dsh-cwl

dsh-cwl

CWL — Context Window Lifecycle for DeepSeek Harness: structured context eviction for long-horizon agents.

Paradigm: Beyond Compaction: Structured Context Eviction for Long-Horizon Agents (arXiv:2606.11213, Kiz8)

English | 简体中文

Why not summarization compaction?

Compaction (the standard response to context pressure) summarizes history with an LLM. Four structural problems (per the CWL paper):

• Unpredictable lossiness — the summarizer decides what matters, not the task.
• Structural destruction — causal chains (tool call → output → decision → action) collapse into prose.
• Blocking cost — a full LLM call fires mid-task, under token pressure.
• Compression-induced hallucination — summarization under length pressure is a known failure mode.

CWL treats the transcript as a structured record of work and evicts deterministically: the agent's trajectory is inferred into a typed episode graph (exploration expl / action act, with dependency edges), and when context pressure exceeds budget, a zero-LLM, deterministic policy strips content in graduated levels — exploration episodes first (pure context, safest), then action episodes whose effects are already persisted. User messages are never evicted.

How it works

1. Episode inference (automatic, no agent annotation needed): consecutive same-type tool batches merge into semantic episodes (expl for pure read/search — including read-only bash like grep/cat — act for anything with side effects: edit/write/write-style bash). Each user message closes the current episode (a turn boundary), and episodes are capped at a batch limit, so even a single-request long autonomous run (dozens of tool calls) splits into bounded, evictable segments instead of collapsing into one giant episode. An act that touches files an earlier expl read gets a dependency edge.
2. Pressure metering: real context pressure = input + cacheRead + output + reasoning tokens (accumulated from assistant/message usage events — tokenMeter.measure().totalTokens omits cacheRead, which dominates long sessions).
3. Graduated eviction on the agent/pre-step waterfall (before every LLM call), from fine to coarse:
   • content stubbing (fine): large tool-result contents in expl episodes are rewritten to a short stub first ([cwl-stub: …]) — structure kept, tokens cut, tool pairing intact
   • whole-episode eviction (coarse): expl episodes first (pure context, one-line "explored: …" marker), then completed act episodes; executed as positional blocks in the surface (positions are the invariant that survives replaces — an eviction never splits a tool-call/result pair into orphans)
   • never touch the newest tail (preserve-recent) or user messages
   • evicted ranges are replaced with a lightweight marker via the official surface-replace seam (original events stay in the log; cwl_recall can restore file paths)

Install

dsh plugin --profile <name> add dsh-cwl                 # from npm
dsh plugin --profile <name> add github:kalifun/dsh-cwl  # or from GitHub

Or vendor the directory and add to your composition:

- id: dsh-cwl
  name: ./dsh-cwl/index.js

Usage

No configuration needed. It stays completely inert while context is under budget (default 80% of the model's context window), and starts evicting only when pressure exceeds budget.

# Optional: override the budget (tokens) — for testing pressure behavior
DSH_CWL_BUDGET=30000 dsh web

Eviction policy (deterministic cache-replay validation: eviction −24% cacheRead, strategy-independent; batch best mean −24.7%, consistent across 7 sessions → defaults below; override via env):

# back to the conservative config (oldest, per-episode replaces)
DSH_CWL_EVICT_ORDER=oldest DSH_CWL_EVICT_BATCH=0 dsh web

Session analysis (per-round token breakdown + "cacheRead of the round after an eviction"):

node tools/analyze-session.mjs <session.jsonl>

Agent-facing tools:

Observability:

Verification

node check.js          # pure-function unit checks (episode inference, eviction policy, strip, pairing)

Live capability benchmarks (helmsman platform): BENCHMARKS.md — the fixed test plan (scenario A: 12-round long conversation; scenario B: single-request long autonomous task ×3) with per-version data rows, refreshed after every behavioral change.

Offline regression tools (run on your own local sessions — no data leaves your machine): tools/cache-replay.mjs (deterministic cacheRead), tools/replay-real.mjs --apply (engine apply-layer with real surface fold + tool-pairing assertion), tools/eval-episodes.mjs.

License

MIT