# Execution and model selection

For each new build, debug, review or devops task, including README edits, choose
**normal execution** or **役小角(enno-oduno)** in the native DSH question cards.
An explicit instruction such as “without enno-oduno” avoids a duplicate mode
question. Chat and existing non-Enno research/writing tasks keep their ordinary
flow. Continuations, retries and replanning retain the logical task's choice.
Cancel or skip preserves the request and draft configuration for resumption.

Normal execution uses the current DSH model, memory, applicable Skills, native
permissions and focused verification. It creates no Enno contract, WorkUnit,
plan approval or automatic Enno continuation.

On Enno and Deep selection cards, use the displayed **Cmd+1–9** shortcuts on macOS
or **Ctrl+1–9** on Windows/Linux, then press Enter to confirm. With the card
focused, you can also type an option number and press Enter to submit. For options
above nine, type consecutive digits (`1`, `2`, Enter selects option 12); Backspace
corrects the number. Digits typed in the search field remain search text, and Enter
submits the search. IME composition and repeated key events do not select or submit.

Enno offers twelve versioned templates across **OpenAI, DeepSeek, OpenCode Go,
OpenCode Zen, OpenRouter, OrcaRouter and Ollama**, plus a custom configured-model selector. The complete
[template table and configuration reference](model-selection.ja.md) lists exact
model IDs. OpenAI assigns Astra to ideal/Zenki/check, Sol to the Goki head and
Luna to workers. Go and OpenRouter provide GLM and Qwen variants; Zen uses GLM.
Ollama uses `qwen3-coder:30b` for every role, limits children to one, and never
downloads models. Reasoning uses model defaults; reflection uses the check model.

All DeepSeek recommendations use **V4.1 Flash for all five roles**. Exact IDs differ:
`deepseek-flash` on [DeepSeek's API](https://www.deepseek.com/en/news/deepseek-v4-1-flash/),
`deepseek-v4.1-flash` on [OpenCode Go](https://opencode.ai/docs/go/), and
`deepseek/deepseek-v4.1-flash` on [OpenRouter](https://openrouter.ai/deepseek/deepseek-v4.1-flash)
and [OrcaRouter](https://www.orcarouter.ai/models/deepseek/deepseek-v4.1-flash).
The direct DeepSeek template uses DSH's `deepseek-official` connection. These templates
never substitute V4 Pro or an older Flash ID when V4.1 Flash is unavailable.
The [OpenCode model registry](https://models.opencode.ai/api.json) was also checked
on 2026-09-11. Its `opencode-go` entry includes V4.1 Flash; its Zen entry (`opencode`)
does not. Its OrcaRouter entry also lacks V4.1 Flash, although OrcaRouter's own model
page lists the ID above. Registry entries can lag provider changes. Zen and Ollama
use the configured-model selector for models without a verified template ID.

For [dsh-codex](https://github.com/askdkc/dsh-codex), choose **OpenAI Codex・推奨（dsh-codex）**
from the template menu, then confirm the configuration. It assigns Astra to
ideal/planning/check, Sol to the Goki head and Luna to workers, following the
[OpenAI model guide](https://developers.openai.com/api/docs/models). This is a
quality/workload recommendation, not a measured per-task latency or quota optimum.
The plugin-specific template binds the exact `openai-codex` / `codex` / `responses`
contract from [dsh-codex's implementation](https://github.com/askdkc/dsh-codex/blob/e2e61b6d8f3b511cf2dac99688e97b8728cb122c/src/index.ts)
without asking for the connection again. Native DSH model validation takes
precedence over legacy `modelRoutes` declarations; the plugin configuration file
remains untouched. Recommended models must exist in the current DSH catalog;
missing models leave roles unset and prevent starting. The catalog is rechecked
immediately before adoption.

Bind generic templates to actual configured DSH providers; route IDs are not assumed.
Distinguish OpenAI API from Codex authentication and Go from Zen. The selector
supports provider/model search, pagination, copying roles, editing, back and
cancel. Exact IDs or declared aliases resolve models; similar display names
never authorize substitutions. Edited templates are custom but retain their
original template ID/version. Configuration confirmation shows every role's
provider and model. Provider-catalog failure, missing routes/models and required
compatibility verification have distinct presentations.
Unset roles start with enno-ideal's provider and go straight to model selection;
editing an assigned role keeps its own provider. Use “接続を変更” (change connection)
on the model card to choose another provider. If the preferred provider is unset
or no longer registered, the selector opens the provider list.
Use “検索をクリア” to clear an empty search and “一覧を再取得” to retry a failed list.
Free-text digits remain search terms. Selecting a model returns directly to role
review. Provider family, authentication and protocol are taken from DSH rather
than requested again. Back and cancel retain confirmed model assignments.

To choose by model name, select **DSHに設定済みのモデルから選ぶ**, edit a role,
select its connection, then search or select a model name and confirm the configuration.
This works with configured **OpenCode Go, OpenCode Zen, OpenRouter, Ollama and OrcaRouter**
connections. Every entry displays its name and exact ID; the same model on OpenRouter
and OrcaRouter remains a distinct binding. Lists come from the current DSH profile.
Register a missing connection in DSH and select **一覧を再取得** to refresh it.
The public registry is a recommendation reference; it does not register models in
DSH. Its `ollama-cloud` catalog is separate from a local Ollama installation.

DSH owns provider authentication, transport and model validation. Kiokuko passes
exact provider/model IDs to the native adapter and uses DSH's configurable-provider
directory for template matching and local concurrency. Unknown provider metadata
is not a reason to ask for connection details again. Authentication, quota and
unavailable-model failures retain completed work and require reselection rather
than automatic fallback.

Optional legacy `modelRoutes` entries remain readable for older hosts and provider
plugins without directory metadata. Current native metadata takes precedence;
these entries never override DSH's wire settings. On hosts without native model
validation, existing Astra/Go declarations retain their compatibility checks.
Local validation and recording-adapter tests do not prove paid account access
or provider acceptance of a request.

Resolved configuration and the ordinary model are persisted per logical run.
Template updates do not rewrite active tasks. Agent-scoped assembly/request hooks
freeze one selection per request and restore ordinary routing at completion;
the Web `selectModel()` command and deployment defaults are not changed.

`enno_delegate` accepts an `instruction` for the current approved WorkUnit. The
host validates the lease, chooses the approved worker model and uses native
in-process `spawn`. Children are bound before their first request and cannot
restart intake, delegate or report parent acceptance. Initially workers use
scoped file tools; Goki runs commands and focused verifiers and then reports with
`enno_work_report`. Child completion is evidence, not WorkUnit acceptance.
Interrupted delegations remain uncertain rather than being automatically replayed.
Child concurrency is counted from durable state across restarts. Each child request
and file operation rechecks the current lease. A hard process exit can leave a
delegation marked as running; inspect its native child evidence before additional
delegation. Restarting does not erase that reservation.

Native tests cover request model IDs, ordinary README completion, cancel/reload,
input recovery, independent sessions, routing restoration and actual adapter
transport. Recording adapters do not verify paid account access, subscription
limits or successful live provider responses. Existing advisors, leases, final
verification, memory finalization and Orca behavior remain in place.