dsh-compaction-tools-gate

A DeepSeek Harness plugin that stops automatic compaction from silently losing its summary.

Mitigates discussion #5521, confirmed by a maintainer as "a real, reproducible data loss defect". This plugin is a stopgap. The correct fix is upstream, and when it lands you should delete this.

The defect

compaction-basic's summarizer replays the conversation's tool schemas into the summarization call — deliberately, to reuse the warm prefix cache (summarizer.ts, comments 111–114):

...input.tools === undefined ? {} : { tools: [...input.tools] },   // summarizer.ts:158

But GenerateOptions (llm/src/types.ts:393) has no toolChoice field — no way to say "you have tools, don't call one." Its own doc comment says so:

GenerateOptions sampling is temperature/maxTokens/stop only — no tool_choice, top_p, or penalty fields

So a tool-capable model, deep in a tool-calling session, can answer the "you are now acting as a compaction engine" instruction with another tool call. summaryText() keeps text blocks only, finds none, and throws summarization produced no text summary content (summarizer.ts:171). It retries until the provider returns 400.

Nothing in the session log distinguishes that from a slow compaction. The rawOutput is discarded before the throw, so one generic string covers three different bugs: a genuinely empty completion, a truncation, and a model that chose the wrong behaviour.

What this does

Registers one listener on the llm/stream waterfall and, for calls tagged purpose: 'compaction' only, removes the tools array.

- insert:
    - id: compaction-tools-gate
      name: 'dsh-compaction-tools-gate'

Two things make the seam the right place rather than a config change:

• It reaches the acting instance. Under dsh web the app disables the host-plane compaction-basic and the standard preset mounts its own inside an isolated group — so a home-patch config row configures a disabled instance. Under --profile headless the host-plane one is live. The llm/stream waterfall reaches both.
• No preset fork, no patched vendor code. The plugin has zero dependencies and imports nothing from @deepseek-ai.

The cost, measured

The tools are replayed for cache reuse, so dropping them costs a re-prefill. On one box — qwen3.8-27b (Q5_K_XL) on llama.cpp across 2× RTX A4000, 131k window, one slot, ~815 t/s prefill:

Over 74 compactions in real sessions:

compactions that produced no summary	25 of 64 (39%)
of those, this defect's signature	15 (12 in a single session)
trigger point, median input	65k tokens of a 131k window
total cost of one compaction	~208 s, once per ~54k tokens of growth
— summarizing	131 s (decode-bound, ~22 t/s)
— re-prefill caused by dropping the tools	77 s

From one captured compaction (a single sample, not a distribution): 87% of the summarization request was tool results, not schemas or assistant text; capping them made the request 43% smaller and the re-prefill 88 s → 50 s.

So the trade is roughly 77 s per compaction against a 39% chance of losing the summary entirely. On this hardware that is not a close call. Your mileage depends on prefill speed: at 815 t/s the re-prefill is a minute; on a faster lane it is seconds, and on a slower one you should think harder.

Options

config:
  # Cap oversized tool results inside the summarized region. 0 disables.
  # The trade is real: a summary can miss what only lived in a truncated tail.
  # Every cut is marked in-place; assistant text is never touched.
  toolResultMaxChars: 2000

  # Send compaction to a different model. A THINKING lane spends the
  # summarizer's token budget on reasoning and truncates; its nothink sibling
  # of the same weights writes the summary with the whole budget, faster.
  reroute:
    'provider/thinking-model': 'nothink-model'

  quiet: true   # no log line per gated call

Scope, and what it does NOT fix

The gate keys on options.purpose !== 'compaction' and passes everything else through. That is deliberate — a llm/stream listener sees every request, and a broad tool-stripping rule there is a footgun.

It is also why this cannot substitute for the upstream fix: any other synthetic call that carries tools and must not use one has the same exposure. A general optional toolChoice on GenerateOptions is the right shape, and if it lands compaction-specific this plugin is still needed for the next call of that class.

Install

npm i dsh-compaction-tools-gate

Then add the insert row above to your cordis.patch.yml, or apply the bundled patch via the dsh.bundle field.

Test

npm test        # 36 checks, no network, no model

Verified against dsh 0.1.2-alpha.5 and 0.1.2-rc.1; the defect is present and unchanged in both. An offline end-to-end run (stub provider, a 64k window, automatic compaction fired twice) asserts the summarization request carries no tool schemas while main-loop requests keep all 33 of theirs, and that the resulting compaction/summary holds text.

License

MIT