Skip to main content

Roadmap

This roadmap is directional. It explains product sequencing, not date commitments.

Shipped: local task flow and adapter files​

DevSpecs v1.1 centers on the local CLI:

  • ds task creates bounded task workspaces with packed context;
  • ds apply emits one-slice prompts without launching an agent;
  • ds map, ds recent, and ds find provide the diagnostic trust layer;
  • ds init can write small repo-local files for Codex, Cursor, Claude, or Windsurf that route agents back through ds task and ds apply.

The current generated entry points are $ds-task / $ds-apply for generic agent skill files and /ds-task / /ds-apply for Cursor, Claude, and Windsurf-style files.

Short-Term: adopt existing intent​

Repos already have useful PRDs, RFCs, ADRs, plans, specs, and agent notes. ds adopt should make those artifacts addressable without asking teams to rewrite them.

Expected shape:

  • detect candidate intent artifacts during map/find flows;
  • show a dry run before writing anything;
  • create thin wrapper artifacts for addressability and metadata;
  • leave original files untouched by default;
  • make wrappers easy to version when they should become durable team context.

Short-Term: richer agent wrappers​

Initial adapter files are shipped through ds init. The next step is polish:

  • clearer tool-specific recipes for task creation and apply prompts;
  • better generated instructions for one-target work;
  • safer defaults for agents that restart shells or run inside sandboxes;
  • future local MCP adapter over bounded commands such as ds task next, ds task prompt, ds task checkpoint, ds map, ds recent, and ds find.

The CLI remains the source of truth. Integrations should wrap it, not create a parallel workflow.

Mid-Term: workspace support​

Many teams work across more than one repo. Workspace support should make cross-repo intent easier to search, address, and hand to agents without forcing every team into a single monorepo or shared folder layout.

Likely areas:

  • workspace-level config;
  • cross-repo task and intent views;
  • clearer storage choices for personal versus team artifacts;
  • safer series naming when multiple people create task tracks.

Long-Term: research-backed context quality​

We are actively researching better ways to identify the right source, test, and intent context from local evidence. The bar for shipping this into DevSpecs is high: improvements should be deterministic, explainable, and useful without requiring cloud sync.

Expected shape:

  • better context packing and scouting from local repository evidence;
  • clearer positive and negative evidence for why files were included or excluded;
  • public writeups when results are reproducible enough to support product claims;
  • no broad retrieval superiority claims until the evidence is narrow, current, and public-safe.