# 03 — RTK v0.45.0: current design (https://jackin.tailrocks.com/research/context/tools/03-rtk-design/) ## Summary [#summary] RTK is a narrow, deterministic shell-observation reducer with strong command coverage, but it cannot reduce native reads or conversation history. Cutoff: **2026-08-12**. Current stable: [v0.45.0](https://github.com/rtk-ai/rtk/releases/tag/v0.45.0), released 2026-08-07. RTK is a deterministic command proxy. An agent hook rewrites eligible shell commands through one Rust binary; RTK runs the real command, preserves its behavior, and emits a smaller observation using command-specific rules. It does not compress the model's native file/search tools or the conversation history. ## Question and scope [#question-and-scope] Which shell observations can RTK v0.45.0 reduce deterministically, and which context sources remain outside its reach? ## Method [#method] The review fixes RTK to v0.45.0, inspects its command coverage and deterministic transforms, and distinguishes payload counters from independent task-level evidence. ## Findings [#findings] ### Current facts [#current-facts] | Field | Current value | | ----------------- | --------------------------------------------------------------------------------------------- | | Repository | [rtk-ai/rtk](https://github.com/rtk-ai/rtk) | | Stable release | [v0.45.0](https://github.com/rtk-ai/rtk/releases/tag/v0.45.0) | | Adoption snapshot | 75,756 stars / 203 subscribers / 4,758 forks / 1,950 open issues and PRs | | Current scope | 100+ recognized development commands plus custom filters | | Form factor | Small Rust binary + per-agent shell hook; no compression model | | Current headline | Up to 90% reduction in Bash output; current README explicitly says this is not bill reduction | | License | Apache-2.0 | Metadata source: [GitHub API](https://api.github.com/repos/rtk-ai/rtk). Current scope and caveats: [README](https://github.com/rtk-ai/rtk). ### Current interception [#current-interception] ```text agent asks shell tool to run command ↓ hook classifies and rewrites eligible command ↓ rtk runs the real command ↓ command-specific deterministic filter ↓ exit status + compact observation enter context ``` Because compression occurs before the new observation enters history, it is cache-safe by construction. No learned model can hallucinate the summary. The trade is reach: the current README states built-in Read, Grep, and Glob tools bypass the hook. RTK only helps when the agent invokes shell commands RTK recognizes or the user explicitly calls RTK wrappers. ### Current release evidence [#current-release-evidence] v0.45.0 adds Mistral Vibe hook support and multiline command rewriting. It also hardens GitHub Copilot hook installation/healing while preserving user configuration. These are integration improvements; the fundamental shell-only boundary remains. Release source: [v0.45.0 notes](https://github.com/rtk-ai/rtk/releases/tag/v0.45.0). ### Current evidence [#current-evidence] RTK provides current per-command examples and a local gain ledger. These establish mechanics, not end-to-end efficiency. No independent accepted-task benchmark explicitly testing **v0.45.0** was found at the cutoff. Results from superseded versions are excluded. ## Implications for jackin❯ [#implications-for-jackin] ### Benefits [#benefits] * Deterministic, inspectable rules; no ML model or inference latency. * Small single artifact and low per-command overhead. * Zero MCP tool-schema rent. * Write-time cache safety. * Broad current command catalog and custom filter support. * Explicit raw command remains available. * Current README correctly scopes headline percentages to Bash output. ### Best fit [#best-fit] RTK is a good candidate when traces show supported shell output is a large fraction of **fresh** input: verbose tests, builds, Git status/diff, package managers, containers, or infrastructure CLIs. It is also the simplest option when deterministic behavior, minimal runtime, and zero model dependency dominate. Do not install it as a generic token remedy when the agent mainly uses native reads/search, commands are already quiet, or prompt-cache reads—not new shell output—dominate consumption. ### Current verdict [#current-verdict] RTK v0.45.0 is the narrowest and easiest-to-reason-about input tool. That is both strength and limit. It is objectively good for large supported shell observations; there is no current evidence that it reduces every coding task. Status: **use after a reach audit, not by headline**. Next: [04 — lean-ctx v3.9.18](/research/context/tools/04-leanctx-design/). ## Limitations and unknowns [#limitations-and-unknowns] ### Costs and risks [#costs-and-risks] * Structural reach ceiling: native agent file/search tools bypass it. * Command rewriting can conflict with other PreToolUse hooks or shell wrappers. * A successful command may still contain a detail removed by the filter; raw rerun is the recovery path. * Filtering can cause the agent to issue follow-up commands, erasing gross savings. * `rtk gain` estimates removed output using approximate token arithmetic; it cannot observe counterfactual turns, cache changes, or task quality. * Current documentation conflicts on telemetry default: privacy copy says anonymous telemetry is opt-in and disabled by default, while `DISCLAIMER.md` says telemetry is enabled by default. Resolve from the exact installed build/config before deployment. * Current README's verification example still names an older version, a documentation freshness defect. ## Sources [#sources] Evidence and repository references are cited inline beside the claims they support. ## Related work [#related-work] * [Token-optimization tools dossier](/research/context/tools/) * [Token-optimization techniques](/research/context/techniques/) * [Context engine](/research/context/engine/)