Skip to content

progress.md

Latest commit

 

History

History
152 lines (120 loc) · 12.3 KB

progress.md

File metadata and controls

152 lines (120 loc) · 12.3 KB
Raw

AgentForge — Progress Tracker

Last Updated: June 16, 2026 (UI Bug Fix Session — Save/Auth/Dark Mode)

Current Phase: All Features Implemented — UI Polish + Bug Fixes

Phase 0–3: All Done (see STRUCTURE.md for details)

Phase 4–8: All Done (see STRUCTURE.md for details)

Roast Review Fixes (from ROAST_REVIEW.md)

Critical (Kill List)

Issue Status Fix
eval() in RouterNodeExecutor Fixed Replaced with simpleeval library
HITL auto-approves everything Fixed Real Redis polling, timeout rejects
Git history reveals 2-evening build Deferred Requires interactive rebase

Major — All Fixed

  • OTel spans wired into executors + worker
  • Engine tests: 25+ tests (compiler, all executors, safeeval security)
  • SQLite → testcontainers-postgres with fallback
  • No live demo URL — deferred

Minor — All Fixed

  • WebSocket Redis leak → pub/sub relay
  • Docker Compose version removed, stop_signal added
  • .dockerignore added, README URL fixed
  • Evaluator cost tracking, 5 ADRs created

Runtime Fixes (Live Debug — June 15)

Infrastructure

Issue Root Cause Fix
Backend fails to start langfuse.callback removed in v4.x Use Langfuse client directly
pip install -e . fails setuptools multi-package detection [tool.setuptools.packages.find] include = ["app*"]
Register returns 500 passlib incompatible with bcrypt >= 4.1 Direct bcrypt.hashpw()/checkpw()
GET /workflows returns 422 Header(...) required → 422 Header(None) + explicit 401

E2E Workflow Bugs (found during live testing)

Issue Root Cause Fix Commit
Nodes don't appear on canvas React Flow useNodesState creates local state that doesn't sync with Zustand store Removed useNodesState/useEdgesState, use Zustand store directly with applyNodeChanges/applyEdgeChanges dag_canvas_fix
Save returns "Network Error" workspace_id: "default" not a valid UUID → FK violation → 500 → axios shows as network error Auto-create workspace via API if none exists; extract error.response.data.detail for meaningful messages e856afe
Save creates new workflow instead of updating After first save, URL stays at /editor (no ?id=xxx) → store workflowId lost on navigation useEffect watches storeWorkflowId → router.replace(/editor?id=xxx) a962df1
Saved workflow nodes disappear on reload WorkflowResponse model didn't include dag_json field → loadWorkflow() gets undefined Added `dag_json: dict NonetoWorkflowResponse+_to_response()`

UI Bug Fix Session (June 16)

Issue Root Cause Fix
Save button stays active after save (v1→v2 without changes) React Flow fires onNodesChange for select and dimensions events (not just structural edits), which all triggered setNodes → isDirty: true Filter non-structural changes in DAGCanvas.tsx; use useWorkflowStore.setState() directly for select/dimensions without setting isDirty
No auth redirect on first load (goes to dashboard instead of login) Dashboard and editor pages had no auth guard — loaded immediately without checking access_token Added useEffect auth guards with router.replace("/login"), loading screen prevents flash, also wrapped useSearchParams pages in Suspense for Next.js build
Dashboard flashes briefly before login redirect useEffect runs after first render — server pre-renders full dashboard HTML before JS can redirect Next.js middleware (src/middleware.ts) checks access_token cookie server-side before any page renders; login sets cookie alongside localStorage; simplified client-side guards
Dark mode not implemented All 8 pages + 5 DAG components used hardcoded light colors (bg-white, text-gray-900) with no dark: Tailwind variants Added dark: classes to all pages/components; added @custom-variant dark to globals.css for Tailwind v4 class-based dark mode; DarkModeToggle added to dashboard + editor headers
Dark mode toggle stuck on dark (can't switch back to light) (1) Tailwind v4 defaults to prefers-color-scheme media query, ignoring .dark class; (2) Next.js React reconciliation resets <html> className on re-renders, wiping the .dark class (1) Added @custom-variant dark (&:where(.dark, .dark *)) to globals.css; (2) Toggle targets <body> instead of <html> since body isn't reconciled by React

UX Polish + Execution Fix (June 18)

Issue Root Cause Fix
Node numbering is global (Agent 1, Tool 2, Router 3) Single nodeCounter incremented for all node types Per-type counting: filter existing nodes by type, increment per-type (Agent 1, Tool 1, Agent 2)
Edge routing looks curved/pulled upward Handles at Top/Bottom positions with bezier edges Moved handles to Left/Right + smoothstep edge type for clean horizontal flow
No warning when clicking Back with unsaved changes "Back" was a plain <a> link with no dirty check Converted to <button> with window.confirm() when isDirty; added beforeunload listener for tab close
Workflows stay in "queued" status forever Execution worker was a separate process that was never started Embedded worker as asyncio task in FastAPI lifespan with lazy import; worker starts automatically with backend
LLM calls fail with "OPENAI_API_KEY not configured" No LLM provider configured in .env Added OPENAI_BASE_URL config for OpenAI-compatible APIs; integrated Ollama Cloud (https://ollama.com/v1)

Platform Completeness Sprint (June 19)

Issue Root Cause Fix
Templates page is empty No seed data in agent_templates table Created seed_templates.py with 8 starter templates (Summarizer, Code Reviewer, Data Extractor, Sentiment Analyzer, Q&A Agent, Content Writer, Translation Agent, Research Assistant); seeded on startup via lifespan
MCP Servers page is empty No servers registered in mcp_servers table Created seed_mcp_servers.py with 3 built-in servers (filesystem, web-search, postgres); seeded on startup
Cost dashboard shows $0.00 Ollama Cloud models had $0 pricing Updated MODEL_PRICING with realistic Ollama Cloud rates ($0.50–$4.50/1M tokens); costs now flow through execution → CostRecord → dashboard
Dashboard workflow cards lack execution data List endpoint only returns workflow metadata Added /workflows/dashboard/summary endpoint with execution count, last run status, total cost per workflow; enhanced frontend cards with stats display
Dashboard stats cards show static data No aggregation queries Computed totalExecutions and totalCost from summary endpoint; wired to stats cards
Template "Install" button does nothing handleInstall was a fake alert() with no API call Added POST /templates/{id}/install endpoint that creates a workflow with Input → Agent (template config) → Output DAG; frontend calls API and redirects to editor
Template workflow persists after "Leave" Install creates workflow in DB; clicking "Leave" only navigates away Added from_template URL flag; editor marks as dirty on load; handleBack deletes workflow if user leaves without saving
Langfuse not running Container not started, no DB, no API keys Created langfuse DB, started container, configured LANGFUSE_PUBLIC_KEY/SECRET_KEY in .env; added setup guide to STRUCTURE.md

Changelog

Date Session Change
June 11 2026 Phase 0–3 Full build: monorepo, backend, frontend, engine, tests
June 11 2026 Session 2 Frontend polish: save wiring, login, workspaces, node config, execution history
June 11 2026 Session 3 LLM client, MCP client, checkpointer, rate limiting, API keys, webhooks, OAuth, WebSocket, templates, MCP/template/cost UI, E2E tests
June 11 2026 Session 4 Budget enforcement, OTel tracing, Langfuse integration, HITL WebSocket, error boundaries, versioning UI, dark mode, agent memory, security hardening
June 13 2026 Roast Fixes Kill eval(), real HITL, OTel spans wired, engine tests, testcontainers, WebSocket leak, Docker fix, ADRs
June 15 2026 Live Debug langfuse v4, setuptools, bcrypt, auth 401, Docker + migrations + full stack verified
June 15 2026 E2E Bug Fixes 4 critical E2E bugs fixed: React Flow state sync, auto-create workspace, URL sync after save, dag_json in API response
June 16 2026 UI Bug Fixes Save isDirty fix, auth redirect guard, full dark mode across all pages/components, dark mode toggle fix
June 18 2026 UX Polish + Execution Per-type node numbering, left/right handles, unsaved warning, execution worker embedded in lifespan, Ollama Cloud LLM integration
June 19 2026 Platform Completeness 8 seed templates, 3 seed MCP servers, cost tracking, dashboard summary, template install flow, Langfuse observability
June 11 2026 Roast Review V2 Fixes 11 production-ready improvements from comprehensive technical review

Roast Review V2 Fixes (June 2026)

Critical Issues Fixed

Issue Root Cause Fix Files Changed
Compiler router always takes first branch _make_router_fn evaluates conditions locally, doesn't read executor output Read selected_route from node_results[source_id]["output"] instead of dummy evaluation backend/app/engine/compiler.py
Redis connection churn in worker _emit_ws_event creates new Redis client per WebSocket event Added shared singleton Redis client with _get_shared_redis() backend/app/workers/execution_worker.py
LLM client fragile to transient errors No retry logic — single 429 kills entire workflow Added tenacity retry (3 attempts, exponential backoff 2s→4s→8s) backend/app/engine/llm_client.py, pyproject.toml
Context window overflow in prompts Dumping entire upstream state into LLM prompts Added 4000-char truncation with _truncate_context() helper backend/app/engine/executors.py

Major Issues Fixed

Issue Impact Fix Files Changed
E2E tests not in CI pipeline Playwright tests existed but never ran automatically Added complete E2E job to .github/workflows/ci.yml .github/workflows/ci.yml
PostgreSQL testcontainer leak Container never stopped after tests, orphaned processes Added session-scoped cleanup fixture in conftest.py backend/tests/conftest.py
Model pricing hardcoded in Python dict Would go stale quickly, hard to update Moved to external YAML config file loaded dynamically backend/app/engine/model_pricing.yaml, llm_client.py
OAuth callback URL hardcoded to localhost Would break on any non-local deployment Made configurable via settings.oauth_redirect_base_url backend/app/api/v1/oauth.py, config.py
No WebSocket disconnection notification Auto-reconnect exists but no user-facing feedback Created WebSocketStatus component + ExecutionMonitor with reconnect tracking frontend/src/components/execution/WebSocketStatus.tsx, ExecutionMonitor.tsx, useExecutionWebSocket.ts

Deferred (Would Require Migration)

Issue Decision
Agent memory model unused (dead code) Left as-is — removing requires database migration

Files Created

  • backend/app/engine/model_pricing.yaml — External model pricing configuration
  • frontend/src/components/execution/WebSocketStatus.tsx — WebSocket connection status indicator with error boundary
  • frontend/src/components/execution/ExecutionMonitor.tsx — Real-time execution monitoring panel

Files Modified

  • backend/app/engine/compiler.py — Fixed router to use executor output
  • backend/app/workers/execution_worker.py — Shared Redis client for WebSocket events
  • backend/app/engine/llm_client.py — Tenacity retry logic, YAML config loading
  • backend/app/engine/executors.py — Context truncation helper
  • backend/app/api/v1/oauth.py — Configurable OAuth redirect URLs
  • backend/app/core/config.py — Added oauth_redirect_base_url field
  • backend/tests/conftest.py — PostgreSQL container cleanup
  • .github/workflows/ci.yml — E2E test job
  • backend/pyproject.toml — Added tenacity dependency
  • frontend/src/lib/useExecutionWebSocket.ts — Added reconnectAttempts tracking