Skip to content

v9.54.0

Choose a tag to compare

View all tags
@nyldn nyldn released this 18 Jul 23:57
v9.54.0
5e03202
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Added

  • Anti-slop design taste layer (skills/blocks/design-taste.md): binary, mechanically checkable rules distilled from the highest-starred public taste rulebooks — the three banned AI-slop looks, banned default fonts (Inter, Roboto, Space Grotesk, Fraunces, Instrument Serif), palette and layout tells, content tells (placeholder personas, fake-perfect stats), and a pre-ship checklist. Wired into /octo:design-ui-ux as generation constraints, a new SLOP critique dimension in the three-way adversarial review, and a Deliver-phase gate. The Design Shotgun example variants themselves shipped two banned looks (AI-purple gradient, Space Grotesk default) and were replaced.

  • WCAG contrast validator (scripts/helpers/contrast-check.py, stdlib-only): computes WCAG 2.x contrast ratios for FG:BG hex pairs with normal/large thresholds, JSON output, and pairs-file input; exit 1 on any failure. The design skill's Deliver phase now runs it on every text/background token pair instead of prose-only "validate against WCAG AA".

  • Design dials in /octo:design-ui-ux: a fourth intake question maps to ui-ux-pro-max v2.11.0 --variance/--motion/--density flags (conservative/balanced/expressive/maximal presets), passed to searches and recorded in the design direction.

  • Design-system persistence: the design skill's final step now writes the design system to ~/.claude-octopus/designs/<slug>/ under the skill-design-lineage contract (branch-stamped, supersedes-chained), and downstream skills are directed to read the newest design document before inventing new tokens.

  • Enforcement patterns block (skills/blocks/enforcement-patterns.md): documents the three patterns that hold under generation pressure — one Iron Law per discipline, rationalization tables grown from observed excuses, and terminal states that name the successor skill instead of offering a next-steps menu. skill-verification-gate gains a rationalization table; the four flow skills gain explicit Terminal State sections chaining discover → define → develop → deliver → ship.

  • Skill Quality Gate (docs/DEVELOPER.md): new or substantially changed skills require a baseline-vs-with-skill eval pass (≥85% with-skill pass rate), a description trigger test (~10 should / ~10 should-not phrasings), and token budgets, before shipping.

Changed

  • Skill descriptions rewritten to triggers-only for six skills (staged-review, review-response, verification-gate, verify, native-escalation-routing, intent-contract). Descriptions that summarize the workflow create a shortcut the model takes instead of loading the skill; the two verification skills also had identical descriptions, which broke routing between them.
  • Vendored ui-ux-pro-max design intelligence refreshed from v2.0.1 to v2.11.0 (nextlevelbuilder/ui-ux-pro-max-skill, MIT). Brings 11 releases of new data and features into /octo:design-ui-ux: expanded databases (84 styles, 192 palettes, 74 font pairings, 161 UX reasoning rules), a Google Fonts collection, motion presets, design dials (--variance/--motion/--density), and --persist MASTER.md design-system persistence. The vendored subset is now slimmer (src/ plus license and docs; screenshots and the npm CLI are no longer copied) and carries a VENDOR.json manifest recording the pinned upstream tag. scripts/check-vendor-updates.sh was rewritten for plain-file vendoring (the old version silently exited because it still expected .gitmodules, dead since #253) and now compares the manifest tag against the latest upstream release; a nightly Vendor Freshness CI job fails when a vendored dependency goes stale.

Fixed

  • Auto-router no longer coerces skill routing and never routes on system events (#632). The UserPromptSubmit auto-router injected "MANDATORY: Invoke Skill(...) before responding" even when intent detection mis-scored a prompt, pressuring the agent to hijack the turn into an unrelated multi-provider workflow; it also fired on harness-generated turns (task notifications, system reminders) that are not user input. Routing context is now explicitly advisory on both sides of the contract (user-prompt-submit.sh message and auto-router-inject.sh session instruction), and prompts beginning with system-event markers ([SYSTEM NOTIFICATION, <task-notification>, <system-reminder>, <local-command-stdout>) are skipped before intent detection.
Assets 2
Loading