Skip to content

omit v0.4.0 — fail-closed enforcement, and receipts that are checked

Choose a tag to compare

@sriinnu sriinnu released this 18 Sep 21:33
· 28 commits to main since this release
v0.4.0
34302e0

A four-lens security review of the previous work produced 34 findings, 30 of them confirmed by execution. A fifth pass re-verified each one independently and dropped 2 whose stated impact did not survive checking. This release is the fix.

The architectural finding

Roughly half the confirmed items were one bug in different clothes: errors degraded to let it through instead of refuse.

  • git() returned null on every failure and every call site coalesced it to '', so "git could not answer" and "there is nothing here" were the same value. A diff past Node's 1 MiB buffer threw ENOBUFS, was swallowed, and the gate reported hazards: 0 ✅ with a healthy-looking line count while nothing had been scanned.
  • A hook that threw exited 1 — which the harness reads as a non-blocking error — so an objection became a stack trace the model moved past. A receipt citing a directory was enough to make the new-dependency objection vanish.
  • A committed symlink to /dev/zero hung git commit, the hooks and CI forever.

Every sentinel now fails closed, and "could not determine" is never rendered as an empty result.

Receipts are checked, not recorded

The gate validated receipts.includes(dep) — a substring match against a file the agent writes itself. Evidence is now bound to the claim it supports:

  • run must name what it exercises. Exit status alone proves nothing: ["true"] exits 0 and ["false"] exits 1 while testing neither the standard library nor anything else.
  • A tried entry cites a search, not a location. It names the symbol it looked for and omit searches the tree — tracked and untracked — for itself, so a symbol that is there refutes the author's own conclusion.

Also

  • Secrets are genuinely non-suppressible: the secret rules now run before any marker is consulted. The README claimed this; the code did not do it.
  • omit-allow: requires a trailing comment carrying a real reason. Bare, reason-less, or inside a string literal, it is ignored.
  • Dependency parsing covers BOMs, Cargo [workspace.dependencies], and four more manifest formats; the TOML array reader is linear (1,241 ms → 8 ms on 40k lines).
  • A >1 MiB diff, a rename brace path, a -diff git attribute, and a subdirectory working directory can no longer hide a change.
  • action.yml was invalid YAML at the previous release — an unquoted ": " in the description — so the Action may never have loaded.
  • SKILL.md now asks for a Final Draft the stop gate can actually parse; it previously asked for content in a shape the gate rejected, so honest reports were refused.

Breaking

Receipts written before 0.4.0 read as unverifiable rather than being silently accepted. omit verify names each one individually.

Verification

178 tests, from 54. Every original reproduction re-run end to end after the final changes. Fixes proven by mutation testing — un-doing one fix at a time to confirm its test actually bites. test/docs.test.mjs pins the rule files to the checker, so every documented receipt example must verify against a real fixture.