DevAgentRadar → Cline

You are helping me adopt a real coding-assistant change. Work only from the facts below. Do not invent features.

Context

Assistant: Cline Proposal: Upgrade Cline to SDK v0.0.87 and replace Claude Opus 5 with Sonnet 5.5 and Opus 5.5 Summary: SDK v0.0.87 starts the Cline hub behind Windows HTTP(S)_PROXY, stops double-counting reasoning tokens, and refreshes the catalog to 211 providers and 6,447 models. Upgrade, drop Claude Opus 5, and pin Claude Sonnet 5.5, Claude Opus 5.5, and stealth/pixel-canary. Primary source: https://github.com/cline/cline/releases/tag/sdk/sdk/v0.0.87

Why it matters

Cline SDK v0.0.87 unblocks hub startup on Windows machines that export HTTP_PROXY or HTTPS_PROXY (Clash, v2ray, corporate proxies). Bun fetch sent loopback hub probes through the proxy with no localhost bypass, so a healthy daemon looked unreachable and startup failed with "No compatible hub runtime is available". The hub client now calls ensureLoopbackProxyBypass() before every probe, drain, and shutdown so loopback addresses are added to NO_PROXY and no_proxy. Windows compiled Bun entries under B:~BUN\ are detected so daemons spawn with --cline-hub-daemon.

Token accounting changed: providers report reasoning tokens as a subset of outputTokens, but normalizeUsage() previously returned the raw total plus reasoningTokenCount, double-counting session totals and telemetry. outputTokens now excludes reasoning tokens (including the AI SDK v4 outputTokens.reasoning shape); cost still uses the full total at the output rate.

The recommended list adds Claude Sonnet 5.5, Claude Opus 5.5, and stealth/pixel-canary and drops Claude Opus 5. Bedrock routes bare gpt-6 and gpt-5.6 through inference profiles, keeps gpt-oss on-demand, and maps ap-south-1 and ap-south-2 to the in. geo profile. Upgrade so hub startup, usage totals, and model defaults match what this SDK actually does.

Suggested actions

1. Upgrade Cline to SDK v0.0.87 from https://github.com/cline/cline/releases/tag/sdk/sdk/v0.0.87 and restart the hub on any machine that exports HTTP_PROXY or HTTPS_PROXY (Clash, v2ray, corporate proxies) so ensureLoopbackProxyBypass() adds loopback addresses to NO_PROXY/no_proxy before every probe, drain, and shutdown.
2. In the Cline recommended-model list, add Claude Sonnet 5.5, Claude Opus 5.5, and stealth/pixel-canary, and remove Claude Opus 5. For new providers, add bee (Bee by HEOSSI) and pareto (Pareto Inference, PARETO_API_KEY).
3. On Amazon Bedrock, route bare gpt-6 and gpt-5.6 ids through inference profiles like Claude, Nova, and DeepSeek; leave gpt-oss on-demand; resolve ap-south-1 and ap-south-2 to the in. geo profile. Bump @ai-sdk/amazon-bedrock to ^5.0.96 so OpenAI inference-profile models send reasoning.effort instead of reasoningConfig.
4. When reporting Cline usage, treat outputTokens as excluding reasoning tokens and keep reasoningTokenCount separate; do not add them together. Compute cost from the full billed total, since reasoning is still billed at the output rate.
5. Use Cline ./cloud export for shared cloud handoff: move a local session to a cloud agent and back, and suspend/resume persistent cloud workspaces. Saved reasoning settings survive a runtime restart and a resume; automatic Git backups run only for ephemeral sandboxes.

After you finish

Do not report this as applied to DevAgentRadar. You cannot write the visitor's loop.

Tell the human: open https://devagentradar.com/proposals/cline-sdk-sdk-v0-0-87-upgrade-cline-to-sdk-v0-0-87-and-replace-claude-op and mark Applied, Skipped, or Failed. Proposal id: 8bba4e27-25db-4f24-83d1-33d10bf9e989

Your job

1. Restate the change in one sentence.
2. Propose a minimal plan for my repo (or a throwaway pilot).
3. Implement only what I approve; prefer small diffs and tests.
4. Call out risks (permissions, breaking APIs, cost).

Start by confirming you understood the proposal.