@jiesou/dsh-stream-rules
在需要时注入规则，避免浪费上下文。opencode-stream-rules 的 DSH 移植版。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @jiesou/dsh-stream-rules@0.1.7
Stream Rules 以 @jiesou/dsh-stream-rules 发布，当前版本为 0.1.7。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Inject rules when needed, without wasting context.
-6336866371853030306_121You can write custom streaming rules for the agent.
These rules are injected only as a steering notice after a pattern match, then agent retry from the same point. This allows you to control the boundaries of agent behavior, without wasting context.
Port of my jiesou/opencode-stream-rules to DSH. Similar to oh-my-pi's "Time-traveling stream rules", but with a very simple and compact code implementation.
A rule fires on a tool call (tool name + serialized arguments) when its match returns true:
SYSTEM NOTICE steering message into the agent via agent.inject() (DSH's non-waking "queue model-facing context for the next pre-step"). The agent retries from the same point, now knowing the rule.reject: true — denies the FIRST tool call ({ kind: 'deny' }); later attempts are allowed. Steering without over-restricting, e.g. letting pip install through when it's already in a container.Each rule fires at most once per session (per agent), mirroring the original's notified dedup.
From npm (prebuilt, recommended):
dsh plugin --profile <name> add @jiesou/dsh-stream-rules
Or from GitHub (runs prepare to build on install):
dsh plugin --profile <name> add github:jiesou/dsh-stream-rules
Or add the row to your profile's cordis.patch.yml:
- id: stream-rules
name: '@jiesou/dsh-stream-rules'
You need to write the rules in your own .js file. This plugin won't work by default until you edit the rules.
$DSH_HOME/profiles/<name>/node_modules/@jiesou/dsh-stream-rules
where $DSH_HOME defaults to ~/.dsh.
mv rules/rules.js.example rules/rules.local.js
_ are skipped.config.rules:- id: stream-rules
name: '@jiesou/dsh-stream-rules'
config:
rules: /path/to/your/rules
reject: true will only be rejected on the first toolcall; subsequent attempts by the agent will be allowed. This provides steering while avoiding overly restricting the model (e.g., allowing pip install if it's already in a container).// rules/rules.local.js
export default [
{
match: (v) =>
v.includes('pip') &&
v.includes('install') &&
!v.includes('uv pip') &&
!v.includes('uvx'),
reject: true,
prompt: 'Use `uvx` or `uv venv` + `uv pip` instead of `pip install` directly',
},
{
match: (v) => v.includes('curl') && v.includes('api.github.com'),
prompt: 'Prefer using `gh` cli over `curl https://api.github.com/...`. gh offers more requests limits.',
},
{
match: (v) => v.includes('pdf'),
prompt: 'Use the `markitdown` skill to read PDF files.',
},
// add your rules here
]
| field | required | description |
|---|---|---|
match | ✅ | (v: string) => boolean; every tool call is flattened to a string and matched |
prompt | ✅ | The prompt for steering |
reject | If true, prevent the tool call first, instead of just steering |
src/index.ts (~60 lines).tools/pre-execute waterfall (deny) and agent.inject() (steering), the documented native extension points. No core changes, no monkey-patching.