agent-loop: new crate — LLM loop kernel + Model clients (phase 0)

Extract the LLM agent loop into a standalone workspace crate with zero
deps on skald-core/core-api (blueprint project-loop.md, D13-D15):

- kernel: round loop, model fallback with rebuild, parallel tool fan-out
  (ordered id alloc / bounded concurrent exec / ordered record), streaming
  deltas drained before outcomes, sticky cancellation
- models: OpenAiModel/AnthropicModel/OllamaModel/LmStudioModel ported from
  llm-client onto the Model trait; ModelError carries the HTTP status;
  is_retriable default = the 401/403/404/422 rule
- DTL as crate protocol (ToolRendering Inline/DeferredToolReference/
  SystemToolBlock; Anthropic conversions + Kimi system+tools passthrough),
  host catalog behind ActivationSource/ToolActivator
- HistoryStore durability contract + InMemoryStore; LinearAssembler with
  well-formed projection (incl. DTL injection, summary, crash survivors)
- LoopManager singleton (broadcast bus + live registry), one live loop
  per conversation, orphan-marking on start_turn
- 32 tests green (kernel §13 suite, assembler DTL, SSE/Anthropic ports),
  clippy clean
This commit is contained in:
2026-07-25 23:40:41 +01:00
parent 5081ec2afe
commit b8cc6d263b
26 changed files with 5684 additions and 0 deletions
+77
View File
@@ -0,0 +1,77 @@
//! `Gate` — the pre-execution decision point (policy and/or human). It MAY
//! block waiting for a human: the implementation decides (oneshot, UI, …).
//! Before suspending, an implementation marks the call `AwaitingHuman` via the
//! store (durability) and emits `LoopEvent::ApprovalRequired`.
use async_trait::async_trait;
use serde_json::Value;
use crate::events::EventSink;
use crate::ids::{FrameId, ToolCallId};
use crate::tool::Extensions;
/// A tool call awaiting a gate decision.
#[derive(Debug, Clone)]
pub struct PendingCall {
pub id: ToolCallId,
pub name: String,
pub args: Value,
pub frame: FrameId,
pub agent: String,
/// Host free-form (source, permission group, …).
pub extensions: Extensions,
}
/// The gate's verdict.
#[derive(Debug, Clone)]
pub enum GateDecision {
Allow,
Reject { reason: String },
}
#[async_trait]
pub trait Gate: Send + Sync {
/// Decide on a call. MAY block awaiting a human — in that case the
/// implementation marks the call `AwaitingHuman` first (via the store the
/// host gave it) and emits `ApprovalRequired` on `events`.
async fn check(&self, call: &PendingCall, events: &EventSink) -> GateDecision;
}
/// Everything runs. The default for simple hosts and tests.
pub struct AllowAll;
#[async_trait]
impl Gate for AllowAll {
async fn check(&self, _call: &PendingCall, _events: &EventSink) -> GateDecision {
GateDecision::Allow
}
}
/// Reject calls whose name matches a pattern: exact, or `prefix*`.
pub struct DenyList {
patterns: Vec<String>,
}
impl DenyList {
pub fn new(patterns: impl IntoIterator<Item = impl Into<String>>) -> Self {
Self { patterns: patterns.into_iter().map(Into::into).collect() }
}
fn matches(&self, name: &str) -> bool {
self.patterns.iter().any(|p| match p.strip_suffix('*') {
Some(prefix) => name.starts_with(prefix),
None => name == p,
})
}
}
#[async_trait]
impl Gate for DenyList {
async fn check(&self, call: &PendingCall, _events: &EventSink) -> GateDecision {
if self.matches(&call.name) {
GateDecision::Reject { reason: format!("tool '{}' denied by policy", call.name) }
} else {
GateDecision::Allow
}
}
}