v3.7.0 — M8: TUI Implementation

Goal: Implement the comprehensive Text User Interface (TUI) and all TUI-dependent features using Textual >= 1.0.

Scope:

Textual-based MainScreen with 3 sidebar states (conversation, permissions, settings)
Persona system (YAML-based actor configuration with export/import)
Reference/command input system (@, /, ! modes)
TuiMaterializer A2A integration layer bridging TUI to the application layer
Session persistence (SQLite) with multi-session tab support
Dracula theme with configurable display settings
Settings screen, content pruning, and safety behaviors
Shell danger detection for destructive command interception
Permission Question Widget for interactive approval flows
First-run experience with actor setup wizard
Prometheus metrics export and operational observability
Session export/import (JSON + Markdown formats)
ACMS graph backend integration (Blazegraph/Neo4j)
PostgreSQL backend support for multi-user deployments
Helm chart for Kubernetes deployment

Key ADRs: ADR-044 (TUI Architecture), ADR-045 (Persona System), ADR-046 (Reference/Command System)

No deadline assigned. Development effort should focus on M1-M6 milestones first.

v3.8.0 — M9: Server Implementation

Goal: Implement the CleverAgents server and all server-dependent capabilities using the A2A (Agent-to-Agent) protocol standard.

Scope:

A2A JSON-RPC 2.0 wire format and _cleveragents/ extension method routing
A2A stdio transport (local mode — agent as subprocess)
A2A HTTP transport (server mode — single JSON-RPC endpoint)
LangGraph Platform RemoteGraph integration for server-side actor execution
Server application structure (FastAPI + A2A SDK)
Authentication and authorization (API tokens, team RBAC)
Entity sync via _cleveragents/sync/* extension methods
Multi-device experience, shared namespaces, PostgreSQL backend
Server deployment (Docker + Kubernetes + Helm)

Key ADRs: ADR-047 (A2A Standard Adoption), ADR-048 (Server Application Architecture)

No deadline assigned. Development effort should focus on M1-M6 milestones first.

v3.9.0

66%

v3.9.0 documentation and feature updates

v3.2.0 — M3: Decisions + Validations + Invariants

Goal: Decisions are recorded during Strategize and Execute phases and persisted to the database. Users can view the decision tree (plan tree), inspect individual decisions (plan explain), manage invariants (invariant add/list/remove), and correct decisions (plan correct --mode revert|append) with selective subtree recomputation.

Acceptance Criteria

Decision records are created and persisted during Strategize and Execute phases
agents plan tree <plan_id> renders the full decision tree with dependencies
agents plan explain <decision_id> shows decision rationale, context snapshot, and alternatives
agents invariant add / invariant list / invariant remove manage global, project, action, and plan-scoped invariants
agents plan correct --mode revert <decision_id> reverts a decision and recomputes affected subtrees
agents plan correct --mode append <decision_id> appends new guidance without discarding prior decisions
Validation pipeline runs required and informational validations on resource-attached validations
agents validation attach / validation detach manage validation attachments to resources
Definition-of-done gating blocks plan apply when validations fail in required mode
Test coverage >= 97%

Technical Criteria

Decision domain model with 12 decision types (prompt_definition, invariant_enforced, strategy_choice, subplan_spawn, etc.) and context snapshots persisted to SQLite.
Influence DAG tracks downstream decision dependencies for selective subtree recomputation.
Invariant precedence chain enforced: plan > action > project > global.
Correction engine supports revert and append modes with BFS impact analysis.
Validation pipeline distinguishes required (blocking) from informational (non-blocking) modes.
Test coverage remains >= 97%.

v3.3.0

43%

v3.3.0 — M4: Corrections + Subplans + Checkpoints

Goal: Plans can spawn child plans (subplans) during execution. Subplans execute in parallel with configurable concurrency limits. Results are merged back using three-way merge strategies. Correction engine supports revert and append modes. Checkpointing enables rollback to previous plan states.

Acceptance Criteria

Execution actor can spawn subplans via subplan_spawn and subplan_parallel_spawn decision types
Subplans execute in sequential, parallel, or dependency-ordered modes with configurable concurrency limits
Subplan results merge back into parent plan using git three-way merge, sequential apply, or last-wins strategies
agents plan correct --mode revert <decision_id> reverts a decision and all affected downstream decisions
agents plan correct --mode append <decision_id> adds correction guidance without discarding the subtree
agents plan rollback <plan_id> <checkpoint_id> restores a plan to a prior checkpoint state
Checkpoints are created automatically at configurable intervals during plan execution
Phase reversion: a constrained apply transitions back to Strategize for re-planning
agents plan errors <plan_id> shows error recovery hints and retry options
Test coverage >= 97%

Technical Criteria

Subplan orchestration with domain models for config, status, failure handling, and three merge strategies.
Correction model with revert/append modes, BFS impact analysis, and dry-run reporting.
Checkpoint scopes with pruning and cleanup hooks integrated throughout the execution pipeline.
Phase reversion state machine: constrained apply → Strategize transition.
Concurrency locks (plan and project advisory locks) with plan resume and state recovery.
Test coverage remains >= 97%.

v3.4.0

46%

v3.4.0 — M5: ACMS v1 + Context Scaling

Goal: The Advanced Context Management System v1 is operational. Projects with 10,000+ files can be indexed and queried. The context assembly pipeline produces scoped, budget-constrained context views for actors. Hot/warm/cold storage tiers manage context lifecycle.

Acceptance Criteria

ACMS context assembly pipeline with 10 pluggable components is operational (Strategy Orchestration, Fragment Fusion, Context Finalization phases)
Repository indexing service can index projects with 10,000+ files
Context request protocol (CRP) models allow actors to declare needed information and depth
Hot/warm/cold storage tiers manage context lifecycle with per-actor scoped views
Context strategies (keyword search, semantic embedding, graph navigation) execute in parallel
UKO (Universal Knowledge Ontology) Layer 0 and Layer 1 domain ontologies operational
Skeleton compressor propagates compressed context from parent plans to child plans
agents project context set / context show / context inspect / context simulate commands functional
Multi-level config resolution with project-scoped overrides via config service
Test coverage >= 97%

Technical Criteria

ACMS pipeline with ConfidenceWeightedSelector, ProportionalBudgetAllocator, ParallelStrategyExecutor, CircuitBreaker, ContentHashDeduplicator, MaxDepthResolver, WeightedCompositeScorer, GreedyKnapsackPacker components.
UKO Layer 1 with 4 domain ontologies: uko-code, uko-doc, uko-data, uko-infra.
Context strategies: SimpleKeyword, SemanticEmbedding, BreadthDepthNavigator.
Strategy coordinator and fusion engine orchestrate parallel strategy execution.
Config service with multi-level resolution (global → project → action → plan).
Test coverage remains >= 97%.

v3.5.0

27%

v3.5.0 — M6: Autonomy Hardening

Goal: The system can autonomously execute a large-scale task (e.g., porting a substantial codebase) using hierarchical plan decomposition with 4+ levels of subplans, decision correction with selective subtree recomputation, parallel execution scaling to 10+ concurrent subplans, and validation-gated apply.

Note: Server stubs have been moved to M9 (v3.8.0) following the ACP→A2A protocol adoption and server architecture redesign (ADR-047/ADR-048). TUI features moved to M8 (v3.7.0).

Acceptance Criteria

System can autonomously execute a large-scale porting task using hierarchical subplan decomposition (4+ levels)
Parallel execution scales to 10+ concurrent subplans with configurable concurrency limits
Automation profiles (8 built-in: manual through full-auto) control autonomy thresholds per operation
Safety profiles enforce hard constraints: sandbox required, checkpoint required, unsafe tool gating, skill allow-lists, cost/retry limits
Cost and risk estimation actor produces estimates before plan execution
Autonomy guardrails (max steps, tool budget, required confirmations) enforced with audit trail
Semantic validation (syntax, import, and reference checks) runs during Strategize and Execute phases
A2A local facade enables in-process routing (groundwork for server mode in M9)
LSP server stub with JSON-RPC transport registered and accessible from actor graphs
Test coverage >= 97%

Technical Criteria

Automation profiles with 11 automatable task flags and confidence thresholds (0.0–1.0).
Safety Profile sub-model as composed component of AutomationProfile per ADR-041.
Large-project decomposition with 4+ level hierarchical subplans and clustering heuristics.
Estimation actor producing EstimationReport with cost and risk projections.
A2A local facade implementing in-process routing for local mode.
LSP runtime stub with initialize/shutdown handshake and JSON-RPC stdio transport.
Test coverage remains >= 97%.

v3.6.0

51%

v3.6.0 — M7: Advanced Concepts & Deferred Features

Goal: Advanced concepts that were not needed for a basic working version. Extends beyond core MVP but does not require TUI (M8) or Server (M9).

Scope:

Advanced context strategies and additional LLM backends
Additional resource types: cloud infrastructure, database resources, virtual resource types with equivalence tracking
ACP → A2A module rename (groundwork for server mode)
Container tool execution via devcontainer resource types
Pluggable scope chain resolution and plugin architecture extensions
Cost/session budgets and safety profile enforcement
E2E workflow specification tests covering all major workflows
Domain event wiring for all 38 domain events
UKO Layers 2 and 3 (paradigm/format and technology-specific ontologies)
FAISS and Tantivy search backends for semantic and keyword context retrieval

Target Date: 2026-03-28 (Day 48)

Key ADRs: ADR-042 (Resource Type Inheritance), ADR-043 (Devcontainer Integration), ADR-047 (A2A Standard Adoption)