Heliocraft
← All projects

Heliocraft Agent OS

An agentic pipeline that starts from a spoken request and ends with tested, verified, committed code, with a human in control at the steps that matter.

StatusAdvanced prototype
RoleDesign and engineering
Period2026 — now

The problem

Handing a whole project to a coding agent in full autonomy doesn't work: it loses track, rewrites what worked, or reports done without doing it. Steering it prompt by prompt, on the other hand, throws away most of the gain.

I wanted a system with the discipline of a real team: a spec before code, a triaged backlog, tests, review, and a human decision before any change lands.

Architecture

Every request goes through the same circuit, driven by an explicit state contract. A failing step is retried once, then the story is marked blocked and I'm notified.

  1. Voice inputLocal wake word, on-device transcription (faster-whisper).
  2. SpecificationThe agent writes up the need and creates the matching Confluence page.
  3. Breakdown and triageThe AI governance platform generates the Jira story and rates it “ai-ready” or “needs review”.
  4. Human approvalNothing is coded without explicit confirmation, by voice or on Telegram.
  5. CodingCoding agent on a dedicated branch and worktree, with code context from semantic search.
  6. Tests, verification, securityThree checks in a row; an “OK, but…” verdict does not count as a pass.
  7. Commit and notificationLocal commit only, then a summary sent on Telegram.

Guardrails

  • Never an automatic push: merging happens on explicit command.
  • Anti-deletion guard, added after an early run removed a function while tests stayed green.
  • Sequential queue: only one story is being coded at a time.
  • A cancel command that archives, then reverts a story's work.
  • Versioned execution log, to resume cleanly after a crash.

Stack

  • Python
  • Claude Code (headless)
  • LiteLLM
  • Ollama
  • Qdrant
  • faster-whisper
  • openWakeWord
  • Telegram Bot API
  • Confluence
  • Jira
  • GitLab
  • Obsidian (state memory)

Results

  • One story shipped end to end with no intervention beyond approval, in 7 min 40.
  • 349 automated tests on the pipeline.
  • Multi-model routing: local models for simple tasks, cloud models for judgement and code.

It's still a prototype: it's my testing ground on my own projects before I offer it to teams.

Next project : AI governance platform