Openharness Daily: Unbaking the Sandbox
A five-PR stack this morning walked back the practice of pre-baking harnesses and tools into the sandbox image, replacing it with runtime provisioning and a single home mount — fixing a broken install path and simplifying volume management along the way.
Duration: PT2M15S
Episode overview
This episode is a short developer briefing from Openharness Daily.
It explains recent repository work in plain language.
- Show: Openharness Daily
- Published: 2026-08-31T12:00:31Z
- Audio duration: PT2M15S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's August 31st, and today's activity is one connected story: the sandbox image is getting lighter, and provisioning is finally happening where it's supposed to.
The core insight, laid out clearly in PR 905: an earlier fix shipped behind a bake flag that pre-installed every default harness. That meant the new install path never actually ran — not in CI, not in a normal boot. Four defects the team had audited were sitting in code nobody was executing. PR 905 turns that bake…
That connects directly to PR 903, which fixed the real bug underneath: the command-line tool is supposed to provision the sandbox with harnesses and tools once you're inside it, but Claude Code and Codex were configured to install as root, which broke on the standard execution path. PR 907 extends the same logic to…
The second theme is mount simplification. PR 899 replaced eleven separate named volumes — one per tool — with a single mount at the sandbox home directory. Under the old model, every new agent CLI meant a new volume, a new mount line, and a new entry in the entrypoint's chown list, and anything written outside those…
Worth noting: PR 901 shows the team testing its own assumption and…
What'…