Feature Matrix

This page shows what users can rely on today and what is planned next.

Agents & Workflows (v0.3.0)

Feature	Status	Notes
create_agent() + profiles	Implemented	Composable agents; define_agent_profile(), init_agent().
Workflows	Implemented	workflows/**/run(ctx), pyflue run <wf>, POST /workflows/{name} (accepted / ?wait=result / SSE).
FlueContext	Implemented	ctx.id, ctx.payload, ctx.env, ctx.req, ctx.log, ctx.init(agent). (PyFlueContext aliased.)
Persistent agent instances	Implemented	create_agent default-export served at POST /agents/{name}/{id} with session continuity; no run id.
Session stores	Implemented	SessionStore protocol + InMemorySessionStore + SQLiteSessionStore.
dispatch()	Implemented	Async agent input + DispatchReceipt; POST /agents/{name}/{id}/dispatch. Process-memory admission.
Chat integration pattern	Implemented	Verified webhooks dispatch to continuing agent sessions; outbound replies are explicit tools. See Chat.
Operation events	Implemented	operation_start/operation with operation_id/instance_id.
Generation telemetry	Implemented	turn_request/turn events carry model and token usage; mapped to gen_ai.* spans.
OpenTelemetry	Implemented	create_opentelemetry_observer() (pyflue[otel]); workflow, operation, generation, tool, task, and compaction spans.
Model/provider configuration	Implemented	Provider-qualified model strings, thinking levels, and provider endpoint overrides. PyFlue delegates catalog-level routing to harness backends.
Host local() sandbox	Implemented	Real fs + subprocess shell, opt-in env allowlist.
Subagent profiles	Implemented	task(agent="name") selection; profile_to_role()/role_to_profile() bridge.
Packaged skills	Implemented	load_skill(path) imports a SKILL.md as a reusable Skill.
WebSocket	Implemented	Agent (multi-prompt) + workflow (one run); client agents.connect/workflows.connect.
ToolDefinition	Implemented	Canonical name for ToolDef (aliased).
Source layouts	Implemented	Discovers src/agents and src/workflows plus legacy root/.agents/.pyflue.
Durable execution	Not ported	Cloudflare-specific run recovery. See Parity with Flue.
Durable dispatch admission	Planned	dispatch() currently uses process-memory admission; use an application queue for production delivery.

Core

Feature	Status	Notes
Python package	Implemented	pyflue package with console script.
Pydantic AI backend	Implemented (default)	Typed, model agnostic loop, no LangChain. Included with PyFlue.
DeepAgents backend	Implemented (optional)	pyflue[deepagents]. Built on LangChain and LangGraph.
OpenAI Agents backend	Planned	Dependency pinned, runtime not implemented.
Google ADK backend	Planned	Dependency pinned, runtime not implemented.
Markdown skills	Implemented	.agents/skills/**/*.md.
Project instructions	Implemented	AGENTS.md and CLAUDE.md from project files and active sandbox context.
Sessions	Implemented	SQLite-backed history.
Automatic compaction	Implemented	Threshold compaction before turns and one prompt overflow retry.
Roles	Implemented	Markdown roles from .agents/roles/**/*.md.
Task sessions	Implemented	session.task() creates isolated child history with shared sandbox.
Built-in agent tools	Implemented	Filesystem, shell, search, glob, and task tools.
Abort/cancel	Implemented	session.abort() cancels active prompt, stream, task, and shell operations.
Event callbacks	Implemented	on_event emits lifecycle, command, task, and compaction events.
Virtual sandbox	Implemented	Persistent per-session workspace with path boundary checks, metadata helpers, binary file helpers, and policy-gated shell execution.
Shell policy	Implemented	Requires allow_shell=True; optional allowed_commands grants and compound-command blocking.
Secret grants	Implemented	Secrets are only mounted into sandbox env for calls that request them.
Write policy	Implemented	Requires allow_write=True.
DeepAgents file transfer	Implemented	Upload and download methods.
Typed outputs	Implemented	Pydantic TypeAdapter with retry repair loop and free-form JSON extraction.
Structured commands	Implemented	PyFlueCommand exposes reusable shell or callable commands as prompt tools.
Python client	Implemented	PyFlueClient supports health, route listing, prompt, stream, and agent calls.
Model override	Implemented	session.prompt(..., model="...") and session.skill(..., model="...").
File-based agent routing	Implemented	agents/*.py and .agents/*.py expose /agents/{name}/{id} routes.
Route triggers	Implemented	Agent files can declare triggers = {"webhook": True}.
HTTP error envelopes	Implemented	Webhook requests return stable JSON error envelopes.
CLI init	Implemented	Scaffolds project files.
CLI run	Implemented	Runs a prompt with optional --stream event output.
CLI skill new	Implemented	Scaffolds a skill.
CLI build	Implemented	Generates Docker/FastAPI and selected CI/platform artifacts.
CLI dev	Implemented	Starts the FastAPI app with Uvicorn reload.
CLI routes	Implemented	Lists discovered file-based agent routes.
CLI deploy	Implemented	Generates target artifacts and can invoke known provider CLIs when installed.
Remote sandboxes	Implemented	Daytona, E2B, Modal, and Runloop adapters.
Monty Python backend	Implemented	Safe host-side Python execution via pyflue[monty].
Monty state dump/load	Implemented	Serialize and restore Monty REPL state.
Monty dataclass registry	Implemented	session.register_python_dataclass(...).
Monty resource limits	Implemented	resource_limits={...} forwards to Monty.
Streaming/events	Implemented	session.stream(...), pyflue run --stream, and SSE endpoint.
MCP	Implemented	Direct mode and search/execute mode.

Sandbox Providers

Provider	Status	Notes
Virtual	Implemented	Persistent per-session workspace with path boundary checks.
Daytona	Implemented	Optional dependency: pyflue[daytona].
E2B	Implemented	Optional dependency: pyflue[e2b].
Modal	Implemented	Optional dependency: pyflue[modal].
Runloop	Implemented	Optional dependency: pyflue[runloop].

Live provider smoke tests are available behind PYFLUE_LIVE_SANDBOX_TESTS=1 and skip automatically unless the matching provider credentials are present.

Deployment Targets

Target	Status	Notes
Docker/FastAPI	Implemented	Python replacement for a JavaScript server target.
GitHub Actions	Implemented	Generates a manual workflow.
GitLab CI/CD	Implemented	Generates a manual pipeline job.
Railway	Implemented	Uses the Docker/FastAPI app.
Render	Implemented	Uses the Docker/FastAPI app.
Fly.io	Implemented	Uses the Docker/FastAPI app.
Cloudflare	Partial	Generates wrangler.toml; full Python container guide is still needed.
Vercel	Implemented	Generates vercel.json plus Python app artifacts.
Netlify	Implemented	Generates netlify.toml plus Python app artifacts.