default-bash-compress

Turns on RTK's default pipeline with 1 built-in rule changed. Adds nothing else.

aeppling 0 Cuts cost Reviewed by RTK

Result

Total cost

−6.2%

per trial · quality about the same

Author's benchmark Claude Opus 4.8 · 48 trials per arm

You need

Nothing extra

Your data

Stays on your machine

Your agent

Not instructed

Install pinned to the reviewed version

rtk pipeline add github:aeppling/rtk-pipeline-default-bash-compress@ef1b85b59cb670b70a40a7cb0dbb8e699bb6ba9f

Try it first without changing anything

rtk pipeline add github:aeppling/rtk-pipeline-default-bash-compress@ef1b85b59cb670b70a40a7cb0dbb8e699bb6ba9f --shadow

Records what it would have done. Then rtk pipeline eval default-bash-compress shows the numbers; promote or --discard.

Benchmark 12 tasks · SkillsBench

Mean cost per trial by task, without and with the pipeline
Task	stock Claude Code	default-bash-compress	Change	reward
parallel-tfidf-search	$1.37	$0.92	−32.5%	1 → 1
react-performance-debugging	$1.22	$0.97	−20.7%	0.5 → 0
python-scala-translation	$1.22	$1.01	−17.5%	0 → 0
dialogue-parser	$0.37	$0.32	−12.2%	0.75 → 0.79
fix-build-google-auto	$2.11	$1.91	−9.6%	1 → 1
jax-computing-basics	$0.28	$0.27	−4.6%	1 → 1
spring-boot-jakarta-migration	$0.91	$0.86	−4.9%	—
debug-trl-grpo	$0.59	$0.63	+6.4%	0.42 → 0.25
simpo-code-reproduction	$0.97	$1.04	+7.1%	0 → 0
llm-prefix-cache-replay	$0.63	$0.68	+7%	0 → 0
azure-bgp-oscillation-route-leak	$0.49	$0.54	+10.8%	0 → 0
fix-visual-stability	$1.92	$2.19	+13.8%	0.75 → 1
All tasks, mean per trial	$12.10	$11.35	−6.2%	0.49 → 0.46

SkillsBench · Claude Code on Claude Opus 4.8 · 48 trials per arm · Aug 10, 2026. Declared by the author in benchmark.toml, not reproduced by RTK · Write-up · Per-trial data

Author's note: 12 tasks, 4 trials per arm per task, both arms in the same batch, RTK v0.45.0 PreToolUse hook only; runtime-error trials excluded; data-to-d3 ran too but is left out of the aggregates (no verifier score, unbalanced arms). Pooled over all trials with a permutation test: -4.8%, p = 0.305, i.e. cost-neutral to slightly cheaper. Pass rate is the mean reward over the 11 tasks whose verifier produced a score; spring-boot-jakarta-migration has none.

How it works no rules of its own · turns on RTK's default pipeline

Based on RTK's default pipeline (rtk:default).

Adds no rules of its own.

Switches the built-in rule normalize from rtk:default on or off.

pipeline.toml 46 lines

Exact file at commit ef1b85b, reviewed Oct 25, 2026. Comments are the author's own; numbers in them are self-reported. View on GitHub

# default-bash-compress — RTK's original behaviour (pre-1.0): rewrite shell commands to their rtk
# equivalent before they run, so the agent reads the filtered output. Nothing else.
#
# What it does
# - PreToolUse hook: `git status` -> `rtk git status`, `cargo test` -> `rtk cargo test`, and so on
#   for every command RTK has a filter for; commands without a filter run untouched.
# - The filters drop noise the agent never reads (progress bars, ANSI codes, repeated lines,
#   boilerplate) and keep every line that carries signal: errors, test names, paths, diffs.
# - No PostToolUse compression, no delegation, no routing. The output you get is what
#   `rtk <cmd>` prints in your own terminal.
#
# Install needs nothing but rtk itself (`rtk init -g` sets up the hook).
#
# Benchmark — SkillsBench, Claude Code on claude-opus-4-8, 12 tasks, 4 trials per arm,
# 2026-08-10, OFF (stock Claude Code) in the same batch, RTK v0.45.0 hook. Mean cost per
# trial, runtime-error trials excluded (see RESULTS.md):
#
#   task                                OFF     RTK  change
#   parallel-tfidf-search             $1.37   $0.92   -32%
#   react-performance-debugging       $1.22   $0.97   -21%
#   python-scala-translation          $1.22   $1.01   -17%
#   dialogue-parser                   $0.37   $0.32   -12%
#   fix-build-google-auto             $2.11   $1.91   -10%
#   jax-computing-basics              $0.28   $0.27    -5%
#   spring-boot-jakarta-migration     $0.91   $0.86    -5%
#   debug-trl-grpo                    $0.60   $0.63    +6%
#   simpo-code-reproduction           $0.98   $1.04    +7%
#   llm-prefix-cache-replay           $0.63   $0.68    +7%
#   azure-bgp-oscillation-route-leak  $0.49   $0.54   +11%
#   fix-visual-stability              $1.92   $2.19   +14%
#   total                            $12.10  $11.35  -6.2%
#
# Pooled over all trials: -4.8% (permutation p = 0.305), 7 tasks cheaper / 5 costlier:
# cost-neutral to slightly cheaper. Pass rate 0.49 -> 0.46 over the 11 scored tasks. RTK removed
# about 99 KB (~25K tokens) from 154 rewritten commands in the run. Per-trial rows: results.tsv.
schema_version = 1

[pipeline]
name    = "default-bash-compress"
extends = "rtk:default"

# rtk:default also normalizes tool results through the proxy daemon when one runs.
# Off here: this pipeline is the hook alone.
[override.normalize]
enabled = false