Face the Bug

On the table

• Orca — open-source IDE built for running coding agents side by side in isolated git worktrees — Claude Code, Codex, OpenCode — Ghostty-inspired terminal, embedded browser with design mode, mobile companion — onorca.dev
• Herdr — background runtime that keeps coding agents alive when your machine closes or disconnects — persistent terminals across workspaces, working/blocked/idle status tracking, drops into 19+ agent CLIs with no modification — herdr.dev
• beads — the ticket system for the agents we left open at AA11, except someone already built it — a dependency-aware issue graph in a Dolt-backed SQL database, where bd ready returns only the claimable frontier (open beads with no open blockers) so an agent never picks up step three before step one lands. Hash IDs like bd-a1b2 so parallel agents and branches don’t collide on merge, discovered-from edges so work found mid-task gets recorded instead of dropped, and semantic compaction that summarises old closed beads to save context. Wires into Claude Code as a SessionStart hook running bd prime rather than MCP — the docs claim ~1-2k tokens against MCP’s 10-50k. Is the dependency graph the thing markdown plans were always missing, or a project-management layer we’ll resent maintaining? — beads.gascity.com — how it works — github.com/gastownhall/beads
• /powerup — Claude Code’s built-in interactive tutorial, shipped back in v2.1.90 and quietly sitting there ever since — ten animated lessons on what power users actually do: @-mentions, plan mode, /rewind, background tasks, MCP, CLAUDE.md, skills and hooks, subagents, /remote-control, /model and /effort. Five minutes a lesson, progress saved as you go. Show of hands — who’s run it, and what’s in there that the rest of us are still doing the hard way? — docs — commands — v2.1.90 release

Claude 1337 micro

• ⭐ Ctrl+_
  Undo the last input edit
• ⭐ Alt+P
  Switch model without eating your prompt
• ⭐ Shift+Tab
  Cycle permission modes — Manual, acceptEdits, plan, then bypassPermissions or auto where available
• ⭐ Ctrl+L ×2
  /clear and a fresh conversation, within two seconds — fullscreen rendering only; a single press just redraws
• ⭐ Ctrl+K — Ctrl+U — Ctrl+W — Ctrl+Y
  The kill ring — kill to end of line, to line start, or the previous word, then Ctrl+Y pastes the last kill back and Alt+Y cycles the history. Ctrl+U repeats across lines in multiline input
• Ctrl+S
  Stash the prompt — press again on an empty prompt to get it back whole, cursor position and pasted images included
• Esc Esc
  Empty input opens the rewind menu — with text, clears the draft and files it in history so Up recalls it
• Ctrl+O
  Transcript viewer — then [ spills the conversation into native scrollback for Cmd+F and tmux copy mode, v opens it in $EDITOR, { and } jump prompt to prompt
• Ctrl+G
  Open the prompt in $EDITOR — Ctrl+X Ctrl+E is the readline-native binding
• Ctrl+R
  Reverse-search command history
• Ctrl+X Ctrl+K ×2
  Stop every background subagent in the session, within three seconds
• Ctrl+B
  Background a running task — tmux users press twice
• Alt+T
  Toggle extended thinking — no effect on Fable 5, which always thinks
• Alt+O
  Toggle fast mode
• Ctrl+T
  Show or hide the task checklist
• Ctrl+Z
  Suspend to your shell — fg to come back
• Ctrl+D ×2
  Exit, within 800ms — with text in the prompt it deletes forward instead
• ?
  On an empty input, print the shortcut panel for your terminal
• !
  Shell mode — run a command and drop its output into the session
• Ctrl+J
  Insert a newline — the control sequence works in any terminal with no configuration at all
• \ + Enter
  The other universal newline escape — Shift+Enter is native in iTerm2, Ghostty, Kitty, WezTerm, Warp, Apple Terminal and Windows Terminal, and Option+Enter works on macOS once Option is set as Meta

What we talked about

Pydantic

The type-hints-as-schema library that quietly became the substrate under half the agent stack — validation core in Rust, models that emit JSON Schema, and from there Pydantic AI for agents and Logfire for watching what they cost. Ties straight back into the structured-output thread: if the schema is the contract you hand a model, Pydantic is where most Python people already write it. Worth a lap around the table on whether typed validation is the right place to catch a model’s bad output, or just where the error surfaces loudest. — pydantic.dev/docs — github.com/pydantic/pydantic

n8n

The node-graph automation tool — fair-code, self-hostable, 400-odd integrations — that’s spent the last couple of years quietly growing native AI nodes and becoming the place non-agent people wire LLMs into real systems. Sits at the opposite end of the spectrum from the CLI harnesses we keep comparing: the workflow is drawn, explicit and inspectable, rather than decided by a model at runtime. Is a visual graph the honest form for agent plumbing, or the thing we’ll rebuild in code the moment it gets complicated? — n8n.io — github.com/n8n-io/n8n

Video editing 101

A whole lap of the table on where an agent can actually touch video — Remotion for the parts that are really just React rendered to frames, one of the many DaVinci Resolve MCP servers for driving a real NLE from the harness (most want Studio, since external scripting is paywalled on the free edition), colour correction as the step everyone agrees is the taste-dependent one, and Kling for generating the footage in the first place. The split that emerged: code-shaped work automates cleanly, eye-shaped work doesn’t. Where’s your line — what would you hand an agent, and what stays on your own timeline? — remotion.dev — kling.ai

Nano Banana keyframes

The workflow that came straight out of the video thread — don’t ask a video model to invent the whole shot, generate the keyframes as still images with Nano Banana (Gemini’s image family, now four models deep) where you get consistent characters and multi-reference control, then hand those frames to a video generator to fill the motion between them. Storyboard first, interpolate second, which puts the taste decisions back in the cheap, fast, editable step. Does frame-first beat prompting a video model directly, or are you just moving the incoherence one layer down? — deepmind.google — Gemini Image — Gemini API — image generation

Claude Design

Anthropic Labs’ visual surface — prompts, images, docs or a whole codebase in, prototypes, slides, mockups and landing pages out, with your own design system applied. Launched in April, took a chunk out of Figma’s share price on the day, then got overhauled in June with brand-compliance controls, code round-trips and a /design workflow from the terminal, after everyone discovered how fast the research preview ate a weekly token allowance. Same question as the video thread, different medium: is this the layout-and-boilerplate step automated away, or the taste step handed to a model? — claude.ai/design — anthropic.com — Introducing Claude Design

Stitch

Google Labs’ answer to the same itch as Claude Design — describe a UI in a prompt or hand it a reference image, get screens back, then push them out to Figma or straight to frontend code. Immediately the comparison the table wanted: two big labs shipping the same “prompt in, designed screens out” product within a year of each other, both aiming at the handoff rather than the canvas. Has anyone got real screens out of either that survived contact with a design review — or is this still the fastest way to make something that looks finished and isn’t? — stitch.withgoogle.com

Agentic commerce

The plumbing for letting an agent actually buy something — ACP, the OpenAI-and-Stripe spec open-sourced in late 2025 and now on dated snapshot releases, with cart, feeds, orders, auth and MCP folded in; Google’s UCP arrived in January with Shopify, Target and Walmart behind it, aiming at the whole journey rather than just checkout. Standards race, same as MCP two years ago, except the failure mode here spends your money. Who at this table would give an agent a payment credential, and what’s the smallest blast radius you’d accept? — agenticcommerce.dev — github.com/agentic-commerce-protocol

Userback

The other end of the design thread — a widget your users click to file feedback in place, with the annotated screenshot, the console log and the session context attached, so the bug report arrives already reproducible. Interesting now less as a support tool and more as an input format: that bundle is roughly what an agent needs to go fix the thing without a round of “which button, on which page?”. Is structured user feedback the next obvious agent input, and would you let one open the PR straight off a report? — userback.io

Frustration-triggered ticketing

The idea the Userback thread ran straight into — stop waiting for people to file bugs. Watch the webcam for the face someone makes when the app breaks, and when it fires, cut the last minute of session recording and open the ticket automatically. Nobody has to describe anything; the agent gets the replay, the console and the timestamp. Half the table’s reaction was “that’s the missing input”, the other half’s was “you want to put a camera on your users”. Rate it — genuinely the best bug report you’ll ever get, or the demo that gets you a privacy review and a headline?

News since last assembly

Floor: 2026-08-01 (AA11 — Guardrails Off). Generated 2026-08-14.

New Claude Code commands & features

• Todo/task tools retired on the newest models (v2.1.233, 2026-08-14) — TaskCreate/Get/Update/List and TodoWrite are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and newer; CLAUDE_CODE_ENABLE_TODO_TOOLS=1 brings them back — release
• Subagent forking on by default (v2.1.232, 2026-08-13) — a subagent_type: "fork" subagent inherits the full conversation and prompt cache; non-teammate spawns in interactive sessions now run in the background — release
• Cross-session @ mentions (v2.1.232, 2026-08-13) — type @ to mention another Claude session by name, and SendMessage delivers straight to an exact live-session match — release
• claude self-hosted-runner (v2.1.224, 2026-08-07) — turns your own machines or containers into somewhere web, mobile and desktop sessions can run — Team and Enterprise — release
• 200-subagent cap removed (v2.1.224, 2026-08-07) — long sessions no longer refuse new subagents; concurrency and depth limits still apply — release
• /review aliases /code-review, with PRs (v2.1.224, 2026-08-07) — /code-review <level> <pr#>, and the effort level persists between runs — release
• archive plugin source (v2.1.224, 2026-08-07) — install plugins from an HTTPS zip with optional SHA-256 pinning — release

Codex

• [2026-08-07] 0.147.0 — portable Agent Plugins searchable across local, personal, workspace and remote catalogues, --approve-for-me for auto-reviewed approvals, MCP 2026-07-28 with paginated discovery — and codex exec --full-auto is gone, use --sandbox workspace-write — release

Adjacent tools

• [2026-08-14] SpaceX closed its $60B all-stock acquisition of Anysphere — Cursor folds into a new SpaceXAI division alongside Grok — cursor.com/blog — CNBC
• [2026-08-13] Cursor cloud agents start ~3x faster — prebuilt “builds” keep an environment warm instead of booting and cloning per session — cursor.com/changelog
• [2026-08-12] DeepSeek publicised a “Harness Team” — a verified WeChat account and Beijing job listings for an agentic coding product aimed at Claude Code, with V4 Pro updated for better agentic work the same day — Bloomberg

Simon says

• [2026-08-07] “Now we have a timeline of the OpenAI accidental attack against Hugging Face” — OpenAI’s late-added Black Hat talk, video now up — post
• [2026-08-04] LLM 0.32 — reasoning traces to stderr, server-side provider tools, content-addressable SQLite logs, GPT-5.6 Luna as the new default — post

Research & papers

• [2026-08-04] “Self-Evolving Coding Agents” — a taxonomy of what evolves, when it evolves, and which software-specific evidence drives it — arxiv.org/abs/2608.03392

Notable posts

• [2026-08-14] “Why does Opus 5 feel worse to work with?” — 704 points and climbing: the argument is that benchmark pressure selects for models that make bold assumptions instead of stopping to ask. The thread’s fix is prose-level — “only report to me in ASD-STE100 Simplified Technical English” — and, simpler still, just typing tldr — post — HN
• [2026-08-05] Black Hat USA — one GitHub issue from an account with no repo privileges reached the CI runners behind Anthropic’s and Google’s own agent repos, in the configurations both ship by default — CVE-2026-54316 and Gemini CLI’s CVE-2026-12537 at CVSS 10.0, both patched — The Hacker News
• [2026-08-08] CSA research note on the same work — the failure sat in the harness, not the model: one layer marks a value safe and a later layer runs it anyway — CSA

Topics worth a 5-min slot

1. Task tools retired on the newest models — Claude Code just took away its own todo list on Opus 4.8+. Does the plan live in the model’s head now, or in something like beads?
2. Fork-by-default subagents — children inherit the conversation and the prompt cache, and the 200-spawn cap is gone. What does that change about how we fan out?
3. Opus 5’s prose — the loudest thread on HN this fortnight, and the community’s answer is a 1980s aerospace writing standard. What’s actually in your CLAUDE.md that works?