94681289e6
OpenPay (1 crate, ~2287 LOC, 15 tests)
op-dispute-evidence Chargeback lifecycle + evidence packaging across Visa
VCR (10.x/11.x/12.x/13.x), Mastercard Mastercom
(4837/4853/1240/1442), Amex SafeKey + DRR, Discover DRR,
PayPal. CE3.0 compelling-evidence qualifier (120-365d
window, ≥2 qualifiers, ≥2 linking matches, 10.4-only).
Phase state machine + ReasonCodeCatalog + sealed
EvidencePackageBuilder + win-rate scoring.
Smart Byte (1 crate, ~2479 LOC, 44 tests)
smart-byte-edv Encrypted Data Vaults (DIF v0.10).
JWE ECDH-ES P-256 + HKDF-SHA256-Concat-KDF + AES-256-GCM,
multi-recipient; HMAC-SHA-256 encrypted index with
AND-equality queries; full ZCAP-LD chain validation;
chunked AES-256-GCM streaming with per-chunk AAD;
Vault trait + InMemoryVault + one-way pull replication.
EOC (1 crate, ~2065 LOC, 29 tests)
eoc-agent Agentic loops with per-step joule accounting.
ReAct, Reflexion (PASS/FAIL judge + episodic memory),
Tree of Thoughts (BFS/DFS + self-scoring), LATS
(MCTS-UCT select/expand/simulate/backprop),
Plan-and-Execute, CodeAct (fenced-block + pluggable
runtime). Budget enforcement (tokens/joules/tool-calls),
trace spans, JouleCounter integration.
Three other batch-5 deferrals (op-ewallets-asia, smart-byte-zk, eoc-rag)
will land in wave A3.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
62 lines
2 KiB
Rust
62 lines
2 KiB
Rust
//! Integration test: Reflexion turns a failed first attempt into a
|
|
//! retry that uses the judge's critique.
|
|
|
|
use std::sync::Arc;
|
|
use std::sync::Mutex;
|
|
|
|
use async_trait::async_trait;
|
|
use eoc_agent::agent::{Agent, Budget, LlmProvider, LlmReply};
|
|
use eoc_agent::error::AgentResult;
|
|
use eoc_agent::reflexion::ReflexionLoop;
|
|
use eoc_agent::StopReason;
|
|
|
|
struct Scripted {
|
|
replies: Mutex<Vec<&'static str>>,
|
|
}
|
|
|
|
#[async_trait]
|
|
impl LlmProvider for Scripted {
|
|
async fn complete(&self, _prompt: &str) -> AgentResult<LlmReply> {
|
|
let text = {
|
|
let mut q = self.replies.lock().expect("lock");
|
|
if q.is_empty() {
|
|
"(empty)".to_string()
|
|
} else {
|
|
q.remove(0).to_string()
|
|
}
|
|
};
|
|
Ok(LlmReply::new(text, 8, 500))
|
|
}
|
|
}
|
|
|
|
#[tokio::test]
|
|
async fn reflexion_retries_after_failure() {
|
|
// Actor: first attempt is wrong, second attempt is right.
|
|
let actor = Arc::new(Scripted {
|
|
replies: Mutex::new(vec!["answer: 41", "answer: 42"]),
|
|
});
|
|
// Judge: FAILs the first, PASSes the second.
|
|
let judge = Arc::new(Scripted {
|
|
replies: Mutex::new(vec!["FAIL\noff by one", "PASS\nlooks right"]),
|
|
});
|
|
let mut agent = ReflexionLoop::new(actor, judge, Budget::unlimited(10), 5);
|
|
let (out, stop) = agent.run("what is 6 * 7?").await.expect("ok");
|
|
assert_eq!(out, "answer: 42");
|
|
assert_eq!(stop, StopReason::Done);
|
|
assert_eq!(agent.memory.lessons.len(), 1);
|
|
assert!(agent.memory.lessons[0].contains("off by one"));
|
|
}
|
|
|
|
#[tokio::test]
|
|
async fn reflexion_gives_up_after_max_attempts() {
|
|
let actor = Arc::new(Scripted {
|
|
replies: Mutex::new(vec!["bad", "bad", "bad"]),
|
|
});
|
|
let judge = Arc::new(Scripted {
|
|
replies: Mutex::new(vec!["FAIL a", "FAIL b", "FAIL c"]),
|
|
});
|
|
let mut agent = ReflexionLoop::new(actor, judge, Budget::unlimited(10), 2);
|
|
let (_out, stop) = agent.run("goal").await.expect("ok");
|
|
assert_eq!(stop, StopReason::MaxIterations);
|
|
assert_eq!(agent.memory.lessons.len(), 2);
|
|
}
|