IronClaw pins subagent lineage drop invariants and updates asynchronous networking stack

IronClaw maintainers merged PR #8098 on September 20, 2026, codifying strict regression test contracts for subagent execution state. In IronClaw's Rust-based multi-agent framework, subagent turns track execution depth, activation provenance, and descendant execution caps within active run metadata. However, when converting active run state into persistent TurnRunState snapshots, these descendant lineage fields are intentionally omitted to avoid unbounded serialization bloat and recursive dependency cycles in state storage. PR #8098 establishes an inverse regression test alongside the existing terminal-rewrite suite, verifying that state snapshots deliberately drop all three lineage fields and ensuring future refactors do not accidentally restore unbounded nested metadata. Concurrently, PR #8078 updates the Tokio asynchronous ecosystem across the workspace, bumping tower-http to version 0.7.1 and updating tokio-tungstenite to enhance WebSocket stability and directory serving. In parallel, PR #7834 upgrades WebAssembly runtime dependencies, while issue #8093 catalogues daily failure taxonomies across model tool turn drifts.

The facts

- PR #8098 adds inverse regression testing to guarantee TurnRunState snapshots drop depth and descendant lineage fields. - Prevents state database serialization bloat and recursive reference cycles across deeply nested subagent execution trees. - PR #8078 upgrades tower-http to 0.7.1 and updates tokio-tungstenite across IronClaw's asynchronous networking layer. - PR #7834 bumps WebAssembly execution dependencies to improve sandboxed tool execution stability and memory safety. - Issue #8093 documents daily operational failure patterns, highlighting MCP tool discovery drifts and turn execution timeouts.

Why it matters

In deeply nested multi-agent systems, state serialization can easily suffer from exponential memory growth if child agents retain full historical parentage and descendant trees in every persistent snapshot. Explicitly pruning lineage metadata upon state persistence bounds storage footprint while maintaining clean boundary contracts. Backing this pruning behavior with explicit regression tests prevents well-intentioned refactors from inadvertently reintroducing recursive state tracking.

Current

Inspected on 2026-09-21. The IronClaw stable-channel baseline is ironclaw-v1.4.0 published 2026-08-28T11:19:24Z. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.

Evidence

Primary evidence: nearai/ironclaw PR #8098 (https://github.com/nearai/ironclaw/pull/8098). Supporting context: nearai/ironclaw PR #8078 (https://github.com/nearai/ironclaw/pull/8078); nearai/ironclaw PR #7834 (https://github.com/nearai/ironclaw/pull/7834); nearai/ironclaw issue #8093 (https://github.com/nearai/ironclaw/issues/8093). The source bodies and linked context were inspected. Test results quoted from contributors remain attributed reports, not independently reproduced experiments.

Operator take

Contract boundaries in distributed agent state must be defended by negative and inverse test suites. Knowing what state must be dropped during persistence is just as important as knowing what to retain. Teams building hierarchical multi-agent architectures should verify their state snapshot models do not store unbounded execution trees, keeping serialization footprints predictable and lightweight.

Caveat

PR #8098 represents test contract coverage and does not alter production runtime logic. IronClaw's core release line remains at v1.4.0, with daily automated failure audits continuing to document tool execution drift in complex multi-step reasoning tasks.