Skip to content

Releases: theam/claude-dev-kit

v0.19.4

Choose a tag to compare

@atamanvega atamanvega released this 14 Aug 10:41
b79f51f

Added

  • Listing screenshots wired into the plugin interface (mirrored to the portable manifest): the GitHub PR (AI-generated badge + verification evidence), the Codex plan-approval gate, and the updated Jira ticket. Listing metadata is now complete — the package is fully submission-ready pending OpenAI identity verification. (Raw captures; design can drop in cropped/anonymized versions under the same filenames.)

v0.19.3

Choose a tag to compare

@atamanvega atamanvega released this 13 Aug 18:18
d2e9c59

Added — directory-listing metadata (submission-ready)

  • Plugin interface now carries the TAM brand assets (logo/logoDark icons, composerIcon = the t• symbol, brandColor #000000), termsOfServiceURL (Apache LICENSE), and 3 defaultPrompt examples — mirrored into the portable manifest. Assets in plugins/fullstack-dev-kit/assets/ (icons/symbol only, no licensed fonts).
  • SUBMISSION.md runbook for the OpenAI / Cursor / VS Code directories (skills-only strategy + 5 positive / 3 negative review test cases).

Additive metadata + assets — no plugin behavior change. Publishing to the directories still waits on OpenAI identity verification (a manual portal step); this release just makes the package submission-ready.

v0.19.2

Choose a tag to compare

@atamanvega atamanvega released this 10 Aug 10:50
1575929

Changed

  • Wizard install step is now agent-aware. It no longer assumes Claude Code: the "install for Claude Code" step appears only when the claude CLI is detected, and the closing hint shows the right entry point per host (Claude /fullstack-dev-kit:work-story vs Codex/Cursor $work-story). Sections renumbered (Claude · Codex · Cursor); intro reworded for the multi-client kit.

Wizard-only UX change — no host's plugin behavior changes. The create-dev-kit npm package (→ 0.1.2) must be republished for users to get it.

v0.19.1

Choose a tag to compare

@atamanvega atamanvega released this 10 Aug 09:43
387304a

Added

  • Wizard auto-installs for Cursor. npm create @theagilemonkeys/dev-kit now detects Cursor and drops the portable plugin into ~/.cursor/plugins/local/fullstack-dev-kit (skills auto-load on restart; enable MCP from Cursor Settings → Tools & MCP). Additive, idempotent, opt-in. No telemetry on Cursor — tracking stays Codex/Claude only. Copilot install stays manual (VS Code UI action).

No change to Claude Code / Codex behavior. (To reach users, the create-dev-kit npm package must be republished — bumped to 0.1.1.)

v0.19.0

Choose a tag to compare

@atamanvega atamanvega released this 10 Aug 09:34
7c9ab34

Added — portable, multi-client packaging (Agent Plugins 1.0.0)

  • Dual-emit manifest. The plugin now carries both a Codex-native .codex-plugin/plugin.json and a portable root plugin.json conforming to the open Agent Plugins 1.0.0 standard, so the same plugin loads in Codex and other compatible clients (Cursor, GitHub Copilot in VS Code, …). Generated from one source; interface rides under the com.theagilemonkeys.dev-kit extensions namespace.
  • Portable work-story orchestration skill. The end-to-end ticket→PR flow now travels to hosts without an orchestrator subagent: a work-story playbook the host agent runs itself. Shipped to the plugin bundle (Codex + portable clients) only — Claude Code is unchanged and keeps its richer /work-story command + coding-agent subagent.
  • README install guide for Codex, Copilot (VS Code) and Cursor, with verified steps and honest status caveats.
  • Validator extended to cover the portable manifest and Agent Skills naming; runs on every build.

Verified end-to-end on codex-cli 0.147. Cursor/Copilot are format-compliant but not yet verified by us. No change to Claude Code behavior. Telemetry remains wizard-only and opt-in.

v0.18.1

Choose a tag to compare

@atamanvega atamanvega released this 10 Aug 08:47
faee776

Changed — Codex plugin packaging hardening (Phase 1 toward the Agent Plugins standard)

  • Validator (scripts/validate-codex-plugin.mjs, runs at the end of build-codex-plugin.mjs, zero-dep): checks the marketplace + plugin manifests against the enums confirmed on codex-cli 0.147 and enforces a self-contained bundle — every instructions/… a bundled skill references must ship in the plugin. Would have caught the packaging bug pr-review found on #28.
  • Enriched the plugin interface (privacyPolicyURLTELEMETRY.md) toward Plugins Directory listing readiness.

Removed

  • The unverified codex/hooks.json template — Codex session hooks are feature-flagged on the desktop app and version-specific; the kit delivers Codex telemetry via the hook-independent sweep.

No change to Claude Code behavior — Codex packaging only.

v0.18.0

Choose a tag to compare

@atamanvega atamanvega released this 05 Aug 15:11
baaab43

Added — OpenAI Codex support (experimental)

The kit now targets OpenAI Codex (CLI + desktop app) alongside Claude Code.

  • Native Codex plugin. The repo doubles as a Codex marketplace — install with codex plugin marketplace add theam/claude-dev-kitcodex plugin add fullstack-dev-kit@claude-dev-kit. Skills + the instructions/ they reference are generated from the canonical top-level copies (scripts/build-codex-plugin.mjs) — one source of truth, self-contained bundle.
  • Telemetry differentiation. New agent property (claude-code · codex) so PostHog can split usage by tool. One telemetry.mjs serves both hosts; on Codex it reads token totals from the session rollout JSONL and derives the surface (cli/vscode). Delivery via a hook-independent sweep (launchd on macOS) since the Codex desktop app feature-flags session hooks. Consent/install-id/org/dedup shared across both hosts.
  • Unified installer. create-dev-kit installs natively for whichever of Claude Code / Codex it detects, installs the curated connector your tracker choice implies (Jira → atlassian-rovo, Linear → linear, Figma → figma; GitHub/Azure use their CLIs), and schedules the telemetry sweep.

Validated end-to-end against real codex-cli 0.147: plugin install, in-app skill invocation (the kit's own $pr-review reviewed the PR and caught a real bug), Atlassian connector + issue-fetch reading a Jira ticket, and the telemetry parser/sweep on a real rollout.

Note: the telemetry relay must be redeployed for the agent property to reach PostHog (contract change) — otherwise events still arrive, just without the agent tag. No impact on Claude Code users: the Codex packaging lives in paths Claude Code doesn't load.

v0.17.0

Choose a tag to compare

@atamanvega atamanvega released this 29 Jul 18:26
8330a11

Added

  • Accessibility review for frontend changes — folded into pr-review as a conditional dimension (no new agent, no extra pass):
    • Runs only when a diff changes user-facing UI in a frontend stack; zero token cost on backend/non-UI changes.
    • Targets WCAG 2.2 AA and cites the specific Success Criterion per finding. Uses the repo's own a11y tooling (axe-core, eslint-plugin-jsx-a11y, Lighthouse) when present; never scaffolds one.
    • Automatic, non-blocking by default, no setup. Optional "a11y" key in .claude/dev-kit.json: auto (default) · required (blocking) · off — hand-edit or just ask the kit to change it.
    • coding-agent also nudges the implement step to write accessible markup up front, so it's covered proactively — not only caught at self-review.

v0.16.1

Choose a tag to compare

@atamanvega atamanvega released this 29 Jul 18:26
e06d0eb

Fixed

  • Telemetry duplicate events: a session could be reported twice (Claude Desktop firing SessionStart more than once on reopen). Sends are now guarded by an at-most-once claim (exclusive sent-<id> sentinel, GC'd), and the per-session dedup id is derived deterministically from install_id + session_id so every code path yields the same id and PostHog collapses re-sends.
  • Client-only change; no relay redeploy needed.

v0.16.0

Choose a tag to compare

@atamanvega atamanvega released this 28 Jul 12:42
92932a3

Rolls up 0.14 → 0.16 (v0.13.0 was the previous release). Full detail in CHANGELOG.md.

Telemetry robustness (0.14)

  • Faster delivery (outbox stale threshold 30→10 min) so sessions that never fire SessionEnd (Claude Desktop "new chat") still send.
  • Per-session dedup (event_uuid) so re-sends count once; org attribution falls back to the user-level config (works in worktrees); relay suppresses GeoIP (it only ever saw the relay's IP, never yours).

PR review loop (0.15)

  • PR intent statement as the scope ruler; fix-pr triages every item to FIX_NOW / DEFER_TO_ISSUE / DISCARD (deferrals file a tracked issue via the tracker adapters); a late-feedback watcher waits out slow bot/CI reviews and loops; consent-first replies. pr-review gains performance/duplication dimensions.

Lightweight review facets (0.16)

  • Optional, proportional adversarial check in pr-review (high-stakes diffs only) and an optional "Delivery readiness" section in create-pr — folded into existing skills, no new agents, no extra passes by default (keeps token cost down).

Update via auto-update at session start, or claude plugin marketplace update claude-dev-kit.