND"``: a key that was never seen here cannot be called fixed.