Skip to content

AGENT_LOG.md

Latest commit

 

History

History
117 lines (102 loc) · 20.1 KB

AGENT_LOG.md

File metadata and controls

117 lines (102 loc) · 20.1 KB
Raw

AGENT_LOG

A log of human–AI collaboration on DevYoga. Short one-line entries grouped by date. Maintained by the agent-log skill (proactively) or by hand.

2026-06-07

  • plan Chose the stack: NestJS + SQLite (CRUD) backend, React Native Expo frontend, AI features later.
  • scaffold Agent created the agent-log project skill to maintain this log.
  • decision Log entries are English, one-line, tagged by interaction type; agent work and manual user edits are distinguished.
  • decision Made the agent-log skill proactive-by-default — it auto-logs in the background after agent interactions.
  • scaffold Added a Stop hook (.claude/hooks/agent-log-checkpoint.sh + .claude/settings.json) that nudges a log entry only on a real checkpoint, not every turn.
  • chore Initialized the git repo with a .gitignore and made the initial commit via the conventional-commit skill.
  • edit Added an auto-commit step to the agent-log skill: after appending entries it commits only AGENT_LOG.md.
  • decision Chose a flat backend/ + mobile/ monorepo layout — minimal ceremony over workspaces; avoids Expo/Metro friction for the test task.
  • scaffold Initialized NestJS 11 backend in backend/ via Nest CLI (npm, strict TS, no nested git).
  • decision Adopted Context7 (ctx7/find-docs) proactively for library/API docs, codegen, and setup — no need to ask each time.
  • decision Chose TypeORM + better-sqlite3 with synchronize (dev-only, env-gated) over migrations for the first DB iteration.
  • scaffold Added the DB layer: TypeOrmModule.forRootAsync + ConfigModule, plus a Task entity and TasksModule CRUD.
  • chore Wired a global ValidationPipe (whitelist + transform) and .env/.env.example for DB config.
  • review Verified end-to-end: build clean, CRUD smoke test passes, validation returns 400, baseline tests green.
  • edit Refined the agent-log skill: never log the commit/revert action itself — log the work it captures.
  • decision Chose to use Expo plugin/skills for further mobile app development.
  • scaffold Created backend/scripts/seed.js — bare better-sqlite3 seed with 7 sample tasks across all statuses; added npm run seed and seed:clean scripts.
  • config Added Postman MCP server (postman-mcp-server) to global Claude Code config via claude mcp add.
  • scaffold Created "DevYoga API" Postman collection in the DevYoga workspace with a GET /tasks "Get All Tasks" request; saved workspace ID to project memory.
  • scaffold Initialized Expo SDK 56 mobile app in mobile/ (Expo Router, TS): NativeTabs (Tasks + Chat) with per-tab Stacks, task list/detail/create-as-formSheet screens, typed fetch helpers to the backend; verified via tsc + iOS export.
  • review Agent reviewed initialized mobile code against vercel-react-native-skills; identified list perf, missing StyleSheet.create, and unstable callback issues.
  • edit Applied review fixes: memoized TaskCard, extracted TaskItem + EmptyState from FlatList renderItem, hoisted all styles to StyleSheet.create, wrapped onSubmit/handleSubmit/status handlers in useCallback; lint clean.
  • edit Removed 11 Expo template leftover assets (react-logo, expo-badge, tabIcons, tutorial-web, logo-glow); kept only the 9 files referenced in app.json.
  • decision Switched mobile off Expo Go to a local dev build (expo-dev-client) — needed for native modules (@expo/ui, expo-glass-effect); kept it local-only due to Apple Developer account constraints.
  • scaffold Added expo-dev-client (~56.0.19) and built/ran the iOS Simulator dev build via expo run:ios; verified the app + dev-launcher render on the booted iPhone 17 sim.
  • edit Pointed npm start at expo start --dev-client and rewrote the README around the local dev-build flow (free-Apple-ID iPhone 7-day caveat, Android steps).
  • manual Flattened the tab route groups — (tasks)/(chat) → index/chat folders — and added the iOS bundle id + Apple team to app.json; committed separately from the dev-build switch.
  • feature Made the server address env-configurable for LAN access from a real iPhone: backend binds to HOST in app.listen, mobile reads EXPO_PUBLIC_API_BASE (localhost fallback); added .env/.env.example both sides and gitignored mobile/.env.

2026-06-08

  • feature Added priority (low/medium/high, default medium) to Task across backend entity/DTO and mobile types/form/card; generalized form's StatusButton into a generic OptionButton<V> to host both selectors.
  • feature Planned and implemented optional ?status= filter for GET /tasks: new FindTasksQueryDto, conditional TypeORM where, updated controller + service; mobile getTasks(status?) + StatusFilter chip row wired to index screen; backend committed, mobile UI pending David's edits.
  • edit Reworked mobile status filter per David: dropped the body chip row for a native @expo/ui Menu bar-button in the header (funnel SF Symbol, fills when active, checkmark on selected) placed before the ＋.
  • fix Replaced the header ＋ text glyph with a native SwiftUI Button(systemImage="plus") so both header buttons share one native style, tint, and vertical alignment.
  • fix Plus button rendered empty — a systemImage-only Button needs labelStyle('iconOnly') (with a real label kept for accessibility) to show just the icon.
  • feature Moved add-task out of the header into a FAB: a circular borderedProminent SwiftUI Button pinned bottom-right above the tab bar; header now keeps only the filter.
  • feature Added sorting to GET /tasks: sortBy (createdAt|priority) + sortOrder (asc|desc) enums on FindTasksQueryDto; service switched to QueryBuilder with a CASE rank for priority (low<medium<high) and a createdAt tie-break, default unchanged (createdAt DESC). Backend only.
  • fix Centered the filter icon in its header glass button — the Menu's empty label="" baseline-shifted the SF Symbol up/left; switched to a real label + labelStyle('iconOnly') (verified on the simulator).
  • test Added a minimal TasksService unit spec (mock repo via getRepositoryToken, chainable QueryBuilder stub): covers create, findOne found/missing, remove-missing, and findAll status filter + priority CASE ordering; 6 tests green.
  • feature Added a Sort menu to the tasks header beside the filter: SwiftUI Menu with "Sort by" (Date/Priority) and "Order" (Descending/Ascending) sections, filled icon when non-default; getTasks now sends sortBy/sortOrder.
  • edit Increased header icon size (font({ size: 22 })) and added tap-area padding (padding({ all: 4 })) to both SortMenu and StatusFilter.
  • edit Changed default sort to priority DESC in backend DTO, mobile index state, and SortMenu's DEFAULT_SORT_BY constant.
  • edit Moved priority and status badges to the bottom of TaskCard — priority left, status right via justifyContent: 'space-between'.
  • scaffold Added a root CLAUDE.md — high-level orientation for new sessions, centered on the north star (easing developers' task-triage pain with AI as the amplifier), plus stack/layout, data model & API, run commands, and conventions.

2026-06-09

  • edit Added AppLightTheme/AppDarkTheme to theme.ts derived from Colors palette; wired into root ThemeProvider so nav headers and screen backgrounds match the app palette; cleared task-detail dynamic title; removed redundant Pressable wrapper from TaskCard; fixed Link.MenuAction child syntax.
  • scaffold Installed Redux Toolkit, react-redux, redux-persist, react-native-mmkv (v4) for offline-first state management; created src/store/mmkv.ts MMKV-backed redux-persist storage adapter.
  • scaffold Implemented steps 3–6 of the state-management plan: tasks-api.ts (RTK Query endpoints + tag invalidation + rehydration), filters-slice.ts (persisted filter/sort state), store/index.ts (persistReducer + MMKV + setupListeners), and _layout.tsx wired with <Provider> + <PersistGate>.
  • edit Migrated all three screens (list, detail, create) off imperative fetch/useState onto RTK Query hooks; Redux filters slice drives list query; tag invalidation replaces focus-refetch hack; removed all imperative functions from utils/api.ts, keeping only types.
  • edit Extracted delete-confirmation Alert into shared useDeleteConfirm hook; list screen now shows a confirmation dialog before deleting (previously deleted immediately with no prompt).
  • decision Brainstormed the AI layer and converged on a single agentic chat (no separate "Today" screen): stateless backend (messages[] round-tripped, no conversation/memory tables in v1), human-in-the-loop proactive task creation with clarifying questions + "too big → propose split", batch creation via a create_tasks tool, staged bottom-up (backend then frontend); chose the Vercel AI SDK over LangChain (lighter, declarative zod tools, agent loop via stopWhen: stepCountIs, tool-without-execute for HITL) and over raw SDK.
  • plan Finalized the staged AI-layer plan (chat on Vercel AI SDK): Stage 0 backend POST /ai/chat with a create_tasks tool (zod batch input, execute → TasksService) and a system-prompt gate (clarify when vague, propose flat split when too big, never create before explicit confirm); Stage 1 mobile chat; Stage 2 more tools (list/prioritize/update/delete/status); Stage 3 optional persistence + read-only memory. Decided: too-big split = flat separate tasks (no schema change), confirmation round-trips through /ai/chat, stateless backend, mock mode when no API key.
  • scaffold Built AI-layer Stage 0: AiModule with POST /ai/chat, ChatRequestDto (validated by the global pipe), and ChatAgentService using the Vercel AI SDK (generateText + create_tasks zod tool with execute, stopWhen: stepCountIs(5)) on @ai-sdk/anthropic; reuses TasksService via DI (TasksModule now exports it).
  • decision Put the agent's intelligence in the system prompt (clarify/decompose/confirm), not in code — the never-create-before-explicit-"yes" gate lives there too; default model claude-sonnet-4-6 (env-overridable via AI_MODEL) for fast/cheap interactive triage.
  • scaffold Added mock mode: with no ANTHROPIC_API_KEY the agent returns canonical draft→confirm→create responses so the server boots and the flow is demoable without credentials; added ANTHROPIC_API_KEY + AI_MODEL to .env.example.
  • review Verified Stage 0 live via curl: vague request → textual draft (no tasks created) → "yes" → create_tasks fires → task persisted and visible in GET /tasks; plus validation 400s (empty/bad-role messages) and mock mode both pass.
  • scaffold Stage 1 frontend chat: chat-api.ts RTK Query slice (sendChat mutation, cross-slice Task.LIST invalidation via onQueryStarted/tasksApi.util.invalidateTags), chat types in utils/api.ts, store wired; full ChatScreen with useReducer transcript, @expo/ui/swift-ui TextField input bar, quick-reply chips ("Create a task", "Plan my day"), user/assistant message bubbles, and error handling.
  • decision Used RTK Query mutation (not @ai-sdk/react useChat) for Stage 1 — the backend returns plain JSON (generateText, no streaming), and RTK Query gives free cross-tab invalidation without a backend rewrite.
  • manual Removed Stop hook (agent-log-checkpoint.sh) from .claude/settings.json — hook was auto-logging AI collaboration on session end.
  • scaffold Moved the chat transcript off local useReducer onto a persisted Redux chat-slice (appendMessage/clearChat/selectMessages, MMKV-whitelisted) so conversations survive app restarts.
  • fix Chat list rendered in wrong order / native large-title fade looked broken: replaced the inverted FlatList (its scaleY(-1) transform flipped the iOS large-title scroll-edge fade) with a bottom-pinned non-inverted list — flexGrow:1 + justifyContent:'flex-end' content container, natural message order, auto-scrollToEnd on content growth.

2026-06-10

  • fix Chat input bar hid behind the keyboard: set keyboardVerticalOffset to the native Stack header height (via useHeaderHeight from expo-router's bundled @react-navigation/elements, not a standalone install — that would spawn a second HeaderHeightContext returning 0).
  • scaffold Added a destructive trash button to the Chat header (headerRight via Stack.Screen) that confirms via Alert then dispatches the existing clearChat action; hidden when the transcript is empty. Reused the @expo/ui/swift-ui Button/Host header-button pattern from the Tasks tab.

2026-06-11

  • review Agent reviewed backend AI architecture against David's wishlist (more tools, quick replies, today-plan, status, memory): verdict — current shape holds; everything is a new tool on the one agent, not new agents/infra.
  • plan Agreed the decomposition: split chat-agent.service.ts into orchestration + AgentToolsService (per-request tool factory closing over a ChatTurnEffects collector) + prompt file + mock service; quick replies via a no-op suggest_quick_replies tool, no second LLM call.
  • decision Prioritized remaining AI work 1–6: decompose → list_tasks+today-plan (meets the ≥2-features bar) → update/delete tools → quick replies → status generator → memory (first to cut); explicitly descoped on-device AI (key exposure) and update-hooks auto-reprioritization.
  • scaffold Agent wrote plans/ (index + 6 staged plan docs, 01–06) — each with goal, steps, done-criteria, effort, and cut order.
  • edit Executed plan 01: decomposed the AI module into orchestration-only ChatAgentService, AgentToolsService.build(effects) per-request tool factory, chat-turn.types.ts, prompts/system-prompt.ts, and MockAgentService — pure refactor, verified live draft→confirm→create and mock mode end-to-end.
  • scaffold Executed plan 02: added read-only list_tasks tool (compact shape, truncated descriptions) and a PLAN-THE-DAY prompt section weighing priority + status + age; verified live — an in_progress low task correctly outranked a medium todo.
  • scaffold Executed plan 03: added update_task/delete_task tools with id-resolution-via-list_tasks and a hard delete-confirm gate; mobile refreshes the list and shows Updated/Deleted confirmations; verified live — direct status flip, two-turn delete, ambiguous reference triggers a question.
  • edit Added long-press copy context menu to chat message bubbles: expo-clipboard + @expo/ui/swift-ui ContextMenu on iOS; Pressable + ToastAndroid fallback on Android.
  • scaffold Executed plan 06 (agent memory): MemoryFact entity + AgentMemoryService (in AiModule, not tasks), buildSystemPrompt(facts) folding durable facts into the prompt, remember/forget tools, and savedMemories/forgotMemories effects surfaced in the mobile chat; onboarding nudge when memory is empty.
  • fix forget tool let the model fabricate success: soft phrasing ("forget the Friday thing") picked a match phrase absent from the stored fact, matched nothing, yet the model still claimed removal — switched to word-overlap matching and made a no-match return the current facts so it can't lie.
  • review Verified memory end-to-end against the live Anthropic key: remember saves, dedupe skips a restatement, a fresh transcript recalls the fact, and forget removes the right one; documented memory's shape and limits (flat, no retrieval, single user) in the backend README.
  • scaffold Submission docs pass: agent read the brief PDF and wrote the root README — quick start (.env setup, seed, LAN-IP real-device testing), architecture + SQLite trade-offs, AI features mapped to the brief, David's decisions, descoped ideas, and the agentic-first development principles.
  • edit Rewrote backend/README (API + env table + storage choice + agent decomposition) and mobile/README (env-first get-started, structure/state overview); refreshed mobile/AGENTS.md with orientation rules.
  • fix Root CLAUDE.md "Current state" was stale (claimed AI not built) — updated to reflect the shipped AI layer and the remaining optional plans (04/05).
  • edit Added npm run dev aliases to backend and mobile package.json so the brief's npm install && npm run dev starts both apps as-is.

2026-06-17

  • scaffold TDD: extracted pure formatChatReply helper from chat screen inline logic; test written first (failed: module not found), then implementation, 4/4 passing.
  • scaffold TDD: made ChatAgentService stateful — persists user/assistant messages (with effects) via ConversationsService and AI-titles the first turn; spec written first (3 failing), then implemented, 4/4 passing.
  • edit Task 8: replaced chat-api with conversation-aware RTK Query endpoints (list/get/delete/sendChat + Conversation tags); retargeted formatChatReply to MessageEffects, ChatMessageBubble to backend Message type; deleted chat-slice.ts and removed it from store + persist whitelist.
  • edit Wired the stateful chat end-to-end: controller passes the whole DTO, mock agent returns AgentTurnOutput, AiModule imports ConversationsModule; full suite green (18/18), build clean, E2E smoke shows POST /ai/chat returning conversationId + AI title with persisted user/assistant messages.
  • plan Adapted the backend-conversations plan's mobile phase (Tasks 7–10) to reuse the now-merged decomposed chat units (ChatMessageBubble/ChatTypingBubble/ChatEmptyState/ChatInputBar/useKeyboardVisible) instead of re-inlining them — thread screen is pure orchestration; only ChatMessageBubble + formatChatReply change to consume backend types.
  • decision Dropped the plan's active-conversation Redux slice as unused (YAGNI): the open thread is the [id] route param and conversation data lives in the RTK Query cache, so nothing extra needs persisting.
  • review Reviewed Task 8 diff (chat-api overhaul, chat-slice deletion, formatChatReply retargeting); identified critical dead-import bug in chat/index.tsx still referencing deleted chat-slice and stale ChatMessage type.
  • scaffold Task 9: created chat/[id].tsx thread screen (send/delete/new-thread routing via router.replace) and updated _layout.tsx to register the [id] route; tsc clean on new files.
  • edit Mobile: chat tab is now a backend-backed conversation list + thread screens (RTK Query).

2026-06-23

  • scaffold Added ios-screenshot skill: capture.sh grabs a booted iOS Simulator screenshot (+ optional devyoga:// deep-link nav) so the agent can Read its own UI and close the implement→observe→fix loop; gitignored mobile/.screenshots, plus SKILL.md and a CLAUDE.md pointer.
  • fix Quick-reply chips now render on the new-chat screen: dropped the !isNew guard in chat/[id].tsx's isEmpty (it forced /chat/new to skip the empty state); now messages.length === 0 && !isConversationLoading && !isLoading. Verified via simulator screenshot.
  • scaffold maestro-e2e skill: SKILL.md documents the build→drive→observe→fix loop with run-flow.sh, curated flows in mobile/e2e/, and optional Maestro MCP wiring via .mcp.json.
  • scaffold mobile/e2e/: added Maestro flows (smoke, create-task, chat-send) and run-flow.sh runner under mobile/e2e/ (moved out of .claude/skills/); also added testIDs to task-form and chat-input-bar so flows can target controls by id.
  • fix Root-caused why Maestro assertVisible couldn't see task cards: expo-router's native Link.Preview/Link.Menu host strips its trigger subtree from the iOS accessibility tree (live hierarchy dump showed the FlatList present but zero card nodes — also invisible to VoiceOver). Confirmed testID/accessibilityLabel on the inner Pressable can't surface it.
  • fix Re-exposed task cards by wrapping the <Link> in an outer accessible View (accessibilityLabel={title}, testID="task-<id>"); keeps tap-nav + long-press preview/Delete menu and fixes VoiceOver. Verified end-to-end with a gitignored repro flow (mobile/.maestro/repro-link-a11y.yaml).
  • fix Replaced task-detail header "Delete" text button with a SwiftUI trash icon (@expo/ui/swift-ui Button, label "Delete task") to remove text ambiguity with the Alert's "Delete" button; updated create-task.yaml cleanup to tap "Delete task" then "Delete" — flow is now PASS and self-cleaning.
  • decision Removed legacy ios-screenshot skill — superseded by maestro-e2e (which covers static screenshots via takeScreenshot and interactive flows); kept both was confusing agents about which to reach for.