Pi agent harness: CLI, AI SDK, Pi Durable vs kill -9
I ran the Pi agent harness three ways (pi --print, the AI SDK's HarnessAgent, Cloudflare's PiHarness) and killed each mid-task. What came back, and why.
orbi.build · author
Lawrence Liu
Maintains Orbi: the AI coding agent that turns labelled GitHub Issues into independently reviewed, merged PRs and tagged releases. Every post on this blog is written by him.
Pi + DeepSeek Flash, 169 merged PRs: 8–16¢ each
In model tokens at DeepSeek off-peak prices, 169 merged deliveries cost a median of $0.082 each. One issue through four Pi sessions; what Pi leaves out.
Run Claude Code unattended: headless mode is step one
Claude Code headless mode (claude -p) runs one session unattended, logged in with an API key or setup-token. Picking work, review and merge are on you.
Searching for Orbi's harness, part 2: guard cost
On eight unseen bugs whose Issue shows only a narrow example, Orbi's original harness failed a third of its runs. Five designs to cut the cost of the fix.
Searching for Orbi's harness, part 1: review gaps
Orbi's review passed two fixes that each broke something. We rebuilt the harness one rule at a time and reran the loop on failed and unseen issues.
GitSpawn and the agent that never asks
A git flaw hit seven coding agents by skipping their approval prompts. Orbi has none to skip, so we shipped a fix, checked who runs git, and reverted it.
DeepSeek coding agent cost per merged PR
See how Orbi uses DeepSeek as an unattended coding agent, what each merged PR costs, why cache reads matter, and how to connect your own API key.
Why Orbi is open source again
Orbi began on Apache-2.0 and switched to the Sustainable Use License on Sep 4. From v0.5.44 it is AGPL-3.0, or the SUL if AGPL is ruled out. Here is why.
The k8e etcd ticket that closed too early
Orbi wrote an etcd test suite for k8e, a 497-star Kubernetes distro. It was rejected, fixed and merged, and the four-phase issue closed after phase one.
How one slash command got its shape
A close look at the name, syntax, regex anchor, and registry behind one in-ticket command, compared with Prow and bors-ng and checked against their source.
The engine waited for an answer users could not give
Our engine paused for a human decision, then asked tenants to use a command they did not have. Three versions of the same mistake in one afternoon.
Seven steps of an Orbi Cloud setup, on video
An 83-second video follows the Orbi Cloud setup from GitHub sign-in to a tagged release and subscription. The final step starts paid Cloud delivery here.
What an autonomous coding agent does when it says no
Most autonomous coding agents stop at a pull request. Here, a second model rejected the diff twice, while the commits and release remain inspectable.
Claude Code in Actions: who presses merge?
Claude Code Action can implement and review pull requests in your runner. It stops before merge and release. Here is the evidence-backed layer for both.
The Docker image made it out on try three
The orbi image reached GHCR and Docker Hub on its third run. The first two failures were build context and Dockerfile paths; the last an invalid tag.