Roadmap¶
This roadmap reflects the current shipped state of AION and the remaining work required to move from a local control-plane prototype toward production-grade self-healing.
Current status¶
The repository now ships:
- context-aware Python scanning with repository profiling and Semgrep triage
- deterministic repair artifacts plus standalone verification
- repair evaluation with persisted repair records and metrics
- sandbox orchestration for single events, event queues, inbox items, and webhooks
- release candidates with approval, staged advancement, rejection, and rollback
- runtime-first defense planning for containment before rollout
- drift detection, knowledge base, and a continuous watch loop for self-evolution
- knowledge-base confidence boosts wired into the policy engine
Completed phases¶
Phase 0: Trusted foundation — shipped in v1.0.0¶
- Findings are upgraded into structured incidents.
- Repair, verification, orchestration, and release flows share JSON models.
- Fixtures cover the incident lifecycle end to end.
Phase 1: Automatic repair closed loop — shipped in v1.0.0¶
- Deterministic repair is implemented for the first supported issue classes.
- Verification runs syntax checks, Semgrep re-scan, assertions, and optional sandbox commands.
repair,verify, andrun-incidentprovide a local closed loop.
Phase 2: Self-verification and learning — shipped in v1.0.0¶
- Repair attempts can be persisted as audit records.
repair-evalcomputes repair success, verification pass, false-fix, and rollback rates.- Failure outcomes are captured for later template and policy refinement.
Phase 3: Pre-production autonomy prototype — shipped in v1.0.0¶
- The orchestrator accepts JSON events, queue payloads, inbox items, and webhooks.
- Policy gating decides whether an incident can enter automatic sandbox remediation.
- Repository-level sandbox execution supports project-specific verification commands.
- Release candidate management now covers approval and staged rollout decisions.
Phase 4: Self-evolving engine — shipped in v1.1.0¶
- Drift detection:
snapshotsaves a point-in-time security state;driftcompares the current codebase against any saved snapshot and reports new incidents, resolved incidents, regressed files, and a numeric health delta. - Knowledge base: every successful repair is persisted as a
RepairPatternin.aion/knowledge/patterns.json; confidence boosts derived from historical success rates are applied by thePolicyEnginebefore the auto-repair threshold check, closing the self-evolving feedback loop. - Continuous watch loop:
watchpolls a target directory on a configurable interval, auto-repairs newly detected incidents, and refreshes the baseline after each successful fix. - Engine status dashboard:
statusshows all saved snapshots and the full knowledge-base summary in a single view. - Expanded LLM provider support: Gemini and Azure OpenAI are now supported alongside Anthropic and OpenAI, with automatic provider auto-detection from available environment variables.
Next steps¶
Phase 5: Production adapters¶
The next major work is no longer core modeling. It is integration:
- authenticated webhook and queue adapters for real event sources
- deployment adapters for promotion, rollback, and rollout telemetry
- provider adapters for WAF, gateway, and feature flag execution
- approval and audit integration with external systems
- richer repository test selection instead of only configured command lists
Guiding principle¶
AION is intentionally conservative. Every step favors deterministic artifacts, auditable state, and reversible rollout decisions over opaque automation.