buildanything 1.5.0 → 1.7.0

package/commands/add-feature.md

@@ -0,0 +1,98 @@
+---
+description: "Add a single feature to an existing project — lightweight build cycle using existing architecture, design system, and CLAUDE.md context"
+argument-hint: "Describe the feature to add. --autonomous for unattended mode."
+---
+
+<HARD-GATE>
+YOU ARE AN ORCHESTRATOR. YOU COORDINATE AGENTS. YOU DO NOT WRITE CODE.
+
+"Launch an agent" = call the Agent tool. For implementation agents, set mode: "bypassPermissions". For parallel work, put multiple Agent tool calls in ONE message.
+</HARD-GATE>
+
+Input: $ARGUMENTS
+
+If the input contains `--autonomous` or `--auto`, skip user approval gates and log decisions to `docs/plans/build-log.md`.
+
+---
+
+## Phase 1: Context Gathering
+
+Read these files directly (no agent needed — this is fast):
+
+1. `CLAUDE.md` — product context, tech stack, rules
+2. `docs/plans/architecture.md` — current architecture
+3. `docs/plans/sprint-tasks.md` — existing user journeys and scope
+
+If any file is missing, proceed with what exists. If the codebase is unfamiliar or the feature touches unknown areas, spawn an Explore agent:
+
+Call the Agent tool — description: "Explore codebase for [feature area]" — prompt: "Find all files related to [feature area]. Report: directory structure, key files, patterns used, relevant components/routes/APIs. Be concise."
+
+---
+
+## Phase 2: Plan the Feature
+
+You do this yourself — no agent needed.
+
+1. **Break the feature into 1-5 tasks** (most features are 1-3). Each task should be one commit-sized unit of work.
+2. **Define behavioral acceptance criteria** for each task — what must be true when the task is done.
+3. **Define the user journey** — the end-to-end flow the user will experience with this feature.
+4. **Present the plan to the user for approval.** In autonomous mode, log the plan to `docs/plans/build-log.md` and proceed.
+
+---
+
+## Phase 3: Build
+
+**For EACH task:**
+
+### Step 3.1 — Implement
+
+Call the Agent tool — description: "[task name]" — mode: "bypassPermissions" — prompt: "TASK: [task description + acceptance criteria]. HANDOFF — Architecture context: [paste ONLY the relevant section from architecture.md]. Style guide: the living style guide at /design-system shows component styling — match it. Implement with real code and tests. Commit: 'feat: [task]'. Report what you built, files changed, and test results."
+
+Set `[COMPLEXITY: S/M/L]` based on task scope.
+
+### Step 3.2 — Cleanup
+
+Skip if trivial (< 20 lines, single file). Otherwise:
+
+Call the Agent tool — description: "Cleanup [task name]" — mode: "bypassPermissions" — prompt: "Clean up these files: [list from implementation]. Fix: naming, dead code, unused imports, style, DRY. Do NOT add features, change architecture, or touch other files. If cleanup breaks acceptance criteria, revert."
+
+### Step 3.3 — Smoke Test
+
+Skip if this task has no UI surface. Otherwise run the Smoke Test Protocol (`protocols/smoke-test.md`): open the affected route, execute behavioral acceptance criteria via agent-browser, collect evidence. On FAIL: spawn fix agent with evidence. Max 2 fix-and-retest cycles.
+
+### Step 3.4 — Verification
+
+Run the Verification Protocol (`protocols/verify.md`). All 7 checks. If FAIL, fix before starting the next task.
+
+---
+
+## Phase 4: End-to-End Verification
+
+### Step 4.1 — Run the User Journey
+
+Call the Agent tool — description: "E2E: [feature name]" — mode: "bypassPermissions" — prompt: "Verify the full user journey for [feature name]: [paste the user journey from Phase 2]. Use agent-browser to walk through each step. For each step: interact, verify the expected outcome, capture evidence. Report PASS/FAIL per step with screenshots."
+
+### Step 4.2 — Dogfood Affected Pages
+
+Call the Agent tool — description: "Dogfood [feature area]" — prompt: "Open every page affected by [feature name]. Check for: broken layouts, console errors, missing data, dead links, regressions. Report issues with screenshots."
+
+### Step 4.3 — Fix Loop
+
+If issues found in 4.1 or 4.2: spawn a fix agent with the evidence. Re-run the failing check. Max 2 fix-and-retest cycles. After 2 failures:
+- **Interactive:** present evidence to the user.
+- **Autonomous:** log to `docs/plans/build-log.md` and proceed with a warning.
+
+---
+
+## Phase 5: Done
+
+Report to the user:
+
+```
+FEATURE COMPLETE: [feature name]
+Tasks: [done]/[total] | Tests: [count] passing
+User journey: PASS/FAIL
+Evidence: [paths to screenshots/logs]
+```
+
+If the feature expands the product scope, update `CLAUDE.md` to reflect the new capability.