dravix-agent 0.1.0

package/ROADMAP.md

@@ -0,0 +1,117 @@
+# Aegis-v2 Roadmap
+
+Phase 0 ships in this session. Phase 1-3 ship in their own sessions, each
+phase fully production-ready — no stubs left behind.
+
+---
+
+## Phase 0 — Real-time Phase-0 gate (SHIPPED)
+
+**Status:** Done.
+
+**What ships:**
+- `aegis` CLI: `mcp`, `hook`, `scan`, `doctor`, `--version`.
+- MCP server (stdio) exposing `precheck_change`, `validate_edit`, `explain_risk`.
+- Engines: `secret-scan` (built-in), `treesitter`, `eslint`, `pyright`, `tsc`, `semgrep`.
+- Orchestrator: parallel run, per-engine + total timeout, LMDB cache.
+- Risk scoring: confidence × severity → ALLOW / WARN / BLOCK; exit 0 / 2.
+- Path safety: deny home-secret + system paths; reject outside-root.
+- Audit log: append-only `.aegis/audit.jsonl`.
+- Claude Code hook config (`.claude/settings.example.json`).
+- Tests: findings schema, lang, risk, secret-scan, tree-sitter, eslint, orchestrator.
+
+**Exit criteria (met):**
+- `npm install && npm run build && npm test` passes on this machine.
+- `aegis doctor` prints engine availability.
+- `aegis scan tests/fixtures/with_secret.ts` blocks (exit 2).
+- `aegis scan tests/fixtures/clean.py` allows (exit 0).
+- Hook config attaches to Claude Code and blocks a `Write` with secrets.
+
+---
+
+## Phase 1 — Project-aware context (NEXT)
+
+**Goal:** raise recall on cross-file bugs by giving every engine the
+project graph + symbol resolution + similar-code retrieval.
+
+**Deliverables:**
+- `src/index/scip.ts` — wrapper around `scip-typescript` and `scip-python`;
+  builds an incremental SCIP index of the project; LMDB-backed lookup
+  for `defOf(sym)` / `refsOf(sym)` / `callersOf(fn)` in O(log n).
+- `src/index/joern.ts` — Joern CPG: invoke `joern-parse` once per project,
+  store CPG in OverflowDB; query API for `dataFlow(src, sink)` and
+  `cpg.method.name("X").caller`. Incremental update on file save.
+- `src/index/embeddings.ts` — local `CodeT5p-110M` (ONNX) embedder via
+  `@xenova/transformers` (no Python dep); FAISS-style brute-force top-K
+  on small projects, `hnswlib-node` on larger ones.
+- `src/engines/context.ts` — uses the three indexes to enrich the
+  Tier-1 prompt with focused, per-symbol context (not generic dump).
+- Updated `validate_edit` returns a `project_context` block alongside
+  the findings list — the LLM critic in Phase 2 will consume it.
+
+**Exit criteria:**
+- Indexing fixme/argus (~100 files) completes in < 5 s cold, < 200 ms incremental.
+- Cross-file lookup (`callers_of("require_admin")`) returns the right call site.
+- Latency budget: p50 ≤ 600 ms (Phase 0 + Phase 1 combined).
+- New tests: SCIP index, Joern query, embedding round-trip, focused-context block.
+
+**Cite-tools:** RepoGraph (ICLR 2025), SCIP (Sourcegraph), Joern, CodeT5+.
+
+---
+
+## Phase 2 — Deep semantic + critic
+
+**Goal:** catch the bugs no static analyzer alone can — taint reachability,
+race conditions, deep logic — by combining best-of-class OSS analyzers
+with an LLM critic that reasons over a *retrieved evidence trail*
+(Vul-RAG pattern; +16-24% over pure-LLM per Du 2024).
+
+**Deliverables:**
+- `src/engines/pysa.ts` — Pyre/Pysa taint analysis for Python; pre-warmed
+  function summaries cached per commit; sub-second per-file query.
+- `src/engines/racerd.ts` — Joern-CPGQL race queries for Python/TS/JS
+  modeled on RacerD's `NoThread` / `AnyThread` lattice; static candidates
+  + post-filter.
+- `src/engines/vulrag.ts` — local vector store of CVE knowledge (cause +
+  vuln-pattern + fix-pattern + CWE), populated from the existing 313K-row
+  argus corpus; retrieval by `(cwe, semantic)`.
+- `src/engines/router-150m.ts` — local 150M classifier (ONNX, INT8) that
+  routes each edit to `pass` / `lint` / `deep_review`.
+- `src/engines/llm-critic.ts` — calls Claude/GPT with structured evidence
+  (diff + Joern path + Vul-RAG knowledge + callers); returns
+  `{verdict, cwe, evidence, fix, confidence}`. Ensemble vote with a
+  second-vendor model.
+
+**Exit criteria:**
+- Catches 4/6 planted bugs in the argus smoke test (vs. 1/6 today).
+- p50 ≤ 1.0 s for non-critic path; p95 ≤ 4 s with critic.
+- LLM critic budget cap: 10% of edits / $0.05 max per call.
+
+**Cite-tools:** Pysa, Infer/RacerD, Vul-RAG (arxiv 2406.11147), IRIS (ICLR 2025).
+
+---
+
+## Phase 3 — Production hardening
+
+**Goal:** suppression learning + telemetry + property-based testing gate +
+policy-as-code — what turns a working prototype into something a team can rely on.
+
+**Deliverables:**
+- `src/learning/suppressions.ts` — every `accepted` / `rejected` /
+  `edited-then-saved` outcome appended; weekly re-tunes the router and
+  per-rule confidence priors.
+- `src/learning/telemetry.ts` — local p50/p95/p99 dashboard via a small
+  CLI command (`aegis stats`).
+- `src/engines/property-tests.ts` — for high-suspicion functions
+  (money/auth/perm), LLM emits a Hypothesis (Python) or fast-check (JS/TS)
+  property; runner executes in subprocess sandbox; counterexample = proven bug.
+- `src/policy.ts` — YAML-driven policy: `CWE-89 → BLOCK`,
+  `CWE-704 → WARN`, etc.; per-project override.
+
+**Exit criteria:**
+- Suppression-aware re-training reduces FP rate by ≥ 20% over baseline.
+- Property-test gate catches at least one of: tax-before-discount,
+  cache stampede, off-by-one in argus smoke test.
+- `aegis stats` shows true p50/p95 from real audit log.
+
+**Cite-tools:** Hypothesis ghostwriter, fast-check, Fuzz4All (arxiv 2308.04748).