Roadmap

Roadmap

Completed — Phase 0-2 (v1.0.0 — v1.16.1)

• Core architecture (orchestrator, router, watchdog, job runner)
• 4-type memory system with provenance model + consolidation + persistent conversation
• 9-layer processing pipeline (dispatch → cache → RAG → classify → LLM → escalation → learning → filter → explain)
• Docker sandbox (read-only, no-network, resource limits)
• Encrypted vault (Fernet AES-128, PBKDF2 480K iterations)
• Agent-to-Agent API (HTTP, port 8420, replay protection)
• Learning system (skill tracking, model escalation, prompt augmentation)
• Security: deny-by-default tool policy, channel enforcement, approval queue
• Finance: budget policy (hard/soft caps), risk templates, approval workflow
• Review bounded context: repo audit, PR review, release review, golden cases
• Build bounded context: 10 mutation types, verification loop, acceptance criteria
• Control plane: policies, intake, gateway, delivery, evidence export, reporting
• CI: mypy, ruff, 1300+ test gate, architecture invariants, release-readiness gate

Phase 3 — Operatorization (v1.17.0 — current)

Closed

• v1.17.0 — 7 audit-driven bug fixes (cron, telegram, memory, sandbox, job runner, learning)
• v1.18.0 — Security hardening (sandbox pip validation, shell escape, CSV injection guard)
• v1.19.0 — Operator Telegram surface: /intake, /report, /build
• v1.20.0 — Runtime capability binding (T4-E2-S3) + /jobs, /deliver
• v1.21.0 — Cost accuracy feedback loop (T6-E1-S1) + unified policy boundary (T5-E1-S1)
• v1.21.1 — Deployment portability: dynamic identity in consolidation/redaction

Open

• T6-E2-S1 — Deeper runtime telemetry (time-series, /telemetry command)
• T7-E2-S5 — Seller-side Obolos publishing and wallet-topup
• T7-E1-S1 — Expand gateway contract beyond one provider
• T8-E1-S3 — Stronger architecture invariants for cross-domain boundaries

Future (Phase 4: Enterprise Hardening)

• Live operator backend/UI on top of current CLI and TS contracts
• Service extraction readiness
• Stronger observability and persistent approvals
• Earning module (agent finds work, proposes, human approves, executes)

Known Limits

• CLI token overhead: ~26k tokens per call (Claude adds context)
• No live operator dashboard UI (CLI + Telegram + mock TS surface only)
• Builder has no general code generation or semantic requirement engine
• Only one external provider (obolos.tech) — seller-side not implemented
• Semantic model uses ~1.5GB RAM
• trycloudflare URLs change on restart (no named tunnel)