| name | squash |
| description | Analyze branch commits and consolidate iterative modifications. |
| disable-model-invocation | true |
Squash — Consolidate Branch History
Analyzes commits on the current branch and consolidates them into clean,
logical units. The original branch is untouched until the user explicitly
confirms the result. Heavy git operations are delegated to Python scripts
(scripts/ directory) that enforce safety checks.
Instructions
- NEVER operate on the default branch (main/master).
- All user-facing text must match the language the user has been using.
- When presenting choices, use interactive dialogs with selectable options
rather than plain text.
- Scripts are at
.codebuddy/skills/squash/scripts/ relative to repo root.
Use python3 to run them. NEVER read or modify the script contents.
Cleanup procedure
Whenever stopping due to failure or user discard, extract these values from
the latest successful step output (if available), then run:
git worktree remove --force {worktree_path}
git branch -D {tmp_branch}
rm -rf {session_dir}
| Field | Source |
|---|
worktree_path | Step 4 output (on Step 4 failure or Step 5 discard) |
tmp_branch | Step 4 output (on Step 4 failure or Step 5 discard) |
session_dir | Step 2 output (on Step 2/3 failure) |
If these values cannot be determined (e.g., Step 2 failed before creating
session_dir), skip the cleanup that depends on them. Errors during cleanup
can be safely ignored—resources will eventually be removed by resolve_scope.py
on the next run.
Step 1: Cleanup
If $ARGUMENTS is empty, run:
python3 scripts/resolve_scope.py {repo_dir}
The script validates that the current branch is not the default branch, and
cleans up stale sessions from previous runs. If it fails, inform the user
and stop.
If $ARGUMENTS is non-empty (a branch name), skip this step — validation
is handled by the collect script in Step 2.
Step 2: Collect Commits
Two modes depending on $ARGUMENTS:
If the script fails, inform the user and stop.
Record {session_dir} from the output — all temporary files and the worktree
will live inside this directory throughout the session.
If commits is empty, inform the user and stop.
Present the commit_table field to the user.
Step 3: Analyze & Decide
Use commits (with subject, files, is_merge, quality_flag) and
overlaps (adjacent pair file overlap ratios) from Step 2.
Grouping
Group adjacent commits belonging to the same logical change when either:
- File overlap ≥ 30 % (from
overlaps).
- Message similarity — same module/feature keyword, or follow-up pattern.
A "revert" commit should be grouped with the commit it reverts. If both are
adjacent, squashing them cancels them out.
Never group across a merge commit. Never include a merge commit in a
squash group. Never reorder commits.
Messages
Synthesize a new commit message for each squash group. Rewrite any commit with
a non-null quality_flag.
If nothing to merge or rewrite, inform the user the history is clean and
stop.
Output — Decision Table
Write to {session_dir}/decision.json:
{
"session_dir": "...",
"repo_dir": "...",
"branch": "...",
"squash_base": "<full_hash>",
"squash_end": "<full_hash>",
"upstream_remote": null,
"upstream_merge": null,
"actions": [
{"type": "keep", "commits": ["<hash>"]},
{"type": "squash", "commits": ["<hash1>", "<hash2>"], "message": "..."},
{"type": "reword", "commits": ["<hash>"], "message": "..."},
{"type": "merge", "commits": ["<hash>"]}
]
}
Copy session_dir, repo_dir, branch, squash_base, squash_end,
upstream_remote, upstream_merge directly from Step 2 output.
Rules: chronological order (oldest first), every commit in exactly one action,
full 40-char hashes, squash/reword require message, merge must have
exactly one commit, squash must never contain a merge.
Step 4: Build New History
python3 scripts/build_history.py {session_dir}/decision.json
On failure → run cleanup procedure and stop (error output includes
worktree_path and tmp_branch).
On success, parse the JSON output and record {worktree_path}, {tmp_branch},
and {current_head} from it. The script automatically handles any commits
added after squash_end — no manual git operations needed.
Step 5: User Confirmation
Mandatory interaction point. Do NOT skip.
python3 scripts/format_report.py {session_dir}/decision.json
Present the report (showing planned changes), then ask: Confirm / Discard.
Wait for response.
Note: The reported "new commit count" in the format report is an estimate
based on the decision table. The actual count after execution (which accounts
for zero-diff groups being skipped) will be shown in Step 6 results.
Discard → run cleanup procedure and stop.
Step 6: Replace Branch
Write {session_dir}/replace.json using values from prior steps, then run:
python3 scripts/replace_branch.py {session_dir}/replace.json
| Field | Source |
|---|
repo_dir | Step 2 output |
branch | Step 2 output (branch field — may be a non-current branch) |
tmp_branch | Step 4 output |
worktree_path | Step 4 output |
current_head | Step 4 output |
upstream_remote | Step 2 output |
upstream_merge | Step 2 output |
The script automatically appends any new commits added since {current_head},
verifies tree equality, and atomically updates the branch ref. On success it
removes the worktree and tmp branch.
On failure → run cleanup procedure and stop.
On success:
rm -rf {session_dir}
Report using new_commit_count from Step 4 output (which accounts for commits
skipped due to zero-diff groups):
Squash complete: {original_commit_count} commits → {new_commit_count} commits
Step 7: Offer to Push
If the branch has upstream tracking and the remote tracking point is an
ancestor of the new HEAD, ask: Push / Skip.
Otherwise, end silently.