docs: describe the shadow-call model set as a set

The docs named a single intercepted slug, which was already wrong for clients
from 0.145.0 and would go wrong again on the next bump. They now describe the
prefix set and where it is configured.

Author: lidge-jun

docs-site/src/content/docs/ko/reference/cli.md

@@ -196,7 +196,7 @@ GUI가 필요하지 않습니다. `add`, `remove`, `list-custom`은 설정 파
 | `provider <name> <on\|off>` | `--json` | 한 프로바이더의 모든 모델을 한 번의 쓰기로 켜거나 끕니다. |
 | `selected <provider>` | `--set <id,id...>`, `--clear`, `--json` | 프로바이더 모델 허용 목록을 읽거나 교체합니다. `--clear`는 목록을 지워 전체 모델을 제공합니다. |
 | `context <status\|value <tokens>\|provider <name> <on\|off>\|all <on\|off>>` | `--json` | 컨텍스트 윈도 상한을 전역 또는 프로바이더 단위로 읽고 설정합니다. |
-| `shadow <status\|set> [model\|-]` | `--enabled <on\|off>`, `--json` | 백그라운드 shadow-call 모델을 읽거나 설정합니다. `-`는 모델을 지웁니다. |
+| `shadow <status\|set> [model\|-]` | `--enabled <on\|off>`, `--json` | Codex 백그라운드 헬퍼 호출을 대체할 모델을 읽거나 설정합니다. `-`는 모델을 지웁니다. `status`는 프록시가 가로채는 헬퍼 슬러그 `sourceModels`도 함께 보여줍니다(기본값 `gpt-5.4-mini`, 그리고 Codex 0.145.0부터 쓰이는 `gpt-5.6-luna`). |
 
 ```bash
 ocx models live --json                                  # 지금 Codex가 실제로 보는 목록

docs-site/src/content/docs/reference/cli.md

@@ -247,7 +247,7 @@ require the proxy to be running (`ocx start`, or an installed service).
 | `provider <name> <on\|off>` | `--json` | Enable or disable every model of one provider in a single write. |
 | `selected <provider>` | `--set <id,id...>`, `--clear`, `--json` | Read or replace the provider model allowlist. `--clear` removes the allowlist so every model is offered. |
 | `context <status\|value <tokens>\|provider <name> <on\|off>\|all <on\|off>>` | `--json` | Read or set the context-window cap, globally or per provider. |
-| `shadow <status\|set> [model\|-]` | `--enabled <on\|off>`, `--json` | Read or set the background shadow-call model. `-` clears the model. |
+| `shadow <status\|set> [model\|-]` | `--enabled <on\|off>`, `--json` | Read or set the replacement model for Codex's background helper calls. `-` clears the model. `status` also reports `sourceModels`, the helper slugs the proxy intercepts (default `gpt-5.4-mini` and `gpt-5.6-luna`, which Codex 0.145.0 switched to). |
 
 ```bash
 ocx models live --json                                  # what Codex can actually see right now

docs/shadow-call-intercept.md

@@ -2,17 +2,24 @@
 
 ## What are shadow calls?
 
-Codex Desktop App makes background API calls using `gpt-5.4-mini` for internal helper tasks:
+Codex Desktop App makes background API calls with a hard-coded helper model for internal tasks:
 
 - **Thread title generation** — auto-generates 3-8 word titles after your first prompt
 - **Commit message generation** — generates git commit messages
 - **Skill orchestration** — internal orchestration turns
 
 These calls happen independently of your selected main model and use `reasoningEffort: low`.
 
+The helper model is not stable across client versions. Codex used `gpt-5.4-mini` up to 0.144.x and
+moved to `gpt-5.6-luna` in 0.145.0, which silently disabled a single-literal intercept
+([#311](https://github.com/lidge-jun/opencodex/issues/311)). The intercept therefore matches a
+**set** of source-model prefixes — `gpt-5.4-mini` and `gpt-5.6-luna` by default — so a client bump
+does not quietly turn the feature off. Routed ids (`provider/model`) are never matched: a shadow
+call is always a bare native slug, and an explicit routed selection must not be hijacked.
+
 ## The problem
 
-- Non-OpenAI providers (Bedrock, Azure, etc.) may not support bare `gpt-5.4-mini`, causing 404 errors
+- Non-OpenAI providers (Bedrock, Azure, etc.) may not support the bare helper slug, causing 404 errors
 - Users have no control over which model handles these helper tasks
 - Shadow calls consume API quota without user awareness
 
@@ -27,6 +34,9 @@ Related GitHub issues: [#26288](https://github.com/openai/codex/issues/26288), [
 3. Toggle the switch to enable
 4. Enter a replacement model (e.g., `gpt-5.5`)
 
+The panel badge and tooltip render the source models the running proxy actually intercepts, read
+from `GET /api/shadow-call-settings` (`sourceModels`), so the UI never names a stale slug.
+
 ### Via config.json
 
 ```json
@@ -38,13 +48,29 @@ Related GitHub issues: [#26288](https://github.com/openai/codex/issues/26288), [
 }
 ```
 
+`sourceModels` is an optional override. When set to a non-empty array of strings it **replaces**
+the defaults rather than extending them:
+
+```json
+{
+  "shadowCallIntercept": {
+    "enabled": true,
+    "model": "gpt-5.5",
+    "sourceModels": ["gpt-5.6-luna"]
+  }
+}
+```
+
 ### Behavior
 
-- When enabled, ALL requests with model IDs starting with `gpt-5.4-mini` are rewritten to the configured model
+- When enabled, ALL requests whose bare model id starts with one of the source-model prefixes
+  (default `gpt-5.4-mini`, `gpt-5.6-luna`) are rewritten to the configured model
 - Reasoning effort is forced to `low` (matching the original behavior)
 - The original model ID is logged as `shadowCallRewrittenFrom` in request logs
 - When disabled (default), no interception occurs
 
 ### Warning
 
-Enabling this redirects ALL gpt-5.4-mini requests. If you intentionally use gpt-5.4-mini for other tasks, those will also be redirected.
+Enabling this redirects every request for a source model, not just Codex's background helper turns.
+`gpt-5.6-luna` is also a selectable chat model, so if you pick it as your main model while the
+intercept is on, those turns are redirected too — narrow `sourceModels` if that matters to you.

structure/05_gui-and-management-api.md

@@ -77,7 +77,7 @@ this document owns is which module holds which area and what invariant that area
 | System | `POST /api/system/restart` restarts the proxy in place. `GET /api/system/memory` — service-process runtime/memory identity (pid, Bun version/revision, platform, RSS/heap/external/ArrayBuffers scalars, observed memory = max(RSS, external, ArrayBuffers), `bun:jsc` heap context, streamMode + eager-relay gate decision, watchdog snapshot sliced to the last 60 samples). Scalar-only payload; rides the standard management auth gate and must never move to unauthenticated `/healthz`. Consumed by `ocx doctor`'s Memory/runtime section and the dashboard Memory observability card. |
 | Stop | `POST /api/stop` — restore native Codex, stop any installed service, and exit the proxy. |
 | Diagnostics/sync | `src/server/management/config-routes.ts` — `GET /api/diagnostics/project-config` reports project-level Codex config that bypasses managed routing; `POST /api/sync` re-runs catalog/config sync. The diagnostic reports the bypass; it does not rewrite the project file. |
-| Sidecar/shadow-call settings | `src/server/management/config-routes.ts` — `GET/PUT /api/sidecar-settings` and `GET/PUT /api/shadow-call-settings`. PUT accepts model and backend plus optional `webSearch.reasoning` and `vision.maxDescriptionsPerTurn`; the read and PUT-response payload reports model, backend, and the vision per-turn limit. Credentials live in the provider and OAuth stores instead. |
+| Sidecar/shadow-call settings | `src/server/management/config-routes.ts` — `GET/PUT /api/sidecar-settings` and `GET/PUT /api/shadow-call-settings`. PUT accepts model and backend plus optional `webSearch.reasoning` and `vision.maxDescriptionsPerTurn`; the read and PUT-response payload reports model, backend, and the vision per-turn limit. Credentials live in the provider and OAuth stores instead. Both shadow-call responses also report the resolved `sourceModels` — the prefixes the runtime actually intercepts (`src/lib/shadow-call.ts`, default `gpt-5.4-mini` + `gpt-5.6-luna`), so no client hard-codes a helper slug that a Codex release can invalidate. |
 | Storage | `src/server/management/logs-usage-routes.ts` — `GET /api/storage`, `POST /api/storage/cleanup/preview` and `/api/storage/cleanup`, `GET /api/storage/trash`, `POST /api/storage/trash/restore`, and `GET/PUT /api/storage/cleanup-policy` plus `POST /api/storage/cleanup-policy/run`. `GET /api/storage/cleanup-policy/test-stream` and `GET /api/storage/trash/restore/test-stream` exist for progress-stream testing. Cleanup takes an explicit `mode`: `quarantine` moves to trash and is restorable, `permanent` is not. The caller must name the mode — there is no default that silently deletes. |
 | Provider quotas and tests | `src/server/management/provider-routes.ts` — `GET /api/provider-quotas`, `POST /api/providers/test`, `GET/PUT /api/provider-context-caps`, `GET /api/provider-presets`. A quota read may be served from cache or force-refreshed; absent quota data is reported as unknown rather than as a measured zero. |
 | Models and visibility | `src/server/management/model-routes.ts` — `GET /api/models`, `PUT /api/disabled-models`, `PUT /api/model-visibility`, `PUT /api/selected-models`, `GET/POST /api/custom-models`. Visibility writes trigger catalog sync through the owning server path. |