Skip to content

The verification loop

A threat model written from reading code is a list of guesses. Some are right. The loop is how Bravos tells which: each claim is turned into a test, run against a target you own, and recorded with what happened. It is then checked again when the code under it changes.

Step Skill What it leaves for the next
Learn what the app is for bravos-context .guardlink/intent.md: invariants (INV-n) and intended-by-design behaviour (INT-n) every claim is judged against
Write the claim bravos-annotate @exposes, @mitigates, @flows and the rest, beside the code, passing guardlink’s lint gate
Make it testable bravos-templatize a validated cxg template, or a retest kit, per open exposure; classes with no oracle recorded as gaps
Bring up a target bravos-env a target you own, loopback by default, a reviewed scope.yaml and your typed authorization
Run it bravos-prove cxg’s verdicts, written back only through guardlink’s hypothesis ledger
Check it again bravos-recheck stale claims re-verified, retest kits replayed after a fix
Write it up bravos-report an assessment, one finding, a retest or a brief, bound to the gathered evidence

The code graph and SAST make each step sharper. With a graph, bravos-templatize targets the route and auth that reach each sink, and bravos-prove tests the most reachable exposures first. SAST findings joined to a claim are taint evidence for it. Without them the skills read the code and say so.

Keep scanner checks and replay kits separate

Section titled “Keep scanner checks and replay kits separate”

bravos-templatize uses GuardLink’s in-session translate instructions: guardlink_translate or guardlink translate --stdout. Ordinary scanner checks live under .guardlink/cxg-templates/; reviewed JavaScript replay kits live under .guardlink/cxg-replay/. The replay helper binds a kit’s bytes to the current GuardLink claim before execution.

bravos-prove runs bound kits through ordinary cxg scan with zero model calls. The helper checks the native engine result, tested source, target build and claim binding before writing a matching confirmed/refuted outcome through GuardLink and reading it back. Ambiguity, missing identities, changed inputs or failed writeback stay pending with a reason. A scanner match is a lead, and an empty confirmed list is not a refutation. Every probe is written and reviewed in the active agent session. Bravos uses only cxg scan for tests, including Electron and Tauri proof; it does not invoke cxg’s pentest orchestration or model generation. cxg pentest auth remains the login-capture interface. Electron scans use --target-type electron with --app-cmd or --app-binary; Tauri in-process proof uses --tauri-prove. A probe needing no login declares @requires_identities: {} or @requires_auth_count: 0, because an undeclared identity requirement defaults to one captured login.

The words for each state are on reading a finding.

Code changes. A claim tested last month is about code that may no longer exist. guardlink marks a claim stale when the code under it changes, and bravos-recheck takes it from there. Autopilot, the pre-PR gate and the nightly keep the model moving with the code. See Run Bravos in CI/CD.

The agent in your session does every step, with the real CLIs: guardlink for the model and its ledger, cxg for the runs. Bravos never spawns another agent CLI and has no server. It keeps local run evidence, claim bindings, receipts and report data so a later check can identify exactly what was tested. After a recorded engine confirmation, the agent offers to promote the verified kit to security/ (or the configured security_tests_dir) as a regression test. That needs your agreement and a review of scope, oracle and runner; credentials and raw run evidence stay outside git.