v1.1.0 - Depth Modes, Hive Voting, and Per-IDE Onboarding

First tagged release of the slim UniGrok gateway. 1.0.0 was the untagged initial public candidate (dual-plane Grok harness — Build CLI subscription + metered xAI API — workspace-neutral courier model, sessions and durable facts in SQLite, media/search/code specialists, benchmark telemetry); 1.1.0 adds everything below. Releases before 1.0.0 refer to the retired multi-provider platform, not the current gateway.

Added

• GitHub Copilot client onboarding: gh Copilot CLI ~/.copilot/mcp-config.json
  merge entry (repo-level .copilot/mcp-config.json at project scope, .vscode/mcp.json
  alternative for VS Code), a namespaced .github/instructions/unigrok.instructions.md
  routing rule, and --allow-tool 'grok(agent)' session flags for auto-approve.
• Per-IDE "never prompt for @grok" auto-approve in the onboarding plan, each via its
  native mechanism: Claude Code permissions.allow (per-tool mcp__grok__agent), Codex
  config.toml MCP tool approval_mode = "auto", Gemini/Antigravity server trust: true,
  Cursor the beforeMCPExecution hook. GitHub Copilot/generic get none (no verified format).
• Cursor client onboarding: grok_mcp_onboard_client now emits a .cursor/mcp.json merge entry (points Cursor at the Grok gateway with X-Client-ID: cursor, carries no credentials), a .cursor/rules/using-unigrok.mdc routing rule, and a .cursor/hooks.json + before-unigrok-agent.py beforeMCPExecution hook that auto-approves ONLY the agent tool so @grok never stalls on a permission prompt. Cursor is a client, not an execution plane — ported from the old public version's static .cursor/ setup.
• Depth modes deep (cached j-space harness prompt, harness-leak guard, final polish loop) and hive (draft → parallel persona votes with numbered-line dif-vote anchors → always-on xhigh merge; voters split across CLI/API planes; per-stage plane+cost receipts)
• Public level ladder none→ultra via level parameter and voters override
• Auto++ router: three flat-rate parallel intent votes (route/depth/voter-count) replacing the metered routing pass on unclear tasks, with dynamic voter sizing
• Agent job persistence to SQLite (results survive restarts; interrupted jobs return status "lost")
• Verifying benchmark suite: benchmark_deep.py (executed-code checks, level sweeps, --voters sweep), persona_bench.py tournament, parallel_probe.py, triage_optimize.py scout
• Dogfood optimizer loop (dogfood_optimize.py) with anti-Goodhart counter-metric gate (no new imports, bounded diff) and 8% noise floor
• Six gateway functions hive-optimized with measured wins (+52.8%, +22.6%, +20.8%, +18.9%, +18.0% ×2)
• Shadow done-vote experiment flag UNIGROK_SHADOW_DONE_VOTE (off by default)
• Metered-API output cap on micro-emit votes (UNIGROK_VOTE_MAX_OUTPUT, default 128)

Changed

• xhigh effort auto-downgrades to high on the API plane instead of erroring
• Requested/resolved depth and level echoed in payloads and telemetry

Fixed

• CLI-plane tool-permission cancellations (grok CLI 0.2.101 lowercase tool IDs; ACP session/request_permission reject handler; one tool-less retry in-session) that silently converted subscription work into metered API spend via cross-plane fallback