Explore fable-os durability and honesty patterns for agents/automations

[ReBoticsAI]

Context

fable-os is a research QEMU kernel with Claude-driven tools. GodMode should not port bare-metal self-patching or ring-0 LLM ideas. This issue tracks steal-worthy product and reliability patterns that may fit the OSS personal-OS core (agents, automations, tools, CI).

Upstream: https://github.com/robiot/fable-os

High priority inspirations

Explore and decide what lands in core vs plugins vs follow-up Ready issues:

1. Non-forgeable tool audit trail - Host-emitted truth for tool calls and outcomes, distinct from model prose, so logs and UI cannot be spoofed by the model.
2. Automation crash / reentrancy rails - Attempting markers, consecutive-failure retirement, and nested-turn guards so unattended runs fail closed instead of looping or double-applying.
3. Mock-LLM turn-loop CI - Tests that drive a turn loop with a fake model and assert tool traces plus DB/FS side effects.
4. Discoverability that updates every turn - Skills, plugins, and capabilities refreshed in-session so agents see current host state each turn.

Medium (optional)

• Named, versioned callable capabilities and compose graphs
• Richer evidence packs for unattended automation runs
• Honest failure taxonomy (transport vs store corruption, etc.)
• Declarative-first agent UI surfaces
• Schema / tool budget as a first-class constraint

Out of scope

• Ring-0 / in-kernel LLM
• Live .text patching or self-modifying kernel code
• NL-only OS as the primary interface
• In-kernel C compiler
• Driver VM / full QEMU research stack

Acceptance

Ship a short exploration note, and/or spike PRs, and/or follow-up Ready issues for the high-priority items that fit GodMode OSS core.