Roadmap

A proposal for enterprise partners — adaptable per team’s needs.

Non-goals

Dandori is the outer harness. It does not build:

Non-goal	Owned by
Sandbox / filesystem / bash	Coding agent runtime
Sub-agent spawning	Coding agent runtime
Context window compaction	Coding agent runtime
Chat UI / IDE	Claude Code / Codex / Copilot
Model provider gateway	AI Gateway or provider direct
MCP server hosting	MCP Hub — Dandori governs, doesn’t host

Foundation (F1-F4) — must come first

#	Module	~Effort	Why first
F1	Service layer + REST API + Web UI + auth	1 sprint	Everything surfaces through it
F2	Audit Log + middleware	½ sprint	Retrofitting compliance later is painful
F3	Task Board CRUD (no DAG yet)	1 sprint	Root entity for runs, cost, approval
F4	Adapter layer + lifecycle event emission (Claude Code + run record)	1 sprint	Without runs, nothing has data; lifecycle hooks later plug into these events

End state: create task → spawn Claude Code → see output + cost in Web UI. Audit log records every mutation.

Milestones

M0 — Skeleton (1 sprint): F1 + F2. Web UI, audit log, API keys.

M1 — Single-task MVP (2 sprints): F3 + F4. Create task → agent runs → run record stored. Foundation complete.

M2 — Knowledge Flow (2 sprints, parallelizable):

Track A: Context Hub (5-layer assembly + versioning)
Track B: Skill Library (basic, no progressive disclosure)
Track C: Agent Templates (clone / customize / promote)

M3 — Dual-audience value (2-3 sprints, parallelizable):

Track A: Quality Gates + Cross-agent Analytics (+ sub-agent rollup)
Track B: Approval Workflow + Slack approvals
Track C: Cost Attribution dashboard
Track D: Fleet Operations Dashboard

End of MVP. Both engineers and leadership get real value.

M4 — Advanced orchestration + evaluation (2-3 sprints): Task DAG + auto-wakeup + phases. Codex adapter. Evaluation Suite (golden task sets + regression detection).

M5 — MCP + platform primitives (3-4 sprints, parallelizable): Built-in MCP server (pre-req), then: Inline Sensors, Tool Governance, Skill progressive disclosure, user-configurable Lifecycle Hooks (promote F4’s infrastructure emission into org-wide, versioned hooks).

Full outer harness. 13 user-facing features.

M6 — Ecosystem integrations (2-3 sprints): Jira, Confluence, GitHub Enterprise, Copilot MCP registration. (Google Drive + Slack already in M3.)

Three paths

Path	Scope	Timeline (1 dev / 2-3 devs)
Minimal pilot	M0 → M1 → M2-A → M3-B → M6 (Jira+Slack)	~3 mo / ~2 mo
Full vision	M0 → M1 → M2 → M3 → M4 → M5 → M6	~11 mo / ~6 mo
Enterprise	Full vision + AI platform integration	~14 mo / ~8 mo

Stop at any milestone — each delivers standalone value.

Per milestone we deliver

Working code merged to main
Demo session (engineers + leadership invited)
Audit log evidence of usage during milestone
Retrospective report

What we ask from partners

Pilot team of 5-15 engineers (already using Claude Code or Codex)
Read-only access to existing tools (Jira, Confluence, GitHub, Drive, Slack)
A sponsor in engineering leadership
Weekly 30-min sync for fast feedback
Honest feedback — “this doesn’t help me” is most valuable

Customization

This roadmap is a proposal, not a contract:

Skip modules the partner doesn’t need
Promote urgent modules (e.g., Cost Attribution before Skill Library)
Add integrations not in M6 (e.g., Linear instead of Jira)
Pause between milestones for adoption
Combine milestones with multiple devs

First conversation: “what does your team need first?” — roadmap re-orders around the answer.