Commit Graph
7 Commits
Author SHA1 Message Date
Marius CiepluchandClaude Opus 4.8 867bd536a2 feat(fusion): Phase 13 — two-panel fusion adjudication
OpenUltraCode-style deepening for findings that need more reasoning than the
ranker→hunter→verifier→mapping loop provides.

- fusion.py (zero-dep core + lazy LLM path): a deterministic trigger policy
  (critical/high severity, verifier disagreement, static-vs-semantic conflict,
  gates-risky-fix, high-assurance request); two panels that steel-man the
  vulnerability and false-positive cases; a deterministic, auditable decider that
  reconciles to one of five dispositions (accepted/rejected/mitigated/deferred/
  blocked); full disclosure (votes, model IDs, degradations, warnings, decision
  source). fuse_findings_dispatch routes panels through the configured HarnessX
  provider when panel_model + extra are present, else deterministic + a degradation.
- config [fusion]: enabled (default true), panel_model, decider_model,
  high_assurance.
- cli: standard-mode fusion stage runs on triggered findings, writes fusion.json
  and a manifest `fusion` block; quick mode never fuses. reports.write_manifest
  gained the fusion summary param.
- 8.4 zero-dep guard extended to cover fusion.py (lazy harnessx import only).
- README fusion section marked implemented + documented; openrouter-sast-harness
  Phase 13 checked off.

tests/test_fusion.py (16 tests) + config tests. 169 passed, 4 skipped; ruff + mypy
clean; detection gate PASS.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
2026-06-22 18:57:03 +02:00
Marius CiepluchandClaude Opus 4.8 ef59c0ae3d feat(harnessx): Phase 3 task 6 — hunter sub-harness + llm-judge verifier + fallback
Adopt HarnessX at the hunter loop and verifier, capability-gated so the default
zero-dependency scan is byte-identical and CI (no extra) exercises the fallback.

- 6.1/6.2 build_sast() (single, lazy HarnessX composition site in harness_ext.py,
  mirrors DocuHarnessX make_docgen) + HxScanOrchestrator (hunter_harness.py): a
  drop-in for run_hunter_pool that runs one bounded HarnessX agent task per ranked
  target (ModelConfig(main=AnthropicProvider).agentic(build_sast()).run(BaseTask)),
  parses findings from a JSON structured-output contract, returns the existing
  HunterPoolResult, maps scan_id->session_id and emits per-target exit_reason/cost/
  steps to the outer trace.
- 6.4 verify_judge.verify_findings_dispatch: a drop-in superset of
  verification.verify_findings — structural verifier verbatim by default (rejects
  unknown reachability), HarnessX llm-judge (build_judge_prompt/parse_judge_response)
  when the extra + a verifier model are present; verdict mapped to VerificationResult
  evidence-gated (plausible>=0.6 accepted; unsupported/refused rejected; infra
  failure never auto-rejects).
- 6.5 cli capability gate (hx_hunter/hx_verify): selects HxScanOrchestrator /
  judge only in standard mode with a configured model AND has_harnessx(); otherwise
  falls back and records a {stage,reason:harnessx_extra_unavailable} degradation in
  the manifest (new append-only write_manifest degradations kwarg). With no model
  configured the manifest stays byte-identical.
- pyproject: mypy override ignore_missing_imports for harnessx.* (absent in the
  type-check env).

Verified the HarnessX path constructs against the real extra in a throwaway venv
(build_sast -> HarnessConfig; orchestrator -> Harness; build_judge_prompt/
parse_judge_response/verdict mapping) — only the live agent.run()/provider.complete()
LLM calls are offline-untestable and stay gated. 109 tests pass (2 construction
smokes skip without the extra), ruff + mypy clean, detection gate 93.6%/0% intact.
Tasks 6.1/6.2/6.4/6.5 done; 6.3 (deterministic stages as HarnessX processors) deferred.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
2026-06-22 16:29:52 +02:00
Marius CiepluchandClaude Opus 4.8 ae365f6cb7 feat(governance): Phase 1 — CWE policy store + project score (zero-dep)
First implementation slice of the approved harnessx-self-improving-rulesets spec
(Foundation: governance + scoring planes). Additive and zero-dependency; findings
stay byte-identical and the >=90% recall / <10% FP detection gate is unchanged.

- PolicyStore (policy/verycode.py): vendors verycode-policies CWE_Score.tsv (167
  CWEs, CRLF + trailing-space header parsed defensively); CwePolicy(severity 0-5,
  static/dynamic); load_policy, resolve_severity (severity is policy-governed,
  keyed on CWE — one CWE = one severity), assert_rules_resolve (fail-loud if an
  enabled rule's CWE is unmapped). Remap the C/C++ buffer rules CWE-120 -> CWE-121
  (CWE-120 is absent from verycode; 121 = Stack-Based Buffer Overflow, severity 5).
- ScoreModel (scoring/project_score.py): SEV_WEIGHT, REACH_MULT, finding_penalty,
  project_score (0-100 exponential decay; one sev-5 reachable -> ~43), two-condition
  gate (sev-5+reachable always fails; score<min_score fails only when blocking),
  build_score_artifact.
- Wire into _run_scan: policy_load + fail-loud policy_check stages; score stage
  writes score.json and merges a score block into the manifest (advisory-first).
- ScoreConfig (k/min_score/block_severity_reachable/blocking) in config.py.
- pyproject uv_build ships the vendored TSV in the wheel (verified).
- Tests: test_policy, test_scoring, test_score_pipeline (artifact + fail-loud).
  85 passed; ruff + mypy clean. Spec phases 1.1-2.4 checked off; spec approved.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
2026-06-22 12:31:35 +02:00
Marius Ciepluch be65aed96e feat(hunter): add standard-mode scheduling 2026-06-18 10:59:25 +02:00
Marius Ciepluch cad70721cb feat(reports): add SARIF manifest artifacts 2026-06-18 10:16:25 +02:00
Marius Ciepluch f8c690999a feat(harness): add runtime and reachability mapping 2026-06-18 07:57:06 +02:00
Marius Ciepluch e7df721d16 feat(scan): add initial SAST pipeline 2026-06-17 21:13:51 +02:00