TERMINUS

TERMINUS is an autonomous coding agent that turns natural-language into working software through codebase understanding, intelligent planning, code execution, testing, and iterative debugging.

Comment

TERMINUS — Where instructions become execution.

TERMINUS is an autonomous coding agent that lives inside your repository. It doesn't just write code and hope — it researches unfamiliar technology with cited web sources, maps the codebase, plans the change, asks for your approval, edits files through a deterministic permission gate, runs the project's own checks, and reports completion from observed evidence rather than model prose.

Pipeline: research → map → plan → approve → edit → run checks → (fail → diagnose → fix) → report

Inspiration

Coding agents that "just edit" are hard to trust: they clobber files, skip tests, and declare victory in prose. We wanted an agent whose completion claim is proven by real check results, with every action passing a policy gate the model cannot override.

How we built it

  • Monorepo (bun workspaces, TypeScript strict): shared, model-gateway, tool-runtime, permission-engine, context-engine, session-store, agent-runtime, research, apps/cli.
  • Model gateway: OpenAI-compatible client to Nebius Token Factory, default nvidia/nemotron-3-super-120b-a12b (256K context), streaming with tool_calls delta assembly, token usage tracked end-to-end. Anthropic fallback via --provider anthropic. Single seam selectGateway.
  • Deep research: Tavily web_research tool stores results as untrusted, cited evidence ([title](url)); search results never reach the permission engine.
  • Context engine: repository map + evidence store (hashes, provenance, trust, freshness) + stale-write guard (re-hash before edit; never clobber externally changed files). Progressive exploration only — no whole-repo preload.
  • Permission engine: deny-regex → allowlist → risk-tier → user approval, decided on the proposed action only. Session-based [a]lways, never global; local overrides tighten-only.
  • Agent runtime: state machine queued → investigating → planning → awaiting_approval → implementing → verifying → reviewing → completed, with implementing → failed_check → diagnosing → retrying self-correction.
  • Tool runtime: read_file, write_file, edit_file (exact-string replace), list_dir, search, run_command (bounded, timeout, cancellation, secret redaction), git_status, git_diff — all workspace-restricted, zod-validated.
  • Session store: local SQLite (node:sqlite/bun:sqlite), append-only audit log, --resume/--sessions.
  • CLI: responsive dashboard (suit palette), state timeline, per-event tool feed with verdicts, usage lines, approval/report cards.

Challenges

  • Getting Nemotron to emit reliable structured tool_calls on Token Factory (validated via spike).
  • Keeping research evidence untrusted so it can never authorize commands.
  • Stale-write safety under concurrent external edits.
  • Making completion a function of observed check results, not model prose.

What we learned

Model-independent harness (tools, permissions, context, verification) lets us swap models with a flag; the agent's reliability comes from the loop, not the model.

Built With

Share this project:

Updates

Submission history