Voice
Context Lookup
By OpenClaw plugin: function-call style on-demand context lookup tool. Replaces auto-injection with explicit topic queries.
Configuration Example
{
"topics": {
"<name>": {
"file": "shared/foo.md",
"section": "Optional default heading substring",
"description": "Optional human-readable description",
"aliases": ["optional", "alternate-names"]
}
}
}README
# context-lookup
OpenClaw plugin that registers a single function-call style tool —
`context_lookup` — for agents to fetch shared documentation, tool guides,
and any other topic-indexed reference material **on demand**.
## Why
Auto-injection plugins (such as the predecessor `progressive-context`)
push large blocks of shared context into every agent prompt up-front,
even when the agent doesn't need it. `context-lookup` flips the model:
agents query by topic name, the plugin returns just that file (or a
single section of it).
This plugin **does not** touch OpenClaw core's standard bootstrap files
(`AGENTS.md`, `TOOLS.md`, `MEMORY.md`, `USER.md`, `IDENTITY.md`,
`SOUL.md`, `HEARTBEAT.md`) — core injects those natively each turn.
## Tool spec
```
context_lookup({
topic?: string, // topic name from registry (or alias)
section?: string, // optional heading substring to slice
list_topics?: boolean, // enumerate all topics
list_sections?: boolean // list headings of the resolved file
})
```
Returns one of:
- `{ topic, file, content, section?, truncated? }` — body (full file or
section). `truncated:true` means the response was clipped at
`maxBytes`.
- `{ topic, file, sections: [...] }` — when `list_sections:true`.
- `{ topics: [{ name, description, file, default_section, aliases }] }` —
when `list_topics:true`.
- `{ error: string }` — on failure (unknown topic, missing file, etc.).
Errors include the available topic list so the agent can self-correct.
The tool description registered with the host dynamically lists every
configured topic + its description, so agents see options without
needing to call `list_topics` first.
## Topic registry
Topics live in a JSON file (default: `topics.json` at workspace root,
overridable via `topicsFile` config). Schema:
```json
{
"topics": {
"<name>": {
"file": "shared/foo.md",
"section": "Optional default heading substring",
"description": "Optional human-readable description",
"aliases": ["optional", "alternate-names"]
}
}
}
```
Edit the registry freely; the plugin reads it on gateway startup. A
`_readme` key at top level is ignored.
## Plugin config (`openclaw.json`)
```json
"context-lookup": {
"topicsFile": "topics.json",
"toolName": "context_lookup",
"maxBytes": 64000,
"skipAgents": []
}
```
All fields optional — defaults shown.
## Build
```sh
npm run build
```
Produces `index.js` (CommonJS, bundled, `fs`/`path` external).
## Compared to `progressive-context`
| feature | progressive-context | context-lookup |
| --- | --- | --- |
| auto-inject standard 6 bootstrap files | filters/compacts | leaves alone (core handles it) |
| inject shared on first turn | yes (`injectSharedOnFirstTurn`) | no |
| inject shared every turn | yes (`injectSharedEveryTurn`) | no |
| compact reminder turn 2+ | yes | no |
| on-demand `workspace_context` tool | yes (path-based: `workspace:<key>`, `shared:<name>`) | replaced by `context_lookup` (topic-based) |
`context-lookup` does only the last row — and does it via a topic
registry instead of file paths, so the agent sees semantic names.
voice
Comments
Sign in to leave a comment