anolisa

Framework-agnostic companion layer: Token-less compresses tool responses/schemas between any Agent framework and the model with no harness changes, so it can sit alongside DeepSeek Harness workflows; the README does not mention dsh by name

• Token-less tool-output compression

  tool schemas and tool responses between the Agent and the model→compressed context with field-blacklist drops; dropped array items stay retrievable through reversible `<<tokenless:KEY>>` markers; `tokenless stats summary` / `tokenless stats list` show savings

  runs as a pipeline between the Agent and the model, so it intercepts tool traffic in flightadapter configuration writes into the connected agent (e.g. Claude Code) and requires restarting it
  4624909d619e…4624909d619e…

• cosh-ng AI-native terminal

  natural language and shell commands in one terminal→an Agent that understands intent, uses tools and Skills, and asks for approval before risky work, while keeping Bash/Zsh behavior

  installed in system mode via `sudo anolisa --install-mode system install cosh-ng`
  4624909d619e…

• Agent Sec Core sandbox & Skill verification

  agent operations and signed Skills→risky operations isolated in a sandbox; changed signed Skills reported as `drifted` and blocking findings recorded as `deny` on rescan

  isolates/intercepts risky operations at runtime and records security findings for Skills
  4624909d619e…4624909d619e…

• AgentSight observability

  a running Agent on Linux→end-to-end view of user input through model and tool calls, with Token use and sub-agent branches, observed via eBPF without changing agent code

  uses eBPF kernel-level tracing on Linux to observe the agent process
  4624909d619e…