Turns your agent into a trainable employee. Learns rules from your corrections (with your approval), and skills earn autonomy through demonstrated reliability — proposed → verified → trusted → autonomous. No daemon. No cloud.
Teaches Claude Code how to use clrun for interactive terminal sessions — TUI prompts, select lists, checkboxes, confirms, long-running processes, and env persistence.
The meta plugin for the agent-plus framework. Workspace bootstrap, env-var readiness, identity cache, marketplace lifecycle (install/list/update/remove/search/prefer), extensions. One call to init creates .agent-plus/; envcheck reports which sibling-plugin env vars are set (names only); refresh resolves project.
One-call structured triage of a git diff — file-by-file role classification (source/test/config/doc/generated/build/fixture/migration), risk tier (low/medium/high) with reasons, public-API touch detection, co-changed-test detection, secret-risk path flagging, and aggregate stats. Replaces the 5–20 Read calls an agent…
The fifth universal primitive of the agent-plus framework. Mines real Claude Code session transcripts to surface repeated tool-use patterns, scaffolds them into project-scoped Claude skills (.claude/skills/ /), audits existing skills against a quality contract, and closes the loop with a cross-source feedback ag.
The trust layer for AI-assisted coding. Enforces Google Style Guides across 21 languages via MCP, with persistent project memory, fix-verify loops, and workspace heatmaps.
RoT DTD GOAL - a /goal engine where completion is earned by re-running real acceptance criteria, not declared by the model. Simultaneous verification, clock-free flake scoping, a fenced trust boundary, a dependency-ordered goal queue, and a Lean 4 specification the kernel re-verifies.
A /goal engine whose completion is EARNED, not announced. Acceptance criteria are real shell commands; the Stop hook re-runs every one of them and only exits 0 lets a goal finish. Verification defends itself: a sealed integrity ledger re-hashes each criterion before the gate trusts it, a negative-control red team…
Telegram channel for the opencode CLI — a long-running relay over opencode's headless HTTP server (opencode serve) and /event SSE stream, with access control and pairing. Unlike the codex/grok/agy MCP forks it needs no waitformessage loop, re-arm watchdog, or file-IPC permission bridge (server.heartbeat/session.idle/.
Telegram channel for the pi CLI (earendil-works/pi-coding-agent) — a long-running relay that HOSTS pi in-process via its SDK (createAgentSession), with access control, pairing, streaming edit-in-place, and a sandboxed-by-default permission gate. pi ships no permission system of its own, so every mutating tool…