What happened across your projects this week.
"Use flex: what did we build this week?"Claude reads the schema, writes SQL to group sessions by project, drills into user prompts and file hotspots, and comes back with a narrative — which projects were active, what files got the most edits, what the key sessions accomplished. A weekly digest runs about 16 queries behind the scenes. A file lineage question might take 25. A quick lookup might take 7.
Who created a file, what changed, what it became.
"Use flex: what's the history of worker.py?"Every file operation is indexed with timestamps and session context. Claude finds 531 operations across 67 sessions, then reconstructs the arc — the predecessor that was broken for weeks, the rewrite that moved it into the module system, the package rename, the 14-session feature sprint, the stabilization phase. It identifies the heaviest session and surfaces the conversation context around each change. Files are tracked across renames automatically.
What happened in a session — tool calls, files, agents, decisions.
"Use flex: tell me everything about session 7a4f427"Every session is a unit of work — tool calls, user prompts, agent delegations, files touched. Claude returns a phased narrative — context loading, then a 5-agent fan-out for target discovery, then enrichment with 4 more agents, then a conference submission drafted at the end. Tool breakdown, file counts, duration, key findings. It detects the phases from tool call patterns and groups them into the arcs of work that actually happened. Sessions persist even after Claude prunes the original files.
Sessions grouped by repo, time, and work patterns.
"Use flex: what did we accomplish this week?"Flex maps sessions to repos automatically. Returns sessions grouped by day, tools used, files touched, and project breakdown. Flex detects work sprints — natural blocks of focused work separated by 6+ hour gaps — and identifies hub sessions — the ones that connect the most ideas.
Why it was done this way.
"Use flex: how did we create the curl install script?"The hardest question in software: why was it done this way? Claude finds the session where the decision happened, then reconstructs the path — the Mac test that revealed a name collision, the three install approaches considered, which one failed and why, the validation against other curl installers in the ecosystem, and the deployment to Cloudflare Pages. The reasoning lives in the conversation. Flex finds it.
Find sessions by meaning, not just keywords.
"Use flex: tell me 5 things we talked about this week outside the main project"Two retrieval modes. Semantic search finds sessions by meaning — it searches across everything indexed. Keyword search finds exact terms. Combine both for hybrid results. Claude returns the outliers — competitive intelligence sessions, debugging a separate project, meta-cognition on the project's own origins. It can also filter to only your messages, search by exact term, or explore the topic communities that emerged across your work.
Search tokens let Claude reshape results per query:
- diverse — spread results across subtopics instead of returning 10 variations of the same answer
- suppress:oauth — suppress a dominant theme so the edges surface
- decay:7 — weight toward the last 7 days with exponential decay
- from:single file to:modular system — find content along a conceptual arc
These compose freely. diverse suppress:oauth decay:7 is three operations composed in one query. Claude selects which tokens to use based on the question.
Named queries that ship with Flex.
Presets are discovered automatically when Claude reads the schema — it uses them when they fit. You rarely need to name them directly.
| Preset | What it does | Parameters |
|---|---|---|
@orient | Read the database — schema, available queries, presets | |
@digest | Multi-day activity summary — sessions, tools, files | days (default: 7) |
@file | Find sessions that touched a file (tracks renames) | path (required) |
@story | Session narrative — timeline, artifacts, agents | session (required) |
@sprints | Work sprints detected by time gaps | limit (default: 20), gap_hours (default: 6) |
@genealogy | Trace a concept's lineage — timeline, hubs, key excerpts | concept (required) |
@delegation-tree | Recursive parent → child agent tree | session (required) |
@bridges | Cross-community connector sessions | |
@file-search | BM25 search over file content (Write/Edit/Read) | query (required) |
@health | Pipeline health check — counts, coverage, embedding freshness |
"what did we do?" is too broad. "what did we build this week in the auth module?" gets better results. Name files when you can — file lookups are more precise than semantic search.
For complex research, delegate to trace — a retrieval sub-agent that runs multiple queries without filling your context window.
"Use trace: reconstruct how the deployment pipeline evolved"You can also query from the terminal without Claude:
$ flex search "@digest days=3"
$ flex search "@file path=src/worker.py"
$ flex search "SELECT COUNT(*) FROM sessions WHERE project = 'myapp'"