What's Changed
- Answer a stale draft and a malformed skills project id as 4xx by @ignaciojimenezr in #5832
- Evals: pick reasoning efforts when running a single test case by @ignaciojimenezr in #5877
- Fix/abort signal fallback by @ignaciojimenezr in #5872
- Refuse unattended approval hosts up front; show a reset when a harness runtime changed by @chelojimenez in #5874
- Docs: harness tool approval refusal + runtime-changed chat notice by @mintlify[bot] in #5880
- Swarms: show the model a goal card ran by @ignaciojimenezr in #5879
- Run Codex on the app-server adapter everywhere; remove the exec transport by @chelojimenez in #5875
- Keep the app up when a billing or notification read fails by @chelojimenez in #5882
- Bump harness adapters to the 1.0.117 core line (supersedes #5769) by @chelojimenez in #5843
- deps: clear the open Oneleet dependency-scanning findings (undici, provider-utils, x/sys and friends) by @olartgabo in #5823
- Record eval trials our infrastructure failed instead of scoring them by @chelojimenez in #5893
- mcp: signed-in only by default, and hold beta tools off the catalog by @chelojimenez in #5805
- Evals: rerun a run's failed cases over the API and CLI; keep subset reruns out of client aggregates by @chelojimenez in #5890
- Bake the hosted harness runtimes and run them in Docker CI by @chelojimenez in #5891
- Drop stream chunks the browser would reject, and report the ones it does by @chelojimenez in #5906
- Remove dead trace repair (Auto fix) by @chelojimenez in #5907
- fix: repair upstream runtime pack fingerprint snapshot by @ozboms in #5903
- Move the inspector, SDK and chat-ui to AI SDK 7 by @chelojimenez in #5912
- Pin published local harness runtime packs 1.0.1 by @ozboms in #5911
- Swarms: credits dialog only on real exhaustion, bounded host output, busy-reservation retry by @chelojimenez in #5717
- Fix image-history and disabled-effort assertions in CI by @ozboms in #5918
- fix: hide composer reasoning picker in multi-model mode by @ignaciojimenezr in #5909
- Pin hosted pnpm to 12.8.1 so the harness bake fits a 1 GiB box by @chelojimenez in #5913
- Keep the feature-unavailable denial machine-readable by @chelojimenez in #5923
- Local harness runtimes: dependable updates (PRs 1–6 and acceptance tests) by @chelojimenez in #5914
- evals: "Where it runs" picker for suites without environments by @ignaciojimenezr in #5905
- Pack workflow: generate the Codex bridge bundle before the Inspector layer by @chelojimenez in #5929
- CI: stop Dependabot rebases from flooding the runner queue by @chelojimenez in #5934
- mcp: stop served skills from naming held tools by @chelojimenez in #5936
- fix(sdk): use version-specific logging conformance checks by @ignaciojimenezr in #5865
- Support durable swarm Stop requests in runners, API, and UI by @ignaciojimenezr in #5922
- Arbitrary-port Inspector: verified guest authority, per-instance sessions, resolved runtime config by @chelojimenez in #5821
- Bump the example-ext-apps group across 4 directories with 1 update by @dependabot[bot] in #5845
- chore: release changeset for inspector, CLI, and SDK by @ignaciojimenezr in #5946
- Conformance timing: launch the bridge through the Inspector layer by @chelojimenez in #5945
- chore(local-harness): pin codex pack 1.0.2 by @chelojimenez in #5943
- Publish local runtime packs as part of starting a release by @chelojimenez in #5939
- Evaluate: remember the last run's iteration count in Setup Run by @nachocossio in #5937
- Keep hosted harness bridges alive between turns by @chelojimenez in #5942
- Pack pipeline: retry GitHub reads on a transient 5xx by @chelojimenez in #5955
- Turn off Claude Code background tasks in harness turns by @chelojimenez in #5944
- Require annotations.title in Claude readiness; set it on our MCP server by @chelojimenez in #5941
- Relay API key management to the hosted app from local builds by @chelojimenez in #5957
- Accept MCP audience for background run auth by @chelojimenez in #5959
- Hosted MCP: directory-ready hints and descriptions by @chelojimenez in #5950
- docs: MCPJam in Claude (hosted MCP connector) by @chelojimenez in #5951
- Block a second suite rerun while servers connect by @nachocossio in #5901
- Refuse eval launches whose server is not connected before charging by @nachocossio in #5896
- chore: patch release inspector, CLI, and SDK by @ignaciojimenezr in #5969
- Fix inspector image build: call the Playwright CLI by path by @nachocossio in #5967
- Fix three server tests left stale by #5957 and #5959 by @nachocossio in #5970
- Hand a keyed eval retry its existing run when the server preflight fails by @nachocossio in #5899
- chore(release): version packages (3.14.0) by @chelojimenez in #5977
- Pull request for mintlify/docs-update-pr-5977-1791321666924 by @mintlify[bot] in #5978
- Build local runtime packs in the background, ahead of the release by @chelojimenez in #5979
- Layer digests: generate the Codex bridge bundle first (fixes release contract) by @chelojimenez in #5980
- Skip forge's unused renderer build when packaging the desktop app by @chelojimenez in #5981
- release: wait up to 60 min for the backend prod deploy by @chelojimenez in #5982
Full Changelog: v3.13.0...v3.14.0