System Changelog

Checkout success stops being a confetti screen and starts being a real provisioning flow — payment confirmed → sandbox API key issued via /v1/keys/issue → key + endpoint + curl displayed → "Run in browser" button that fires a live call against the endpoint so you see the typed response right there. Use the agents as MCP servers in Claude Desktop / Cursor via a single-file Node stub at /mcp/vulcan-agent.mjs — drop the path into your MCP config and a new tool appears.

• New DeliveryFlow component on checkout success. Three live steps: Payment confirmed → Provisioning sandbox key (auto-calls /v1/keys/issue, surfaces the bundle in localStorage) → Verify on this device.
• "Run in browser" button on the verify step fires a live POST against the agent endpoint with your fresh Bearer token. Whether 200 success or a typed refusal lands, you see the parsed JSON envelope inline.
• Bundle card shows the API key, the endpoint, the daily quota, and a curl snippet pre-wired with the real key. Copy buttons on each row.
• Slug resolution: success screen now looks up the canonical slug from the agent name via the registry, so the issued key matches the correct backend endpoint (capital-intelligence-agent → capex-agent, etc.).
• Razorpay-mode badge: shows "Razorpay test mode" honestly while we run on the test key; flips to "Live Razorpay" when the live key is set in env.
• New /mcp/vulcan-agent.mjs — a single Node 18+ file with zero dependencies that wraps any Vulcan agent endpoint as an MCP server over stdio. Implements initialize / tools/list / tools/call. Takes --slug, --endpoint, --tool plus VULCAN_API_KEY from env.
• Setup page MCP tab updated: shows real Claude Desktop / Cursor config (not the imaginary @vulcan/mcp-client npm package). Adds a 4-step install block with a direct download link for vulcan-agent.mjs.
• Apollo learned MCP. Ask "how do I use Vulcan agents as MCPs" and Apollo returns the 3-step recipe with a working config block. The default suggestion chips now surface this question on first open.

Every agent now ships with a dedicated setup page at /workflows/<slug>/setup — endpoint, sandbox API key, copy-pasteable cURL / Python / JavaScript / MCP-config snippets, sample payload, and the typed response shape. A "Setup guide" CTA is visible right on the agent detail page above the integration-method picker. Checkout polished: Razorpay branding fixed (Vulcan, not GetBot), accent colour aligned to the marketplace, animated spinner during processing, success state hands you straight to the setup guide instead of just the dashboard.

• New /workflows/[id]/setup route with five numbered sections — Endpoint, API key, First call (4 tabs), Sample payload, Response shape — plus a footer CTA stack (Try in playground · Open workbench · Purchase).
• New src/lib/agents/setup-snippets.ts — per-agent payload + response shape + 4-language snippet builder. Covers all 17 agents including AnswerBot, Customer Support, and the 5 new Intelligence agents.
• "Issue sandbox key" button on the setup page calls the existing /v1/keys/issue endpoint, caches the bundle in localStorage, and rewrites every snippet on the page with the real key.
• Visible "Setup guide for <agent>" CTA card on the detail page (above the integration-method picker) with endpoint + snippet preview hint.
• Checkout: Razorpay name corrected to Vulcan, accent colour to #00C878, image swapped to favicon, description simplified.
• Checkout success: now reads "<Agent> is provisioned and ready. Your sandbox API key is waiting on the setup guide." with two CTAs (Open setup guide / Go to dashboard).
• Checkout pay button: animated spinner during processing, plan/cadence visible in the price ("Pay ₹499 / monthly"), prominent "Secured by Razorpay · 256-bit SSL · PCI-DSS compliant" trust footer.
• When Razorpay SDK fails to load (demo / offline), the simulated flow still flips the purchased flag so the dashboard + setup-page reflect the upgrade.

Major polish + capability pass driven by user feedback. Every demo tagline + capability chip rewritten in plain business English — no more "MVP — lint + security scan, v3-25 brings PR-aware review". Investment Intelligence gets a live market snapshot endpoint with per-ticker price + day change %. Contract Intelligence ships playbook-aware redline suggestions + a customer/vendor protections breakdown. AI Engineering surfaces function-count, average cyclomatic complexity, hotspots, and test detection. Customer Support, Communication, Product Adoption, and Executive Operations gain a file-upload affordance for grounding context. Apollo updated.

• Repositioning pass: every demo tagline rewritten in plain English ("A staff engineer reviewing every line. Catches eval(), bare excepts, console.log noise…") and capability chips made outcome-driven ("Live market snapshot", "0-100 thesis fit", "Buy / Consider / Monitor / Avoid"). No more internal phase markers (v3-22 / v3-25) shown to end users.
• Investment Intelligence — new POST /v1/investment-agent/market-snapshot endpoint. Returns deterministic seeded per-ticker price + day_change_pct + asset_class, with a timestamp and source label. Sandbox feed for the demo; production swap to a live data provider is a single function replacement.
• Contract Intelligence — playbook-aware suggested_redline per high/medium-risk clause. New library covers termination, liability, indemnity, auto_renewal, payment, governing_law across vendor / buyer / standard playbooks. ProtectionBreakdown surfaces customer_friendly + vendor_friendly + neutral clause counts. top_redlines lifts the 3 highest-impact rewrites to the top of the result.
• AI Engineering — module metrics. functions_count, average_cyclomatic across detected functions, up to 5 FunctionHotspot entries (name + start_line + length_lines + cyclomatic_estimate + note when >= CC 15 or >= 80 lines), tests_detected (pytest / jest / vitest / describe / it / def test_*). metrics_note summarises the worst hotspot in human English.
• New shared AgentFileDrop component — drag-and-drop or click upload, indexed locally for the demo session with a "X of Y indexed" footnote. Wired into CustomerSupportDemo (FAQ / policy files), EmailDemo (signature / tone guide), OnboardingDemo (product guide / activation playbook). ManagerDemo intentionally left for the Antigravity UI pass.
• Apollo Assistant — investment-agent purpose line updated to lead with "AI portfolio analyst + due-diligence reviewer" rather than "rule-based summary".
• agents-backend now serves 27 routes (+1 for market snapshot).

Closes out the v3 deepening sprint with four agents and the full refusal-coverage audit. Communication Intelligence (email) now triages inboxes with per-email intent + sentiment + priority + SLA + escalation. Product Adoption (onboarding) detects friction, scores churn risk per step, and prescribes a concrete next product action. A new companion endpoint at /v1/customer-support/score takes conversation threads and returns churn band + escalation + next action. Executive Operations gained an execute=true mode that simulates running the plan against sandbox-agent stubs and returns a consolidated output. Z14 audit verifies every agent returns the typed envelope on every refusal path.

• Communication Intelligence (email): per-email EmailAnalysis block — intent (12 labels), sentiment (positive/neutral/frustrated/angry) with -1..+1 score, priority (low/normal/high/urgent), urgency_minutes SLA, needs_escalation + escalation_reason, follow_up_after_days. Plus an inbox triage_summary with total / urgent / high / needs_escalation / average_sentiment counts.
• Product Adoption (onboarding): activation_score 0–100, friction_detected (confusion, error, blocked, slow, missing_info, abandonment, negative), churn_risk (low/med/high), suggested_next_action — concrete product-team moves like "trigger save-our-relationship: offer a 1:1 onboarding call within 24h".
• Customer Success Intelligence scoring (new endpoint /v1/customer-support/score): per-message sentiment with score + excerpt, overall_sentiment + score, churn_risk band (low/med/high/critical), needs_escalation + escalation_reason (legal/press/lawsuit auto-trigger), next_action recommendation (auto_resolve / agent_followup / csm_call / executive_escalation / drop_in_nurture / send_csat). Plan-tier aware — enterprise customers escalate to executive level.
• Executive Operations (manager): optional execute=true flag. Simulates running each plan step against deterministic stubs for the sandbox agents (Scraper / Summariser / Notifier); returns execution_trace[] with per-step status + output + duration_ms plus a consolidated_output string. Cross-agent live HTTP fan-out across the v3 catalogue is scoped for v3-Phase-K.
• Z14 — refusal-coverage audit: ran a curl matrix across all 22 agent endpoints. Every refusal path returns the typed AgentResponse envelope with ok:false + a reason_code. Verified empty / oversized / invalid / unreachable / blocked / out-of-range / domain-specific edge cases (e.g. logger no_matches, investment unknown_ticker, onboarding step_out_of_range). No 500s.
• Apollo Assistant knowledge base updated for all four deepened agents — "how do I use" + "what does X do" answers describe the new capabilities with refreshed sample inputs.
• agents-backend now serves 26 routes (was 25 yesterday); pm2 process restarted; sync rsync'd cleanly.

Resume Processor stops being a parser and starts being a hiring decision system. Every result now ships estimated years of experience, seniority tier, career trajectory, culture signals, and 5 generated interview questions. Pass an optional JD as a form field and the agent scores skill alignment (matched / missing / bonus), produces a 0–100 jd_match_score weighted 60/25/15 across skills/seniority/years, and recommends strong_hire / hire / interview / pass with a rationale. Apollo updated.

• estimated_years_experience — explicit "X years" first, then merged year-range spans across roles.
• seniority_tier across 7 levels (intern → executive) inferred from titles + fallback to years bands.
• trajectory — rising / flat / lateral / declining — derived from rank progression of role titles in chronological order.
• culture_signals — leadership, mentoring, open_source, startup, remote, research, publication, speaking, ownership, collaboration. Each with a 240-char evidence snippet.
• New optional `jd` multipart form field on POST /v1/resume-processor/run.
• skill_alignment block: matched + missing + bonus, extracted from a 60-token JD skill vocab (langs, frameworks, data, cloud, ML).
• jd_match_score 0–100 weighted 60% skill coverage + 25% seniority alignment + 15% years alignment. Tier mapping: ≥85 strong_hire, ≥70 hire, ≥50 interview, <50 pass.
• interview_questions — up to 5, drawn from categories: technical (top skill), system_design (most recent role), behavioural (standard), skill_gap (when JD has missing skills), trajectory-aware.
• rationale string explains the composite score breakdown for hiring managers.
• Pure rule-based — no LLM. Every claim traces back to a literal line in the parsed resume.
• Verified live: synthetic 8-year senior backend engineer → 94/100 strong_hire against a Python/Go/Kafka JD, 5 culture signals detected, 4 interview questions including trajectory probe.
• Apollo Assistant knowledge base updated — "how do I use Talent Intelligence Agent" / "what does Resume Processor do" both describe the new flow with multipart sample.

Lead Gen stops being a contact lookup and starts being a sales decision system. Every result now ships detected buying signals (funding, hiring, leadership change, product launch, tech adoption, expansion, press, IPO path) with intensity + evidence + score contribution. The new intent_score (0–100) maps to a cold / warm / hot / very_hot tier. Pass an optional ICP descriptor and the agent scores criterion-by-criterion. Every result ends with a concrete next-step. Apollo Assistant updated to teach the new flow.

• 8 signal kinds: funding, hiring, leadership_change, product_launch, tech_adoption, expansion, press, ipo_path. Each detected by a small + transparent regex vocabulary — every match returns an evidence snippet, no fabricated signals.
• Per-signal score contribution capped at 30. Intent score is the sum, capped at 100. Tiers: very_hot (≥70), hot (≥45), warm (≥20), cold (<20).
• New optional `icp` field on the request. Parser supports SaaS / Fintech / AI sectors, US / India / Europe regions, Seed / Series B+ / Public stages, hiring activity. Each criterion in the supplied ICP becomes an IcpCheck with evidence.
• Next-step recommendation generated from signals + tier + ICP fit. Examples: funding + very_hot → personalised note within 48h targeting the role; new leadership → congratulations note; weak ICP (<40) → drip cadence.
• Mock corpus enriched — Satya Nadella, Sundar Pichai, Jensen Huang, Lisa Su, and a Series-B DemoCo profile each carry a synthetic company_signals_text + company_meta block.
• Scrape-path lead enrichment also scans the fetched HTML for signals (best-effort, confidence stays low to be honest).
• Apollo Assistant knowledge base updated — "how do I use Revenue Intelligence Agent" / "what does Lead Gen do" both describe signals, intent, ICP, and next-step with new sample input.
• Verified live: DemoCo (Series B) returns intent 83/100 very_hot with funding + leadership + hiring signals, ICP 75/100 (3 of 4 criteria matched), next-step references the role.

Crawler Agent stops blindly BFS-ing and starts asking the site for its own map. Auto-discovers sitemap.xml + RSS / Atom feeds, falls back to BFS only when neither exists. Every result now includes a discovery field (source URL + method + counts), per-page keywords pulled from title + headings, and topic clusters that partition the crawled pages by shared keyword overlap. Apollo updated to teach the new flow.

• Sitemap discovery — probes /sitemap.xml, /sitemap_index.xml, /sitemap-index.xml. Parses sitemap-index recursively (one level). Same-domain filtered. Replaces BFS when found.
• RSS / Atom feed discovery — probes /feed, /feed.xml, /rss, /rss.xml, /atom.xml. Article links seeded into the crawl queue.
• New DiscoveryInfo on every result: method (sitemap | rss | bfs | mixed), the source URL used, sitemap_urls_found + rss_items_found counts. Honest about how the page list was assembled.
• Per-page keywords field — top tokens from title + headings, stopword-filtered, deduped, capped at 12 per page.
• Topic clustering — greedy partition by ≥2 shared keywords across crawled pages. Up to 5 clusters returned with label (top shared tokens) + member URLs + keywords. No LLM in the path.
• Verified live on vercel.com (sitemap → 20 URLs, 2 clusters), overreacted.io (RSS → 20 articles, "explorer / overreacted / introducing" cluster), example.com (BFS fallback).
• Apollo Assistant knowledge base updated — "how do I use Knowledge Intelligence Agent" and "what does Crawler do" both describe sitemap / RSS / clustering with new sample input.

Web Scraper deepened from a single-pass HTML parser into the full 5-layer extraction stack we promised. Layer 3 statically discovers XHR / REST / GraphQL endpoints from inline scripts + preload tags. Layer 4 parses __NEXT_DATA__ and inline JSON islands into structural summaries. Layer 5 lets you supply user-defined schema_fields and walks every JSON source for matches — sensible defaults baked in for price / location / name / rating / bedrooms. Apollo Assistant updated to teach the new layers.

• Layer 3 — network analysis: pattern-matches fetch() / axios / XHR endpoint URLs from inline scripts, detects GraphQL endpoints, scans <link rel="preload" as="fetch"> tags. Up to 25 hints returned with source attribution (preload | script | graphql | next_data).
• Layer 4 — backend source discovery: parses Next.js __NEXT_DATA__ into a structural summary (buildId, page, query, pageProps keys). Walks the dehydrated state DFS for apiUrl / apiBaseUrl signals. Pulls up to 12 inline application/json blobs and summarises their shape.
• Layer 5 — schema-driven extraction: pass schema_fields: [{name, hints[]}]. Agent walks JSON-LD + __NEXT_DATA__ + inline JSON DFS and returns the first match per field under schema_results. Defaults for price / amount / cost, location / city / address, image, rating, description, bedrooms, currency.
• Open Graph aggregation — clean og:* dict via the new "open_graph" field.
• New extractable fields exposed in the request type: next_data, inline_json, api_hints, open_graph.
• Apollo Assistant updated — "how do I use Data Acquisition" + "what does Web Scraper do" now describe all 5 layers with sample input including schema_fields.
• Hard limits preserved: 10 s fetch timeout, 2 MB body cap, content-type guard, typed RefusalResponse on every edge case.

Three agents leveled up from MVP to v3-real. Capital Intelligence now ships full DCF projection tables + ±20% best/base/worst scenario stress test with a board-ready recommendation. Engineering Intelligence (formerly Logger) added a real RCA correlation engine — detects cascades, surfaces causal chains, names the upstream root cause. Investment Intelligence gained a pitch-deck intake mode that extracts financials with regex, scores against a user-supplied investment thesis, and outputs a defensible Buy / Consider / Monitor / Avoid recommendation. Apollo Assistant's knowledge base updated to teach every new capability.

• Capital Intelligence: DCF projection table (per-year after-tax cash flow, discount factor, present value).
• Capital Intelligence: 3-scenario stress test — best (+20% cash flow), base (0%), worst (-20%) — each scenario re-runs NPV / IRR / payback with a verdict (accept / marginal / reject).
• Capital Intelligence: headline recommendation — proceed / monitor / reject — derived deterministically from base + worst case verdicts. Frontend renders a status pill + verdict colour badge per scenario row.
• Engineering Intelligence: RCA correlation engine. Mode 1 detects multi-service same-minute cascades; Mode 2 detects single-service sustained bursts (5 events in 5 min) and elevated-volume incidents (≥10 events). No-LLM path — every chain claim grounds in a literal log entry.
• Engineering Intelligence: plain-English causal chain with topology-aware arrows ("auth → api: Auth failures cascading into API rejections"). Rule-based RCA summary names the likely upstream root cause + suggested action.
• Investment Intelligence: new POST /v1/investment-agent/analyze-deck endpoint. Regex extraction surfaces company name, sector, region, revenue, ARR, growth %, burn, runway, valuation, raise amount.
• Investment Intelligence: optional thesis scoring — parses criteria like "AI cos, ARR > 2M, growth > 30%, US market" into yes/no checks, scores 0–100, recommends strong_buy / consider / monitor / avoid with rationale.
• Investment Intelligence: risk detection — short runway, high burn multiple (>2× ARR), legal/regulatory signals, valuation > 50× ARR.
• InvestmentDemo: new "Deck analysis" mode tab — paste deck text + thesis, see extracted financials KV grid + thesis check list + recommendation badge.
• Apollo Assistant knowledge base updated for all three deepened agents — "how do I use" + "what does it do" answers now describe the new capabilities including DCF, scenarios, RCA correlation, causal chain, deck intake, thesis scoring.

Apollo went from a stub chatbot to a deterministic agent advisor that knows every Vulcan agent — features, how-to-use, pricing, sample inputs, refusal codes. Auto-nudge prompts idle users every 30 seconds with "Need help? Ask Apollo". The long-standing invisible-input-text bug is fixed.

• New src/lib/apollo/knowledge.ts — structured help base covering all 17 agents. Per-agent: purpose, capabilities, how-to-use steps, sample input JSON, refusal codes, pricing.
• Deterministic intent classifier (no LLM) handles: list agents, list by layer, how-to-use, features, pricing, refusals, sample-input, about Vulcan, greetings.
• Suggestion chips render under every reply so users can drill in without typing.
• Auto-nudge: first pulse 30 s after page load, then every 60 s while idle, exactly as requested. Stops on first interaction.
• Invisible-input-text bug fixed — forced color + -webkit-text-fill-color + caret-color on the input plus same treatment on user message bubbles, so text shows correctly regardless of theme cascade.
• Inline rendering of **bold**, `code`, and ```pre``` blocks in Apollo responses, with HTML-injection-safe whitelisting.
• Apollo answers under 300 ms — pure lookup, zero hallucination, consistent with every other Vulcan agent's guardrail policy.
• AnswerBotStudio rebuilt as a 3-panel workbench: left rail "Chats" (sessions persisted to localStorage, new + delete), center "Data Ingestion" (drag-drop zone) plus "Knowledge Library" (file list with delete), right "Chat" (conversation with prompt chips + cited sources). Replaces the cluttered 2-panel layout flagged earlier.
• AnswerBot input bug fixed too — same forced colour treatment so typed messages remain visible.

The buy-to-use loop closes. Purchased agents now issue HMAC-signed API keys on demand from the dashboard. Bearer-authenticated requests bypass the free-tier rate limit and grant a 1,000-call daily quota. The Delivery panel surfaces a cURL / JavaScript / MCP-config snippet per agent so a customer can integrate from any stack within a minute of purchase.

• New backend endpoints: POST /v1/keys/issue and POST /v1/keys/validate on agents.aidesktop.in.
• HMAC-SHA256 signed tokens bind {slug, account_id, issued_at, nonce} — keys can't be forged, replayed across agents, or surface beyond the signing secret.
• BearerAuthMiddleware on the agents-backend attaches AuthState to every request — valid keys mark requests as paid and bypass the slowapi free-tier limit.
• Per-agent daily quota for purchased customers (default 1,000 calls/day, in-memory MVP — durable Redis counter ships in Phase G+).
• Frontend: src/lib/agents/credentials.ts manages the credential bundle lifecycle in localStorage (cache, rotate, clear).
• runAgent() auto-attaches Authorization: Bearer <token> when a bundle is cached — purchased demos hit the paid quota path with zero code changes.
• Dashboard: new DeliveryPanel component rendered below the demo for purchased agents. Shows endpoint, masked API key with copy-to-clipboard, three integration tabs (cURL / JavaScript / MCP config), and a rotate-key button.
• Per-browser account identifier auto-created (acct_<random>) so keys are device-scoped; real auth handshake lands in Phase G+ with the signed-JWT exchange against the marketplace.
• All 15 demo titles aligned with v3 canonical names (Capital, Engineering, Investment, Data Acquisition, Revenue, Product Adoption, Talent Intelligence) — earlier rename missed these AgentDemoShell title props.
• Honest disclosure: this MVP is a buy-to-use credential bundle. Real OAuth, signed-JWT exchange, rotating signing keys, and webhook delivery channels remain scoped for v3 Phase G+.

Marketplace rebranded from "Get A Bot" to Vulcan. Five new Intelligence Agents shipped from Coming Soon → live MVP backends: Research, Contract, Analytics, AI Engineering, Compliance. Each runs as a typed deterministic endpoint with a real demo (rule-based / lookup MVP; full enterprise capabilities ship per docs/v3-backend-phases.md). Global emoji ban enforced — every agent icon now renders as an SVG via a central icon system, no emoji left on any display surface.

• Marketplace renamed to Vulcan across page title, OpenGraph + Twitter metadata, Footer, Logo alt, and structured-data org info.
• Research Intelligence Agent live (MVP) — query routes against a fixed topic catalogue (AI market, SaaS metrics, India hospitality), returns cited findings + confidence + suggested follow-ups.
• Contract Intelligence Agent live (MVP) — regex + keyword clause classifier with high/medium/low risk flagging across 9 clause families + playbook awareness.
• Analytics Intelligence Agent live (MVP) — NL → SQL template generator over 3 demo schemas (sales / ops / marketing) with interpretation + sample rows.
• AI Engineering Agent live (MVP) — heuristic code review for Python + TypeScript with severity-graded findings (correctness / security / style / complexity) and an overall grade.
• Compliance Intelligence Agent live (MVP) — framework + question → mapped controls across SOC2, GDPR, ISO27001, HIPAA, PCI-DSS with evidence guidance + next steps.
• Every demo wired in both /workflows/<slug> and /dashboard?agent=<slug>. The five Coming-Soon panels are replaced with live demos using the same AgentDemoShell + typed envelope as the v2 agents.
• SVG-only icon system: new src/lib/agents/icons.tsx maps every AgentSlug to a lucide-react SVG component (DollarSign, Workflow, Globe, Code2, Shield, …). Every emoji removed from registry, MOCK_AGENTS, dashboard rail, and detail-page hero.
• AnswerBot confirmed in the Data Layer (was missing from the original v3 category map). Icon: BookOpen.
• Marketplace agent count badge in the catalog stats reads 17 (was 12 in v2.8).
• Honest MVP disclosure: every new agent's tagline + demo header surfaces "MVP — rule-based / lookup table" plus the v3-phase ID where the real engine ships.

Repositioned the entire marketplace from tool-named agents (Web Scraper, Logger, Email Automation) to outcome-driven AI Operators (Data Acquisition Intelligence, Engineering Intelligence, Communication Intelligence). Every agent now owns a business function instead of an implementation. Catalog restructured into five layers: Intelligence, Data, Communication, Experience, Orchestration. Seeded 5 new agents (Research, Contract, Analytics, AI Engineering, Compliance) with full positioning + Coming Soon UX.

• All 12 existing agents renamed to outcome-driven Intelligence Agent positioning — Capex Agent → Capital Intelligence Agent, Logger Agent → Engineering Intelligence Agent, Web Scraper → Data Acquisition Intelligence Agent, Manager Agent → Executive Operations Agent, and 8 more.
• Long-form positioning copy rewritten across every detail page — reads like a service pitch from a domain expert, not a tool spec sheet.
• Catalog organised into 5 marketplace layers: Intelligence (10 agents), Data (3), Communication (2), Experience (1), Orchestration (1).
• New "Layer" filter group added as the primary catalog axis. Status filter gained a "Coming Soon" option.
• 5 new agents seeded with full positioning + Coming Soon UX: Research Intelligence, Contract Intelligence, Analytics Intelligence, AI Engineering, Compliance Intelligence.
• Coming Soon agents render a typed waitlist panel on detail pages with a disabled Buy button; backend phases scoped in /docs/v3-backend-phases.md.
• Single source of truth registry (lib/agents/registry.ts) now drives slug → name → tagline → layer → status across catalog, detail pages, and dashboard — eliminates the previous naming drift.
• All 10 demo component titles updated to match the new outcome names — no inconsistency between detail page hero and embedded demo header.
• Backend slugs unchanged in v3 — existing /v1/<slug>/run endpoints keep working. URL contract preserved during the rebrand.
• 26-phase v3 backend buildout scoped in docs/v3-backend-phases.md — every Intelligence Agent has a concrete capability roadmap with effort estimates.

Shipped agents.aidesktop.in — a typed, rate-limited mono-service backing real working demos for all 10 previously-unbuilt agents. Rewired /dashboard from a 4-mock-bot stub into a 12-agent workbench. Hardened every demo with a typed RefusalResponse envelope so edge cases surface as structured UI instead of 500s.

• agents.aidesktop.in mono-backend live: one FastAPI service hosting Capex, Logger, Web Scraper, Resume Processor, Investment, Lead Gen, Email Automation, Onboarding, Crawler, and Manager agents.
• Every agent returns a typed AgentResponse[T] envelope — success data or RefusalResponse with a machine-readable reason_code. No 500s on expected edge cases.
• Deterministic agents (Capex, Logger, Investment, Onboarding, Crawler) bypass the LLM entirely — pure formulas + structured logic, zero hallucination surface.
• LLM-backed agents (Email, Manager) validate every output against Pydantic schemas before returning; mismatches auto-refuse as low_confidence instead of guessing.
• Per-IP rate limiting (slowapi): 4 free runs per agent per day, resets at UTC midnight. Purchased agents bypass the limit.
• /dashboard rewritten as "Agent Workbench" — left rail lists all 12 agents from a canonical registry; selecting one mounts its live demo in the main pane.
• Removed the broken MCP / SDK / External APIs / Webhook configuration panel — those integrations were stubs and never connected. Replaced with an honest Delivery card (endpoint URL, runs-left, "API key delivery in v2.7" disclaimer).
• Customer Support live chat extracted into a standalone component so the same RAG chat surfaces in both /dashboard and /workflows/customer-support.
• Manager Agent frontend demo: task input + sandbox-agent toggles + plan visualisation with per-step expected outputs.
• Runs counter now syncs across the dashboard header and every demo shell via a custom RUNS_CHANGED_EVENT (same-tab writes don't fire native storage events).
• VPS hardened with a third pm2 process (agents-backend on :8002), Let's Encrypt cert on agents.aidesktop.in, nginx reverse proxy.

Launched AnswerBot as a dedicated 12th agent with a multi-panel demo studio (file upload, OCR, audio transcription, document library). Rebuilt the integration UX with real 3-state status, fixed theme inconsistencies, added a notification-driven coupon system, and shipped curated 6-hour hero word rotation.

• AnswerBot listed as 12th agent — universal doc Q&A with OCR + audio transcription.
• Multi-panel AnswerBot Studio: document library, live chat with citations, capability chips.
• Cloud migration: OCR via GPT-4o Vision, audio via Whisper API (no GPU needed on VPS).
• Real integration UX: paste API key / MCP config / SDK snippet, status verified before showing "Connected".
• Coupon system: VULCAN40 for 40% off, applied at checkout with server-side validation.
• Notifications: pinned offer at bottom, top 3 recent releases auto-fed from the changelog.
• Theme polish: killed rogue blue across filter chips, dashboard toggles, command palette.
• Hero word now rotates every 6 hours from a 60+ word curated pool (EXQUISITE, ATELIER, MOMENTUM…).
• Footer: GitHub + LinkedIn + portfolio icon links next to Devaansh credit.

Expanded from 4 placeholder agents to 11 production-ready agents with a multi-category tag system and grouped filter chips.

• Email Automation, Web Scraper, Resume Processor, Lead Gen, AI Onboarding, Crawler, Capex Intelligence, Investment Intelligence, Manager, and Logger agents added.
• Tags now span 5 categories: Delivery, Domain, Capability, Type, and Status.
• Filter bar is grouped with scrollable chip UI — toggle any filter to narrow the catalog.
• WorkflowCard redesigned: MCP, Webhook, and Full Setup delivery badges with distinct colors.
• Status tags (Live Demo, New, Beta, Popular) render with distinct glow-style colors.

AnswerBot RAG pipeline deployed to production at support.aidesktop.in. SSL-certified, PM2-managed, and live with a real-time chat demo embedded in the marketplace.

• FastAPI + Uvicorn running on port 8001, managed by PM2.
• Heavy ML deps (torch, easyocr, faster-whisper) stripped for VPS — optional try/except fallbacks added.
• Let's Encrypt SSL certificate issued for support.aidesktop.in.
• Nginx reverse proxy configured with 50MB upload limit.
• Live chat demo panel embedded on /workflows/customer-support — real GPT-4o responses, no mock data.

Replaced static "How to access" cards with 4 interactive, expandable delivery option cards on every agent detail page.

• API Integration: live code snippet with copy button, Node.js example pre-filled.
• WhatsApp Integration: no-code setup flow with step-by-step guidance.
• MCP Server: claude_desktop_config.json snippet for Claude Desktop and Cursor IDE.
• Full Setup Service: +₹4,999 add-on with checkbox toggle and dynamic total price.
• Selected delivery method synced to sidebar with colored dot indicator.

Marketplace deployed to Hostinger VPS (Ubuntu 24.04) at vulcan.aidesktop.in with Nginx routing, PM2 process management, and Let's Encrypt SSL.

• Nginx reverse proxy with subdomain routing: vulcan.aidesktop.in → :3000, support.aidesktop.in → :8001.
• PM2 managing the Next.js marketplace and FastAPI AnswerBot processes with auto-restart.
• SSL certificates issued for all active subdomains via Certbot.
• UFW firewall: only ports 80, 443, and 22 exposed.
• Domain aidesktop.in configured with Hostinger nameservers and 5 A-record subdomains.

Fixed 10 build-breaking errors across the codebase introduced by stale code generation. App compiles cleanly with no TypeScript or ESLint errors.

• Removed duplicate component exports across 7 files.
• Fixed toDataStreamResponse() → toTextStreamResponse() for ai@6.x SDK compatibility.
• Moved top-level new Razorpay() inside POST handler to prevent build-time crash.
• Wrapped useSearchParams() in Suspense boundary on checkout page.
• Pinned duckdb==0.10.3 for duckdb-engine compatibility.

Initial production build with Razorpay payment integration, Clerk authentication, and MongoDB purchase records.

• Razorpay order creation and signature verification API routes.
• Clerk authentication with sign-in/sign-up flows and protected purchase gate.
• MongoDB via Mongoose for purchase records and workflow storage.
• Checkout page with dynamic agent/price parameters from URL.
• 4-theme system (charcoal, light, launch, launch-light) via CSS custom properties.