rov-autonomy/CLAUDE.md

# CLAUDE.md — Rules for Claude on the Argonaut 3 Project
**Version:** 3.5
**Updated:** 8 July 2026

**Changelog 3.5:** Added `[VSC]` terminal label. Added SESSION LOG section + write triggers
(bridges Claude Code work back to claude.ai chat, which has no native access to Claude Code
sessions).

---

## PRIME DIRECTIVES — These override everything else

### 1. Research before acting. Always.
Before any technical decision, code change, or command:
- Read project docs (handover, UI design, DIR, this file) and search past chat history
- Read source code and official docs before acting; GitHub source/issues are more
  authoritative than official docs for Cockpit/BlueOS specifics
- Confirm facts from sources, not from memory

If uncertain: **search first, then act**. Never guess.

### 2. Time and tokens are finite and valuable
- **Time wasting is FORBIDDEN. Token wasting is FORBIDDEN.**
- Short replies. No preamble padding. To the point.
- Never repeat an approach that has already failed
- Never ask for information already in the project docs
- Never issue a command without knowing what it will do and why

### 3. Facts only. Never assumptions.
- Do not proceed based on what "should" work
- Do not assume port, routing, or API behaviour — measure it
- Do not infer system state — verify it before issuing commands
- If a claim cannot be verified from a source, say so explicitly
- Never present a prior session's confirmed state as the current running state

---

## Working Style

### One command at a time
Give one command, wait for feedback, then give the next. Never batch commands unless asked.

### Lead with context and corrections BEFORE instructions, never after.

### Copy-paste friendly commands
Commands inside terminal blocks must be immediately copy-pasteable and suitable for the
terminal in use (Git Bash on the laptop, bash on the Pis/server). Never include:
- Placeholders like `<YOUR_IP>` or `path/to/file`
- Hashed `#` comments inside the command block
- Options that require editing before running

### Code comments
When writing code (Python, config, scripts), always include full inline comments
explaining what each section does and why. This is a standing user preference.
NOTE: this applies to CODE files only — NOT to terminal command blocks, which must
stay clean and copy-paste ready with no inline comments.

### Terminal labels
Every command block is preceded by a plain-text terminal label above the block.
Grant pastes terminal output including the shell prompt — use these to identify
which device the output came from:

| Label | Shell prompt | Device | IP |
|---|---|---|---|
| `[SERVER]` | `grant@symbytech:~$` | Dell PowerEdge R210 | 192.168.1.175 |
| `[RPi5]` | `ubuntu@rov-brain:~$` | Raspberry Pi 5 autonomy brain | 192.168.1.101 |
| `[RPi4]` | `pi@blueos:~$` | Raspberry Pi 4 BlueOS gateway | 192.168.1.100 |
| `[LAPTOP — Git Bash]` | `grant@BB8 MINGW64 ...` | Windows field laptop BB8 | — |
| `[VSC]` | — | VS Code integrated terminal on BB8 (local Windows session at `C:\Users\grant\Desktop\Autonomous ROV System`) | — |
| `[BROWSER]` | — | Web browser on laptop | — |
| `[COCKPIT]` | — | Cockpit native desktop app on laptop | — |

`[VSC]` is a **local Windows session on BB8, NOT a Remote-SSH session to rov-brain**, unless
explicitly stated otherwise. Do not assume `[VSC]` == RPi5 working copy.

Never put the label inside the bash block.

> The BlueOS KVM VM (`blueos-vm`) was decommissioned once RPi4/RPi5 hardware became
> operational. `[VM]` is no longer a valid label. Any VM artefacts remaining on the
> server are for reference only.

### Verify before modifying
Always read a file before modifying it. Always check current state before changing it.
State what you observed before proposing a change.

### File operations
Read the exact file first, then `str_replace` with exact matching text. Remove old
versions before writing new ones. Complete files, not fragments — no partial code.

### No guessing at code
When writing code that interfaces with an external system:
1. Find the official documentation or source code
2. Find a working example from the official repo or community
3. Build on the confirmed example
Never invent API calls, message formats, or connection strings.

### Own mistakes once
Acknowledge an error once and fix it. No repeated apologies, no grovelling.

### KISS
Recommend the simple correct path, not a menu of options.

---

## CREATE-THEN-BACKUP (non-negotiable)
When any new component (device, drive, service, config, OS install) is created or
configured, a backup must be put in place immediately as part of that work — not
deferred. Backups and foundational architecture (storage, recovery, monitoring) are
prerequisites, prioritised BEFORE new feature/build work. The RPi4 SD failed with no
backup because this was deferred — do not repeat.

---

## Project-Specific Rules

### Cockpit DIY widgets
- Cockpit is **v1.18.2** (native desktop app mandatory — browser/Lite extension does
  NOT support DIY widgets)
- JSON format with `html`/`css`/`js` string fields
- CSS hard constraints: no `body {}`/`:root {}` global blocks, no `@keyframes`, no
  unicode in JS strings, no `position: fixed`, no global tag selectors
- No `async/await` in widget JS — use `.then()/.catch()`
- Use `var` not `let`/`const` in widget JS for compatibility
- Correct live-value getter: `window.cockpit.getDataLakeVariableData(variableId)`
  (CONFIRMED v1.18.2). `getDataLakeValue()` and `listenToDatalakeVariable()` do NOT
  exist in DIY widget scope. `getAllDataLakeVariablesInfo()` is metadata-only.
- Widget JSON files in the repo use underscore naming: `w6_altitude.json` (matches
  W0–W5). The Cockpit-loaded widget is stored by internal config, not filename, so the
  repo filename is the canonical reference.

### Data injection into Cockpit data lake
- ONLY working external-injection method is the Generic WebSocket Connections feature
  (Settings → General). Production path is the `cockpit_bridge` ROS2 node (rov_mission,
  WebSocket :9001).
- Message format: `variable-name=value` (one per send). Data lake key: `external/{name}`.
- ALWAYS use hyphens — Cockpit normalises underscores to hyphens in IDs.
- HTTP POST to mavlink2rest does NOT work for data lake injection. MAVLink UDP injection
  from external scripts does NOT work. Tested and confirmed — do not suggest them.
- Current data lake variables published by `cockpit_bridge` (19 total — 8 telemetry + 11 mode-profile flags):

| Variable | Source topic | Meaning |
|---|---|---|
| `external/rov-failsafe` | `/rov/failsafe` | 0=GREEN, 1=AMBER, 2=RED |
| `external/rov-depth` | `/rov/depth` | Depth in metres |
| `external/rov-voltage` | `/mavros/battery` | Battery voltage |
| `external/rov-heading` | `/mavros/mavros/compass_hdg` | Compass heading degrees |
| `external/rov-ms` | `/rov/mission/status` | Mission state integer |
| `external/rov-mp` | `/rov/mission/status` | Mission progress 0-100% |
| `external/rov-recording` | `/rov/recording/active` | 1=recording, 0=not recording |
| `external/rov-altitude` | `/ping2/altitude` | Ping2 altitude metres (-1.0=NO SONAR) |
| `external/rov-mode` | `/rov/mode/profile` | Active profile name: ROV / AUV / hybrid |
| `external/rov-mode-base` | `/rov/mode/profile` | Base mode: 0=ROV, 1=AUV |
| `external/rov-hilt` | `/rov/mode/profile` | HILT: 1=present, 0=absent |
| `external/rov-health-role` | `/rov/mode/profile` | W1: 0=advisory, 1=gate |
| `external/rov-rec-control` | `/rov/mode/profile` | Recording: 0=manual, 1=auto |
| `external/rov-nogo-gate` | `/rov/mode/profile` | No-go: 0=relaxed, 1=enforced |
| `external/rov-rec-reminder` | `/rov/mode/profile` | ARM reminder: 1=on, 0=off |
| `external/rov-return-mode` | `/rov/mode/profile` | Return: 0=operator-selectable, 1=autonomous |
| `external/rov-live-telem` | `/rov/mode/profile` | Telemetry: 0=on, 1=predive-only |
| `external/rov-nav-display` | `/rov/mode/profile` | Nav: 1=shown, 0=autonomy-only |
| `external/rov-camera-role` | `/rov/mode/profile` | Camera: 0=pilot-primary, 1=record-only |

### FastAPI backend (rov_api)
- `rov_api` package provides the HTTP-to-ROS2 bridge on port 8081. Runs as
  `argonaut-api.service` (manual-start in dev, same policy as argonaut.service).
- Architecture: uvicorn/FastAPI in the main thread, an embedded rclpy node
  (`RovApiNode`) spinning in a daemon background thread. HTTP handlers call thread-safe
  methods on the node. Do NOT call `rclpy.spin()` more than once.
- Core endpoints (Stage 2, live): `GET /health`, `POST /abort` (W3 RETURN TO SAFE,
  publishes Bool to `/rov/mission/abort`), `POST /mission/start` (MissionCommand START),
  `POST /mission/stop` (MissionCommand ABORT).
- Deferred to Stage 2b: `/return_budget` (W5), `/backup/status` + `/backup/delay` (W7,
  Option A design), `/mission/upload`.
- **W3 abort path is vehicle-layer:** publishes Bool(true) to `/rov/mission/abort`, which
  failsafe_monitor latches as `flag_manual_abort` → FSM NORMAL → HOLD_AND_RECOVER. This is
  DIR-5 return-to-safe, NOT the mission ABORT service. Do not confuse the two.
- **Python dependencies:** fastapi + uvicorn installed system-wide on RPi5 via
  `sudo pip3 install fastapi uvicorn --break-system-packages --ignore-installed typing_extensions`.
  The `--ignore-installed typing_extensions` flag is required — the Debian-managed
  typing_extensions has no pip RECORD file and blocks the install otherwise.

### Deploy files (version-controlled)
- systemd units and launch wrappers live outside the workspace but are version-controlled
  in the repo `deploy/` folder (copies; authoritative running copies at deploy locations).
- `deploy/argonaut.service` → `/etc/systemd/system/argonaut.service`
- `deploy/argonaut-stack.sh` → `/usr/local/bin/argonaut-stack.sh`
- `deploy/argonaut-api.service` → `/etc/systemd/system/argonaut-api.service`
- `deploy/argonaut-api.sh` → `/usr/local/bin/argonaut-api.sh`
- `deploy/README.md` documents locations, pip deps, and install procedure.
- `/home/ubuntu/ros2_ws` is a symlink to `/data/ros2_ws` — both wrappers rely on it.

### Recording architecture (DIR-9a — ARM-bounded black-box)
- Recording is a **single ARM-bounded tier** for black-box/debug/ML/fault-finding via
  `recording_manager`. **AUV:** starts on ARM, stops on DISARM. **ROV:** operator
  start/stop; the ARM edge fires a reminder ("armed, not recording"). Keyed to the
  `recording_control` flag, not the mode. (Supersedes the old continuous DIR-9 model.)
- **No-go coupling:** AUV `enforced` — recording fail → mission stops, vehicle recovered.
  ROV `relaxed` — recording fail after ARM → pilot alerted; operator resolves in-water
  or recovers.
- Certified deliverable capture is OUT of scope pending Class-approved devices; recorded
  data is dev/debug/ML only, never a client deliverable until then.
- The recorder manages a `ros2 bag record` subprocess — MCAP format, zstd compression,
  500MB bag splitting, 10GB free-space retention on `/data`.
- **`foxglove_bridge` is visualisation only — it does NOT record MCAP.** Correct this
  wherever the old "MCAP via foxglove_bridge" claim appears.
- Bags written to `/data/bags/dive_<timestamp>/` on NVMe.
- No camera in the topic set currently — recording ~2.5 MB/s (nav/MAVLink only).
- **Storage (no reliable in-field server backup):** primary lever is LOWER PUBLISH RATES
  on heavy streams (keep topics present for debug); ARM-bounded runs self-limit; retention
  cap tunable; NVMe expansion later. Topic/rate review is an open action.

### Operating-mode profiles (DIR-1/2, mode_profiles.yaml)
- One widget set; behaviour is set by flags from a mode profile. Dominant axis is HILT
  (tether present = ROV / live interfaces; absent = AUV / headless).
- **Source of truth:** `mode_profiles.yaml` in `rov_mission/config` — ROV/AUV defaults
  plus custom hybrid via "Save As". Hybrid is NOT a third mode; it is any saved profile
  with overrides.
- `mode_profile_loader` (rov_mission) resolves the active profile into a `ModeProfile`
  message (rov_interfaces), publishes it LATCHED on `/rov/mode/profile`, reloadable via
  the `/rov/mode/reload` (std_srvs/Trigger) service. Fail-safe: invalid flag values are
  rejected and the last-good profile is retained.
- `cockpit_bridge` surfaces the 11 flags to the data lake as `external/rov-*`. Widgets
  read the flags, NOT the mode.
- Flags: hilt, health_role (advisory|gate), recording_control (manual|auto), nogo_gate
  (relaxed|enforced), record_reminder_on_arm, return_mode (operator_selectable|autonomous),
  live_telemetry (on|predive_only), nav_display, camera_role (pilot_primary|record_only).
- **YAML gotcha:** YAML 1.1 parses bare `on`/`off`/`yes`/`no` as booleans — quote enum
  values like `"on"`. The loader also normalises bool→string as a guard.

### BlueOS / MAVLink router
- **MAVLinkRouter** is the confirmed router. Do NOT recommend MAVP2P — BlueOS itself
  warns "MAVP2P has been presenting issues and is not currently recommended."
- MAVP2P was tried and reverted. Do not suggest it again.
- **ardupilot-manager FD exhaustion — CORRECTED ROOT CAUSE (9 Jul 2026).** The earlier claim that this was "fixed permanently" by `/etc/docker/daemon.json` `nofile: 65536`, with root cause "UI polling TCP sockets," is WRONG on both counts. The `nofile` change only raised the ceiling (failure at ~10 min → ~11 h of cumulative exposure); it fixed nothing. The leaked FDs are `/dev/i2c-1` handles, NOT TCP sockets. Measured mechanism: BlueOS's Autopilot Firmware / board-config page polls `GET /v1.0/available_boards`; each request triggers 20 `flight_controller_detector.linux.detector:detect_boards` probes for Navigator / NavigatorPi4 / NavigatorPi5 / Argonot. Each Linux-board probe opens `/dev/i2c-1` and never closes it on failure. We run a Pixhawk 2.4.8 on USB serial (PlatformType.Serial), so every probe fails and every probe leaks. Measured leak rate ~100–108 FD/min with the board-config page open; exactly zero with it closed (4630 flat over 65 min, then +90 FD in 50 s on opening the page, flat again immediately on closing). Confirmed cause-and-effect, reversible. Note: FD count accumulates across sessions and only resets when `blueos-core` is recreated (core restart/reboot) — cumulative page-open time, not wall-clock, drives time-to-failure. At 65536 FDs, ardupilot-manager can no longer open new handles → AUTOPILOT_VEHICLE_TYPE_FETCH_FAIL, AUTOPILOT_ENDPOINT_FETCH_FAIL, BAG_OF_HOLDING_GET_DATA_FAIL, "No board running". Core restart recreates the container with a fresh FD table, which is why restarting always "fixed" it. This is an upstream BlueOS 1.4.3 bug; do NOT patch BlueOS (updates would revert it).
- **Mitigation (dev):** do not leave the BlueOS Autopilot Firmware / board-config page open. Open, act, close. Other BlueOS pages and Cockpit are unaffected.
- **Mitigation (field) — NOT YET IMPLEMENTED:** operator discipline is not a field-safe control. The exposure is launching with a nearly-exhausted FD table after pre-dive setup, not leakage during the dive itself. Required: (a) FD-count check on ardupilot_manager as part of the AUV pre-dive gate, blocking launch above a threshold; (b) failsafe_monitor must detect FC comms loss so W1 goes RED when the board is unreachable. See open items.

### Hardware facts (do not re-question)
- **RPi5 storage:** boots from SD card (`/dev/mmcblk0`). NVMe holds DATA ONLY (`/data` —
  ros2_ws, bags, logs). Backup images the SD for OS, rsyncs `/data` for data.
- **Flight controller:** Pixhawk 2.4.8 on `/dev/ttyACM0` is the active flight controller.
  SITL is NOT used in production. BlueOS `preferred_board` = Pixhawk1.
- **Tailscale on RPi5:** installed but stopped and disabled (`tailscaled.service` disabled
  2 Jul 2026). Re-enable with `sudo systemctl enable --now tailscaled` if remote access
  to RPi5 is needed.

### Hostname resolution (DNS)
- Hostnames `symbytech` (192.168.1.175), `blueos` (192.168.1.100), and `rov-brain`
  (192.168.1.101) are resolvable from the laptop (Windows hosts file) and RPi5
  (`/etc/hosts`). Grant prefers hostnames over IPs.
- Git remotes use the hostname form: `ssh://git@symbytech:222/administrator/rov-autonomy.git`
- If a "Could not resolve hostname symbytech" error appears on a machine, that machine is
  missing the hosts entries — add them rather than switching remotes to IPs.

### Backup system (built 30 Jun 2026)
- Push model: each Pi self-backs-up via root cron, pushes to server over SSH (no inbound
  access to Pis).
- RPi4: `/usr/local/bin/rov-brain-pi4-backup.sh`, 00:00 daily.
- RPi5: `/usr/local/bin/rov-brain-pi5-backup.sh`, 00:30 daily (staggered).
- Server monitor `check_rov_backup` alerts on FAILED (always) and OVERDUE (≥2 days,
  online or offline, throttled weekly).
- Field/ops backup model is operator-aware, never unattended mid-dive (DIR-11, W7).

### argonaut.service / argonaut-api.service (RPi5)
- Manual start is deliberate for both. Verify BlueOS heartbeat first, then
  `sudo systemctl start argonaut.service`. Do NOT enable auto-start until field-ready.
  Restate this rule each time it comes up.
- `argonaut-api.service` orders After=argonaut.service but does not Require it — the API
  can start independently for /health checks; mission commands return a clear
  "service unavailable" error if the stack is down.

### ROS2 code
- Check `rov_interfaces` message definitions before writing any node code
- Use `self.get_logger()` not `print()`
- Do not hardcode IPs, ports, or thresholds
- Do not deploy `rov_simulation` in production launch files
- Editing an installed launch/config requires `colcon build --packages-select <pkg>` —
  the running launch uses the installed copy under install/, not src/
- ARM64 Docker builds on the laptop only (buildx), never on the server

### Git workflow
- Commit to `master` on `administrator/rov-autonomy`
- Gitea SSH remote (hostname form): `ssh://git@symbytech:222/administrator/rov-autonomy.git`
- Both laptop and RPi5 resolve `symbytech` via hosts entries (see Hostname resolution).
- Widget JSON files are re-imported from the laptop (source of truth for widgets) and
  pushed to Gitea — never overwritten from chat history.
- UFW rules before starting new services on new ports.

### Security — deferred to pre-field hardening pass
- **UFW:** currently inactive on RPi5. Deferred because there is no online exposure
  (Tailscale disabled, no proxy rules to the Pis). When enabling: allow SSH (22),
  operator TCP (8765, 9001, 8081), and all traffic from LAN subnet 192.168.1.0/24
  (covers ROS2 DDS dynamic ports — do NOT try to pin them individually), then default
  deny incoming. Bundle with SSH password change. Trigger: before remote ops or field.
- **RPi5 SSH password:** weak dev password, password auth enabled. Change to strong
  password before field deployment.

---

## What Claude MUST do
- State confirmed facts before proposing an action; cite the source of any technical claim
- Acknowledge when something has already been tried and failed
- Provide complete, working code (fully commented) based on confirmed examples
- Verify system state before and after changes
- Update the handover doc at the end of each session (version-controlled, changelog)

## What Claude MUST NOT do
- Guess at how a system works and propose commands based on that guess
- Repeat a failed approach
- Issue commands whose outcome is uncertain
- Provide partial code requiring manual editing to work
- Assume a system is stable without checking
- Use `print()` in ROS2 nodes; hardcode addresses/ports/thresholds
- Build x86 Docker images for vehicle deployment
- Change the failsafe priority order without design review
- Present prior-session state as current running state
- Recommend MAVP2P as the MAVLink router
- Commit docs to Gitea — only code and README go to Gitea; docs live in Claude project knowledge

---

## Documentation discipline
- All project docs (handover, UI design, DIR, this file, sim/ML plan) are maintained as
  COMPLETE replacement files uploaded to Claude project knowledge — never patches or
  fragments, never committed to Gitea.
- Version numbers and changelogs maintained on every update.
- Only code and README files go to Gitea.

---

## SESSION LOG (bridges Claude Code → claude.ai chat)

**Why this exists:** Claude Code (VS Code) and the claude.ai chat are separate systems with
no shared memory. Claude Code reads this file at session start but the claude.ai chat cannot
see Claude Code's session history at all. This log is the bridge.

**Format:** `YYYY-MM-DD — one-line summary of decision/change (commit hash if applicable)`

**Claude Code MUST append an entry here at the end of any session where one of these
happened** — not for routine edits, refactors, or in-progress work:
- A design/architecture decision was made or changed (DIR-level, mode-flag behaviour,
  failsafe priority, recording behaviour)
- A hardware fact was confirmed or changed (wiring, storage layout, board config)
- A security-relevant change was made (SSH, UFW, credentials, exposure)
- A backup/recovery-relevant change was made
- A bug was root-caused and fixed (not just patched)
- A previously-deferred item was completed or its trigger condition changed
- Anything Grant would reasonably expect the claude.ai chat to already know next time he
  opens it

**Skip:** typo fixes, comment-only changes, WIP/incomplete work, anything already fully
captured by the git commit message with no extra context needed.

**Workflow:** Grant pastes new entries into the claude.ai chat when relevant; Claude folds
them into chat memory via the memory tool. This log is append-only — do not edit past
entries; correct forward with a new dated entry if something needs revising.

- 2026-07-05 — Widget housekeeping: added W6 altitude widget, removed probe diagnostics
  widget (`785ec56`)
- 2026-07-08 — rov_api: added GET /mode and POST /mode endpoints (validate-before-write +
  /rov/mode/reload trigger); deployed and tested on RPi5, all paths verified (9109439)
- 2026-07-08 — Root-caused chronic widget "no data": cockpit_bridge (feeds Cockpit data
  lake on :9001) launched but had no respawn, so a boot-time race left it permanently
  dead. Added respawn=True to its launch Node. W8 mode indicator built + validated live
  (render+data). W1/W8 confirmed GREEN/AUV once bridge up.
- 2026-07-08 — FLAGGED (open): recording_manager.py still implements superseded DIR-9
  (continuous-from-boot recording), NOT the governing DIR-9a (ARM-bounded black-box:
  AUV ARM-start/DISARM-stop, ROV operator manual + arm reminder). Design decision
  exists only in handover + CLAUDE.md; code never updated. foxglove_mcap.launch.py
  comments also still say DIR-9. Also: argonaut-stack.sh passes record:=${RECORD} but
  rov_full.launch.py declares no such arg, so RECORD=false is silently ignored and
  bags record regardless. Both deferred to a dedicated DIR-9a recorder-rework session
  — do NOT wire RECORD into the current DIR-9 model.
- 2026-07-08 — Added respawn=True/respawn_delay=5.0 to failsafe_monitor,
  motion_controller, ekf_filter_node, depth_node, mission_executor. Deliberately NOT
  applied to: mavros_node (respawn could mask/worsen the chronic FCU-link instability
  — belongs in that investigation), recording_manager (pending DIR-9a rework),
  camera_node/feature_detector (perception not operational yet), foxglove_bridge
  (viz-only). NOTE: failsafe_monitor respawn has a coverage gap during the 5s respawn
  delay — no failsafe assessment in that window; flagged for future design
  consideration.
- 2026-07-09 — Root-caused the recurring "No board running" / AUTOPILOT_*_FETCH_FAIL
  failure. NOT the previously-documented TCP-socket/UI-polling cause. Real cause:
  /dev/i2c-1 FD leak in ardupilot_manager's Linux board detector, driven by the BlueOS
  board-config page polling available_boards (20 probes/request, ~100-108 FD/min).
  Zero leak with page closed; confirmed reversible by controlled test (4630 flat 65min,
  +90 FD in 50s on open, flat on close). nofile=65536 only delayed failure to ~11h
  cumulative page-open time. Upstream BlueOS 1.4.3 bug — not patching. Corrected the
  stale root cause in this file.
- 2026-07-09 — FLAGGED (open, safety-critical): failsafe_monitor reports GREEN with the
  flight controller unreachable. W1 is the AUV pre-dive GO/NO-GO gate (health_role=gate)
  — it would pass a dive with a dead FC. failsafe_monitor must consume /mavros/state
  connection + heartbeat and drive W1 RED on FC comms loss. W1 status line should carry
  cause text (e.g. "Vehicle Disconnected", "Heartbeat Lost", "No Board Running"). Also
  flagged: no FD-count check on ardupilot_manager in the pre-dive gate.

---

## Key Reference Documents (current versions)

| Document | Purpose |
|---|---|
| `ROV_Project_Handover_v2_21.md` | Master reference — architecture, environment, commands, status board |
| `design_intent_register_DRAFT.md` | DIR-1…DIR-11 (+ DIR-9a) — owner/operator intent, source of truth for *why*; recording now ARM-bounded per DIR-9a |
| `ROV_Failsafe_Design_v2_0.md` | Failsafe state machine, sensor roadmap, priority order |
| `Argonaut3_UI_Design_v1_9.md` | Cockpit widget specs (W0–W7) and confirmed data paths |
| `Mission_and_Data-Recording_Lifecycle_for_AUV_ROV_Systems...md` | Lifecycle research backing DIR-9/10 |
| `Argonaut3_Simulation_ML_Tooling_Plan_v1_0.md` | Simulation + ML tooling: dev-PC tiers, cloud options/cost, training assets/datasets/models, PCIe Gen 3 decision |

> When referencing these, use the latest version present in the project folder — do not
> cite a version number from memory if a newer file exists.

---

## Safety Notice
This is a production-oriented autonomous safety system. All code must be treated as if it
will run on a real vehicle in open water with no operator present. Safety, correctness,
and reliability are non-negotiable. Shortcuts taken during development become risks during
field operations.