Issue #1128: nikaia run: parallel runs of the same script race on the shared run/rlib cache (Bus error in ld, missing file)

Open and answer this issue on GitHub.

ready · opened 2026-11-01 · updated 2026-11-01 · 0 comments

The full suite on c60babc2 failed once with tests/compiler: agent_effort ... FAILED (the settings disagree). At user-parallelism=no, the first dry run of nikaia run scripts/agent.nika exited 1 with empty stdout. The QA agent could not reproduce it on a warm cache.

With --no-cache, it found a race. Four tests that all run nikaia run scripts/agent.nika at the same moment (agent_effort, agent_script, agent_facts, focus_rotation) failed in 1 of 3 runs. The errors were:

• ld terminated with signal 7 [Bus error]
• no such file ... tokens.txt

Probe

1. Copy those four tests into a scratch project.
2. Run nikaia test --both-settings --no-cache.
3. Repeat a few times.

Suspects

• Concurrent writes to the shared rlib cache.
• Concurrent writes to ~/.cache/nikaia/run/<name>-<key> (write_if_changed, project.rs around l.694-715). Two runs build and link the same binary into the same place without a lock or an atomic rename.

The disk may also have been close to full during one probe.

To build

• A test (several programs running the same script at once, cold cache) that fails today.
• One of the following, for both the run cache and the rlib cache:
  • build into a private temp path and rename it into place atomically, or
  • take a lock per cache key.

Back to Open Issues