anuna-research/hence
1
0
Fork 0
Code Issues 1 Pull requests 3 Projects Packages Wiki Activity Actions

feat: implement experiment, refinement, and race loop patterns with watch loop integration #16

Open
hugooconnor wants to merge 41 commits from hence/plan-IMPL-015 into main
pull from: hence/plan-IMPL-015
merge into: anuna-research:main
Conversation 0 Commits 41 Files changed 26 +3449 -85
hugooconnor commented 2026-03-20 09:57:02 +00:00 (Migrated from codeberg.org)
Copy link

Summary

  • Repeatable tasks — new (repeatable "true") metadata field enables automatic respawn after completion/failure, with consecutive failure counter reset so transient failures don't exhaust the retry budget
  • Auto-merge control — (auto-merge "false") metadata skips branch merge after completion, preserving branches for downstream inspection (race judge, experiment comparison)
  • Plan templates — hence plan init --template {refinement,experiment-loop,race} scaffolds full plans with companion shell scripts (check-quality.sh, check-metric.sh, record-result.sh, judge.sh, merge-winner.sh)
  • Context injection — context constructor now assembles refinement context (iteration number, quality feedback, improvement guidance) and experiment context (metric name, direction, best value, results history) into agent prompts
  • Env var injection — HENCE_* environment variables (metric, direction, time budget, modifiable file, plan) are injected into evaluate-check commands, lifecycle scripts, and agent PTY sessions
  • Watch loop hardening — permanently-failed guard prevents infinite respawn loops, infrastructure failure resets cached state, and permanently-failed repeatable tasks are treated as terminal
  • Race judge integration — end-to-end tests for 3-approach race flow including judge verdict validation and merge-winner branch merging
  • Refinement defeater fix — corrected undercutting defeater syntax so quality-passed properly blocks ready-refine

Test plan

  • 6 watch loop integration tests (permanently-failed guard, state reset, terminal detection)
  • 6 refinement defeater CLI integration tests + 3 watch unit tests
  • 4 integration branch invariant tests (keep/discard/cumulative)
  • Race judge end-to-end tests (readiness, completion, verdict, merge)
  • Unit tests for repeatable metadata, auto-merge, failure counter reset, context construction
## Summary - **Repeatable tasks** — new `(repeatable "true")` metadata field enables automatic respawn after completion/failure, with consecutive failure counter reset so transient failures don't exhaust the retry budget - **Auto-merge control** — `(auto-merge "false")` metadata skips branch merge after completion, preserving branches for downstream inspection (race judge, experiment comparison) - **Plan templates** — `hence plan init --template {refinement,experiment-loop,race}` scaffolds full plans with companion shell scripts (check-quality.sh, check-metric.sh, record-result.sh, judge.sh, merge-winner.sh) - **Context injection** — context constructor now assembles refinement context (iteration number, quality feedback, improvement guidance) and experiment context (metric name, direction, best value, results history) into agent prompts - **Env var injection** — `HENCE_*` environment variables (metric, direction, time budget, modifiable file, plan) are injected into evaluate-check commands, lifecycle scripts, and agent PTY sessions - **Watch loop hardening** — permanently-failed guard prevents infinite respawn loops, infrastructure failure resets cached state, and permanently-failed repeatable tasks are treated as terminal - **Race judge integration** — end-to-end tests for 3-approach race flow including judge verdict validation and merge-winner branch merging - **Refinement defeater fix** — corrected undercutting defeater syntax so `quality-passed` properly blocks `ready-refine` ## Test plan - [x] 6 watch loop integration tests (permanently-failed guard, state reset, terminal detection) - [x] 6 refinement defeater CLI integration tests + 3 watch unit tests - [x] 4 integration branch invariant tests (keep/discard/cumulative) - [x] Race judge end-to-end tests (readiness, completion, verdict, merge) - [x] Unit tests for repeatable metadata, auto-merge, failure counter reset, context construction
This pull request has changes conflicting with the target branch.
  • plans/IMPL-015-experiment-loop.spl
View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin hence/plan-IMPL-015:hence/plan-IMPL-015
git switch hence/plan-IMPL-015

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch main
git merge --no-ff hence/plan-IMPL-015
git switch hence/plan-IMPL-015
git rebase main
git switch main
git merge --ff-only hence/plan-IMPL-015
git switch hence/plan-IMPL-015
git rebase main
git switch main
git merge --no-ff hence/plan-IMPL-015
git switch main
git merge --squash hence/plan-IMPL-015
git switch main
git merge --ff-only hence/plan-IMPL-015
git switch main
git merge hence/plan-IMPL-015
git push origin main
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
anuna-research/hence!16
No description provided.