cerredz/Vidbyte-SDK

Vidbyte is an agent engineering platform for building, evaluating, instrumenting,
and distributing AI workflows. The Vidbyte SDK is the Python package surface for
that platform: it gives developers composable agents, tools, middleware, context
management, MCP server integration, prompts, evals, provider adapters, pipelines,
validated workflows, and tracing primitives.
This repository is intentionally focused on reusable SDK abstractions. Private
Vidbyte service logic, proprietary learning systems, hosted scoring, and database
access remain outside this package.
What You Can Build
- Agent applications with explicit system prompts, provider/model configuration, tools, context, and trace behavior.
- Tool-using agents with local Python functions, MCP-backed tools, permission policies, and provider-native schemas.
- Runtime policies with deterministic middleware for rate limits, budgets, retries, audit logs, compaction, and safety checks.
- Durable agent sessions that checkpoint, resume, fork, batch fork, tag, export/import, and summarize usage across long-running work.
- Open harness implementations with config-as-source-of-truth identity, Session-backed capture, and consented, redacted trajectory datasets.
- MCP Studio servers that expose Vidbyte agents, tools, prompts, and pipelines to MCP-compatible clients.
- Local eval suites with reusable graders, concurrency controls, and run registries.
- Agent pipelines that compose specialized agents through sequential, parallel, conditional, and map-reduce topologies.
- Typed state-machine workflows with validation gates, conditional branches, cycles, retries, and declared jumps.
- Prompt libraries, context-window algorithms, and trace artifacts that make long-running agent work easier to inspect.
JevAgent continuation checks
JevAgent can run named done checks whenever its main agent attempts to finish. Enable whole-task completion evidence alongside other checks through JevContinualSettings.checks:
COMPLETION_EVIDENCE checks whether the final answer's overall complete, incomplete, or blocked status matches the requested outcomes and observations in the run. An unqualified final answer implies completion, even when it does not say “done.” Honest incomplete or blocked reports can pass when the run evidence supports them. The final answer's own claim that external work happened does not count as evidence for that work. The result is available as agent.response.done[JevDoneCheck.COMPLETION_EVIDENCE]; if Jev cannot evaluate it, the check fails open.
Layer Guide
| Layer | Role |
|-------|------|
| vidbyte.agents | Executable agent actors, runtimes, inferred runner selection, handoff, and agent registries |
| vidbyte.config | Safe YAML parsing into declarative agent settings (with nested tools/middleware) or a harness spec, plus agent construction from those settings alone |
| vidbyte.cli | Unified console command for SDK developer surfaces, currently vidbyte-sdk skills |
| vidbyte.context | Structured context items, context windows, compaction, algorithms, and handoff models |
| vidbyte.evals | Local eval cases, suites, runners, graders, registries, and result summaries |
| vidbyte.harnesses | The Harness base class: config-as-source-of-truth identity, Session-backed capture, and consented, redacted trajectory export |
| vidbyte.lib | Shared dataclasses, enums, registries, errors, runners, config, and tracing contracts |
| vidbyte.mcp_server | Stdio MCP Studio server for exposing agents, tools, prompts, and pipelines |
| vidbyte.middleware | Deterministic runtime hooks and built-in policy, safety, retry, budget, and compaction middleware |
| vidbyte.paradigms | Thin runnable paradigm harness scaffolding built from agents, tools, context, prompts, middleware, trace, pipelines, and evals |
| vidbyte.pipelines | Multi-agent pipeline topologies with a string-in/string-out contract |
| vidbyte.prompts | Static prompt assets, enum-keyed lookup, direct imports, and prompt families |
| vidbyte.providers | Provider adapter factories for text, image, video, audio, embeddings, and streaming |
| vidbyte.sessions | Durable checkpoint-DAG persistence, stores, scope, usage rollups, and portable bundles |
| vidbyte.sources | Artifact-to-context loaders for public documents, llms.txt, fetch/cache, hashing, and selection |
| vidbyte.shared | Reserved shared namespace; currently no stable public symbols |
| vidbyte.tools | Tool contracts, decorators, catalogs, execution, MCP bridges, and permissions |
| vidbyte.trace | Trace facade, debug tracer, provider tracers, and continual trace artifacts |