YunaAloft/dsh-rtk-workbench

English · 中文 · Español · हिन्दी · Português

Bring rtk (Rust Token Killer) to DeepSeek Harness — and ship the binary with it.
_{rewrite what you can · compress what is left · tell the model what happened}

Why this exists

rtk works by sitting between the agent and its tools: it rewrites a command into a filtered variant, and it compresses tool output before the model reads it. It ships its own adapters for 20 agents — for Claude Code that means writing a PreToolUse hook into its config, and the other 19 work the same way. DeepSeek Harness had none of those adapters; this plugin is the missing one.

The existing DSH plugins close that gap only halfway — they all implement command rewriting and stop there. That leaves two thirds of rtk's value unused:

Gap	Consequence
No output compression	Commands rtk cannot rewrite (node build.js, cat big.log, MCP calls) still dump their full output into the context
No awareness section	The model sees [... elided ...] and re-runs the command to "get the real output" — paying twice
rtk must be installed by hand	A plugin that only works if you install something else first is a plugin that silently does nothing

This plugin closes all three, and keeps every failure path open: if rtk is missing, slow, or answers something unexpected, the tool call behaves exactly as it would without the plugin.

What you get

Capability	Default	How it works
Command rewriting	on	Wraps ToolDefinition.execute; git status runs as rtk git status, and a result above 500 B starts with [rtk] ran: … so the executed command is visible
Tool-output compression	on (≥ 2048 B)	Runs rtk hook claude with a synthetic PostToolUse event in tools/post-execute; the condensed text replaces the content while the canonical value stays intact for run_code
Awareness section	on	One static system-prompt section explaining [rtk] ran:, the elision markers, and rtk recall
Bundled rtk	on	The binary ships as a per-platform optional package; nothing to install, rtkSource: 'path' opts out
rtk deny rules	off	Feeds rtk hook claude a PreToolUse event and honours permissionDecision: deny
/rtk command group	always	status, tools, doctor, gain, recall <hash>, and on / off per switch

Nothing here registers a new tool, replaces a shell executor, or touches the model API.

Install

From the plugin panel — search dsh-rtk-workbench in Add plugin and install.

With the CLI:

dsh plugin --profile web add dsh-rtk-workbench

Or straight from a GitHub Release — no registry involved. Pick the asset for your platform:

Platform	Asset
Windows x64	dsh-rtk-workbench-win32-x64.tgz
macOS Intel	dsh-rtk-workbench-darwin-x64.tgz
macOS Apple silicon	dsh-rtk-workbench-darwin-arm64.tgz
Linux x64	dsh-rtk-workbench-linux-x64.tgz
Linux arm64	dsh-rtk-workbench-linux-arm64.tgz

# `latest/download/` always resolves to the newest release, so the URL survives upgrades
dsh plugin --profile web add https://github.com/YunaAloft/dsh-rtk-workbench/releases/latest/download/dsh-rtk-workbench-win32-x64.tgz

Asset names carry no version on purpose: a versioned file name would break every releases/latest/download/ link on the next release. Which version you got is visible in /rtk status (version : rtk 1.1.0) and in the release tag.

Restart the host (a bundle change is not a page reload). You do not need to install rtk separately — the platform binary is an optional dependency of this package:

#	Resolution order
1	rtkPath config
2	RTK_BIN / DSH_RTK_BIN
3	the bundled platform package (or vendor/ inside a release tarball)
4	rtk on PATH

A platform package that fails to download simply loses its turn, so the plugin degrades to "use the rtk you installed" instead of failing.

Quick start

/rtk status     # binary, provenance, pipeline, switches, what was patched
/rtk doctor     # deep self-check: version, pipeline, a real round trip, recall
/rtk tools      # how each visible tool resolved (canonical / alias / explicit)

Then just work normally:

• ask the agent to run git status in a big repository → the result begins with [rtk] ran: … and contains the compact one-line-per-file form (tiny clean trees stay un-annotated on purpose: the line costs more than it saves there);
• ask it to cat a 100 KB log → the result is condensed, ending with [full output: rtk recall <hash>];
• paste that hash back and ask for rtk recall <hash> → the original comes back in full.

How it works

flowchart LR
  M[model calls a tool] --> P[tools/pre-execute]
  P -->|patch once| X[ToolDefinition.execute]
  X -->|rtk rewrite| R{rewritten?}
  R -->|yes| E1["run rtk <cmd>"]
  R -->|no| E2[run original cmd]
  E1 --> C[tools/post-execute]
  E2 --> C
  C -->|await next first| D[downstream listeners]
  D --> G{">= 2048 B and unmarked?"}
  G -->|yes| H[rtk hook claude → updatedToolOutput]
  G -->|no| I[pass through]
  H --> J[content replaced, value preserved]

Three details matter for correctness, and each cost a real bug to learn:

1. Patching, not re-dispatching. A tools/execute listener that skips next() loses the registry's signal fusion and bodyInvoked bookkeeping, degrading cancellation. Patching execute keeps that machinery intact and substitutes one argument.
2. await next() first. A tools/post-execute listener that returns early short-circuits every listener behind it — including other plugins' notices.
3. The annotation never eats the body. If a result cannot be read as text, or the tool is not compressible, the annotation is added as its own block or skipped — never the other way round.

Measured effect

Reproduce with node scripts/benchmark.mjs: it builds a deterministic fixture repository, then measures each case twice — once raw, once through the same path the plugin uses (rtk <command> for rewriting, a synthetic PostToolUse event through rtk hook claude for compression).

Case	Path	Before	After	Saved	Est. tokens
git status	rewrite	377 B	47 B	−87.5 %	95 → 12
git log, 20 commits	rewrite	3 578 B	1 698 B	−52.5 %	895 → 425
git diff, 100 changed lines	rewrite	20 511 B	6 945 B	−66.1 %	5 128 → 1 737
ls -la	rewrite	508 B	138 B	−72.8 %	127 → 35
grep -rn TODO src, 300 hits	rewrite	20 364 B	2 190 B	−89.2 %	5 091 → 548
pytest output, 2 001 lines	compress	89 505 B	11 287 B	−87.4 %	22 377 → 2 822
application log, 5 000 lines	compress	272 640 B	13 633 B	−95.0 %	68 160 → 3 409
package-lock.json, read shape	compress	76 896 B	4 654 B	−93.9 %	19 224 → 1 164
Total, one call each		484 379 B	40 592 B	−91.6 %	121 095 → 10 148

These are bytes, not tokens: the token column is bytes / 4, an estimate for ASCII-heavy text — read it as an order of magnitude. The fixture is synthetic on purpose (repetitive logs, a 300-TODO file, a generated lockfile), so re-run the script on your own workload before quoting a number.

Latency per round trip (30 samples, process.hrtime):

Operation	p50
rtk --version (process start only)	50 ms
rtk hook claude, 8.4 KB payload	96 ms
rtk rewrite "git status"	206 ms
First spawn in a fresh host process	2.7–5.8 s (the platform scanning a 17 MB unsigned binary)

That is why the compression threshold is 2048 bytes, why an already-rtk command never spends a spawn, and why the start-up probe gets its own budget instead of blocking a tool call.

Coexistence

The plugin is built to sit next to whatever else you run:

• It does not take ctx.shell. Rewriting happens by wrapping the registered tool, so a plugin that replaces the shell executor keeps working.
• It always awaits next() in tools/post-execute, so a safety plugin's "this command was not checked" notice and the spill policy still run.
• It never names or blocks other plugins. If another listener already replaced a result's content, the plugin notices (a content diff, nothing more), keeps working, and reports observed: another result transformer changed N result(s) in /rtk status.
• deny is off by default, because command blocking usually belongs to a dedicated safety plugin; rtk's allow rules are what this plugin honours when it is on.

Configuration

Every knob is a Schemastery schema; invalid values fail at load instead of at runtime.

Key	Type	Default	Description
rtkPath	string	''	Explicit binary; empty = RTK_BIN / DSH_RTK_BIN / bundled / PATH
rtkSource	string	'auto'	auto prefers the bundled binary, path always resolves through PATH
timeoutMs	number	2000	Per-invocation budget; a timeout means no rewrite and no compression
rewriteEnabled	boolean	true	Master switch for rewriting
shellTools	string[]	['bash','pwsh']	Tool names treated as shells for detection
rewriteTools	string[]	['bash','ops_bash','pwsh']	Eligible for rewriting; a local pre-filter skips PowerShell cmdlets and $-expressions, which rtk always defers on
compressEnabled	boolean	true	Master switch for compression
compressTools	string[]	['*']	Eligible tools; * = every tool's text result
excludeTools	string[]	[]	Never touched, regardless of other lists
compressThresholdBytes	number	2048	Below this the ~96 ms spawn is not worth it
maxPayloadBytes	number	2097152	Above this skip (rtk caps hook stdin at 64 MiB)
awarenessEnabled	boolean	true	Inject the awareness section
denyEnabled	boolean	false	Honour rtk's deny verdict
pipelineMode	string	'plugin-owned'	plugin-owned injects RTK_CONFIG_DIR; user relies on your own pipeline file
annotateRewrites	boolean	true	Prepend the [rtk] ran: line
annotateMinBytes	number	500	Skip that line when the result is shorter than this — it names a full path, so on a tiny result it costs more than the rewrite saved

RTK_DB_PATH and RTK_TEE_DIR are deliberately not overridden: rtk recall is run by the model through the ordinary shell, and redirecting the tee/DB would make recovery impossible.

Troubleshooting

/rtk status says rtk is not usable. The line names the reason (spawn, timeout or exit). The first spawn in a fresh process can take seconds — a timeout there is usually the platform scanning the binary, not a broken install. If it says via path, the bundled package for your platform did not install; set rtkPath or install rtk yourself.

Compression never fires. Check /rtk tools: a tool without a compress mark is not in compressTools, or the result is below compressThresholdBytes. If the command was rewritten, skipping compression is correct — rtk already produced the compact output.

Rewriting never fires. Get-ChildItem and $env:PATH are PowerShell grammar; rtk defers on them by design, so the plugin does not even spend a process. git status, cargo test, docker ps and ls are rewritten normally.

rtk recall fails with a store error. The tool shell is sandboxed (workspace-write) and cannot reach rtk's database in %APPDATA%; the plugin itself runs outside that sandbox. Ask the user to run /rtk recall <hash> — a slash command runs in the host process and reads the same store.

Two compressors at once. /rtk status shows observed: another result transformer … when something downstream replaced a result's content. Nothing breaks, but keeping one compressor is usually better — /rtk off compress turns this one off for the session.

Nothing works in a brand-new profile. Windows profiles may ship only the pwsh tool. That is enough for compression and for rewriting real programs; a bash tool appears only if a shell plugin provides one.

Development

pnpm install
node node_modules/typescript/bin/tsc --noEmit
node node_modules/vitest/vitest.mjs run
node node_modules/tsdown/dist/run.mjs          # build lib/
node scripts/build-vendor.mjs                  # bundle rtk (downloads ~31 MB, cached)
node scripts/release.mjs --dry-run             # show the publish plan
npx dsh-plugin-dev check --cwd . --strict
npx dsh-plugin-dev verify --cwd . \
  --base @deepseek-ai/dsh-base@0.2.1-rc.5 \
  --headless @deepseek-ai/dsh-headless@0.2.1-rc.5

scripts/smoke-rtk.mjs <binary> proves a bundled binary runs on the current platform. CI runs it on every platform we ship — Ubuntu x64/arm64, macOS arm64/x64 and Windows — while the unit suite runs on Ubuntu, macOS arm64 and Windows.

Pushing a change under .github/workflows/ needs the workflow token scope; run gh auth refresh -s workflow once and the next push goes through.

Releasing. Build every platform (node scripts/build-vendor.mjs with no --only), then publish the five platform packages and the main package in that order:

node scripts/release.mjs --yes --registry https://registry.npmjs.org
gh release create v0.1.0 dist/github/*.tgz dist/github/checksums.txt dist/github/manifest.json \
  --title "dsh-rtk-workbench v0.1.0 (rtk 1.1.0)" --generate-notes

release.mjs then prints a one-shot registry snapshot: a new package is accepted with HTTP 202 and served as 0.0.0-stage until its staged release is approved, so a successful publish is not yet an installable one. Add --verify to wait for the staged releases instead of reading the snapshot.

The order matters: the main package lists the platform packages as optionalDependencies, so publishing it first would make every install resolve zero binaries. The platform packages are named dsh-rtk-workbench-<platform> and carry the same version as the main package; the bundled rtk version lives in dshRtkVersion.

Credits

The compression rules, filters and the hook protocol are rtk's work; this plugin is the DeepSeek Harness adapter and ships the unmodified rtk binary. That binary is Apache-2.0 and links rtk-engine under BUSL-1.1: source-available, with an additional use grant that makes production use free and that turns into Apache-2.0 on 2030-09-09, but which does not permit offering rtk itself as a paid or hosted product. Both texts ship in licenses/, with a one-page map of what covers what.

License

Apache License 2.0 © 2026 dsh-rtk-workbench contributors.