43 Task Loop Harness

43 Task Loop Harness

Source: docs/design/43-task-loop-harness.md

Status: P0–P4 implemented (2026-10-21, D93)

Built-in pi tasks get coding-agent harness pieces:

1. Auto context compression — deterministic trim of old tool output → light-model summary → deterministic fallback; never split tool_call/result pairs
2. update_plan — structured plan injected as <current_plan> each turn; progress in run_steps
3. Soft budget + loop detection + stall watchdog — repeated identical tool outcomes warn then fail LOOP_DETECTED; stall → stalled; configurable maxWallMinutes (default 240; 0 = unlimited)

Settings → General → Task Loop (general.task-loop). External agent tasks keep their native harness. Failure feedback: length/aborted/deferred settle as failure and wake; user cancel from card also wakes.