Product roadmap
From governed SDD core to a local agentic development control plane.
The roadmap deliberately keeps workflow authority, evidence and human governance ahead of autonomy. Later capabilities must preserve the stable core rather than bypass it.
0.14
Control Plane & Guided Studio
Local App Server, guided setup, task timelines, Review Center, Cursor ACP, Codex app-server, project checks, themes, help, validation discovery and deterministic risk assessment.
0.15
Verified Agent Harness
Context handles, ExecutionRuntime, validation integrity, eval registry v0, execution policy outside the model, and measurable skill/handle comparisons.
0.16
Governed Agent Teams
Plan-DAG serial/parallel waves, file-disjoint leases, claim/fan-in and reviewer/tester as contracts. DynosAI does not spawn extra provider processes.
0.17
Eval Intelligence
Failure attribution, local-trace mining, inbox-only improvement tasks and offline regression evidence. Predictive routing stays in shadow mode. No live-provider leaderboard.
0.18
Secure Autonomous Runtime
Strict/Balanced/Autonomous host-owned profiles, runtime-only secret vault and policy evidence. Human gates stay required. Docker/VM and OS-level network interception are not shipped.
0.19
Ecosystem
Capability manifests for the existing Cursor ACP and Codex app-server transports. Additional clients and extension packs are refused, not silently certified.
1.0 RC1
Contract freeze
MCP names, public CLI, schema v6 and Studio compatibility are documented and tested. Optional harness switches degrade safely. This is not production-ready 1.0: no installer, no live 1.0 matrix, no OS sandbox.
1.0 RC2
Certification evidence
MATRIX_1.0 cells exist and stay not_run. Two host-opened sessions can claim file-disjoint leases. Historical 0.13 Quality 100 is not copied forward.
1.0 RC3
Eval maturity
Acceptance ZIP importer, governed-change cost aggregates and a stable authority prefix. Unused MCP tools are not waste. Still not production-ready 1.0.
1.0 RC4
Installation
Studio doctor, verified local wheel (not PyPI), installed-package CI on Ubuntu/Windows/macOS, and 0.19.0 schema-v6 upgrade preservation. Still not production-ready 1.0.
1.0 RC5
Freeze
Threat model (enforced vs not enforced), contribution/certification process, and Studio loopback honesty. External PRs stay closed. Still not production-ready 1.0.
1.0 RC6
Standards & Release Evidence
MCP 2026-07-28 plus legacy 2025 hosts, unified release gate, eval corpus, PyPI name ADR, Agent Plugins ADR. MATRIX_1.0 stays honest.
1.0 RC7
Certification Integrity
Archive safety, MCP 2026 strictness/MRTR, MATRIX runner honesty. Live Codex greenfield failed because managed CODEX_HOME sat under OS temp.
1.0 RC8
Provider Runtime Portability
User-local workspaces, Codex temp-home abort, candidate identity. Live Codex greenfield failed: App Server started, DynosAI MCP did not.
1.0 RC9
Codex MCP Observability
Deterministic App Server MCP preflight without model turns. Opaque Codex _meta no longer aborts a 2025 session. Live Codex cells reached implementing then work_not_done.
1.0 RC10
Governed Session Continuity
Bounded multi-turn Codex trials in one App Server session. Codex is Certification pending. Cursor is Preview and not 1.0 certified. No RC11 is created automatically.
1.0.0
Stable Governance Core
Promotion with minimal delta from the accepted RC. Requires green gates, live Tier-A matrix, current MCP compatibility and coherent public surfaces. No features during promotion.
1.1
Enforced Secure Runtimes
Move filesystem, network, credential and process-tree controls from logical decisions to real runtime backends. Stay local-first. No mandatory Docker for Core.
1.2
Portable Agent Plugins
Consume Agent Plugins (plugin.json + extensions.com.dynosai). Local directory and Git URL first. No marketplace. No silent authority.
1.3
ACP Governance Gateway
Investigate DynosAI as an ACP-consumed governance layer in front of coding agents. Certify clients with greenfield, brownfield, refusal, gate and recovery tests.
1.4
Long-Horizon Governed Operations
Durable background work, pause/resume, quotas and parked review jobs. Not a generic swarm. Background never means auto-merge.
1.5
Eval flywheel & adaptive routing
Failure to eval to shadow recommendation. Predictive routing stays recommendation-only until evidence. No silent quality downshift.