nWave CLI Reference
The nwave-ai command-line tool installs, configures, and manages the nWave framework in supported agentic hosts. For hands-on setup, see the Installation Guide. For per-project activation, see the Activation Guide.
Invocation
Commands are invoked as:
nwave-ai <command> [options] [arguments]
Alternatively, use Python's module syntax:
python -m nwave_ai.cli <command> [options] [arguments]
Synopsis
Observed --help behaviour: nwave-ai status --help exits 0 but prints the status report instead of usage; nwave-ai install --help and nwave-ai uninstall --help exit 0 and print their own usage (the uninstall text names python uninstall_nwave.py); nwave-ai rigor --help exits 2 and prints its usage line after WHAT/WHY/HOW lines; nwave-ai attribution --help and nwave-ai mode --help exit 1.
nwave-ai install [--platform <tool>] [--target <path>] [--yes]
[--dry-run] [--backup-only] [--restore]
nwave-ai uninstall [--target <path>]
nwave-ai doctor [--json] [--fix]
nwave-ai attribution <on|off|status>
nwave-ai model set --provider {codex,claude} --model MODEL [--role ROLE] [--project]
nwave-ai rigor [--show] [--yes] [--models keep | --model-set GROUP_OR_ROLE=PROVIDER/MODEL ...] [--review-rounds 1-5|current] [--examine-level quick|standard|extended|current] [--documents yes|no|keep [--document-language LANG]] [--hosts HOST[,HOST...]|none|keep]
nwave-ai project <enable|disable>
nwave-ai mode <all|opt-in>
nwave-ai status
nwave-ai completion <bash|zsh>
nwave-ai version
nwave-ai --version
nwave-ai outcomes [--registry PATH] register | check
nwave-ai plugin <list|install|uninstall> [name]
nwave-ai tests run [--root DIR] [--declaration FILE] [--class NAME] [--timeout SECONDS]
[--receipt FILE] (NAME [ARG ...] | -- COMMAND [ARG ...])
nwave-ai tests select --use NAME [--root DIR] [--declaration FILE] [--add TEST]...
[--json] [PATH ...]
nwave-ai install
Install nWave into the selected host's configuration directories. --target selects a custom Claude configuration directory; native OMP installation uses its own discovery roots. Installation does not ask for or write documentation-density preferences.
Synopsis
nwave-ai install [--platform <tool>] [--target <path>] [--yes] \
[--dry-run] [--backup-only] [--restore]
Flags
| Flag | Required | Type | Description |
|---|---|---|---|
--platform |
no | enum | Target agentic tool: auto, claude-code, codex, cursor, opencode, copilot, omp, devin, or all; auto infers available hosts and refuses when none are detected. Forwarded to the installer subprocess. |
--target |
no | path | Install into ~/.claude/. Must not be $HOME (rejected with exit 2). Sets CLAUDE_CONFIG_DIR for subprocess. Default: ~/.claude/. |
--yes |
no | bool | Non-interactive mode. Recommended for CI. |
--density-only |
no | bool | Run only the first-run documentation density prompt and exit (test-driving fixture; not a user flag). Exit 0 on completion. |
--dry-run |
no | bool | Preview changes without modifying the filesystem. Passed through to install script. |
--backup-only |
no | bool | Create a backup only without installing. Passed through to install script. |
--restore |
no | bool | Restore nWave from a previous backup. Passed through to install script. |
Pass-through behavior: All flags except --yes and --target are forwarded to scripts/install/install_nwave.py. --target sets CLAUDE_CONFIG_DIR for the subprocess; --yes selects non-interactive mode.
For --platform omp, public skills and native task agents use OMP's separate discovery roots. PI_CONFIG_DIR selects the configuration-root name under the user home; PI_CODING_AGENT_DIR selects the native skill/data root when set. Named OMP_PROFILE/PI_PROFILE installations are refused rather than silently writing to the default profile. The installer reports capabilities that OMP cannot project, including Claude-specific hooks and unsupported scoped tool permissions. User-owned skills and agents are not nWave-managed assets.
Exit codes
| Code | Condition |
|---|---|
| 0 | Installation succeeded. Package manager recorded in global config if needed. |
| 1 | Installation failed (subprocess error). |
| 2 | Configuration error: --target points to $HOME, or --target argument missing. |
Output
Installation reports its actual installer result; it does not report a documentation-density preference being written. stderr on error:
nwave-ai: --target must point to a config directory (e.g. ~/.claude-nwave or ./.claude), not your home directory
Example
# First-time install, without a density prompt
nwave-ai install
# CI environment: no user interaction
nwave-ai install --yes
# Install to a custom Claude config directory
nwave-ai install --target ~/.claude-nwave
# Dry-run to see what would be changed
nwave-ai install --dry-run
# Restore from backup after a failed upgrade
nwave-ai install --restore
nwave-ai uninstall
Remove the nWave framework from configured tool directories. Project files
(.nwave/, source code) are not touched; only the framework itself is
removed.
Synopsis
nwave-ai uninstall [--target <path>]
Flags
| Flag | Required | Type | Description |
|---|---|---|---|
--target |
no | path | Uninstall from $HOME (rejected with exit 2). Sets CLAUDE_CONFIG_DIR for subprocess. |
Exit codes
| Code | Condition |
|---|---|
| 0 | Uninstallation succeeded. |
| 1 | Uninstallation failed (subprocess error or I/O failure). |
| 2 | Configuration error: --target points to $HOME. |
Output
stdout on success (subprocess-dependent, typically minimal).
stderr on error (subprocess-dependent).
Example
nwave-ai uninstall
nwave-ai uninstall --target ~/.claude-nwave
nwave-ai doctor
Run diagnostics on the nWave installation. Checks for missing files, broken configuration, and hook integrity. Use --json for machine-readable output or --fix to attempt repairs (currently not yet implemented).
Synopsis
nwave-ai doctor [--json] [--fix] [--help]
Flags
| Flag | Required | Type | Description |
|---|---|---|---|
--json |
no | bool | Emit JSON output instead of human-readable text (suitable for scripting). |
--fix |
no | bool | Attempt to fix detected issues. Currently not implemented; exits 2. |
--help |
no | bool | Print help message and exit (exit 0). |
Exit codes
| Code | Condition |
|---|---|
| 0 | All diagnostic checks passed. |
| 1 | One or more checks failed (unhealthy installation). |
| 2 | Misconfiguration or unimplemented flag (--fix without support). |
Output
stdout on check pass (human):
✓ python_version: OK
✓ framework_files: OK
✓ hooks_registered: OK
✓ hook_python_path: OK
✓ config_ssot: OK
✓ attribution: OK
✓ version_sync: OK
The report lists one line per installed check; a failing check shows its
diagnostic and a remediation line instead of OK.
stdout on check failure (human):
✗ config_ssot: the configuration file could not be read or parsed
stdout on --json:
{
"checks": [
{
"name": "config_ssot",
"passed": true,
"message": "OK"
}
],
"summary": "All checks passed"
}
Example
nwave-ai doctor
nwave-ai doctor --json | jq '.summary'
nwave-ai doctor --fix
# → --fix not yet implemented. Run `nwave-ai install` to restore a broken installation.
# → (exit 2)
nwave-ai model
Select the provider and model for nWave roles. Records an explicit
provider/model pair in model_runtime (model_runtime.default, or
model_runtime.roles.<role> with --role) instead of deriving a model from
a competence label.
Synopsis
nwave-ai model set --provider {codex,claude} --model MODEL [--role ROLE [--competence COMPETENCE]] [--project]
Without --project the global default is updated; with --project the
project's .nwave/config.json is written instead. With --role, that role
is updated instead of the tier default. The supported providers are codex
and claude; whether your host consumes model_runtime depends on that
host's model-configuration support (see
Assigning models to roles).
nwave-ai rigor writes its per-role entries into the same model_runtime.roles store that model set --role writes, and it reads both the project and the global file. A per-role entry from model set --role (without --competence) appears in nwave-ai rigor --show. A default from model set without --role does not: --show lists per-role entries only. model set keeps its meaning; rigor does not replace it.
nwave-ai rigor
Choose this project's quality scope: the model for each role, the review-round limit, the examine level, and whether readable documents for people are produced. Run it from the repository root. It is not the nw-rigor skill, which keeps using nwave-ai model set. On a terminal the command asks each missing question one at a time, showing the current value and its source. A script answers every question with an option. It writes nothing until the final summary is confirmed, or until --yes is given.
Synopsis
nwave-ai rigor [--show] [--yes] [--models keep | --model-set GROUP_OR_ROLE=PROVIDER/MODEL ...] [--review-rounds 1-5|current] [--examine-level quick|standard|extended|current] [--documents yes|no|keep [--document-language LANG]] [--hosts HOST[,HOST...]|none|keep]
Answers
--models, --review-rounds, --documents and --hosts each need one answer. On a terminal the command asks each missing one, one question at a time: Enter keeps the value in effect, and a wrong answer asks the same question again. With --yes it never asks: a missing answer refuses with exit 2 and names every missing option. Without --yes and without a terminal (stdin at end of input, for example </dev/null or an empty pipe), it refuses with exit 2 and names every question that has no answer, for example no answer arrived for 5 questions: --models, --review-rounds, --examine-level, --documents, --hosts. The terminal closed. The examine level is listed there because it is asked as well, although --examine-level is optional. A fully closed stdin (0<&-) was not measured after this change. Scripts must therefore pass --yes and all four answers, for example --yes --models keep --review-rounds current --documents keep --hosts keep. On a terminal the models question takes one target=provider/model pair; for several, use the options.
Flags
| Flag | Required | Type | Description |
|---|---|---|---|
--show |
no | bool | Print the effective value of every role model, the review rounds, the examine level and the documents choice, each with its source. Writes nothing. Other options are not applied when --show is given. |
--yes |
no | bool | Confirm the final summary without asking. Every answer must then be given as an option. A host named by --hosts is written under this confirmation too. |
--models |
answer | keep |
Leave every role model unchanged. Only keep is accepted here; set models with --model-set. |
--model-set |
answer | GROUP_OR_ROLE=PROVIDER/MODEL |
Repeatable; several values may follow one option. GROUP_OR_ROLE is producers (roles whose tools include Write or Edit), reviewers (names ending -reviewer), executors (names ending -crafter), or one published role name. Groups overlap. Values apply in the order given, so give a group first and a single-role override after it. The provider is the text before the first / and the model is the rest; both are stored as given, with no availability check and no fallback. |
--review-rounds |
answer | 1-5 or current |
The review limit per finding: a whole number from 1 to 5. current keeps the value in effect. |
--examine-level |
no | quick, standard, extended or current |
The examine level: how many examinations a delivery admits. quick admits 1 examination and no repair round; standard (the built-in default) admits 2 examinations and one repair round; extended admits 3 examinations and two repair rounds. A level limits only the examinations, as the review rounds limit reviews. current keeps the value in effect, and an omitted option means current. A terminal session asks this question after the review rounds; Enter keeps the level in effect, and a wrong answer asks again. |
--documents |
answer | yes, no or keep |
Whether to produce a readable HTML brief for people at the charter and design closes. keep leaves the setting unchanged. |
--document-language |
no | text | Language of the readable brief. Default en, stored as given. With --yes, --documents yes also needs this option. |
--hosts |
answer | HOST[,HOST...], none or keep |
Host configuration files to write the model facts into. Host names: omp, claude-code, codex, copilot, opencode, devin. none and keep write no host file. See Host files. |
Exit codes
| Code | Condition |
|---|---|
| 0 | The project file was written, or --show printed. A host that cannot take per-role models prints its refusal; the project file is still written and the command still exits 0. |
| 1 | The project file could not be written (not writable, or a config migration conflict). Nothing was written, host files included. |
| 2 | Refusal, with WHAT, WHY and HOW lines on stderr: an unknown or repeated option; a role name that is not published; review rounds outside 1 to 5; an answer missing with no terminal to ask; an unconfirmed summary; an unreadable project or global config file; a model fact that matches no published role; an unknown host name; --documents yes with --yes and no --document-language. Nothing was written. |
Output
--show prints one line per published role, then the review limit and the documents choice:
role nw-acceptance-designer: claude-opus-5 (built-in default)
role nw-acceptance-designer-reviewer: claude/opus-4-6 (project)
review rounds: 3 (project)
examine level: standard (built-in default)
gives: 2 examinations, 1 repair round
not covered at this level: more than 1 repair round
not covered at any level: security, load or checks on other machines
a row not examined is never green
documents: yes, language en (project)
The built-in defaults are 2 review rounds, examine level standard and documents yes in English. The default applies if no level declares documents.readable_brief. If the readable brief is on, the command checks each markdown brief that it writes. A diagram without exactly one non-empty %% alt: line, or with more than 25 distinct nodes, refuses the command. The refusal names the brief file and the diagram. The command writes no file. --documents no turns the brief and the check off. See document for the briefs that are checked.
On a terminal, each question prints its name, what it is, why change it, its answers, its default and where to see it later, then its answer line. The models question gives one example, producers=claude/opus-4-6, and says that running nwave-ai rigor again changes the models. The documents question says the language question, number 5, is asked only after yes. The review rounds question explains the trade-off: few rounds can leave defects unseen; many rounds can find new things and never stop. When a finding is still open at the limit, the work stops and the finding goes to its owner. Reaching the limit is never an approval.
A write prints the final summary first. The Git sentence comes before the confirmation question, so you read it before you answer yes. A kept examine level shows as keep, with its source:
Final summary - read before you confirm
review rounds: 3 (was 2, built-in default)
examine level: quick (was standard, built-in default)
gives: 1 examination, no repair round
not covered: any repair round, security, load or checks on other machines
a row not examined is never green
documents: yes, language en (was no, built-in default)
model nw-documentarist: codex/gpt-5
hosts: none
The project file .nwave/config.json: this file may be tracked by Git,
depending on your .gitignore.
Nothing is written until you answer yes.
confirm the final summary? (yes):
Written /path/to/repo/.nwave/config.json (project).
Whether the project file is tracked by Git depends on the repository's .gitignore. Run git status before you commit it.
Files written
The project file .nwave/config.json in the current directory receives:
review.max_rounds, when the rounds answer is a number.currentwrites nothing.examine.level, as{"level": NAME}, when the examine answer names a level.currentand an omitted--examine-levelwrite nothing.documents.readable_briefas{"enabled": true|false, "language": LANG}, when--documentsisyesorno.model_runtime.roles.<role>as{"provider": ..., "model": ...}, for each role a model fact names. A group names every published role of that group.
Top-level keys that rigor does not write are kept. When it writes review or documents, it replaces that whole object, so a nested key inside it that rigor does not write is dropped. The global file is read but never written. documentation, rigor.profile and model_runtime.default are not written.
Host files
Only the hosts you name and confirm are written. Paths are under your home directory.
| Host | File | Format status |
|---|---|---|
omp |
.omp/agent/config.yml, key task.agentModelOverrides. The file must already exist: without it the adapter reports a refusal naming cannot access local variable 'task', the command still exits 0, and the project file is still written (observed). |
Verified on this machine for an existing file. Comments in that YAML file are lost. |
claude-code |
.claude/agents/nw/<role>.md, a model: line; the file is created when absent |
Verified. Accepts only the claude provider and writes claude-<model>. An nWave reinstall regenerates these files. |
opencode |
.config/opencode/opencode.json, key agents.<role>.model |
Documented, not verified on this machine (version 2 format). |
copilot |
.copilot/agents/<role>.agent.md, a model: frontmatter line |
Documented only for provider-qualified local models; not verified. |
codex |
none | Refuses: the config has one top-level model and no per-role surface. |
devin |
none | Refuses: the documented config has one default model and no per-role surface. |
The summary shows each host's path before the write. Unverified formats are labelled format not verified: format NAME assumed from source URL. The previous bytes are kept beside the file as NAME.nwave-bak, and an existing backup is never overwritten. A host file whose content does not match the documented shape is not written.
Which routes read these settings
- The spine route (
nwdes prompt) readsreview.max_rounds,examine.level,model_runtime.rolesanddocuments.readable_brieffrom the project file and the global file, project first. It prints them in each role bundle:review rounds: N (source),examine level: NAME (source) - N examinations, ..., andmodel <role>: provider/model (source)ornone declared in project or global. The examiner bundle states the level in examinations and repair rounds only. The readable-brief line appears only in the charter-close and design-close bundles, and only when enabled. A review value outside 1 to 5 or an examine level outside quick, standard and extended makes the bundle refuse, naming the key. The bundle states the model;nwdesstarts no model.
Example
# What is in effect (writes nothing)
nwave-ai rigor --show
# Group, then one role; rounds 3; readable brief in English; no host file
nwave-ai rigor --yes --model-set reviewers=claude/opus-4-6 --model-set nw-documentarist=codex/gpt-5 --review-rounds 3 --documents yes --document-language en --hosts none
# Keep every current value; still writes the project file
nwave-ai rigor --yes --models keep --review-rounds current --documents keep --hosts keep
nwave-ai version
Print the installed nWave version and exit.
Synopsis
nwave-ai version
nwave-ai --version
nwave-ai -V
Exit codes
| Code | Condition |
|---|---|
| 0 | Version printed successfully. |
Output
stdout:
nwave-ai 1.1.0
Example
nwave-ai version
nwave-ai --version
nwave-ai -V
nwave-ai attribution
Enable or disable automatic commit attribution. When enabled, nWave adds a Co-Authored-By: trailer to commits made inside Claude Code hooks.
Synopsis
nwave-ai attribution <on|off|status>
Positional arguments
| Argument | Required | Type | Description |
|---|---|---|---|
| action | yes | enum | One of on, off, or status. |
Exit codes
| Code | Condition |
|---|---|
| 0 | Action completed successfully. |
| 1 | Invalid action (not on, off, or status). |
Output
on and off write the machine preference. A project override takes
precedence; credit also requires an active project and a supported commit
hook. These commands do not change project activation or overrides.
stdout on attribution on:
Attribution enabled. New Claude commits will carry the nWave credit via the universal handler (observes the preference at commit time).
stdout on attribution off:
Attribution disabled. New Claude commits will not carry the nWave credit (the universal handler observes the preference at commit time).
stdout on attribution status with an active project override:
Attribution is currently on.
Preference source: project.
Attribution is active for this repo.
status reads the effective preference (project, global, or default) and
reports attribution as inactive if the project is inactive or the preference
is off. It does not change any configuration.
stderr on invalid action:
Unknown attribution action: invalid
Usage: nwave-ai attribution <on|off|status>
Example
nwave-ai attribution on
nwave-ai attribution status
nwave-ai attribution off
nwave-ai project
Enable or disable nWave activation for the current project repository. Writes .nwave/config.json with enabled: true|false as a version-controlled marker and reconciles .gitignore so the marker stays tracked.
Synopsis
nwave-ai project <enable|disable>
Positional arguments
| Argument | Required | Type | Description |
|---|---|---|---|
| action | yes | enum | One of enable or disable. |
Exit codes
| Code | Condition |
|---|---|
| 0 | Marker written and gitignore reconciled. |
| 1 | Invalid action (not enable or disable). |
Output
stdout on success:
nWave activation for this project: enabled.
or
nWave activation for this project: disabled (sticky opt-out).
stderr on invalid action:
Usage: nwave-ai project <enable|disable>
Details
- enable: Writes
"enabled": trueinto.nwave/config.jsonand fixes.gitignoreto allow the marker file. - disable: Writes
"enabled": falseinto.nwave/config.json(sticky opt-out; the recorded value signals explicit non-participation).
See the Activation Guide for the full concept and workflow.
Example
nwave-ai project enable
nwave-ai project disable
nwave-ai mode
Set the global activation mode for nWave across all projects. Writes to ~/.nwave/config.json under the key activation.mode.
Synopsis
nwave-ai mode <all|opt-in>
Positional arguments
| Argument | Required | Type | Description |
|---|---|---|---|
| mode | yes | enum | One of all or opt-in. |
Exit codes
| Code | Condition |
|---|---|
| 0 | Mode written successfully. |
| 1 | Invalid mode (not all or opt-in). |
Output
stdout on success:
Global nWave activation mode set to 'all'.
or
Global nWave activation mode set to 'opt-in'.
stderr on invalid mode:
Usage: nwave-ai mode <all|opt-in>
Details
- all: Enable nWave for all projects by default (unless explicitly disabled per-project).
- opt-in: Disable nWave globally; activate only per-project via
nwave-ai project enable.
See Global Config Reference for the full schema.
Example
nwave-ai mode all
nwave-ai mode opt-in
nwave-ai status
Show effective activation, attribution, verbosity, running CLI version, and
detected Claude Code/Codex assets with the source of each setting. Labels
distinguish detected host files from a validated installation. Persisted
update_check is retired and ignored. Documentation density is diagnostic-only;
the global expansion preference controls optional explanation after completed
waves. This command reads local state without writing files or checking releases.
Synopsis
nwave-ai status
Exit codes
| Code | Condition |
|---|---|
| 0 | Status printed. |
| 1 | Invalid selected home override; no status printed. |
Output
stdout, for example:
Global activation mode: opt-in (default)
This project is active.
Activation source: project.
Version: 4.0.0 (running CLI; check releases: nwave-ai update --check)
Attribution: off (global); nWave co-author credit will not be added here.
Verbosity: terse (global).
Claude Code: framework files found.
Codex: installation manifests not detected.
Retired update_check: ignored (global); no automatic update prompt.
Documentation density: diagnostic-only (global); does not reshape waves.
Wave-end explanation: ask-intelligent (default; after completed waves only).
To credit future nWave commits: nwave-ai attribution on (machine-wide).
Host detection is not validation; use nwave-ai doctor for Claude Code or the Codex installer's validation.
Activation and other settings can come from a project override, the machine
config, or a default. An inactive repo instead shows
nwave-ai project enable. Status does not turn attribution on automatically.
Example
nwave-ai status
nwave-ai completion
Print a shell-completion script for Bash or Zsh to stdout. Source the output in your shell startup file.
Synopsis
nwave-ai completion <bash|zsh>
Positional arguments
| Argument | Required | Type | Description |
|---|---|---|---|
| shell | yes | enum | One of bash or zsh. |
Exit codes
| Code | Condition |
|---|---|
| 0 | Completion script printed successfully. |
| 1 | Unsupported or missing shell argument. |
Output
stdout (Bash example):
_nwave_ai_completion() {
local cur="${COMP_WORDS[COMP_CWORD]}"
local prev="${COMP_WORDS[COMP_CWORD-1]}"
# Completion logic here
COMPREPLY=( $(compgen -W "install uninstall doctor version status" -- "$cur") )
}
complete -o bashdefault -o default -o nospace -F _nwave_ai_completion nwave-ai
stderr on unsupported shell:
Unsupported completion shell: 'fish'
Usage: nwave-ai completion <bash|zsh>
Example
# Bash: add to ~/.bashrc
nwave-ai completion bash >> ~/.bashrc
source ~/.bashrc
# Zsh: add to ~/.zshrc
nwave-ai completion zsh >> ~/.zshrc
source ~/.zshrc
Advanced commands
The following subcommands are supported but are not part of the everyday user surface. Refer to their sources or existing documentation for details:
- outcomes — Register / check shipped outcomes (Tier-1 collision detection). See Outcomes CLI Reference.
- plugin — Install, uninstall, or list nWave tool plugins.
- tests — Run a test command through the test queue of the account (
tests run; one queue per account, two accounts on one machine do not coordinate) and list the tests a change selects (tests select). See Test Queue and Test Selector Reference.
nwdes
nwdes is the delivery tool. It builds role prompts and packets, writes handoff and authority records, and manages lanes. It starts no model and runs no verification command. The command list, arguments, refusals and an example for each verb come from nwdes --help and nwdes VERB --help.
Route facts
| Fact | Behavior |
|---|---|
| Roots | --repo-root defaults to the current directory and --project-root to the repository root. Lane verbs resolve the main worktree. |
| First recording | The first add, set or remove of a role kind creates its scoped draft. |
| Revision | open --from-current seeds a draft from published authority. It is not a first call. |
| Close | close --outcome OUTCOME --diagnostic TEXT publishes the draft and removes it. A refusal keeps the draft for correction. |
| Abort | abort removes the draft. |
| Values | Roles record typed values with CLI flags or --FIELD-file PATH. The root submits no role output. |
| Lane | nwdes lane integrate fast-forwards trunk to CANDIDATE, then removes the lane worktree and branch without force. |
Prompt
nwdes prompt --role ROLE --handoff PATH [--jev {authorized,unauthorized}] [--task TASK] [--slice ID] [--phase PHASE] [--diff PATH] [--root-context TEXT]
--jev is optional and authorized by default: when neither the flag nor a
configuration tier declares jev, the resolved authorization is authorized
(the flag wins over the repository tier, which wins over the global tier; a
malformed value is refused; nwdes never writes or persists the key).
The output is one complete inline role bundle that ends with the generated verb reference for that role. See the prompt contract and the nwdes command reference.
Source-blind examiner prompts
The examiner works from the user's point of view and needs only public product facts. Pass only public text in --root-context; it is rendered verbatim. The examiner prompt replaces the preloaded nw-typesafe-system-one client recipe with the skill's source-blind EXAMINE limit notice (do not call Jev; report typesafe: unavailable).
Unavailable or inconsistent required inputs produce a nonzero result with WHAT/WHY/HOW on stderr, not a partial prompt.