Inline local rollout images in PrimeMonitor samples

[mrmoxon]

Summary

• Convert local file:// image references in PrimeMonitor rollout samples back into data:image/...;base64,... URLs before writing the sample Parquet.
• Make offload_images_to_disk() emit absolute file URIs via Path.as_uri(), so relative output_dir configs do not produce malformed file://outputs/... URLs.
• Preserve the source image media type when offloading images, so JPEG/WebP/etc. inputs are not re-uploaded as data:image/png with mismatched bytes.
• Recursively handles image URLs in prompt, completion, and trajectory payloads without mutating the original rollout objects during monitor serialization.
• Adds a 2 MiB per-image inline cap so sample uploads do not accidentally absorb large local artifacts.
• Keeps the PR limited to production code changes; test-file changes were removed per reviewer request.

Why

offload_images_to_disk() rewrites inline multimodal images to local paths under the orchestrator output directory, e.g. file:///data/outputs/.../assets/images/.... That is fine inside the training pod, but those paths are later stored in PrimeMonitor sample Parquet. The hosted dashboard only sees that file:// pointer and cannot load the image from a browser.

Inlining small local images during PrimeMonitor serialization keeps future uploaded samples browser-renderable without adding a new backend artifact upload/signed URL path.

Validation

• uvx ruff check src/prime_rl/orchestrator/trajectories.py src/prime_rl/utils/monitor/prime.py
• uvx ruff format --check src/prime_rl/orchestrator/trajectories.py src/prime_rl/utils/monitor/prime.py
• uv run --no-sync python -m py_compile src/prime_rl/orchestrator/trajectories.py src/prime_rl/utils/monitor/prime.py
• Confirmed git diff --name-status origin/main...HEAD only lists production files under src/.
• Remote test machine root@95.133.252.95 at current head: ruff check/format and py_compile passed for touched production files.
• Remote test machine root@95.133.252.95: custom offload → PrimeMonitor Parquet smoke passed for relative output dirs, monitor non-mutation, completion image inlining, and >2 MiB image skip behavior.
• Remote test machine root@95.133.252.95: real-image MIME smoke verified PNG stays data:image/png and JPEG stays data:image/jpeg after offload → PrimeMonitor Parquet serialization.
• Remote test machine root@95.133.252.95: live platform E2E with image sample upload passed via real PrimeMonitor registration, sample presign, R2 PUT, confirm, and final summary; uploaded Parquet bytes were asserted to contain data:image/png;base64,... and no file:// reference. Run id: yg12nqyecvzfdcyaokiscgb2.
• Remote test machine root@95.133.252.95: realistic live platform E2E passed with run_config.max_steps=20 for 20 steps × 5 image rollouts/step. API verification showed maxSteps=20, stepsWithSamples=[0..19], total=100 samples, and sampled prompts contain data:image/png;base64,... with no file://. Run id: j66iit3cb0vlyb93v12kxhx6.
• Remote test machine root@95.133.252.95 at current head: real renderable PNG/JPEG live platform E2E passed with run_config.max_steps=10 for 10 steps × 4 image rollouts/step. API verification showed maxSteps=10, stepsWithSamples=[0..9], step 9 has 4 decoded valid 96×64 PNG/JPEG samples with matching data:image/png / data:image/jpeg media types and no file://. Run id: e9l2830u8xfydm0rh5oyp0gg.

---

Note

Low Risk
Changes are limited to image offload paths and PrimeMonitor sample serialization, with a 2 MiB inline cap and no training or auth logic touched.

Overview
Rollout image offloading now writes files with the correct extension from the source data:image/... media type (not always .png), dedupes by full on-disk path, and stores absolute file:// URIs via Path.resolve().as_uri() so relative output dirs do not produce broken references.

PrimeMonitor sample Parquet serialization recursively re-embeds eligible local file:// image URLs in prompt, completion, and trajectory JSON as data:image/...;base64,... (without mutating in-memory rollouts), with per-URL caching and a 2 MiB per-image inline cap so the hosted dashboard can render images while large artifacts stay as file pointers.

^{Reviewed by Cursor Bugbot for commit 27ec93d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3ccf1a6. Configure here.}