flock-spawn-test / src /flock /models.py
cmpatino's picture
cmpatino HF Staff
deploy controller (2e85dec)
3ed7ad0 verified
Raw History Blame Contribute Delete
17.1 kB
"""Shared contracts between controller, worker and CLI.
This file is the contract. Everything in docs/design.md section 5 ("Contracts")
is expressed here. Change it deliberately and update the design doc.
"""
from __future__ import annotations
import re
from datetime import datetime, timedelta
from enum import StrEnum
from typing import Any, Literal
from pydantic import BaseModel, Field, field_validator
# --------------------------------------------------------------------------- durations
_DURATION_RE = re.compile(r"^\s*(\d+)\s*(ms|s|m|h|d)\s*$")
_UNIT_SECONDS = {"ms": 0.001, "s": 1, "m": 60, "h": 3600, "d": 86400}
def parse_duration(value: str | int | float | timedelta) -> timedelta:
"""Parse '30s', '10m', '24h', '7d' (or a number of seconds) into a timedelta."""
if isinstance(value, timedelta):
return value
if isinstance(value, (int, float)):
return timedelta(seconds=float(value))
m = _DURATION_RE.match(value)
if not m:
raise ValueError(f"bad duration {value!r}; use e.g. 30s, 10m, 24h, 7d")
return timedelta(seconds=int(m.group(1)) * _UNIT_SECONDS[m.group(2)])
# --------------------------------------------------------------------------- enums
class Harness(StrEnum):
claude = "claude"
codex = "codex"
class AgentState(StrEnum):
working = "working" # a harness process is running a turn
idle = "idle" # no process; idle_since set
stuck = "stuck" # a turn exceeded its timeout and was killed; cleared on next turn
unknown = "unknown" # worker heartbeat older than 90 s
dead = "dead" # heartbeat older than 180 s, or job terminal
class WorkerState(StrEnum):
launching = "launching"
running = "running"
draining = "draining"
unknown = "unknown"
dead = "dead"
class TypeHealth(StrEnum):
ok = "ok"
crashlooping = "crashlooping"
# --------------------------------------------------------------------------- spec
AGENT_NAME_RE = re.compile(r"^[a-z][a-z0-9-]*-\d{4,}$")
TYPE_NAME_RE = re.compile(r"^[a-z][a-z0-9-]*$")
class TypeSpec(BaseModel):
"""One agent type inside a swarm. See design.md 'Spec'."""
harness: Harness
model: str
image: str | None = None # None -> default worker image
system_prompt_file: str | None = None # written into the workdir as CLAUDE.md / AGENTS.md
# The file's contents, inlined by `flock apply` (the controller and workers cannot
# read the caller's files). Wins over system_prompt_file on the worker.
system_prompt: str | None = None
env: dict[str, str] = Field(default_factory=dict)
secrets: list[str] = Field(default_factory=list) # names of secrets the job receives
labels: dict[str, str] = Field(default_factory=dict)
replicas: int = 0
agents_per_worker: int = 50 # measured default, experiments/packing
max_working_per_worker: int = 20 # measured default, experiments/packing
# Worker slots for this type, declared only (docs/spawn.md). Required for a type that
# `spawn:` lets others create; None keeps v1's placement.
workers: WorkersSpec | None = None
flavor: str = "cpu-basic"
worker_lifetime: str = "24h"
lifetime_jitter: str = "4h"
drain_window: str = "10m"
idle_grace: str = "120s"
turn_timeout: str = "30m"
max_turns: int | None = None # per prompt, passed to the harness
max_budget_usd: float | None = None # per prompt, passed to the harness
persist_workdir: bool = True
resume_prompt: str | None = None
msgboard_invite: str | None = None # name of the secret holding the invite code
@field_validator(
"worker_lifetime", "lifetime_jitter", "drain_window", "idle_grace", "turn_timeout"
)
@classmethod
def _valid_duration(cls, v: str) -> str:
parse_duration(v)
return v
@field_validator("replicas", "agents_per_worker", "max_working_per_worker")
@classmethod
def _non_negative(cls, v: int) -> int:
if v < 0:
raise ValueError("must be >= 0")
return v
@field_validator("workers")
@classmethod
def _replicas_fit(cls, v: WorkersSpec | None, info) -> WorkersSpec | None:
"""Declared workers are a cap (docs/spawn.md): the replicas must fit in them."""
replicas, per = info.data.get("replicas"), info.data.get("agents_per_worker")
if v is None or v.initial is None or replicas is None or per is None:
return v
room = (v.initial + v.max_extra) * max(1, per)
if replicas > room:
raise ValueError(
f"replicas {replicas} do not fit in workers.initial + workers.max_extra = "
f"{v.initial + v.max_extra} workers x agents_per_worker {per} = {room} agents"
)
return v
class SwarmSpec(BaseModel):
swarm: str
namespace: str # HF namespace jobs are launched under
bucket: str # e.g. hf://buckets/cmpatino/flock
types: dict[str, TypeSpec]
@field_validator("swarm")
@classmethod
def _valid_swarm(cls, v: str) -> str:
if not TYPE_NAME_RE.match(v):
raise ValueError("swarm name must be lowercase [a-z0-9-], starting with a letter")
return v
# Run policy for spawned agents (docs/spawn.md), separate from what a type is. A type
# absent from `spawn` cannot spawn.
spawn: dict[str, SpawnRule] = Field(default_factory=dict)
limits: LimitsSpec = Field(default_factory=lambda: LimitsSpec())
@field_validator("spawn")
@classmethod
def _valid_spawn(cls, v: dict[str, SpawnRule], info) -> dict[str, SpawnRule]:
types = info.data.get("types")
if types is not None:
validate_spawn(v, types)
return v
@field_validator("types")
@classmethod
def _valid_types(cls, v: dict[str, TypeSpec]) -> dict[str, TypeSpec]:
for name in v:
if not TYPE_NAME_RE.match(name):
raise ValueError(f"type name {name!r} must be lowercase [a-z0-9-]")
return v
def agent_name(type_name: str, slot: int) -> str:
return f"{type_name}-{slot:04d}"
# --------------------------------------------------------------------------- spawn
class WorkersSpec(BaseModel):
"""Worker slots a type may use: `initial` (default: what v1 computes from replicas and
agents_per_worker) plus up to `max_extra` opened on demand for spawned agents."""
initial: int | None = None
max_extra: int = 0
@field_validator("initial", "max_extra")
@classmethod
def _non_negative(cls, v: int | None) -> int | None:
if v is not None and v < 0:
raise ValueError("must be >= 0")
return v
class SpawnRule(BaseModel):
"""What an agent of one type may spawn. `types` is a list (with one total `max`) or a map
of per-type caps (`max` then an optional total cap). Caps count the parent's existing
children. `depth` is the generations allowed below the agent (0: cannot spawn). Unknown
keys are rejected, so a misspelled or unsupported policy field never passes silently."""
model_config = {"extra": "forbid"}
types: list[str] | dict[str, int]
max: int | None = None
depth: int = 1
@field_validator("max", "depth")
@classmethod
def _non_negative(cls, v: int | None) -> int | None:
if v is not None and v < 0:
raise ValueError("must be >= 0")
return v
@field_validator("types")
@classmethod
def _valid_caps(cls, v: list[str] | dict[str, int]) -> list[str] | dict[str, int]:
if isinstance(v, dict) and any(n < 0 for n in v.values()):
raise ValueError("per-type caps must be >= 0")
return v
@property
def allowed(self) -> list[str]:
return list(self.types)
def cap_for(self, type_name: str) -> int | None:
return self.types.get(type_name) if isinstance(self.types, dict) else None
class LimitsSpec(BaseModel):
"""Swarm-wide limits checked when an agent spawns (None: no limit). `alive` counts agents
not removed, replicas included; `spawned_total` the spawn counter; `budget_usd` the cost
of every agent in the swarm, removed ones included."""
alive: int | None = None
spawned_total: int | None = None
budget_usd: float | None = None
@field_validator("alive", "spawned_total", "budget_usd")
@classmethod
def _non_negative(cls, v: float | None) -> float | None:
if v is not None and v < 0:
raise ValueError("must be >= 0")
return v
def validate_spawn(spawn: dict[str, SpawnRule], types: dict[str, TypeSpec]) -> None:
"""Apply-time checks: known types, no cycle, `workers` on every spawnable type, a total
`max` for the list form."""
for parent, rule in spawn.items():
if parent not in types:
raise ValueError(f"spawn.{parent}: unknown type")
if isinstance(rule.types, list) and rule.max is None:
raise ValueError(f"spawn.{parent}: a list of types needs `max`")
for child in rule.allowed:
if child not in types:
raise ValueError(f"spawn.{parent}.types: unknown type {child!r}")
if types[child].workers is None:
raise ValueError(f"types.{child}: spawnable by {parent!r}, so it needs `workers`")
state: dict[str, int] = {} # 1 on the DFS stack, 2 done
def visit(t: str, path: list[str]) -> None:
state[t] = 1
for child in spawn[t].allowed if t in spawn else []:
if state.get(child) == 1:
cycle = " -> ".join([*path[path.index(child):], child])
raise ValueError(f"spawn: cycle {cycle}")
if child not in state:
visit(child, [*path, child])
state[t] = 2
for t in sorted(spawn):
if t not in state:
visit(t, [t])
def child_name(parent: str, type_name: str, n: int) -> str:
return f"{parent}-{type_name}-{n:04d}"
def initial_workers(ts: TypeSpec) -> int:
"""v1's worker count for a type's replicas, or the declared `workers.initial`."""
if ts.workers is not None and ts.workers.initial is not None:
return ts.workers.initial
return -(-ts.replicas // max(1, ts.agents_per_worker))
def max_workers(ts: TypeSpec) -> int:
return initial_workers(ts) + (ts.workers.max_extra if ts.workers else 0)
def root_depth(spec: SwarmSpec, type_name: str) -> int:
"""Spawn depth of a replica: its type's `spawn.depth`, 0 if the type cannot spawn."""
rule = spec.spawn.get(type_name)
return rule.depth if rule else 0
def child_depth(spec: SwarmSpec, parent_depth: int, type_name: str) -> int:
return max(0, min(parent_depth - 1, root_depth(spec, type_name)))
TypeSpec.model_rebuild()
SwarmSpec.model_rebuild()
# --------------------------------------------------------------------------- usage
class TokenUsage(BaseModel):
input: int = 0
output: int = 0
cache_read: int = 0
cache_write: int = 0
reasoning: int = 0
def __add__(self, other: TokenUsage) -> TokenUsage:
return TokenUsage(
input=self.input + other.input,
output=self.output + other.output,
cache_read=self.cache_read + other.cache_read,
cache_write=self.cache_write + other.cache_write,
reasoning=self.reasoning + other.reasoning,
)
# --------------------------------------------------------------------------- worker -> controller
class AgentHeartbeat(BaseModel):
"""One record per agent inside a worker heartbeat (~100 bytes)."""
name: str
state: Literal["working", "idle", "stuck"]
turn_id: str | None = None
idle_since: datetime | None = None
session_id: str | None = None
resumed: bool | None = None # None until the first (re)start on this worker
tokens: dict[str, TokenUsage] = Field(default_factory=dict) # by model, cumulative
harness_estimate_usd: float | None = None # Claude's client-side total, stored not trusted
turns: int = 0
last_sync: datetime | None = None
msgboard_principal: str | None = None
class WorkerHeartbeat(BaseModel):
worker: str
deadline: datetime
draining: bool = False
agents: list[AgentHeartbeat]
class Event(BaseModel):
"""One line of an agent's event stream.
The on-disk / over-the-wire shape is pi session format v3 (see swarmlab's
export.py for the reference encoding). `type`, `ts`, `agent` are required by us;
the rest of the pi record lives in `data`. Harness-neutral types the worker adds:
session_started, session_resumed, session_restarted, turn_started, turn_ended,
turn_interrupted, sync.
"""
type: str
ts: datetime
agent: str
turn_id: str | None = None
data: dict[str, Any] = Field(default_factory=dict)
class EventBatch(BaseModel):
worker: str
events: list[Event]
# --------------------------------------------------------------------------- controller -> worker
class PromptCommand(BaseModel):
kind: Literal["prompt"] = "prompt"
agent: str
prompt_id: str
text: str
timeout: str | None = None
reply_to: str | None = None # opaque, for the caller's bookkeeping
class StopCommand(BaseModel):
kind: Literal["stop"] = "stop"
agent: str # kill the agent's harness process if any; agent stays assigned
class DrainCommand(BaseModel):
kind: Literal["drain"] = "drain" # worker-wide: finish in-flight turns, sync, exit 0
class AssignCommand(BaseModel):
"""Assign an agent to this worker; the worker restores it from the bucket if present."""
kind: Literal["assign"] = "assign"
agent: str
type: str
spec: TypeSpec
restore: bool = True # restore workdir + session from bucket before first prompt
resume_prompt: str | None = None # sent once after restore when a turn was interrupted
parent: str | None = None # the spawning agent, None for a replica
# The agent's scoped controller token, injected into its harness env as FLOCK_AGENT_TOKEN.
agent_token: str | None = None
class UnassignCommand(BaseModel):
kind: Literal["unassign"] = "unassign"
agent: str # stop, final sync, forget
Command = PromptCommand | StopCommand | DrainCommand | AssignCommand | UnassignCommand
class CommandBatch(BaseModel):
commands: list[Command]
# --------------------------------------------------------------------------- API shapes
class PromptRequest(BaseModel):
text: str
timeout: str | None = None
reply_to: str | None = None
class PromptAck(BaseModel):
prompt_id: str
agent: str
status: Literal["queued", "pending"] # queued: worker is up; pending: held until a worker is
turn_id: str | None = None # set later via the event stream, may be None here
class SpawnRequest(BaseModel):
type: str
prompt: str | None = None # first prompt, held until the child is placed
key: str | None = None # idempotency key, unique per parent
class SpawnAck(BaseModel):
name: str
created: bool
parent: str
type: str
depth: int
class AgentStatus(BaseModel):
name: str
swarm: str
type: str
harness: Harness
model: str
labels: dict[str, str]
state: AgentState
idle_since: datetime | None = None
turn_id: str | None = None
worker: str | None = None
session_id: str | None = None
resumed: bool | None = None
pending: int = 0
tokens: dict[str, TokenUsage] = Field(default_factory=dict)
cost_usd: float = 0.0 # derived by the controller from its price table
harness_estimate_usd: float | None = None
turns: int = 0
created_at: datetime
last_heartbeat: datetime | None = None
last_sync: datetime | None = None
msgboard_principal: str | None = None
parent: str | None = None # spawning agent; None for a replica
children: list[str] = Field(default_factory=list)
depth: int = 0 # generations this agent may still spawn below itself
class WorkerStatus(BaseModel):
id: str
swarm: str
type: str
job_id: str | None
state: WorkerState
deadline: datetime | None
agents: list[str]
last_heartbeat: datetime | None
launched_at: datetime
crashes: int = 0
class TypeStatus(BaseModel):
name: str
health: TypeHealth
replicas: int
agents_by_state: dict[str, int]
workers_by_state: dict[str, int]
cost_usd: float = 0.0
class SwarmStatus(BaseModel):
swarm: str
namespace: str
bucket: str
types: dict[str, TypeStatus]
agents_total: int
workers_total: int
pending_prompts: int
cost_usd: float = 0.0
# --------------------------------------------------------------------------- constants
HEARTBEAT_INTERVAL = timedelta(seconds=30)
UNKNOWN_AFTER = timedelta(seconds=90)
DEAD_AFTER = timedelta(seconds=180)
CRASHLOOP_THRESHOLD = 5
LAUNCHES_PER_WINDOW = 400 # per 5-minute Hub rate-limit window, PRO assumption
RATE_WINDOW = timedelta(minutes=5)
JOB_LABELS = ("swarm", "type", "worker") # every worker job carries these labels