Switch docker build telemetry from plain-text regex parsing to rawjson + metadata-file

[simonrosenberg]

Problem

The current docker build telemetry in openhands-agent-server/openhands/agent_server/docker/build.py is extracted by regex-parsing BuildKit plain-text progress output from docker buildx build. That works, but it is inherently tied to human-readable progress strings like exporting cache to, pushing layers, and exporting manifest, which is a fragile boundary for SDK telemetry.

Proposed change

Replace the regex-based plain-text parsing with Buildx/BuildKit’s structured outputs:

• pass --progress=rawjson to docker buildx build
• pass --metadata-file <tmpfile> to capture final build metadata
• parse JSON-line progress events for timing/progress telemetry
• parse the metadata file for final build metadata
• remove the regex parser rather than keeping a plain-text fallback

Why

• rawjson is the machine-oriented progress mode exposed by docker buildx build
• --metadata-file gives structured final result metadata
• this is a cleaner and more reliable telemetry boundary than parsing terminal-oriented text output

Scope

• update build_with_telemetry() to request rawjson progress and a metadata file
• replace _parse_buildkit_telemetry() with JSON-based parsing
• thread any useful metadata-file fields into the telemetry/result model where appropriate
• update tests to validate JSON progress parsing instead of regex matching plain-text stderr

Acceptance criteria

• no regex parsing of BuildKit plain-text progress remains in the telemetry path
• telemetry still includes the currently useful phase timings/counts where available
• final build metadata is captured from --metadata-file
• tests cover successful builds and failed builds under the new parsing path

References

• Docker buildx build progress modes, including rawjson: https://docs.docker.com/reference/cli/docker/buildx/build/
• Docker buildx build --metadata-file: https://docs.docker.com/reference/cli/docker/buildx/build/