Roadmap
BotMinter is developed through incremental milestones. Each milestone builds on the previous one, adding features to the CLI and validating them with synthetic test tasks before operational use.
Pre-Alpha
BotMinter is under active development and not yet ready for production use. The information on this page reflects the current plan and is subject to change.
Overview
| Milestone | Status | Description |
|---|---|---|
| Structure + human-assistant | Complete | Profile skeleton, scrum profile, workspace model, first member |
| Autonomous Ralph | Complete | Spike validating persistent loop pattern for pull-based agents |
| Architect + First Epic | Complete | Second member, epic lifecycle, two-member coordination |
| GitHub Integration | Complete | Replaced file-based coordination with real GitHub via gh CLI |
bm CLI |
Complete | Rust CLI, single operator interface, workzone model |
| Minty and Friends | Complete | Chief of Staff role, profile externalization, Minty assistant |
| Team Bridge | Complete | Bridge plugin system, Matrix (default) + Telegram + Rocket.Chat bridges |
| Full Team + First Story | Future | Dev, QE, reviewer members, full story kanban, TDD flow |
| Eval/Confidence System | Future | Formalized eval framework, scored confidence, HIL graduation |
Completed
Structure + human-assistant
Built the foundational infrastructure:
- Profile skeleton — process-agnostic directory structure
scrumprofile — PROCESS.md, CLAUDE.md, team knowledge/invariants- human-assistant member skeleton with three-hat model
- Workspace model — clone project, embed team repo, surface config
- Human-in-the-loop (HIL) via RObot — validated during development; available as an optional bridge on any profile
Proved: Inner loop works (Ralph + hats). Workspace model works (clone, surface, run). HIL validated (human <> human-assistant via Telegram).
Autonomous Ralph
Spike validating how Ralph runs autonomously in a persistent loop, pulling work from a board instead of a single-objective prompt:
persistent: truewithtask.resumerouting- Self-clearing scratchpad/tasks between work items
- Idle behavior via
LOOP_COMPLETE
Deliverable: validated ralph.yml pattern adopted directly by later milestones.
Architect + First Epic
Added the architect as a second team member and validated two-member coordination:
- Architect member skeleton (ralph.yml, PROMPT.md, CLAUDE.md, five hats)
- human-assistant evolution — new hats for epic creation, design gating
- Epic lifecycle statuses in PROCESS.md
- Two-member outer loop coordination validated with synthetic epics
Proved: Outer loop works (GitHub issues, status labels, knowledge resolution). Pull model works (architect picks up work via status watch). Two-member coordination works (PO creates, architect designs, PO gates).
GitHub Integration
Replaced file-based coordination with real GitHub (pulled forward from original planning):
ghCLI calls replaced file operations (1:1 mapping)- Unified
ghskill as single interaction point - Coordination model unchanged — only backing store moved
bm CLI
Replaced Justfile-based tooling with a Rust CLI binary (bm):
- Single operator interface for managing agentic teams
- Workzone model with known, discoverable workspace directory
- Profile restructuring — collapsed into single layer with
botminter.yml+.schema/ - Profiles shipped with the binary, extracted to disk on first use (
bm profiles init) - Event-driven daemon with webhook and poll modes (
bm daemon start/stop/status) - Knowledge management commands (
bm knowledge list/show+ interactive mode) - Formation-aware deployment (
bm start --formation) - Commands:
bm init,bm hire,bm start(alias:bm up),bm stop,bm status,bm teams list/show/sync,bm members list/show,bm roles list,bm profiles list/describe,bm projects add/list/show/sync,bm knowledge list/show(+ interactive mode),bm daemon start/stop/status,bm completions - Integration test suite covering full lifecycle, hire, sync, schema guard, multi-team
Proved: CLI-driven team management works. Versioned profile model enables future upgrades. Workzone model provides discoverability.
Minty and Friends
Multiple UX enhancements to improve the operator experience:
- Chief of Staff role — a new team-scoped role for process improvement tasks, operating independently from dev workflow
- Profile externalization — profiles extracted to disk on first use, editable and customizable without rebuilding the binary
- Workspace repository model — dedicated git repo per agent with submodules for team repo and project forks, replacing the earlier embedded workspace model
- Minty — BotMinter's interactive assistant persona with composable skills
- Coding-agent-agnostic cleanup — abstracted Claude Code-specific assumptions from profiles and CLI
Proved: Role-as-skill pattern works. Skill-driven architecture viable. Disk-based profiles enable operator customization. Stack is coding-agent-agnostic end-to-end.
Team Bridge
Bridge plugin system for connecting team members to messaging platforms:
- Bridge plugin contract — Knative-style resource format (bridge.yml + schema.json + Justfile), no Rust code needed for new bridges
- Matrix / Tuwunel bridge (local, default) — single Podman container with Tuwunel homeserver, pre-selected during
bm init - Telegram bridge (external, experimental) — operator-managed bot tokens, per-member identity
- Rocket.Chat bridge (local, experimental) — Podman pod with RC + MongoDB, auto-provisioned bot accounts
- Bridge CLI —
bm bridge start/stop/status,bm bridge identity add/show/rotate/remove,bm bridge room create/list - Per-member start/stop —
bm start <member>andbm stop <member>for individual member lifecycle - Credential storage — system keyring with env var fallback, formation-aware via CredentialStore trait
- E2E test coverage — operator journey scenarios for Matrix and Telegram bridges
Proved: Bridge plugin model is extensible. Local bridges can self-provision. Credential management works across formation types.
Future
Full Team + First Story
Adds dev, QE, and reviewer as team members:
- Dev, QE, reviewer member skeletons
- Full story kanban statuses in PROCESS.md
- TDD flow: QE writes tests, dev implements, QE verifies, reviewer reviews, architect signs off, PO merges
- Codebase access model (project fork, agent-cloned)
- First real knowledge accumulation
Eval/Confidence System
Formalizes the distributed eval framework:
- Eval framework across recursive scopes (team, project, member, member+project)
- Hard-gate vs advisory eval distinction
- Scored confidence metrics
- Evidence chain verification
- HIL graduation path (training, supervised, autonomous)
Future ideas
- Extract human-assistant from profiles (infrastructure, not a team role)
- Access control per scope + SLSA-style attestation for PRs
ralph.ymlhot-reload (eliminate need for sync + restart)- Knowledge observation mechanism for automatic knowledge capture
- Hat-level skill filtering (pending Ralph runtime support)
Related topics
- Architecture — profile-based generation model and two-layer runtime
- Member Roles — current and planned role definitions
- Bridges — bridge types and plugin contract