dsh Codex

English | 中文

Use a ChatGPT subscription in DeepSeek Harness through OpenAI's Codex sign-in flow—no OpenAI Platform API key required and no dsh source patch required.

dsh-codex is an independent dsh bundle. It adds:

• ChatGPT OAuth from the dsh Settings panel or a standalone CLI, with automatic token refresh
• the Codex GPT catalog, including vision-capable models when the account offers them
• streaming, tool calls, reasoning replay, prompt caching, and dsh compaction through the normal LLM service
• Codex standalone web search through dsh's existing web_search tool
• optional HTTP(S) URL input added to Harness's existing read_image tool
• an imagegen tool backed by gpt-image-2, with workspace or conversation reference images and automatic workspace output
• browser image input through dsh's existing paste and drop controls
• a per-conversation Fast Mode switch and compact weekly quota indicator in the Web composer

ChatGPT subscription authentication and usage-based OpenAI API access are different products. This plugin uses the ChatGPT Codex backend only; it does not turn a subscription into a general-purpose OpenAI API credential.

Install

Install the prebuilt bundle from npm into the selected dsh profile:

dsh plugin --profile web add dsh-codex
dsh web

From a DeepSeek Harness source checkout, use pnpm dsh plugin --profile web add dsh-codex. A local plugin checkout can still be installed with link:/absolute/path/to/dsh-codex for development.

Open Settings → OpenAI Codex → Sign in with ChatGPT. The plugin opens OpenAI's authorization page and completes the localhost callback. The account page shows live Codex quota bars and exact remaining percentages; exact credit balances or workspace limits appear only when the account API supplies them.

Loopback Web pages are trusted automatically. If dsh runs on another machine, the account page shows the exact origin command that must be approved on the dsh host, for example dsh plugin --profile web exec dsh-openai-codex trust-origin http://host:port. The allowlist is exact-origin, stored separately from OAuth credentials, and can be inspected or revoked with trusted-origins and untrust-origin.

The CLI remains available for terminal and headless installations:

dsh plugin --profile web exec dsh-openai-codex login
dsh plugin --profile web exec dsh-openai-codex login --device-code
dsh plugin --profile web exec dsh-openai-codex status
dsh plugin --profile web exec dsh-openai-codex doctor --json
dsh plugin --profile web exec dsh-openai-codex logout

For dsh-tui, install the bundle into the same profile:

dsh plugin --profile dsh-tui add dsh-codex

After restarting the TUI, /model lists the openai-codex catalog. With no explicit route or saved selection, the TUI adopts the bundle's gpt-5.6-sol default. Use /codex status|login|logout|usage|config for the account and live settings; the four boolean settings can be changed with /codex set <read-image|imagegen-other-models|websocket-context|native-compaction> <on|off>. Browser login shares the same dsh credential file used by the Web profile.

Codex, Claude Code, and other automation agents should follow INSTALL.md. It is a complete, idempotent runbook and does not require reading this repository's source or design notes.

The bundle selects openai-codex / gpt-5.6-sol for new agents and selects the Codex search provider. A model already saved in dsh settings still takes precedence; the model picker can select any other Codex model visible to the signed-in account.

Model catalog

By default, the model picker advertises the complete openai-codex catalog. Open Settings → OpenAI Codex and use the model checkboxes to choose which entries remain visible. The selection is live and durable; dsh refreshes the Web and TUI model directories after it changes.

The same initial subset can be seeded through models on the llm-openai-codex entry while preserving provider order:

- id: llm-openai-codex
  config:
    models:
      - gpt-5.6-luna
      - gpt-5.6-sol
      - gpt-5.6-terra

The checkboxes and models setting control discovery only. A hidden model already stored in an existing session or supplied explicitly remains resolvable, so narrowing the picker does not invalidate older records. Omit models to start with the full catalog; an empty list advertises no models.

Images

Image support uses dsh's durable attachment path:

• paste an image into the Web composer with Ctrl+V, or drag and drop it;
• on Windows, paste a clipboard image with Ctrl+V in the adapted dsh-tui, or enter @relative/image.png; clipboard images go straight to the attachment store, while path images use the active workspace filesystem;
• ask the model to call read_image with either file_path for a workspace image or url for an HTTP(S) image;
• PNG, JPEG, WebP, and GIF are accepted within the active dsh attachment limits;
• only a model that explicitly advertises image input may receive an image.

imagegen is available to any vision-capable conversation model. The current model writes an ordinary prompt and may select either referenced_image_paths or num_last_images_to_include; the plugin reads the bytes from ctx.fs or the attachment store and sends them to gpt-image-2. The model never emits base64. Every result is shown inline, saved as a durable attachment, and written to the active workspace. output_path chooses the destination; omitting it creates a unique generated-<timestamp>-<id>.png file. Local saving is included in this plugin, while dsh-remote-ssh supplies the remote AHP write path when that plugin owns the workspace.

The Settings page has separate Enhance read_image and Image generation for other models toggles. Both default on. Turning off the first removes the plugin's agent-scoped override and restores Harness's original local-only read_image schema. Turning off the second keeps imagegen available to Codex vision models and rejects calls from other model providers at execution time.

read_image stores validated bytes as a dsh attachment before returning the actual image block. Local paths are delegated unchanged to Harness, including its configured filesystem and sandbox behavior. The URL extension bounds redirects and bytes, rejects credentials embedded in URLs, rejects local/private/special network targets, and pins each validated public address across the corresponding HTTP hop.

For an eligible Codex GPT conversation, the Web composer also exposes a session-local Fast Mode switch. Enabling it adds the provider's priority service tier only to that conversation; it does not change saved model settings. A neighboring quota bar shows the applicable weekly limit and provider-declared reset time.

Search

The provider connects dsh's web_search tool to the standalone search protocol used by Codex. It returns ordinary dsh text and HTTP(S) citations, so later turns and compaction retain the tool history.

Configure the llm-openai-codex row in a profile patch:

- id: llm-openai-codex
  config:
    searchMode: live
    searchContextSize: medium

Each resolved, secret-free auxiliary request is recorded before dispatch as the dedicated web/openai-codex-search-llm-request session event. The event is owned and registered by this plugin; no generic search event or dsh fork is required.

Responses API experiments

The Settings page provides two Codex-only switches. Both are off by default:

• WebSocket context reuse keeps store: false and selects pi-ai's Codex WebSocket continuation transport. While the same session keeps a reusable connection and the next request is an exact extension, it sends previous_response_id with only the new input. History edits, compaction, Fork, connection loss, and process restarts fall back to a full request. With the switch off, ordinary turns use SSE and always send the full Harness context.
• Native Responses compaction follows Codex's current V2 flow: it sends the existing history plus a compaction_trigger item through codex/responses, retains recent client messages with the returned encrypted compaction item inside the Harness checkpoint, and restores those native items on later requests. Existing checkpoints remain readable after the switch is disabled. If V2 compaction is unavailable or fails, the same call falls back to the existing Harness model summary.

The switches are independent. Every ordinary Codex request keeps store: false; the default uses SSE with the text-summary path from dsh-compaction-basic.

Credentials and privacy

dsh keeps this login separate from Codex CLI/Desktop:

• credentials are stored at $DSH_HOME/.openai-codex-auth.json (~/.dsh by default);
• writes are atomic and token refresh is locked across local dsh processes;
• browser status and diagnostics never return token values;
• ~/.codex/auth.json is never copied or modified.

Keeping the stores separate prevents two clients from racing the same rotating refresh token. Removing the bundle does not delete the credential; use the account page or logout command when the local account should be removed.

Compatibility notes

• This branch targets the DSH 0.1.0-rc.7 plugin surfaces and @earendil-works/pi-ai 0.82.1. The adapter migrates the earlier pi-ai replay envelope while reading history so existing reasoning/tool metadata remains usable after the rc.7 upgrade.
• The plugin runs on released dsh plugin surfaces and does not require a modified Harness checkout. It can generate attachments and save local output when installed alone.
• ChatGPT plan eligibility, model access, quotas, and backend behavior are controlled by OpenAI and may change.
• The Codex endpoint does not enforce the ordinary Responses max_output_tokens field. Compaction works, but its configured summary cap cannot be imposed server-side on this route.
• Filesystem, shell, skills, MCP, subagents, permissions, attachments, compaction, and the web_search tool itself still come from the active dsh profile.
• The standalone search endpoint is not a public OpenAI Platform API. Compatibility follows the pinned Codex/pi-ai implementation.

See the design document for protocol, persistence, and lifecycle details.

Development

pnpm install
pnpm run check

The check performs strict Host and browser TypeScript checking, focused tests, and both runtime bundles.

License

Apache-2.0