mancitrus/system-prompts-and-models-of-ai-tools

mirror of https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools.git synced 2026-06-17 23:09:35 +00:00

Dealix Builder bcf545c22e feat(self-improving): Hermes-inspired Agent Platform — 6 layers + 30 endpoints + 76 tests + Private Beta launch

Security Curator (4 modules) — جدار الحماية الأول
- secret_redactor: 11 patterns (GitHub PAT, OpenAI/Anthropic/Supabase/WhatsApp/Moyasar/Sentry/Google/AWS/private keys); never returns raw secret
- patch_firewall: blocks .env / credentials.json / RSA keys; scans added lines for secret patterns
- trace_redactor: masks phones (+966...) and emails for PII safety
- tool_output_sanitizer: cleans tool outputs before they hit ledger/Proof Pack/UI/observability

Growth Curator (5 modules) — التحسين الذاتي
- message_curator: grades Arabic messages (0..100), detects 8 risky phrases, suggests Saudi-tone skeleton
- playbook_curator: scores playbooks by outcome (accept/reply/meeting/deal); winner/promising/needs_work/archive
- mission_curator: scores completed missions; ship_it_widely/iterate/rework_or_retire
- skill_inventory: deterministic 23-skill catalog across 5 layers
- curator_report: weekly Arabic summary "ماذا تعلمنا هذا الأسبوع"

Meeting Intelligence (5 modules) — ذكاء الاجتماعات
- transcript_parser: accepts Google Meet entries OR plain "Speaker: text" format
- meeting_brief: 6-section pre-meeting brief in Arabic (objective/questions/objections/offer/next-step)
- objection_extractor: 8 categories (price/timing/authority/trust/integration/competitor/results/complexity)
- followup_builder: email + WhatsApp drafts; live_send_allowed=False always
- deal_risk: 0..100 score from objections + missing next-step + decision-maker absence + days-since-touch

Model Router (5 modules) — موجّه النماذج
- provider_registry: 7 providers (Claude Sonnet/Haiku, GPT-4-class, GPT-4o-mini, Gemini Pro, Azure OAI KSA-region, Local Qwen Arabic-tuned)
- task_router: 10 task types × routing decisions with reasons_ar
- cost_policy: bulk → low; output > 1500 tokens → high
- fallback_policy: high-sensitivity workloads prefer KSA-region/self-hosted FIRST
- usage_dashboard: deterministic demo of all task routes

Connector Catalog (3 modules) — كتالوج التكاملات
- 14 connectors (WhatsApp Cloud, Gmail, Calendar, Google Meet, Moyasar, LinkedIn Lead Forms, Google Business Profile, X API, Instagram, Sheets, CRM, Website Forms, Composio, MCP Gateway)
- Each has launch_phase (1-4), risk_level, allowed_actions, blocked_actions, Arabic risk dossier
- WhatsApp blocks cold_send_without_consent; Moyasar blocks store_card_number; MCP requires allowlist

Agent Observability (5 modules) — مراقبة الوكلاء + التقييمات
- trace_events: SHA256-hashes user/company IDs; sanitizes payload/output before logging
- safety_eval: 7 rules (guarantee, scarcity_fake, medical_claim, financial, regulatory, personal_data, urgency); 0..100 → safe/needs_review/blocked
- saudi_tone_eval: positive markers (هلا, لاحظت, يناسبك) vs negative (تحية طيبة وبعد, synergy, leverage); arabic_ratio bonus
- eval_pack: 5 curated cases with expected verdicts
- cost_tracker: per workflow/provider/task_type aggregation

Routers (6 new) — 30 endpoints
- /api/v1/security-curator/{demo, redact, inspect-diff, sanitize-output}
- /api/v1/growth-curator/{skills/inventory, messages/grade, messages/improve, messages/duplicates, missions/next, report/weekly, report/demo}
- /api/v1/meeting-intelligence/{brief, brief/demo, transcript/summarize, followup/draft, deal-risk}
- /api/v1/model-router/{providers, tasks, route, cost-class, usage/demo}
- /api/v1/connector-catalog/{catalog, summary, status, risks, {key}}
- /api/v1/agent-observability/{trace/build, safety/eval, tone/eval, evals/run}

Tests (6 new files, 76 tests)
- test_security_curator: 16 tests (PAT detect, key redact, env diff block, payload scan, trace mask)
- test_growth_curator: 16 tests (Arabic grade, risky phrases, dup detect, playbook scoring, mission recommend, weekly report)
- test_meeting_intelligence: 13 tests (transcript parse, brief sections, objection extract, followup drafts, deal risk)
- test_dealix_model_router: 11 tests (every task → ≥1 provider, KSA-region for high sensitivity, cost class, primary override)
- test_agent_observability: 12 tests (trace hashing, safety verdicts, tone scoring, eval pack)
- test_connector_catalog: 11 tests (≥12 connectors, every has risk/blocked actions, WA cold-send blocked, Moyasar card-storage blocked)

Docs (8 new + 1 updated)
- AGENT_SECURITY_CURATOR.md (Arabic)
- GROWTH_CURATOR_STRATEGY.md (Arabic)
- MEETING_INTELLIGENCE.md (Arabic)
- MODEL_PROVIDER_ROUTER.md (Arabic)
- CONNECTOR_CATALOG.md (Arabic)
- AGENT_OBSERVABILITY_EVALS.md (Arabic)
- PRIVATE_BETA_LAUNCH_TODAY.md (Arabic) — go-checklist + offer + risks
- DEMO_SCRIPT_12_MINUTES.md (Arabic) — minute-by-minute demo flow
- FIRST_20_OUTREACH_MESSAGES.md (Arabic) — 7 personas + 3 follow-ups, all under safety/tone evals
- DEALIX_100_PERCENT_LAUNCH_PLAN.md — added §34 Self-Improving Agent Platform + §35 Private Beta Launch

Landing
- landing/private-beta.html — Arabic RTL, dark theme, pricing, 11 demo endpoints, safety banner

Test results
- 76/76 new tests pass
- Full suite: 663 passed, 2 skipped (missing API keys, unrelated)
- 0 existing tests broken

Safety
- All 6 layers honor approval-first, draft-only, no-live-send
- Hash user/company IDs before any trace
- No secrets in logs/embeddings/traces (3-layer defense: redactor + sanitizer + firewall)
- Saudi tone eval rejects "تحية طيبة وبعد" + "synergy" auto-corporate language
- Safety eval blocks "ضمان 100%" + medical claims + fake urgency
- Connector Catalog: WhatsApp blocks cold-send, Moyasar blocks card storage, MCP requires allowlist

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-01 16:30:18 +03:00

4.2 KiB

Raw Blame History

Security Curator — منظومة حماية وكلاء Dealix

القاعدة الأولى: لا سرّ يخرج من Dealix إلى log/trace/embedding/patch. الـ Security Curator هو الجدار الأول، يعمل قبل أي اتصال بأي قناة خارجية.

1. لماذا هذه الطبقة قبل أي tool live؟

Dealix يربط أدوات حساسة: WhatsApp Cloud, Gmail, Calendar, Moyasar, Google Meet, CRM. كل أداة فيها token، كل token خطر إذا تسرب. سابقاً تعرضنا لـPAT مكشوف، لذا قبل أي ربط حي:

يجب أن يمر كل log/trace من redactor.
يجب أن يمر كل diff من patch firewall.
يجب أن يمر كل tool output من sanitizer.
يجب ألا تخزّن أي assets مع secrets في الـembedding store.

2. الوحدات

الوحدة	الدور
`secret_redactor`	كشف وإزالة 11 نمط سر (GitHub PAT، OpenAI/Anthropic keys، Supabase JWT، WhatsApp/Moyasar/Sentry/Google API keys، AWS، private keys).
`patch_firewall`	يفحص الـunified diff قبل commit ويرفض الـ.env و service-account JSON و RSA keys.
`trace_redactor`	بالإضافة للأسرار، يخفي phones وemails داخل القيم النصية.
`tool_output_sanitizer`	يعقّم مخرجات الأدوات قبل إظهارها للمستخدم أو حفظها في الـledger.

3. أنماط الأسرار المكشوفة

github_pat              ghp_***
github_pat_legacy       github_pat_***
openai_key              sk-***
anthropic_key           sk-ant-***
supabase_service_role   eyJ.***.***
whatsapp_token          EAA***
moyasar_secret          sk_***_***
langfuse_secret         lf_sk_***
sentry_dsn              https://***@***/***
aws_access_key          AKIA***
google_api_key          AIza***
private_key_block       BEGIN PRIVATE KEY *** REDACTED ***

ومفاتيح JSON الحساسة تُستبدل بـ*** بناءً على substring match (case-insensitive) لـ: api_key, apikey, secret, token, password, authorization, access_token, refresh_token, client_secret, private_key, ssn, credit_card, card_number, cvv, iban, moyasar_secret.

4. Patch Firewall

أي PR قبل ما يدخل الريبو:

ملفات محظورة: .env, .env.local, .env.staging, .env.production, credentials.json, service-account*.json, id_rsa, *.pem, *.p12, *.pfx.
أسرار في الأسطر المضافة: أي line يبدأ بـ+ يُمرر من detect_secret_patterns.
الناتج: PatchFirewallResult{safe, reasons_ar, blocked_files, secret_findings}.

GitHub Push Protection يقبض الأسرار قبل push، لكن لا تعتمد عليه وحده — Patch Firewall يعمل في طبقة التطوير المحلية + CI.

5. Tool Output Sanitizer

قبل أن يصل أي مخرج إلى:

الـAction Ledger
الـProof Pack
الواجهة (UI / WhatsApp / Email)
Langfuse / Sentry

يمر عبر sanitize_tool_output(output) الذي يُرجع:

safe: bool
redacted: <نفس الشكل، مُعقّم>
notes_ar: ["تمت إزالة قيم حساسة من المخرج: ..."]

6. Endpoints

GET  /api/v1/security-curator/demo
POST /api/v1/security-curator/redact
POST /api/v1/security-curator/inspect-diff
POST /api/v1/security-curator/sanitize-output

7. اختبارات الأمان (16 test)

detect_github_pat لا يُرجع السر الخام أبداً.
redact_openai_key يستبدل بالـmask.
scan_payload يخفي api_key وtoken.
inspect_diff يحظر .env.
inspect_diff يحظر سراً مكتوباً داخل سطر مضاف.
redact_trace يخفي phones/emails مع الحفاظ على الـdomain للسياق.
sanitize_trace_event يحفظ event_type/agent_name/latency_ms ويعقّم payload.

8. ما لا تفعله هذه الطبقة

لا تكشف السر الخام في الـlogs أبداً.
لا تُرجع payload فيه token.
لا توقع على diff فيه secret.
لا تستبدل أو تعطّل GitHub Push Protection — هذه الطبقة إضافة، لا بديل.

4.2 KiB Raw Blame History Unescape Escape