Skip to content

Latest commit

 

History

History
54 lines (44 loc) · 2.68 KB

roadmap.md

File metadata and controls

54 lines (44 loc) · 2.68 KB

Roadmap

M0 — foundation (this init)

  • pnpm workspace with React/Vite and a reusable TypeScript Agents API client;
  • live same-origin Core connection to an independently deployed backend;
  • Agent list/create and a Session conversation vertical slice;
  • fetch-based SSE, durable Item recovery, cancel, and function-result handoff;
  • architecture, protocol coverage, and private self-iteration workspace rules;
  • a public, read-only agent-ready issue intake gate that emits a normalized task artifact without shipping private prompts or a privileged runner.

M1 — pinned Parsar Core acceptance

  • keep the Web contract fingerprinted to an immutable services/agents-api revision and record upgrade evidence;
  • add raw HTTP fixtures for every UI-used response/event variant;
  • harden same-ID durable/live Item precedence so a stale projection cannot replace a finalized durable value;
  • persist pending-operation idempotency keys across reconciliation-assisted manual retry decisions;
  • verify disconnect recovery, active steering, cancellation races, function action/result/error, pagination, and error presentation;
  • add browser E2E with the real core and a controlled executor.

M2 — runtime-capability integration

  • track upstream Parsar Core runtime and Environment work without implementing Core adapters in this repository;
  • define a namespaced, optional Web capability-discovery contract only when an upstream implementation is ready to expose it;
  • add compatibility fixtures and UI acceptance for allocation, connection, lease, recovery, cancellation, and cleanup behavior;
  • expose Docker, E2B, AWS Bedrock AgentCore Runtime, or another provider in Web only after the connected Core advertises a versioned capability and passes the relevant lifecycle tests;
  • report missing Core behavior as an upstream gap rather than emulating it in browser or client state.

M3 — issue-driven self-iteration pilot

  • connect a trusted runner to successful intake artifacts;
  • consume only structured issues carrying a maintainer-applied agent-ready label;
  • plan in an isolated branch/worktree, run repository checks, and open a draft PR;
  • require human review; never auto-merge, release, or expand issue scope;
  • keep prompts, scratch decisions, and run state in the ignored .agents/ workspace while committing only code, tests, and public design decisions.

M4 — separate Web Cloud service/product

  • introduce a BFF, user sessions, managed runtime inventory, audit, and secrets in a separate Web Cloud service rather than this open-source TS client;
  • design organization/role/billing features as a separate product layer instead of coupling them to the open Agents API client.