Problem

run_single_query() in skills/skill-creator/scripts/run_eval.py spawns the nested eval claude -p via subprocess.Popen(cmd, stdout=..., stderr=..., cwd=..., env=...) without redirecting stdin. The child inherits the parent's stdin file descriptor.

When run_eval.py/run_loop.py itself runs nested inside a claude session (a supported, intended use — the file explicitly strips CLAUDECODE from env to allow this), the child process can block briefly waiting on/polling the inherited stdin before it gives up and proceeds, adding an observed ~3s stall per query. At --num-workers > 1 this multiplies across concurrent subprocesses and inflates wall-clock eval time without any corresponding signal.

Repro

Run any eval from inside a claude -p or Claude Code session:

python -m scripts.run_loop --eval-set eval.json --skill-path <skill> --model <m>

Observe each run_single_query call taking ~3s longer than the same call run from a plain (non-nested) shell.

Fix

Single-line change — explicitly discard the child's stdin instead of inheriting it:

--- a/skills/skill-creator/scripts/run_eval.py
+++ b/skills/skill-creator/scripts/run_eval.py
@@ -84,6 +84,7 @@ def run_single_query(
 
         process = subprocess.Popen(
             cmd,
+            stdin=subprocess.DEVNULL,  # nested claude -p must not wait on/inherit our stdin
             stdout=subprocess.PIPE,
             stderr=subprocess.DEVNULL,
             cwd=project_root,

Note for maintainers

Neither open PR #1323 nor #1298 (both reworking this same function for the 0%-recall bug) touches the Popen call — this fix is orthogonal and should be folded into whichever lands, or applied standalone if merged first.

Disclosure: this issue was researched and written by Claude (Claude Code)