Release v3.3.0

What's Changed

• fix(sessions): a session permalink outruns the rollout flag by @chelojimenez in #4427
• fix(agent-turn): namespace server skills by server name, not raw id by @chelojimenez in #4428
• Skill loads follow the host's approval policy, not a forced prompt by @chelojimenez in #4429
• harness: fix the Claude Code bootstrap on pnpm 11, and stop depending on one setting by @chelojimenez in #4430
• Skills tab: address hosted servers by id, release the MCP half on its own flag by @chelojimenez in #4432
• Clarify Skills tab feature gating in docs by @mintlify[bot] in #4434
• Skills over MCP: inspection on every agent surface by @chelojimenez in #4433
• The hosted skills request must declare the skills capability by @chelojimenez in #4435
• server doctor: don't blame the server when we're the reason skills went uninspected by @chelojimenez in #4436
• Skills: review follow-ups from #4433, and two docs claims that stopped being true by @chelojimenez in #4437
• Versioned Skills: surface skill revisions in run comparisons and the environment picker by @chelojimenez in #4431
• Skills rows carry their origin, and re-list when the handshake changes by @chelojimenez in #4440
• Document server skills UI: inline badges and unified refresh by @mintlify[bot] in #4442
• WebMCP Inspector V1 (local) by @chelojimenez in #4422
• Fix/mcpjam elicitation default by @ignaciojimenezr in #4443
• feat(evals): wire B5c intent and stage measurements by @chelojimenez in #4441
• Document intent field on EvalTestConfig and EvalResultInput by @mintlify[bot] in #4448
• Remove the non-prod lockdown gate; move staging access to the edge by @chelojimenez in #4449
• Hosted skills routes connect before asking whether the extension is active by @chelojimenez in #4446
• Skills: one source, every surface by @chelojimenez in #4445
• Connector Bench (IN-1): the relay in front of the backend's benchmark routes by @chelojimenez in #4410
• feat(bench): the Connector Bench run worker and its eval children by @chelojimenez in #4415
• feat(bench): the conformance and auth-probe children by @chelojimenez in #4421
• bench: the score-site benchmark screens, driven by the run row by @chelojimenez in #4423
• feat(bench): write-manifest enforcement, the artifact ledger, and cleanup by @chelojimenez in #4425
• bench: a third insights scope, and a diagram that waits to be asked by @chelojimenez in #4426
• Keep synthetic loadSkill tool-call ids inside the provider's charset by @chelojimenez in #4451
• Carry a Cloudflare Access service token through the staging smoke lane by @chelojimenez in #4450
• Skills tab: an honest count, and rows that say only what distinguishes them by @chelojimenez in #4453
• Auto-connect by default, without the wall of red by @chelojimenez in #4447
• Revert "Auto-connect by default, without the wall of red" by @chelojimenez in #4456
• Keep Playwright out of the Electron main bundle (fixes the red macOS release build) by @chelojimenez in #4457

Full Changelog: v3.2.0...v3.3.0