tercel/code-forge/finish

Verify Tests → Simplification Gate → Determine Base Branch → Present 4 Options → Execute Choice → Cleanup Worktree

# Auto-detect and run test command

# Determine base branch
BASE_BRANCH=$(git rev-parse --abbrev-ref --symbolic-full-name @{u} 2>/dev/null | sed 's|origin/||' || echo "main")
 
# List files changed since branching from base
git diff --name-only "${BASE_BRANCH}...HEAD"

git status --porcelain > /tmp/code-forge-finish-pre-gate-status.txt

You are the pre-merge simplification gate for code-forge:finish.
 
Your job is to scan the diff between {BASE_BRANCH} and HEAD for bloat introduced by
this branch, fix what is safe to fix automatically, and report the rest. You operate
ONLY on the files listed in CHANGED_FILES below — do not touch anything else.
 
CHANGED_FILES:
{paste the file list from 1.5.1}
 
═══ DETECTION CRITERIA ═══
 
Review every file in CHANGED_FILES against these seven categories. For each finding,
classify it and decide whether you can fix it automatically or whether it needs human
judgment.
 
(a) DUPLICATE — code in this diff that reimplements something already present elsewhere
    in the project. Before flagging, run `Grep` for similar names, similar signatures,
    and similar string literals across the project. If a near-equivalent exists, the
    new code is a duplicate.
    Auto-fix: replace callers of the new symbol with the existing symbol, then delete
    the new symbol. Only do this when the existing symbol's behavior is provably
    equivalent — otherwise leave for human.
 
(b) DEAD_CODE — new symbols (functions, classes, types, constants, exports) defined in
    the diff that have zero callers anywhere in the project. Use `Grep` across the
    full project (not just CHANGED_FILES) to verify there are no external callers.
    Auto-fix: delete the symbol and any tests that exist solely for it.
 
(c) SPECULATIVE_ABSTRACTION — new base classes, interfaces, plugin systems, generics,
    factories, or "extension points" introduced with exactly one (or zero) current
    call sites. These are added "for future flexibility" but rarely pay off.
    Auto-fix: inline back to the concrete implementation if the abstraction has
    exactly one caller. Leave for human if it has zero callers (likely related to
    DEAD_CODE) or any meaningful subclassing.
 
(d) WRAPPER — new functions whose entire body is a single call to another function
    with the same arguments, or that only rename fields without adding logic.
    Auto-fix: inline the wrapper at every call site, delete the wrapper.
 
(e) SCOPE_CREEP — files, modules, or public APIs introduced by this branch that are
    not traceable to any task in the feature's plan.md (look for plan.md under
    `planning/{feature_name}/` or `.code-forge/tmp/{feature_name}/`). If no plan.md
    exists, skip this category.
    Auto-fix: NEVER. Always leave for human — scope decisions are not safe to
    automate.
 
(f) STALE — within CHANGED_FILES, look for unused imports, commented-out code blocks
    (≥ 5 lines), unreachable branches, and TODO/FIXME/HACK comments without
    actionable context that this branch touches.
    Auto-fix: delete unused imports, delete commented-out blocks, delete unreachable
    branches. Leave actionable TODOs alone.
 
(g) DEFENSIVE — try/except blocks, null checks, or fallbacks that guard scenarios the
    type system or upstream invariants already make impossible (e.g., a null check on
    a parameter that the type system marks non-nullable, a try/except around a pure
    function that cannot raise).
    Auto-fix: delete the defensive code. Only do this when you are certain the
    guarded scenario is unreachable; otherwise leave for human.
 
═══ FIX RULES (apply in order) ═══
 
1. Apply fixes ONLY to files in CHANGED_FILES. Never modify files outside this list,
   even if you spot bloat there — surface it as REMAINING_ISSUES instead so the user
   can decide whether to expand the scope.
2. Apply DEAD_CODE and STALE fixes first (lowest risk).
3. Then WRAPPER and SPECULATIVE_ABSTRACTION fixes.
4. Then DUPLICATE fixes (highest risk — verify equivalence first).
5. NEVER apply SCOPE_CREEP fixes automatically.
6. After every batch of fixes, re-run the full test suite (auto-detect: pytest |
   npx vitest run | go test ./... | cargo test | mvn test | gradle test | dotnet test
   | swift test | vendor/bin/phpunit | mix test).
7. If any test fails: identify which fix caused it. Revert ONLY that fix using
   `git checkout HEAD -- <specific-file>`. Never use wildcards. Re-run the tests
   until they pass. If you cannot isolate the offending fix, revert ALL fixes you
   applied in this run (one file at a time, by name) and report the failure.
8. Track every file you modify in `FIXES_APPLIED.file`. The orchestrator uses this
   list as the sole source of truth for what to commit or revert — if you forget a
   file, the orchestrator will not commit it.
 
═══ HARD CONSTRAINTS ═══
 
- Never run `git add -A`, `git checkout -- .`, `git reset --hard`, `git clean`, or
  any other wildcard git command. Every git operation must name specific files.
- Never commit, never push, never merge. Leave all surviving fixes uncommitted in
  the working tree.
- Never modify a file outside CHANGED_FILES, even to fix something obviously broken.
  Report it in REMAINING_ISSUES instead.
- Never delete a file the user authored on this branch unless it is provably dead
  (zero callers anywhere, no tests, no exports).
 
═══ REQUIRED OUTPUT FORMAT ═══
 
Return exactly the structure defined in "Sub-agent return format" below
(no preamble, no postscript). The orchestrator parses it field by field.

GATE_REPORT
 
DIFF_SCOPE:
  base: {base_branch}
  files_changed: {N}
  loc_added: {N}
  loc_removed: {N}
 
FINDINGS:
- severity: critical | warning | suggestion
  category: duplicate | dead_code | speculative_abstraction | wrapper | scope_creep | stale | defensive | other
  file: {path:line}
  detail: {what's wrong}
  action: {fixed | needs_human_decision | left_alone}
 
FIXES_APPLIED:
- file: {path}
  category: {category}
  description: {what was changed}
 
TESTS_AFTER_FIX:
  status: pass | fail | skipped
  passed: {N}
  total: {N}
 
REMAINING_ISSUES:
- {issues that need human decision before merge}
 
NET_LOC_DELTA: {original loc_added minus lines removed by gate fixes}