Harden Claude subscription path + restore coding quality through the proxy

[0xPuncker]

Summary

Three intertwined fixes to the Claude subscription (OAuth) provider path, plus a Docker build fix. Coding requests routed through the proxy were noticeably worse than talking to Claude directly — the root cause was the proxy silently stripping the request settings that drive coding intelligence, plus a couple of correctness bugs on the subscription path.

Background

Live log analysis (podman logs cc-proxy) showed this on every request:

stripped fields: context_management, output_config
stripped N historical thinking block(s) for subscription

Changes

1. Coding-quality passthrough (src/index.ts — the headline fix)

cleanBody was a hardcoded allow-list that omitted output_config, so every request lost it. Claude Code puts effort (it sends effort: \"xhigh\" for Opus 4.7/4.8 — the setting Anthropic calls out as best for coding/agentic work), task_budget, and format inside output_config. Stripped → the API silently fell back to high, reducing reasoning/agentic depth on every coding request.

• cleanBody is now provider-aware: for subscription/anthropic it passes the body through untouched (only model-remap + message-clean); the strict allow-list is kept only for zai/openrouter, which genuinely need the translated shape. Future-proofs against new client fields.
• cleanMessages no longer strips all historical thinking blocks on the subscription path. It now keeps signed thinking blocks + all redacted_thinking (preserving cross-turn reasoning continuity — what direct Claude does), dropping only unsigned (contaminated) blocks that would 400. Gated by STRIP_SUBSCRIPTION_THINKING (unsigned default | all legacy | none).
• Fixed mapModel OpenRouter mapping: claude-opus-4-8/4-7/4-6 + claude-opus now map to ~anthropic/claude-opus-latest instead of being silently downgraded to Sonnet on OpenRouter fallback.

2. Tool-use ID sanitization (src/index.ts)

A conversation whose earlier turns were served by a fallback provider (GLM/Z.AI) accumulates tool-use IDs that fail Anthropic's strict ^srvtoolu_ / ^toolu_ validation, 400-ing every replay (messages.N.content.M.server_tool_use.id: String should match pattern ...). Added sanitizeToolUseIds: rewrites non-conforming IDs and remaps every paired tool_use_id reference in the same pass (subscription + anthropic paths).

3. OAuth auto-refresh (src/index.ts, src/types.ts)

The proxy now redeems the refresh token itself when the access token is near expiry or on a 401, persisting rotated credentials atomically (temp file + rename) under a single-flight mutex (Anthropic rotates the refresh token on every use). Removes dependence on the external credential-sync job. Configurable via CLAUDE_OAUTH_REFRESH_URL / CLAUDE_OAUTH_CLIENT_ID / CLAUDE_OAUTH_REFRESH_SKEW_MS.

4. Docker build fix (Dockerfile)

The production image re-ran npm ci --only=production, which dropped the pg runtime dependency (and the generated Prisma client). Reuse the builder's node_modules instead and fail the build loudly if pg is missing.

Verification

• tsc --noEmit + eslint clean
• 102/102 tests pass (93 prior + 9 new: passthrough for subscription/anthropic, still-strip for zai/openrouter, all three thinking modes, Opus→Opus mapping, sanitization pairing, refresh behavior)
• Built and deployed; live logs confirm output_config is no longer stripped, thinking-strip drops only contaminated blocks, zero new subscription 400s after keeping signed thinking.

Operational notes

• The proxy must mount the credentials file read-write so rotated tokens can be persisted (compose already has :rw).
• To revert the thinking-strip change instantly if a 400 appears: set STRIP_SUBSCRIPTION_THINKING=all and recreate the container (no rebuild).