Aegis Client v0.6.1 - Execution Guidance Helpers

Adds SDK support for Aegis execution guidance returned by the backend under scope_data.execution.

What changed:

• Added result.execution helper
• Added result.model_tier
• Added result.context_mode
• Added result.max_retries
• Added result.allow_escalation
• Included execution guidance in summary and log records
• Added defensive handling for malformed or missing execution payloads
• Updated result model documentation

Why it matters:
Aegis can now expose tier-based execution guidance for downstream runtimes, including budget-aware model routing, context mode selection, retry limits, and escalation control.

Validation:

• python -m pytest -q
• 106 passed

Aegis 0.6.0 — Pipeline Sidecar & Simulation

Aegis 0.6.0

Aegis Shell has been redesigned as a zero-integration sidecar for AI pipelines.

You can now run your existing pipeline through Aegis without modifying your code:

aegis attach --cmd "python run_agent.py"

Aegis observes the run and generates a simulation report showing:

• instability signals observed
• controls Aegis would have issued
• projected impact (iterations avoided, retries prevented, scope reduced)
• recommended SDK integration points

New features:

• Added aegis attach command for pipeline simulation
• Exportable reports via --report
• Attach run history and comparison
• Improved control-state visibility via .aegis/control.json
• Enhanced session tracking and summaries

Positioning:

• Aegis Shell now demonstrates value without requiring integration
• Aegis SDK remains the enforcement layer for production pipelines

This release introduces a faster way to evaluate Aegis in real workflows.

v0.5.0 - Aegis Shell

Adds Aegis Shell: a tool-agnostic background control layer for AI-assisted workflows.

Highlights:

• aegis start background mode
• loop, scope drift, and diff growth detection
• Aegis-routed control recommendations
• session summary and aggregate stats
• doctor/reset lifecycle tools
• new Aegis Shell docs

Tests: 68 passed

Aegis Python SDK v0.4.0

Aegis Python SDK v0.4.0

Aegis 0.4.0 expands the SDK into the full runtime control surface for AI systems.

What’s new

Added two new public auto scopes:

• client.auto().context(...)
• client.auto().agent(...)

Aegis now supports five runtime control scopes:

• llm → control model-call behavior
• rag → control retrieved evidence and context
• step → control one workflow/action
• context → control information state
• agent → control multi-step workflow loops

Context control

Use context scope to clean, prioritize, and structure information before the next model or workflow action.

It supports:

• message prioritization
• tool result handling
• protected context
• carry-forward context
• structured context output
• trace and metrics

Agent control

Use agent scope to control multi-step workflow loops on top of your existing pipeline.

It supports:

• planned steps
• session-aware state
• carry-forward context
• tool-result integration
• stop/retry/escalation observability

Aegis does not replace your agent framework, model, RAG system, or tools. It controls runtime behavior around them.

SDK updates

New methods:

• client.auto().context(...)
• client.auto().agent(...)

Existing methods remain unchanged:

• client.auto().llm(...)
• client.auto().rag(...)
• client.auto().step(...)

Docs added/updated

• Expanded README for all five scopes
• Added request shape documentation
• Added implementation examples
• Updated architecture, scopes, result model, SDK overview, and migration docs
• Added examples for OpenAI-style LLM calls, RAG pipelines, and agent workflows

Package

Install or upgrade:

pip install --upgrade scelabs-aegis

Notes

• Legacy fallback remains only for llm, rag, and step
• context and agent require a newer Aegis backend with /v1/auto/context and /v1/auto/agent

Aegis Client v0.3.0

Aligns the public Aegis SDK with the live runtime scope model.

Highlights

• New primary SDK surface: AegisClient, AegisConfig, AegisResult
• Scope-first runtime interface:
  • client.auto().llm(...)
  • client.auto().rag(...)
  • client.auto().step(...)
• Unified result envelope with inspectable:
  • actions
  • trace
  • metrics
  • used_fallback
  • explanation
  • scope
  • scope_data

This release moves the client away from the old plan-first workflow and aligns it to the current runtime stabilization model.

Aegis v0.2.0 — SDK Integrations + Control Layer Upgrade

This release upgrades the Aegis SDK into a full runtime control layer for AI systems.

Highlights:

• AegisPlan abstraction for structured control
• Integrations:
  • OpenAI
  • LangChain
  • LangGraph
  • HuggingFace (local models)
• Framework-agnostic design
• Control helpers:
  • generation
  • retry
  • validation
  • coordination
  • tool usage
• Clean examples for all integrations
• Updated README for faster onboarding

Aegis now works as a control layer across AI systems, not just a single integration point.