Sick is a self-improving, open-source coding agent (published on PyPI as
sick-agent) built on the NVIDIA OO Agents Framework. It reads, writes,
and edits code in your workspace — and unlike other agents, its host
process enforces the engineering process instead of prompting for it.

Every task runs an evidence-gated SDLC lifecycle (requirements →
design → implement → verify → deliver): phases are observed from real
tool results, and a run with edits but no passing verification after
the last edit ends BLOCKED — never “success”. Every run persists
auditable artifacts (.sick/runs//: plan, verification.json,
summary), so any claim of “done” can be inspected after the fact.

Sick also learns in a verifiable way. Failures become structured
lessons (trigger → root cause → fix) that are promoted only when a fix
actually verifies, and marked ineffective when a failure recurs.
Successful tasks can be distilled into reusable SKILL.md playbooks
(/distill), and when a capability is missing the agent grows its own
tool: it writes a Tool class, smoke-tests it in a subprocess,
registers it live, and keeps it for future sessions (.sick/tools/).

It ships the real Ponytail and Aftermath plugins (vendored, pinned):
destructive commands are blocked and forward edits freeze after
unresolved failures, enforced in code. A repo guardian mode
(sick watch) re-verifies your repo on change and makes one bounded,
checkpointed fix attempt when tests go red.

The TUI streams everything live — model thinking, every tool call,
the SDLC phase tracker and gate state — across five switchable model
providers, switchable mid-session with /connect.