dsh-context-proxy

Thin on-demand context-retrieval layer for the DeepSeek Harness.

Automatic truncation, spill-to-disk, compaction, and token metering are already owned by output-retention, spill-policy, compaction-basic, and session-query. This package adds only the missing piece: three model-facing tools that read already-persisted history back on demand.

Install

dsh plugin --profile <name> add github:EvilIrving/dsh-context-proxy

Or, from a checkout:

dsh plugin --profile <name> add ./dsh-context-proxy

The bundle patch inserts one plugin row (dsh-context-proxy). sessionQuery and subprocess are optional backends (ctx.get): without a sessionQuery backend each tool degrades to an isError result instead of waiting forever.

Tools

Citations are replay-safe: context_query/context_slice cite session:<seq> (rebuildable from the canonical log), and context_grep cites either a spill path (path:line, fast local path) or session:<seq> (semantic-text fallback when no spill path is supplied or available).

Config

export interface Config {
  readWindowDefault: number  // default before/after for context_slice (default 0)
  grepMaxBytes: number       // rg stdout in-memory cap (default 1 MiB)
  rgGraceMs: number          // rg termination grace (default 5000)
}

Dependencies

• sessionQuery and subprocess are optional services read via ctx.get(...). A missing sessionQuery backend degrades each tool to an isError result (session query backend unavailable) instead of leaving the plugin waiting forever; the same applies to subprocess for context_grep with an explicit path.
• tools is a hard dependency (inject).
• The packaged ripgrep binary comes from @vscode/ripgrep — the same mechanism dsh-tool-fs-search uses — so no system rg install is required and no shell layer exists between the argv vector and ripgrep.

Model Experience

Request context and condition

What the model sees

Three tool schemas (context_query, context_slice, context_grep) are registered into the agent's tool set. Each result is a compact JSON object of matches/items plus citations; no prose is returned for the model to parse ids out of.

Token effect

Zero-direct effect until a tool is invoked; each invocation's cost is bounded by the query/slice window and the grepMaxBytes cap.

KV Cache effect

Append-only: tool results append to the surface; they never rewrite earlier request tokens.

Known Limitations and Deferred Work

• context_grep line parse is path:line:text — a spill path containing a newline (not expressible in the durable locator) would confuse the parser; absolute spill paths never contain one.
• Fallback text scan is literal, not regex — when no spill path is supplied, context_grep falls back to sessionQuery's literal, case-insensitive, whitespace-flexible scan, not a regex search.
• Bundle writing is not sandboxed — nothing here writes files; retrieval only reads through the seams.