devspecs

Persistent context for coding agents

Keep AI coding work usable after the chat ends

Persistent coding-agent context is the durable, repo-local record of why work exists, what evidence supports it, where it stands, and what should happen next.

Direct answer

The useful memory is the part your repository can keep

Agent sessions are good at holding a working conversation. They are a fragile place to keep the reason for a change, the scope an agent was given, or the evidence behind the next decision.

DevSpecs indexes the intent and engineering evidence already around your code, then keeps bounded plans, results, and checkpoints alongside it. A returning human or agent can recover the thread without treating an old chat transcript as the project record.

Mechanism

One local thread across intent, code, and change history

DevSpecs builds a local SQLite index from the repository you already use. Source files stay authoritative; durable task records can stay versioned with the code.

Intent
Plans, ADRs, PRDs, RFCs, specs, runbooks, and project notes preserve why the work exists.
Evidence
Source, tests, documentation, and Git history show where the change lives and what has actually moved.
Lifecycle
Task slices, results, checkpoints, and decisions keep attempted work attached to its current state and next gate.
Handoff
Recent, find, task, and apply turn the local record into a focused continuation for the next human or coding agent.

Product proof

Return to a topic, not a blank prompt

This site uses DevSpecs itself. After the search-page work moved through several commits, one query recovered the current foundation and the decisions that led to it.

ds recent "search workflow pages"
  • Search Page Foundation · 2a244c4 · 11 files
  • Align Pages Search Intent · d296444 · 5 files
  • Approve Search Page Intent · 0a1191c · docs/growth/seo-page-intent-map.md
  • Next: inspect the current task state before continuing

Boundaries

What stays outside the record

  • DevSpecs does not capture every private prompt, tool call, or hidden model state.
  • It does not observe an agent automatically; humans and agents record lifecycle evidence through the CLI.
  • It does not upload your code, call an LLM, or require cloud sync.

Continue

See the workflow, then use the exact docs.

Review an agent's work after returningMatch a physical diff to its recorded goal, evidence, and human gate.Keep an agent inside one bounded changeCreate explicit slices and checkpoint the evidence before moving on.Resume a structured task after time awayRecover the current slice, last decision, and planned follow-up.

Try DevSpecs

Keep the thread in your repository.

Install the local CLI. Use the command docs when you need exact flags and lifecycle details.

Install CLIRead command docs