Skip to content

Releases: WePwn/demo_cli

Release list

v0.4.0-beta.8

v0.4.0-beta.8 Pre-release
Pre-release

Choose a tag to compare

@Y0ung-DST Y0ung-DST released this 19 Jul 17:03

Pre-release. Manual, verify-first install pinned to this tag — read the source and confirm the published SHA-256 before running anything. curl | sh remains available as convenience, not the safe path.

Highlights

  • Escalate everywhere. Removed the SANDBOX low-blast exception: an unrecoverable, non-recursive mutation in a development/test/sandbox/staging workspace no longer returns SANDBOX (SAFE) and proceeds with no recovery point. It escalates like everywhere else. The invariant "recoverable or escalate" now holds without exception; the SANDBOX disposition is gone.
  • Install pinned to a release tag, not the moving beta branch. install.sh / install.ps1 and every README snippet now reference v0.4.0-beta.8.
  • Relicensed to MIT across LICENSE, pyproject.toml, and the README (was Apache-2.0).
  • Fail-open is now loud on every path. The last silent path (unparseable hook input) now writes a visible stderr warning instead of a silent allow. Decisions themselves remain fail-closed.

Honesty fixes

  • Shared-receipt install command pinned to the release tag. A receipt --share card's verify command still pointed at @beta; it now derives the tag from the package version via version.release_tag() (0.4.0b8v0.4.0-beta.8), so a stranger verifies against exactly the code the receipt was written by. Locked by a regression test.
  • Fixed a self-defeating trust check in the README. The no-telemetry grep suggestion returned hits (urllib.parse, imported twice for pure string handling). The grep now covers actual network I/O (urlopen, requests, httpx, socket, http.client) and returns nothing; the two benign urllib.parse uses are named explicitly.
  • Corrected the opening comparison. Claude Code and Cursor checkpointing does cover the agent's own file edits — what it doesn't cover is shell commands and databases. The README now says that.
  • Corrected the recovery-demo caption. It implied the referenced incident's files were "gone for good"; file-carving recovers only a partial, luck-dependent fraction. The caption now states the honest case.

Docs & platform

  • Documented two standing limitations: shell parsing is pattern-based (full AST is roadmap; ambiguous parses treated as unrecoverable), and database coverage is SQLite + Postgres only (other engines escalate).
  • Added SECURITY.md with a responsible-disclosure contact.
  • Windows drive-letter detection is now cross-platform (via ntpath), so a multi-drive delete is recognised as having no common capture root on any host; its test runs everywhere instead of Windows-only.

Verify before you run. Download the .whl / .tar.gz and check them against SHA256SUMS.txt attached to this release:

sha256sum -c SHA256SUMS.txt