dsh-anthropic-oauth

Bridge your existing Claude Code OAuth session (Pro / Max / Team) into DeepSeek Harness (DSH) — with zero hardcoding.

• Reads ~/.claude/.credentials.json claudeAiOauth live (same file Claude Code writes)
• Auto-refreshes accessToken via https://platform.claude.com/v1/oauth/token when <5 min remains
• Syncs ANTHROPIC_OAUTH_TOKEN (+ CLAUDE_CODE_OAUTH_TOKEN alias) into ~/.dsh/.credentials.yaml — the credential seam dsh-llm-pi-ai already resolves
• Ensures llm-pi-ai.providers.anthropic (api: anthropic-messages) exists in ~/.dsh/settings.yaml
• Pulls the model catalog live from GET /v1/models (id, display name, context window, max output, thinking/effort capabilities) — no hardcoded model list; new Claude releases appear automatically
• Sends the current Claude Code request identity (up-to-date OAuth beta set + claude-cli impersonation headers) with no real user identifier — cch is a zeroed placeholder
• Billed to your subscription, not a console API key (Bearer + oauth-2025-04-20)
• Live quota panel: shows your subscription's 5-hour and 7-day usage with a reset countdown, read from the anthropic-ratelimit-unified-* headers on a cheap max_tokens:1 probe (these headers ride only on /v1/messages, never /v1/models)

This does not embed or persist your tokens beyond the two standard DSH/Claude files.

Quota display

In the DSH web UI, the plugin's client half docks a themed quota panel in the composer's ambient readout slot (conversation.composer.dock) — non-invasive (touches no shell chrome) and theme-aware (uses --dsw-* tokens, so it adapts to light/dark):

• 5h / 7d bars with per-window utilization (green → amber ≥70% → red ≥90%)
• live reset countdown to the representative (usually 5-hour) window
• Check quota button forcing a fresh probe (otherwise auto-refreshes every 60s)
• model-aware: the panel only appears while the session's active model routes through the Anthropic provider. Switch to a non-Anthropic model (Qwen, DeepSeek, GLM, …) and it hides itself; switch back and it returns — read live from the shared model-selection directory, so it reacts immediately. If the model-selection service is unavailable, the gate is skipped (panel shown) rather than hidden forever.

The client half is authored directly in the DSH client-bundle format (window.__ModuleLoader__.load) and served verbatim — no bundler or build step. It reads two JSON endpoints the host half serves:

The quota probe spends a negligible amount (Haiku, 1 output token) and is throttled to 60s + manual refresh.

Install

# stable — from npm (once published):
dsh plugin --profile headless add dsh-anthropic-oauth
dsh plugin --profile web add dsh-anthropic-oauth   # optional; bridge is headless but harmless in web

# bleeding edge — from GitHub:
dsh plugin --profile headless add github:TrueNix/dsh-anthropic-oauth

Prerequisites

• Claude Code logged in: claude login or claude setup-token at least once — ~/.claude/.credentials.json must contain claudeAiOauth
• DSH ~0.1.0-rc.6+ (has dsh-llm-pi-ai, credentials, settings, fs)

How it works

~/.claude/.credentials.json  --watch-->  dsh-anthropic-oauth Host plugin
   claudeAiOauth {accessToken, refreshToken, expiresAt}
                                         │
                               refreshIfNeeded() if expiry <5m
                               POST https://platform.claude.com/v1/oauth/token
                               {grant_type: refresh_token, client_id: 9d1c25…f5e}
                                         │
                                         ├─► ~/.claude/.credentials.json (updated expiry)
                                         └─► ~/.dsh/.credentials.yaml
                                              ANTHROPIC_OAUTH_TOKEN: sk-ant-oat01-…
                                              CLAUDE_CODE_OAUTH_TOKEN: sk-ant-oat01-…

                                          └─► GET /v1/models  (live catalog)
                                               → llm-pi-ai.providers.anthropic.models
                                                 [{id, name, contextWindow, maxTokens,
                                                   reasoning, effortLevels}, …]

~/.dsh/settings.yaml llm-pi-ai.providers.anthropic {apiKeyEnv: ANTHROPIC_OAUTH_TOKEN}
         │
         └─► pi-ai anthropicMessagesApi — Authorization: Bearer sk-ant-oat01-…
             anthropic-beta: oauth-2025-04-20,claude-code-20250219,…

The model catalog is refreshed from GET /v1/models (≤6h TTL, or on demand via
POST /api/anthropic-oauth/sync); if discovery fails the provider keeps a minimal
Sonnet-4.5 seed so it still resolves offline. Quota is probed from
POST /v1/messages (max_tokens:1, Haiku — fractions of a cent, 60s TTL)
and exposed at /api/anthropic-oauth/quota.

Set as default model

# ~/.dsh/settings.yaml
agent-default-model:
  provider: anthropic
  model: claude-sonnet-4-5

Or pick per-session in Settings → Models after a reload.

Logs & manual sync

# DSH console (bridge + quota probe):
# [anthropic-oauth] pulled N models from /v1/models
# [anthropic-oauth] refreshed, new expiry …

# Force-refresh the model catalog (also syncs quota):
curl -X POST http://127.0.0.1:3080/api/anthropic-oauth/sync

# Live quota (60s TTL; ?force=1 bypasses cache):
curl http://127.0.0.1:3080/api/anthropic-oauth/quota
curl "http://127.0.0.1:3080/api/anthropic-oauth/quota?force=1"

Troubleshooting

• MISSING_CREDENTIAL llm-pi-ai: no credential for provider route "anthropic" — bridge hasn't synced yet; re-login with claude login.
• 401 from Anthropic — token expired and refresh failed; claude login again.
• Quota ok:false with a refresh error — subscription may have changed; re-login, then check /api/anthropic-oauth/quota?force=1.

Security

• Never commits tokens — .gitignore excludes *.oat*, *.credentials.*.
• Tokens live only in 0600 files owned by you: ~/.claude/.credentials.json and ~/.dsh/.credentials.yaml.
• A Feb-2026 Anthropic docs note declares Pro/Max OAuth for official clients only — personal use at your own risk; use a console ANTHROPIC_API_KEY for production.

License

MIT