yoza10635/dsh-argp
C3Guarded context compaction for DeepSeek Harness (dsh): the LLM proposes, deterministic guards dispose — eager per-atom shrink (extract/summary/false under verbatim guards) + lazy reference-graph evict
★ 5+ · yoza10635/dsh-argp source on GitHub · this plugin in the registry
yoza10635/dsh-argp is a DeepSeek Harness plugin rated C3 — powerful capability combined with sensitive behavior. It patches the dsh runtime, can rewrite the system prompt, executes system commands, uses eval / new Function.
Installable plugin — declares a dsh.bundle manifest
What it can do
| Capability | Flag | Evidence |
|---|---|---|
| patches the dsh runtime | runtime_patch | ./cordis.patch.yml |
| can rewrite the system prompt | prompt_surface | systemPrompt |
| executes system commands | exec | ×8 in authored code, e.g. spike/07-real-trace.ts:27, spike/08-game-trace.ts:32 |
| uses eval / new Function | eval | ×1 in authored code, e.g. spike/08-game-trace.ts:77 |
| starts a network server | net_server | ×1 in authored code, e.g. spike/llm-log-proxy.mjs:16 |
Services it injects
locale slots systemPrompt tools
Hooks it attaches
agent/pre-step agent/request agent/request-error agent/status agent/turn-stopping session/event
Outbound domains
api.deepseek.com api.example.com
How to read this
Levels measure capability surface and transparency, not maliciousness. A C3 plugin can be entirely legitimate — a desktop shell genuinely needs subprocesses. The point is that you can see this before installing. See the levels explained and how dsh plugins work.
Findings come from static analysis of shipped code; nothing is executed. Think a flag is wrong? Open an issue — every flag cites the file and line it came from.