v0.2.0 - always-on hooks, real deps, and the honest null benchmark

D Strange v0.2.0.

What's new:

• Always-on hook. A UserPromptSubmit hook (hooks/inject-ruleset.mjs, Node built-ins only, no network / secrets / spawn) makes the method available on every prompt in Claude Code. On every other supported agent, install is still just copying files.
• Real, declared dependencies. commander, js-yaml, zod at runtime plus vitest for tests. v0.1's zero-execution / zero-dependency posture is retracted in-repo and replaced by an honest one: one small auditable hook, and a declared, Dependabot-monitored dependency graph.
• checkHookSafety gate, a mode system (lite / full / ultra / off), and an uninstall script.
• A hash-locked benchmark harness (benchmarks/) and grounding-lint, a standalone linter that exits non-zero on any uncited claim.

Gates: 25 tests pass, validator OK, npm audit 0 vulnerabilities, semgrep 237 rules / 0 findings.

D Strange v0.1.0

Stop burning tokens on trial-and-error. D Strange makes your agent lay out the real options, ground each in evidence, name what is genuinely impossible, and converge on a recommended path (plus the conditions that flip it) instead of committing to the first idea.

Three tools, one brand:

• d-strange (decide): the constraint-breaker method.
• d-strange-council: stress-test a plan or claim with deliberately opposed lenses.
• d-strange-extract: verified extraction, every claim traces to an exact quote.

Zero dependencies, zero install execution, semgrep-clean, MIT. Works across Claude Code, Codex, Cursor, Windsurf, Cline, Gemini CLI, and OpenCode.

Honest note: this is v0.1. The method is sound and all three tiers run end to end, but efficacy versus a no-skill agent is not yet benchmarked. No fabricated metrics anywhere; see SECURITY.md.