惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

M
MIT News - Artificial intelligence
罗磊的独立博客
Hugging Face - Blog
Hugging Face - Blog
Apple Machine Learning Research
Apple Machine Learning Research
Last Week in AI
Last Week in AI
S
SegmentFault 最新的问题
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
美团技术团队
人人都是产品经理
人人都是产品经理
WordPress大学
WordPress大学
The Cloudflare Blog
IT之家
IT之家
雷峰网
雷峰网
小众软件
小众软件
博客园 - 叶小钗
博客园 - 聂微东
爱范儿
爱范儿
博客园 - 司徒正美
博客园 - 三生石上(FineUI控件)
V
Visual Studio Blog
博客园 - 【当耐特】
V
V2EX
博客园_首页
T
Tailwind CSS Blog

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders
GitHub - human-again/orbit: Mission control for AI coding...
varmahaj · 2026-05-19 · via Hacker News - Newest: "AI"

CI Python 3.11+ License: MIT v0.1.0

Mission control for AI coding agents.

Orbit is the harness your coding agent actually needs — structured loops, real validation gates, rubric-based evaluation, checkpoint resumability, and a full audit trail. Define the mission. Orbit lands it.

Orbit rescuing a broken auth flow

↑ Orbit detects a failing test, selects the right task, runs the agent, validates the fix, and marks the orbit complete — no hand-holding required.


The problem with AI coding agents

They drift. They hallucinate changes. They skip validation. They leave no record of what they did or why. You give them a task and cross your fingers.

You:   "Fix the auth bug."
Agent: ✅ Done!
You:   "Did you run the tests?"
Agent: ✅ All passing!
You:   *runs tests* ❌❌❌

Sound familiar?


Orbit fixes that

Every agent run becomes a structured orbit — a bounded, validated, auditable loop. If the agent can't prove its work, the orbit doesn't close.

The Orbit loop

↑ The Orbit loop: task selection → agent execution → validation gate → rubric evaluation → review → next orbit.

What you get:

  • 🎯 Mission files — human-owned goal, scope, constraints, definition of done
  • 📋 Dependency-aware task queue — priority-sorted, one task per orbit, respects your DAG
  • 🚦 Validation gate — real pytest/lint/typecheck runs; nothing advances without evidence
  • 📐 Rubric evaluator — structured scoring on task focus, completion, change signal, and validation
  • 🔁 Retry policy — auto-retry on failure or validation miss, configurable limits
  • 💾 Checkpoint resumability — durable state; pick up mid-mission without losing progress
  • 💰 Budget management — cap runs, failure count, and estimated cost per session
  • 📡 Full telemetry — every orbit writes agent-result.json, evaluation.json, review.json, and a timestamped progress log

Works with any agent

Orbit is adapter-based. Swap in whatever coding agent you have:

Adapter Agent
adapters.claude_cli:ClaudeCliAdapter Claude Code (claude --print)
adapters.codex_cli:CodexCliAdapter OpenAI Codex CLI
adapters.cursor_cli:CursorCliAdapter Cursor
adapters.cli_json:CliJsonAdapter Any agent that reads stdin and writes JSON
adapters.mock:MockAgentAdapter Local testing / dry runs

What can you do with Orbit?

Use case How Orbit helps
Nightly bug triage Point Orbit at your failing CI jobs. It selects the highest-priority fix, validates the patch, and leaves a full artifact trail for review in the morning.
Incremental feature backlog Break a feature into tasks in backlog.json. Orbit works through them in dependency order, one per orbit, never skipping validation.
Automated code health Schedule Orbit to run overnight on your tech-debt backlog. Cap spend with max_estimated_cost_usd. Wake up to a progress log with diffs and scores.
Self-healing repos Wire Orbit into CI as a bot that auto-fixes known flaky tests, restores broken migrations, or patches lint failures — with full review artifacts before merge.
Dependency upgrades Model each version bump as a task. Orbit upgrades them in order, runs your full test suite after each one, and halts if validation breaks.
Multi-agent handoffs Run one orbit with Claude for implementation, a second with Codex for review, and a third with your custom CLI for deployment checks — same harness, swappable adapters.

Quickstart

git clone https://github.com/human-again/orbit
cd orbit
python -m venv .venv && .venv/bin/pip install pytest pillow

Drop three files in your project:

mission.md        ← what to build, and done means done
agent-rules.md    ← safety rules, coding standards, escalation policy
backlog.json      ← tasks with acceptance criteria and dependencies

Then launch:

python orchestrator.py --config runtime/config.json

Orbit selects the next task, hands it to your agent, validates the output, evaluates the result, and loops — until the mission is complete or it needs you.


See it in action

Demo 1 — Auth Rescue

A real auth bug is introduced into the codebase. Orbit selects the right task, runs the agent, turns red tests green, and closes the orbit with a full evaluation artifact.

Auth rescue demo

MOCK=1 ./replay.sh auth-rescue

Demo 2 — Issue Search

A new feature (scoped search) is added to a tiny issues dashboard. Orbit validates the feature with tests and a browser preview before marking it complete.

Issue search demo

MOCK=1 ./replay.sh issue-search

Both demos use MOCK=1 — fully deterministic, no API key needed, runs in CI. Swap in a real adapter (DEMO_ADAPTER=claude or DEMO_ADAPTER=codex) to use your agent.


How it works

┌─────────────────────────────────────────────────────────┐
│                      Orbit Loop                         │
│                                                         │
│  mission.md + backlog.json                              │
│        │                                                │
│        ▼                                                │
│  choose_task()  ──── dependency-aware, priority-sorted  │
│        │                                                │
│        ▼                                                │
│  build_task_bundle()  ── mission + rules + context      │
│        │                                                │
│        ▼                                                │
│  adapter.run_agent()  ── your agent, any CLI            │
│        │                                                │
│        ▼                                                │
│  run_validation()  ── real tests, lint, typecheck       │
│        │                                                │
│        ├── FAIL ──► retry_policy ──► loop or stop       │
│        │                                                │
│        ▼                                                │
│  evaluate()  ── rubric scoring → verdict                │
│        │                                                │
│        ▼                                                │
│  review()  ── accept / iterate                          │
│        │                                                │
│        ▼                                                │
│  update backlog + progress + telemetry                  │
│        │                                                │
│        └────────────────────────────► next orbit        │
└─────────────────────────────────────────────────────────┘

Configuration

{
  "adapter": "adapters.claude_cli:ClaudeCliAdapter",
  "adapter_args": {
    "command": "claude",
    "args": ["--print", "--output-format", "json", "--dangerously-skip-permissions"]
  },
  "validation_commands": ["pytest tests/ -v"],
  "max_iterations": 10,
  "run_evaluator": true,
  "run_reviewer": true,
  "require_validation": true,
  "retry_policy": {
    "max_attempts": 2,
    "retry_on_statuses": ["failed"],
    "retry_on_validation_failure": true
  },
  "budgets": {
    "max_runs": 50,
    "max_failures": 5,
    "max_estimated_cost_usd": 10.0
  }
}

Adapter contract

Two methods. That's it.

class MyAdapter(BaseAgentAdapter):
    def prepare_prompt(self, task_bundle: str) -> str: ...
    def run_agent(self, prompt: str, cwd: str, timeout_s: int) -> AgentResult: ...

AgentResult fields: status, changed_files, notes, metadata, raw_output.


Repo layout

orchestrator.py        ← mission control — the main loop
adapters/              ← pluggable agent adapters
evaluator.py           ← rubric scoring
reviewer.py            ← accept/iterate recommendation
validation_runner.py   ← runs test/lint commands
retry_policy.py        ← retry on failure or validation miss
checkpoint_manager.py  ← durable resume state
budget_manager.py      ← run/failure/cost caps
diff_evaluator.py      ← git diff detection
risk_guard.py          ← command classification and gating
observability.py       ← structured telemetry log (events.jsonl)
mission.md             ← your goal, scope, definition of done
agent-rules.md         ← coding rules, escalation rules
backlog.json           ← tasks with acceptance criteria
progress.md            ← append-only mission log
runtime/               ← configs, run artifacts, checkpoints
research/              ← bounded research artifacts per task
examples/              ← end-to-end demos
docs/media/            ← demo GIFs

Advanced

Codex self-heal demo — Orbit uses its own Codex adapter to fix a real Orbit regression inside a temp copy of the repo, then records validation, evaluation, review, and backlog evidence for that orbit.

.venv/bin/python examples/codex-self-heal-demo/run_demo.py

Regenerate the demo GIFs:

scripts/make-demo-media.sh auth-rescue
scripts/make-demo-media.sh issue-search
scripts/make-demo-media.sh orchestrator-pseudocode

⭐ Star it

If Orbit has saved you from an agent that said "tests passing" when they weren't — star it. It helps others find it.

Star on GitHub


Credits

The agent loop design draws inspiration from Andrej Karpathy's thinking on agentic systems and long-running LLM workflows.


License

MIT