Account
Analytics and large exports
Inspect hosted ContextStream operating signals and run paged exports or aggregates when the result is too large for one response.
These are hosted-product capabilities for administrators, support work, and large data jobs. They are not part of the normal local connection, and many people will never need them.
Operational analytics
Hosted ContextStream can report a small allowlisted set of operating signals:
| Signal | What it helps answer |
|---|---|
| Request latency | Are ContextStream-backed operations getting slower? |
| Acceleration cache hit rate | How often is the faster cached path serving a request? |
| Degraded provider rate | How often is a fallback or degraded provider path being used? |
| Archive search health | Are older retained records still searchable? |
| Signal delivery health | Are indexing and operation signals arriving successfully? |
These are scoped analytics, not arbitrary database queries. You can choose a time range and supported rollup interval, then narrow by an allowed dimension such as the operation being measured.
Use them when diagnosing a slow or degraded ContextStream experience. For the health of your own source code—cycles, unused code, complexity, drift, and dependency impact—use Code health and dependencies instead.
Diagnose a slow afternoon
Suppose search and recall feel slow between 14:00 and 15:00, but only for some requests. Start with request latency for that window, then compare acceleration cache hit rate and degraded-provider rate over the same interval.
Those signals answer different questions. Higher latency with a lower cache hit rate suggests more requests took the primary path. A rise in degraded-provider rate suggests a fallback was involved. Normal archive-search health tells you the slowdown was not specifically caused by older retained records.
This is diagnostic evidence, not an automatic root cause. Narrow by an allowed operation dimension, compare with a nearby healthy interval, and preserve the time range in the support or incident record. The charts cannot run arbitrary queries over application data, and they do not report the health of your own codebase.
Exports that do not fit in one response
A large export can run asynchronously over four durable collections:
- captured transcripts;
- decisions;
- lessons;
- documents.
The job is submitted first, then checked until it completes. Results are read in pages so a large workspace does not have to fit into one assistant response. Transcript exports can omit full message content when only metadata and summaries are needed.
This is different from a capsule. A capsule is a curated, policy-controlled context package for a recipient. An export is a data job over a selected collection.
Export for an audit without over-collecting
For a quarterly architecture review, you may need every decision and lesson for one project, but not private transcripts. Submit separate scoped exports for decisions and lessons, check each job until complete, and read every page before calling the collection finished.
For a retention audit, transcript metadata and summaries may be enough. Omit full message content when the audit does not require it. That reduces the amount of conversation data handled without pretending the transcript records do not exist.
If the question is only “how many decisions were recorded each month?”, use a monthly aggregate. Exporting every decision and counting it elsewhere adds data handling without improving the answer.
Aggregates
The same hosted job system can aggregate one of those collections into hourly, daily, weekly, or monthly buckets. Use that for questions such as “how many decisions were recorded each week?” without streaming every underlying record.
An aggregate is not an AI summary. It groups records by time; the source records remain the authority for what they say.
Availability and permissions
Analytics and large jobs appear only when the connected hosted or remote deployment enables them. A local standard-input MCP process can be completely healthy without exposing either capability.
Every query remains scoped to the authenticated workspace and, when selected, project. Exporting data does not bypass personal records, team membership, or the retention and access rules described in Privacy and what gets indexed.
If it doesn't work
The capability is not present. The current connection may be local, the hosted layer may not be enabled, or your account may not include it. This is an availability question, not an indexing failure.
A job is still running. Large exports are intentionally asynchronous. Check its progress rather than submitting the same job again.
The result stops after one page. Continue from the returned position until there are no more records.
A chart is missing. Hosted deployments can expose only a subset of the allowlisted charts. List what the current deployment provides.