Skip to content

Releases: Da7-Tech/SureForge

Release list

SureForge 1.0.0

SureForge 1.0.0 Latest
Latest

Choose a tag to compare

View all tags
@Da7-Tech Da7-Tech released this 09 Sep 13:00
v1.0.0
dadf511

First public release.

SureForge is an instruction-only skill for complex work: research before asking, ask before planning, plan before building, verify before delivering, and get an independent review before calling anything done. The installed skill is the skills/sureforge/ folder, plain text, no runtime.

Install

npx skills add Da7-Tech/SureForge

Or copy skills/sureforge/ into your host's skill directory. Per-host directories are in the README.

What this release contains

  • The skill text reviewed in three earlier rounds, with the version identifier and one installer paragraph in the platform notes updated.
  • Fixes from a fourth review round of the 1.0.0 candidate (four independent reviews of the same snapshot). The list is in the changelog.
  • The evaluation kit, package checks, tests, and the seeded-fault audit script.

Checks made on this tag (2026-09-09)

Recorded here because they can only be made once the tag exists.

  • Remote installation: npx skills@1.5.23 add Da7-Tech/SureForge --copy from an empty project, targets claude-code, cursor, codex, devin, hermes-agent. Four directories were written (Cursor and Codex share .agents/skills/); each installed copy was compared with skills/sureforge/ at this tag by scripts/verify_install.py and matched in all 13 files. Skill directory fingerprint 08b15d1765ba1216….
  • Commit identity: every commit on main has author and committer Da7-Tech; no co-author trailers; the contributors list shows one account.
  • Source snapshot: python3 -B -m scripts.check_package passes with 52 files; source fingerprint a165abb13a7b…. 121 unit tests pass on Python 3.11.15 and 3.14.6. All 28 faults in scripts/mutation_audit.py are caught. skills-ref validate accepts the skill.

Behavior

The two pilots described in the README (GLM-5.2, 38 sessions; Grok 4.6 with and without the skill, 24 runs) confirmed that the models follow the workflow, declare their tier, say when no independent reviewer is available instead of pretending, and report checks they could not perform as blocked. No harmful behavior was observed in any run. The three-arm comparison on the maintainer's own task set is the next step and will be reported in a later release.

Assets 2
Loading