编程

Alazab Global Context

试用

Shared production context, operational paths, orchestration policies, agent routing, approvals, deployment and execution standards for all Alazab Portal AI agents, models, sessions and nodes.

它能做什么

Shared production context, operational paths, orchestration policies, agent routing, approvals, deployment and execution standards for all Alazab Portal AI agents, models, sessions and nodes.

技能文档

Alazab Portal AI — Global Skills

Purpose

This file defines the shared operational skills loaded by default for all agents, models, runtimes, sessions, projects, and execution nodes in the Alazab AI Portal.

Canonical global context:

/etc/portal-ai/shared/alazab-portal-ai-global-context.json

The same file must be used by both production servers:

portal-ai.alazab.com
portal-ai.alazab.cloud

Scope

scope: global
auto_attach: true
applies_to:
  agents: ["*"]
  models: ["*"]
  runtimes: ["*"]
  sessions: ["*"]
  projects: ["*"]
  nodes: ["*"]

Context Loading Order

Every agent or model must load context in this order:

  1. Global context
  2. Organization policies
  3. Project context
  4. Agent role
  5. Mission context
  6. Task context
  7. Current user instruction

Later layers may override earlier layers only within their permitted scope.

Global Skills

1. Mission Planning

Convert a user request into:

  • Mission
  • Workstreams
  • Tasks
  • Dependencies
  • Risks
  • Acceptance criteria
  • Rollback plan

No execution starts before the target environment, node, paths, and success criteria are known.

2. Agent Routing

Select the correct agent, model, or runtime based on:

  • Task domain
  • Required tools
  • Target node
  • Data sensitivity
  • Cost
  • Latency
  • Context-window requirements
  • Review requirements

Supported execution targets include:

  • Codex CLI
  • Gemini
  • Azure AI Foundry agents
  • Azure-hosted models
  • Ollama models
  • PTY shell executor
  • Specialized review agents

3. Multi-Agent Orchestration

Support:

  • Sequential execution
  • Parallel execution
  • Supervisor-worker execution
  • Independent review
  • Retry and recovery
  • Escalation to owner approval
  • Consolidation of outputs

Agents must not duplicate work unless parallel comparison is explicitly requested.

4. Project Context Management

Every active project should maintain:

PROJECT.md
ARCHITECTURE.md
DECISIONS.md
CURRENT_STATE.md
KNOWN_ISSUES.md
SERVICE_MAP.md
EXECUTION_PLAN.md
TASKS/
REPORTS/
RUNS/
ARTIFACTS/
BACKUPS/

Agents must update project state after material changes.

5. Production Path Awareness

Core server

/var/www/core/portal-ai
/etc/portal-ai/core
/var/lib/portal-ai/core
/var/log/portal-ai/core
/var/backups/portal-ai/core

Cloud server

/var/www/apps/portal-ai
/etc/portal-ai/cloud
/var/lib/portal-ai/cloud
/var/log/portal-ai/cloud
/var/backups/portal-ai/cloud

Shared paths

/etc/portal-ai/shared
/etc/portal-ai/shared/skills
/etc/portal-ai/shared/policies
/etc/portal-ai/shared/schemas
/etc/portal-ai/shared/templates

Agents must never invent production paths when the global context already defines them.

6. Safe Production Execution

Before changing production systems:

  1. Inspect current state.
  2. Identify affected services and files.
  3. Create backup or snapshot.
  4. Record rollback procedure.
  5. Apply the smallest valid change.
  6. Run validation tests.
  7. Verify service health.
  8. Record the result.

Non-destructive, reversible operations may proceed automatically.

Destructive or irreversible operations require explicit owner approval.

7. Approval Control

Approval is required for operations including:

DROP DATABASE
DROP SCHEMA
rm -rf
pg_dropcluster
initdb on existing data
pg_resetwal
force push
DNS deletion
firewall flush
secret rotation

Approval requests must state:

  • Exact command
  • Target node
  • Affected resources
  • Expected impact
  • Backup status
  • Rollback procedure

8. Evidence-Based Diagnosis

Agents must:

  • Inspect before modifying.
  • Use logs, configuration, process state, and service state as evidence.
  • Avoid assumptions.
  • Stop when evidence indicates corruption or an unsafe state.
  • Distinguish temporary runtime files from critical data or catalog files.

9. API and Integration Design

All Core-to-Cloud communication must use authenticated internal APIs over the private network.

Default internal endpoints:

Core:  http://10.77.0.1:8200
Cloud: http://10.77.0.2:8100

Required request identity:

Authorization: Bearer 
X-Alazab-Node: 
X-Alazab-Timestamp: 
X-Alazab-Signature:

Execution nodes must not write directly to the Core database.

10. Codex Session Management

Codex execution must support:

  • Session creation
  • Task submission
  • Follow-up messages
  • Transcript retrieval
  • Interrupt
  • Resume
  • Timeout handling
  • Exit-code capture
  • Artifact collection
  • Independent review

Default Codex paths:

/var/lib/portal-ai/cloud/workspaces/codex
/var/lib/portal-ai/cloud/sessions/codex
/var/log/portal-ai/cloud/codex

11. Model Selection

Use the smallest model that can complete the task reliably.

Recommended routing:

  • Architecture and orchestration: az-model-core
  • Maintenance operations: az-model-maint
  • Financial analysis: az-model-finance
  • Embeddings and semantic retrieval: az-models-text
  • Speech-to-text: az-modelspeech
  • Voice interaction: az-model-voice
  • Local coding: deepseek-coder-v2:16b
  • Local vision: llama3.2-vision:11b
  • Local OCR: deepseek-ocr:latest

12. Quality Review

No task is complete until the result is checked against:

  • User request
  • Acceptance criteria
  • Current architecture
  • Security boundaries
  • Production paths
  • Test results
  • Rollback readiness

High-impact tasks require independent review by a different agent or model.

13. Logging and Audit

Every execution event should include:

timestamp
node_id
project_id
mission_id
task_id
agent_id
session_id
event_type
status
message

Logs must be structured as JSONL where supported.

14. Error Handling

When an operation fails:

  1. Stop dependent steps.
  2. Preserve logs and outputs.
  3. Capture exit code.
  4. Determine whether rollback is needed.
  5. Do not conceal partial failure.
  6. Report the exact failing component.
  7. Resume only from a verified safe state.

15. Secret Handling

Secrets must be resolved from Infisical at runtime.

Production secret manager:

https://env.alazab.com

Rules:

  • Never write secrets into this file.
  • Never commit secrets.
  • Refer to secrets by environment-variable name.
  • Redact credentials from logs and reports.

16. Deployment

Production deployment uses release directories and an atomic current symlink.

Core:

/var/www/core/portal-ai/releases/
/var/www/core/portal-ai/current

Cloud:

/var/www/apps/portal-ai/releases/
/var/www/apps/portal-ai/current

Required deployment checks:

  • Configuration validation
  • Backup
  • Dependency lock verification
  • Tests
  • Security audit
  • Health check
  • API contract check
  • Worker heartbeat
  • Smoke test

17. Rollback

Application rollback:

Switch current symlink to the previous verified release.
Restart affected services.
Run health and smoke tests.

Database rollback must use an approved recovery plan and a verified backup.

18. Communication Standard

Agent responses must be:

  • Direct
  • Evidence-based
  • Actionable
  • Specific to the current request
  • Free of invented details

Operational reports must clearly separate:

  • Confirmed state
  • Changes made
  • Validation results
  • Remaining risks
  • Next required action

Mandatory Runtime Variables

PORTAL_AI_NODE_ID
PORTAL_AI_NODE_ROLE
PORTAL_AI_ENVIRONMENT
PORTAL_AI_GLOBAL_CONTEXT
PORTAL_AI_SERVICE_TOKEN
PORTAL_AI_HMAC_SECRET
PORTAL_AI_DATABASE_URL
PORTAL_AI_REDIS_URL
PORTAL_AI_INFISICAL_PROJECT

Canonical Skill Installation Path

/etc/portal-ai/shared/skills/alazab-global-context/SKILLS.md

Repository development path:

skills/alazab-global-context/SKILLS.md

Final Rule

The global context file is the source of truth for infrastructure, paths, nodes, services, agents, models, and production rules.

This skills file defines how every agent and model must use that context during planning, execution, review, and reporting.

相关技能

ALab Global Admin

Use when administering an ALab home with root authority, including home bootstrap, root and project-admin credential management, project initialization and h...

6 次安装

AI Agent Governance

Register an autonomous agent's identity, check a runtime policy decision before it acts, and log a hash-chained attestation of what it did. Backed by A2Z SOC...

ALab Project Controller

Use when managing one existing ALab project with a project admin key to create and coordinate experiments, validate and adjust project configuration, manage...

6 次安装

Global Agent Rules

Shared behavioral rules for ALL Donald's agents. Universal: correction protocol, time check, pre-action checklist, learning loop, mistake breaker.

ALab Experiment Worker

Use when operating inside one ALab experiment worktree with that worktree token context to inspect status, edit candidate source, run evaluations, submit fin...

5 次安装