dsh-perf-patch
Runtime-only performance patches for DeepSeek Harness (DSH) when third-party plugins make
the host O(N) slow. Nothing in node_modules is edited: the patches wrap host services at load
time, and each one fails safe — on any doubt it falls back to the original code path instead
of changing behaviour.
Built and measured on DSH Desktop 0.2.1-rc.2 (Electron 44 / Node 24 / Windows 11) with
dsh-ops and dsh-mcp-connector installed.
What it fixes
1. The agent-mount tools/change storm
Mounting an agent registers ~161 tools. Every registration fires tools/change, and every
listener that reacts (discovery reconcile, PTC reconcile, MCP governance) replays its full
O(tools) work. One agents.create therefore emits hundreds of notifications, each listener
re-running hundreds of times.
Fix: while agents.create / agents.resume runs, tools.layers.onChange is suppressed and
the notification is emitted once, synchronously, when the outermost call returns.
Suppress + synchronous flush is deliberate. Flushing via queueMicrotask starves the event
loop (listeners queue more microtasks than the loop drains) and was observed to freeze the UI;
that variant was removed and is not shipped.
2. tools.schemas() deep-clones on every read
schemas(scope) clones every tool's JSON Schema; schemaOf(def, true) clones one. Both are pure
functions of the tool revision / definition, yet listeners call them in loops.
Fix: memoize schemas(scope) per tool revision (WeakMap keyed by scope, invalidated on
tools/change) and schemaOf(definition, true) by definition + parameter identity. Returned
objects are shared between callers, so treat them as read-only.
3. dsh-ops asks a one-boolean question with a full prompt assembly
dsh-ops/lib/presentation.js decides "is this agent on PTC presentation?" by calling
systemPrompt.assemble({scope, signal}) and then reading exactly one thing:
assembly.sections.some(s => s.name === 'tools:ptc-only' && s.text.trim() !== '')
assemble() renders ~36 sections and structuredClones every tool schema (~250–310 ms
measured). dsh-ops does this 6× per live agent, per reconcile.
Fix: wrap systemPrompt.assemble. When — and only when — the call is one of those probes,
return a minimal assembly built from the same layers:
{ sections: [{ name: 'tools:ptc-only', text }], contexts: [], tools: [] }
Probe detection uses dsh-ops' own marker (isPresentationProbe, a WeakSet reached through the
module cache); if that import is unavailable it falls back to "the stack contains
dsh-ops/lib/presentation.js". Every other caller — including the real prompt assemblies in
ptc/index.js and discovery/index.js — passes straight through.
The section text is recomputed on every call. Caching it was tried and reverted: a stale
control input turns reconcile into a fixed point that never settles, and freezes the host.
Measured on a 3-agent instance: per probe 254–312 ms → 0 ms (~4.9 s per reconcile sweep),
with the PTC decision byte-identical for all 3 agents.
Install
cd ~/.dsh/profiles/desktop
npm install ChenXu233/dsh-perf-patch
Then list "dsh-perf-patch" in the dsh.profile.bundles array of the profile's
package.json. The package ships dsh.bundle.patch → cordis.patch.yml, which inserts the
loader entry. Adding the name to bundles alone is not enough — without the patch layer the
plugin is never loaded.
From a checkout: clone this repo into <profile>/node_modules/dsh-perf-patch and add the name to
bundles the same way.
Kill switches
| Variable | Effect |
|---|
DSH_PERF_PATCH=0 | Disables all three patches at load time. |
DSH_PERF_PATCH_VERIFY=1 | Probe short-circuit also computes the full assembly and compares. On divergence it logs an error and permanently disables the short-circuit for the session. |
Runtime counters:
globalThis.__dshPerfPatchProbe // { hits, bypass, mismatches, enabled }
hits = probes answered short, bypass = calls that went to the real assemble. If hits
stays 0, detection is not matching and the plugin is a no-op.
Expectations
- Intended to become inert once upstream fixes land: the schema cache is keyed by revision
(correct regardless), and the probe short-circuit stands down when dsh-ops stops matching.
- The batch gate is the only behavioural change: listeners fire once per mount instead of once
per tool. That is the point, but a plugin relying on every intermediate
tools/change during
mount would now see one.
- Do not mutate objects returned by
schemas() / schemaOf().
Related upstream issues
- dsh-ops #333 — discovery reconcile recomputes presentation for every agent
- dsh-ops #369 —
tools.schemas() deep copy with 258 tools, linear slowdown
- dsh-mcp-connector #178 — governance reconcile cost
If this plugin causes a problem: set DSH_PERF_PATCH=0, confirm it disappears, and open an issue
with the numbers.
License
MIT
