test/e2e/e2e-cloud-experimental/features/skill/add-sandbox-skill.sh-19-20 (1)

19-20: ⚠️ Potential issue | 🟡 Minor

Fix the fixture path in this usage comment.

The script's actual default is features/skill/fixtures/skill-smoke-template.SKILL.md on Line 38, so the path shown here points readers to the wrong file when they update the smoke template.

🤖 Prompt for AI Agents

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

In `@test/e2e/e2e-cloud-experimental/features/skill/add-sandbox-skill.sh` around
lines 19 - 20, Update the usage comment at the top of add-sandbox-skill.sh that
mentions the default template: the comment currently points to the wrong
fixture; change the path text to the actual default template path used by the
script (features/skill/fixtures/skill-smoke-template.SKILL.md) so SKILL_FILE /
SKILL_BODY documentation matches the behavior in the script.

test/policies.test.js-90-96 (1)

90-96: ⚠️ Potential issue | 🟡 Minor

Restore the previous env var value instead of always deleting it.

At Line 95, unconditional delete can leak state across tests when NEMOCLAW_OPENSHELL_BIN was pre-set by the runner/environment.

Suggested fix

   it("uses the resolved openshell binary when provided by the installer path", () => {
+    const previousOpenshellBin = process.env.NEMOCLAW_OPENSHELL_BIN;
     process.env.NEMOCLAW_OPENSHELL_BIN = "/tmp/fake path/openshell";
     try {
       const cmd = policies.buildPolicySetCommand("/tmp/policy.yaml", "my-assistant");
       assert.equal(cmd, "'/tmp/fake path/openshell' policy set --policy '/tmp/policy.yaml' --wait 'my-assistant'");
     } finally {
-      delete process.env.NEMOCLAW_OPENSHELL_BIN;
+      if (previousOpenshellBin === undefined) {
+        delete process.env.NEMOCLAW_OPENSHELL_BIN;
+      } else {
+        process.env.NEMOCLAW_OPENSHELL_BIN = previousOpenshellBin;
+      }
     }
   });

🤖 Prompt for AI Agents

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

In `@test/policies.test.js` around lines 90 - 96, The test mutates
process.env.NEMOCLAW_OPENSHELL_BIN and currently unconditionally deletes it in
the finally block, which can leak state; change the test to save the previous
value (e.g., const prev = process.env.NEMOCLAW_OPENSHELL_BIN) before assigning
"/tmp/fake path/openshell" and in the finally block restore it: if prev is
undefined delete process.env.NEMOCLAW_OPENSHELL_BIN else set
process.env.NEMOCLAW_OPENSHELL_BIN = prev; keep the assertion using
policies.buildPolicySetCommand unchanged.

.agents/skills/nemoclaw-reference/references/architecture.md-44-52 (1)

44-52: ⚠️ Potential issue | 🟡 Minor

Reconcile the runtime model in this doc.

This section says the blueprint runtime lives in nemoclaw/src/blueprint/, but the unchanged intro still describes a Python blueprint subprocess, and the lifecycle section below still omits the newly listed rollback phase. Please update those sections together so the architecture reads consistently.

🤖 Prompt for AI Agents

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

In @.agents/skills/nemoclaw-reference/references/architecture.md around lines 44
- 52, Update the architecture doc to consistently reflect the TypeScript
blueprint runtime under nemoclaw/src/blueprint/: remove or replace the outdated
"Python blueprint subprocess" wording in the intro with a clear statement that
the blueprint runtime is implemented in TypeScript in the plugin source tree
(referencing runner.ts, ssrf.ts, snapshot.ts, state.ts), and amend the lifecycle
section to include the missing "rollback" phase and note that runner.ts
implements plan/apply/status/rollback behavior; keep descriptions aligned (e.g.,
mention snapshot.ts handles migration snapshot/restore and state.ts handles
persistent run state) so all mentions match the new runtime model.

test/security-binaries-restriction.test.js-23-29 (1)

23-29: ⚠️ Potential issue | 🟡 Minor

Ignore top-level comments when leaving network_policies.

This exit condition stops scanning on any column-0 text, so a # ... comment between entries can end the section early and let later policy blocks skip the binaries assertion.

Suggested fix

-      if (inNetworkPolicies && /^\S/.test(line) && line.trim() !== "") {
+      if (inNetworkPolicies && /^\S/.test(line) && line.trim() !== "" && !/^\s*#/.test(line)) {

🤖 Prompt for AI Agents

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

In `@test/security-binaries-restriction.test.js` around lines 23 - 29, The exit
condition for leaving the network_policies block incorrectly treats any column-0
text (including top-level comments) as the end, so update the check in the block
that uses inNetworkPolicies, currentBlock and the /^\S/ test (the if with
inNetworkPolicies && /^\S/.test(line) && line.trim() !== "") to ignore top-level
comments: first compute trimmedLine = line.trim(), and only treat a column-0
line as the section end if trimmedLine is non-empty and does NOT start with '#'
(i.e. skip lines where trimmedLine.startsWith('#')), leaving comment lines
inside network_policies and preventing premature exit so later policy blocks
still get validated for binaries.

test/e2e/e2e-cloud-experimental/test-port8080-conflict.sh-25-26 (1)

25-26: ⚠️ Potential issue | 🟡 Minor

Fix the usage example path.

The documented command drops the e2e-cloud-experimental/ directory, so a copy/paste run from the repo root will miss this script.

Suggested fix

-#   NEMOCLAW_NON_INTERACTIVE=1 NVIDIA_API_KEY=nvapi-... bash test/e2e/test-port8080-conflict.sh
+#   NEMOCLAW_NON_INTERACTIVE=1 NVIDIA_API_KEY=nvapi-... bash test/e2e/e2e-cloud-experimental/test-port8080-conflict.sh

🤖 Prompt for AI Agents

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

In `@test/e2e/e2e-cloud-experimental/test-port8080-conflict.sh` around lines 25 -
26, Update the usage comment at the top of test-port8080-conflict.sh so the
example command includes the correct relative path to this script
(e2e-cloud-experimental/test-port8080-conflict.sh) instead of just
test/e2e/test-port8080-conflict.sh; edit the comment lines near the
shebang/usage block to show: NEMOCLAW_NON_INTERACTIVE=1 NVIDIA_API_KEY=nvapi-...
bash e2e-cloud-experimental/test-port8080-conflict.sh so copying from the repo
root runs the correct script.

test/e2e/e2e-cloud-experimental/test-inference-local-chat.sh-20-21 (1)

20-21: ⚠️ Potential issue | 🟡 Minor

The usage example points at the wrong script name.

The file is test/e2e/e2e-cloud-experimental/test-inference-local-chat.sh, but the comment still tells users to run demo-inference-local-chat.sh.

Suggested fix

-#   bash test/e2e/demo-inference-local-chat.sh
+#   bash test/e2e/e2e-cloud-experimental/test-inference-local-chat.sh

🤖 Prompt for AI Agents

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

In `@test/e2e/e2e-cloud-experimental/test-inference-local-chat.sh` around lines 20
- 21, Update the usage comment that currently references
"demo-inference-local-chat.sh" to the correct script name
"test-inference-local-chat.sh" so the example reflects the actual file; locate
the top-of-file usage block containing the string "bash
demo-inference-local-chat.sh" and replace it with "bash
test-inference-local-chat.sh".

test/e2e-gateway-isolation.sh-37-48 (1)

37-48: ⚠️ Potential issue | 🟡 Minor

Don't reuse the default image tag for local runs.

When NEMOCLAW_TEST_IMAGE is unset, this still skips the build if nemoclaw-isolation-test already exists locally. A failed or interrupted previous run can then make this script validate an old image instead of the current checkout.

♻️ Suggested guard

-# Skip build if image already exists (e.g., loaded from CI artifact)
-if docker image inspect "$IMAGE" >/dev/null 2>&1; then
+# Only reuse a preloaded image when the caller explicitly provided one.
+if [ -n "${NEMOCLAW_TEST_IMAGE:-}" ] && docker image inspect "$IMAGE" >/dev/null 2>&1; then
   info "Using pre-built image: $IMAGE"
 else
   info "Building sandbox image..."

🤖 Prompt for AI Agents

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

In `@test/e2e-gateway-isolation.sh` around lines 37 - 48, The script currently
reuses a potentially stale local image when NEMOCLAW_TEST_IMAGE is unset because
IMAGE defaults to a fixed tag; change the logic so that when NEMOCLAW_TEST_IMAGE
is empty you generate a unique tag (e.g., include git commit SHA or a
timestamp/short UUID) and assign it to IMAGE before the docker image inspect
check, or alternatively skip the inspect check and always run the docker build
when NEMOCLAW_TEST_IMAGE is unset; update the code paths around the IMAGE
variable and the docker build block so the build is forced for local runs unless
an explicit NEMOCLAW_TEST_IMAGE is provided.

.github/workflows/docs-preview-deploy.yaml (1)

100-106: Optional: gate sticky-comment deletion to same-repo PRs.

This avoids unnecessary API calls on fork PR closures, since preview comments are only created for same-repo PRs.

Suggested diff

-      - name: Remove preview comment
-        if: steps.meta.outputs.action == 'closed'
+      - name: Remove preview comment
+        if: steps.meta.outputs.action == 'closed' && steps.meta.outputs.same-repo == 'true'
         uses: marocchino/sticky-pull-request-comment@v2
         with:
           header: pr-preview
           number: ${{ steps.meta.outputs.pr-number }}
           delete: true

🤖 Prompt for AI Agents

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

In @.github/workflows/docs-preview-deploy.yaml around lines 100 - 106, Update
the "Remove preview comment" step so it only runs for closed PRs that originate
from the same repository; change the step's if condition (currently if:
steps.meta.outputs.action == 'closed') to also verify the PR is not from a fork
by comparing the PR head repo to the workflow repo (for example: ensure
github.event.pull_request.head.repo.full_name == github.repository in the if
expression) so the marocchino/sticky-pull-request-comment deletion only executes
for same-repo PRs.

scripts/docs-to-skills.py (1)

1350-1373: Fragile path matching using string containment.

The check if ".agents/skills" in str(out_dir) can match unintended paths like .agents/skills-backup/, .agents/skills2/, or subdirectories like .agents/skills/foo/. Consider using a more precise check.

♻️ Proposed fix for precise path matching

             # Only create symlink if output is under .agents/skills
-            if ".agents/skills" in str(out_dir):
-                agents_skills = Path(out_dir)
+            agents_skills = Path(".agents/skills")
+            try:
+                # Check if out_dir is exactly .agents/skills or a child of it
+                out_dir.resolve().relative_to(agents_skills.resolve())
+                is_agents_skills_output = True
+            except ValueError:
+                is_agents_skills_output = out_dir.resolve() == agents_skills.resolve()
+            if is_agents_skills_output:

Alternatively, for an exact match only:

if out_dir.resolve() == Path(".agents/skills").resolve():

🤖 Prompt for AI Agents

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

In `@scripts/docs-to-skills.py` around lines 1350 - 1373, The current fragile
check uses string containment ("if '.agents/skills' in str(out_dir)") which can
match unintended paths; update the logic around args.output_dirs and out_dir to
perform a precise Path comparison instead (e.g., compare out_dir.resolve() to
Path(".agents/skills").resolve() or test
out_dir.samefile(Path(".agents/skills"))), locating the check in the loop that
iterates args.output_dirs and references claude_skills, agents_skills, and rel;
replace the substring match with the exact Path equality/samefile check so only
the intended .agents/skills directory triggers symlink creation.

docs/_ext/search_assets/modules/SearchEngine.js (1)

235-237: Consider logging skipped documents during index build.

At Line 235, swallowing per-document indexing failures without any trace makes data-quality regressions hard to detect.

Proposed refactor

-                } catch (_docError) {
-                    // Skip documents that fail to index
+                } catch (docError) {
+                    console.warn('Skipping document during index build', {
+                        docId: doc?.id,
+                        error: docError
+                    });
                 }

🤖 Prompt for AI Agents

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

In `@docs/_ext/search_assets/modules/SearchEngine.js` around lines 235 - 237, The
catch block currently swallows per-document indexing errors (catch (_docError))
so failures are invisible; update that catch to log the error and document
identifier/metadata (e.g., doc.id or the document object) using the module's
logger if available (this.logger or similar) or console.error, including a clear
message and _docError details, then continue to skip the document as before.

nemoclaw/src/onboard/config.ts (2)

72-85: Consider improving error handling in the fallback path.

The fallback logic in ensureConfigDir catches all errors silently (line 77) and the fallback creation (lines 79-80) can also throw without being caught. If both paths fail, the error from the fallback will propagate with no context about the original failure.

♻️ Suggested improvement

 function ensureConfigDir(): void {
   if (configDirCreated) return;
   if (!existsSync(configDir)) {
     try {
       mkdirSync(configDir, { recursive: true });
-    } catch {
+    } catch (primaryError) {
       configDir = join(tmpdir(), ".nemoclaw");
       if (!existsSync(configDir)) {
-        mkdirSync(configDir, { recursive: true });
+        try {
+          mkdirSync(configDir, { recursive: true });
+        } catch (fallbackError) {
+          throw new Error(
+            `Failed to create config directory at primary (${process.env.HOME}/.nemoclaw) and fallback (${configDir}): ${String(fallbackError)}`
+          );
+        }
       }
     }
   }
   configDirCreated = true;
 }

🤖 Prompt for AI Agents

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

In `@nemoclaw/src/onboard/config.ts` around lines 72 - 85, The ensureConfigDir
function currently swallows the original error in the first mkdirSync attempt
and does not guard the fallback mkdirSync, so failures lose context; update
ensureConfigDir to catch the initial error into a variable (e.g., err1) when the
first mkdirSync fails, then attempt the fallback (set configDir = join(tmpdir(),
".nemoclaw")) inside its own try/catch so you can capture a second error (e.g.,
err2) if that also fails, and finally throw a new Error that includes both err1
and err2 messages (or attach them as cause) to preserve full context; reference
ensureConfigDir, configDir, configDirCreated, mkdirSync, existsSync, tmpdir, and
join when making the changes.

---

8-8: Module-level mutable state may cause test isolation issues.

The let configDir is mutable and modified by ensureConfigDir(). Once the fallback path is used, subsequent calls in the same process will use the modified path. This could cause unexpected behavior in tests or long-running processes.

🤖 Prompt for AI Agents

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

In `@nemoclaw/src/onboard/config.ts` at line 8, The module-level mutable variable
configDir causes shared-state leaks; replace it with a pure accessor and stop
mutating module state: remove the let configDir and add a function
getConfigDir() that returns join(process.env.HOME ?? tmpdir(), ".nemoclaw"),
update ensureConfigDir() to call getConfigDir() (or accept an optional path
param) and remove any code that reassigns configDir, and update all other
references to use getConfigDir() so tests and long-running processes no longer
rely on mutated module state.

package.json (1)

14-14: Consider extracting the complex prepare script to a separate shell script.

The inline prepare script handles multiple conditional paths (git presence, prek binary, prek package). While functional, extracting this to a dedicated scripts/prepare.sh would improve readability and maintainability.

🤖 Prompt for AI Agents

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

In `@package.json` at line 14, The inline "prepare" entry in package.json is doing
multiple conditional checks and should be moved to a dedicated script for
clarity and maintainability: create a new executable scripts/prepare.sh that
contains the current conditional logic (checking for .git, testing for the prek
binary, and handling the presence of node_modules/@j178/prek), make it
executable, and then replace the package.json "prepare" value with a simple
invocation of that script (e.g., calling scripts/prepare.sh). Ensure the new
script preserves identical behavior and exit codes and reference the same
symbols (prek, node_modules/@j178/prek) used in the original prepare logic.

test/e2e/e2e-cloud-experimental/skip/04-nemoclaw-openshell-status-parity.sh (1)

52-65: Consider simplifying the Ready check condition.

Line 61 checks cols.includes("Ready") && !cols.includes("NotReady"). If columns include "Ready" as a status, the !cols.includes("NotReady") check may be unnecessary unless a row can contain both states simultaneously. This is minor and the current implementation is safe.

🤖 Prompt for AI Agents

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

In `@test/e2e/e2e-cloud-experimental/skip/04-nemoclaw-openshell-status-parity.sh`
around lines 52 - 65, The Ready-check inside the Node one-liner can be
simplified by removing the redundant "!cols.includes('NotReady')" check; update
the predicate that sets ok (the arrow function used in
clean.split(...).some(...)) to only verify cols.includes("Ready") (or a
word-boundary check like matching /\bReady\b/ against the trimmed line) so the
readiness determination uses a single clear condition; modify the lambda that
assigns ok accordingly.

test/credentials.test.js (1)

33-42: Prefer behavior assertions over source-regex assertions.

At Line 39–Line 41, this test is tightly coupled to code formatting/order and can fail on harmless refactors. Consider exercising the prompt error path and asserting observable behavior (rejection + SIGINT intent) instead of matching source text.

🤖 Prompt for AI Agents

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

In `@test/credentials.test.js` around lines 33 - 42, The test currently inspects
credentials.js source via regex (matching "return new Promise((resolve, reject)
=> {" and 'reject(err); process.kill(process.pid, "SIGINT");'), which is
fragile; instead, mock/stub the secret prompt used by the credentials.js module
so it throws an error, spy on process.kill, import and invoke the exported
function that triggers the secret prompt, then assert the returned promise
rejects and that process.kill was called with process.pid and "SIGINT" (replace
the source-regex assertions with these behavioral assertions).

scripts/check-spdx-headers.sh (2)

38-46: Consider adding .jsx to the // comment style pattern.

The pattern handles .tsx but omits .jsx, which should also use // style comments.

♻️ Suggested fix

   case "$base" in
     Dockerfile | *.dockerfile | *.Dockerfile) echo "#" ;;
-    *.ts | *.tsx | *.js | *.mjs | *.cjs) echo "//" ;;
+    *.ts | *.tsx | *.js | *.jsx | *.mjs | *.cjs) echo "//" ;;
     *) echo "#" ;;
   esac

🤖 Prompt for AI Agents

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

In `@scripts/check-spdx-headers.sh` around lines 38 - 46, The case in function
comment_style_for currently maps *.ts, *.tsx, *.js, *.mjs, *.cjs to "//" but
omits JSX files; update the pattern in comment_style_for to include *.jsx
alongside *.tsx so files like *.jsx also return "//". Locate the case arm inside
comment_style_for and add "*.jsx" to the list of alternatives that echo "//"
(keeping the existing order and spacing).

---

61-85: Minor: temp file not cleaned up on partial failure in insert_spdx.

If chmod or mv fails on Line 84, the temp file in $tmp is left behind. Consider adding cleanup logic or a trap within the function.

♻️ Suggested improvement

 insert_spdx() {
   local file=$1
   local style
   style=$(comment_style_for "$file")
   local tmp
   local mode
   tmp="$(mktemp "${TMPDIR:-/tmp}/nemoclaw-spdx.XXXXXX")"
+  trap 'rm -f "$tmp"' RETURN
   mode="$(stat -c '%a' "$file" 2>/dev/null || stat -f '%Lp' "$file")"

🤖 Prompt for AI Agents

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

In `@scripts/check-spdx-headers.sh` around lines 61 - 85, The insert_spdx function
can leave the temporary file ($tmp) behind if chmod or mv fails; update
insert_spdx to create a cleanup trap or explicit error-path removal that deletes
"$tmp" on any failure and disables the trap after successful mv, ensuring we
still preserve the original file mode (mode variable) and that the trap
references the same tmp variable; target the tmp creation/cleanup around the
mktemp, chmod and mv sequence inside insert_spdx and ensure the trap is cleaned
up on success.

test/validate-blueprint.test.ts (2)

22-28: Consider adding error handling for blueprint parsing.

Parsing blueprint.yaml at module scope means any parsing errors will crash the entire test suite rather than reporting as individual test failures. Consider wrapping the parsing in a try-catch or moving it inside a beforeAll hook.

♻️ Suggested improvement

-const bp = YAML.parse(readFileSync(BLUEPRINT_PATH, "utf-8")) as Record<string, unknown>;
-const declared = Array.isArray(bp?.profiles) ? (bp.profiles as string[]) : [];
-const defined =
-  (bp?.components as Record<string, unknown> | undefined)?.inference != null
-    ? ((bp.components as Record<string, unknown>).inference as Record<string, unknown>).profiles as
-        Record<string, Record<string, unknown>> | undefined
-    : undefined;
+let bp: Record<string, unknown>;
+let declared: string[];
+let defined: Record<string, Record<string, unknown>> | undefined;
+
+beforeAll(() => {
+  bp = YAML.parse(readFileSync(BLUEPRINT_PATH, "utf-8")) as Record<string, unknown>;
+  declared = Array.isArray(bp?.profiles) ? (bp.profiles as string[]) : [];
+  defined =
+    (bp?.components as Record<string, unknown> | undefined)?.inference != null
+      ? ((bp.components as Record<string, unknown>).inference as Record<string, unknown>).profiles as
+          Record<string, Record<string, unknown>> | undefined
+      : undefined;
+});

🤖 Prompt for AI Agents

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

In `@test/validate-blueprint.test.ts` around lines 22 - 28, Parsing blueprint.yaml
at module scope (YAML.parse(readFileSync(BLUEPRINT_PATH, "utf-8")) producing bp
and derived declared/defined) can crash the whole test run on parse errors; move
the file read and YAML.parse into a beforeAll hook or wrap them in a try-catch
inside beforeAll so you can surface parsing errors as test failures (use
BLUEPRINT_PATH, readFileSync, YAML.parse, bp, declared, defined) and call fail()
or throw a controlled error from the hook to report the failure without aborting
the test runner.

---

51-60: Clarify endpoint validation logic for dynamic_endpoint profiles.

When dynamic_endpoint === true, the test only verifies the endpoint key exists (Line 56) but doesn't check its value. When false, it requires a truthy value (Line 58). This asymmetry is intentional per the summary but could benefit from a brief comment explaining why dynamic endpoints allow empty/placeholder values.

📝 Suggested clarification

       for (const field of REQUIRED_PROFILE_FIELDS) {
         it(`has non-empty '${field}'`, () => {
           const cfg = defined?.[name];
           if (!cfg) return; // covered by "has a definition"
+          // dynamic_endpoint profiles may have placeholder endpoint values resolved at runtime
           if (field === "endpoint" && cfg.dynamic_endpoint === true) {
             expect(field in cfg).toBe(true);
           } else {
             expect(cfg[field]).toBeTruthy();
           }
         });
       }

🤖 Prompt for AI Agents

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

In `@test/validate-blueprint.test.ts` around lines 51 - 60, The test in the loop
over REQUIRED_PROFILE_FIELDS treats "endpoint" specially when
cfg.dynamic_endpoint === true by only asserting the key exists rather than
asserting a truthy value; add a short explanatory comment inside that branch
(near the conditional checking dynamic_endpoint and the expect(field in cfg)
line) clarifying that profiles with dynamic_endpoint may provide a
placeholder/empty endpoint at config time and are validated dynamically at
runtime, so the spec only requires the key to be present and not a truthy value;
reference REQUIRED_PROFILE_FIELDS, cfg.dynamic_endpoint, and the endpoint check
in the test to locate where to add the comment.

docs/reference/inference-profiles.md (2)

56-64: Colons used as general punctuation between clauses.

Per the style guide, colons should only introduce a list. Lines 56, 58, 60, and 62 use colons to separate a label from its explanation. Consider restructuring these items.

Suggested restructure

-- OpenAI-compatible providers:
-  NemoClaw tries `/responses` first, then `/chat/completions`.
+- **OpenAI-compatible providers** — NemoClaw tries `/responses` first, then `/chat/completions`.

Or use a definition list / nested structure where the provider type is a sub-heading.

🤖 Prompt for AI Agents

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

In `@docs/reference/inference-profiles.md` around lines 56 - 64, The documentation
uses colons incorrectly as general punctuation in the list items describing
provider behaviors; update the sentences for the
OpenAI/Anthropic/NVIDIA/Compatible endpoint lines so the provider labels are not
followed by colons (e.g., rephrase "OpenAI-compatible providers: NemoClaw tries
`/responses` first..." to "For OpenAI-compatible providers, NemoClaw tries
`/responses` first, then `/chat/completions`") or convert the list into a
definition/nested structure with provider types as sub-headings (refer to the
terms "OpenAI-compatible providers", "Anthropic-compatible providers", "NVIDIA
Endpoints manual model entry", and "Compatible endpoint flows") so each label
introduces the explanation without using a colon as general punctuation.

---

86-88: Missing "Next Steps" section.

Per page structure guidelines, documentation pages should include a "Next Steps" section at the bottom linking to related pages. The current ending abruptly references another document without providing additional navigation options.

🤖 Prompt for AI Agents

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

In `@docs/reference/inference-profiles.md` around lines 86 - 88, Add a "Next
Steps" section after the "Runtime Switching" paragraph that provides clear
navigation to related docs; include the existing "Switch Inference Models" link
(link text "Switch Inference Models") plus 1–2 other relevant pages such as
"Inference Providers" and "Model Selection" (or similar) to guide readers, using
the same relative-link style as the current reference to ensure consistent
formatting and placement at the bottom of the page.

test/onboard.test.js (1)

427-436: Consider using findLast for cleaner JSON payload extraction.

The .slice().reverse().find() pattern works but findLast (available in Node 18+) is more idiomatic.

Suggested simplification

-    const payloadLine = result.stdout
-      .trim()
-      .split("\n")
-      .slice()
-      .reverse()
-      .find((line) => line.startsWith("{") && line.endsWith("}"));
+    const payloadLine = result.stdout
+      .trim()
+      .split("\n")
+      .findLast((line) => line.startsWith("{") && line.endsWith("}"));

🤖 Prompt for AI Agents

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

In `@test/onboard.test.js` around lines 427 - 436, Replace the verbose
reverse-find pattern used to extract the JSON payload from stdout (the
payloadLine computation that uses
result.stdout.trim().split("\n").slice().reverse().find(...)) with the more
idiomatic Node 18+ Array.prototype.findLast call to locate the last line that
starts with "{" and ends with "}", keeping the subsequent
JSON.parse(payloadLine) and assertions (payload, payload.liveExists,
payload.sandbox) unchanged.

.agents/skills/nemoclaw-reference/references/inference-profiles.md (1)

51-53: Consider varying the list item phrasing.

Static analysis flagged three successive items beginning with "Local". While not incorrect, varying the structure improves readability.

Suggested alternative

-- Local Ollama
-- Local NVIDIA NIM
-- Local vLLM
+- Ollama (local)
+- NVIDIA NIM (local)
+- vLLM (local)

🤖 Prompt for AI Agents

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

In @.agents/skills/nemoclaw-reference/references/inference-profiles.md around
lines 51 - 53, The three list items ("Local Ollama", "Local NVIDIA NIM", "Local
vLLM") use the same leading word which hurts readability; change the phrasing to
vary structure (for example: "Ollama (local)", "NVIDIA NIM — local deployment",
"vLLM (local runtime)") or similar alternations so each item reads distinctively
while preserving meaning; update the items where these exact strings appear.