Roadmap

Roadmap

Status snapshot of what's shipped and what's next. Everything here sits on top of the pipeline and must keep its guardrails intact (no live host by default, no patching the target, read-only repo, no auto-submit) — see Guardrails & Safety.

Shipped

Phase	Feature	Status
Precision & depth uplift	Ground-truth recon, variant hunting, downgrade-don't-delete, SCA	✅ Done
0	Backend API & job orchestration	✅ Done
1	Core web UI (paste → run → live progress → results)	✅ Done
2	Settings + "Let the AI choose" + light/dark theme	✅ Done
3	Interactive chat / interrogation layer	✅ Done
4	Context enrichment (vuln index, Stage-0 research)	◑ Partial — AST metadata deliberately deferred
5	Multi-backend runners (Codex / local-OSS)	✅ Done
6	Remediation (fix) pipeline	✅ Done
7	Benchmarks & evaluation	✅ Done (core harness)
8	Cost model & economics	✅ Done (cost side)
9	Dynamic/runtime analysis (opt-in, sandboxed)	✅ R1–R4 done
10	Live target interaction (opt-in, scope-locked)	✅ L1–L3 done

See the individual feature pages linked from Home for what each of these actually does.

Deliberately deferred

• Light static metadata (tree-sitter AST/call-graph). Decided against for now — unproven ROI on top of a strong LLM. Trigger to revisit: the benchmark shows a recall loss attributable to missed inter-procedural/multi-file data flow. See Design Decisions.
• Deep static analysis (CPG/PDG/CFG, e.g. Joern). Decided against — build-based CPG must compile the target (guardrail conflict) and would confound the research study (can't separate LLM findings from graph findings). Would only follow the parse-only step above, sandboxed, gated on evidence.
• Gemini backend. Investigated and parked — the gemini CLI currently returns persistent errors on the agentic path (Google's Code-Assist → Antigravity migration); the raw API works but is text-only (no repo tools), so it could only serve as a fallback for text-reasoning stages, not the agentic audit/recon stages.
• Streaming chat replies. Real architectural blocker — every backend is a single blocking call today; deferred behind higher-value work.
• Per-run subprocess isolation. Low priority for a local single-user tool; boundary-level cancellation already works.

Constraints that shape every decision here

1. Cost is a first-class UX concern — a full real run can cost tens of dollars; every UI path needs a cost preview, a hard budget control, and a free mock mode.
2. Run time is minutes to 30+ minutes — no blind spinners; stage timeline + live cost + partial results as they land.
3. Guardrails are non-negotiable — any new execution path must route through or mirror the existing chokepoint.
4. The AgentRunner is already swappable — adding a backend is mostly re-validating guardrails for it, not the interface. See Multi-Backend.

Related

• Design Decisions — the why behind the biggest deferred item (no CPG/AST).
• Home — the feature-by-feature index of everything already shipped.