### Fixes

- Anthropic: reject uppercase provider-prefixed forward-compat model ids locally instead of sending malformed dynamic ids upstream. Fixes #73715.

- OpenAI/embeddings: pass configured output dimensionality through single and batched embedding requests so memory embedding indexes can request smaller vectors. Fixes #55126.

- CLI/infer: normalize HEIC/HEIF image files to JPEG before model-run requests, avoiding providers that reject Apple image container formats. Fixes #50081.

- OpenRouter: keep the default `openrouter/auto` model ref canonical while preventing TUI and Control UI catalog pickers from displaying or submitting `openrouter/openrouter/auto`. Fixes #62655.

- Exec approvals/node: let trusted backend node invokes complete no-device Control UI approvals after the original request connection changes, while keeping node, command, cwd, env, and allow-once replay bindings enforced. Fixes #78569. Thanks @naturedogdog.

- Agents/subagents: keep background completion delivery on the requester-agent handoff/queue-retry path instead of raw-sending child results directly, and strip child-result wrapper or OpenClaw runtime-context scaffolding from queued outbound retries. Fixes #78531. Thanks @EthanSK.

- CLI/completion: guard the shell-profile source line written by `openclaw completion --install` with a file existence check (`[ -f ... ] && source ...` for bash/zsh, `test -f ...; and source ...` for fish) so uninstalling OpenClaw no longer makes new login shells error on a missing completion cache. (#78659) Thanks @sjf.