Configuration

In the Web UI, open Settings → Plugins → Plugin configuration and expand Codex App Server to edit image generation, Codex Web Search takeover, its default model, and its result cap. The card writes through DSH's llm-codex-app-server settings namespace. Saved values apply to the next call; a process-capability change alters the request epoch so a stale cached thread is replaced automatically.

Later profile patch layers can replace the llm-codex-app-server row. A replacement must restate the complete config because Harness patch rows do not deep-merge.

Key	Default	Meaning
`provider`	`codex-local`	Harness provider route.
`displayName`	`Codex (local login)`	Selector label.
`modelProvider`	`openai`	Codex App Server model-provider id.
`models`	Six visible App Server catalog entries	Advisory model metadata, input modalities, and reasoning choices. The shipped snapshot declares image input for five routes and text-only for Spark; omitted/custom modalities default to text-only. Unlisted safe model ids remain routable as text-only.
`timeoutMs`	`300000`	Wall-clock limit for one App Server turn.
`disposeGraceMs`	`3000`	Process-tree termination grace.
`maxJsonRpcLineBytes`	`8388608`	Maximum bytes accepted from App Server stdout for one newline-delimited JSON-RPC message. Parsed messages do not count toward a cumulative stdout limit.
`maxRequestImageBytes`	`20971520`	Maximum projected base64 image payload hydrated into one model request; oldest model-visible images become deterministic omission text first.
`maxStderrBytes`	`65536`	Maximum retained diagnostic output.
`maxRetries`	`1`	Harness-visible retries for transient process or provider failures. A model-capacity rejection waits five minutes, then retries the original model without rerouting.
`maxCachedSessions`	`8`	Maximum idle or tool-waiting session leases retained before least-recently-used eviction. Active requests may temporarily exceed it.
`sessionIdleTimeoutMs`	`600000`	Idle lifetime of a cached App Server session thread.
`imageGenerationEnabled`	`true`	Whether ordinary model turns enable native image generation in the pinned App Server. Compaction turns always keep it off.
`webSearchEnabled`	`true`	Whether this provider conditionally takes over its agents' `web_search` calls; disabling it delegates to the existing DSH Web provider chain.
`webSearchModel`	follow main model	Plugin-owned search-only model override. Blank uses the initiating Codex model; official Codex config exposes no separate Web Search model key.
`webSearchMaxResults`	`8`	Source cap applied after conditional Codex-native searches are merged.
`env`	`{}`	Explicit child environment layered over Harness's scrubbed parent environment. Use this to pass `CODEX_HOME` when it is nonstandard.

Example override:

- id: llm-codex-app-server
  name: dsh-llm-codex-app-server
  config:
    provider: codex-local
    displayName: Codex (local login)
    modelProvider: openai
    models:
      - id: gpt-6-astra
        name: GPT-6-Astra
        contextWindow: 272000
        inputModalities: [text, image]
        reasoningEfforts: [low, medium, high, xhigh, max, ultra]
        defaultReasoningEffort: medium
    timeoutMs: 600000
    disposeGraceMs: 3000
    maxJsonRpcLineBytes: 8388608
    maxRequestImageBytes: 20971520
    maxStderrBytes: 65536
    maxRetries: 1
    maxCachedSessions: 8
    sessionIdleTimeoutMs: 600000
    imageGenerationEnabled: true
    webSearchEnabled: true
    webSearchModel: gpt-5.6-luna
    webSearchMaxResults: 8
    env:
      CODEX_HOME: !!js process.env.CODEX_HOME

dsh-llm-codex-app-server

AI 分析

安裝

說明文件

Configuration