anomalyco/opencode merged Kit Langton's June 20 commit 4f1a9d7a, changing configured agent step limits so the final turn disables tools and asks for a text summary instead of failing as a runner error.
Facts
- OpenCode moved the max-step prompt into core runner code and added tests that the final limited turn sends
toolChoiceasnonewith no tools available. openai/codexcommit d1209bdd added first, previous, and current context-window IDs that survive compaction, resume, and rollback.openclaw/openclawcommit 03b022b8 routed Copilot prompt, tool, output, and agent-end events through standard harness lifecycle hooks.NousResearch/hermes-agentcommit 47fadc24 added an in-place compaction option that keeps one session ID across summarization.
Evidence
The receipts are runner tests, context-window state changes, lifecycle-hook wiring, and compaction code paths. Together they show stop, trace, and resume behavior becoming explicit runtime contracts.
Context
For agent builders and workspace operators, the practical question is no longer only whether the model can keep working. It is whether the system can say where the work is, why it stopped, and what state carries forward.
Limits
These are independent patches, not a shared standard. The evidence shows local engineering pressure around long-running agents; it does not prove that one common runtime model has arrived.