Blog on hiway-kit

The Last Task Never Closes: Completion Marks That Evaporate at the Turn Boundary, and Three Layers of Defense

Tue, 07 Jul 2026 00:00:00 +0000

A user reported an odd pattern: “Everything’s finished, but the last task keeps getting left behind, never marked complete.”

Everything worked. The commits were in, the gates were green. Yet the task board still showed something like Wrap-up — doc sync + adversarial review + push sitting at in_progress forever. It looks like a cosmetic glitch, but from a durable-execution point of view it’s serious: if the task ledger lies, the next session either redoes finished work or skips unfinished work.

Done Is a Command's Output, Not a Claim: Designing a Durable Completion Gate for Long-Running Loop Agents

Fri, 03 Jul 2026 00:00:00 +0000

The previous two posts mapped the landscape of harness and loop engineering (research notes) and then applied it to git isolation for parallel work. This one is about the next question: when an agent runs a loop across many sessions over a long time, what decides that something is “done”?

Key takeaways

Hand a loop agent the authority to judge completion and it will mark unverified work as done. Completion has to be the output of a command, not a claim made by the model.

Green Lights Lie: A Six-Way Adversarial Audit of the Completion Gates I Built

Fri, 03 Jul 2026 00:00:00 +0000

The three earlier pieces — a map of harness and loop engineering, parallel git isolation in practice, and designing a durable completion gate — all converged on one direction: turn verification into a machine. This post is about the uncomfortable place that direction leads: who verifies the verification machine?

Key takeaways

A completion gate returning exit 0 (green) is not the same thing as the code being right.

When I let six independent adversarial reviewers loose on the whole project, the most serious defects weren’t in feature code — they were in the verification machinery itself: false-greens, failures reported as success. The gates were lying.

Harness Engineering in Practice: Hardening Git Isolation in a Parallel-Agent Toolkit with Two Adversarial Reviews

Thu, 02 Jul 2026 00:00:00 +0000

Harness engineering reads clearly on paper. The real test starts when you put those principles on top of a tool people actually run. This post records what happened when we applied the ideas from our harness and loop engineering landscape to our own project, hiway-kit. It is also the story of how two rounds of adversarial review showed that a flaw we “knew” we had fixed was still there.

Key takeaways