refactor(integrations): evaluate agent plugin baseline and shared lifecycle

[PsiACE]

Draft for architecture evaluation, related to #1410.

Use integrations/agent-plugin/powercontext/ as the editable Skills and MCP baseline. One generator projects it into native resources and derives the minimum toolkit from the client contract. Scope, Memory, current-work Handoff, and candidate inspection share the same methodology; adapters handle native formats, events, permissions, and output.

Host rules, Target profiles, setup, and discovery stay under integrations/distribution/, loaded from the selected repository source/ref. They require neither a separate management installation nor a client release to change. The installed client supplies Core Hook execution and shared Server diagnostics.

Targets

Targets (15)
+-- Python hosts: Codex, Claude Code, Hermes, WorkBuddy, MiniMax
+-- TS hosts:     Pi, DSH, OpenCode, OpenClaw
+-- Python apps:  Bub, LangChain, LangGraph, Pydantic AI, OpenDAL
`-- Portable:     Agent Plugin

Setup and generation

Agent Plugin baseline + API contract
                 |
setup <target> -> Python HostAdapter
                 +-- client / uvx / npx + transport policy
                 `-- generate -> native install -> verify

Native packages carry generated tool metadata instead of embedding duplicate schemas in bundles. Existing private MCP settings survive regeneration. Direct host loading from a checkout requires make agent-resources. Ordinary work transfer returns a temporary Handoff; durable commit requires explicit milestone intent.

Doctor

doctor <target> -> saved source -> HostAdapter
                 +-- prerequisites + discovery + transport
                 `-- --server --+
/pc doctor (connection) ---------+-> Python Server doctor
                                    transport -> live -> ready -> capabilities

See the architecture and guide. Superseded Skill/MCP copies, schema declarations, installation instructions, and staged-transfer evaluation fixtures are removed.

Validation

• make check; License Eye 0.9.0; full Python unit and end-to-end CI matrices (3.11–3.14).
• Pi 106, DSH 213, OpenCode 64, and OpenClaw 75 tests; DSH also runs 9 HTTP and 6 real-host acceptance tests. Hermes and common Handoff qualification cover native behavior.
• All 15 distributions regenerate deterministically; native package contents and documentation links checked.
• Latest upstream merged; no HTTP contract or persisted format changes.

AI usage

Implementation, refactoring, validation, and PR preparation used OpenAI Codex (GPT-6).