feat: persistent Python console — working memory outside the context

[KevRojo]

What

A new Python tool: a persistent REPL whose namespace survives across calls. Scan a large structure once into a variable and it stays alive in the kernel's heap — filter, count, and aggregate it across turns while the model only ever prints the small slice it needs.

Why

When an agent explores something big (a directory tree, a parsed file, an API dump), the result normally lands in the conversation and gets re-sent every turn for the rest of the session — you end up paying to re-read the search forever. Keeping the data in a live kernel heap instead of the context means only small slices ever cost tokens. Working memory outside the context window.

Example: 101,619 files scanned into a variable, then queried three ways across separate calls — the context only ever saw a number, an 8-line slice, and a one-line summary.

How it's built

• Isolated subprocess kernel — pyconsole.py doubles as the worker via --pykernel-worker; an infinite loop or crash kills the kernel, never the agent. The parent enforces a wall-clock timeout and auto-restarts a dead/killed kernel on the next call.
• Output capped at the source by characters (not just lines) with a bounded reprlib echo, so one enormous line or a huge repr() can't flood the context and get re-billed every turn.
• input() is neutralised (clean EOFError, never hangs the protocol), tracebacks are trimmed to the caller's frames, and a trailing bare expression echoes like a REPL.
• Self-registers through the tools/ auto-import — adds "pyconsole" to the tuple, matching the existing files/browser/email submodules. The cheetahclaws.tool_registry import stays lazy so the worker subprocess never drags in the package.

Verified

Registration through tool_registry, state persistence across separate execute_tool calls, REPL echo, giant-line / huge-repr caps, input() no-hang, and timeout-kills-a-runaway-loop + auto-restart. Pyright-clean.

Scope: two files — new cheetahclaws/tools/pyconsole.py, one line in cheetahclaws/tools/__init__.py.