Datarim 1.15: Research Phase — Know Before You Build
Version 1.15.0 adds a structured research phase to the pipeline. A new researcher agent investigates library versions, best practices, and security advisories before planning begins — and fills knowledge gaps during implementation.
Every developer has been there: you plan a feature, start coding, and halfway through discover that the library you chose deprecated its core API two months ago. Or that two dependencies you picked are incompatible. Or that a similar problem was already solved in another part of the project.
Datarim 1.15.0 addresses this with a structured research phase that runs automatically inside /dr-prd. No new commands to remember. No extra pipeline steps to manage. Research is now built into the process you already use.
How It Works
When you run /dr-prd on a Level 2 or higher task, a new Phase 1.3 activates between context analysis and the discovery interview. A dedicated researcher agent investigates external context and produces an insights document.
The research runs through a structured checklist:
- Versions and dependencies — current stable releases of your stack
- Breaking changes — migration guides, deprecated APIs
- Best practices — recommended approaches from official docs and community
- Stack documentation — relevant sections loaded via context7 or web search
- Architectural patterns — reference implementations of similar tasks
- Compatibility — peer dependency conflicts, runtime compatibility
- Security advisories — known CVEs in your dependency tree
- Past experience — relevant lessons from previous tasks (via LTM, if available)
- Existing codebase — reusable components already in your project
- Infrastructure constraints — ports, resources, server limits
Level 2 tasks get a lighter version (5 checkpoints). Level 1 quick fixes skip research entirely.
Insights Document
Research results go into a dedicated datarim/insights/INSIGHTS-{task-id}.md file — separate from the PRD. This matters because the PRD is a frozen requirements document, while insights are a living artifact that grows during implementation.
The discovery interview that follows Phase 1.3 reads the insights document, so its questions and proposed answers are informed by actual research rather than assumptions.
Gap Discovery During Implementation
Research does not end at the PRD stage. When the developer agent encounters an unknown during /dr-do — an import failure, unexpected API behavior, documentation that does not match reality — it automatically spawns the researcher agent to investigate the specific gap.
Findings are appended to the insights document's "Gap Discoveries" section. If the gap is fundamental (wrong technology choice, impossible requirement), the agent stops and recommends returning to /dr-prd.
This closes the loop: research happens before building, and new unknowns trigger targeted research without leaving the implementation flow.
Adaptive Tooling
The researcher agent works with whatever tools are available in your environment:
- context7 MCP — token-efficient library documentation (preferred)
- WebSearch — version checks, CVE lookups, architectural patterns
- Scrutator LTM API — past task experience from long-term memory
- Codebase tools — always available: grep, glob, read for local analysis
No hard dependencies on any specific MCP server. If nothing external is available, the agent falls back to codebase-only research and flags findings based on training data as unverified.
By the Numbers
| Component | Before | After |
|---|---|---|
| Skills | 21 | 22 |
| Agents | 16 | 17 |
| Commands | 19 | 19 |
| Templates | 12 | 13 |
One new skill (research-workflow), one new agent (researcher), one new template (insights-template). Eight existing files updated. Zero new commands — research is integrated, not bolted on.
Design Decision: Inside PRD, Not a Separate Command
We considered adding a standalone /dr-research command. That would have required updating 5+ routing files, changing the pipeline topology, and adding cognitive load (operators would need to decide when to research vs when to proceed). Instead, research became an internal phase of /dr-prd — the command you already run for requirements. The pipeline stays the same. Research just happens.
Full Change List
- Research Phase 1.3 in /dr-prd for L2+ tasks (TUNE-0029)
- Gap Discovery Step 7.5 in /dr-do — auto-spawns researcher on unknowns (TUNE-0029)
- New researcher agent (sonnet model, adaptive tooling)
- New research-workflow skill (10-point full checklist, 5-point lite mode)
- New insights-template (7 sections including Gap Discoveries)
- /dr-plan auto-loads insights document as context
- Discovery interview reads insights before proposing answers
- Visual maps updated: pipeline routing, stage flows, dependency graphs
- path-and-storage updated: datarim/insights/ directory
- backlog-and-routing: embedded phases documented
See the changelog for the complete version history.