Snipara

Companion can resume saved workflow state when the next coding tool has access to the project folder with its saved .snipara state and artifacts, and the project’s Snipara MCP. One local Claude Code to Codex test preserved the plan, completed tasks, attempt counters and evidence, then passed 21 independent checks. This is checkpoint continuity, not migration of a live process. MCP-only restoration and Orchestrator controller transfer were not tested.

Snipara is a persistent, model-agnostic harness for coding agents. It combines project context and reviewed memory with Companion workflows and Orchestrator execution, external validation, independent review and controlled integration.

A coding agent harness is the system around a model that supplies context, scopes its work, manages execution and checks changes. The model proposes code; the harness defines the workflow and the conditions for accepting it.

Your model and coding client still reason and use tools. Snipara supplies project knowledge and workflow controls. MCP context access and native execution support are different capabilities; native execution depends on the configured adapter.

Context and memory ground the work. Companion tracks the plan and phases; Orchestrator can execute bounded tasks and bind validation and review to a candidate before controlled integration. A passing test alone does not approve a change.

An isolated Claude CLI journey reached acceptance with a test-local launcher. Qwen3 Coder Next through LM Studio passed external validation, then independent review refused a defect and blocked integration. These tests use one exposed development fixture and do not establish general model reliability.