Skip to main content

Crate superfluid_daemon

Crate superfluid_daemon 

Source
Expand description

superfluid — the daemon.

Re-exports§

pub use codec::MockCodec;
pub use codec::NoCodec;
pub use codec::TextCodec;
pub use codec::Utf8Stream;
pub use runtime::EngineHost;
pub use runtime::WorkerSpec;
pub use store::CommittedEvent;
pub use store::SessionStore;
pub use wal::EventBody;
pub use wal::GenParams;
pub use codec::BundleCodec;

Modules§

anthropic
api-anthropic.
api
api-native.
apiweb
api-web.
atem
batches
bus
The session event bus.
capabilities
What a loaded model can do, as its runtime reported it in the Link W hello.
cli
The superfluid command line: its help text, the serve flag table and the defaults that let superfluid serve <model> start with no other flag.
codec
The daemon-side text codec seam.
compat
superfluid compat [--json]: what this build was built and tested for, per runtime.
completion_bucket
Keystroke-storm protection for FIM completions.
dylib_tokenizer
A runtime’s tokenizer, loaded from its package at run time.
export
files
File store for the OpenAI /v1/files API (the substrate for /v1/batches).
fleet
The head daemon’s fleet edge.
fleet_join
superfluid node join: this machine serves a model to a fleet head it dials, and the head takes nodes that dial in.
fleet_manager
Head-side multi-node orchestration.
fleet_mdns
The head on the local network, as _superfluid-fleet._tcp, so a node finds it with no address given.
fleet_openai
api-openai over a FLEET (multi-node head).
fleet_token
superfluid fleet token: the secret a head and its nodes authenticate and encrypt their link with, kept at $SUPERFLUID_HOME/fleet/token.
harmony
Harmony, the response format of OpenAI’s gpt-oss models.
installs
Runtime installs.
keypolicy
Per-API-key serving policy (--key-policy <file.json>).
launch
superfluid launch <agent>: run a coding agent pointed at a local server, starting one when none answers. Each agent is a row of AGENTS whose Plan is what would run.
linked
The runtimes this build links, for source builds (--features basert, llamacpp, mlx).
media
The content-addressed media pool.
metrics
nodeagent
The node-agent.
ollama
Ollama inference over the shared daemon, policy checks and generation core.
openai
api-openai.
otlp
park
Park artifacts.
pins
Pinned prefixes.
pressure
OS memory-pressure sources.
pulls
Models by id.
qos
ratelimit
Token-bucket rate limiter keyed by client IP (--rate-limit N).
registry
Multi-model host management.
response_store
Stored responses for the OpenAI /v1/responses API: which session each response ran in, so previous_response_id can continue it.
runtime
runtime_options
--<runtime>-<name> <value> on superfluid serve: an option of one runtime, checked against what its adapter declares and forwarded to its worker as --<name> <value>; set once per process, as BASERT_LIB is.
runtime_pick
Which runtime serves a model.
runtimes
Where each runtime’s worker is, and what it says about itself.
runtimes_report
superfluid runtimes [<model>] [--json].
scheduler
The multi-lane scheduler.
selfupdate
superfluid self update|list|use|check|install: releases side by side under $SUPERFLUID_HOME/versions, a current link the commands on PATH go through, and the daily check serve runs.
serve
superfluid serve: the server, as a library function so that another program can run it.
speculate_resolve
--speculate without knowing drafter formats.
store
store_id
A session store’s persisted identity; see StoreId.
telemetry
template_codec
TemplateCodec.
tool_fragment
Streaming a tool call while the model is still writing it.
tui
superfluid serve terminal UI — an htop/nvtop-style live monitor.
tui_top
superfluid top — a multi-node fleet monitor.
versions
superfluid runtime versions <id> [--json], and the INSTALLABLE column of superfluid runtimes.
wal
workerd
superfluid-workerd, the worker process of a runtime linked into this build, as a library function so that another program can be the worker.

Structs§

CancelSet
Sessions whose generation should stop. A cancel also asks the running tick to end at its next step boundary, so the lane stops within a step; a queued chat asks the same.
Completion
Daemon
DaemonOptions
FimRequest
GenerateOutcome
Inspection
ModelSamplingDefaults
Refusal
SessionSummary
SpeculationOptions
SyncClassScope

Enums§

DaemonError
StageOutput
Tier

Functions§

with_sync_class