Repository navigation
v4.4.0
v4.4.0
Hand one task to another coding agent through a saved profile with the new agent-profile skill, and run an agent fleet in the Herdr desktop app with the new hc-agent skill, which issue-work-loop now uses there. product-launch plans a launch around one launch story with opt-in campaign tracks, appstore-review-checker catches the patterns behind post-approval App Store removals, and appstore-assets, aso-marketing, landing-page-generator and idea-validator improve. lstack 0.8.0 now includes the two new skills, for 48 skills.
Download the attached lstack bundle, Claude Code plugin, or Codex plugin ZIP, each version 0.8.0 with 48 skills, and follow the included installation and prerequisite guide.
What's new
- New skill
agent-profile1.1.0 hands one task to another coding agent through a named profile: a harness (Claude Code, pi, opencode, Codex, Factory Droid, Devin CLI, Grok Build or Gemini CLI), an optional harness profile, a model, a thinking level and a permission mode. New profiles offer only the CLIs, profiles, models and thinking levels it finds installed. Each run asks where to run (the current directory or a new worktree) and how (foreground or background), shows the exact command, and needs a separate yes for full permissions. opencode profiles can be confined toask(read-only tools) oredits(edits, no shell), and the run refuses to start unless opencode confirms those rules. A killed or timed-out run stops the harness's whole process tree. (#497, #503) - New skill
hc-agent1.0.0 bringsherdr-agent's fleet workflow to the Herdr desktop app, where theherdrCLI reaches a separate TUI server. It drives the app only throughhcandherdr-app: workers open as labelled split panes, a prompt is refused while its target is working, blocked or has no agent status, andcloseneeds--confirmedand closes only this root's workers. It was verified live on macOS with a Claude Code worker; Linux, Windows and other worker kinds are not verified yet.herdr-agent3.3.2 now stops in the desktop app and points there. (#498) issue-work-loop: the loop runs in the Herdr desktop app. In the app, 1.8.0 starts and drives its workers throughhc-agent; the Herdr TUI keepsherdr-agentwith unchanged behavior, and with neither runtime the run stops with No Herdr runtime detected. Permission bypass stays off. The app's on-screen check of a claude or opencode worker's permission mode is not verified live yet, so it fails closed with App boot mode unverifiable whenever the expected mode does not show. (#500)product-launch: one launch story, a copy test and opt-in campaign tracks. 1.3.0 builds the launch around one story (hero, villain, contrarian take, lovers, haters), each part a sourced fact or a founder input, and tests copy for being visualizable, falsifiable and something only you can say. Eight campaign tracks (video, friends and a launch-day invite, a private supporter list, co-launch partners, earned media, an in-person event, founder-run paid amplification and a story owner) are off by default; turning one on sets launch day far enough out for its lead time, and a date you chose is flagged rather than moved. Any paid, gifted or relationship-based post needs a clear in-post disclosure, and fake accounts, bots, engagement pods and undisclosed paid posts are banned. Reddit guidance (1.2.0) starts weekly participation with no product name or link from the account-setup day, drafts mod mail where a subreddit confines promotion, and writes a separate post for each subreddit. The skill still never posts.landing-page-generator1.9.0 applies the same copy test to its hero and reuses the launch story when a launch kit exists. (#462, #501)appstore-assets: real app UI inside the header's safe area. The header image and video now default to a layout that keeps the caption and one hero device inside Apple's art safe area, so a narrow display's crop shows the app and not only the caption, and the review sheet outlines that area. Re-rendering a set made before 1.4.0 withoutcreatives.layoutchangesheader.pngandheader.mp4; setcreatives.layout: "wide"to keep the old composition. (#502)appstore-review-checker: the rejection patterns behind post-approval removals. 1.7.0 checks paywalls closely and fails only what the evidence shows: guaranteed results, invented or random social proof, fake countdowns, hidden billed prices, and missing Terms of Use or Privacy Policy links, while a remote paywall gets a warning to check by hand. It also flags "Allow" buttons and rewards on pre-permission screens, over-the-air update SDKs, "Coming soon" features and marketing in Live Activities, and it audits a live app as strictly as a new one. (#467)aso-marketing: lessons from "The Ultimate Guide to ASO". 1.6.0 puts the keyword before the brand in the title until users search the brand, picks keywords the app can win at its current size, writes benefit-first screenshot captions, times the review prompt after a moment of value, and localizes keywords and screenshots from local research. A new optional Localized Pricing section only recommends tiers and never changes a price, and a Measurement Plan tracks one lever per store submission over at least six weeks. (#488)idea-validator: paid market players and clearer verdicts. The research phase now maps the paid market: the leader, challengers and long tail with dated public figures, pricing tiers, share estimates with confidence labels, the market structure and a market-size estimate, from free public sources only.validate.mdgets a fixed evaluation shape, every step prints a numbered step report, and the verdict inputs are defined, so a free product with demand evidence can reachBuild itand every hard risk is named under Concerns. (#455, #461, #466)- lstack 0.8.0: two new skills. lstack adds
agent-profileandhc-agent, for 48 skills (46 top-level).
Other changes
- Docs, evals and tests out of the plugin folder. Each skill's
docs/README.md,evals/andtests/now live inmeta/<name>/, outside theskills/plugin folder, which holds 425 files, under the plugin directory's 512-file limit. Nothing a skill loads at runtime moved. Pointers to those files inside 16 skills now name themeta/path, each with a patch bump, and the lstack ZIPs no longer include each skill'sdocs/README.md. (#485)
The catalog is version 4.4.0; lstack is version 0.8.0, and individual skills keep their own versions. Detailed developer entries are in CHANGELOG.md.
Thanks to @luongnv89.
Full Changelog: v4.3.0...v4.4.0