…IDIA#2221 tests

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

… confirms

Follow-up to NVIDIA#2221 addressing the remaining credential-persistence window
CodeRabbit flagged: setupNim's credential prompt writes to
~/.nemoclaw/credentials.json via saveCredential as soon as the user types
a key, so a later abort at the review gate still left that key on disk.
The abort message worked around it by directing users to
`nemoclaw credentials reset <KEY>`.

Introduce a module-level defer flag so credential-prompt writes are
held in memory (process.env only) and only flushed to disk after the
review gate confirms.

Mechanism

- `beginDeferredCredentialPersistence()` — arm the flag and clear the
  pending set. Called once at the top of the inference retry loop in
  onboard(), only when the review gate will actually fire
  (interactive + not --dangerously-skip-permissions).
- `replaceNamedCredential()` — when the flag is armed, set process.env
  and add the env name to `pendingDeferredCredentials`, skip the
  saveCredential call. Print "Key held in memory; will be saved to
  credentials.json on confirm." instead of the usual "Key saved".
- `flushDeferredCredentials()` — iterate the pending set, call
  saveCredential(envName, process.env[envName]) for each, clear the
  set. Flag stays armed so retry-loop iterations after an inference
  validation failure continue to defer their replacement keys.
- `discardDeferredCredentials()` — clear the set and disarm the flag.
  Called on gate abort — nothing ever reached disk.
- `endDeferredCredentialPersistence()` — clear + disarm. Called after
  the retry loop breaks successfully so downstream prompts (Brave,
  messaging tokens — already after the gate, already after flush) use
  the normal immediate-persist path.

Abort flow

The gate's "n/no" branch now calls discardDeferredCredentials() and
prints "No credentials were persisted to disk." — replacing the earlier
"credentials are stored in ~/.nemoclaw/credentials.json; clear them
with `nemoclaw credentials reset <KEY>`" workaround.

Non-interactive and --dangerously-skip-permissions flows skip the
deferral entirely (no gate to flush at) and use the existing
immediate-persist path.

Scope

Touches only onboard.ts. The other saveCredential call sites
(configureWebSearch, setupMessagingChannels, sandbox-channels,
sandbox-config) already run after the gate and don't need deferral.

Tests

- test/onboard.test.ts — 134 tests pass (no regressions in the
  non-deferred or resume paths).

The defer helpers themselves are module-scoped with no test seams that
don't require stdin simulation; their correctness is covered by the
existing onboard integration tests which exercise both the interactive
(deferral on) and non-interactive (deferral off) code paths.

Stacks on NVIDIA#2221 — depends on the gate and formatOnboardConfigSummary
landing first.

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

… confirms

Follow-up to NVIDIA#2221 addressing the remaining credential-persistence window
CodeRabbit flagged: setupNim's credential prompt writes to
~/.nemoclaw/credentials.json via saveCredential as soon as the user types
a key, so a later abort at the review gate still left that key on disk.

Introduce a module-level defer flag so credential-prompt writes are held
in memory until the gate confirms, plus follow-up fixes addressing
CodeRabbit review on b20e3b3.

Mechanism

- beginDeferredCredentialPersistence() arms the flag, clears the pending
  set, resets the flush-tracking bit. Called once at the top of the
  inference retry loop in onboard(), only when the review gate will
  actually fire (interactive + not --dangerously-skip-permissions).
- replaceNamedCredential(), when the flag is armed, sets process.env
  and adds the env name to pendingDeferredCredentials, skipping the
  saveCredential call. Prints "Key held in memory; will be saved to
  credentials.json on confirm." instead of the usual "Key saved".
- flushDeferredCredentials() iterates the pending set, calls
  saveCredential(envName, process.env[envName]) for each, clears the
  set, and marks deferredCredentialsFlushed so a subsequent abort after
  a confirmed-then-failed retry iteration tells the truth.
- discardDeferredCredentials() now also clears the pending env vars
  from process.env, not just the set. Called on gate abort.
- endDeferredCredentialPersistence() clears, disarms, resets the flush
  bit. Called after the retry loop breaks successfully so downstream
  prompts (Brave, messaging tokens) use the normal immediate path.
- discardStagedCredential(envName) removes a single envName from the
  pending set and process.env. Used when setupNim goes back to
  selectionLoop after a credential was staged for an abandoned provider.

setupNim changes (close CodeRabbit Majors at lines 761 and 6798)

- In deferred mode only, route the NVIDIA API key prompt through
  replaceNamedCredential (respects the flag) instead of ensureApiKey
  (which writes immediately). Non-deferred mode still calls ensureApiKey
  so existing mocks and single-path flows are unchanged. The onboarding
  box UI is preserved inline; validateNvidiaApiKeyValue runs the same
  starts-with-nvapi- check ensureApiKey did.
- Track stagedCredentialEnvThisIteration across selectionLoop iterations
  and discardStagedCredential() at the top of each new iteration so a
  user who enters key A for provider X, navigates back, then picks
  provider Y with key B only has B in pendingDeferredCredentials when
  the gate flushes.

formatOnboardConfigSummary changes (close CodeRabbit Minor at line 3378)

- New credentialStorageState parameter distinguishes "deferred" (key
  held in memory) from "persisted" (already on disk). The summary now
  says "held in memory — will be saved to ~/.nemoclaw/credentials.json
  on confirm" when the key is still pending so the gate's promise
  matches reality. The call site passes the state based on
  pendingDeferredCredentials membership.

Abort flow (closes CodeRabbit Minor at line 6883)

- The gate's n/no branch now checks haveDeferredCredentialsBeenFlushed()
  before printing its after-message. If no flush ever happened it says
  "No credentials were persisted to disk." as before. If an earlier
  retry iteration flushed and this iteration aborted, it acknowledges
  that "credentials confirmed in an earlier attempt are already in
  credentials.json" and points to nemoclaw credentials reset <KEY>.

Non-interactive and --dangerously-skip-permissions flows skip the
deferral entirely and use the existing immediate-persist path.

Scope

Touches only onboard.ts. The other saveCredential call sites
(configureWebSearch, setupMessagingChannels, sandbox-channels,
sandbox-config) already run after the gate and dont need deferral.

Tests

- test/onboard.test.ts — 135 tests pass.
- test/onboard-selection.test.ts — 33 tests pass (verifies setupNim
  back-nav + NVIDIA Endpoints path).
- Typecheck clean.

Stacks on NVIDIA#2221 — depends on the gate and formatOnboardConfigSummary
landing first.

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

… confirms

Follow-up to NVIDIA#2221 addressing the remaining credential-persistence window
CodeRabbit flagged: setupNim's credential prompt writes to
~/.nemoclaw/credentials.json via saveCredential as soon as the user types
a key, so a later abort at the review gate still left that key on disk.

Introduce a module-level defer flag so credential-prompt writes are held
in memory until the gate confirms, plus follow-up fixes addressing
CodeRabbit review on b20e3b3.

Mechanism

- beginDeferredCredentialPersistence() arms the flag, clears the pending
  set, resets the flush-tracking bit. Called once at the top of the
  inference retry loop in onboard(), only when the review gate will
  actually fire (interactive + not --dangerously-skip-permissions).
- replaceNamedCredential(), when the flag is armed, sets process.env
  and adds the env name to pendingDeferredCredentials, skipping the
  saveCredential call. Prints "Key held in memory; will be saved to
  credentials.json on confirm." instead of the usual "Key saved".
- flushDeferredCredentials() iterates the pending set, calls
  saveCredential(envName, process.env[envName]) for each, clears the
  set, and marks deferredCredentialsFlushed so a subsequent abort after
  a confirmed-then-failed retry iteration tells the truth.
- discardDeferredCredentials() now also clears the pending env vars
  from process.env, not just the set. Called on gate abort.
- endDeferredCredentialPersistence() clears, disarms, resets the flush
  bit. Called after the retry loop breaks successfully so downstream
  prompts (Brave, messaging tokens) use the normal immediate path.
- discardStagedCredential(envName) removes a single envName from the
  pending set and process.env. Used when setupNim goes back to
  selectionLoop after a credential was staged for an abandoned provider.

setupNim changes (close CodeRabbit Majors at lines 761 and 6798)

- In deferred mode only, route the NVIDIA API key prompt through
  replaceNamedCredential (respects the flag) instead of ensureApiKey
  (which writes immediately). Non-deferred mode still calls ensureApiKey
  so existing mocks and single-path flows are unchanged. The onboarding
  box UI is preserved inline; validateNvidiaApiKeyValue runs the same
  starts-with-nvapi- check ensureApiKey did.
- Track stagedCredentialEnvThisIteration across selectionLoop iterations
  and discardStagedCredential() at the top of each new iteration so a
  user who enters key A for provider X, navigates back, then picks
  provider Y with key B only has B in pendingDeferredCredentials when
  the gate flushes.

formatOnboardConfigSummary changes (close CodeRabbit Minor at line 3378)

- New credentialStorageState parameter distinguishes "deferred" (key
  held in memory) from "persisted" (already on disk). The summary now
  says "held in memory — will be saved to ~/.nemoclaw/credentials.json
  on confirm" when the key is still pending so the gate's promise
  matches reality. The call site passes the state based on
  pendingDeferredCredentials membership.

Abort flow (closes CodeRabbit Minor at line 6883)

- The gate's n/no branch now checks haveDeferredCredentialsBeenFlushed()
  before printing its after-message. If no flush ever happened it says
  "No credentials were persisted to disk." as before. If an earlier
  retry iteration flushed and this iteration aborted, it acknowledges
  that "credentials confirmed in an earlier attempt are already in
  credentials.json" and points to nemoclaw credentials reset <KEY>.

Non-interactive and --dangerously-skip-permissions flows skip the
deferral entirely and use the existing immediate-persist path.

Scope

Touches only onboard.ts. The other saveCredential call sites
(configureWebSearch, setupMessagingChannels, sandbox-channels,
sandbox-config) already run after the gate and dont need deferral.

Tests

- test/onboard.test.ts — 135 tests pass.
- test/onboard-selection.test.ts — 33 tests pass (verifies setupNim
  back-nav + NVIDIA Endpoints path).
- Typecheck clean.

Stacks on NVIDIA#2221 — depends on the gate and formatOnboardConfigSummary
landing first.

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

## Summary
Catches up the user-facing docs for changes merged since v0.0.24 and
bumps `docs/versions1.json` so the next docs build publishes as 0.0.25.

## Changes
- `docs/reference/troubleshooting.md`: corrected the JetPack 7 (L4T
39.x) entry to reflect the new conditional `br_netfilter` auto-load on
R39 hosts that lack it (from #2419), and added a new "DNS resolution
from inside docker fails (corporate firewall)" section covering the new
container-DNS preflight probe and platform-specific remediation paths
(from #2349).
- `docs/get-started/quickstart.md`: added a "Review the Configuration
Before the Sandbox Build" subsection documenting the new `[Y/n]` review
gate that fires after the sandbox-name prompt (from #2221).
- `docs/versions1.json`: promoted 0.0.25 to `preferred: true`, demoted
0.0.24, and trimmed to ten entries to match what
`scripts/bump-version.ts` writes. `docs/project.json` regenerates from
`versions1.json` at build time and now reports `0.0.25`.
- `.agents/skills/nemoclaw-user-get-started/` and
`.agents/skills/nemoclaw-user-reference/`: regenerated via
`scripts/docs-to-skills.py`.

## Type of Change
- [ ] Code change (feature, bug fix, or refactor)
- [ ] Code change with doc updates
- [x] Doc only (prose changes, no code sample modifications)
- [ ] Doc only (includes code sample changes)

## Verification
- [x] `npx prek run --all-files` passes
- [x] `npm test` passes
- [x] Tests added or updated for new or changed behavior
- [x] No secrets, API keys, or credentials committed
- [x] Docs updated for user-facing behavior changes
- [x] `make docs` builds without warnings (doc changes only)
- [x] Doc pages follow the [style
guide](https://github.com/NVIDIA/NemoClaw/blob/main/docs/CONTRIBUTING.md)
(doc changes only)
- [ ] New doc pages include SPDX header and frontmatter (new pages only)

## AI Disclosure
- [x] AI-assisted — tool: Claude Code

---
Signed-off-by: Miyoung Choi <miyoungc@nvidia.com>

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Onboarding now shows a configuration review and requires explicit
confirmation before building a sandbox; non-interactive mode prints the
summary without prompting.

* **Documentation**
* Expanded Jetson troubleshooting for JetPack 7 / L4T 39.x and
conditional br_netfilter guidance.
* Added corporate-firewall DNS troubleshooting with platform-specific
remediation and a verification nslookup step.
  * Quickstart updated to document the new confirmation flow.

* **Chores**
  * Documentation version index updated.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… confirms

Follow-up to NVIDIA#2221 addressing the remaining credential-persistence window
CodeRabbit flagged: setupNim's credential prompt writes to
~/.nemoclaw/credentials.json via saveCredential as soon as the user types
a key, so a later abort at the review gate still left that key on disk.

Introduce a module-level defer flag so credential-prompt writes are held
in memory until the gate confirms, plus follow-up fixes addressing
CodeRabbit review on b20e3b3.

Mechanism

- beginDeferredCredentialPersistence() arms the flag, clears the pending
  set, resets the flush-tracking bit. Called once at the top of the
  inference retry loop in onboard(), only when the review gate will
  actually fire (interactive + not --dangerously-skip-permissions).
- replaceNamedCredential(), when the flag is armed, sets process.env
  and adds the env name to pendingDeferredCredentials, skipping the
  saveCredential call. Prints "Key held in memory; will be saved to
  credentials.json on confirm." instead of the usual "Key saved".
- flushDeferredCredentials() iterates the pending set, calls
  saveCredential(envName, process.env[envName]) for each, clears the
  set, and marks deferredCredentialsFlushed so a subsequent abort after
  a confirmed-then-failed retry iteration tells the truth.
- discardDeferredCredentials() now also clears the pending env vars
  from process.env, not just the set. Called on gate abort.
- endDeferredCredentialPersistence() clears, disarms, resets the flush
  bit. Called after the retry loop breaks successfully so downstream
  prompts (Brave, messaging tokens) use the normal immediate path.
- discardStagedCredential(envName) removes a single envName from the
  pending set and process.env. Used when setupNim goes back to
  selectionLoop after a credential was staged for an abandoned provider.

setupNim changes (close CodeRabbit Majors at lines 761 and 6798)

- In deferred mode only, route the NVIDIA API key prompt through
  replaceNamedCredential (respects the flag) instead of ensureApiKey
  (which writes immediately). Non-deferred mode still calls ensureApiKey
  so existing mocks and single-path flows are unchanged. The onboarding
  box UI is preserved inline; validateNvidiaApiKeyValue runs the same
  starts-with-nvapi- check ensureApiKey did.
- Track stagedCredentialEnvThisIteration across selectionLoop iterations
  and discardStagedCredential() at the top of each new iteration so a
  user who enters key A for provider X, navigates back, then picks
  provider Y with key B only has B in pendingDeferredCredentials when
  the gate flushes.

formatOnboardConfigSummary changes (close CodeRabbit Minor at line 3378)

- New credentialStorageState parameter distinguishes "deferred" (key
  held in memory) from "persisted" (already on disk). The summary now
  says "held in memory — will be saved to ~/.nemoclaw/credentials.json
  on confirm" when the key is still pending so the gate's promise
  matches reality. The call site passes the state based on
  pendingDeferredCredentials membership.

Abort flow (closes CodeRabbit Minor at line 6883)

- The gate's n/no branch now checks haveDeferredCredentialsBeenFlushed()
  before printing its after-message. If no flush ever happened it says
  "No credentials were persisted to disk." as before. If an earlier
  retry iteration flushed and this iteration aborted, it acknowledges
  that "credentials confirmed in an earlier attempt are already in
  credentials.json" and points to nemoclaw credentials reset <KEY>.

Non-interactive and --dangerously-skip-permissions flows skip the
deferral entirely and use the existing immediate-persist path.

Scope

Touches only onboard.ts. The other saveCredential call sites
(configureWebSearch, setupMessagingChannels, sandbox-channels,
sandbox-config) already run after the gate and dont need deferral.

Tests

- test/onboard.test.ts — 135 tests pass.
- test/onboard-selection.test.ts — 33 tests pass (verifies setupNim
  back-nav + NVIDIA Endpoints path).
- Typecheck clean.

Stacks on NVIDIA#2221 — depends on the gate and formatOnboardConfigSummary
landing first.

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

… confirms

Follow-up to NVIDIA#2221 addressing the remaining credential-persistence window
CodeRabbit flagged: setupNim's credential prompt writes to
~/.nemoclaw/credentials.json via saveCredential as soon as the user types
a key, so a later abort at the review gate still left that key on disk.

Introduce a module-level defer flag so credential-prompt writes are held
in memory until the gate confirms, plus follow-up fixes addressing
CodeRabbit review on b20e3b3.

Mechanism

- beginDeferredCredentialPersistence() arms the flag, clears the pending
  set, resets the flush-tracking bit. Called once at the top of the
  inference retry loop in onboard(), only when the review gate will
  actually fire (interactive + not --dangerously-skip-permissions).
- replaceNamedCredential(), when the flag is armed, sets process.env
  and adds the env name to pendingDeferredCredentials, skipping the
  saveCredential call. Prints "Key held in memory; will be saved to
  credentials.json on confirm." instead of the usual "Key saved".
- flushDeferredCredentials() iterates the pending set, calls
  saveCredential(envName, process.env[envName]) for each, clears the
  set, and marks deferredCredentialsFlushed so a subsequent abort after
  a confirmed-then-failed retry iteration tells the truth.
- discardDeferredCredentials() now also clears the pending env vars
  from process.env, not just the set. Called on gate abort.
- endDeferredCredentialPersistence() clears, disarms, resets the flush
  bit. Called after the retry loop breaks successfully so downstream
  prompts (Brave, messaging tokens) use the normal immediate path.
- discardStagedCredential(envName) removes a single envName from the
  pending set and process.env. Used when setupNim goes back to
  selectionLoop after a credential was staged for an abandoned provider.

setupNim changes (close CodeRabbit Majors at lines 761 and 6798)

- In deferred mode only, route the NVIDIA API key prompt through
  replaceNamedCredential (respects the flag) instead of ensureApiKey
  (which writes immediately). Non-deferred mode still calls ensureApiKey
  so existing mocks and single-path flows are unchanged. The onboarding
  box UI is preserved inline; validateNvidiaApiKeyValue runs the same
  starts-with-nvapi- check ensureApiKey did.
- Track stagedCredentialEnvThisIteration across selectionLoop iterations
  and discardStagedCredential() at the top of each new iteration so a
  user who enters key A for provider X, navigates back, then picks
  provider Y with key B only has B in pendingDeferredCredentials when
  the gate flushes.

formatOnboardConfigSummary changes (close CodeRabbit Minor at line 3378)

- New credentialStorageState parameter distinguishes "deferred" (key
  held in memory) from "persisted" (already on disk). The summary now
  says "held in memory — will be saved to ~/.nemoclaw/credentials.json
  on confirm" when the key is still pending so the gate's promise
  matches reality. The call site passes the state based on
  pendingDeferredCredentials membership.

Abort flow (closes CodeRabbit Minor at line 6883)

- The gate's n/no branch now checks haveDeferredCredentialsBeenFlushed()
  before printing its after-message. If no flush ever happened it says
  "No credentials were persisted to disk." as before. If an earlier
  retry iteration flushed and this iteration aborted, it acknowledges
  that "credentials confirmed in an earlier attempt are already in
  credentials.json" and points to nemoclaw credentials reset <KEY>.

Non-interactive and --dangerously-skip-permissions flows skip the
deferral entirely and use the existing immediate-persist path.

Scope

Touches only onboard.ts. The other saveCredential call sites
(configureWebSearch, setupMessagingChannels, sandbox-channels,
sandbox-config) already run after the gate and dont need deferral.

Tests

- test/onboard.test.ts — 135 tests pass.
- test/onboard-selection.test.ts — 33 tests pass (verifies setupNim
  back-nav + NVIDIA Endpoints path).
- Typecheck clean.

Stacks on NVIDIA#2221 — depends on the gate and formatOnboardConfigSummary
landing first.

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

… confirms

Follow-up to NVIDIA#2221 addressing the remaining credential-persistence window
CodeRabbit flagged: setupNim's credential prompt writes to
~/.nemoclaw/credentials.json via saveCredential as soon as the user types
a key, so a later abort at the review gate still left that key on disk.

Introduce a module-level defer flag so credential-prompt writes are held
in memory until the gate confirms, plus follow-up fixes addressing
CodeRabbit review on b20e3b3.

Mechanism

- beginDeferredCredentialPersistence() arms the flag, clears the pending
  set, resets the flush-tracking bit. Called once at the top of the
  inference retry loop in onboard(), only when the review gate will
  actually fire (interactive + not --dangerously-skip-permissions).
- replaceNamedCredential(), when the flag is armed, sets process.env
  and adds the env name to pendingDeferredCredentials, skipping the
  saveCredential call. Prints "Key held in memory; will be saved to
  credentials.json on confirm." instead of the usual "Key saved".
- flushDeferredCredentials() iterates the pending set, calls
  saveCredential(envName, process.env[envName]) for each, clears the
  set, and marks deferredCredentialsFlushed so a subsequent abort after
  a confirmed-then-failed retry iteration tells the truth.
- discardDeferredCredentials() now also clears the pending env vars
  from process.env, not just the set. Called on gate abort.
- endDeferredCredentialPersistence() clears, disarms, resets the flush
  bit. Called after the retry loop breaks successfully so downstream
  prompts (Brave, messaging tokens) use the normal immediate path.
- discardStagedCredential(envName) removes a single envName from the
  pending set and process.env. Used when setupNim goes back to
  selectionLoop after a credential was staged for an abandoned provider.

setupNim changes (close CodeRabbit Majors at lines 761 and 6798)

- In deferred mode only, route the NVIDIA API key prompt through
  replaceNamedCredential (respects the flag) instead of ensureApiKey
  (which writes immediately). Non-deferred mode still calls ensureApiKey
  so existing mocks and single-path flows are unchanged. The onboarding
  box UI is preserved inline; validateNvidiaApiKeyValue runs the same
  starts-with-nvapi- check ensureApiKey did.
- Track stagedCredentialEnvThisIteration across selectionLoop iterations
  and discardStagedCredential() at the top of each new iteration so a
  user who enters key A for provider X, navigates back, then picks
  provider Y with key B only has B in pendingDeferredCredentials when
  the gate flushes.

formatOnboardConfigSummary changes (close CodeRabbit Minor at line 3378)

- New credentialStorageState parameter distinguishes "deferred" (key
  held in memory) from "persisted" (already on disk). The summary now
  says "held in memory — will be saved to ~/.nemoclaw/credentials.json
  on confirm" when the key is still pending so the gate's promise
  matches reality. The call site passes the state based on
  pendingDeferredCredentials membership.

Abort flow (closes CodeRabbit Minor at line 6883)

- The gate's n/no branch now checks haveDeferredCredentialsBeenFlushed()
  before printing its after-message. If no flush ever happened it says
  "No credentials were persisted to disk." as before. If an earlier
  retry iteration flushed and this iteration aborted, it acknowledges
  that "credentials confirmed in an earlier attempt are already in
  credentials.json" and points to nemoclaw credentials reset <KEY>.

Non-interactive and --dangerously-skip-permissions flows skip the
deferral entirely and use the existing immediate-persist path.

Scope

Touches only onboard.ts. The other saveCredential call sites
(configureWebSearch, setupMessagingChannels, sandbox-channels,
sandbox-config) already run after the gate and dont need deferral.

Tests

- test/onboard.test.ts — 135 tests pass.
- test/onboard-selection.test.ts — 33 tests pass (verifies setupNim
  back-nav + NVIDIA Endpoints path).
- Typecheck clean.

Stacks on NVIDIA#2221 — depends on the gate and formatOnboardConfigSummary
landing first.

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

…ox build (NVIDIA#2221)

## Summary

`nemoclaw onboard` collected 8 inputs then immediately began a ~6-minute
sandbox image build + provider/credential persistence with **no final
confirmation**. Users who hit Enter by reflex at the sandbox-name prompt
(where `[my-assistant]` is pre-filled) were locked into a build they did
not intend to start.

This PR adds a compact summary + `[Y/n]` gate right after the
sandbox-name prompt in `createSandbox()`, before any destructive action
(docker build, `patchStagedDockerfile`, provider creation).

### Before

```
  [6/8] Creating sandbox
  ──────────────────────────────────────────────
  Sandbox name (lowercase, starts with letter, hyphens ok) [my-assistant]:
  ← Enter pressed →
  [7/8] Building sandbox image...   ← 6 minutes locked in
```

### After

```
  [6/8] Creating sandbox
  ──────────────────────────────────────────────
  Sandbox name (lowercase, starts with letter, hyphens ok) [my-assistant]:
  ← Enter pressed →

  ──────────────────────────────────────────────────
  Review configuration
  ──────────────────────────────────────────────────
  Provider:      gemini-api
  Model:         gemini-2.5-flash
  Web search:    disabled
  Messaging:     none
  Sandbox name:  my-assistant
  ──────────────────────────────────────────────────
  Apply this configuration? [Y/n]:
```

Default is `Y` so a confident user still just hits Enter once. Answering
`n`/`no` prints `"Aborted. Re-run nemoclaw onboard to start over."` and
exits `0` (clean abort, not an error).

Non-interactive runs (`NEMOCLAW_NON_INTERACTIVE=1`, or
`--dangerously-skip-permissions`) still print the summary for log
clarity but skip the gate — automated flows opted out of prompts.

## Related Issue

Closes NVIDIA#2165 (priority:high, `Getting Started` label)

## Changes

- **`src/lib/onboard.ts`** — added `formatOnboardConfigSummary()` (pure,
exported for testing) and an inline confirm prompt in `createSandbox()`
right after the sandbox-name validation. Non-interactive /
dangerouslySkipPermissions skip the prompt but still emit the summary.

## Testing

- [x] `npx vitest run test/onboard.test.ts` — 134 tests pass (+1 for new
summary test)
- [x] `npx vitest run test/runner.test.ts` — 60/60 pass (exercises
create-sandbox e2e)
- [x] `npm run build:cli` + `npm run typecheck:cli` clean

Executed:
- New test covers: full-field render
(provider/model/web/messaging/sandbox name), empty channels → `"none"`,
null webSearch → `"disabled"`, and null-field → `"(unset)"` fallback
(guards against stringified `"undefined"` leaking).

## Checklist

- [x] Follows [Conventional
Commits](https://www.conventionalcommits.org/)
- [x] Commit is signed (SSH)
- [x] DCO Signed-off-by trailer present

Signed-off-by: latenighthackathon
<latenighthackathon@users.noreply.github.com>

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Added a configuration review summary showing provider, model, API
credential hint (or "not required"), web search status, messaging
channels (or "none"), sandbox name (prompted if missing), and optional
notes.
* Interactive confirmation prompt lets users approve or abort applying
the configuration; abort prints guidance and exits onboarding
(non-interactive runs skip the prompt but still display the summary).

* **Tests**
* Added tests validating the configuration summary formatting and
placeholder rendering.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: latenighthackathon <latenighthackathon@users.noreply.github.com>
Co-authored-by: latenighthackathon <latenighthackathon@users.noreply.github.com>

## Summary
Catches up the user-facing docs for changes merged since v0.0.24 and
bumps `docs/versions1.json` so the next docs build publishes as 0.0.25.

## Changes
- `docs/reference/troubleshooting.md`: corrected the JetPack 7 (L4T
39.x) entry to reflect the new conditional `br_netfilter` auto-load on
R39 hosts that lack it (from NVIDIA#2419), and added a new "DNS resolution
from inside docker fails (corporate firewall)" section covering the new
container-DNS preflight probe and platform-specific remediation paths
(from NVIDIA#2349).
- `docs/get-started/quickstart.md`: added a "Review the Configuration
Before the Sandbox Build" subsection documenting the new `[Y/n]` review
gate that fires after the sandbox-name prompt (from NVIDIA#2221).
- `docs/versions1.json`: promoted 0.0.25 to `preferred: true`, demoted
0.0.24, and trimmed to ten entries to match what
`scripts/bump-version.ts` writes. `docs/project.json` regenerates from
`versions1.json` at build time and now reports `0.0.25`.
- `.agents/skills/nemoclaw-user-get-started/` and
`.agents/skills/nemoclaw-user-reference/`: regenerated via
`scripts/docs-to-skills.py`.

## Type of Change
- [ ] Code change (feature, bug fix, or refactor)
- [ ] Code change with doc updates
- [x] Doc only (prose changes, no code sample modifications)
- [ ] Doc only (includes code sample changes)

## Verification
- [x] `npx prek run --all-files` passes
- [x] `npm test` passes
- [x] Tests added or updated for new or changed behavior
- [x] No secrets, API keys, or credentials committed
- [x] Docs updated for user-facing behavior changes
- [x] `make docs` builds without warnings (doc changes only)
- [x] Doc pages follow the [style
guide](https://github.com/NVIDIA/NemoClaw/blob/main/docs/CONTRIBUTING.md)
(doc changes only)
- [ ] New doc pages include SPDX header and frontmatter (new pages only)

## AI Disclosure
- [x] AI-assisted — tool: Claude Code

---
Signed-off-by: Miyoung Choi <miyoungc@nvidia.com>

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Onboarding now shows a configuration review and requires explicit
confirmation before building a sandbox; non-interactive mode prints the
summary without prompting.

* **Documentation**
* Expanded Jetson troubleshooting for JetPack 7 / L4T 39.x and
conditional br_netfilter guidance.
* Added corporate-firewall DNS troubleshooting with platform-specific
remediation and a verification nslookup step.
  * Quickstart updated to document the new confirmation flow.

* **Chores**
  * Documentation version index updated.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

## Summary
Catches up the user-facing docs for changes merged since v0.0.24 and
bumps `docs/versions1.json` so the next docs build publishes as 0.0.25.

## Changes
- `docs/reference/troubleshooting.md`: corrected the JetPack 7 (L4T
39.x) entry to reflect the new conditional `br_netfilter` auto-load on
R39 hosts that lack it (from NVIDIA#2419), and added a new "DNS resolution
from inside docker fails (corporate firewall)" section covering the new
container-DNS preflight probe and platform-specific remediation paths
(from NVIDIA#2349).
- `docs/get-started/quickstart.md`: added a "Review the Configuration
Before the Sandbox Build" subsection documenting the new `[Y/n]` review
gate that fires after the sandbox-name prompt (from NVIDIA#2221).
- `docs/versions1.json`: promoted 0.0.25 to `preferred: true`, demoted
0.0.24, and trimmed to ten entries to match what
`scripts/bump-version.ts` writes. `docs/project.json` regenerates from
`versions1.json` at build time and now reports `0.0.25`.
- `.agents/skills/nemoclaw-user-get-started/` and
`.agents/skills/nemoclaw-user-reference/`: regenerated via
`scripts/docs-to-skills.py`.

## Type of Change
- [ ] Code change (feature, bug fix, or refactor)
- [ ] Code change with doc updates
- [x] Doc only (prose changes, no code sample modifications)
- [ ] Doc only (includes code sample changes)

## Verification
- [x] `npx prek run --all-files` passes
- [x] `npm test` passes
- [x] Tests added or updated for new or changed behavior
- [x] No secrets, API keys, or credentials committed
- [x] Docs updated for user-facing behavior changes
- [x] `make docs` builds without warnings (doc changes only)
- [x] Doc pages follow the [style
guide](https://github.com/NVIDIA/NemoClaw/blob/main/docs/CONTRIBUTING.md)
(doc changes only)
- [ ] New doc pages include SPDX header and frontmatter (new pages only)

## AI Disclosure
- [x] AI-assisted — tool: Claude Code

---
Signed-off-by: Miyoung Choi <miyoungc@nvidia.com>

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Onboarding now shows a configuration review and requires explicit
confirmation before building a sandbox; non-interactive mode prints the
summary without prompting.

* **Documentation**
* Expanded Jetson troubleshooting for JetPack 7 / L4T 39.x and
conditional br_netfilter guidance.
* Added corporate-firewall DNS troubleshooting with platform-specific
remediation and a verification nslookup step.
  * Quickstart updated to document the new confirmation flow.

* **Chores**
  * Documentation version index updated.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>