scaryrawr Finalize autoresearch
Read .auto/log.jsonl (legacy: autoresearch.jsonl) and .auto/prompt.md.
Consider only experiments whose status is keep.
Analyze
- Expand every kept short SHA with
git rev-parse. - Discover the repository's trunk branch and compute
git merge-base. - Inspect incremental diffs:
base..<first kept>and<previous kept>..<next kept>. - Group commits into logical changesets in application order.
- Merge groups that touch the same file. Independent branches cannot safely own overlapping files.
- Flag cross-file dependencies. Merge tightly coupled groups; otherwise state the required landing order.
- Present proposed branch titles, commits, files, and metric changes. Obtain approval before creating branches.
Execute
Write .auto/finalize-groups.json:
{
"base": "<full merge-base hash>",
"trunk": "main",
"final_tree": "<full current HEAD hash>",
"goal": "short-slug",
"groups": [
{
"title": "Switch to forks pool",
"body": "Why and what changed.\n\nExperiments: #3, #5\nMetric: 42.3s -> 38.1s (-9.9%)",
"last_commit": "<full hash of the last kept commit in this group>",
"slug": "forks-pool"
}
]
}
Use full hashes and at least one group. Then run:
bash <this-skill-directory>/finalize.sh .auto/finalize-groups.json
The script validates commit order and file ownership, stashes a dirty worktree, creates every branch from the merge base, and verifies that cherry-picking all created branches reproduces the final non-session tree.
Remove .auto/finalize-groups.json after success.
Report
Report:
- each branch and its focused change;
- baseline-to-best metric improvement;
- dependencies and required landing order;
- the script's explicit cleanup commands;
- useful untried entries from
.auto/ideas.md.
Do not include discarded experiments or unrelated commits. Preserve any pre-existing dirty work by stashing before branch operations and restoring it afterward. If creation fails, return to the original branch and remove only branches created by this finalization attempt.