Run as a bifrost LLMPlugin¶

Embed the context-guru pipeline inside an existing bifrost deployment — no separate process, no transport of its own.

Steps¶

Build a pipeline and store from your config:

pipe, err := cfg.Build(emitter)   // cfg is a *config.Config
st := cfg.NewStore()

Construct the plugin and register it:

p := bifrost.New(pipe, st)
// add p to BifrostConfig.LLMPlugins

Wrap the transport with the expand loop if you need reversibility (see below).

Where it sits¶

PreRequestHook runs the pipeline over req.ChatRequest in place — the canonical mutate phase. Non-chat requests pass through. It never aborts a request; fail-open lives inside the pipeline.
Session id comes from the context-guru-session context value (set by the transport from the request header or Anthropic metadata.user_id), falling back to a content hash.
PreLLMHook / PostLLMHook are pass-throughs. The expand loop cannot be a hook — it must re-invoke upstream. Put it in a transport wrapper: run the pipeline, forward, then resolve <<cg:HASH>> markers with the shared expand/ package and the store before returning, capped at 3 rounds. That is exactly what the shipped proxy does.

Troubleshooting

Markers reach my clients unexpanded. The plugin alone does not recover anything — the expand loop lives in the transport wrapper, not in a hook. Either add the wrapper or run a lossless-only pipeline (format, toon, cachesplit) so no marker is ever written.

Every request lands in the same session. The transport has to set the context-guru-session context value. Without it the fallback is a content hash, which changes whenever the head of the conversation changes.

Non-chat requests are untouched. By design — PreRequestHook passes them through.

For the standalone proxy and the AuthBridge plugin, see Host adapters.

Edit this page