dsh-context-decay

简体中文

Model-driven, lossy context decay for DeepSeek Harness.

The model decides when and what to compress. Finished spans are replaced by self-written summaries; the current turn is always kept. There is no recovery: the summary is the only surviving record.

Install note: This is a manual-install plugin: it requires a DSH checkout with the modelMessageDecoration seam and is not a drop-in plugin-market bundle (see Requirements).

What it does

• Inline position anchors. Every model-addressable message with a visible text block is prefixed with ⟦p<number>⟧; tool-call-only assistant turns are unanchored by design. Compression ranges use those anchors directly: ⟦p<from>⟧ summary ⟦p<to>⟧. Anchors are presentation, never part of the content.
• Copied anchors removed. If the model starts a reply with one or more anchors, the leading run is stripped at the request boundary and again before the reply is committed, so the transcript and later requests stay clean.
• Model-selected spans. The model calls compress with one or more half-open ranges and writes each summary itself. No extra summarizer request. Split around content that must stay live; pass all consumed spans in one call so the prefix is not invalidated repeatedly.
• Lossy by design. Compressed content is gone. There is no decompress, no original-text search, and no external table. The summary must be self-contained.
• Tool results are compressible. All results of one assistant tool-call message must stay in the same span. A boundary inside that group is rejected; the plugin never silently widens a span.
• Current turn protection. The active turn and its tool results cannot be compressed. If the turn starts with unanchored system/runtime messages, a to boundary at its first visible anchor keeps those messages and compresses only the history before the turn; "end" remains the end-of-compressible-history sentinel.
• Tiered re-compression. Compressing a live summary block again produces a tier-2 summary; larger merges increase the tier. drop removes a block completely with no tombstone.
• Pressure without automation. The plugin reports context pressure and injects a reminder on each growth step; the reminder level follows the current pressure band, and the critical band uses a shorter step. It never compresses automatically; the model decides.
• Live panel + audit. A sidebar panel shows pressure, blocks, positions, request counts, and the reminder ledger. An optional JSONL audit records each request and compression.

Requirements

• DeepSeek Harness 0.2.1-rc.2 with the local host patch in patches/dsh-model-message-decoration.patch applied. Stock DSH does not expose the decoration seam; there is no runtime fallback. Installation alone is not enough: rebuild the host after patching.
• Node.js >= 22.19.
• A web profile. The client half needs the sidebar/session/locale UI packages listed in package.json.
• The bundle patch disables official automatic compaction (compaction-basic.auto: false), keeps the manual /compact command, and disables the in-place surface rewriters (tool-result-pruner, image-offload) when those rows exist. /compact remains an emergency host fallback; its checkpoint is not a context-decay block.

Install

Manual patched install:

# 1. Apply the host patch to a DSH checkout and rebuild the host.
git -C /path/to/deepseek-harness apply /path/to/dsh-context-decay/patches/dsh-model-message-decoration.patch
pnpm -C /path/to/deepseek-harness install
pnpm -C /path/to/deepseek-harness build

# 2. Install this plugin into a profile (checkout or tarball).
dsh plugin --profile web add /path/to/dsh-context-decay
# or: dsh plugin --profile web add ./dsh-context-decay-0.1.0.tgz

# 3. Restart the web server and open the profile.
dsh web --profile web

See docs/INSTALL.md for the full patched-checkout workflow, Windows commands, verification, updating, and rollback.

Tools

Tool	Purpose
compress({spans:[{from,to,summary,topic?}]})	Replace one or more half-open ranges with self-written summaries. from/to are visible anchor texts such as "⟦p12⟧" (bare "p12" and integer 12 also work); to may also be "end". Several non-overlapping spans per call are allowed. Tool-group edges are rejected.
drop({from,to})	Remove exactly one live summary block. Use the block's printed anchor pair; "end" is allowed for a tail block. No tombstone, no recovery.

A live summary block prints its pair inline as ⟦p<from>⟧ summary ⟦p<to>⟧; use that pair for drop. The sidebar panel shows pressure and the block ledger, so no extra status tool is needed.

Examples:

# one span
compress({spans:[{from:"⟦p4⟧", to:"⟦p9⟧", summary:"Read the three requirements files; constraints are ..."}]})

# two non-overlapping spans in one call
compress({spans:[
  {from:"⟦p4⟧", to:"⟦p9⟧", summary:"Read the requirements; constraints are ..."},
  {from:"⟦p9⟧", to:"end", summary:"Implemented the parser and tests; ..."}
]})

# drop one live summary block
drop({from:"⟦p4⟧", to:"⟦p9⟧"})

The anchor ids above are illustrative; use anchors visible in the current session.

Configuration

Key	Default	Meaning
enabled	true	Enable the plugin.
contextWindowTokens	route value	Override the total context window.
reservedOutputTokens	route value	Override the reserved output budget.
defaultContextWindowTokens	128000	Fallback when the route has no window.
growthTokens	20000	Periodic reminder step below the critical band.
criticalGrowthTokens	10000	Shorter periodic reminder step while at the highest warning level.
warnRatios	[0.70, 0.81, 0.95]	Warning levels as fractions of the effective window; the reminder level follows the current band.
reminderCooldownTurns	1	Minimum turns between normal reminders; warnings are not delayed.
maxSpansPerCall	8	Maximum spans in one compress call.
minSummaryChars	20	Minimum summary length.
maxSummaryChars	8000	Maximum summary length.
maxMissTokens	null	Optional one-time miss guard.
maxPaybackRequests	null	Optional payback guard.
audit	false	Write request/compression audit JSONL.
debug	false	Verbose plugin logging.

The effective window is total window - reserved output. The model sees pressure as a fraction of the effective window. The panel labels that percentage as effective pressure, shows the total and reservation, and adds a total-window occupancy line for direct comparison with the built-in context meter. Pressure, reminder, and request facts are kept per session, so switching sessions updates the panel.

How it works

The host patch adds a modelMessageDecoration service. The plugin registers a decoration provider that prefixes the messages selected by the replay state; the agent loop, token meter, and request payload all apply the same decoration. Positions come from the replayed surface, not from message text. Compressions append compaction/start, compaction/summary, a replacement user message, and compaction/end; replay rebuilds blocks, tiers, and the free list.

See docs/ARCHITECTURE.md for the state model, invariants, and failure modes.

See docs/PROMPTS.zh-CN.md for a Chinese translation of the model-facing prompts, tool descriptions, reminders, and error copy.

Limitations and non-goals

• No recovery. There is no decompress, search, or spill archive. This is deliberate; it is also the main trade-off.
• Requires a patched host. The decoration seam is not upstream. Without it the plugin cannot safely present anchors or replay positions.
• No automatic compression. Reminders are advisory. A model may ignore them; the budget stays protected by the request guard, not by silent compaction. The official automatic policy is off; manual /compact remains available.
• /compact is a manual fallback. It runs the official backend and writes an official checkpoint, not a context-decay tier block. Use it only when a session is already too large for a normal request.
• Manual install, not a market bundle. A market index cannot express the host-patch requirement, and engines.dsh cannot distinguish a patched host from a stock one. Distribute the checkout or tarball with the patch instructions until the seam is upstream.
• One compression backend per profile. Do not run another compaction backend or an in-place surface rewriter alongside it.
• Replay depends on the current event format. Compression replay uses the current surfaceOp and compaction event shape.

Development

npm install          # if building outside a DSH checkout
npm run build        # src/ -> lib/ (JS + declarations), client/ -> lib/client.js
npm run check        # typecheck only
npm test             # node:test suite against lib/

Local development can reuse a DSH checkout's TypeScript: set DSH_HARNESS_ROOT or place the checkout next to this package.

License

MIT. See LICENSE.