Repository navigation
feat(minimal): add scene capture, compact results and observation windows - #32
Merged
Merged
Conversation
Project model-facing state, feedback and calibration without changing execution, image attachments or stored evidence. Preserve recovery diagnostics and share the result view across all minimal integrations.
Bound historical tool images and paired calibration while preserving text, reasoning, action history and native compaction input. Keep the existing identity and sliding-window policies independent.
Keep per-camera image paths in compact results and aged history. Align snapshot access guidance, age already-windowed metadata, and handle nullable hold IK statistics without invalidating completed motion.
# Conflicts: # CHANGELOG.md # uni_robot_agent/tools/scene/provider.py
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Minimal manipulation needs fresh observations without dummy motion, readable robot geometry, and smaller model-facing feedback. This change makes its tool set
move,hold,capture_scene, andfinish, supporting calibrated RGB or RGB-only observations with depth disabled.--context-policy observation_window --context-observation-horizon 2. It ages recognized observation images and calibration together while preserving file references, action/state/feedback history, assistant text, reasoning and continuation summaries. Initial/reference images and images with unknown ownership remain protected; this is not a total-context token cap. Existing context policies remain available.--reasoning-effort noneby sendingenable_thinking=false. The adapter's non-thinking output limit is 8192 tokens, compared with 65536 with thinking enabled.Validation: package build/install and CLI configuration checks completed. RoboDojo RGB-only and calibrated 42-episode runs exercised the compact results and observation window; saved requests confirmed readable image references, calibration delivery and image eviction. Qwen requests also confirmed thinking disabled with zero reported reasoning tokens. These runs establish execution and observation behavior, not an improvement in task success or billed cost. Native coverage of other environments remains limited.