feature: Help Context Guru end users understand what Context Guru does for them

Feature Description

We can now use the rossoctl CLI to run Context Guru locally on a laptop without Docker or Kubernetes access.

When I run

LOG_LEVEL=debug rossoctl authbridge exec \
   --config https://raw.githubusercontent.com/rossoctl/rossoctl-cli/refs/heads/main/examples/context-guru-tls-bridge.yaml \
   -- claude "explain this repo"

... I see a lot of interesting items in /tmp/authbridge.log, for example

time=2026-09-01T17:30:21.900-04:00 level=INFO msg="context-guru compacted request" binary=rossoctl-cortex-exec session=default tokensBefore=3322 tokensAfter=2608 tokensSaved=714 pctSaved=21.5% components=[extract]
time=2026-09-01T17:30:21.901-04:00 level=INFO msg="context-guru rewrote request body" binary=rossoctl-cortex-exec provider=anthropic path=/v1/messages session=default bytesBefore=143305 bytesAfter=140935 pctSaved=1.7%
time=2026-09-01T17:30:35.477-04:00 level=WARN msg="forward-proxy: text/event-stream response with WritesBody plugin — falling back to buffered path" binary=rossoctl-cortex-exec host=ete-litellm.ai-models.vpc-int.res.ibm.com

This gives a developer a hint that perhaps costs were reduced by 21.5%, or perhaps 1.7%, or perhaps something in the middle. However, this raw log isn't easy to understand and I wouldn't know what Context Guru is doing for me or provide justification for use.

Proposed Solution

Provide more information to end-users and enterprises about the user benefit.

• This could be as simple as logging statistics for the entire lifetime at plugin shutdown
• This could be sending metrics via OpenTelemetry about token savings and volume
• This could be a Context Guru-specific log that only logs things an end user should be aware of

Want to contribute?

• I would like to work on this issue.

Additional Context

The warning

time=2026-09-01T17:35:54.628-04:00 level=WARN msg="forward-proxy: text/event-stream response with WritesBody plugin — falling back to buffered path" binary=rossoctl-cortex-exec host=ete-litellm.ai-models.vpc-int.res.ibm.com

... is troubling. I don't know if this comes from Context Guru or some other AuthBridge Component. Would it make sense to log the plugin name or that the log does not come from a plugin, to help the deployer understand what has gone wrong?

cc @huang195

Start by reproducing the rossoctl authbridge exec example and reviewing /tmp/authbridge.log to identify which messages come from Context Guru versus other AuthBridge components. The issue needs a maintainer decision on whether the finished work should provide shutdown statistics, OpenTelemetry metrics, user-facing logs, or clearer plugin attribution before implementation can begin.