Release v0.3.0 — consent-gated triggers, review fixes, plugin packaging · TheRealClodius/fram-loop

Trigger surface (minor-version change)

Consent-gated, not magic-phrase-gated. Agents may suggest fram-loop when a task fits (multi-phase deliverable, spec + plan present); any clear request to run the loop counts as the go-ahead — exact phrasing doesn't matter. A run still never starts without the user's yes. Validated with subagent trigger tests (5/5 scenarios).

Frontmatter compatibility YAML folding corruption (mid-word line breaks reached installers' parsers)
Claude Code subagent narrowing recipe: the agent-definition field is tools: (comma-separated, harness-enforced), not allowed-tools:; granular Bash(...) specifiers are unsupported there — templates.md + pre-flight.md corrected
Run #1 token totals corrected (~1.4M → ~2.1M) in EVALS.md and tradeoffs.md
Model routing reframed by capability tier (concrete IDs rot) with June-2026 examples

.claude-plugin/plugin.json — the repo now works as a Claude Code plugin (and is submitted to the community marketplace)
LICENSE.txt bundled inside the skill folder so installs carry the license
SKILL.md metadata (author / homepage / LinkedIn) + Provenance footer
README: skills.sh install path (npx skills add TheRealClodius/fram-loop), loop diagram, kickoff example, Evidence section with run #1 metrics and honest caveats