AI Forge Apply

AI Forge Apply (HITL)

Throw the lasso, pull it tight, move to the next. One change at a time, one approval at a time — no bundling, no skipping ahead.

Core question for every item: What's the smallest change that makes this finding false? Apply only that — nothing more. Verify: if any part of the change were removed, would the finding still be true? If yes, the change is too large.

Workflow

1. Read the list

Parse the numbered improvement list from context. If there are multiple lists or it's unclear which to apply, ask once before starting — not mid-loop. If the response does not resolve the ambiguity, stop: "Still unclear which list to apply — please re-invoke with the target list quoted directly."

If improvements reference specific files, verify they still exist and haven't changed substantially since the evaluation. If a target file is missing or heavily modified, note it at the top of the board: "⚠ Target may be stale — has changed since evaluation. Items may not apply cleanly. (c)ontinue / (Q)uit?"

If ALL items reference stale or missing targets, skip the board entirely: "⚠ All targets stale — re-run the evaluation against the current state rather than forcing through outdated findings." Then stop.

If items have dependencies (item 3 requires item 1), note the dependency at the board level. Do not re-order — apply in sequence and mark dependent items [○] if their prerequisite was skipped.

Check git status before showing the board. If the working directory is not a git repo, note this at the top of the board and skip all commit steps:

ℹ No git repo — changes will be applied but not committed.

Display the full status board upfront: all items with [ ] markers, then (A)pprove all / (s)tart options. Symbols: [ ] pending, [✓] done, [→] in progress, [–] skipped, [○] obsolete.

If the user responds A: apply all items in sequence, then show the final board per step 3.

If the user responds s: begin the per-item loop from item #1.

2. Apply each one

Repeat for every item in order:

a. Announce

Show ─── #N of M ─── and the full improvement description.

b. Apply

If the smallest change requires touching more than one logical unit (function, section, rule), the improvement is under-scoped — note this in the change summary and apply only the first unit.

Do not touch adjacent code, bundle items, or anticipate the next improvement. If applying this item overlaps with a later item (same lines, same function), apply only what's needed now and note the overlap in the change summary.

If an improvement requires non-edit actions (adding a dependency, creating a new file, running a command), describe the action in the change summary and let the user decide whether to perform it.

c. Show the change

Changed: <file>:<lines> — <one-line description>
<concise diff summary — not the full file>

d. Ask

(a)pprove, (r)evise, (s)kip, (o)bsolete, (A)ccept-all, (k)skip-all, (Q)uit?

Wait. Do not advance until the user responds. Any input not in the table below is a stop signal: show the current board and ask "Stop here? (y) to pause, (c)ontinue from #N?"

e. Update the board — after each decision, show the current state of all items using the status symbols. For accept-all (A): skip intermediate board updates — apply each remaining item silently, then show the final board once in step 3.

3. Done

When all items are processed, show the final board and a one-liner: HITL complete. Applied: N Skipped: M Revised then applied: P

Git Workflow

Default: no git operations. Commits are opt-in only — never commit unless the user explicitly requests it.

Before any changes (after showing the board)

Check git status. If uncommitted changes exist, warn:

⚠ Uncommitted changes detected. (s)tash / (c)ontinue / (Q)uit?

Then check if the user wants git integration:

Git: (n)o git — just apply changes / (c)ommit when done? [default: n]

On n (or no response): skip all git operations for this session. Apply changes to working tree only.

On c: enable git mode — then determine branch:

• If already on an ai/* branch, check relevance (does the branch name relate to the skill/agent being modified?):

  On branch ai/<name>. Relevant to this change? (y)es — commit here / (n)ew branch / (Q)uit?
  

  On y: use current branch. On n: create ai/<skill-or-agent-name>-<two-word-description> from current HEAD.

• If on any other branch:

  Create ai/<suggested-name> branch? (y)es / (n)o — abort git mode
  

NEVER commit or push on a non-ai/ branch.* AI-related changes always go on a dedicated ai/* branch.

After all items are processed (git mode only)

If git mode was enabled and at least one item was approved:

All changes applied. Commit? (y)es / (n)o — leave staged

On y: stage and commit once. Message format: [-] <summary of all applied changes> — use ticket number from branch name if available.

On n: leave changes staged but uncommitted.

NEVER push automatically. The human decides when to push.

NEVER

• NEVER commit unless git mode was explicitly opted into — the default is no git operations; committing without explicit opt-in violates user trust and creates unwanted history.
• NEVER commit after each individual item — bundle all approved changes into one commit after the full loop completes; per-item commits create noisy git history and force the user to squash later.
• NEVER commit or push on a non-ai/ branch* — AI-related changes always go on a dedicated ai/* branch to keep work branches clean.
• NEVER assume an existing ai/ branch is relevant* — ai/refine-codebase-skills is not the right branch for unrelated agent changes; always check relevance before committing there.
• NEVER apply two improvements in one change — if both items touch the same code, the user loses the ability to approve them independently, which defeats the purpose of the loop.
• NEVER advance past an item without an explicit approve, revise, or skip response — silence is not approval.
• NEVER ask clarifying questions about an improvement's intent mid-loop — ambiguities must be resolved in step 1 before the first item.
• NEVER show the full file diff — show only the changed lines and a one-line description. Full files bury the signal.
• NEVER let a revision loop run silently past 3 attempts — a stuck item means the improvement is under-specified, not that the agent needs to try harder.
• NEVER offer accept-all (A) before the status board has been shown — the status board IS the plan review; once the user has seen all items listed, offering approve-all at the board level or from item #2 onwards is acceptable.
• NEVER treat lowercase q as a quit signal — only uppercase Q quits; lowercase q is keyboard-adjacent to a and falls through to the unknown-input stop signal.