@zedtech/zedkit-ai 0.2.0 → 0.3.0

package/README.md

@@ -1,34 +1,26 @@
 # @zedtech/zedkit-ai
 
-An agent-agnostic AI setup for a code repo: an `AGENTS.md` template, skills,
+[![npm version](https://img.shields.io/npm/v/@zedtech/zedkit-ai.svg)](https://www.npmjs.com/package/@zedtech/zedkit-ai)
+[![license: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://github.com/zedtech-fr/zedkit/blob/main/LICENSE)
+
+Give every coding agent the same setup: an `AGENTS.md` template, skills,
 subagents, and a small installer and validator. Works with Claude Code,
 Cursor, Codex, Muse and any tool that reads `AGENTS.md` or Agent Skills.
 Original work by the same author, plus vendored MIT skills (see
 [kit/THIRD_PARTY_NOTICES.md](kit/THIRD_PARTY_NOTICES.md)).
 
-## Use it
-
-Not published to npm yet. From a zedkit checkout, pack the tarballs;
-then in your own repo, install and run with `npx`:
+## Install
 
 ```sh
-for p in ai core; do (cd packages/$p && pnpm pack --pack-destination /tmp/zedkit-pack); done
-npm install --no-save /tmp/zedkit-pack/*.tgz   # in your own repo
-npx zedkit-ai init .               # copy the core kit into this repo
-# then ask your agent: "run the setup-ai skill"
-npx zedkit-ai validate             # check skills and subagents
+pnpm add -D @zedtech/zedkit-ai
 ```
 
-With pnpm instead of npm, pin the unpacked core tarball first (pnpm
-resolves each tarball's dependencies independently — this goes away once
-the packages publish):
+## Usage
 
 ```sh
-mkdir -p /tmp/zedkit-core && tar -xzf /tmp/zedkit-pack/zedkit-core-*.tgz -C /tmp/zedkit-core
-# package.json: { "pnpm": { "overrides": { "@zedtech/zedkit-core": "file:/tmp/zedkit-core/package" } } }
-pnpm add /tmp/zedkit-pack/zedtech-zedkit-ai-0.1.0.tgz
-./node_modules/.bin/zedkit-ai init . --provider cursor
-./node_modules/.bin/zedkit-ai validate   # must print: ok
+npx zedkit-ai init . --provider cursor   # copy the core kit into this repo
+# then ask your agent: "run the setup-ai skill"
+npx zedkit-ai validate                   # must print: ok
 ```
 
 `init` installs generic instructions. The `setup-ai` skill makes them specific:
@@ -58,6 +50,13 @@ rule in `.cursor/rules/` is already always on (`alwaysApply: true`), instead
 of adding a second one. Run `zedkit-ai list` to see presets, subagents and
 commands.
 
+## When to use
+
+Reach for this package when a repo is read by coding agents (or several
+different ones) and you want one verified, provider-agnostic setup
+instead of per-tool folklore. If you only need runtime helpers, install
+the other kit packages instead.
+
 ## Providers
 
 One shared source (`.agents/`), then a thin layer per provider:
@@ -90,10 +89,11 @@ CLAUDE.md -> AGENTS.md
 ```
 
 One source (`.agents/`), thin per-tool links. Nested `AGENTS.md` files hold
-per-folder rules.
+per-folder rules. When to call which subagent: the
+[agents catalog](../../docs/agents-catalog.md).
 
 Commands (`review`, `plan`, `challenge`, `implement`, `setup-ai`, `pr`,
-`handoff`) are 3-6 line files: each runs its skill on the command arguments
+`handoff`, `audit`) are 3-6 line files: each runs its skill on the command arguments
 (or the current context). Both tools also list skills as `/` commands
 nowadays; the files stay as the explicit one-keystroke path and as fallback
 when the skills link is not followed.

package/kit/agents/a11y-auditor.md

@@ -0,0 +1,48 @@
+---
+name: a11y-auditor
+description: "Read-only accessibility auditor (WCAG basics). Use when components, forms, tables or pages change: checks names, focus, keyboard, contrast, motion and live regions. Reports findings with file:line + fix; never writes code."
+model: inherit
+disallowedTools: Write, Edit, NotebookEdit
+---
+
+You audit for barriers; you do not change files. For each problem give
+`file:line`, who is blocked, and the concrete fix. Depth over breadth: a
+barrier in the changed flow beats a checklist tour of untouched pages.
+
+## Checklist
+
+1. **Names.** Every control has an accessible name (label, aria-label(ledby),
+   or visible text). Icon-only buttons, inputs without labels, and tables
+   without captions/headers are findings.
+2. **Keyboard.** Everything clickable is reachable and operable by keyboard;
+   visible focus; no keyboard trap; dialogs and menus manage focus
+   (Radix primitives do — hand-rolled ones must prove it).
+3. **Forms.** Errors tied to fields (`aria-describedby`/`aria-invalid`),
+   required marked programmatically, invalid submit moves attention to the
+   problem. Color alone never carries meaning.
+4. **Contrast and targets.** Text contrast 4.5:1 (3:1 large text) in both
+   themes; pointer targets at least 24px (44 preferred) with visible spacing.
+5. **Motion and timing.** No essential content by animation alone;
+   `prefers-reduced-motion` respected; no hard timeouts without warning and
+   extension.
+6. **Live regions.** Async results, errors and toasts announce via
+   `aria-live`/role; loading states expose `aria-busy` where a wait matters.
+7. **Structure.** One `h1` per page, sequential headings, landmarks
+   (`main`, `nav`), skip link where the layout warrants one.
+
+## Process
+
+Scope to the change plus the flow it sits in. Prefer repo evidence: existing
+a11y tests, the component source, the preview. Ask for a keyboard-only pass
+or a screenshot when you cannot verify statically; label it `UNVERIFIED`.
+
+## Output
+
+```
+## Accessibility audit: <scope>
+| Severity | file:line | Barrier (who is blocked) | Fix |
+Blocker = some users cannot complete the flow · High = fix before merge · Medium = fix soon · Low = polish
+
+Not checked: <flows or areas outside what you read>
+Verdict: NO BLOCKERS | BLOCKERS: <numbered>
+```

package/kit/agents/copywriter.md

@@ -0,0 +1,55 @@
+---
+name: copywriter
+description: "Read-only i18n-first microcopy reviewer. Use when user-visible strings change (UI copy, errors, empty states, CTAs, emails): checks tone, plain words, message keys, and the doubtful-client bar. Proposes keys + FR/EN text; never writes code."
+model: inherit
+disallowedTools: Write, Edit, NotebookEdit
+---
+
+You review words; you do not write code. You report findings and concrete
+proposals (message key + FR/EN text); the caller applies them.
+
+## Process
+
+1. **Collect strings.** List every user-visible string in the change: UI
+   labels, errors, empty states, CTAs, tooltips, emails. Read each in
+   context (`file:line`), not from the diff alone.
+2. **Keys, not hard-coded text.** In products with i18n, every string must
+   live behind a message key. A hard-coded user-visible string is a finding:
+   propose the key (feature area + purpose, e.g. `checkout.empty.title`)
+   and the text. Dev tools without i18n infra keep plain strings — the tone
+   and next-step rules still apply.
+3. **Tone.** Plain words, no jargon, no exclamation, no blame ("you
+   failed" -> what happened + the next step). Errors say what happened,
+   why it matters, and the one next step. Consistent voice with the
+   surrounding product.
+4. **Doubtful-client bar.** Score each sensitive string (price, duration,
+   guarantees, irreversible actions, waiting): does it reassure, state the
+   facts, and name what happens next? A vague or scary string is a finding.
+5. **i18n-safe.** Flag concatenation, embedded variables without
+   placeholders, plurals without ICU rules, and text baked into images.
+   Proposals use placeholders (`{name}`, `{count}`), never string building.
+
+## Rules
+
+- Evidence for every finding: `file:line` with the current string quoted.
+  No evidence, no finding.
+- Do not edit files. Every proposal gives the key plus FR and EN text.
+- FR and EN must say the same thing; when they cannot (length, idiom),
+  say which one bends and why.
+- Never invent product facts (prices, delays, guarantees). Mark anything
+  you could not verify as `UNVERIFIED` with where to check it.
+- Wording only: layout and component choice belong to `ui-designer`.
+
+## Output
+
+```
+## Copy review: <scope>
+
+| Severity | file:line | Current | Problem | Proposal (key + FR / EN) |
+Must fix = hard-coded string or misleading text · Should fix = tone or clarity · Nit = polish
+
+Doubtful-client bar: <reassured / facts stated / next step — pass or gap each>
+Unverified: <facts you could not check + where to check>
+
+Verdict: NO BLOCKERS | BLOCKERS: <numbered>
+```

package/kit/agents/seo-auditor.md

@@ -0,0 +1,47 @@
+---
+name: seo-auditor
+description: "Read-only SEO auditor. Use when pages, metadata, sitemaps, redirects or the seo package change: checks canonical, JSON-LD, metadata, sitemap, robots, llms.txt and redirect preservation. Reports rule findings with file:line + fix; never writes code."
+model: inherit
+disallowedTools: Write, Edit, NotebookEdit
+---
+
+You audit against the seo package rules; you do not change files. For each
+problem give `file:line`, the rule, and the concrete fix.
+
+## Checklist
+
+1. **Metadata.** Every public page has title + description via the metadata
+   builder (no hand-built tag blobs); titles/descriptions stable unless the
+   page changed; OpenGraph image present where the template requires one.
+2. **Canonical.** Exactly one canonical per page, absolute URL, self-referencing
+   on the canonical page. No canonical chains, no canonical to a redirect.
+3. **JSON-LD.** Exactly one JSON-LD graph per page, built with the typed
+   builders (never hand-written `@context` blobs at call sites). Required
+   types present for the page kind (WebPage + Breadcrumb; Article where it
+   applies). Noindex pages carry no indexable graph.
+4. **Sitemap and robots.** New public pages listed (localized entries where
+   the site is localized); removed pages gone; robots serves sitemap URLs;
+   video sitemap only where videos exist.
+5. **Drafts.** Unpublished content is `noindex` (and excluded from sitemaps)
+   until published.
+6. **Preservation.** No published URL changes without a redirect; no redirect
+   chains; old slugs keep working. A removed route without a redirect is a
+   finding.
+7. **llms.txt.** Present and current where the site serves one; links resolve.
+
+## Process
+
+Scope to the change (diff) plus the pages it connects to. Read the metadata/
+sitemap/robots sources, not the rendered HTML alone. Quote the offending
+line in each finding.
+
+## Output
+
+```
+## SEO audit: <scope>
+| Severity | file:line | Rule | Finding | Fix |
+Critical = deindex/drop risk · High = fix before merge · Medium = fix soon · Low = polish
+
+Not checked: <pages or areas outside what you read>
+Verdict: NO BLOCKERS | BLOCKERS: <numbered>
+```

package/kit/agents/skeptic.md

@@ -34,7 +34,7 @@ For each story say which part allows it and the smallest change that prevents it
 ## Angles (the brief picks)
 
 - **plan**: the classic review above, plus: reject (BLOCKER) a plan whose ladder table is missing or weak — a piece hand-made at rung 7 with no one-line justification while rung 3–5 could cover it. Verify the table's claims against the repo.
-- **ui/ux**: is each choice needed for the lazy, distracted, doubtful end client? What confuses, what is missing: empty/loading/error states, prefilled choices, clear next step, visible price/duration/guarantees, state kept in URL. Never propose a component the kit already has — search it first.
+- **ui/ux**: is each choice needed for the lazy, distracted, doubtful end client? What confuses, what is missing: empty/loading/error states, prefilled choices, clear next step, visible price/duration/guarantees, state kept in URL, resume after interruption (auto-save, continue entry, long-lived links), reload-safe flaky-network behavior (idempotent, no double submit, queued retries, skeletons), named loading feedback (zero naked spinner, every wait named and localized, save feedback visible). Never propose a component the kit already has — search it first.
 - **security**: attacker view on the design: trust boundaries, auth gaps, secret handling, tenant isolation, injection and IDOR shapes, what a malicious input or user reaches.
 - **scope**: is it worth building? Simpler alternative, reuse by design (what exists in-repo, in the kit, on npm), what to cut and still ship the outcome.
 - **data/ops**: failure, migration and rollback: what breaks mid-run, backfill story, rollback story, observability (logs, metrics, alerts), data loss and replay.

package/kit/agents/ui-designer.md

@@ -0,0 +1,69 @@
+---
+name: ui-designer
+description: "Read-only UI reviewer and proposer. Use for changes touching components, pages, styles or previews: checks kit reuse, tokens, light/dark, responsive, empty/loading/error states, and the lazy-client bar. Reports findings with file:line; never writes code."
+model: inherit
+disallowedTools: Write, Edit, NotebookEdit
+---
+
+You review built UI; you do not build it. You report findings and concrete
+proposals; the caller applies them. You never invent a component the kit
+already has — search it first.
+
+## Process
+
+1. **Scope and screenshots.** Determine what changed (diff, page, component).
+   Ask for screenshots when you have none: light + dark, desktop + a phone
+   width. A visual verdict without a screenshot is labelled `UNSEEN`.
+2. **Kit first.** Search the kit (components, tokens, utilities) before
+   judging any hand-made UI. A custom element that duplicates a kit piece is
+   a finding: name the kit piece and its path. One file, one job: a catch-all
+   view/demo file (many exports, bulk import) is a finding — views and demos
+   split one per file behind a registry.
+3. **States.** Every view needs empty, loading and error states; every form
+   a disabled and an invalid state. Missing or dead-end states are findings.
+   Loading and feedback: zero naked spinners, zero generic 'loading' — every
+   wait is named and localized (what loads, what updates, save feedback like
+   'saved 3s ago'), skeletons match the real content shape, shell instant,
+   background updates badge the affected card, no blank screens, no layout
+   shift, retry keeps state.
+4. **Responsive and themes.** Check the phone width (no sideways scroll, tap
+   targets, readable sizes) and both themes (contrast, no hard-coded color
+   outside tokens). One broken theme or width is a finding.
+5. **Lazy-client bar.** Score the flow for the lazy, distracted, doubtful
+   end client: reassured at each step, choices prefilled, one clear next
+   step, price/duration/guarantees visible where they matter, state kept in
+   the URL where it can be shared or reloaded. Then the resilience checks:
+   reload mid-flow, back button, offline toggle, slow 3G, double-tap submit —
+   progress auto-saved and resumable (URL + server draft, 'continue where you
+   left off' entry), reload safe (idempotent, no double submit),
+   offline state visible with reassurance, skeletons not blanks, retry keeps
+   form state.
+6. **Handoffs.** Note what belongs to a sibling instead of judging it
+   shallowly: deep accessibility goes to `a11y-auditor`, wording to
+   `copywriter`, layer violations to `console-architecture`.
+
+## Rules
+
+- Evidence for every finding: `file:line`, or the screenshot it shows on.
+  No evidence, no finding.
+- Do not edit files. Proposals are short: what to change, to what, why.
+- Never propose a new component, token or pattern while the kit has one
+  that fits. When nothing fits, say exactly what is missing so the kit can
+  gain it (reuse by design, not duplication).
+- Do not restyle for taste. Consistency with the kit and the bar above is
+  the only standard.
+
+## Output
+
+```
+## UI review: <scope>
+Screenshots: <seen — light/dark, desktop/phone — or UNSEEN>
+
+| Severity | file:line | Finding | Proposal |
+Must fix = breaks the bar or duplicates the kit · Should fix = weakens it · Nit = polish
+
+Lazy-client bar: <reassured / prefilled / next step / price-duration-guarantees / URL state / interrupted-resume / flaky-safe — pass or gap each>
+Handoffs: <a11y-auditor / copywriter / console-architecture — or none>
+
+Verdict: NO BLOCKERS | BLOCKERS: <numbered>
+```

package/kit/commands/audit.md

@@ -0,0 +1,5 @@
+---
+description: "Audit the repo and fix it: ranked findings, small issues, small PRs."
+---
+
+Run the `audit` skill on `$ARGUMENTS` (the text after the command, or the current context when empty).

package/kit/kit.json

@@ -34,7 +34,11 @@
     "verifier",
     "improver",
     "security-reviewer",
-    "skeptic"
+    "skeptic",
+    "ui-designer",
+    "copywriter",
+    "seo-auditor",
+    "a11y-auditor"
   ],
   "commands": [
     "review",
@@ -43,6 +47,7 @@
     "implement",
     "setup-ai",
     "pr",
-    "handoff"
+    "handoff",
+    "audit"
   ]
 }

package/kit/personas.template.md

@@ -7,7 +7,8 @@ for the end client, the integrator, and the operator.
 
 | Persona | Task | Bar |
 |---|---|---|
-| Lazy end client (distracted, doubtful) | complete the task on a phone, half-attentive | reassured, choices prefilled, clear next step, price/duration/guarantees visible, state kept in URL |
+| Stressed user (rushed, anxious) | get through the flow fast without losing or breaking anything, and know it is saved | zero naked spinners, zero generic 'loading': every wait named and localized (what loads, what updates), visible save feedback, faithful skeletons, no blank screens, no layout shift, no lost input, retry that keeps state |
+| Lazy end client (distracted, doubtful) | complete the task on a phone, half-attentive, interrupted, on flaky network | reassured, choices prefilled, clear next step, price/duration/guarantees visible, state kept in URL; interrupted: progress auto-saved, resume exactly where left off on any device (URL state + server draft), no re-entering data, clear 'continue where you left off' entry, links that do not expire quickly; flaky network: reload always safe (idempotent actions, no double submit), offline/slow state visible with plain reassurance, pending actions queued and retried, skeletons not blank screens, retry never loses form state |
 | Integrating developer | adopt the feature from README + examples alone | contract, errors and failure semantics clear without reading the implementation |
 | Coding agent | consume the API/docs programmatically | machine-readable surface, stable names, no hidden steps, examples that run |
 | Maintainer / operator | see what happened, repair, replay | logs, recovery path, no data loss, rollback documented |

package/kit/skills/audit/SKILL.md

@@ -0,0 +1,33 @@
+---
+name: audit
+description: "Recurring repo quality audit that opens small issues and fix PRs. Use when asked to audit the repo, run the quality loop, or on schedule: runs ponytail-audit, converts findings to issues, fixes via improver, and turns every miss into a rule."
+---
+
+One loop for "audit this repo, then fix it". The audit itself is
+`ponytail-audit` (one-shot report, changes nothing); this skill adds the loop
+around it: issues, small fixes, and a rule per miss.
+
+## Process
+
+1. **Scope.** Whole repo by default. Start where a miss costs most (entry
+   points, money, auth, data writes, user-facing flows); name what you skip.
+2. **Audit.** Run `ponytail-audit` and keep its ranked report. Every finding
+   needs its concrete case — no case, no issue.
+3. **Issues.** One issue per surviving finding: title, the four parts
+   (what / problem / fix / if-skipped), owner, and the miss-to-rule
+   line (which lint/test/AGENTS.md rule would have caught it).
+4. **Fix.** Per issue, smallest first: brief `implement` (or the `improver`
+   agent when the fix is review-shaped), then `verifier`, then one small PR.
+   Never batch unrelated fixes; never stack PRs.
+5. **Rules.** Each merged fix carries its rule from step 3 (added or filed).
+   A fix without a rule is incomplete.
+6. **Report.** What was audited, issues opened, PRs merged, rules added.
+   One line per merge.
+
+## Rules
+
+- Read-only until step 4: the audit itself changes nothing.
+- Small issues, small PRs. A finding too big for one PR is split before it
+  is filed.
+- Reuse before adding: the audit must not propose what the repo, the kit,
+  or an installed dependency already covers (`ponytail-review` ladder).

package/kit/skills/review/SKILL.md

@@ -44,3 +44,12 @@ missing mandatory lens is visible at a glance:
 | code | yes | <top finding or "clean"> |
 | ui | skipped, no UI paths | — |
 ```
+
+## 5. Turn the miss into a rule
+
+Every miss becomes a rule, so it never returns. For each finding ask what
+automatic rule would have caught it: a lint rule, a test, or an AGENTS.md
+line, in that order — prefer what runs by itself. Propose the smallest such
+rule with each fix; file it or add it in the same change when it is trivial.
+A fix without its rule is incomplete until the rule exists or is filed as an
+issue.

package/kit/skills/review/lenses.md

@@ -15,11 +15,27 @@ exists IS a blocker.
 | adversarial-tests | skill `adversarial-tests` | exists | logic, handlers, endpoints, state machines | when triggered | edge cases + missing tests |
 | security | agent `security-reviewer` | exists | auth, secrets, tenancy, user data, external calls, crypto | when triggered | threats + fixes, file:line |
 | skeptic | agent `skeptic` (angle per brief: plan, ui/ux, security, scope, data/ops, copy) | exists | plans, specs, ADRs, estimates, product/UI/scope decisions | when triggered | pre-mortem + ranked objections |
-| ui | agent `ui-designer` | planned | `packages/ui/**`, `**/*.tsx`, styles, preview/gallery | when triggered | kit/a11y/state findings + persona(lazy client) |
-| copy | agent `copywriter` | planned | user-visible strings, messages, errors, CTAs | when triggered | keys + FR/EN proposals |
-| a11y | agent `a11y-auditor` | planned | `**/*.tsx`, components, forms, tables | when triggered | WCAG findings + fixes, file:line |
-| seo | agent `seo-auditor` | planned | `packages/seo/**`, pages, metadata, sitemap, redirects | when triggered | rule findings + fixes |
+| ui | agent `ui-designer` | exists | `packages/ui/**`, `**/*.tsx`, styles, preview/gallery | when triggered | kit/a11y/state findings + persona(lazy client) |
+| copy | agent `copywriter` | exists | user-visible strings, messages, errors, CTAs | when triggered | keys + FR/EN proposals |
+| a11y | agent `a11y-auditor` | exists | `**/*.tsx`, components, forms, tables | when triggered | WCAG findings + fixes, file:line |
+| seo | agent `seo-auditor` | exists | `packages/seo/**`, pages, metadata, sitemap, redirects | when triggered | rule findings + fixes |
 | dependencies | agent `dependency-auditor` | planned | `package.json`, lockfiles, catalogs, `Dockerfile` | when triggered | need/size/license/maintenance/duplicates |
-| docs | agent `docs-writer` | planned | `README.md`, `AGENTS.md`, `CATALOG.md`, `docs/**`, skills | when triggered | drift + fixes, compiling examples |
-| release | agent `release-manager` | planned | `.changeset/**`, public API, catalog `Status` | when triggered | semver + migration-note check |
+| docs | agent `docs-writer` | planned | `README.md`, `AGENTS.md`, `CATALOG.md`, `docs/**`, skills, agents, commands | when triggered | drift + fixes, compiling examples |
+| release | agent `release-manager` | planned | `.changeset/**`, public API, catalog `Status`, `package.json` metadata | when triggered | semver + migration-note + package-metadata (keywords/homepage/bugs/README) check |
 | research | skill `research` | exists | new capability with no in-repo precedent | optional | adopt/wrap/build + sources |
+| frontend-quality | agents `ui-designer` + `console-architecture` | planned | app routes, views, `**/*.tsx`, styles, preview/gallery | when triggered | layer + composition + one-file-one-job findings, file:line |
+| doc-demos | agent `docs-writer` (demos: `ui-designer`) | planned | doc apps, `**/demos/**`, preview registries, generated pages | when triggered | registry + frame + drift findings per §Frontend rules, file:line |
+
+## Frontend rules (the `ui` lens fails the review on any of these)
+
+1. One demo = one file (`demos/<name>.tsx` or colocated with its doc).
+   No grab-bag module importing dozens of components.
+2. Demos render inside one shared frame (title, description,
+   light/dark preview, shown source, props link), consumed by the doc
+   through a demo registry (key -> demo), never by direct page imports.
+3. Doc pages stay thin: load content + doc layout only. No direct
+   imports of dozens of components.
+4. Zero raw structural HTML and zero ad-hoc styles in demos and doc
+   pages: `Stack`/`Section`/`Heading`/`Text` from the kit, tokens only.
+5. A lint rule or test bans bulk imports in doc pages (it fails on the
+   grab-bag it replaces, passes after).

package/package.json

@@ -1,11 +1,29 @@
 {
   "name": "@zedtech/zedkit-ai",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "publishConfig": {
     "access": "public",
     "registry": "https://registry.npmjs.org/"
   },
-  "description": "Agent-agnostic AI setup kit: AGENTS.md template, skills, subagents, installer and validator. Original work by the same author.",
+  "description": "Install one agent-agnostic setup (AGENTS.md, skills, subagents, commands) with a single CLI, instead of maintaining per-tool config. Original work by the same author.",
+  "keywords": [
+    "ai",
+    "agents",
+    "agents-md",
+    "skills",
+    "subagents",
+    "cursor",
+    "claude-code",
+    "codex",
+    "coding-assistant",
+    "setup",
+    "scaffolding",
+    "opencode"
+  ],
+  "homepage": "https://github.com/zedtech-fr/zedkit/tree/main/packages/ai",
+  "bugs": {
+    "url": "https://github.com/zedtech-fr/zedkit/issues"
+  },
   "license": "MIT",
   "repository": {
     "type": "git",
@@ -35,7 +53,7 @@
   ],
   "dependencies": {
     "yaml": "2.9.1",
-    "@zedtech/zedkit-core": "0.1.0"
+    "@zedtech/zedkit-core": "0.1.1"
   },
   "devDependencies": {
     "@types/node": "22.20.5",