Nushell was slower on every timed task in a trial pre-registered for improvement, regression, or no difference

Ground truth was recomputed before every run of the tuned corpus, and the command that got scored was pulled from the transcript rather than from the agent's own account of what it had done ^[16]. That second choice is the one worth stealing. An agent that claims a Nushell pipeline and then quietly runs something else is a failure the author already hit: in that environment Claude Code substituted some commands with different implementations ^[9].

The stated design also reconstructs its own totals. Four tasks, two variants, 25 repetitions per variant gives the 200 microbenchmark runs ^[11][1]. Five task families, ten repetitions per family per arm, two arms gives 100 ^[15][2]. Add the 50 held-out runs and the 30 observations from the aggregate case and the headline 380 closes exactly ^[5][3]. Multiplication is the cheapest audit available on someone else's benchmark, and plenty of tables fail it.

Nushell came out slower on every timed task, and it did not always reduce context ^[12]. The place it paid was real aggregation, where a container summary fell from 3,353 bytes to 566, a saving of 2,787 bytes or about 83 percent ^[13][4]. The semantic case is separate and survives the timing loss: comparing sizes and dates as typed values carries fewer implicit assumptions than text columns you manufactured and then decided how to sort ^[2][14]. For the byte figure to transfer, your agent's task has to be an aggregation whose text form is mostly repeated structure, and the saving has to be worth a slower run. Nothing here replaces zsh; Nushell is a selective route with narrow activation rules ^[3][8]. If the task is one command that already works, the author's own policy says the structured route is unnecessary ^[7].

The wrapper is where that policy is enforced rather than described. nu-query returns JSON, imposes a timeout, and reports whether it truncated rows ^[8]. The truncation flag matters more than the JSON, because a summary that silently drops rows fails in the same direction as the text pipeline it replaced. nu-query is also deliberately absent from the auto-approved permission list, on the grounds that nu -c can modify the system ^[8].

The most useful material predates the measurement. In the test tree, ls **/* found roughly 46,400 files while ls -a **/* found about 127,900, so the glob without -a was hiding some 81,500 paths, near 64 percent ^[10][5]. The author presents this as a dated observation about that environment rather than a property of Nushell ^[10]. Any file comparison run before it surfaced was scoring two different trees.

Which is why the arm that would settle the pre-registration is the one still to read. The three outcomes were defined in terms of accuracy against cost ^[4], and the accuracy figures for the 100-run corpus are incomplete in the material available, breaking off at "rose from 23/", with no numbers given for the held-out or aggregate arms ^[18]. The author's own summary is bounded exploratory evidence, with part of the integration tuned during the process ^[6]. The reusable part is the harness: a named regression case, two of five families where the tool should be refused ^[15], ground truth recomputed per run, commands read from transcripts ^[16]. The timings belong to one file tree, in Spanish, on a Sonnet-family model ^[15][16].