v0.121.4

What's Changed

• fix(frontend): stop a palette scroll race from failing web unit tests by @mmabrouk in #7184
• fix(observability): correct tracing analytics figures by @mmabrouk in #7176
• fix(sdk): keep one context across streamed pulls and pin usage to the workflow span by @mmabrouk in #7175
• fix(api): roll up costs and tokens for agent batches, zero costs and summary spans by @mmabrouk in #7179
• fix(api): recompute trace totals when a trace arrives in several OTLP requests by @mmabrouk in #7181
• fix(api): price LLM spans correctly and mark spans that cannot be priced by @mmabrouk in #7178
• fix(runner): correct token and cost reporting for the Pi, Claude and Codex harnesses by @mmabrouk in #7177
• fix(runner): keep usage across an approval pause, and mark custom-connection model spans by @mmabrouk in #7183
• fix(runner): mark spans served by a custom model connection by @mmabrouk in #7187
• chore(runner): bump the Pi harness to 0.87.1 so Pi prices calls with current rates by @mmabrouk in #7180
• feat(runner): use OpenRouter's billed cost on Pi chat spans by @mmabrouk in #7185
• fix(mobile): fill the trace drawer's data slots so tabs show span data by @mmabrouk in #7182
• fix(frontend): wait for the build kit before the first /m message by @mmabrouk in #7190
• fix(api): keep chat app inputs when an evaluation invokes it by @mmabrouk in #7196
• fix(api): download litellm's latest price list at startup again by @mmabrouk in #7197
• fix(runner): keep a late Claude cost reading from overwriting the turn's usage by @mmabrouk in #7203
• [release] v0.121.4 by @github-actions[bot] in #7191

Full Changelog: v0.121.3...v0.121.4