Repository navigation
ha 0.9.12
·
67 commits
to master
since this release
ha 0.9.12
From four papers (docs/architecture/COST_AND_SKILLS.en.md, with the gap tables):
- Re-reads not sent twice (arXiv 2611.12122). Within a turn, a
read_fileof a file the model still holds unchanged answers with a pointer to that result (fresh: truereads it again); a re-read after a small change sends only the changed lines; runs of 4+ identical lines of command output are shown once with their count.HA_REREAD=off;ha exec --jsonreportsreread. - Cost by part (arXiv 2611.12122).
ha exec --jsongivescost_usdsplit into cache read, cache write, uncached input and output for the run, its children and the memory passes, each call at its own model's price;/costand the bug report show the split. - Learned skills revised like weights (SkillOpt, arXiv 2605.23904). A revision is an
editof at most 4 structured edits; a whole rewrite of a proven skill is refused. The revision is on trial with the incumbent kept: two judged uses better accept it, two worse roll it back by itself (files, absorbed siblings and counters); a fix that replaces a failing revision keeps the sound incumbent. Refused, rolled-back and superseded revisions are shown to the refiner.ha learning reportcounts them. - What a skill costs, and its prune (EfficientSkill, arXiv 2611.19823). Steps, tool calls and output tokens per skill in play, against a baseline;
ha learning prune [--skill] [--model] [--thinking](and the curator's cadence,HA_SKILL_PRUNE=off) shortens proven skills that are long or costly, on trial and rolled back at more than 1.10x the incumbent's cost. - Each step judged in hindsight (arXiv 2611.17167). After a turn of 6+ tool calls, the helper model scores every action with the whole turn in view: did it help achieve the request? Kept in
step-judgments.jsonl, given to the next review as data, shown per learned skill and inha learning report.HA_STEP_JUDGE=off.
Fixes:
ha exec: a learned skill the model activated is in play for the review (the review job replaced the run's list with an empty one), and a skill on trial in play is always judged.- Goal: only the person answers the agent's question, and the host's own children do not wake the conversation.
- A message the user sends with
ha sendanswers the agent's open question (it was queued behind a run that was not running, andha send --waithung).
Live (openai-codex/gpt-6-luna, thinking max): reread.sh 16/16, skill-trial.sh 13/13, step-judge.sh 10/10, strategy.sh 31/31, queue.sh 19/19, children.sh 14/14, goal.sh 6/6. Control on 0.9.11: reread 8 of 16 fail, skill-trial 9 of 12 fail, step-judge 6 of 10 fail.