operon-openhands-gates 0.1.0a3

Structural reliability critics for the OpenHands Agent SDK — certified stagnation detection, backed by the Operon categorical framework.

operon-openhands-gates

In-loop structural reliability critics for the OpenHands Agent SDK — drop-in, cert-emitting.

OpenHands' own docs flag an architectural gap in iterative refinement:

"the current implementation relies solely on threshold/iteration limits rather than monitoring improvement velocity or convergence rates — suggesting this is an architectural gap where monitoring logic could plug in." — https://docs.openhands.dev/sdk/guides/iterative-refinement

This package ships the missing monitor as a CriticBase subclass. It replaces an LLM-judged success score with a Bayesian stagnation signal computed over the conversation's message history. When the agent goes in circles, the critic's score drops below threshold, iterative refinement terminates, and a replayable behavioral_stability certificate is emitted.

At a glance:

• OperonStagnationCritic — epiplexic_integral-based detection (Paper 4 §4.3, 0.960 convergence accuracy with real embeddings) that plugs directly into Agent(critic=...).
• One certificate per detection transition, self-verifiable via certificate.verify().
• Zero-dep NGramEmbedder default — bring your own neural embedder for paraphrase-robust detection.

Install

pip install operon-openhands-gates

Requires operon-ai>=0.34.4 and openhands-sdk>=1.15.

Quickstart

from openhands.sdk import Agent, Conversation, LLM
from openhands.sdk.critic.base import IterativeRefinementConfig
from operon_openhands_gates import OperonStagnationCritic

critic = OperonStagnationCritic(
    threshold=0.2,
    critical_duration=3,
    iterative_refinement=IterativeRefinementConfig(
        success_threshold=0.2,  # match the critic's threshold
        max_iterations=5,
    ),
)

agent = Agent(llm=LLM(model="anthropic/claude-sonnet-4-5"), tools=[...], critic=critic)
conversation = Conversation(agent=agent, workspace=workspace)
conversation.send_message("Fix the failing test in ...")
conversation.run()  # iterative refinement terminates on sustained stagnation

if critic.certificate is not None:
    # Replayable evidence of what the gate saw.
    verification = critic.certificate.verify()
    assert verification.holds

Why the non-default success_threshold

OpenHands' default success_threshold=0.6 is tuned for LLM probability-of-success scores. OperonStagnationCritic returns the epiplexic_integral directly — in [0, 1] where low = stagnant. Paper 4 §4.3 uses δ=0.2 as the stagnation threshold, so match it on the refinement config.

Sibling package

• operon-langgraph-gates — same Paper 4 substrate, same behavioral_stability_windowed certificate, targeted at LangGraph's StateGraph with .wrap() / .edge() node APIs. Two packages, one core — this is the framework-portability claim from Paper 5 §3 in code.

Certificate theorem name and verification

Certificates emitted by this package carry the theorem name behavioral_stability_windowed (not the core's shared behavioral_stability). The two differ in how they verify:

• behavioral_stability (shared core): mean(severities) < threshold. Loses the per-window structure that rolling-integral detection operates on.
• behavioral_stability_windowed (shared core, since operon-ai 0.36.0): max(per_window_severity_means) <= stability_threshold. Mirrors detection exactly.

Both verifiers are registered in operon_ai.core.certificate._THEOREM_FN_PATHS, so deserialized certificates resolve through _resolve_verify_fn without this package needing to be imported. Any consumer with operon-ai>=0.36.0 can round-trip a behavioral_stability_windowed certificate correctly.

Breaking change from pre-alpha prototypes

Earlier pre-release builds emitted certificates with theorem name behavioral_stability (the shared core name), bound to a locally-attached _verify_fn. That shape was semantically wrong — the shared verifier is flat-mean-based, so any cert round-tripped through serialization would silently revert to the wrong replay logic. Consumers that key on certificate.theorem == "behavioral_stability" or metadata["certificate_theorem"] == "behavioral_stability" must update to "behavioral_stability_windowed". No migration path is provided; alpha.

Citations

Backed by Paper 4 §4.3: convergence/false-stagnation accuracy 0.960 with real sentence embeddings (all-MiniLM-L6-v2, N = 300 trials). Full numbers and reproduction commands in the Operon repo at eval/results/benchmarks_real_embeddings/multi_model_summary.json. Paper 5 §3 establishes the preservation-under-compilation framework that the certificate follows.

Status

Alpha. API may change before 0.1.0 stable. Feedback welcome via Issues.

License

MIT — see LICENSE.