Autonomous repo guardian. A multi-agent crew, powered by Xiaomi MiMo V2.5, that nightly hunts bugs, drafts minimal patches, verifies them in a sandbox, and opens PRs with audio + visual reviews.
Most "AI bug-fixer" tools stop at suggesting a diff. orbiter goes the last mile: it closes the loop end-to-end, in five sequential agents, all coordinated under one orchestrator with persistent run-state and deterministic recovery.
+-----------+
input repo ---> | Detective | --> findings ---+
+-----------+ |
v
+-------------+
| Reproducer | --> regression tests +
+-------------+ |
v
+------------------+ verdict +--------------+
| Surgeon | <-------+ | Verifier |
| (max N rounds) | | | (sandboxed) |
+------------------+ | +--------------+
| passed patches |
v |
+--------------+ |
| Reviewer | -----------> artefacts
+--------------+ (PR md + PNG + audio)
| Agent | Role | MiMo capability |
|---|---|---|
| Detective | Scan source, surface high-confidence defects | reasoner (JSON-strict) |
| Reproducer | Author a pytest case that proves the bug | reasoner |
| Surgeon | Draft the smallest correct patch | reasoner |
| Verifier | Apply patch in sandbox, run repro and full suite | – (pure execution) |
| Reviewer | Compose PR title + body, render diagram, narrate audio | reasoner + multimodal + TTS |
The Surgeon ↔ Verifier loop is the closed-loop spine: a rejected patch
hands the failure transcript back to the Surgeon as additional context,
up to ORBITER_MAX_PATCH_ITERATIONS retries (default 3) before the
finding is dropped.
Every fix that ships green produces a 60–90 second narrated walkthrough attached to the PR. This is the unique angle MiMo TTS makes possible: async teams listen to PR reviews on a commute or during deep-work breaks instead of staring at another GitHub tab.
git clone https://github.com/suisenn1/orbiter.git && cd orbiter
python3.11 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
# Sanity check (works without an API key)
orbiter doctor
# Run the offline demo against the planted-bug fixture
orbiter run --mock examples/buggy_repo --out outYou'll see the table:
orbiter run · 20260507T2110-6280
┏━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ phase ┃ count ┃ notes ┃
┡━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ detective │ 2 │ high-confidence findings │
│ reproducer │ 2 │ regression tests authored │
│ surgeon │ 2 │ patches accepted │
│ verifier │ 2 │ patches verified green │
│ reviewer │ 2 │ PR artefacts emitted │
│ tokens │ 0 │ billed against MiMo budget │
│ phase │ done │ │
└────────────┴───────┴────────────────────────────────┘
…and out/<run_id>/ will contain real artefacts:
out/20260507T2110-6280/
├── PR_F01-timing-attack.md ← markdown PR body
├── F01-timing-attack.png ← root-cause diagram (1280×720)
├── F01-timing-attack.wav ← 60–90s audio walkthrough
├── PR_F02-bare-except.md
├── F02-bare-except.png
└── F02-bare-except.wav
cp .env.example .env
# fill in ORBITER_MIMO_API_KEY from https://platform.xiaomimimo.com
orbiter run examples/buggy_repo --out outIn live mode every agent calls MiMo: reasoner for analysis, multimodal for diagrams, TTS for narration. Mock-mode artefacts are byte-for-byte identical in structure to live ones, so reviewers and CI can lean on the deterministic offline brain.
# crash mid-run? pick up exactly where it stopped:
orbiter recover
# or target a specific run id:
orbiter recover 20260507T2110-6280| Command | Description |
|---|---|
orbiter run <repo> |
Start a fresh pipeline against <repo>. |
orbiter recover [<id>] |
Resume the latest unfinished (or a specific) run. |
orbiter list |
Tabulate every persisted run. |
orbiter show <id> |
Pretty-print a run's findings + token usage. |
orbiter doctor |
Validate env, deps, and credentials. |
All commands support -v / --verbose and --out <dir>.
Every knob lives in environment variables (ORBITER_*) or .env. See
.env.example for the full set; the highlights:
| Variable | Default | Purpose |
|---|---|---|
ORBITER_MIMO_API_KEY |
– | Empty → mock mode only. |
ORBITER_MIMO_BASE_URL |
https://api.xiaomimimo.com/v1 |
Override for self-hosted MiMo. |
ORBITER_MAX_TOKENS_PER_RUN |
2_000_000 |
Hard budget cap (raises TokenBudgetExceeded). |
ORBITER_MAX_PATCH_ITERATIONS |
3 |
Surgeon ↔ Verifier retry rounds. |
ORBITER_SANDBOX_TIMEOUT_SECONDS |
120 |
Max wall-clock per pytest run. |
- Typed contracts everywhere. Every payload that crosses an agent
boundary is a Pydantic v2 model; every LLM call uses
response_format=json_schemawithstrict=Trueand a one-shot repair pass onValidationError. - Persistent run-state. After every phase, the orchestrator writes
<run_id>.jsonatomically (tmp + rename) to~/.orbiter/. A crash never costs you tokens already spent. - Closed-loop verification. The Verifier snapshots the working tree to a sandbox, applies the Surgeon's patch, drops in the Reproducer's regression test, runs both the new test and the project's existing suite, and only marks the finding GREEN if everything passes.
- Patch hygiene. Surgeon emits explicit
(original, replacement)hunks (not free-form unified diffs); the orchestrator refuses to apply any hunk whoseoriginaldoesn't appear exactly once in the file. - Mock mode = first-class citizen. Each agent ships a deterministic offline brain so the CI suite, the demo, and grant reviewers never need a credential.
See docs/architecture.md for the deep dive.
- GitHub App: open the PR upstream automatically
- Language packs beyond Python (TypeScript first)
- Severity-weighted token budgeting
- Slack / webhook notifier for green PRs
- Self-hosted MiMo deployment guide
orbiter is a finalist for the Xiaomi MiMo Orbit 100T Token Creator Incentive Program. The reasoner-first architecture is a love letter to MiMo V2.5's three-pillar design: reasoning, multimodal, TTS behind one OpenAI-compatible endpoint.
MIT — see LICENSE.