Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@examples/experimental/claude_code_agent_adapter/pyproject.toml`:
- Around line 27-45: Add the example to the root project examples list and
rename the uv source to a nat_-prefixed name: in root pyproject.toml include
"nat_claude_code_agent_adapter" in the examples/installable examples array so
the example is packaged, and in
examples/experimental/claude_code_agent_adapter/pyproject.toml change the
[tool.uv.sources] key from "nvidia-nat" to a nat_-prefixed identifier (e.g.,
"nat_nvidia_nat" or "nat_nvidia-nat") so the source reads like nat_nvidia_nat =
{ path = "../../..", editable = true }; leave the project.name and the entry
point nat_claude_code_agent_adapter = "nat_claude_code_agent_adapter.register"
intact.

In `@examples/experimental/codex_agent_adapter/package.json`:
- Around line 4-6: Update the dependency declaration for `@openai/codex-sdk` in
package.json to pin it to the tested version/range instead of "latest" — replace
the current "`@openai/codex-sdk`": "latest" entry with a stable version (e.g.,
"`@openai/codex-sdk`": "0.135.0" or a constrained range "`@openai/codex-sdk`":
"^0.135.0") so installs align with the package-lock.json and the README's
supported Codex version.

In `@examples/experimental/codex_agent_adapter/pyproject.toml`:
- Around line 30-47: The root pyproject.toml needs the new example registered:
add "nat_codex_agent_adapter" to the top-level examples = [...] array and add a
corresponding [tool.uv.sources] entry named nat_codex_agent_adapter that points
to the example path (nat_codex_agent_adapter = { path =
"examples/experimental/codex_agent_adapter", editable = true }) so the package
declared as name = "nat_codex_agent_adapter" and the entry point nat.components
-> nat_codex_agent_adapter = "nat_codex_agent_adapter.register" in the example
pyproject.toml is discoverable by the root build/configuration.

In
`@examples/experimental/codex_agent_adapter/src/nat_codex_agent_adapter/register.py`:
- Around line 397-400: The `_run_codex_cli` implementation contains a dead
non-relay branch (the `else: command = _build_codex_command(...)` path) that is
never used because `_run_codex` only calls `_run_codex_cli` when
`config.relay_enabled` is true and all non-relay calls go to `_run_node_sdk`;
remove the unreachable branch (and the helper `_build_codex_command` if it
becomes unused) or, if you intended to support non-relay CLI execution, change
`_run_codex` to route non-relay calls into `_run_codex_cli`. Also update the
`FileNotFoundError` handler wording inside `_run_codex_cli` to refer
specifically to the Codex relay binary (and remove “Codex CLI” wording) so the
error message matches the actual execution path. Ensure any removed helper
references (e.g., `_build_codex_command`) are cleaned up and tests/usage updated
accordingly.

In `@examples/experimental/cursor_agent_adapter/pyproject.toml`:
- Around line 27-44: Add the missing example registration: update the root
pyproject.toml to include "nat_cursor_agent_adapter" in the top-level examples =
[...] list, and add a corresponding entry under [tool.uv.sources] named
nat_cursor_agent_adapter that maps the package name to the example package
(using editable = true) so the nat_cursor_agent_adapter component and its
entry-point (nat_cursor_agent_adapter = "nat_cursor_agent_adapter.register") are
discoverable and installable.

In `@examples/experimental/cursor_agent_adapter/README.md`:
- Line 150: The phrase "Relay's raw event file" in README.md uses a possessive
's on an inanimate object; update the sentence to use a non-possessive form such
as "the raw Relay event file" (or "Relay raw event file") so the documentation
follows the guideline against using possessive 's with inanimate objects; search
for the exact string "Relay's raw event file" and replace it accordingly in the
README.md content.

In `@examples/experimental/hermes_agent_adapter/configs/config-relay.yml`:
- Around line 4-11: The config currently enables relay (workflow._type:
hermes_agent, relay_enabled: true) but does not set relay_atof_output_dir, so
ATOF events are written to a temp dir and removed; update the config to set
relay_atof_output_dir to the README's inspect path (e.g.
.tmp/nat-relay-hermes-atof) so the events.jsonl file is preserved for the `cat
./.tmp/nat-relay-hermes-atof/events.jsonl` step, or alternatively change the
README to point to the phoenix config that already sets relay_atof_output_dir.

In `@examples/experimental/hermes_agent_adapter/pyproject.toml`:
- Around line 27-44: The root pyproject.toml must include the example module and
a uv source for nat_hermes_agent_adapter; add "nat_hermes_agent_adapter" to the
root examples array and add a [tool.uv.sources] entry mapping
nat_hermes_agent_adapter to the examples/experimental/hermes_agent_adapter path
(editable true) so the project name nat_hermes_agent_adapter and its existing
[tool.uv.sources] usage in the example match the root config.

In `@examples/experimental/hermes_agent_adapter/README.md`:
- Line 126: The sentence "You can inspect Relay's raw event file:" uses a
possessive 's with an inanimate object; update the README line that currently
reads "You can inspect Relay's raw event file:" (search for that exact string)
to a non-possessive phrasing such as "You can inspect the raw event file of
Relay:" or "You can inspect the Relay raw event file:" to comply with the
documentation guideline.

In
`@examples/experimental/openclaw_agent_adapter/src/nat_openclaw_agent_adapter/register.py`:
- Around line 273-274: The public workflow function openclaw_agent (decorated
with `@register_function` and accepting config: OpenClawAgentWorkflowConfig and
_builder: Builder) must include a return type hint and a Google-style docstring;
update the signature to add the appropriate return type (e.g., ->
YourReturnType) and add a docstring immediately below the def that documents
Args (config, _builder) and Returns (describe return value and type), mentioning
any side effects or exceptions as needed so the function complies with public
API guidelines.

In `@packages/nvidia_nat_core/src/nat/experimental/relay_telemetry_bridge.py`:
- Around line 52-63: The loader load_atof_jsonl currently calls json.loads on
every non-empty line and will raise on a malformed line; change it to open the
file with encoding="utf-8" and wrap the json.loads call in a try/except (catch
json.JSONDecodeError/ValueError) so malformed lines are skipped (optionally log
or count them) and valid dict events continue to be appended to the events list;
keep the function signature and return type the same and only alter the file
open call and the per-line parsing logic in load_atof_jsonl.

---

Nitpick comments:
In
`@examples/experimental/claude_code_agent_adapter/src/nat_claude_code_agent_adapter/register.py`:
- Around line 376-379: The current routing makes the CLI path unreachable unless
config.relay_enabled is true, leaving the else branch in _run_claude_code_cli
and parts of the FileNotFoundError message unused; either remove the dead CLI
fallback code (drop the else branch in _run_claude_code_cli and related
non-relay FileNotFoundError text) or add a clear CLI fallback path by updating
_run_claude_code to call _run_claude_code_cli when relay is disabled (or when a
new config flag indicates CLI mode), and ensure _build_claude_command is invoked
only in the supported branch; modify the logic in _run_claude_code,
_run_claude_code_cli, and any FileNotFoundError messaging accordingly so there
is no dead-code path left.

In `@examples/experimental/openclaw_agent_adapter/README.md`:
- Line 58: Replace possessive "'s" usages with "of" phrasing for inanimate
nouns: change "OpenClaw's plugin system" to "the plugin system of OpenClaw" and
"the example's ..." to "the example settings" (or "settings of the example") in
the README content; search for and update any other occurrences of possessive
"'s" applied to inanimate objects so documentation follows the style guideline.

In `@packages/nvidia_nat_core/src/nat/experimental/relay_telemetry_bridge.py`:
- Around line 66-83: The docstring for relay_events_to_intermediate_steps uses
NumPy-style Parameters/---------- formatting; update it to a Google-style
docstring: replace the Parameters section with an Args: block listing events,
root_parent_id, and function_ancestry (each with a one-line description and
types where helpful), and keep the brief top description and Returns/Examples if
present in Google style; ensure the function name
relay_events_to_intermediate_steps and parameter names root_parent_id and
function_ancestry are referenced exactly as in the signature.

In
`@packages/nvidia_nat_core/tests/nat/experimental/test_relay_telemetry_bridge.py`:
- Around line 124-145: Add a negative-path test that writes a JSONL file
containing both a valid event and a malformed/partial line, then call
load_atof_jsonl and inject_atof_jsonl to verify the loader skips the bad line
and only returns/injects the valid event: create a file with one well-formed
event JSON and one broken line, assert load_atof_jsonl(path) returns only the
valid event, subscribe to Context.intermediate_step_manager, call
inject_atof_jsonl(path, context=...), and assert the captured
IntermediateStep(s) correspond only to the valid event (check event types,
parent_id, and count).

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@pyproject.toml`:
- Around line 129-180: The examples and UV source roots are missing the new
OpenClaw adapter; add "nat_openclaw_agent_adapter" to the top-level examples
array (the examples list shown) and also add the corresponding
"nat_openclaw_agent_adapter" entry under the [tool.uv.sources] section so the
adapter is discoverable via the same root/example install flow as the other
adapters.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In
`@examples/experimental/claude_code_agent_adapter/src/nat_claude_code_agent_adapter/register.py`:
- Around line 213-235: In _run_claude_code_cli validate that
Path(config.working_directory).resolve() exists and is a directory before
creating the subprocess: check cwd.exists() and cwd.is_dir(), call
relay_temp_dir.cleanup() and raise a clear RuntimeError indicating the provided
working_directory is invalid (include the path) so a bad workspace is reported
instead of a FileNotFoundError about the relay/claude command; keep the existing
FileNotFoundError handler for missing executables.

In
`@examples/experimental/cursor_agent_adapter/src/nat_cursor_agent_adapter/register.py`:
- Around line 222-244: The FileNotFoundError except block in _run_cursor_agent
currently conflates a missing working directory with a missing executable;
before attempting asyncio.create_subprocess_exec, validate cwd =
Path(config.working_directory).resolve() (or use Path(config.working_directory)
then resolve only after existence) and raise a clear RuntimeError (and cleanup
relay_temp_dir) if the directory does not exist or is not a directory, so that
the existing except FileNotFoundError only handles missing executable errors
from asyncio.create_subprocess_exec; update the pre-subprocess logic in
_run_cursor_agent to check cwd.exists() and cwd.is_dir() and handle
cleanup/raise accordingly.

In
`@examples/experimental/hermes_agent_adapter/src/nat_hermes_agent_adapter/register.py`:
- Around line 205-227: The FileNotFoundError raised by
asyncio.create_subprocess_exec can be caused by a missing working directory as
well as a missing executable; before calling asyncio.create_subprocess_exec in
_run_hermes_agent, validate that Path(config.working_directory).resolve() (cwd)
exists and is a directory and if not, call relay_temp_dir.cleanup() and raise a
clear RuntimeError that the configured working_directory is invalid (include the
invalid path), keeping the existing FileNotFoundError except-block to handle the
missing executable/uvx case separately.

---

Outside diff comments:
In
`@examples/experimental/claude_code_agent_adapter/src/nat_claude_code_agent_adapter/register.py`:
- Around line 167-170: The current _write_relay_config flattens
ClaudeCodeAgentWorkflowConfig.command and .command_args using " ".join which
re-tokenizes paths with spaces/quotes; instead import shlex and build the full
argv list ([config.command] + config.command_args) and use shlex.join(...) to
produce a properly shell-escaped command string before json.dumps and
path.write_text; update the function _write_relay_config to use shlex.join on
that argv list (and add the shlex import) so path.write_text writes a safe,
quoted command to the TOML.

In
`@examples/experimental/codex_agent_adapter/src/nat_codex_agent_adapter/register.py`:
- Around line 225-247: The FileNotFoundError from asyncio.create_subprocess_exec
can come from a missing working directory, so in _run_codex_cli validate cwd
(from Path(config.working_directory).resolve()) exists and is a directory before
calling create_subprocess_exec; if it does not exist raise a clear RuntimeError
about the invalid working_directory (and still cleanup relay_temp_dir) instead
of falling through to the subprocess FileNotFoundError handler that assumes the
relay binary is missing. Ensure the check uses the same cwd variable and keep
relay_temp_dir.cleanup() on error paths.
- Around line 172-175: The code in _write_relay_config currently flattens argv
with " ".join, losing argument boundaries; instead build the argv list and
serialize it (e.g., JSON/TOML array) so spaces are preserved: construct args =
[config.command, *config.command_args, *_build_codex_root_args(config)] and
write that serialized list into the TOML (use json.dumps(args) or a TOML array
representation) when calling path.write_text; update references to
config.command, config.command_args and _build_codex_root_args to use the list
form rather than a joined string.

In
`@examples/experimental/cursor_agent_adapter/src/nat_cursor_agent_adapter/register.py`:
- Around line 159-163: The current _write_relay_config function flattens argv
with " ".join which loses argument boundaries; instead serialize the command as
a JSON array to preserve argv. Replace cursor_command = "
".join([config.command, *config.command_args]) with using the list
[config.command, *config.command_args] and write it via json.dumps (e.g.,
f"command = {json.dumps([config.command, *config.command_args])}\n") while
keeping patch_restore_hooks written as before; this ensures
CursorAgentWorkflowConfig.command and command_args are preserved exactly in the
Relay config.

In
`@examples/experimental/hermes_agent_adapter/src/nat_hermes_agent_adapter/register.py`:
- Around line 141-144: The current _write_relay_config builds hermes_command
with " ".join(...) which flattens argv boundaries and loses quoting for
whitespace-bearing paths; instead serialize the command as an array to preserve
each argv element. Update _write_relay_config (referencing
HermesAgentWorkflowConfig, config.command and config.command_args, and
Path.write_text) to write the command field as a JSON array (i.e.,
json.dumps([config.command, *config.command_args])) into the TOML snippet so
each argument is preserved exactly.

---

Nitpick comments:
In
`@examples/experimental/claude_code_agent_adapter/src/nat_claude_code_agent_adapter/register.py`:
- Around line 269-287: The public async workflow function claude_code_agent
lacks a Google-style docstring and an explicit return type; add a module-level
docstring if missing and add a Google-style docstring immediately under the
async def claude_code_agent(...) describing purpose, parameters (config:
ClaudeCodeAgentWorkflowConfig, _builder: Builder) and returns. Also add an
explicit return annotation to claude_code_agent such as ->
AsyncGenerator[FunctionInfo, None] (or the project's preferred AsyncGenerator
generic) to reflect that the function yields FunctionInfo; keep existing inner
helpers (_response_fn, _stream_fn) unchanged.

In
`@examples/experimental/codex_agent_adapter/src/nat_codex_agent_adapter/register.py`:
- Around line 281-299: Add a Google-style docstring and an explicit return type
to the public async generator function codex_agent: annotate its signature with
an appropriate AsyncGenerator return type (e.g., -> AsyncGenerator[FunctionInfo,
None] or the project's exact FunctionInfo wrapper type) and include a docstring
that briefly describes the workflow's purpose, its parameters (config:
CodexAgentWorkflowConfig, _builder: Builder), and what it yields (FunctionInfo
containing single_fn and stream_fn). Keep the docstring concise and place it
immediately under the def codex_agent(...) line; you do not need to change the
inner helper functions (_response_fn/_stream_fn) unless they are part of the
public API.

In
`@examples/experimental/cursor_agent_adapter/src/nat_cursor_agent_adapter/register.py`:
- Around line 277-295: Add a Google-style docstring to the exported async
generator function cursor_agent describing its purpose, parameters (config:
CursorAgentWorkflowConfig, _builder: Builder), and what it yields (FunctionInfo
instances) using Args and Yields sections; also add an explicit return type
annotation indicating it is an async generator that yields FunctionInfo (e.g.,
-> AsyncGenerator[FunctionInfo, None]) and import AsyncGenerator from typing if
not present so the signature and docs satisfy the public-API guidelines.

In
`@examples/experimental/hermes_agent_adapter/src/nat_hermes_agent_adapter/register.py`:
- Around line 260-278: The public async function hermes_agent lacks a
Google-style docstring and an explicit return type annotation; add a concise
docstring above hermes_agent describing its purpose, parameters (config:
HermesAgentWorkflowConfig, _builder: Builder), and the yielded value, and
annotate its signature with the correct return type (e.g., ->
AsyncGenerator[FunctionInfo, None]) to reflect that it yields a FunctionInfo;
ensure the docstring also briefly documents what _response_fn and _stream_fn do
and what FunctionInfo.create yields so callers and type checkers can understand
the API.