rtk was wired as a PreToolUse hook in ed04393. rtk discover over the preceding 30 days found 8 of 7348 agent Bash commands going through rtk — 0.1% — so every number collected before that commit describes a machine that was not using the tool. Nothing is worth deciding until a month of post-hook data exists.
Revise on or after 2026-09-12, against three questions.
1. Where does rtk fail outright? rtk gain --failures lists commands that could not be parsed and fell back to raw execution. Empty today. This is the only "works badly" signal that needs no interpretation.
2. What should never be rewritten? The lever is exclude_commands in the rtk config. The evidence for it is not in rtk's own numbers: Saved% measures volume removed, not value lost, and a command that saves 0% is merely useless while one that saves 90% may have dropped the single line the command was run for. That evidence only exists if it is captured at the moment a filtered output costs something. Decide the threshold before reading the numbers.
3. What could rtk cover but doesn't? rtk discover --all. Filters are declarative TOML under src/filters/, so a gap is a small upstream PR, not Rust work — and rtk supports user-global filters, so a candidate can be authored locally and proven on real work before being offered upstream.
Two candidates already visible, once the parser noise (shell fragments split on whitespace) and the single-line commands (nothing to compress) are discounted:
kubectl beyond get — a filter exists but is narrow; auth can-i, config get-contexts and describe all pass through unhandled.terragrunt — absent from the filter set entirely, while terraform-plan and all four tofu-* filters are present. Its plan output is larger than either.
Assisted-By: Claude noreply@anthropic.com
rtk was wired as a
PreToolUsehook in ed04393.rtk discoverover the preceding 30 days found 8 of 7348 agent Bash commands going through rtk — 0.1% — so every number collected before that commit describes a machine that was not using the tool. Nothing is worth deciding until a month of post-hook data exists.Revise on or after 2026-09-12, against three questions.
1. Where does rtk fail outright?
rtk gain --failureslists commands that could not be parsed and fell back to raw execution. Empty today. This is the only "works badly" signal that needs no interpretation.2. What should never be rewritten? The lever is
exclude_commandsin the rtk config. The evidence for it is not in rtk's own numbers:Saved%measures volume removed, not value lost, and a command that saves 0% is merely useless while one that saves 90% may have dropped the single line the command was run for. That evidence only exists if it is captured at the moment a filtered output costs something. Decide the threshold before reading the numbers.3. What could rtk cover but doesn't?
rtk discover --all. Filters are declarative TOML undersrc/filters/, so a gap is a small upstream PR, not Rust work — and rtk supports user-global filters, so a candidate can be authored locally and proven on real work before being offered upstream.Two candidates already visible, once the parser noise (shell fragments split on whitespace) and the single-line commands (nothing to compress) are discounted:
kubectlbeyondget— a filter exists but is narrow;auth can-i,config get-contextsanddescribeall pass through unhandled.terragrunt— absent from the filter set entirely, whileterraform-planand all fourtofu-*filters are present. Its plan output is larger than either.Assisted-By: Claude noreply@anthropic.com