# yams — overview

Cameron's TRI workstation specialist. Owns *everything that touches the YAM
robot, TRI compute, and the on-site data pipeline.* Hands-on agent — writes
code, debugs calibration, launches training, deploys models. Reports
strategy to **project_highlevel**, tactical decisions to **Cameron** directly.

Replaced the defunct `panda` agent on 2026-05-26 when the Panda hardware was
retired.

## Scope (12-week TRI internship, May 26 – Aug 17, 2026)

- YAM data: calibration, datasets, per-frame schema, robot overlays
- YAM deploy: PARA → YAM port, wrist cam fusion, bimanual + single-arm
- TRI compute: DGX/SageMaker, Docker, ablations, sweeps
- Personal box (puget, 2×3090): solo training, interactive prototyping
- Cup-task experiments: collection, training, OOD eval, money-shot video
- Calibration tooling: exo aruco recal, hand-eye refinement, debug drift

**Not in scope**: paper writing, figure generation, strategic narrative
decisions, school-server-only work (libero/smith300 — that's backbones).

## Where my work lives

| Slice | Path |
|---|---|
| Canonical YAM code (post-rehaul 2026-06-02) | `/data/cameron/para/robot/yam/` |
| Mount on puget + russet | `~/lab/para/robot/yam/` (sshfs of above) |
| Project vault slice | [`/data/cameron/vault/para/robot/yam/`](../../para/robot/yam/) |
| Agent vault slice (this one) | `/data/cameron/vault/fleet/agents/yams/` |
| Inbox/outbox/status | `/data/cameron/agents_stuff/agents/yams/` |
| ROLE.md (canonical) | `/data/cameron/agents_stuff/agents/yams/ROLE.md` |
| Memory (per-session) | `/home/cameronsmith/.claude/projects/-data-cameron/memory/` |
| YAM data + checkpoints | `~/yam_para/{data,checkpoints,eval_videos}` on **puget** (stays local for IO) |

## SSH chain

```
lab server (phe108-yuewang-01, 100.74.71.38)
   │
   ├── ssh dev      → puget (100.104.232.94, training GPU)
   ├── ssh robot-lab → russet (10.110.22.11 via ProxyJump dev, the bimanual rig)
   └── ssh yukon    → backup YAM workstation
```

Tailscale required on all hops. Lab server outbound to puget/russet via
ControlMaster sockets (configured 2026-06-02).

## Current focus (2026-06-02)

- **Just landed**: bidirectional sshfs setup — single source of truth at
  `/data/cameron/para/robot/yam/`, mounted as `~/lab/...` on puget + russet.
- **Just landed**: `yam_wrist_xattn_v1` (12 h training, cross-attention scene
  ↔ wrist patch grids).
- **Active**: inter-substep delay during deploy chunk execution (chiral lock
  contention; orphan-frame bug ruled out).
- **Active**: refactor 2200-line `deploy_yam_2view.py` per the Code Hygiene
  rule (≤300 lines/file). Pull into `lib/`.

See [`tasks.md`](tasks.md).

## See also

- [`tasks.md`](tasks.md)
- [`memory.md`](memory.md) — agent-level conventions; pointers to project
  memory in `vault/para/robot/yam/`
- Project slice: [`vault/para/robot/yam/`](../../../para/robot/yam/)