[earakely-scale]

Why. v0.9.1344's agent-build failed. The Claude Code image's pip install asked for agentenv-framework-protocol==0.1.359 at 04:48:35. That was 33 s after the protocol job uploaded it (04:48:02), but PyPI's index, which a CDN caches, still listed up to 0.1.358:

ERROR: No matching distribution found for agentenv-framework-protocol==0.1.359

So the release stopped with the protocol on PyPI and the framework not, the failure mode #288 documents. Re-running the failed jobs finished it a few minutes later, and 0.9.1344 is out, with the agent image in its record. This makes the next release wait for PyPI instead.

Change

• Retry: infra_build.py retry, which already wraps each image's docker build, now also tries the build again while pip can't find the agentenv-framework-protocol version. It uses the rate limit's waits: 30, 60 and 120 s.
• Unchanged: any other missing package still fails at once. Both build jobs keep running the same steps, which the tests require.
• Readability: _until_not_rate_limited is now _tried_again, and _why_again says why a retry can work.

Verified

• The new pattern matches the failed job's log, and the rate-limit pattern doesn't.
• The release workflow tests pass (72), including new ones for this retry and for another missing package failing at once.
• The unit tier passes, except two CLI teardown timing tests that fail only under -n on a loaded machine and pass alone.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge.

Summary

Adds retries when an agent image cannot install the just-published agentenv-framework-protocol version.

• Agent image builds retry when PyPI has not listed the protocol yet.

_{Reviews (1) · Last reviewed commit: "fix(release): build an agent image again..." · Reviewed by Greptile}