Extraction Inventory — Twenty / Agno → @agentkit

What we extract, from where (pinned to a source commit), the target @agentkit package, and the dependency edges that block a clean lift — each with a proposed neutral port. Analysis only.

Sources (pinned)

Source	Repo	Commit	Role
ShareFlow AI backend	Rise-Experts/social_share → ai_backend/	731c93fe	The running Agno runtime being replaced
Twenty	Rise-Experts/twenty	454be88b71	Architectural patterns (not a running agent)
Twenty apps	Rise-Experts/twenty-apps	163ec22	How the social app consumes the runtime

Naming follows Agno's concepts; the target structure follows Twenty's layered module layout, which the @agentkit package map already mirrors.

Agno runtime → @agentkit (file-level)

ai_backend/app/assistant/ — orchestration

Agno file	Concept	Target package	Extract vs rebuild
agent.py	Agno Agent build (history, deps)	runtime + agents	rebuild (drop num_history_runs)
factory.py	assistant factory / composition	runtime (composition root)	rebuild
specialists.py, studio.py	multi-agent Team, per-role specialists	agents (versioned manifests)	rebuild as declarative manifests
context.py	contextvars (workspace_id, user_id)	core ExecutionContext	rebuild (explicit, passed not global)
dependencies.py	deps injected into run	context providers	rebuild
auth.py, middleware.py	request auth / middleware	authorization + transport (graphql)	rebuild
db.py, db_skills.py	persistence + workspace skills	persistence + skills	extract logic, re-port storage
skills.py	skill loading (LocalSkills)	skills	rebuild on the skills port
tools.py	@tool defs (generate_content, search_web)	tools (+ ShareFlow integration)	rebuild on the tool registry
llm_resolve.py	model resolution	models	rebuild

ai_backend/app/agents/ — the specialist team + adapters

Agno file	Concept	Target package	Notes
planner_agent, writer, tailor, reply, researcher, quality, vision	specialist agents	agents manifests + runtime	ShareFlow-specific manifests live in the integration
model.py	complete() provider adapter, LlmOverride	models (provider factory)	rebuild on Vercel AI SDK
memory.py	get_agno_db() → Agno PostgresDb/InMemoryDb	persistence (SessionStateStore)	AVOID Agno session pattern (see gaps)
tools.py	agent tools	tools	rebuild

ai_backend/app/llm/, core/, schemas/

Agno file	Concept	Target package
llm/registry.py	LlmOverride, SUPPORTED_PROVIDERS (pure data, no SDK)	models (registry)
llm/agno_factory.py	build Agno model per provider	models (provider factory)
core/browser,capture,readpage,readpdf,makepdf,safefetch,websearch	web/pdf/vision utilities	files / documents / pdf / rag + research tools
core/config.py	settings	composition/config
schemas/api.py	request/response contracts	graphql

ai_backend/skills/ — built-in skills (ShareFlow-specific)

analytics-reporting, document-generation, mermaid-diagrams, platform-media-rules, post-composition, publishing-safety, research-and-citation → the skill-loading mechanism goes to skills; the skill content is ShareFlow's and lives in the integration. Note: db_skills.py already enforces TEXT-ONLY workspace skills (no executable code) — matches the platform rule "arbitrary skill scripts disabled by default".

Twenty patterns → @agentkit

Twenty source	Pattern	Target
twenty-server/src/modules/workflow/workflow-executor, workflow-tools	step/tool execution, status machine	runtime + tools
twenty-server/src/modules/dashboard/tools	tool descriptor shape	tools
twenty-sdk / twenty-client-sdk (graphql-sse)	transport, subscriptions over SSE	graphql + the SSE adapter (SPEC #27)

Blocking dependency edges → proposed neutral port

Blocking coupling (Agno/Twenty)	Why it blocks	Proposed neutral port
Agno Agent(db=…) binds runtime to Agno's DB	can't swap storage	persistence stores + SessionStateStore
contextvars global request context	hidden global, not tenant-safe by construction	explicit ExecutionContext (host-built)
num_history_runs=20 history injection	token bloat + failures	budgeted context (docs/03) + compaction (docs/13)
SUPABASE_DB_URL / Supabase wiring in memory.py	provider lock-in	persistence (Postgres/Supabase adapters) + BlobStore
LlmOverride + agno_factory tie to Agno models	provider lock-in	models registry + provider factory (Vercel AI SDK)
Agno Team routing by role string	framework-specific	declarative agents manifests + runtime router

Gap notes (keep vs avoid)

• Context window (AVOID). Agno uses add_history_to_context=True, num_history_runs=20 — a fixed, model-unaware window plus manual token pointers. Replace with model-registry limits + per-section budgets + pruning (docs/03).
• Session handling (AVOID). Agno sessions via get_agno_db re-inject 20 turns and drop the rest (no summarization). Replace with session state + versioned thread-summary compaction (docs/13).
• Retry (KEEP). Agno's retry is the pattern worth carrying over — reimplemented Claude-style (backoff+jitter, honor retry-after, transient classes only, bounded attempts) and made safe by idempotency keys (docs/04). Now specified and surfaced to the frontend.