Add a supported domain to a language

Read the extension workflow in language authoring and the submission template. The language's README must describe the new coverage and its verification.

Respect offline-only scope: reuse exact supplied artifacts, prepare requests and run local checks. add-domain with localization.mode="reuse" is offline; mode="generate" calls an LLM. Auditioning even existing voice IDs calls the provider. Mark missing generation/review/live evidence in a draft PR rather than running those steps without authorization or claiming completion.

Extend the pack

1. Identify the existing pack, its exact task bundles, requested supported domain and a new revision. Preserve source pack files. Confirm that the new domain is not already bound; replacing reviewed data requires a rebuilt revision.
2. Run tau2 factory stage-pack using the exact published pack and domain bundles.
3. For the new domain, run tau2 factory init-request --domains <domain> with the language's reviewed author form (--form), --reviewed-by, an explicit --model and --reasoning-effort, and the pack's regional settings. Use its emitted domains[0] entry (frozen-50 tasks, caller genders, task IDs and $artifact references, kept relative to the new request file) as the domain of an AddDomainRequest. Do not hand-assemble task lists.
4. Run add-domain with that bundle and a new revision. Review <output>.review/index.html (content, mappings and before/after flags), and have the reviewer record tau2 factory approve --gate content.
5. When persona/voice content is unchanged, use reuse-voices with the original reviewed assets and their audio directory; --output is a new directory. The original listening approval still covers the reused voices. Otherwise run prepare-voices, audition and tau2 factory approve --gate listening. Imported packs without factory artifacts can audition their existing voice IDs.
6. Run finalize into a new directory and update the language's submission README. Install that local output in the contribution worktree and use tau2 languages render-prompts --language <code> --domain <domain> --output <review-directory> for full runtime prompt review before submission. Record task identity/lookup checks and the smoke gate tau2 run --audio-native --benchmark none --language <code> --domain <domain> --num-tasks 2.

Inspect command help and use the package's typed request models. Keep all source requests and review digests. The stage/build commands own content generation and validation; do not improvise translations or create ad hoc scripts to replace them. Keep simulator prompts about text, spelling conditional on request and authentication guidance in domain policy.

For family-first name conventions, set the pack/form name_order metadata. The shared retail compiler supplies explicit given/family-name facts from the structured caller identity; review that output instead of adding language-specific manual prompt patches. Existing agent-clause assembly stays in the runtime.

Judge coverage

An existing language judge is not automatically validated for newly claimed conditions. Record whether the domain extension falls within its evidence scope. If extending or adding a judge, follow judge validation and keep candidate results separate until validated. User-only coverage can ship without a new judge. Do not copy caller behavior into agent judging rules.

PR

Include the reviewed pack/task change, updated README, reproducible checks and optional judge evidence. Reuse voices when possible; do not recreate them just because domain tasks changed. Do not change the benchmark's frozen language/task selection or run a full benchmark as part of routine domain localization. Report missing review or live evidence accurately; never manufacture approvals.