A frictionless "green zone" sandbox for an autonomous coding/ML agent, plus the PreToolUse policy that classifies every tool call as ungated / green / gated. `sbox` runs a command inside a bubblewrap cage: whole system read-only EXCEPT a few working dirs (~/projects, /tmp, ~/.cache), GPU passed through, NO network, ~/.ssh masked — so in-cage work is auto-allowed and only crossing a boundary (network, a write outside the zone, a broad command) trips a human approval. cage_policy.py is the transport-free classifier (green = read-only tools + one clean `sbox <argv>` + scoped-safe git + writes-in-green-zone); it is the precondition the slack-facade skill's approvals depend on. Trigger on: sandboxing an agent's shell/GPU work, deciding what should auto-run vs ask, standing up the cage on a new host, or debugging why a command was gated.
Stand up a machine-resident agent as a full participant in a Slack workspace: it converses in channels/threads and delegates real work to its machine, with approvals routed OUT-OF-BAND (Slack DM or console) so it never stalls or leaks in a channel. A two-process design — a "consciousness" (Slack-facing, empty tool schema, can't stall/leak) that delegates via an async API to a "body" (an ordinary permission-gated agent session on the host). Ships the consciousness (app.py, Agent SDK), the body relay (body_bridge.py), the Slack app manifest, and the PreToolUse/ arming hooks. Depends on the agent-cage skill for its frictionless green zone. Trigger on: putting an agent into Slack, wiring out-of-band approvals, adding a replica bot to the fleet, or debugging why a face stalled / an approval went to the wrong place. Read ../../slack-facade.md first for the design rationale.
Universal, toolchain-agnostic workflow for authoring LaTeX papers / technical writeups and compiling them to a real PDF on any machine. Detects the LaTeX engine (latexmk → pdflatex → tectonic, never assume), uses an arXiv-style single-column article template with the recurring gotchas baked in (\texorpdfstring in headings, no math in captions, pgfplots compat, hyperref last), and two figure paths (matplotlib from a reproducible venv → vector PDF, or inline TikZ/pgfplots). Self-contained, with no version-control assumptions. Trigger on: writing/editing a paper or writeup, compiling a .tex to PDF, making a figure for a paper.