Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/commands/init.sh`:
- Around line 275-277: The Fish wrapper's hardcoded subcommand list is out of
sync with the Bash list (the Bash block uses COMP_CWORD/COMPREPLY with compgen
and includes "trust"); update the Fish completion template in this file to match
the Bash subcommand list (add "trust" and any other missing entries) so that
running the init fish generator (git gtr init fish) produces the same
completions as the COMPREPLY/compgen block; alternatively, factor the subcommand
list into a single variable and reference it from both the COMP_CWORD/COMPREPLY
block and the Fish template so they stay synchronized.

In `@lib/commands/trust.sh`:
- Around line 13-40: cmd_trust reads and displays hooks then calls
_hooks_mark_trusted which re-hashes the file after the prompt; compute the
reviewed marker (e.g., call the hashing function/_get_marker used by
_hooks_mark_trusted) immediately after reading hook_content and before calling
prompt_yes_no, store that marker (and reviewed path if needed), show the hooks
as before, then after prompt_yes_no pass that stored marker into a new variant
of the marker-writer (or change _hooks_mark_trusted to accept a marker argument)
so you persist the exact reviewed marker instead of re-computing it
post-confirmation; keep _hooks_are_trusted usage unchanged but compare against
the stored marker where appropriate.

In `@lib/hooks.sh`:
- Around line 16-32: The trust marker key is currently content-only (via
_hooks_file_hash) which makes markers shared across repos; update
_hooks_trust_path to incorporate a repository-unique identifier (e.g.,
repository root path, remote URL, or git object like commit/slug) into the
computed key so the path becomes "$_GTR_TRUST_DIR/<repo-id>-<hash>" instead of
just "<hash>", and likewise update the wrapper generation in
lib/commands/init.sh to include the same repo identifier when creating trust
markers so markers are repo-scoped; keep using _hooks_file_hash for the content
part and prepend or append the chosen repo-id consistently.

---

Outside diff comments:
In `@lib/adapters.sh`:
- Around line 307-360: The metacharacter check currently blocks shell operators
but allows path separators, letting relative paths like "./tool" pass; update
the validation in the adapter loader to reject path separators for the
configured command (check either "name" or the extracted "cmd_name" for '/' and
'\' characters) before accepting it, and return an error like other checks;
alternatively, resolve the command with "command -v" and ensure the resolved
path is a PATH-resolved command (not a relative/containment path) before
proceeding — reference the variables "name" and "cmd_name" and the caller
"_run_configured_command" when making this change.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/adapters.sh`:
- Around line 307-318: The override-loading code currently sources the adapter
file using adapter_selector/adapter_file then returns early (return 0), which
drops the original full name (e.g., "nano -w") so flags never reach the generic
runner; change the flow to source the override but do NOT return
immediately—instead set a marker (e.g., ADAPTER_OVERRIDE_LOADED=1 or set
adapter_override_file="$adapter_file") and let the function continue so the full
name variable (the original name variable) is preserved and passed to the
generic runner; apply the same change for the second override branch that
mirrors this logic (the block that uses adapter_selector/adapter_file again).
- Around line 183-193: The validation allows shell builtins and wrapper commands
with flags which can be exploited because _run_configured_command uses eval "set
-- $command_string"; change the validation to use type -P instead of command -v
so only PATH executables (not builtins) are accepted, and add a check that the
command string contains no option flags (tokens starting with -) or disallowed
wrapper executables (e.g., sh, bash, eval, builtin) before allowing it to be
stored/used; update the validation logic that currently checks metacharacters
(refer to the validation function called before _run_configured_command) to
explicitly reject tokens beginning with '-' and to reject commands where type -P
returns empty or where the resolved command name matches a blacklist of shell
wrappers.

In `@lib/hooks.sh`:
- Around line 91-102: _hooks_write_trust_marker() currently creates the target
file path then writes to it, which can leave a partial/empty marker on failure;
change it to write atomically by creating a temp file in the same directory
(e.g. "${trust_path}.tmp"), write the payload to that temp file, flush it (fsync
the file) and then mv the temp into place, and optionally fsync the directory
(using _GTR_TRUST_DIR) to ensure durability; additionally, tighten
_hooks_are_trusted() to validate the stored payload (e.g. check file is
non-empty and matches the expected config_file or pattern) before treating it as
trusted so a truncated/empty marker cannot be accepted.

---

Duplicate comments:
In `@lib/hooks.sh`:
- Around line 22-29: The current hook trust key uses only path+hook content, so
a different commit checked out at the same path reuses approvals; update
_hooks_file_hash to incorporate repository state by including a stable commit
identifier (e.g., git rev-parse --verify HEAD) or a tree hash of the working
tree when present, then pass that combined string into _hooks_content_hash so
any change to checked-out content invalidates prior approvals; adjust callers
like run_hooks to continue using _hooks_file_hash unchanged so the trust key now
reflects both .gtrconfig hook entries and the current repository commit/tree
state.

---

Nitpick comments:
In `@lib/hooks.sh`:
- Around line 41-79: The trust-key logic is duplicated between lib/hooks.sh
(functions _hooks_trust_key, _hooks_trust_key_for_content, _hooks_trust_path,
_hooks_trust_path_for_content) and the generated helper in lib/commands/init.sh;
consolidate to a single source by moving the core trust-key recipe into one
canonical implementation (either a shared template file or a single library
function) and change the generator to emit a thin wrapper that calls that
canonical implementation so both runtime and generated code remain byte-for-byte
aligned; ensure the generator references the canonical symbol name (e.g., the
single source function) when producing the wrapper and remove the duplicate
manual implementations.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/hooks.sh`:
- Around line 117-118: If _hooks_canonical_config_path fails after creating the
temp file, the temp file referenced by temp_path (created via mktemp
"$_GTR_TRUST_DIR/.trust.XXXXXX") is leaked; update the failure path so that when
canonical_config_file=$(_hooks_canonical_config_path "$config_file") returns
non-zero you remove the temp file (rm -f "$temp_path") before returning (or set
a local cleanup/trap that removes temp_path on early exit) to ensure no
temporary files are left behind.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/adapters.sh`:
- Around line 296-304: Remove the overly restrictive pre-parse validations:
delete the helper functions _configured_command_uses_path_arg and
_configured_command_has_safe_syntax and stop rejecting arguments like
"~/.toolrc" or raw metacharacters, and instead only validate the adapter command
name inside _load_adapter(); ensure _load_adapter() no longer calls those two
functions and only checks the command name token (leaving GTR_*_CMD_ARGS
handling to the adapter overrides) so argv-style execution via "$@" can accept
literal path-like args and metacharacters as safe data.

In `@lib/commands/init.sh`:
- Around line 174-180: The regex passed to string replace in the branch handling
linked worktrees is over-escaped ('/\\.git$') so Fish treats the backslash
literally; change the pattern to '/\.git$' (single backslash) in the string
replace -r call that operates on _gtr_git_common_dir so the trailing '/.git'
suffix is correctly stripped (i.e., update the expression string replace -r
'/\\.git$' '' -- "$_gtr_git_common_dir" to use '/\.git$' instead).

In `@lib/hooks.sh`:
- Around line 135-175: The code currently treats failures from
_hooks_current_trust_path as success (returning 0), merging “no hooks” with real
errors; change the error handling so that when _hooks_current_trust_path fails
it returns non-zero (fail closed). Specifically, in _hooks_are_trusted() replace
the trust_path assignment branch that uses "|| return 0" with "|| return 1" so
resolution errors mark hooks untrusted, and in _hooks_mark_trusted() do the same
(replace "trust_path=$(_hooks_current_trust_path \"$config_file\") || return 0"
with "|| return 1") so marker writes are aborted on resolution errors.

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@lib/commands/init.sh`:
- Line 179: The Fish regex used in the printf call still contains an escaped
backslash ('/\\.git$') so Fish matches a literal backslash instead of the '.git'
suffix; update the regex in the printf/string replace invocation (the line
calling string replace -r '/\\.git$' '' -- "$_gtr_git_common_dir") to use
'/\.git$' (i.e., remove the extra backslash) so Fish correctly strips '/.git'
from linked worktree paths.

---

Nitpick comments:
In `@tests/adapters.bats`:
- Around line 131-151: Extract the repeated "mktemp + cat + chmod + PATH"
scaffold into a reusable test helper (e.g., create_mock_bin or setup_mock_bin)
and call it from the tests instead of inlining the block; specifically replace
the repeated scaffolding around mock_bin_dir, creating an executable (tool/bunx)
and prepending it to PATH in the tests named "_load_adapter ..." and
"_load_adapter allows path-like literal arguments", ensure the helper accepts
the binary name and content, makes it executable, updates PATH, and performs
cleanup (or returns the temp dir for later cleanup) so the duplicate code at the
other occurrence (lines 159-165) is removed.