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 `@src/lib/sandbox-config.ts`:
- Around line 429-442: In buildRecomputeSandboxConfigHashScript, avoid
truncating HERMES_STRICT_HASH_FILE directly; instead compute checksums into a
temp file (e.g., HERMES_STRICT_HASH_FILE.tmp in the same dir), verify sha256sum
writes to that temp, set ownership and permissions on the temp, then atomically
mv the temp to HERMES_STRICT_HASH_FILE and only then cp or mv to the
compatibilityHash (also using a temp for the compatibility file if desired) so
that creation, chown, chmod and final replacement are all atomic and cannot
leave a partial/empty strict or compatibility hash; update the sequence that
currently uses `> ${HERMES_STRICT_HASH_FILE}` to use temp files and mv
operations while keeping the existing mkdir, chown and chmod steps.

---

Nitpick comments:
In `@docs/inference/inference-options.md`:
- Line 72: The table cell for "Hermes Provider" contains two sentences on a
single Markdown source line; edit the cell in the inference-options.md table so
each sentence is on its own source line (i.e., break the cell content after
"NemoClaw." so "Available only when onboarding Hermes Agent." and the curated
models sentence are on separate lines) to satisfy the one-sentence-per-line docs
rule.

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

Nitpick comments:
In `@test/e2e/test-hermes-inference-switch.sh`:
- Around line 171-209: The inline Python probe in
test/e2e/test-hermes-inference-switch.sh uses regex-based YAML parsing (see the
Python block that builds the model dict and the regex matches) which is fragile
for nested/multiline YAML; add a brief comment immediately above that Python
heredoc explaining that this is a deliberate lightweight approach for the
controlled E2E test (PyYAML may not be available), call out its limitations
(won't handle nested maps or multiline values), and note that if the config
parsing needs to be robust in future the code should switch to a proper YAML
parser like PyYAML.

In `@test/e2e/test-openclaw-inference-switch.sh`:
- Line 17: The script's shell flags use "set -uo pipefail" without "-e", which
can allow unexpected command failures to continue; either add "-e" (change to
"set -euo pipefail") to enable fail-fast behavior, or explicitly document the
intentional omission by adding a comment near the top explaining that fail() and
manual exit handling are used to collect all results; reference the existing
"set -uo pipefail" line and the fail() invocation so the reviewer can choose the
correct fix.

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 `@test/e2e/test-hermes-inference-switch.sh`:
- Around line 495-505: The registry removal check should be skipped when the
sandbox is intentionally kept (NEMOCLAW_E2E_KEEP_SANDBOX=1); modify the
post-destroy verification so that if the environment variable
NEMOCLAW_E2E_KEEP_SANDBOX is set to "1" the script does not check registry_file
(the "${HOME}/.nemoclaw/sandboxes.json" lookup for "${SANDBOX_NAME}") and does
not call fail/pass for removal—wrap the registry_file existence and grep check
(and the subsequent fail "Sandbox ${SANDBOX_NAME} still in registry after
destroy" / pass "Sandbox ${SANDBOX_NAME} removed") in an if that only runs when
NEMOCLAW_E2E_KEEP_SANDBOX != "1".
- Around line 133-141: The current guard "if session:" allows empty dicts/lists
from deserializing onboard-session.json to pass; change the checks to explicitly
validate that session is a non-empty mapping (e.g., ensure session is a dict and
has keys) before skipping and add an explicit error when it's empty or not a
dict. In other words, replace or augment the "if session:" guard with a check
like "if not isinstance(session, dict) or not session: errors.append('session is
empty or invalid')" and then perform the existing field assertions against
session.get('sandboxName'), session.get('agent'), session.get('provider'), and
session.get('model') (using the variables name, provider, model) so an empty
onboard-session.json fails the test.

In `@test/e2e/test-openclaw-inference-switch.sh`:
- Around line 428-438: The registry existence check assumes the sandbox was
destroyed; skip this check when NEMOCLAW_E2E_KEEP_SANDBOX is set by guarding the
registry_file lookup and the fail/pass logic with the same condition used for
destroy (NEMOCLAW_E2E_KEEP_SANDBOX), i.e., if NEMOCLAW_E2E_KEEP_SANDBOX==1 then
report or skip verification for SANDBOX_NAME instead of running grep on
registry_file; update the block that references registry_file, SANDBOX_NAME,
fail and pass so it only verifies removal when the sandbox was actually
destroyed.