Autonomous agentic harness for residual (24/7 daemon, ACP interface). Working repo; destined for git.merith.xyz/residual/agent.
- Go 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| cmd/agent | ||
| internal | ||
| testdata | ||
| .gitignore | ||
| DESIGN.md | ||
| go.mod | ||
| go.sum | ||
| README.md | ||
agent
Autonomous agentic harness for residual — a 24/7 daemon that does work on its own, built to residual's design constraints.
- Runs on device (e.g. Raspberry Pi with a screen); supervised by
res-initwith respawn for uninterrupted operation. - Pluggable LLM backends: a direct API (OpenAI-compatible or Anthropic)
or an external ACP harness it drives as a client (
opencode acp,claude-code-acp, …). It does not serve ACP itself. - Workspaces & identities: each agent gets
AGENT_DATA/workspace/<name>as its home; persona and config live inAGENT_DATA/identities/(agents/andsubagents/, md+yml pairs). - mcp-sync: MCP servers declared once in agent config are translated into harness-native workspace configs (opencode officially supported).
- Long-term memory in sqlite, exposed through built-in MCP tooling so agents persist efficiently.
- The only external dependency is an LLM provider.
See DESIGN.md for the architecture.
Layout
cmd/agent/ entrypoint
internal/
acp/ DEAD CODE (intentional): ACP agent-side, kept for reference
acpclient/ ACP client: spawn + drive external harnesses
workspace/ AGENT_DATA layout, identity (md+yml) loading
mcpsync/ canonical config → harness workspace configs
memory/ (next) sqlite long-term memory
Status
See DESIGN.md § "Milestones".