SD-160

Captain’s rubric on harness behavior — L6 is more complex than mapped. Empirical data from L12: (1) Without subagents, typing during generation often interrupts the model mid-response and forces recalibration. (2) With subagents running, user inputs queue in a visible FIFO stack in the UI. (3) Double-escape is a force override — hardware-level interrupt, redundancy layer. (4) Compaction is user-controllable at any point 0-200k, not only automatic at saturation. This affords fine-grained control over the weave and context pressure. (5) The harness makes one or more API requests, intercepts tool use, and can interrupt processes mid-flow. The Captain has not stress-tested these systematically — learning as he goes.

Standing observation — map data

---

← all decisions