Skip to content

Latest commit

 

History

History
55 lines (38 loc) · 5.72 KB

PROJECTS.md

File metadata and controls

55 lines (38 loc) · 5.72 KB

Afloat16 Project Guide

English · 简体中文

Independent public tools and skills maintained by Afloat16. Choose by the problem you are solving, then follow the project's installation guide, examples and validation notes.

Geometry and reconstruction

Project What it helps you do Scope
ReconTrail Inspect photo quality, recover sparse geometry and camera poses, reconstruct calibrated stereo, and share offline interactive reports. CPU/local workflow; results depend on input, parallax and calibration. No general dense or watertight-mesh guarantee.
COLMAP Model Audit Check reciprocal tracks, view connectivity and exact-name overlap candidates in legacy COLMAP text models. No newer rig/frame validation; reprojection supports PINHOLE/SIMPLE_PINHOLE. Name overlap does not guarantee alignment or database compatibility.
Pose Convention Bridge Convert explicit w2c/c2w, OpenCV/OpenGL and COLMAP/TUM/JSON pose conventions; export Nerfstudio cameras. Does not infer extrinsics, intrinsics or scale. Apply the same world mapping to scene geometry separately.

Robotics diagnostics

Project What it helps you do Scope
Bag Timeline Audit Inspect ROS 2 SQLite/MCAP or JSONL timestamps for gaps, regressions, duplicates and declared same-clock topic offsets. A gap is not proof of loss; does not infer split-bag ordering or decode SQLite payload headers.
TF Route Audit Inspect exported TF or native URDF topology, required frames, conflicts and observed time windows. Does not simulate live tf2 cache/interpolation or discover publisher authority; movable URDF joints provide no temporal evidence.

Model interfaces and tool execution

Project What it helps you do Scope
LLM Handshake Preview and run bounded OpenAI-compatible Chat Completions endpoint probes; compare reports after changes. Defaults to a model-list GET; inference must be enabled. Not full protocol certification or a model-quality benchmark.
SSE Toolcall Audit Audit saved Chat Completions/Responses SSE tool-call assembly and compare semantic calls with saved non-stream responses. Offline; does not repair generation. Different stochastic completions do not automatically imply a regression.
Tooltrace Audit Inspect execution attempts or conversations for duplicates, overlaps, missing results and identity conflicts; optionally record argument digests. Caller supplies stable IDs and meaningful clocks. Does not implement idempotency, cancellation or scheduler fixes.
Jev MCP Ask four read-only advisory tools for bounded decisions, routing, risk signals and yes/no gates. Unofficial TypeSafe Jev integration; requires service access. State is sent to the external API; advice has no approval authority.

Research and engineering workflow

Project What it helps you do Scope
Research to Fix Find prior fixes and viable approaches, map evidence to local code, apply a small repair and compare the original symptom before/after. Search and coding tools come from the host; record validation does not prove a repair worked.
Debate Direction Turn ambiguous requirements into independent proposals, challenges, revisions, an issue ledger and testable next steps. Consensus is not proof. Native skill needs a host with child tasks; the CLI uses the selected model service.
AI Dev Steward / StewardCheck Track task-scoped Git changes, explicit checks and review evidence; use the complementary artifact/review skill when appropriate. CLI and skill are independent. Evidence does not replace human acceptance criteria; MIT applies only to tools/stewardcheck.
Agent Action Notifier Receive email/local desktop notifications, track waiting requests and retry failed sends. Notifications are not approvals; SMTP is opt-in and imported replies do not establish trusted identity.

Suggested starting points

  • Unfamiliar bug: Research to Fix; add LLM Handshake for endpoint integration symptoms.
  • Ambiguous design: Debate Direction, then validate the selected assumptions in code.
  • Robot data: Bag Timeline Audit for timing; TF Route Audit for exported frame graphs.
  • Model tool calls: SSE Toolcall Audit for saved responses; Tooltrace Audit for actual attempt records.
  • 3D workflow: ReconTrail for reconstruction, COLMAP Model Audit for text-model preflight, Pose Convention Bridge after establishing the camera conventions.

Evidence and community work

The six newer domain utilities were selected from 29 public issue threads and 336 comments, a purposive sample rather than a field-wide urgency ranking. They are experimental MIT projects with 67 local tests and 36 passing OS/Python CI combinations at the recorded release. See the linked runs and limitations in the report.

Outreach and directory submissions record actual PRs and issue replies. A submitted recommendation is not a listing, adoption claim or upstream endorsement.

For selected merged upstream contributions, visit Afloat16's profile.