apply

Goal

Apply actions correctly and safely — the canonical skill for every apply pattern — merged from the former choose-and-apply, validate-then-apply, and delete

Scope

choose: several alternatives (tools, libraries, approaches) → analyze → pick → apply
validate: confirm before applying — code, config, references, architecture, or a decision
delete: delete files/folders safely — dry-run, confirm, backup, update references
No mode given → detect from context/argument, then /ask-me if still unclear

For domain-specific deletes, call the skill directly: /delete-cicd-fails, /delete-git-branch, /delete-git-submodules, /delete-git-worktree, /delete-projects, /delete-project-temp — each skill has its own dry-run + confirm gate

Execute

0. Detect Mode

Goal: Pick the workflow that fits the context

argument/context specifies a mode → use that mode
Several options to choose from → choose
Deleting files/folders → delete
A general apply that needs confirmation first → validate
Unclear → /ask-me

Mode: Choose

Goal: Pick the best of N options and apply it

Capture options — take them from argument/context; none → /ask-me; a single option → /follow-your-suggestion
Analyze — assess scope, cost, effort, risk, compatibility, maintainability; close options/high trade-offs → /rethink
Present and confirm — /report table with columns No. | Option | Pros | Cons | Score | Recommended; mark the recommendation with ✅; high-risk → /ask-me first
Apply — continuing from context → /follow-your-suggestion; a single action → /continue; multiple steps → /manage; affects many places → /all-this-patterns
Validate and report — /run-check or /deep-validate as appropriate → /report table (No. | Option | Status | Output) → /suggest-next-action

Mode: Validate

Goal: Always check before applying — never apply until the criteria pass

Capture context — take the action/context; unclear → /ask-me; define scope, dependencies, expected outcome
Select validation — code → /run-check; skill/rules/config → /deep-validate; reference/path → /check-reference or /review-devin-global-harness; architecture/design → the fitting /review; decision → /rethink
Run validation — record passed/failed, findings, warnings; order by severity; /report table (No. | Check | Status | Findings)
Decide — all pass → apply; low warnings → notify the user and apply once confirmed; error/blocker → fix until it passes; unsure → /ask-me
Apply — suggestion-only → /follow-your-suggestion; file ops in many places → /all-this-patterns; multiple steps → /manage
Re-validate and report — validate again after applying → /report table (No. | Step | Validation | Apply | Status); stuck step → /report-progress → /suggest-next-action

Mode: Delete

Goal: Delete safely — dry-run + confirm + backup + update references

Identify targets — take the list from the user; expand pattern/glob into concrete files; record absolute paths
Assess impact — /report summarizes the state; /scan-codebase or /use-astgrep to find references; check git status for tracked/untracked; identify broken refs; many/important references → /ask-me first
Dry run — show the deletion list + refs that would break + total size → ask proceed/skip/cancel
Backup — under git → recoverable via git, no backup needed; not in git/uncommitted → copy to %TEMP% or .trash/ and record the path
Delete — Remove-Item/rm; directory → recursive once confirmed; important targets (config, keys) → double confirm; deletion fails → report the error and stop
Update references — /update-references fixes/removes refs pointing at the targets; /review-devin-global-harness checks for broken refs (fix up to 3 rounds)
Validate — confirm the deletion actually happened; /run-check; /review-devin-global-harness again; fails → /resolve-errors then retry
Report — deleted list + fixed refs + remaining broken refs → /suggest-next-action

Rules

1. Confirmation Gates

Never apply without confirmation when high-risk
Delete: always dry-run + confirm; never delete outside scope, git history, or .git
Never apply while validation blockers remain — don't skip them because the user is in a hurry

2. Evidence And Context

Every option must be evaluated against the real context — no guessing
Keep the original context throughout the apply
Keep validation results in the report before and after applying

3. Safety First (Delete)

Never delete files containing secrets, API keys, or credentials
Target in global config (~/.config, %APPDATA%) → double confirm
Target in use by another process → stop the process first
Target not in git → back it up before deleting and record the backup path in the report

4. No Broken References

Every broken reference must be updated or deleted
Never leave imports, paths, or links broken

Expected Outcome

Mode detected/selected correctly from context
Choose: full comparison table + recommended option with reasons + apply succeeds
Validate: check before every apply + no apply with blockers + complete report
Delete: targets removed + no broken refs + passes /run-check + rollback possible if backed up
Next action known for every mode