dsh-plugin-oauth-subs

简体中文 | English

Use a ChatGPT / Codex, xAI Grok, Zhipu GLM, AWS Kiro, Google Antigravity, Cursor, Ollama Cloud, Kimi Code Plan, or GitHub Copilot subscription inside DeepSeek Harness. Official OAuth, plus Kiro API keys, Cursor CLI/IDE reuse, Ollama API keys (ollama.com Cloud, not localhost:11434), Kimi device-code / kimi-code.json, and GitHub Copilot device-code / hosts.json. Loopback proxy + llm-pi-ai route sync; each family picks one DSH api from openai-responses | openai-completions | anthropic-messages.

Install

dsh plugin --profile web add https://github.com/xxww0098/dsh-plugin-oauth-subs
dsh web

Open Settings → OAuth subs. One card per account (quota on every card; Ollama Cloud has no quota bars). About Installed re-reads the package.json this process loaded (not a module-load freeze). A leftover dsh web after a wrapper/pidfile restart still serves the old process until that PID is killed (pgrep -lf 'dsh web'). If profile node_modules is newer, About lists On disk and may add …#<tag> even when that file is already latest. Or pnpm dsh web --patch ./cordis.patch.yml (id: oauth-subs).

Families

Tokens: <profile>/data/dsh-plugin-oauth-subs/auth.json (0600). Models: models.json beside it.

How it works

Not a second LLM adapter. After Settings closes, DSH still calls the loopback proxy. Bind is loopback-only; local credential is DSH_OAUTH_SUBS_API_KEY. GLM 150% Coding Plan boost is identity (ZCode Desktop UA), not a protocol claim. Stack and module tree: AGENTS.md. Reference hops (official CLI + community reverse): docs/oauth.md.

Cache

Acceptance on the full session-772f7f3a-… SkillStar turn (oauth-codex / gpt-5.6-terra-fast, 211 calls, 71 min):

Remaining uncached tokens are almost all new tool output (delta) plus expected prefix rewrites: leaving plan mode (step 55, 169k) and DSH compaction (330k); the next call after each rewrite reused ~99%. Healthy: weighted hit ≥ 80%, zero affinity misses, no TRANSPORT. Compaction / request/header rebuild zeros do not fail the session. Details: docs/error.md.

Diagnose

npm run analyze -- path/to/session.jsonl
node --experimental-strip-types scripts/analyze-session.ts --json path/to/session.jsonl
node --experimental-strip-types scripts/analyze-session.ts --fail-below 80 path/to/session.jsonl

The analyzer labels each call cold_start / delta / compaction / rebuild / affinity_miss so a compacted session is not flagged as a shard regression. Import as dsh-plugin-oauth-subs/analyze-session.

Fast / models / reasoning

Login and chat use official client identity; UA / fingerprint live in each src/oauth/<id>/README.md. The repos those fingerprints were copied from are listed in docs/oauth.md. Settings → Models: per-family checkboxes (default all on except 900K). Reasoning is set in the Harness session menu, not Settings → Models. Fast and 900K spend quota faster.

Codex Priority echo created=auto / completed=default is not a confirmation (openai/codex#14204). 2026-08-26 Luna: 88.3 vs 57.5 tok/s (1.54×); 2026-08-30 interleaved mean 1.33× (1.90 then 0.93). Throughput-only; TTFT and cache unchanged.

Quota

Refresh about once a minute, or Refresh quota. Bars: hsl(remaining × 1.2, 78%, 38%). Codex pro → Pro 20x / $200, prolite → Pro 5x / $100. Plus/Pro may bank weekly resets — one confirm button per credit on the Codex card (Harness risk dialog, then POST …/consume with { redeem_request_id } + idempotencyKey). That spend refreshes the weekly window. Grok has no equivalent. Ollama Cloud has no documented quota JSON (/api/quota 404); the card stays idle with no bars.

Options

Develop

npm test
npm run analyze -- path/to/session.jsonl

See CONTRIBUTING.md.