Superpowers-Inspired Hardening
Competitive audit of obra/superpowers v4.3.1 identified critical gaps in enforcement reliability. This release applies P0, P1, and P2 findings.
Highlights
- Context Compaction Survival (P0) — SessionStart hook (
context-reinforcement.sh) re-injects Iron Laws after context compaction. Enforcement rules no longer lost on conversation compression. - Description Trap Audit (P1) — 5 skill descriptions rewritten to opaque, outcome-focused format. Prevents model from skipping full skill reads.
- XML Enforcement Tags (P1) —
<HARD-GATE>tags on 5 Iron Laws for higher model compliance. Applied to skill-deep-research, skill-factory, skill-tdd, skill-verify, skill-debug. - Human-Only Skill Flag (P1) —
invocation: human_onlyon 5 expensive skills (factory, deep-research, adversarial-security, parallel, ship) — prevents auto-triggering without explicit user invocation. - Two-Stage Review Pipeline (P2) — New
skill-staged-review.md— Stage 1 validates spec compliance against intent contract, Stage 2 runs stub detection and code quality. Gate between stages. - EnterPlanMode Interception (P2) — PreToolUse hook (
plan-mode-interceptor.sh) re-injects enforcement rules when entering plan mode.
Files Changed
18 files (5 new, 13 edited), ~740 lines changed
Testing
33 new tests across 7 suites, 62/62 full suite passing