diff --git a/skills/ask/SKILL.md b/skills/ask/SKILL.md
new file mode 100644
index 0000000..7dfb96e
--- /dev/null
+++ b/skills/ask/SKILL.md
@@ -0,0 +1,106 @@
+---
+name: ask
+description: Fetch version-accurate library documentation, source trees, and producer-shipped skills so the agent works against the exact version installed in the project, not training-data guesses. Use this skill whenever the user needs docs for a dependency, wants to read a library's real source, asks "how does X work internally", needs to pin reading to a specific version or ref, mentions ask docs / ask src / ask skills, or any task that would benefit from a library's actual README / source / skill files over recalled knowledge — even when they don't explicitly name the "ask" CLI. Preferred over inferring API shape from memory whenever accuracy matters.
+allowed-tools: Bash(ask:*)
+---
+
+# Version-Accurate Docs & Source with `ask`
+
+`ask` resolves the version from the project's lockfile
+(`bun.lock → package-lock.json → pnpm-lock.yaml → yarn.lock → package.json`
+range fallback), fetches docs or source once, caches them globally at
+`~/.ask/` (override via `ASK_HOME`), and prints absolute paths to stdout so
+the commands compose naturally in shell substitutions. Progress / errors go
+to stderr, paths go to stdout — safe for `$(ask …)`.
+
+## Core Pattern
+
+```bash
+# Docs — one candidate path per line
+cat "$(ask docs zod | head -n1)"/README.md
+rg "parseAsync" $(ask docs zod)
+
+# Source — single absolute path to the checkout root
+rg "ZodError" $(ask src zod)
+fd -e test.ts . $(ask src zod)
+
+# Producer-shipped skills — one /skills/ dir per line
+ls $(ask skills vercel/ai)
+```
+
+`ask docs` emits candidate documentation directories (publish-time
+`dist/docs` first, then any subdirectory whose basename matches `/doc/i`
+up to depth 4, falling back to the checkout root when nothing matches).
+`ask src` emits exactly one path: the checkout root. Both auto-fetch on
+cache miss; pass `--no-fetch` to fail fast (exit 1) on miss instead.
+
+## Spec Grammar
+
+```
+zod                         # bare → npm ecosystem (resolved via lockfile)
+npm:next                    # explicit ecosystem
+npm:@mastra/client-js       # scoped package
+facebook/react              # owner/repo → github:facebook/react@main
+github:vercel/next.js@v14.2.3   # pinned tag
+github:owner/repo@main          # pinned branch
+```
+
+- For `npm:` specs (and bare names), the version comes from the project's
+  lockfile. Append `@version` to pin explicitly: `zod@3.22.0`,
+  `npm:next@14.2.3`.
+- For `github:` specs, `@<ref>` pins a tag or branch. Bare `owner/repo`
+  (no `@ref`) defaults to `main`.
+- Any ref works with these one-shot reading commands — branches, tags,
+  or mutable refs like `main` / `master` are all accepted, since
+  nothing is persisted.
+
+## One-Shot Reading Commands
+
+| Command | Output | Use when |
+|---------|--------|----------|
+| `ask docs <spec> [--no-fetch]` | Candidate doc dirs, one per line | You want README / guides / handwritten docs at the installed version |
+| `ask src <spec> [--no-fetch]`  | Checkout root, single line        | You need to read real source, search all files, follow implementations |
+| `ask skills <spec>` (= `ask skills list`) | `/skills/` dirs, one per line | The library ships its own Claude / Cursor / OpenCode skills |
+
+All three share `ensureCheckout`, so the cached path is reused across
+commands — calling `ask docs`, then `ask src`, then `ask skills list` on
+the same spec fetches once.
+
+## When You Need More
+
+Lazy-load these references only when the situation calls for them:
+
+- **Managing the cache** — disk pressure, stale entries, `--kind` /
+  `--older-than` filters, legacy v1 layout cleanup →
+  [`references/cache.md`](references/cache.md).
+- **Project-level declarative workflow** — `ask.json`, `ask install`,
+  `ask add`, `ask remove`, `ask list`, auto-regenerated `AGENTS.md` and
+  per-library `.claude/skills/<name>-docs/SKILL.md` →
+  [`references/declarative-workflow.md`](references/declarative-workflow.md).
+- **Vendoring producer skills into this project** — `ask skills install`,
+  `--force`, `--agent claude,cursor,opencode,codex`, `ask skills remove
+  --ignore-missing` →
+  [`references/skills-vendoring.md`](references/skills-vendoring.md).
+
+## When to Reach for `ask`
+
+Reach for it when:
+
+- The installed version matters — otherwise the agent risks fabricating
+  API shape from an outdated training snapshot.
+- The answer lives in source, not types — edge cases, error paths,
+  internal helpers, behavior that isn't documented anywhere else.
+- A library may ship its own skills — `ask skills list <spec>` discovers
+  producer-side `skills/` directories without touching the project.
+
+Skip it when TypeScript / LSP / intellisense can answer the question, or
+when the user has already pointed at a specific file path.
+
+## Why This Exists
+
+Training data ages; lockfiles don't. `ask` bridges the two by pinning
+every read to the version the project actually runs, so generated code
+reflects reality instead of last year's docs. The `$(ask …)` idiom is
+the main ergonomic: it turns a cached path into a first-class argument
+to `rg`, `cat`, `fd`, or any tool that accepts a path — no extra API to
+learn.
diff --git a/skills/ask/references/cache.md b/skills/ask/references/cache.md
new file mode 100644
index 0000000..cf82e81
--- /dev/null
+++ b/skills/ask/references/cache.md
@@ -0,0 +1,93 @@
+# Cache Management (`ask cache`)
+
+Read this when disk is filling up, a cached entry is stale, or you need
+to inspect what `ask` has fetched. The global store lives at `~/.ask/`
+by default; set `ASK_HOME` to relocate it. Every `ask install` writes
+`<askHome>/STORE_VERSION` (currently `"2"`) as a layout marker.
+
+## Store Layout (v2)
+
+```
+<askHome>/
+├── npm/<pkg>@<version>/              # npm tarball extractions
+├── github/<host>/<owner>/<repo>/<ref>/   # github shallow clones (PM-unified)
+├── web/<sha256-of-url>/              # crawled doc sites
+├── llms-txt/<sha256>@<version>/      # llms.txt imports
+├── STORE_VERSION                     # "2"
+└── .quarantine/<ts>-<uuid>/          # corrupt entries, preserved for inspection
+```
+
+`host` is currently always `github.com` — the segment exists so
+`gitlab.com` / `bitbucket.org` can slot in without a migration.
+
+## `ask cache ls [--kind <kind>]`
+
+List every entry in the store with its size.
+
+```bash
+ask cache ls                      # everything
+ask cache ls --kind npm           # filter: npm | github | web | llms-txt
+```
+
+Entries print as `<kind>/<key>  <size>`. Total size is summarized at
+the end. If a legacy (pre-v2) `github/db` or `github/checkouts`
+directory exists, its entries are tagged with a `(legacy) ` prefix in
+the key — that's the signal to run `ask cache clean --legacy`.
+
+## `ask cache gc [--dry-run] [--older-than <duration>]`
+
+Remove entries that no project references anymore. `gc` discovers
+references by walking `$HOME` (or each path in `ASK_GC_SCAN_ROOTS`,
+colon-separated) for `.ask/resolved.json` files and treating the
+entries listed there as roots.
+
+```bash
+ask cache gc                          # remove all unreferenced
+ask cache gc --dry-run                # preview only, no deletion
+ask cache gc --older-than 30d         # also require age > 30 days
+ask cache gc --older-than 12h         # supports d / h / m / s
+ASK_GC_SCAN_ROOTS=/repos:/work ask cache gc   # restrict scan
+```
+
+Always try `--dry-run` first when `ASK_GC_SCAN_ROOTS` is unset — the
+default `$HOME` scan may miss project directories stored elsewhere,
+which would mark still-used entries as unreferenced.
+
+## `ask cache clean --legacy`
+
+Remove the pre-v2 github store layout. The flag is required; calling
+`ask cache clean` without it exits with a reminder.
+
+```bash
+ask cache clean --legacy
+```
+
+Deletes:
+- `<askHome>/github/db/`        (bare-clone DB from the old shared layout)
+- `<askHome>/github/checkouts/` (per-ref worktrees from the old layout)
+
+The v2 layout (`<askHome>/github/<host>/<owner>/<repo>/<ref>/`) is
+untouched. Safe to run anytime; it's a no-op when no legacy dirs exist.
+
+## Environment Variables
+
+| Variable              | Purpose                                    | Default        |
+|-----------------------|--------------------------------------------|----------------|
+| `ASK_HOME`            | Override the global store root             | `~/.ask`       |
+| `ASK_GC_SCAN_ROOTS`   | Colon-separated scan roots for `cache gc`  | `$HOME`        |
+
+Both are read at invocation time — set them inline or export per shell.
+
+## Troubleshooting Tips
+
+- "Corrupted store entry … quarantined to …" — `ask install` moved a
+  tamper/missing-stamp entry into `.quarantine/`. Re-run the install;
+  the fresh fetch replaces it. Delete the quarantine dir manually once
+  you've inspected it.
+- `ask docs <spec>` prints nothing — the checkout exists but has no
+  `/doc/i` subdirs AND the walker can't see the root. Try
+  `ask src <spec>` and `ls $(ask src <spec>)` to confirm the checkout
+  is populated.
+- `gc` removed an entry you still needed — either the project wasn't
+  inside `ASK_GC_SCAN_ROOTS`, or `.ask/resolved.json` was missing /
+  out of date. Re-run `ask install` in the project to rebuild both.
diff --git a/skills/ask/references/declarative-workflow.md b/skills/ask/references/declarative-workflow.md
new file mode 100644
index 0000000..bcd0ed8
--- /dev/null
+++ b/skills/ask/references/declarative-workflow.md
@@ -0,0 +1,137 @@
+# Declarative Workflow (`ask.json` + `ask install`)
+
+Read this when the user wants a checked-in list of libraries the
+project's agents should reference, or wants `AGENTS.md` auto-regenerated
+with lazy `ask src` / `ask docs` pointers. The one-shot reading
+commands (`ask docs` / `ask src` / `ask skills list`) don't persist
+anything; this flow does.
+
+`ask install` is **lazy-first**: it resolves versions and writes
+`AGENTS.md` plus a per-library `.claude/skills/<name>-docs/SKILL.md`
+that tells agents to invoke `ask src <spec>` / `ask docs <spec>` on
+demand. No documentation is downloaded during `install` — fetching
+happens the first time an agent calls `ask docs` / `ask src`.
+
+## What Gets Written
+
+- `ask.json` — the declarative input. Array of spec strings. Checked
+  into git; edited by `ask add` / `ask remove` (or by hand).
+- `AGENTS.md` — block between `<!-- BEGIN:ask-docs-auto-generated -->`
+  and `<!-- END:ask-docs-auto-generated -->` maintained by `ask`.
+- `.claude/skills/<name>-docs/SKILL.md` — one skill per declared
+  library, each delegating to `ask docs <spec>` / `ask src <spec>`.
+- Ignore files (`.gitignore`, `.prettierignore`, etc.) — patched via
+  `# ask:start … # ask:end` marker blocks. Don't hand-edit inside
+  those markers; `install` / `remove` overwrite them.
+
+## `ask install`
+
+Reads `ask.json`, resolves each spec to a concrete version, regenerates
+per-library skills, and rewrites the `AGENTS.md` block.
+
+```bash
+ask install
+```
+
+Per-entry behavior:
+
+1. **Version resolution**
+   - `npm:` specs → read lockfile chain
+     `bun.lock → package-lock.json → pnpm-lock.yaml → yarn.lock →
+     package.json` (range fallback). Missing from every lockfile ⇒
+     entry skipped with a warning.
+   - `github:` specs → the ref encoded in the spec
+     (`github:vercel/next.js@v14.2.3`). No lockfile lookup.
+   - Explicit `@version` on any spec wins over lockfile resolution.
+2. **Skill generation** — writes `.claude/skills/<name>-docs/SKILL.md`.
+3. **AGENTS.md regeneration** — rewrites the auto-generated block
+   with every resolved library.
+
+`postinstall`-friendly: per-entry failures emit a warning; the overall
+exit code is always 0, so an unresolvable entry doesn't break
+`bun install`.
+
+## `ask add <spec>` / `ask add` (interactive)
+
+Appends to `ask.json`, then runs install for just that entry.
+
+```bash
+ask add npm:next
+ask add npm:@mastra/client-js
+ask add github:vercel/next.js@v14.2.3
+ask add facebook/react           # bare owner/repo → github:facebook/react
+ask add                          # interactive picker
+```
+
+Bare names without `:` or `/` (e.g. `ask add zod`) are rejected with a
+hint listing the two valid forms. Note the asymmetry: the one-shot
+`ask docs zod` DOES accept bare names — only `ask add` is strict
+because the spec is persisted.
+
+## `ask remove <name>`
+
+Removes from `ask.json`, deletes the generated skill, re-runs install
+to regenerate `AGENTS.md`.
+
+```bash
+ask remove next
+ask remove @mastra/client-js
+ask remove npm:next              # also accepts the full spec
+```
+
+The `name` match is tried against the full spec, the spec body, and
+the derived library slug — so the unscoped name, the full spec, or the
+scoped package name all work.
+
+## `ask list [--json]`
+
+Prints declared libraries with their resolved versions. Unresolved
+entries (declared in `ask.json` but missing from every lockfile) show
+`version: unresolved` so drift is visible.
+
+```bash
+ask list
+ask list --json | jq '.entries[] | select(.version == "unresolved")'
+```
+
+JSON shape (`ListModelSchema`): `{ "entries": [...], "conflicts":
+[...], "warnings": [...] }`. Each entry has `name`, `version`,
+`format`, `source`, `location`, plus optional `itemCount` / `skills`.
+
+## `ask.json` Shape
+
+Strict array of ecosystem-prefixed spec strings — schema is
+`z.array(z.string().regex(/^[a-z][a-z0-9+-]*:.+$/))` with `.strict()`.
+Object entries are rejected.
+
+```json
+{
+  "libraries": [
+    "npm:next",
+    "npm:@mastra/client-js",
+    "npm:zod@3.22.0",
+    "github:vercel/next.js@v14.2.3",
+    "github:vercel/ai@v5.0.0"
+  ]
+}
+```
+
+- `npm:` specs: append `@<version>` to pin, otherwise the lockfile
+  decides.
+- `github:` specs: append `@<ref>` to encode the version inside the
+  spec string. `github:` without a ref defaults to `latest` in the
+  generated output.
+- No separate `ref`, `source`, or `docsPath` keys. Per-library
+  `docsPath` comes from the ASK Registry when `ask docs` / `ask src`
+  actually fetch, not from `ask.json`.
+
+## Intent-Format and Resolved-Cache Notes
+
+Prior versions of `ask` had a `.ask/resolved.json` lockfile, a
+`<!-- intent-skills:start -->` block in `AGENTS.md` for
+TanStack-intent packages, and a `--allow-mutable-ref` flag. In the
+current lazy-first architecture `ask install` does not write
+`.ask/resolved.json` and does not populate intent blocks; the
+underlying code modules still exist but are not wired into the
+install path. Treat the docs, source, and skills commands as the
+supported surface, and `ask.json` as a strict spec-string array.
diff --git a/skills/ask/references/skills-vendoring.md b/skills/ask/references/skills-vendoring.md
new file mode 100644
index 0000000..502d286
--- /dev/null
+++ b/skills/ask/references/skills-vendoring.md
@@ -0,0 +1,109 @@
+# Vendoring Producer Skills (`ask skills install` / `remove`)
+
+Read this when the user wants a library's own skills (the ones the
+library *publishes* for consumers) copied into this project and linked
+into the agent's skills directory. The one-shot `ask skills list
+<spec>` only prints paths; the commands here actually vendor and link.
+
+## The Core Flow
+
+`ask skills install <spec>` does five things:
+
+1. Resolves and fetches the library the same way `ask src` does.
+2. Collects each direct child of every `skills/` (or `skill/`) parent
+   directory shipped by the library — matched case-insensitively
+   against `^skills?$` exactly, so `my-skills/` / `skill-set/` are
+   ignored.
+3. Vendors those children into `.ask/skills/<specKey>/<skill-name>/`
+   in the project. `specKey` is derived from ecosystem + name +
+   resolved version (e.g. `npm__next__14.2.3`).
+4. Symlinks each vendored skill into the detected agent skills
+   directories (`.claude/skills/`, `.cursor/skills/`,
+   `.opencode/skills/`, `.codex/skills/`).
+5. Records the install in `.ask/skills-lock.json` and patches
+   `.gitignore` via the `ask` marker block.
+
+## `ask skills install <spec>`
+
+```bash
+ask skills install vercel/ai
+ask skills install npm:@mastra/core
+ask skills install github:owner/repo@v1.2.3 --no-fetch      # cache hits only
+ask skills install vercel/ai --force                        # overwrite conflicts
+ask skills install vercel/ai --agent claude,cursor          # explicit targets
+```
+
+| Flag              | Effect                                                           |
+|-------------------|------------------------------------------------------------------|
+| `--no-fetch`      | Return cache hit only — exit 1 on miss.                          |
+| `--force`         | Overwrite a conflicting `.claude/skills/<name>` symlink.         |
+| `--agent <csv>`   | Explicit targets: `claude,cursor,opencode,codex`. Skips detection. |
+
+### Agent selection
+
+Without `--agent`:
+
+- 0 agents detected → error asking the user to pass `--agent <name>`.
+- 1 agent detected → auto-use it.
+- 2+ detected → interactive multiselect prompt.
+
+Detection looks for the agent's directory (`.claude/`, `.cursor/`,
+`.opencode/`, `.codex/`). If none exist, pass `--agent` to create the
+target directory and force-install.
+
+### Exit conditions
+
+- No `skills/` (or `skill/`) dir found in the source → exit 1 with the
+  message `no skills/ directories found for <spec>`.
+- No agent selected → exit 1.
+- Symlink collision without `--force` → exit from `linkSkill` with a
+  descriptive error.
+
+## `ask skills remove <spec>`
+
+Reverses a prior install using `.ask/skills-lock.json` as the source
+of truth. Only removes symlinks that point at the vendored copy — a
+user-edited skill in `.claude/skills/<name>` that no longer points at
+`.ask/skills/<specKey>/<name>/` is left alone.
+
+```bash
+ask skills remove vercel/ai
+ask skills remove npm__vercel-ai__5.0.0       # spec-key also accepted
+ask skills remove vercel/ai --ignore-missing  # exit 0 even if no lock entry
+```
+
+| Flag                | Effect                                                |
+|---------------------|-------------------------------------------------------|
+| `--ignore-missing`  | Exit 0 when the spec has no lock entry.               |
+
+On success, the vendor dir `.ask/skills/<specKey>/` is deleted and the
+lock entry removed.
+
+## Directory Layout After Install
+
+```
+.ask/
+├── skills/
+│   └── <specKey>/
+│       ├── <skill-name-1>/
+│       │   └── SKILL.md
+│       └── <skill-name-2>/
+│           └── …
+└── skills-lock.json
+
+.claude/skills/
+├── <skill-name-1> -> ../../.ask/skills/<specKey>/<skill-name-1>
+└── <skill-name-2> -> ../../.ask/skills/<specKey>/<skill-name-2>
+```
+
+Each skill is symlinked individually — no wrapper directory — so the
+agent discovers them like any other project skill.
+
+## Docs vs. Intent-Format Distinction
+
+This command only handles the "skills" channel. Libraries that
+distribute docs (not skills) go through `ask install` instead. If the
+library is a TanStack-intent package (keywords include
+`tanstack-intent`), `ask install` takes the intent path and writes an
+`<!-- intent-skills:start -->` block in `AGENTS.md`, independent of
+the `ask skills install` flow described here.