mirror of
https://github.com/block/buzz.git
synced 2026-10-27 12:40:30 +00:00
d2cfd377e2
## Why The base prompt prescribed a todo lifecycle for every task needing follow-up tools, which added control calls even for work completed in one turn. This keeps the important behavioral contract—continue after publishing pickup—without prescribing the mechanism. ## What - Replace the 40-word todo lifecycle with a concise pickup follow-through rule - Preserve the requirement to publish the outcome or blocker before stopping ## Risk Assessment Low to medium — this changes managed-agent instructions, not runtime code. The terminal benchmark covers single-turn task completion but does not cover restart, compaction, or long-lived multi-turn recovery. ## References - Builds on the prompt simplification in https://github.com/block/buzz/pull/6161 - Benchmark setup: GPT 5.6 Terra at high effort; the same 11-task Terminal-Bench 2.1 slate; four concurrent trials; 4 CPU and 8 GiB per trial; 3× timeout | Prompt | Pass | Active-h | Median active | Tool calls | |---|---:|---:|---:|---:| | PR 6161 baseline | 22/22 | 0.3438 | 0.91 min | 216 | | Benchmarked 14-word rule | 22/22 | 0.2926 | 0.75 min | 198 | The benchmarked rule used 14.9% less active time, 16.9% lower median active time, and 8.3% fewer tool calls. Across the screen and confirmation runs it passed 33/33 trials with every completion report present; results are directional because model and tool behavior is stochastic. --- **Update Aug 18, 10:47 EDT:** Expanded the completion outcomes following review feedback. - The follow-through rule now covers a verified result, blocker, or key decision or information that needs to be surfaced. - The benchmark was not rerun; the table reflects the prior 14-word formulation. This is a completion-taxonomy clarification, not a return to a prescribed todo mechanism. Generated with Codex Signed-off-by: Salman Mohammed <smohammed@squareup.com>