Skip to main content

Crate paigasus_helikon_core

Crate paigasus_helikon_core 

Source
Expand description

Trait surface and core types for the Paigasus Helikon AI SDK.

This crate is the dependency root of the workspace; the facade crate paigasus-helikon re-exports its surface unconditionally.

The seven object-safe traits (Model, Tool, Session, Guardrail, Hook, Agent, Runner) and their carrier types form the contract every other Paigasus Helikon crate depends on.

See the project documentation site for conceptual material; this crate’s rustdoc is the canonical reference for the trait signatures and carrier types.

Re-exports§

pub use graph::GraphAgent;
pub use graph::GraphAgentBuilder;
pub use graph::GraphBuildError;
pub use agent::*;
pub use agent_as_tool::*;
pub use agent_builder::*;
pub use context::*;
pub use guardrail::*;
pub use handoff::*;
pub use hook::*;
pub use item::*;
pub use loop_state::*;
pub use model::*;
pub use permission::*;
pub use runner::*;
pub use session::*;
pub use state::*;
pub use swarm::*;
pub use token_counter::*;
pub use tool::*;
pub use tool_exec::*;
pub use workflow::*;

Modules§

agent
The Agent trait and its carrier types.
agent_as_tool
AgentAsTool — expose any crate::Agent as a crate::Tool.
agent_builder
Typestate builder for crate::LlmAgent. See SMA-319 design for the full rationale.
command_match
Pure, dependency-free shell-command tokenizer used by the permission layer to match Bash sub-commands. Models Claude Code’s pragmatic Bash matcher — not a full POSIX shell grammar. See the crate’s permissions concept page for the enumerated coverage and known bypasses.
context
Run-scoped context types.
control
Interceptors: the run’s control-layer orchestration unit.
graph
GraphAgent — a declared DAG of agents; node execution gated by dependencies (SMA-333, ADR-11).
guardrail
The Guardrail trait and its carrier types.
handoff
Handoff carrier types.
hook
The Hook trait and its carrier types.
item
Canonical wire-format messages and content blocks.
loop_state
Explicit state machine for the agent loop.
model
The Model trait — the single canonical async interface to an LLM provider — and its carrier types.
permission
Permission layer: gate tool calls via deny rules → permission mode → canUseTool policy. See the Permissions, Guardrails & Hooks concept page.
redaction
Secret redaction for tool output. Scrubs secret-shaped strings (by key-name pattern and by known env value) before tool output re-enters the model context or the session trajectory.
runner
The Runner trait and its carrier types.
schema
JSON Schema strict-mode normalization.
session
The Session trait and its carrier types.
state
Run-scoped, in-memory coordination state for workflow agents (SMA-325).
swarm
SwarmAgent — a pool of LlmAgents with full-mesh handoff tools auto-injected; the first member to produce a final output instead of handing off wins (SMA-333, ADR-11).
token_counter
Token estimation for crate::CompactingSession threshold decisions.
tool
The Tool trait and its carrier types.
tool_exec
The public per-tool-call execution pipeline: resolve → authorize → invoke → redact → convert.
workflow
Deterministic workflow agents (SMA-325): SequentialAgent, ParallelAgent, LoopAgent.

Structs§

CompactingSession
A Session that wraps any inner session and triggers LLM-based compaction once the projected token count exceeds threshold.
CompactingSessionBuilder
Builder for CompactingSession.

Enums§

CompactingSessionError
Error constructing a CompactingSession.