Skip to content

DOGFOOD.md

Latest commit

 

History

History
125 lines (83 loc) · 5.68 KB

DOGFOOD.md

File metadata and controls

125 lines (83 loc) · 5.68 KB
Raw

Dogfooding BrightVision (agent-first)

Dogfood means BrightVision validates itself through the same stack agents use: superproject workspace, bright-vision-core-serve on :8741, cecli Session, optional local Ollama — without requiring a human to click Chat/Tasks in the desktop shell.

Humans and Cursor agents share one contract:

Who How
Agent / CI yarn dogfood:agent (or yarn dogfood:gate with optional DOGFOOD_LLM=1)
Human (optional) yarn tauri dev for native shell spot-checks only

The desktop app is not the definition of done; the automated gate is.

What “100% dogfood” means

Layer Requirement
Workspace Repo root (BrightVision/), not cecli/ alone — enforced in test_superproject_dogfood.py, integration e2e, LLM fixtures
Engine Vision HTTP + cecli (pytest, integration e2e, LLM lanes) — same paths as React → SSE
Model Optional for the default gate; DOGFOOD_LLM=1 runs real Ollama turns when reachable
Validation yarn dogfood:agent green; friction → failing test or ROADMAP.md row

Daily loop (no GUI)

cd /path/to/BrightVision
git submodule update --init --recursive cecli
source activate.sh
yarn install

yarn dogfood:agent
# Full bar + Ollama when running:
DOGFOOD_LLM=1 yarn dogfood:agent
# Slow superproject-root LLM lane:
DOGFOOD_LLM=1 DOGFOOD_SUPERPROJECT_LLM=1 yarn dogfood:agent

Agents (Cursor, headless scripts): run the same commands after substantive changes. Do not ask the user to open the UI unless a change is native-only (file picker, keychain, DMG).

What yarn dogfood:agent runs

  1. yarn dogfood:check — layout, verify:submodule, yarn test:fast, core superproject pytest when .venv exists, optional Ollama probe
  2. yarn dogfood:gate — test-local.sh release (mocked e2e + bright-core + integration e2e + verify)
  3. Optional LLM — when DOGFOOD_LLM=1 and Ollama is up: yarn test:llm:core + yarn test:e2e:llm (+ superproject lane when DOGFOOD_SUPERPROJECT_LLM=1)

Shorthand:

Command Use
yarn dogfood:check Fast preflight (~20s)
yarn dogfood:gate Full gate without LLM
yarn dogfood:agent Default — check + gate (alias for agent workflow)
yarn test:everything Max bar — check + release + fixtures + full LLM incl. superproject (see TESTING.md)

One-time machine setup

yarn dogfood:setup    # local-llm.env + M4-friendly Ollama pulls
# or: cp local-llm.env.example local-llm.env && ollama pull …

Ollama is only required for DOGFOOD_LLM=1. The default gate does not need a running model.

Automated coverage map (#19)

Concern Automated proof
Superproject + submodule paths test_superproject_dogfood.py, test_superproject_integration.py, yarn verify:submodule
Vision HTTP / todos / agent import yarn test:bright-core, yarn test:e2e:integration
UI contracts (mocked) yarn test:e2e, shipped-scenarios.spec.ts
Char-split agent todo → Tasks title agent-todo-char-split.spec.ts, integration/import-agent-plan.spec.ts
Real local LLM (opt-in) yarn test:llm:core, yarn test:e2e:llm

Manual GUI checklist: SUBMODULE_VERIFICATION.md — release spot-check only, not daily.

Agent self-dev scenarios

Copy into an agent session (headless or in-app chat). Prefer driving POST /sessions + SSE when not using the GUI.

Smoke (fast tier, read-only):

Read bright_vision_core/session.py and summarize how Session.create sets fnames for the workspace root.

Context + edit (superproject):

/add bright_vision_core/http_api.py — add a one-line docstring that this is the BrightVision Vision HTTP API. Emit SEARCH/REPLACE only.

Tasks + spec (core HTTP):

Create a workspace todo via API or Tasks, run generate-spec, implement the first implementation task only — assert .cecli/todos.json and spec files on disk.

Post-/agent todo sync:

After a turn that calls UpdateTodoList, POST /workspaces/todos/import-agent-plan must yield a real title (not [ from char-split JSON). See e2e/integration/import-agent-plan.spec.ts.

Friction → tests or roadmap

When automated dogfood fails or an agent hits a new blocker:

  1. Add a repro — pytest, integration spec, or mocked e2e (preferred).
  2. Or add a row to ROADMAP.md with workspace path, file path, model, expected vs actual, layer (UI / Vision HTTP / cecli / Ollama).

Watch list: wrong workspace root, proposed vs applied edits, commit in wrong repo, stuck Connecting (:8741), asyncio “Task was destroyed” after Stop, char-split UpdateTodoList titles.

Optional: desktop spot-check

Use when you change Tauri-only behavior (tray, keychain, native apply, file dialogs, resource overlay):

yarn tauri dev

In the app: project = repo root → Terminal → Local LLM → Start → Terminal → Start. Spot-check SUBMODULE_VERIFICATION.md A–D before a release announcement — not before every merge.

Local LLM on Apple Silicon

See LOCAL_LLM.md. Fast + heavy hopper + model router for dogfood under VRAM pressure.

Related