dsh-compaction-micro

Standalone micro-clear tier for DeepSeek Harness compaction. It clears whitelisted, re-runnable tool results to placeholders at its own microThresholdRatio — below the summarize threshold — so the expensive model-backed summary runs less often. It is an observer: it never mutates the model request, makes no model call, and every rewrite is a single-node tool/result content replacement preceded by a compaction/prune shadow-price event, so the durable log stays the reconstructable source of truth.

Snip (trimming) and full (summarization) stay owned by the official dsh-compaction-tool-result-pruner and dsh-compaction-basic.

Composition

Mount this bundle before dsh-compaction-basic. Its agent/pre-step and agent/request-error listeners are registered with prepend: true, so they run first regardless of mount order:

- name: '@deepseek-ai/dsh-token-meter'
- name: '@deepseek-ai/dsh-compaction-tool-result-pruner'
- name: '@deepseek-ai/dsh-compaction-basic'
  config:
    thresholdRatio: 0.9
- name: 'dsh-compaction-micro'
  config:
    clearableTools: [read, glob, grep, web_search, web_fetch]
    retainRecent: 3
    microThresholdRatio: 0.85
    reservedTokens: 0

How it works

1. agent/pre-step (prepend): measure pressure; when totalTokens >= floor((contextWindow - reservedTokens) × microThresholdRatio), clear whitelisted tool results oldest-first, retaining the newest retainRecent visible. The official backend then re-measures the smaller surface and skips summarization when the clear already relieved enough pressure.
2. agent/request-error (prepend): on a confirmed context overflow, clear whitelisted results first, then delegate to the official prune + summarize + retry.

Configuration

Field	Default	Meaning
clearableTools	[]	Tool names whose results the micro tier may clear. Empty disables the tier. Only list read-only, re-runnable tools — the service does not verify that property.
retainRecent	3	Newest clearable results always retained visible.
microThresholdRatio	0.85	Clear at this fraction of the effective window. Must be in (0, 1].
reservedTokens	0	Tokens subtracted from the routed context window before scaling the threshold.
auto	true	Install the automatic listeners; false exposes only ctx.microCompaction.clearSession().

The whitelist is operator-owned and is the safety gate: read/glob/grep and read-only web tools are the intended members; write, edit, shell, subprocess, and workflow tools must stay out.

Development

Build against a local DeepSeek Harness checkout by junctioning its workspace packages into node_modules/@deepseek-ai (see the checkout's packages/*/* and vendor/*), then:

npm run typecheck     # tsc -p tsconfig.json --noEmit
node --import tsx --test tests/micro.test.ts

Runtime smoke (from the DSH checkout):

dsh plugin --profile headless add link:<this-dir>
DEEPSEEK_BASE_URL=http://127.0.0.1:8000/v1 DEEPSEEK_API_KEY=mock-key \
  dsh --profile headless "run the pwsh tool once and report"

License

MIT