Changelog

v5.5.02026-08-21

Security fix: update as soon as you can, and re-run the install once.

Opening someone else's project could run their code on your machine.

v5.4.02026-08-16

Installing the kit no longer removes the other automations you had set up.

Older versions replaced that settings file wholesale, silently, so if you installed before today it is worth opening it and checking that anything you set up yourself is still listed. The quality gate also stops blocking simple questions it should let through, and because it no longer mistakes its own text for the assistant's work, expect it to speak up more often than it used to.

v5.3.02026-06-23

The marketing skill can now help you design the offer itself, not just how you describe it.

v5.2.02026-06-17

Each plan now installs the part of the kit it covers.

v5.1.02026-06-16

The quality gate now works when the kit is installed for your whole account, not only inside one project.

v5.0.02026-06-14

Renamed to SummonAI Kit, and the kit became one curated set instead of a menu of per-tool add-ons.

v4.4.02026-06-10

v4.3.02026-06-07

v4.2.02026-06-06

The generated harness now pushes the agent toward the right engineering instincts without hardcoding any provider, library, or task.

All guidance is language/framework/platform agnostic and integrated into the new architecture (hook context + skill patterns + subagent briefs) — no separate legacy hook file.

v4.1.02026-06-06

Generated kits previously shipped harness agents (implementer/verifier/reviewer/closer/retro) that nothing ever invoked.

The Stop hook now enforces the workflow: the first three gates must each run as their own subagent, in sequence (implementer → verifier → reviewer), before a turn can end; closer/retro remain receipt sections the lead writes.

v4.0.12026-06-06

The agent-body sanitizer now cuts everything before the first markdown heading, so conversational preambles the model occasionally emits ("Now I have enough context.

Let me write…") never make it into generated agent files. Regenerated the shipped harness agents to drop the stray line.

v4.0.02026-06-06

Skill/agent/CLAUDE.md generation no longer runs a per-skill LLM pipeline.

The old generator (analyzer prompts, per-skill generation, interactive generation hooks) is removed (−16k lines).

v3.8.12026-06-03

- fix(cli): revert skill/CLAUDE.md generation back to the Sonnet model.

The v3.8.0 Haiku switch produced unreliable, hallucinated library prose (e.g. oRPC .query/.mutation + @trpc/client instead of .handler(), a fabricated Alchemy class API, an invalid 'use compiler' directive, motion skills omitting the actual animation library). The A/B eval gate missed it because the measured uplift comes from the verbatim-injected managed sections rather than the authored prose. Reverting prioritizes generated-skill correctness over generation speed. Analyzer/discovery calls remain on Haiku; the --effort wiring is retained. - perf(cli): raise subagent generation concurrency to 5 across all engines (was 2 for Claude, 3 for others), matching skill generation.

v3.8.02026-06-01

- perf(cli): run Claude skill and CLAUDE.md generation on the Haiku model instead of Sonnet (~2.7x faster per call).

Quality-gated via the A/B eval harness — Haiku-generated skills match Sonnet's treatment scores because the managed sections are injected verbatim post-generation rather than authored by the model. - Wire a --effort level into the Claude CLI invocation (low for generation, high for the eval judge).

v3.7.02026-05-30

- feat(cli): self-improvement eval loop + hardened Stop-hook + agnostic generator contracts (v3.7.0)

v3.6.12026-05-24

- Make interactive CLI prompts accept line-buffered input in remote terminals, including dirty-worktree confirmation and device authentication shortcuts.

v3.6.02026-05-22

- Add broad subagent advantage guidance so generated and fallback subagents ask clarifying questions when needed, inspect repo evidence, and ship complete repo-consistent slices.

- Generalize skill and hook quality guidance away from task-specific examples toward user-visible correctness, integration quality, reliability, maintainability, and operability. - Run skill generation with five concurrent tasks across interactive and kit update flows. - Make summonaikit workflow catalog fall back to the bundled workflow catalog when the remote premade-skill endpoint returns 404.

v3.5.82026-05-17

- Add UI/UX skill generation contracts with a dedicated ux skill, frontend design quality gates, and advisory UI hook checks.

- Harden webhook/payment task contracts around fail-closed signature verification, atomic event claims, business-level idempotency, and relational uniqueness.

v3.5.72026-05-16

- Make generated Stop hook contract checks advisory so repository-wide dirty diffs cannot trap Claude Code in a completion loop.

v3.5.62026-05-16

- Harden generated agent validation against permission prompts, manual paste artifacts, placeholder bodies, and customization recaps.

- Replace invalid backend and DevOps generated agents across Claude, Cursor, and Codex targets.

v3.5.52026-05-16

- Release the skill generation reliability fixes from v3.5.4.

v3.5.42026-05-16

- Make the skill generation progress view resilient to repeated API failures by moving long retry errors into a compact summary.

- Reduce Claude skill-generation concurrency to avoid bursts of failing requests during large skill sets. - Compact rich per-skill prompts and fall back to compact/local skill generation when transient API failures exhaust retries.

v3.5.32026-05-16

- Fix generated hook self-interference during internal Claude generation.

- Improve skill validation for documents ending with code fences. - Tighten prompt-aware hook activation to avoid unrelated skill noise. - Keep CLAUDE.md and AGENTS.md production contract guidance aligned.

v3.5.22026-05-16

- Add Skill System V3 production contracts, prompt-aware hooks, and generated agent cleanup.

- Regenerate Claude, Cursor, and Codex skills/agents with native-first capability guidance.

v3.5.12026-05-16

- Update commit message instructions - ship rework - Add platform-native guidance to generated CLAUDE docs - Harden CLI skill generation quality checks

v3.5.02026-05-15

- Add platform-native guidance to generated CLAUDE docs - Harden CLI skill generation quality checks - feat(cli): signal-driven, multi-language skill detection + doc-aware generation

v3.4.02026-05-15

- (no CLI commits found, please describe the change)

v3.3.02026-05-15

- (no CLI commits found, please describe the change)

v3.2.02026-05-15

- feat(cli): signal-driven, multi-language skill detection + doc-aware generation - fix - fix

v3.1.12026-05-14

- Handle premade skill catalog 404s with bundled fallbacks - fix(cli): stabilize login URL and add copy shortcut - things

v3.1.02026-05-07

- (no CLI commits found, please describe the change)

v3.0.22026-05-06

- (no CLI commits found, please describe the change)

v3.0.12026-05-06

- fix - workflows

v3.0.02026-05-05

Workflows is here.

v2.0.282026-05-04

Polish and stability.

v2.0.272026-04-30

Generation reliability boost, flaky network blips during a run are now caught and retried automatically.

v2.0.262026-04-30

Smoother generation runs and clearer error messages, when something goes wrong mid-flow you actually see what to fix.

v2.0.252026-04-14

Stricter skill-name validation.

Generated skill folders now follow the Agent Skills spec exactly (lowercase, hyphens only, no leading or consecutive hyphens, ≤ 64 characters).

v2.0.232026-03-27

Behind-the-scenes refinements.

v2.0.222026-03-26

Faster, more reliable skill and subagent generation.

Concurrency is now managed per AI CLI so multiple skills can build in parallel without hitting rate limits.

v2.0.192026-03-26

Sharper detection of which AI CLIs you have installed (Claude Code, Cursor, Codex).

Fewer "not found" messages on machines where the CLI is just in a non-standard path.

v2.0.172026-03-25

Steadier hand during long generation runs.

v2.0.142026-03-23

Real-time progress for skill and subagent generation.

You now see each skill finish individually instead of one global spinner, much easier to tell what's happening.

v2.0.102026-03-11

Updater now handles slow or flaky networks gracefully, no more crashes on startup when the version check times out.

v2.0.92026-03-10

Quality-of-life touches.

v2.0.82026-03-06

The generation model is now picked automatically per AI CLI (Sonnet for Claude Code, Opus for Cursor, Codex 5.3 for Codex).

The legacy model picker is gone, one less question on every run.

v2.0.32026-02-17

Cursor integration runs more reliably across edge cases.

v2.0.22026-02-13

Cursor and Claude Code now share matching argument handling, the same prompts work on both without surprises.

v2.0.12026-02-10

You can now pipe prompts via stdin (echo "..." | summonaikit ...) for shell integration.

v2.0.02026-02-03

Multi-target support is here: install skills into Claude Code, Cursor, Codex, or all three from a single run.

Pick which AI CLI does the generation pass and pick which directories receive the output independently.

v1.2.142026-02-03

Fixed a race condition that occasionally caused empty CLAUDE.md generation when subagents and skills were created in parallel.

v1.2.112026-02-02

First version with multi-target support.

Choose which AI CLIs you want skills installed for during the target-selection step.

v1.2.52026-01-26

The CLI now auto-checks for updates at startup and tells you if a new version is available.

v1.2.42026-01-26

Clearer error messages when an AI CLI is installed but not yet authenticated.

The preflight step now tells you exactly what to fix instead of bailing.

v1.2.22026-01-26

Automatic retry for transient errors during generation.

Network blips, rate-limit pushes, and momentary failures no longer abort the whole flow.

v1.2.12026-01-26

Fixed a race condition when generating multiple skills in parallel - sometimes the first skill's directory write would conflict with the second.

v1.2.02026-01-21

The CLI now picks up MCP servers you have configured in your AI CLI and uses them as context when generating skills.

Skills referencing Context7, Sentry, Stripe, etc. now actually mention the right MCP tool names.

v1.1.42026-01-19

Smoother release pipeline under the hood.

v1.1.32026-01-18

Fallback storage for environments without a system keychain (WSL, containers, CI).

Auth tokens now save to ~/.summonaikit/credentials.json when Bun.secrets isn't available.

v1.1.12026-01-16

Context7 integration.

When the CLI detects you have Context7 set up, it recommends Context7 doc lookups inside the generated skills (so your skills pull live framework docs instead of stale training data).

v1.0.102026-01-15

Foundation tidying, paving the road for the features in 1.1.x.

v1.0.92026-01-14

Better auth flow with clearer error states.

The browser-open prompt is more helpful when the device-flow code expires before the user confirms.

v1.0.82026-01-14

First version with multi-engine support, the CLI can now use Claude Code, Cursor, or Codex to run generation.

The new preflight step verifies your AI CLI is installed and authenticated before starting, saving you from a confusing failure mid-flow.

v1.0.52026-01-09

Temp files now clean up reliably even when you Ctrl+C the CLI mid-run.

No more leftover summonaikit-* directories in your tmp folder.

v1.0.42026-01-04

Installer improvements.

Clearer platform detection (Apple Silicon vs. Intel macOS, Linux x64 vs. arm64) and better error messages when the expected binary isn't available for your platform.

v1.0.32026-01-04

Sharper analyzer prompts.

Skill detection is more accurate across frontend (React, Vue, Svelte), backend (Hono, Express, FastAPI), and testing (Vitest, Jest, Playwright), fewer missed skills, fewer duplicates.

v1.0.12026-01-03

Initial public release.