In @.github/workflows/release-dev.yml:
- Around line 90-96: The Darwin upload step in the release workflow omits the
pilot binary; update the Darwin "Upload artifact" action so its with:path
includes both the existing pilot-darwin-arm64.so entry and the pilot executable
(i.e., mirror the Linux upload behavior), ensuring the Darwin artifact list
contains the .so plus the pilot binary.

In @.github/workflows/release.yml:
- Around line 14-17: The approval step currently always uses
needs.prepare-release.outputs.changelog for the summary and ignores the optional
inputs.changelog input; update the approval summary logic in the approval step
to prefer the provided inputs.changelog when non-empty (fall back to
needs.prepare-release.outputs.changelog only if inputs.changelog is empty or
undefined). Locate the approval step where the summary or body is built
(references to approvals/summary and needs.prepare-release.outputs.changelog)
and change it to use a conditional selection (e.g., use inputs.changelog ?
inputs.changelog : needs.prepare-release.outputs.changelog) so manual runs
reflect the custom changelog; apply the same fix to the other occurrence noted
(the secondary summary usage).

In `@install.sh`:
- Line 9: Normalize the VERSION variable after its initialization by stripping
any leading "v" so that downstream URL constructions using VERSION (e.g., in the
code that builds release URLs) don't produce "vv..." sequences; specifically,
after the existing VERSION="${VERSION:-}" assignment, add a normalization step
that removes a leading 'v' from VERSION (e.g., using shell parameter expansion
like VERSION="${VERSION#v}" or equivalent) so all usages of VERSION (the VERSION
variable) are safe.
- Around line 446-454: The issue is that version_arg="--target-version $VERSION"
combines flag and value into one argv element causing argparse to reject it;
update the calls that use "$version_arg" (the two python -m installer install
invocations inside the is_in_container / saved_mode branch and the else branch)
to pass the flag and value as separate arguments (e.g., --target-version
"$VERSION") or build an array of arguments so the flag and VERSION are separate
elements, ensuring argparse receives "--target-version" and the VERSION as
distinct argv entries.

In `@installer/cli.py`:
- Line 104: The installer references an unreachable subscription URL ("Subscribe
at: https://license.claude-pilot.com") in console.print calls; update all
occurrences of that literal string to a verified, reachable license URL (or to a
configurable constant) so users aren't sent to a 502 page—search for the string
"Subscribe at: https://license.claude-pilot.com" in installer/cli.py and replace
it with the correct URL (or a LICENSE_URL constant used by the CLI output), and
ensure any tests or docs that reference the old URL are updated accordingly.

In `@installer/steps/claude_files.py`:
- Around line 246-276: The cleanup currently erases entire global dirs in
_cleanup_old_directories via _clear_directory_safe which can delete user
content; instead iterate only the managed files listed in the categories maps
(the FileInfo entries under categories["commands"] and categories["rules"]) and
remove those specific paths (or move them to a timestamped backup under
home_claude_dir/"backup") while preserving any other files; update the logic in
_cleanup_old_directories to compute target paths from each FileInfo, check
existence and ownership/marker (if available) before deletion, and call
_clear_directory_safe only on empty dirs or after backing up to ensure
idempotency and cross-platform safe behavior.
- Around line 339-378: The UI success message in _install_category_files
currently reports len(file_infos) which can overstate successes; update the
post-install reporting to use the actual installed list (len(installed)) and, if
failed is non-empty, show a different message or include failure details (e.g.,
count or list) so users see partial failures; modify the logic around
ui.success/ui.spinner in _install_category_files to compute installed/failed and
call ui.success or ui.error accordingly, referencing the installed and failed
variables and the function name _install_category_files to locate the change.

In `@installer/steps/migration.py`:
- Around line 132-152: The _cleanup_global_old_folders function currently always
uses shutil.rmtree on both Path.home() / ".claude" / "bin" and Path.home() /
".claude" / "config"; change it so that you still remove the "bin" folder with
shutil.rmtree (and append its path on success) but for the "config" folder first
check if it exists and is empty (e.g., using any(folder.iterdir()) or similar)
and only remove it with folder.rmdir() if empty (append its path on success);
preserve the existing try/except behavior to ignore OSError/IOError and ensure
the function returns the list of removed paths.
- Around line 155-175: The _migrate_custom_rules function currently uses
shutil.move which will overwrite existing files; change it to skip moving a rule
if dest (new_rules_dir / rule_file.name) already exists to preserve existing
rules and keep the function idempotent, leaving migrated counting unchanged
(only increment when a file is actually moved); update the try/except to catch
and log failures but avoid swallowing all errors silently, and add unit tests
for _migrate_custom_rules that cover: migrating new files, skipping when
destination exists, and running twice to verify idempotency (use the
old_custom_dir and new_rules_dir variables and shutil.move reference to locate
the code to modify).

In `@installer/steps/shell_config.py`:
- Around line 22-27: alias_exists_in_file currently misses unmarked "alias
claude" entries, causing duplicate aliases; update the function to also perform
a regex search for any shell alias definition for "claude" (e.g., match lines
like "alias claude='...'" with optional leading whitespace) in addition to
checking CLAUDE_ALIAS_MARKER and OLD_CCP_MARKER. Concretely, in
alias_exists_in_file use the file content and a multiline regex (e.g.,
(?m)^\s*alias\s+claude\b) to detect unmarked aliases and return True if that
regex matches or either marker is present.
- Around line 30-79: remove_old_alias fails to fully remove fish-style "function
... end" blocks because brace_count logic can clear inside_function prematurely;
update the function to separately track brace-delimited shells vs fish-style
functions (e.g., use two flags like inside_brace_function and
inside_fish_function or branch based on how the function was detected).
Specifically: when you detect "function ccp" or "function claude" set a
fish-specific flag and do not touch brace_count for that case, and only clear
that flag when a line with stripped == "end" is seen; keep the existing
brace_count handling for ccp()/ccp () { style functions but ensure it does not
affect the fish flag. Adjust the loop so lines inside either kind of function
are skipped and removed, and ensure OLD_CCP_MARKER / CLAUDE_ALIAS_MARKER and
alias lines are still skipped as before.

In `@install.sh`:
- Around line 32-48: The get_latest_release function currently assumes tag_name
begins with "v"; update its parsing in both the curl and wget branches to
extract tag_name and strip an optional leading "v" so tags like "6.0.0" produce
a clean version. Locate the lines that pipe API output into grep/sed (inside
get_latest_release) and replace the sed expression with one that captures the
tag value and returns only the inner version (i.e., allow an optional leading
"v" and output the remainder), keeping the same fallback behavior and return
values for the function.

In `@installer/steps/shell_config.py`:
- Around line 22-32: alias_exists_in_file currently misses legacy shell function
definitions (e.g., bash "ccp()" or "claude()" and fish "function ccp" /
"function claude") so ShellConfigStep.run doesn't remove old wrappers; update
alias_exists_in_file to also check for these patterns (e.g., look for "ccp()",
"claude()", "function ccp", "function claude" or use regex to detect
"^\s*(function\s+)?(ccp|claude)\s*\("/line-start variants) OR alternatively call
remove_old_alias unconditionally from ShellConfigStep.run and rely on
remove_old_alias's return value; reference functions: alias_exists_in_file,
remove_old_alias, and ShellConfigStep.run when making the change.

In `@installer/tests/unit/steps/test_shell_config.py`:
- Around line 248-249: The current assertion 'assert "end" not in content or "#
after" in content' can pass when "end" remains; change it to two explicit
assertions so partial removals are caught: keep 'assert "function claude" not in
content', then replace the combined check with 'assert "end" not in content' and
'assert "# after" in content' so the test ensures the fish function body was
removed while the trailing "# after" marker remains.

In `@installer/tests/unit/steps/test_dependencies.py`:
- Around line 222-223: The test stores an unused reference _original_info =
ui.info solely for potential restoration but never restores it and includes an
unnecessary noqa; remove the dead variable and the noqa directive by deleting
the assignment line for _original_info and keep the replacement assignment
ui.info = lambda message: info_calls.append(message) so the test no longer
defines or references _original_info.