nyx/docs/dynamic.md

214 lines
8.3 KiB
Markdown
Raw Normal View History

# Dynamic verification
Nyx verifies every `Confidence >= Medium` finding by default: it builds
a minimal harness, runs your code's entry point against a curated payload corpus
inside a sandbox, and records the verdict in each finding's evidence block.
## Headline metrics
The dynamic-verification overhaul ships with four published acceptance targets,
gated end-to-end by `scripts/m7_ship_gate.sh` (Phase 31) against the eval
corpus (OWASP Benchmark v1.2 + NIST SARD subset + the in-house curated set
from `tests/benchmark/corpus`):
| Metric | Target | Gate | Source |
| --- | --- | --- | --- |
| Unsupported% per `(cap, lang)` cell | < 20% | M7 Gate 1 | `tests/eval_corpus/budget.toml` `[default].unsupported_rate` |
| False-Confirmed% per cap | < 2% | M7 Gate 2 | `~/.cache/nyx/dynamic/events.jsonl` (`kind: feedback`, `wrong: true`) |
| Repro stability | ≥ 95% | M7 Gate 5 | `~/.cache/nyx/dynamic/repro/*/reproduce.sh` exit 0 |
| Wall-clock cost | ≤ 2× static-only | M7 Gate 3 | `benches/fixtures/` (default vs `--no-verify`) |
The corresponding orchestrator is `tests/eval_corpus/run_full.sh`; it bundles
the three corpus sets, writes a canonical `tests/eval_corpus/results.json`,
and propagates the per-cell budget through `tabulate.py` and `report.py`.
A non-zero exit from `m7_ship_gate.sh` is a hard merge blocker for the
default-on flip. Failures map back to the engine follow-ups recorded in
`.pitboss/play/deferred.md` (per-language probe-shim splicing, composite
chain reverifier wiring, telemetry-stability stamping, et al.).
## Default-on semantics
```
nyx scan # verifies Medium+ findings (default)
nyx scan --no-verify # static analysis only, no harness execution
nyx scan --verify # same as default; explicit for clarity in scripts
```
`--no-verify` is the escape hatch. It overrides the config default for a single
run without changing `nyx.toml`.
### What "verified" means
A finding with `dynamic_verdict.status: Confirmed` was successfully triggered
by at least one payload in nyx's corpus. The corpus covers common patterns for
each vulnerability class (SQL injection, XSS, command injection, SSRF, etc.) per
language.
A finding with `dynamic_verdict.status: NotConfirmed` was attempted but no
payload fired. This is not a false-positive signal. It means the corpus did not
have a payload that matched the specific sink variant, or the execution path was
not reachable in the test harness.
A finding with `dynamic_verdict.status: Unsupported` could not be attempted.
Common reasons: confidence below threshold, no flow steps, language or sink type
not yet supported by the harness layer.
### Confidence gate
Only `Confidence >= Medium` findings are verified by default (§5.1). To also
verify low-confidence findings (for corpus building or backfill), pass
`--verify-all-confidence`:
```
nyx scan --verify-all-confidence
```
This is not recommended for production scans because low-confidence findings have
a higher false-positive rate and the harness may produce unreliable verdicts.
## nyx.toml opt-out
If you want static-only scans permanently, set `verify = false` in `nyx.toml`:
```toml
[scanner]
verify = false
```
This survives upgrades. The M7 default flip only changes the inherited default
for projects that have not explicitly set the field.
## Sandbox backends
nyx uses docker when available, then falls back to an in-process runner:
```
nyx scan --backend docker # require docker; fail if unavailable
nyx scan --backend process # in-process runner (no container; less isolation)
nyx scan --unsafe-sandbox # alias for --backend process
```
The docker backend mounts only the entry file's directory and blocks all
outbound network by default. When out-of-band detection is enabled (`oob_listener`
in config), the container gets `--network bridge` with a host-gateway route.
## Repro artifacts
When a finding is `Confirmed`, nyx writes a repro artifact to
`~/.cache/nyx/repro/<stable_hash>/`. The artifact contains the harness spec and
the triggering payload. You can regenerate the verdict with:
```
nyx scan --verify <path> # re-scans and re-verifies
```
See `docs/output.md` for the `dynamic_verdict` field schema.
## Wall-clock cost
Verification adds harness build + sandbox startup time per finding. On typical
codebases with 1050 Medium+ findings, end-to-end overhead is 25× static-only.
If scan time is unacceptable for a given workflow (e.g. IDE integration, quick
pre-commit check), use `--no-verify` for that workflow and rely on the full scan
in CI.
## Event schema
The dynamic layer writes one JSON record per verdict to
`~/.cache/nyx/dynamic/events.jsonl`. Every record begins with a fixed envelope
so older readers fail loudly instead of silently mixing incompatible shapes:
```json
{
"schema_version": 1,
"nyx_version": "0.7.0",
"corpus_version": "4",
"kind": "verdict",
"ts": "2026-05-15T18:42:09Z",
"finding_id": "a3b1...",
"spec_hash": "9f4e...",
"lang": "python",
"cap": "SQL_QUERY",
"status": "Confirmed",
"toolchain_id": "python-3.11",
"toolchain_match": "exact",
"duration_ms": 312,
"build_attempts": 1
}
```
| Field | Type | Meaning |
| --- | --- | --- |
| `schema_version` | integer | Bumped on any breaking change. Readers reject mismatches. |
| `nyx_version` | string | `CARGO_PKG_VERSION` of the writing binary. |
| `corpus_version` | string | Payload-corpus version the verdict was scored against. |
| `kind` | string | `"verdict"` (per-finding) or `"rank_delta"` (rank-score shift). |
| `ts` | RFC-3339 string | Wall-clock at write time. |
| `finding_id` | string | Stable finding identifier. |
| `spec_hash` | string | Hash of the `HarnessSpec` that drove the run. |
| `lang` | string | Language slug; `"unknown"` when spec derivation failed. |
| `cap` | string | Sink capability (e.g. `SQL_QUERY`, `CODE_EXEC`). |
| `status` | string | `Confirmed`, `NotConfirmed`, `Inconclusive`, or `Unsupported`. |
| `inconclusive_reason` | string | Present iff `status == Inconclusive`. |
A `rank_delta` record carries the envelope plus `finding_id`, `status`, and a
signed `delta` applied to the rank score.
### Schema-version mismatch
`scripts/m7_ship_gate.sh` Gate 2 walks every line of the log, requires
`schema_version == EXPECTED_SCHEMA_VERSION`, and exits 3 if any record fails
the check. Programmatic readers use
`crate::dynamic::telemetry::read_events(path)`, which surfaces the same
condition as `TelemetryReadError::SchemaMismatch { expected, found, .. }`.
When schema bumps land, the canonical migration is to roll the log over (move
or delete `events.jsonl`) so new and old records never coexist in a file. The
gate refuses to skip silently on mismatch.
### Sampling
`[telemetry]` in `nyx.toml` controls the on-disk sampling policy:
```toml
[telemetry]
keep_all_confirmed = true # default: retain every Confirmed verdict
keep_all_inconclusive = true # default: retain every Inconclusive verdict
sample_rate_other = 1.0 # 0.01.0 for NotConfirmed / Unsupported
```
`sample_rate_other < 1.0` downsamples NotConfirmed and Unsupported verdicts
deterministically. The decision is seeded by the finding's `spec_hash`, so a
given finding makes the same keep-or-drop call across reruns. Confirmed and
Inconclusive verdicts ignore the rate and are always retained (they gate the
false-Confirmed budget and drive the spec-derivation roadmap).
Rank-delta records (emitted by `emit_rank_delta` when a verdict shifts a
finding's position in the ranked output) are also retained unconditionally and
do **not** consult `sample_rate_other`. They are calibration-critical and small
in volume, so the carve-out is intentional; setting `sample_rate_other = 0.0`
to throttle log growth will still produce rank-delta lines.
`NYX_NO_TELEMETRY=1` disables every write regardless of the policy.
## Opting in to feedback
False positives (nyx says `Confirmed` but you disagree) can be recorded:
```
nyx verify-feedback <finding_id> --wrong "reason"
```
This writes to the local telemetry log (`~/.cache/nyx/dynamic/events.jsonl`)
and contributes to precision monitoring. Feedback is never uploaded automatically.
## nyx serve integration
The browser UI shows `dynamic_verdict` in each finding's detail panel and
uses the verdict in ranking (Confirmed findings surface first). The scan compare
page has a **Verdict Diff** tab that shows which findings changed verification
status between two scans.