OASIS Implementation Patterns

Goal

Apply the code organization and delivery style repeatedly used in OASIS commits.

Workflow

1. Classify the change as feat, fix, or refactor.
2. Identify the bounded module(s) to touch instead of editing a single large file:
   • oasis/oasis.py: CLI and orchestration
   • oasis/analyze.py: SecurityAnalyzer — structured outputs, embedding, cache, and LangGraph hook methods (langgraph_discover_and_publish, langgraph_scan_and_publish, etc.) invoked by the agent layer
   • oasis/agent/: LangGraph-only — compile the DAG, state schema, node wrappers, and invoke_oasis_langgraph (canonical pipeline entry). Keep graph structure and stable node names in one place (see LangGraph layout below).
   • oasis/schemas/: Pydantic models for LLM JSON and canonical vulnerability reports
   • oasis/report.py: JSON-first vulnerability reports; Jinja under oasis/templates/reports/
   • oasis/web.py: dashboard indexing (json/ stats), APIs (/api/report-json, /api/progress, Socket.IO scan_progress, legacy MD preview rules)
   • oasis/backends/: provider-agnostic model backends — factory create_model_manager, the ModelBackend contract, OllamaManager (native Ollama) and OpenAICompatManager (OpenAI-compatible servers: vLLM / LiteLLM via --provider/--api-base/--api-key); oasis/ollama_manager.py remains a backward-compat shim
   • oasis/helpers/: all shared helpers (formatting, parsing, progress helpers, small pure utilities). Do not leave helper-shaped functions in feature modules—extract them here and group by category in dedicated modules (see Design Guardrails).
   • oasis/helpers/embedding/: canonical embed-model normalization/parsing (normalize_embed_models, resolve_embed_models, primary_embed_model, EmbedModelValueError); reuse it for CLI parsing and embedding-manager primary model resolution.
   • oasis/static/js/dashboard/*: web dashboard behavior (JSON modal preview, force=1 on reload for stats/reports/progress, applyProgressPayload / progressState)
3. Keep interfaces compatible unless migration is explicit.
4. Update docs (README.md) when user-facing flags or behavior change.
   • Keep README.md Features as summary-only bullets.
   • Put detailed behavior/usage in the existing relevant section; create a new section when a new feature has no fitting section.
5. Prefer defensive handling around cache/model/network boundaries.
6. For reporting changes, treat canonical JSON schema as the source of truth and keep Jinja templates synchronized.
7. Before considering the task done: verify no duplicated logic remains (same rule in two places, mirrored constants, near-copy functions). Refactor into a single canonical implementation and import it everywhere it is needed.

Dashboard report modals (shared architecture)

When changing any report preview inside #report-modal (vulnerability, executive summary, audit report, future canonical types), follow the single architecture described in the oasis-dashboard-js-patterns skill (Report modal architecture section):

• Compact layout + TOC / section jumps + optional Chart.js blocks, scoped CSS in report_preview.css under #report-modal-content.
• One post-render initializer path from modal.js (_finalizeReportModalView) for kind detection (stem, path, or report_type) — no scattered one-off branches.
• Assistant: one mountReportAssistantPanel pipeline with variants (finding selectors only for vuln JSON; aggregated or report-scoped chat for executive/audit as designed) — duplicate chat stacks per report type are not acceptable.
• Server: meta/preview endpoints and parsing live in web.py + oasis/helpers/; extend shared helpers instead of copying aggregation logic for each modal.

Treat audit modal work as reusing this spine when aligning with the vulnerability/executive experience, not as a greenfield duplicate UI stack.

LangGraph layout (canonical analysis pipeline)

• Product path: AnalysisType.GRAPH in oasis/enums.py is the only orchestration mode; on-disk chunk caches live under graph/deep and graph/scan (see oasis/cache.py).
• Import discipline: Import LangGraph entry points explicitly (e.g. from oasis.agent.invoke import invoke_oasis_langgraph). The oasis.agent package docstring explains avoiding a heavy import at interpreter startup when only the submodule is needed.
• oasis/agent/graph_labels.py: Single source for node ids (GRAPH_NODE_*) and conditional route targets — use these strings in graph.py, routing, and tests instead of scattering literals.
• oasis/agent/graph.py: Builds StateGraph(OasisGraphState) — edges: Discover → Scan → Expand → Deep → Verify → (conditional: Expand or Report) → (conditional: PoC or END when PoC disabled) → END.
• oasis/agent/tools.py: Thin dispatch — each node calls the matching SecurityAnalyzer.langgraph_* method; console UX uses oasis/helpers/langgraph_cli/ (langgraph_emit, langgraph_emit_phase, etc.).
• oasis/agent/state.py: Typed state passed between nodes (all_results, expand_iterations, max_expand_iterations from CLI, PoC markdown accumulator, etc.).
• Helpers tied to the graph:
  • oasis/helpers/progress/ — executive-summary phases rows aligned with LangGraph stages (ProgressPhaseRowId.GRAPH_*, ProgressActivePhase.GRAPH_PIPELINE) plus scan progress JSON/markdown helpers (keep in sync with report-side progress).
  • oasis/helpers/langgraph_cli/ — numbered phases, tqdm-safe logging, post-pipeline / report-delivery banners (also used from report.py), and canonical vuln-type totals for LangGraph (embedding_tasks_vuln_types_total, deep_payload_vuln_types_total).
  • oasis/helpers/context/expand.py — pure line-window expansion around suspicious spans; defaults from oasis/config.py (CONTEXT_EXPAND_*).
  • oasis/helpers/poc/ — compact findings digest JSON for --poc-assist, PoC budgets (re-export from config), hints markdown (build_poc_hints_markdown), chat options (poc_assist_chat_options), stage DEBUG logging (maybe_debug_log_poc_stage_output).
• CLI (extend README.md when behavior changes): --langgraph-max-expand (langgraph_max_expand_iterations), --poc-hints, --poc-assist.
• Tests: tests/test_analyze_orchestration.py for the LangGraph pipeline and PoC helpers; tests/test_oasis_cli.py for LangGraph-related flags.

Design Guardrails

Incremental scan progress (see sibling skills)

Details live in the oasis-python-architecture skill (constants in oasis/helpers/progress/__init__.py, sidecar, helper modules) and the oasis-dashboard-js-patterns skill (REST, Socket.IO, stale updated_at guard). Touch oasis/helpers/progress/__init__.py, report.py, web.py, and dashboard JS together when the wire contract changes; extend tests/test_report_schema.py when behavior is contract-visible.

Audit metrics and dashboard comparison

• Report.generate_audit_report must keep an Audit Metrics Summary markdown table with stable Metric | Value rows (count/similarity/high-medium-low tiers).
• When json is in output formats, OASIS also writes audit_report.json (oasis/schemas/audit_report.py); Markdown and JSON must reflect the same document. The dashboard prefers sibling JSON for listing metrics and modal HTML via /api/report-html when the file exists.
• WebServer parses metrics into audit_metrics for /api/reports (JSON first, then Markdown). Shared md → json rules live in oasis/helpers/dashboard/json_sibling.py and dashboard audit-report-paths.js—keep them aligned with web.py preview routes.
• Dashboard comparison UI (utils.js + views.js + interactions.js) depends on those keys; treat report/web/dashboard as one contract surface and update tests in tests/test_report_schema.py together.

Report storage, executive JSON, and dashboard filters

• Output tree: Default security_reports/<project_slug>/YYYYMMDD_HHMMSS/…; optional --project-name / -pn overrides slug naming—coordinate oasis/helpers/report_project.py, CLI (oasis/oasis.py), exporters, and README.md when behavior shifts.
• Executive summary canonical JSON: Rich schema_version document built in oasis/helpers/executive_summary.py and oasis/report.py (overview KPIs, guidance_markdown, tier_definitions, capped similarity_highlights). HTML preview uses executive_summary_from_json.html.j2 and executive-preview.js (TOC, Chart.js)—change schema, templates, and dashboard consumers together.
• analysis_root: Stored relative to security_reports/ in new JSON; resolution lives in oasis/helpers/analysis_root_path.py—reuse for assistant/RAG and scan_root; never duplicate path guessing in web.py.
• Dashboard: Severity (tier) + project filters and scoped previews (queries wrapped with active filters; server rejects out-of-scope paths). /api/stats exposes severity_finding_totals. Python: oasis/helpers/dashboard/severity_filter.py and web.py; JS: filters.js, api.js, modal.js.
• Web UI theme: Header light/dark toggle and oasis:theme-change in bootstrap.js; charts must follow (views.js, executive-preview.js).

Duplication and centralization (strict)

• Do not repeat code. Treat copy-paste and “almost the same” branches as defects: merge into one function, module, or schema and reuse. KISS means the smallest correct change, not duplicating logic to save a refactor step. DRY is mandatory, not aspirational. SOLID is incompatible with parallel implementations of the same rule in different files.
• Before shipping, ask: “If this behavior changes tomorrow, is there exactly one place to edit?” If not, centralize first (Python: oasis/helpers/ or oasis/schemas/ as appropriate; JS dashboard: shared modules under oasis/static/js/dashboard/).
• Duplicating strings, field names, validation rules, or API shapes across modules is still duplication—use shared constants, Pydantic models, or helpers.

Provider-agnostic LLM calls (model backends)

• Never hardcode Ollama assumptions outside oasis/backends/: feature code goes through the ModelBackend contract — responses are normalized to {"message": {"content": ...}}, and Ollama-style kwargs (options.num_predict, format=<json_schema>) are translated per backend. Adding provider-specific behavior means extending the matching backend (or the contract), not branching on the provider in analyze.py / embedding.py / web.py.
• New provider knobs land in oasis/config.py (OASIS_* env, documented in the module docstring) plus oasis/oasis.py CLI, and stay aligned with tests/test_backends_openai_compat.py and the README Model providers (backends) section when user-visible.

Open-source hygiene (no personal infrastructure in the repo)

• Committed code, help strings, docstrings, tests, and docs stay agnostic of personal infrastructure (server hostnames, SSH aliases, private URLs, keys, container names, local paths): use neutral placeholders (https://llm.example.com/v1, localhost). Personal server documentation/configs live outside the repository (an uncommitted private workspace mirror) and are never referenced from committed files.
• Private strings in reachable history require a history rewrite (amend / cherry-pick rebuild of every affected branch — git branch --contains — then git reflog expire --expire=now --all + git gc --prune=now), not a cleanup commit; verify with git log --all -p | grep … before pushing.

Finding-validation pipeline (POST /api/assistant/investigate)

When changing how the assistant validates findings, follow the contract below:

• Sink resolution lives in one helper: oasis/helpers/assistant/web/sink_resolution.py (resolve_sink_from_finding_indices + coerce_positive_int_line). It handles primary payloads (vulnerability JSON with a files array) and scope payloads from finding_scope_report_path (executive aggregate flow), plus integral-float sink_line clients. Do not re-implement this lookup inline in web.py or other callers.
• Executive aggregate flow: assistant_investigate must run finding_scope_report_path through oasis.helpers.executive.assistant_scope.resolve_aggregate_finding_scope_payload before mapping indices, mirroring what assistant_chat already does for the chat panel — same pure helper, same security guardrails (model-dir match, traversal protection).
• Verdict honesty: compute_verdict runs on the full deterministic evidence (oasis/helpers/assistant/verdict/verdict.py). Never mutate status, confidence, summary, or family based on presentation tweaks; if evidence is thin, insufficient_signal is the correct outcome.
• Presentation-time EP filter: After compute_verdict, call apply_presentation_filter_to_result (oasis/helpers/assistant/web/result_presentation.py) to prune entry_points and rebuild citations:
  • flow: keep EPs linked by an execution_paths.entry_point if any, else EPs whose citation file matches scope.sink_file, else empty.
  • access: keep only EPs whose citation file matches scope.sink_file; never filter control_checks / authz_hits.
  • config: untouched (no EPs).
• LLM synthesis anchoring: compact_investigation_for_llm (oasis/helpers/assistant/think/investigation_synth.py) places a scope_focus block (vuln name, family, sink_file/line, verdict status/confidence) at the head of the dump; the system prompt explicitly references it and forbids fabricating paths or chains. When changing this contract, keep prompt + payload in sync.
• Tests: helper unit tests in tests/test_assistant_validation.py (TestSinkResolution, TestPresentationFilter, TestScopeFocusInLLMPayload) and route end-to-end in tests/test_web_assistant_api.py (TestAssistantInvestigateRoute). Same change set when behavior shifts; verdict-label assertions must stay deterministic.

Other guardrails

• Use KISS: choose the smallest change that solves the issue after consolidation, not instead of it.
• Helper centralization: Anything that matches “helper” characteristics—stateless or lightly stateful utilities, shared formatters/parsers, progress or status row builders, guards, small transforms reused outside one call site—must be implemented under oasis/helpers/, in a module that matches its category (e.g. progress-related code beside scan_progress.py / exec_summary_progress.py, shared types beside progress_types.py). When touching existing code, relocate qualifying functions into the appropriate helper module instead of growing orchestration files.
• Use SOLID: isolate responsibilities and avoid growing god functions; helpers stay thin and focused per module.
• Keep UX fixes localized for dashboard modules and templates.

Done Criteria

• Run the relevant tests/test_<area>.py module(s) (or unittest discover -s tests) for code you changed; add or extend tests in the file that matches the subsystem (same layout as in the project: report contract, CLI, cache, embedding, helpers, web, LangGraph orchestration in tests/test_analyze_orchestration.py, etc.).
• No new duplicated logic (including near-duplicates); consolidation is part of the task, not optional follow-up.
• No obvious module boundary violation; no new helper-shaped logic left outside oasis/helpers/ without a strong, documented reason.
• Any CLI option change is documented.
• Structured output/report changes stay aligned across oasis/schemas/, oasis/report.py, oasis/templates/reports/, and tests/test_report_schema.py.
• Backend/provider changes keep oasis/backends/ contracts, oasis/config.py env docs, and tests/test_backends_openai_compat.py aligned; call sites stay provider-agnostic.
• Incremental progress contract changes stay aligned across oasis/helpers/progress/__init__.py, report.py, web.py, dashboard JS, and contract tests when applicable (same spirit as report schema alignment).
• Change intent can be summarized with a conventional commit subject.