Skip to content

OASIS v0.7.0

Latest

Choose a tag to compare

@psyray psyray released this 09 Sep 11:03
· 1 commit to master since this release
f03e424

πŸš€ [0.7.0] - 2026-09-09

✨ Added

  • 🀝 OpenAI-compatible model backend (--provider openai): run the full pipeline against any OpenAI-compatible local server (vLLM, LiteLLM, LM Studio, llama.cpp, …) via --provider openai --api-base https://your-server/v1 (optional --api-key); structured outputs use response_format schemas with schema-in-prompt fallback, context windows via OASIS_OPENAI_CTX_TOKENS. Env equivalents: OASIS_LLM_PROVIDER, OASIS_OPENAI_BASE_URL, OASIS_OPENAI_API_KEY.
  • 🧱 Model backend abstraction (oasis/backends/): provider-agnostic ModelBackend with OllamaManager (default) and OpenAICompatManager, resolved by a create_model_manager factory; oasis.ollama_manager stays as a compat shim.
  • πŸ”€ Independent embedding backend: chat and embedding workloads are routed to separate backends β€” embeddings inherit the chat backend configuration by default (provider, base URL, API key), and explicit flags override to spread the workloads across servers (e.g. a dedicated RAG server). Flags: --embed-provider / --embed-api-base / --embed-api-key (env OASIS_EMBED_*), dashboard RAG follows with --web-embed-provider / --web-embed-api-base / --web-embed-api-key (env OASIS_WEB_EMBED_*, falling back to the scan-side embedding backend then the chat backend). Every embed model is checked against its backend at startup (/v1/models match for OpenAI-compatible servers, local pull for Ollama) β€” the run aborts with an explicit Model not available error otherwise; chunk-size detection and multiprocessing embedding workers run on the embedding backend, the startup log shows both routed backends, and a RAG backend failure at query time shows a ⚠️ notice on the affected assistant answer.
  • 🧩 Consolidated multi-model report (-rm/--report-model): after a multi-model run (-m a,b), findings from every model's canonical JSON are merged by stable fingerprint into consolidated/consolidated_report.json (+ Markdown sibling) bucketed by confirmation β€” confirmed by all / several / single-model, with per-model severities; the report model then synthesizes an executive narrative from a compact digest via structured output (it never re-detects findings; an LLM failure degrades to a narrative-less document). Digest budget: OASIS_REPORT_CONSOLIDATION_DIGEST_MAX_CHARS (default 24000). Single-model runs skip the feature (logged). The dashboard lists the report under the Consolidated pseudo-model and renders a dedicated JSON modal (TOC, summary buckets, narrative, group tables); the assistant stays off for this kind for now.
  • 🧠 Thinking control on OpenAI-compatible servers: -mt / -smt map to vLLM-style chat_template_kwargs.enable_thinking β€” the default (thinking off) no longer lets reasoning burn the max_tokens budget; OASIS_OPENAI_THINKING_KWARGS=auto|on|off.
  • πŸ“‘ Streaming on OpenAI-compatible servers: stream: true is now actually sent β€” the dashboard assistant no longer returns empty responses on these backends; reasoning deltas render as collapsible thought segments.
  • 🌐 Finding validation: broader language support: the deterministic catalog now covers Go, Rust, Kotlin, Scala, Java, Ruby plus cross-language sinks (C#, Node.js sqlite3/mysql/child-process, JS templates, XML/deserialization); pattern catalog version bumps to 4 (entry-point caches invalidate automatically).
  • πŸ›‘οΈ Scan-time finding validation (automatic): findings are validated after each deep pass and a verdict (status / confidence / summary) is embedded in the canonical reports with a color-coded status badge; full payloads persist as a finding_validations.json sidecar with a flat finding picker, πŸ’¬ Ask AI buttons and Scan-time verdicts in the assistant. Flags: --validate-findings / --no-validate-findings (default: on), --validate-findings-budget SEC, --validate-findings-narrative; report schema version bumps to 6.
vuln-finding-validation vuln-finding-validation-detail
  • πŸ§ͺ Realistic vulnerable test fixtures: test_files/realistic_app/ ships vulnerable mini-apps in Python, C#, Java, JS and PHP, each with safe twin variants to check false positives.
  • 🧹 Transverse finding deduplication: duplicate deep-pass findings (same snippet fingerprint or overlapping line ranges) are merged before validation and report writing.
  • πŸ§ͺ E2E fixture check script (scripts/e2e_fixture_scan.py): dev gate that scans each fixture with the Injection family in an isolated workdir (optional --scan-model for a separate scan model) and prints a pass/fail table; requires a live LLM server, not part of the unittest suite.
  • 🧹 Suppression registry (--suppressions-file, --write-suppression-candidates): JSON registry keyed by finding fingerprints; matching findings get a native SARIF 2.1.0 suppressions entry.
  • 🚫 Inline ignore markers: findings whose source lines carry an ignore marker β€” # oasisignore (OASIS-specific), # noqa, # nosec, # nosemgrep β€” are dropped after dedup, before validation and reports; chunk notes display the skip (N finding(s) skipped via inline ignore marker, original notes preserved). Flags: --inline-ignore / --no-inline-ignore (default: on), --inline-ignore-tokens CSV to customize the marker list. Markers on the line above the snippet are intentionally not honored; unreadable files fail open.
  • 🀝 CI gate (--fail-on SEVERITY): exits with code 3 on findings at or above a threshold (critical|high|medium|low); new README "CI integration" section.
  • πŸ”„ Scan diff vs baseline (--diff-against PATH): writes diff/diff_report.json + a Markdown sibling with new / fixed / persistent buckets, before the CI gate.
  • ⏳ Multi-model scan progress tabs: the dashboard Scan progress card renders an Overall tab (aggregate bar + X/N models complete) plus one tab per deep model β€” emoji + display name, with current (⏳) / done (βœ“) / pending (grayed) states and per-model phase rows. Tabs auto-follow the model currently scanning and return to Overall when the run completes; /api/progress and the scan_progress socket event carry additive models_progress / overall payloads, the latest executive-summary sidecar is selected by its inner updated_at within a run (a second model can no longer be shadowed by the first model's "Complete"), and incoming findings refresh the filter option lists and the report view live. Run-level logs/ directories are no longer mistaken for model directories.
  • 🀝 Consolidated report grouping & context: findings quoted by different models with different amounts of surrounding context (same vulnerable line) now merge into one group β€” a deterministic snippet-containment pass on top of the exact fingerprint β€” instead of splitting into single-model buckets; the consolidated document also carries project / analysis_root (and the dashboard backfills it for artifacts written before), so its card no longer flags the scanned codebase as unreachable.

πŸ› Fixed

  • πŸ“„ KeyError: 'md' with JSON-only output: report post-processing no longer assumes output_files['md'] exists.
  • πŸ›‘οΈ Assistant finding validation: PHP support + reliable sink anchoring: procedural PHP findings are no longer scored unreachable; companion fixes for relative file_path resolution, ripgrep lookaround fallback, caller-hop taint retry, on-path-only mitigations, and the LangGraph entry-point state key.
  • πŸ–₯️ -lm on an unreachable backend: model listing now fails fast with a clear error instead of printing the misleading default model list.
  • πŸ§ͺ Documentation files polluting scan findings: the fixture README moved out of the scanned tree (markdown is analyzed as source) and the pickle import is localized to remove the 'Global Import' false positive.
  • 🧠 Embedding context auto-detection & retry (issue #58): chunk sizing prefers the runtime context from Ollama ps() over declarative num_ctx / GGUF metadata, with a shared retry halving the chunk limit on context-length errors β€” embeddings no longer silently drop a file on input length exceeds the context length.
  • πŸ€– Model selection list: interactive selection (and the preloaded list) shows the full model name including the version tag (e.g. qwen2.5-coder:7b) instead of truncating after : (issue #61).
  • 🏝️ ASCII logo alignment: the startup banner now renders through a raw string, so the palm-tree backslashes display verbatim instead of being halved by Python escape processing β€” the letters of OASIS keep their column alignment.
  • 🧹 Dashboard filter persistence: the model filter now survives a page reload (localStorage save/restore/clear like the other facets, checkboxes pre-checked on render), and the first load no longer prunes the filter option lists β€” every facet (model, vulnerability, severity, format, language, project) keeps all its checkboxes visible even when a restored filter is active (the initial stats fetch omits each facet's own filter; counts stay cross-filtered after interactions).
  • πŸ“‘ Live progress for JSON-only scans: output formats without md now still write the executive summary canonical JSON and its progress sidecar (the canonical path derives from the json artifact when MD is absent), so -of json runs get the same live progress tabs as any other output format.

⚑ Changed

  • πŸ–₯️ Provider-neutral CLI/dashboard messages: banners and errors no longer hardcode Ollama wording; Ollama stays the default provider with unchanged behavior.
  • πŸ“š README: new "Model providers (backends)" section (vLLM workflow + new flags/env vars).
  • 🏝️ Branding: OASIS now stands for Open Automated Security Intelligence Scanner (README tagline, CLI banner, dashboard footer, packaging metadata) β€” reflecting multi-backend support (native Ollama by default, plus any OpenAI-compatible local server).