Friday capability status

This reference describes the reviewed 0.1.0a3 implementation, not a roadmap or a claim of production certification. Framework APIs and their Chat integration can have different maturity. Installation/extras supply dependencies, not runtime enablement. Use the board for current work authority.

Application and composition

Capability	Current status	Evidence / guide
Maintained terminal application	`friday-chat`; explicit YAML; fake default and opt-in real model/memory/tool assembly	CLI, Chat assembly
`friday` executable from CLI component	Declared placeholder entry point, not a functioning application	CLI main module
Headless lifecycle	`RuntimeConfig` and `RuntimeKernel` compose configured services/profiles	Architecture
Automatic Chat first-run setup	No `init`/`doctor` implementation or implicit OS-config lookup	Desktop setup
Root distribution	Metadata-only aggregate, no importable root module	Package map

Models and memory

Capability	Current status	Evidence / guide
Chat/embedding transport	LiteLLM-backed; pinned 1.100.1	LLM README
Canonical model library	Strict versioned endpoint/model profiles, env credential names and role bindings	Configuration
Typed requests/events	Distinct text/reasoning/tool/usage/finish surfaces; control support is profile-specific	Advanced integration
Built-in normalized controls	LM Studio + OpenAI-compatible Chat Completions + Qwen3.8; not universal provider parity	LLM README
Episodic/vector/graph memory	Implemented adapters/controllers; persistence and embeddings explicitly configured	Memory
Extraction/consolidation	Implemented, can call configured models; quality and completion evaluated separately	Memory
Context pruning	Formatting overhead reserved; prior-session history included and pruned first under pressure	Budget guide
Background compaction	Opt-in Chat monitor/compactor; cooldown, severity and result limits apply	Budget guide
Lifecycle and maintenance	Staged scoring, explicit mutation/recovery; capabilities/preflight can refuse operations	Memory
Transcript persistence/export	SQLite/JSONL stores and CSV/JSONL/Markdown views; raw content is sensitive	Transcript example
Pinecone/Neo4j/Ollama-specific adapter classes from old guides	Not shipped under those names; custom backend work is separate	Architecture

Agents, workflows and tools

Capability	Current status	Evidence / guide
Interactive agents	Implemented `InteractiveAgent`/`ReplAgentBase`; compatibility `CopilotAgent` name	Agents
Framework orchestration	Single active agent, cached instances, explicit routing/delegation and configured factories	Multi-agent guide
Chat specialist-agent execution	Deterministic `FakeOrchestrator` in fake and real modes; not independent model agents	Assembly
Autonomous agent	Placeholder; `run()` raises `NotImplementedError`	Source
Workflow hierarchy	Canonical workflow/pipeline/task models, loader/executor, contracts and dispatch state	Workflows
Chat workflow execution	Real mode uses flat engine plus deterministic agent adapter; canonical listing does not imply runnable bundles	Adapter
Namespace-aware sharing	Additive wrapper, policy and promotion journal; host authentication/metadata protection still required	Multi-agent memory
Native tools	Function and typed class surfaces, explicit registry/selection; implementations and stubs coexist	Tools, catalog
Shell/subprocess isolation	Host privileges/policy checks; not arbitrary-code containment	Execution backends
Docker/mock backends	Explicit Docker provisioning; mock does not execute code	Sandbox README
DSPy/knapsack optimizer	Optimization package skeleton only; no implemented training/compiler/evaluator API	Optimization README

MCP and observability

Capability	Current status	Evidence / guide
MCP modern tools	Explicit 2026-07-28 stdio/Streamable HTTP clients and modern server APIs	MCP matrix
MCP compatibility	Explicit 2025-11-25 HTTP client; retained 2024-11-05 legacy APIs/default stdio	MCP matrix
Integrated MCP	Disabled default; canonical filters, budgets, strict HTTPS trust/auth/network policy	Client guide
Catalog refresh	Programmatic generation-bound refresh; integrated stop/start rediscovery	Refresh contract
MRTR/input-required	Unsupported continuation; explicit categorical failure, no callback/state forwarding	Client results
Optional MCP roles	No modern Tasks, roots/sampling/elicitation, subscriptions or integrated resources/prompts; no OAuth discovery/login/refresh	Client, server
Provider-neutral telemetry	Implemented interface/providers/factory/lifecycle and closed LLM/MCP/execution schemas	Telemetry README
Chat diagnostics	Default-disabled bounded local debug observer, turn correlation and borrowed injection	CLI/configuration
Owned OpenTelemetry	Optional validated SDK/API 1.39.1, console/OTLP-gRPC, verified-TLS policy; no automatic global bridge	Advanced telemetry
Provider-neutral Langfuse	Queued, not the legacy callback/extra path	Board
Universal content sanitization	Not guaranteed; closed telemetry, raw transcripts, native helpers and checked MCP have different boundaries	Security

Optional domain/audio integrations

Capability	Current status	Evidence / guide
Finance	Massive data wrapper; plan/freshness dependent; tier probing placeholder and limited exception normalization	Finance README
Moltbook	Typed client with read/mutation operations; application supplies auth and action authority	Moltbook README
Standalone TTS/wake word	ElevenLabs/Picovoice adapters with explicit provider/device setup	Speech
Built-in STT	Picovoice transcription capture not wired; custom adapter required	Speech
Chat provider audio	Deterministic `FakeSpeechAdapter` even in real mode	Speech

Interpreting status

“Implemented” means the named source surface exists with the documented limits. It does not certify every provider, public deployment or semantic output. “Placeholder” means a declared API is not functional. “Queued” describes board work without implementation authority. Guide review dates describe evidence freshness and do not change those product/lifecycle states.

Guide index · Capability status