dsh-terse

Mandatory, always-on terseness for DeepSeek Harness agents. Compresses everything an agent emits and everything it reads, without replacing the deployment's system prompt and without giving up answer quality.

There is no mode and no off switch. Mounting the row is the enabling act.

What it does

Layer	Extension point	Effect
L1 constitution	systemPrompt.section	A static, cache-stable rule set that makes every reply compressed.
L2 standing reminder	systemPrompt.context	A durable <system-reminder> snapshot that survives compaction, where a system-only rule fades.
L3 drift nudge	tools/post-execute -> additionalContexts	A one-line reminder that rides a tool result and catches a reply that ran long. Capped per turn; cannot loop.
L4 input shaping	tools/post-execute content	Collapses repeated lines and elides the middle of oversized tool output, with a locator so the original is one call away.
L5 enforcement	system-prompt/assemble	Asserts the constitution survived assembly; logs a loud warning if it did not. Never rewrites peer contributions.
L6 coverage	per-session state	Cadence and reply observations keyed by the session object, so a resumed or forked session starts clean.

Why it does not hurt quality

Compression is aimed at the chat mouth, never at meaning or at durable artifacts:

• Never cut negations, quantifiers, numbers, units, code, identifiers, paths, commands, or exact error strings. These are the token classes whose loss flips semantics.
• Never fake terseness. Invented abbreviations and arrows cost the same tokens and hurt clarity; if the terse form is not shorter than plain, plain wins.
• Full prose returns for security warnings, irreversible-action confirmations, order-sensitive multi-step sequences, and anywhere compression would create ambiguity.
• Code is never judged as drift, so a long implementation is never punished.
• Tool output is only elided reversibly, never lossy on errors, and the locator lets the agent pull the original back.

Those choices are measured in benchmark/README.md; their design record is a local working artifact, not a shipped doc.

Install

The plugin is one row, and it peers on the harness window >=0.1.5-rc.1 <0.3.0; the releases verified for it are listed in dsh.compatibility.dshReleases of package.json. Add it to a profile:

dsh plugin --profile <name> add @sagmans/dsh-terse

Or insert it directly in the profile's cordis.patch.yml:

- insert:
    - id: terse
      name: '@sagmans/dsh-terse'

The shipped row is in cordis.patch.yml.

Configuration

Every field is optional; the shipped defaults are the product.

Field	Default	Meaning
constitution	shipped text	Replaces the L1 rule set.
standing	shipped text	Replaces the L2 reminder; empty suppresses it.
shaping	true	Whether L4 runs.
drift	see below	L3 thresholds.
shapingConfig	see below	L4 bounds.

interface DriftConfig {
  maxReplyTokens: number           // 220  reply length above which the prose budget is exceeded
  minReplyTokensToJudge: number    // 60   floor below which no reply is judged
  maxProseRatio: number            // 0.72 prose ratio above which a long reply is padding
  minJudgeableProseRatio: number   // 0.4  below this the reply is code, never judged
  idleStepsBeforeNudge: number     // 14   steps with no nudge before the anti-drift backstop fires
  maxNudgesPerTurn: number         // 2    hard cap on nudges, and so on their extra steps
}

interface ShapingConfig {
  maxResultChars: number   // 12000 threshold above which a result may be elided
  headChars: number        // 3000  head kept
  tailChars: number        // 1500  tail kept, where error trails live
  duplicateRunMin: number  // 4     repeated lines before a run collapses
}

Development

pnpm install --frozen-lockfile
pnpm run typecheck
pnpm test
pnpm run build

A profile loads dist/, not src/; run pnpm run build before dogfooding a linked profile.

Try it in the TUI

scripts/dogfood/run-terse-from-worktree.sh clones the real ~/.dsh to a scratch directory, mounts this checkout on the real tui profile inside the clone, and launches it. Every composition input is present — your other bundles, the profile patch, settings, themes, credentials — while every write goes to the throwaway clone.

./scripts/dogfood/run-terse-from-worktree.sh              # clone, build, launch
./scripts/dogfood/run-terse-from-worktree.sh --no-launch  # set up, print the command
./scripts/dogfood/run-terse-from-worktree.sh --status     # show what the clone runs
./scripts/dogfood/run-terse-from-worktree.sh --with-sessions
./scripts/dogfood/run-terse-from-worktree.sh --clean      # remove the clone

What to look for on the surface: the status line shows injected … terse:standing … when the layer is live. Ask a short factual question (expect one line), ask for code inline (expect the full implementation, no elision), and ask something safety-critical (expect full prose). Those three cases are the contract; if the code case is trimmed, that is a bug.

Benchmark

benchmark/benchmark.sh measures this plugin against a control profile that lacks it. See benchmark/README.md; the neutral home is mandatory, because a profile whose global instructions already ask for brevity makes the control terse too.

License

MIT.