Daily Trend Report for Indie Developers — 2026-08-02

1. Cross-Website Trend Synthesis

Five themes cut across Product Hunt, Hacker News and GitHub today:

A. The Agent "Skills & Memory" economy is forming. GitHub is exploding with agent-skill packs: reverse-skill (+1,320★ today) routes AI coding agents through reverse-engineering/pentest playbooks; Korea's k-skill packages 100+ local-life skills; GitHub's own gh-stack ships a "skill" so agents manage stacked PRs correctly; and TencentDB-Agent-Memory turns team experience into reusable Chat Memory/Skill/LLM-Wiki/CodeGraph assets. Skills are becoming the app store of the agent era.

B. Agent supervision, observability and cost transparency are the new trust battleground. Port22 (PH #3) puts agent approvals on your phone; AgentMicro surfaces Codex status in the menu bar; TraceLLM brings OpenTelemetry to AI apps; DepthData tracks company AI spend. Meanwhile HN's #2 thread (302 pts) is Cursor removing cost info from its usage page — users are furious about opaque token billing. And the ripgrep musl segfault was discovered inside the OpenAI Codex bundle — agents now ship core dev tools, raising the quality bar.

C. Self-hosted and open-source alternatives keep winning. kaneo (+760★) is a slick MIT-licensed Linear/Jira alternative; invidious (+435★) keeps growing; NetBSD 11.0 shipped; and the RSS debate (381 pts) shows pushback on platform lock-in.

D. Voice and video generation are entering production. huggingface/speech-to-speech (+442★) provides a local voice-agent pipeline behind an OpenAI Realtime-compatible API; voice-pro is a fully open-sourced ElevenLabs alternative. On video, ByteDance's Seedance 2.5 (135 pts) does 30-second one-take clips with up to 30 images/10 videos/10 audio references, and TRELLIS.2 generates PBR-ready 3D assets in seconds.

E. AI meets security on both sides. reverse-skill weaponizes agents for authorized pentesting, while the Lean kernel soundness postmortem describes an AI-assisted "disproof" of the Collatz conjecture that exploited a kernel bug, and CISA flagged water-sector PLC targeting.

2. Product Deep Dives

• Port22 — The pain point is real: long agent runs stall waiting for approvals. Port22 attaches to existing Claude Code/Codex sessions, shows the *actual* option buttons on your phone (not guessed keystrokes), and relays E2E-encrypted traffic through a no-read relay; free for one Mac + two sessions. Competitors (Conductor, Superset, Remodex) focus on running many agents; Port22 focuses on the human-in-the-loop UX. Commenters' top request — inline diffs on the approval card — is the obvious next feature.
• Cursor's cost-transparency backlash — Removing cost columns from usage/CSV export triggered 302 points and 135 comments. Developers now treat token metering as a right. Third-party meters like DepthData and menu-bar tools like AgentMicro are early winners.
• DeerFlow 2.0 — ByteDance's 78.7k-star agent harness was rewritten from scratch: sub-agents, sandboxes, long-term memory, extensible skills, MCP, IM channels, Langfuse/LangSmith tracing, and a "one-line agent setup" prompt. Open source is commoditizing agent orchestration — the moat moves to *skills and data*, not the harness.
• reverse-skill — An MIT-licensed skill router pack that turns Claude Code/Codex/Cursor into methodical reverse-engineering and authorized-pentest assistants (APK, ELF, JS encryption, CTF, malware, firmware). +1,320 stars in one day proves huge appetite for structured domain playbooks for agents.
• Seedance 2.5 — Moves video gen from "clips" to "creative works": multi-round extension for multi-minute videos, timestamp-level editing, green-screen and reference-based editing; API coming via BytePlus ModelArk. Great raw material for indie post-production and storytelling tools.

3. Market Implications

• Agent UX is still wide open: approval, monitoring, mobile control, diff review, audit.
• Cost transparency will become table stakes; "FinOps for AI" is a real category.
• Skills/playbook packs are a new distribution channel — a "prompt marketplace" for agents.
• Self-hosting demand persists (privacy, lock-in fear, repo-level AI policies).
• Voice + video generation is now cheap enough to embed in everyday products.

4. Actionable Opportunities

1. Build the "approval card with diff" that Port22 commenters begged for — or a multi-agent dashboard with honest cost telemetry.

2. Ship an AI spend/usage transparency layer for Cursor/Codex/Claude (like DepthData) — publish weekly reports, win trust.

3. Create niche skill packs for your domain (finance, healthcare, e-commerce, local services) in the k-skill/reverse-skill format — the agent does the marketing.

4. Add a voice layer to your product with huggingface/speech-to-speech — local, private, OpenAI-Realtime compatible.

5. Build post-production tooling around Seedance 2.5-class models: storyboarding, character/consistency checking, reference management.

6. Speed up your Postgres tests with pgtestdb's template-cloning approach (~100ms setup).

7. Apply Diátaxis to your docs and add AGENTS.md so AI agents can use your project correctly.