v1.1.0

[1.1.0] - 2026-07-05

Changed

• fast-worker — now opens with a machine-readable RESULT: DONE/PARTIAL/BLOCKED line, pastes real verification output, labels unverified claims [inferred]/[unverified], and states explicit default boundaries (no commit/push/PR unless ordered).
• deep-reasoner — now opens with RECOMMENDATION: <one sentence> — confidence: HIGH/MEDIUM/LOW, labels evidence tri-state, and makes its recommend-only / no-side-effects role explicit. Kept on all tools; a local tools: allowlist (e.g. ["Read","Grep","Glob","Bash","WebSearch","WebFetch"]) is a documented opt-in, not the default, because static allowlists block per-machine MCP tools.
• Re-running the installer on a machine with v1.0.0 copies reports [conflict] for these two agents — diff and take the bundled version.

Fixed

• install.sh — agent definitions now install globally into $CLAUDE_CONFIG_DIR/agents (default ~/.claude/agents) instead of per-project (fix landed on main after v1.0.0; first released here).