Add independent Deep Agents migration skill

[adtyavrdhn]

Summary

• add migrate-deep-agents-to-pydantic-ai as an independent marketplace plugin and byte-identical standalone skill
• inventory Deep Agents projects and map prompts, state, tools, planning, subagents, virtual files, approvals, persistence, sandboxes, streaming, and deployment boundaries
• guide migrations through composable Pydantic AI primitives, optional Harness capabilities, and explicit application-owned services
• include focused playbooks for Open SWE and the substantial official Deep Agents research, content, SQL, coding, and GTM projects

Packaging and ownership

This is a repository-owned plugin at plugins/migrate-deep-agents-to-pydantic-ai/. It is not part of the upstream-synced pydantic-ai-harness plugin.

Pydantic AI Harness is an optional, version-sensitive capability library used only when an installed capability contract fits. The migration skill starts from Pydantic AI core abstractions and keeps tenancy, durable state, queues, remote sandbox isolation, artifacts, memory policy, approvals, and side-effect recovery with their application owners.

Skill structure

The 46-line SKILL.md routes into six purpose-named references:

• pydantic-ai-architecture.md
• migration-map.md
• implementation-recipes.md
• project-playbooks.md
• semantic-differences.md
• validation.md

Exploratory spike reports, workaround catalogs, copyable migration assets, and a framework-version-sensitive inventory scanner are intentionally not shipped. Inventory follows repository imports, factories, middleware, configuration, call sites, and application boundaries. Project-specific probes are disposable; accepted behavior is promoted into the migrated project test suite.

Behavioral guidance

The skill is upfront about semantic differences without treating them as dead ends. It shows how to express planning, typed delegation, fan-out, memory, virtual files, background work, approvals, streaming, and durable workflows through Pydantic AI agents, typed deps, outputs, toolsets, capabilities, ProcessHistory, deferred tools, pydantic_graph, durable-execution integrations, and application services.

Current Harness claims were checked against v0.5.0-6-g1ad638f; experimental capabilities are labeled as such, and absent DynamicWorkflow/Memory APIs are not recommended.

Validation

• skill-creator validation for plugin and standalone copies
• byte-identical plugin/standalone sync check
• JSON validation for plugin and marketplace manifests
• Markdown relative-link and heading-anchor validation
• syntax validation for all seven Python documentation fences
• executable Pydantic AI capability construction check
• executable typed-delegation construction check
• executable deferred-approval stop/resume probe
• terminating FunctionModel protocol example probe
• current Pydantic AI/Harness public-export audit
• adversarial skill/spec review and low-effort cold-use migration trial
• verified official LangChain project and example references