dsh-asc

dsh-asc

English | 中文

dsh-asc (full name DeepSeek Harness Agentic Surface Compaction) is a context-compaction plugin for DeepSeek Harness: the model itself decides when and what to compact, and every compaction decision is committed as a durable session-log replacement event (surfaceOp: replace) — replayable, searchable, and reversible.

Inspired by the model-driven compaction philosophy of opencode-acp, but built on DSH's event-sourced log: compaction creates no side-state files, decompression is log replay, and search covers the full log including compacted originals.

Install

Prerequisites: a working DeepSeek Harness installation (dsh CLI available); Node.js ^22.19 or >=24.

Harness compatibility: the built plugin targets core @deepseek-ai/dsh-* 0.2.0-rc.1 or newer (>=0.2.0-rc.1 <0.3.0); its peerDependencies also admit ^0.1.5-rc.2 so an older host still resolves the install. 0.1.0-rc.6 cores need the 0.2.0 release and 0.1.2-rc.1 cores the 0.2.1 release; core changed the Session API between the generations (0.1.5 additionally made the system prompt a surface-eligible system/message event and renamed the replace surfaceOp bounds to startSeq/endSeq), and the plugin fails at runtime when the core is newer than the release supports.

From npm (recommended):

dsh plugin --profile <name> add @internetnutzer/dsh-asc

From GitHub — to use a commit newer than the npm release:

dsh plugin --profile <name> add github:JanEickholt/dsh-asc

dsh plugin adds the plugin to the profile and enables it automatically based on the dsh.bundle declaration in the package; the tools and the system prompt load together with that profile.

Restart required: after installing, restart the running DeepSeek Harness service.

Other install options

From source — to modify the plugin itself, or to contribute:

git clone https://github.com/JanEickholt/dsh-asc.git
cd dsh-asc
pnpm install
pnpm build
dsh plugin --profile <name> add "link:$(pwd)"

Disabling the basic backend

ctx.compaction allows only one provider at a time. Disable the default basic backend in your profile's own cordis.patch.yml:

- id: compaction-basic
  disabled: true

Optionally mount the invariant companion and the full-text-search backend:

- insert:
    - id: dsh-asc-invariant          # runtime invariant checks (optional; DSH <= 0.2.0 only)
      name: "@internetnutzer/dsh-asc/invariant"
    - id: session-query-sqlite       # context_search full-text backend (optional)
      name: "@deepseek-ai/dsh-session-query-sqlite"

Usage

After installing and restarting, no configuration is required — the plugin:

• injects the context-management discipline into the system prompt (judgment rules, tool usage, tiered compaction cadence), so the model actively manages context from the very first turn;
• injects nudge prompts on demand when context usage runs high (cadence-gated; iteration nudges additionally require real token growth — no per-turn nagging);
• provides deterministic degradation (LLM summarization, plus tool-result pruning when the optional upstream pruner is mounted) on overflow or manual compaction, without requiring model cooperation.

The plugin provides six model tools:

Tool	Purpose
context_status	context usage, tiered checkpoints, system/dialogue composition, recommended ranges, recent surface nodes
context_compress	replace a surface range with a checkpoint you write (batching supported; tool-call pairs auto-extended; quality gate)
context_decompress	undo a compaction: the original text returns to the surface at the checkpoint's own position (tier-aware; full: true reaches raw content)
context_recap	re-read checkpoint summaries without decompressing the originals
context_search	full-text search over the whole log (including compacted content)
context_retrieve	return a projected tool-result original byte-exact by its 24-hex sha-256 hash or seq

Compacted content is never lost: the originals stay in the session log and can be decompressed or searched at any time.

Reversible tool-result projection

Separately from model-driven compaction, an optional projection service compresses oversized tool results BEFORE they enter the context. It listens on the tools/post-execute waterfall, measures every candidate's text blocks with the real token meter, and returns every decision unchanged so the original event still lands in the log first. A reversible commit then shadows the original and appends a replacement whose content embeds a retrieval marker:

[dsh-asc projection: structured:json compressed 4200→312 tokens. Full
original (seq 57, stored in this session log): context_retrieve(hash="…").]

• The original stays byte-exact in the session log (single source of truth; no side store), so it survives restarts; a marker is emitted only when the original is durably stored — a marker never dangles.
• Reducers are content-aware: structured explorers for JSON/YAML/XML/delimited/code, git-diff hunk compaction, search-result clipping, CLI rule reduction, and a meter-priced head/tail slice as the last fallback. context_retrieve returns the stored original byte-exact by hash or seq; unknown keys get a diagnostic, never fabricated content.
• The six context_* tools' own results are never projected (retrieval stays byte-exact; status/recap/search reports are already capped).
• The projection mounts as its own cordis service row, independent of ctx.compaction — disabling it never disables compaction, and the overflow-triggered tool-result pruner stays mounted as the fallback.
• Config: projection.enabled (default true), projection.thresholdTokens (default 1000), projection.excludeTools (default []), projection.disabledReducers (default ['code']); see docs/usage.md for the measured trade-off.

The system prompt ties the tools into one operating loop: capture consumed raw work into tier-1 checkpoints, distill settled tier-1 piles into tier-2 decisions and tier-2 piles into a tier-3 fact index. Every checkpoint text carries its topic and Compaction id, so when a visible summary already points at the needed detail the model decompresses that block directly; context_search is used only when no visible summary says where a detail lives, and decompression always proceeds one tier at a time.

How it works

• Event sourcing: a compaction is a transaction in the log (compaction/start → compaction/summary → replaced user/message → compaction/end); no side state.
• Tiered compaction: checkpoints have tiers (T1 full detail → T2 distilled decisions → T3 bare facts); summaries get thinner as they are reused.
• Reversible: decompression replays the events shadowed in the log and commits one in-place replacement event; no side state is needed.
• Auditable: who compacted what, the full summary text, and the token cost are all in the log.

Repository layout

src/
  index.ts      plugin entry: registers ctx.compaction + the six tools
  config.ts     strict config validation
  types.ts      shared config and result types
  events.ts     session-event vocabulary documentation (no custom members)
  invariant.ts  runtime invariant companion (subpath export)
  engine/       the compaction engine core (engine, region, tier,
                quality gate, fallback, prompt, restore)
  policy/       protected-node policy and the nudge state machine
  tools/        the six model tools
  projection/   reversible tool-result projection service + reducers
  analytics/    session-log usage scanner, regret signals, asc-stats command
  utils/        shared text helpers
tests/          vitest suites
scripts/        corpus-stats and bili-cache-join measurement scripts
docs/           usage, design, analysis, e2e-validation, cache-join

Using dsh-asc alongside the official tool-result pruner

dsh-asc aims at post-execute, reversible tool-result compression: every projection keeps a context_retrieve lookup so the full original can be restored at any time. The official @deepseek-ai/dsh-compaction-tool-result-pruner (the tool-pruning patch shipped by the dsh-compaction bundle) complements this as a last-resort overflow guard: when a request still overflows the context window, it truncates oversized tool results to head + tail so the turn survives.

We recommend installing both:

• dsh-asc projection — primary path: reversible post-execute management.
• tool-result pruner — overflow insurance: caps results at head 4096 + tail 1024 characters, but only after a request actually fails on overflow.

Caveat: pruner truncation is irreversible and fires only on overflow — nothing intervenes between a tool's execution and the failed request, so oversized results sit at full length mid-turn. Rely on projection first; the pruner is the safety net, not the standard path.

The two compose cleanly: pruneSession() is idempotent and projection keeps its own retrieval index, so either can run first.

Measured in production

The numbers below come from real sessions on the maintainer's machine — every one is reproducible by running scripts/corpus-stats.mjs against your own DSH session logs (usage: node scripts/corpus-stats.mjs [~/.dsh/sessions]). Session logs are durable, so nothing here is estimated or simulated: folds, shadowed-token counts, and per-request cache usage are all committed events.

Corpus scanned 2026-10-08 — 240 sessions with folds, at least one context_* tool call each (engine attribution: only sessions where dsh-asc's tools were registered count; sessions from the compaction-engine swap experiment are excluded by that fingerprint):

Metric	Value
Sessions with folds	240
Folds (compaction transactions)	721
Tokens shadowed by folds	51.9M
Model-authored summaries	721 (fallback: 10)
Failed folds	16 (2.2%)
Median fold size	60.5K tokens
Prompt-cache hit rate across all metered requests	94.8%
Median first-request-after-fold cache miss	27.6K tokens
context_decompress after a fold	9 calls — 0.7% of folds
context_recap	12 calls
context_retrieve (projection lookups)	1,188 calls
context_search	38 calls

The story the table tells: across ~721 folds covering ~52M tokens of compacted history, the model asked for originals back 9 times. Summaries plus tiered checkpoints carried the work; when detail was genuinely needed, context_retrieve restored it reversibly (1,188 lookups without a single irreversible loss). The 94.8% cache hit rate — measured across all 78,927 metered LLM requests in those sessions, including the re-pay spikes every fold causes — shows compaction coexisting with prompt caching at a healthy level rather than destroying it.

How a fold plays out

flowchart TD
    A[Model calls context_compress] --> B[compaction/start]
    B --> C[Quality gate: summary must pass L1 floor + L2 recall]
    C -->|pass| D[compaction/summary: shadowedRange + tiered summary committed]
    C -->|fail| G[ fold rejected — original stays ]
    D --> E[user/message replace event: surface switches to the summary]
    E --> F[compaction/end]
    F --> H{Model needs detail later?}
    H -->|usually not| I[Work continues on the summary]
    H -->|9 of 721 folds| J[context_decompress / context_retrieve]
    J --> K[Log replay restores byte-exact originals]
    K --> L[One in-place replace event — still no side state]

    style C fill:#f9f
    style J fill:#ff9

Two properties visible in this flow that no wire-level compressor can offer: the quality gate (a fold that fails the floor/recall check never destroys history — it is rejected) and reversibility (decompression is log replay, not a cache of originals; the log is the only source of truth).

Cost of compaction, honestly

Each fold rewrites the message list, so the next request re-pays the prompt cache: the median first-request-after-fold miss is 27.6K tokens (vs a 60.5K median fold). Folds are one-time costs paid for by the per-turn savings of a smaller surface — the cache-join script measures exactly this trade against a wire proxy's ledger, per fold, with PAID BACK / NOT PAID BACK verdicts.

Numbers, not adjectives: the corpus-stats and cache-join scripts exist so the "does compaction pay?" question gets answered from your own logs.

Documentation

Doc	Contents
docs/usage.md	install, configuration, model experience, operations
docs/design.md	implemented contract: events, tools, automatic behavior, protection, invariants
docs/analysis.md	comparison of DSH and opencode-acp context management

License

MIT. Algorithmic inspiration from DeepSeek Harness (MIT); only the ideas of opencode-acp (AGPL) are used, no source code. The tool-result projection is adapted from the MIT-licensed flowctx-dsh port of flowctx. See NOTICE.