Skip to content

v3.4.0 — Earned, not assumed

Choose a tag to compare

View all tags
@MarkusNeusinger MarkusNeusinger released this 12 Oct 15:43
· 271 commits to main since this release
v3.4.0
98dcb7a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The window since v3.3.0 holds 284 merged pull requests (#11839 to #12147). This release covers the review and scoring work and the general fixes; the agent network from the same window follows in v3.5.0. The full record is CHANGELOG.md, and each PR carries its own reasoning.

Added

  • Review retest harness. Re-runs the AI quality review on frozen, hand-labeled sets, so every rubric change is checked before it ships (#11964).
  • Review retest set v3: the 15 first generations of line-tanabe-sugano. The first spec reviewed entirely on Opus (#12038, #12043).
  • Regen gate records. Every regen gate decision carries a reason code (#11950).
  • CI checks every spec's "What a good version looks like" section. (#11948)
  • Check values. A spec's Notes can name one to three numbers a correct implementation reproduces (#12046).
  • Weekly regen review skill. A week of daily-regen regenerations becomes a dated report (#12055).

Changed

  • Every quality review runs on Opus. Whatever model generated the implementation; a manual review can pin another (#11968).
  • The AI review judges each plot against its own plot type. Specs gain a "What a good version looks like" section (#11946).
  • Regenerations only replace a live implementation when they are visibly better. (#11945)
  • The reviewer scores without the approval bar in view. (#12046)
  • A deduction names its defect. (#12046)
  • Compliance is not excellence. (#12046)
  • Polish and optional additions no longer carry a regeneration. (#11977)
  • Implementations compute with the runtime's libraries, not by hand. (#12005)
  • MCP tools return what an assistant needs and nothing it does not. (#12100)
  • The backend test job runs on four workers. (#12136)

Fixed

  • A review that fails after scoring is re-run instead of stranding the PR. (#12004)
  • Repair exhaustion no longer pushes to main. (#12004)
  • MCP tag search scopes values to their category and ANDs the categories. (#12100)
  • .env is read only from the checkout root. (#12105)
  • OG cards render from the 800px derivative again, so the API's memory stays flat. The render peak drops to about 157 MiB from 430-580 MiB (#11867).
  • Approved specifications merge on their own again. (#11868)

Full Changelog: v3.3.0...v3.4.0

Assets 2
Loading