Test Report
Recorded by npm run proof on 2026-11-22T18:43:05.809Z, Node v24.15.0, win32/x64, IntentSutra 0.2.1. Raw output: test-results/test-run.tap. Nothing here is hand-edited.
Summary
| Suite | Result |
|---|---|
Automated tests (node --test) | 118 total, 116 passed, 0 failed, 2 skipped |
| Independent audit (honest demo workspace) | 11 checks: 11 passed, 0 failed; verdict PASS |
| Independent audit (after silent registry edit) | verdict FAIL |
| Fuzzing, fixed seed | 5398 generated inputs; 0 unexpected exceptions or undetected tampering |
The 2 skipped tests create file symlinks, which Windows forbids without Developer Mode. They are reported as skipped, not passed. Directory-link tests use junctions and ran.
Coverage
- Functional: import, stable IDs, duplicate import, revision, approval, links, attestation, impact, reports.
- Proof policy: approval is never inferred; a ticket or PR is not completion; a test file is not a test pass; a failed imported run is not a pass; a revision invalidates approvals and evidence; linking an old revision is rejected.
- Security: traversal, symlink, oversize, HTML and Markdown escaping, tampered audit log, API without token, POST rejected, no CORS, IDE overwrite refused.
- Audit: every AUD check has a pass case and a failing case (forged approval, forged identity flag, upgraded label, tampered chain, drifted evidence, edited content, seal tamper).
- CLI: exit codes,
--version, boolean flags before other flags,--strict,verify.
Fuzzing
Deterministic seed 20261122. 3000 mutated requirement records, 1500 Markdown inputs, 300 hostile import files (null, arrays of junk, invalid JSON, empty, wrong shapes), 398 single-character edits of the audit log (all detected) and 200 renders of injection titles through both HTML and Markdown. Only the intended IntentError is accepted as a failure mode. See fuzz-results.json.
Defects found and fixed during this review
- The CLI parsed arguments outside its error handler, so a stray argument printed a raw stack trace. Now a clean error and exit 1.
--confirm-humanconsumed the next argument when it was not last, silently turning the approval into a refusal. Boolean flags are now explicit.- Invalid JSON,
null, or non-object items in an import threw rawSyntaxError/TypeError. They now raiseINVALID_JSON/INVALID_INPUT. Imports are capped at 500 items. - Five symlink tests failed on Windows with
EPERM. They now use junctions for directories and skip file links when the OS forbids them. auditprinted only a chain status. It now runs 11 checks and writes sealed JSON, Markdown and HTML reports.- The package was marked
privateand had no tarball smoke test; both fixed.
An early run of the new fuzz harness reported 20 failures. They were a harness bug (requirement IDs shorter than the 3-character minimum), not a product defect; the harness was fixed, not the result ignored.
CI matrix
Node 22 and 24 on Ubuntu and Windows, Node 24 on macOS, an npm pack clean-install smoke test, and a secret scan. Hosted-runner results appear in the repository Actions tab; they are not part of the local proof.
Not tested
Native execution in any IDE, signatures, multi-process concurrency, very large registries, accessibility of the HTML reports with assistive technology (structure follows WCAG 2.2 practices, conformance is not claimed).
Reproduce
git clone https://github.com/shahvatsalofficial/intentsutra-ai && cd intentsutra-ai
npm test && npm run proof