deepwiki

Discover, catalog, and rank everything you've built or sketched on a machine via a two-pass parallel read-only sweep: a clustered project catalog plus blind-spot agents (transcripts, other agent CLIs, running services, research artifacts, beyond-home), synthesized into a tiered novelty/leverage ranking with idea/live/dormant tags and evidence-anchored rationale.

Stop suggesting stale package versions from training data. 15 MCP tools over 39 package registries (npm, PyPI, Maven, crates.io, Go, RubyGems, NuGet, Packagist, Docker/GHCR/Quay/GCR, Hex, pub.dev, CRAN, Homebrew, and more): latest-version + package-info lookups, semver compare, install-command generation, batch checks, project dependency scanning, outdated reports, conflict detection, compatibility checks, and a scan -> optimize -> validate -> apply upgrade pipeline that edits manifests with dry-run + automatic backups by default. Local-first, no API key required (an optional GITHUB_TOKEN raises GitHub rate limits).

Seamless context continuity across Claude Code compaction. Four background layers (continuous tool-call capture, mechanical PostCompact handoff, async Opus 4.7 enrichment at --effort max, async Stop-hook refinement, and preemptive snapshot before the next compact) keep the agent's working state intact across every compaction and resume. All summarization is isolated from CLAUDE.md/auto-memory and invisible to the user.

Recover Claude Code sessions after an Anthropic API output-classifier false positive (the "cyber-related safeguards" / "appears to violate our Usage Policy" hard failure). A JSON-aware .jsonl log surgeon that surgically removes the triggering message + refusal turn and re-stitches the parentUuid chain so the session resumes cleanly, cleans accumulated refusals across old sessions, and installs a per-repo CLAUDE.md framing guard to stop recurrence.

A ground-first reflex: the agent judges (from its own training, any domain) when a request is complex/irreversible, grounds before acting, and is hard-blocked from the destructive step until it has.

Eyes and hands on a Linux Wayland desktop: screenshot any monitor and click, type, scroll, drag, and read any visible app over xdg-desktop-portal (RemoteDesktop + ScreenCast), with optional OCR + OmniParser icon grounding. Pure-Python, CPU-only. GNOME/Wayland only. Ships the MCP server plus a drive-screen skill that encodes the locate-ground-act-confirm loop.

The sanctioned OS 'motor cortex' for an agent on a Linux box: control systemd services/timers, query journald, read host resources/processes, send desktop notifications, drive D-Bus, and manage power (suspend/reboot/poweroff) — all through structured interfaces (systemctl/loginctl/journalctl/busctl), never raw PID hacks. Pure standard library, zero pip deps. A self-preservation guard refuses severing actions on units the agent depends on (dbus, logind, sshd, network, tailscaled, the session, goosed) unless force=true. Ships the MCP server plus a control-os skill. The system-service counterpart to screen-mcp's GUI control.

Operate and observe an interactive full-screen TUI on a REMOTE machine over tmux/screen + SSH by driving it like a human: type with send-keys, screenshot with capture-pane, in a type-wait-screenshot-read loop. For the Claude Code TUI, vim, top, curses installers, REPLs, or any program that needs a real PTY over SSH.

Simplify, then add lightness. Removes AI-generated code fat while preserving functionality, then benchmarks the speedup. Multi-pass: measure weight -> trim -> verify equivalence -> benchmark.

Falsification-first investigation workflow: convert every assertion into a labeled falsifiable hypothesis, predict before measuring, run controlled experiments, verify findings adversarially (REFUTE-first), and persist verdicts in a hypothesis ledger so killed ideas are never re-attacked