Strix (STRIX) — report
What this is
strix/report is Strix's memory of what's been found so far in one scan, and the code
that turns that memory into the files a human (or a CI pipeline) actually reads at the
end. Every vulnerability an agent reports passes through here before it's considered
"real" and every artifact on disk is written from here.
What flows through it
Technical details
Package: strix/report/
Depends on: strix.config (dedup model settings), strix.telemetry (cost
callbacks).
Depended on by: strix.tools.reporting (every write goes through
ReportState), strix.tools.finish (finish_scan writes final narrative fields
here), strix.core.runner (reads/writes the global ReportState), strix.interface
(the viewer reads written artifacts back off disk; the CLI reads ReportState for
live display).
Files at a glance
| File | Role |
|---|---|
state.py |
ReportState — the in-memory + on-disk source of truth for one run; module-global singleton via get_global_report_state(). |
dedupe.py |
check_duplicate() — LLM-judged (or fast-path, for dependency findings) duplicate detection. |
writer.py |
write_run_record(), write_executive_report(), write_vulnerabilities() — turns ReportState into on-disk Markdown/CSV/JSON. |
sarif.py |
Builds SARIF 2.1.0 documents for CI/CD (findings.sarif). |
pricing.py |
resolve_litellm_model() — LiteLLM pricing table lookups. |
usage.py |
LLMUsageLedger — cumulative token/dollar cost tracking, including OpenRouter-specific streamed cost capture. |
Files & functions, explained
state.py
- ReportState(run_name) — tracks vulnerability_reports (the accumulating list),
final_scan_result, an LLMUsageLedger, a run_record dict mirrored to
run.json, and a vulnerability_found_callback hook the CLI/TUI register for live
display.
- get_run_dir() — lazily resolves/creates strix_runs/<run_name>/.
- hydrate_from_run_dir() — resume support: reloads run.json and
vulnerabilities.json, raising loudly on corruption rather than silently starting
fresh (avoiding vuln-0001 ID collisions overwriting prior files).
- add_vulnerability_report() — assigns a sequential vuln-NNNN ID and stores the
full structured finding.
dedupe.py
- check_duplicate(candidate, existing) -> dict — builds a dedicated small-model call
(optionally a distinct DedupeSettings model/key, else the main LLM) with a system
prompt instructing the model to judge "same root cause on same asset," parses a
structured is_duplicate/duplicate_id/confidence/reason response.
- _check_dependency_duplicate() — a deterministic fast path for CVE/dependency
findings, keyed on package + version + CVE id instead of LLM judgment.
writer.py
- write_vulnerabilities(run_dir, vulnerability_reports, saved_vuln_ids) — writes
only new reports as individual vulnerabilities/<id>.md files (idempotent — safe
to call repeatedly), while fully rewriting vulnerabilities.csv and
vulnerabilities.json every time. This split is what makes resume safe: previously
saved Markdown files are never re-touched.
sarif.py — maps each vulnerability report to a SARIF result (rule metadata,
CVSS-derived severity, CWE normalization, code-location conversion), applies
repository context (git remote/branch), and computes stable per-finding fingerprints
so re-runs produce comparable/diffable SARIF across CI runs. (Reviewed at the
function-signature level, not verified field-by-field against the SARIF 2.1.0 schema
— see ../design/LLD.md's "not independently verified" note.)
Related: STRIX — Architecture (HLD), tools.md, ../design/data-flow.md.