diff --git a/plugins/heph-expert/skills/heph/references/concepts.md b/plugins/heph-expert/skills/heph/references/concepts.md
index a4599ac..021c655 100644
--- a/plugins/heph-expert/skills/heph/references/concepts.md
+++ b/plugins/heph-expert/skills/heph/references/concepts.md
@@ -8,6 +8,7 @@ The mental model behind heph. Source pages under
 - [Dependencies](#dependencies)
 - [Sandbox](#sandbox)
 - [Caching](#caching)
+- [Scratch caches](#scratch-caches)
 - [Reproducibility](#reproducibility)
 - [Codegen](#codegen)
 - [Runners](#runners)
@@ -119,6 +120,59 @@ are trimmed automatically at the end of the run that writes the new one — no
 action needed. `heph tool gc` sweeps everything else no longer reachable from
 any current target (removed targets, orphaned entries).
 
+## Scratch caches
+
+A **scratch** is a directory a target declares, keeps between runs, and
+shares with every target that references it — the one thing in a sandbox
+that is neither an input nor an output. It's for a tool that already
+maintains its own content-addressed cache (a compiler cache, a package
+download cache, a registry blob store), never for durable state.
+
+The contract: **a target's outputs must be identical whether its scratch is
+warm, cold, or absent.** It never enters `hashin`, so it can never invalidate
+anything — changing its settings rebuilds nothing, deleting it costs only
+time. Prove the contract with `heph --no-scratch inspect hashout <addr>` and
+compare against the ordinary hash.
+
+```python title="BUILD"
+target(
+    name    = "gocache",
+    driver  = "scratch",
+    path    = ".cache/go-build",   # optional; omit for env-var-only
+    env     = "GOCACHE",           # defaults to SCRATCH_<NAME>
+    access  = "shared",            # "exclusive" (default) | "shared"
+    version = "",                  # what the contents depend on, beyond the addr
+    remote  = False,               # may travel through the remote cache
+)
+
+target(name = "build", driver = "bash", scratch = ["//build:gocache"], ...)
+```
+
+| Field | Meaning |
+|---|---|
+| `path` | Mount point inside a consumer's sandbox, relative to its cwd. Optional — omitting it means nothing is placed in the tree, so no output can collect it and no dependency can be shadowed by it. |
+| `env` | Env var a consumer reads the absolute path from. Point it at the tool's own variable (`GOCACHE`, `CCACHE_DIR`). |
+| `access` | `"exclusive"` (default, one consumer at a time, cross-process) or `"shared"` (concurrent — only for a cache safe under concurrent access *by construction*). |
+| `version` | The whole identity beyond the address. heph never guesses; state what the cache depends on (`heph.core.os() + "/" + heph.core.arch()`, a toolchain version, …) or leave empty for a portable cache. |
+| `remote` | May be pulled/pushed through a configured remote cache. |
+| `max_size` | Size cap, e.g. `"10GiB"`; over it the cache is dropped whole rather than trimmed. |
+
+`heph tool scratch ls / head <addr> / path <addr> / rm <addr>|--all / push
+--all / pull --all` inspect, reclaim, and publish caches. `head` explains
+which lineage a build would restore from and why — the go-to when a branch
+starts unexpectedly cold. `.hephconfig`'s `scratch.scope` /
+`scratch.restoreScopes` control per-branch lineages (`${git:branch}` resolves
+automatically); `heph tool gc --scratch-max-size` / `--scratch-max-age-days`
+sweeps by size and age.
+
+A mounted scratch (`path` set) must not sit under a broad output glob in the
+same package — `out = "**/*"` beside a mount fails at pack time, naming the
+mount. The env-var-only form (no `path`) has no such hazard.
+
+Full reference: <https://hephbuild.github.io/docs/concepts/scratch>. The
+`heph-go` plugin covers the Go provider's own `GOCACHE`/`GOMODCACHE` scratch
+caches and `heph.go.gocache_addr()`.
+
 ## Reproducibility
 
 The central promise is **byte-identical outputs**: same inputs ⇒ same artifact,
diff --git a/plugins/heph-expert/skills/heph/references/configuration.md b/plugins/heph-expert/skills/heph/references/configuration.md
index 26a1e10..cf90955 100644
--- a/plugins/heph-expert/skills/heph/references/configuration.md
+++ b/plugins/heph-expert/skills/heph/references/configuration.md
@@ -20,6 +20,7 @@ plugins:
 | `plugins` | list of plugin entries | `[]` | Plugins to register. Each entry sets exactly one of `builtin`, `path`, or `url`, plus an optional `options` map. |
 | `homeDir` | path | unset | Where heph keeps its home and cache. |
 | `memCache` | `{perEntryBytes, capacityBytes}` | unset | In-memory cache sizing. |
+| `scratch` | `{scope, restoreScopes, seedOnFork}` | unset | Which scratch-cache lineage a run reads and writes. |
 | `fuse` | `{enabled: true \| false \| "auto"}` | off | Sandbox overlay mode. |
 | `lock` | `{backend: fs \| mem}` | `fs` | Execute-phase lock backend. |
 
@@ -66,6 +67,21 @@ memCache:
   capacityBytes: 67108864   # total budget; 0 disables the in-memory cache
 ```
 
+## `scratch` — branch-lineage policy
+
+Which lineage of a [scratch cache](./concepts.md#scratch-caches) a run reads
+and writes — the cache's *contents* are declared by its target; this is
+which copy of them a run sees.
+
+```yaml title=".hephconfig"
+scratch:
+  scope: "${git:branch}"       # lineage this run writes to; resolves the current branch
+  restoreScopes: ["master"]    # lineages to read from when scope has nothing yet
+  seedOnFork: true             # default: copy the first warm fallback into a cold scope
+```
+
+`restoreScopes` is read-only — a branch never writes into its fallback.
+
 ## `fuse` — sandbox overlay
 
 Opt-in: assemble sandboxes with a FUSE overlay instead of copying inputs.
diff --git a/plugins/heph-go/skills/heph-go/references/go-plugin.md b/plugins/heph-go/skills/heph-go/references/go-plugin.md
index 8bda8c3..79d058a 100644
--- a/plugins/heph-go/skills/heph-go/references/go-plugin.md
+++ b/plugins/heph-go/skills/heph-go/references/go-plugin.md
@@ -187,6 +187,34 @@ options:
   cctool: "//@heph/bin:cc"    # default; point elsewhere for a hermetic compiler
 ```
 
+## Build and module caches
+
+`GOCACHE` (one per Go module + build variant) and `GOMODCACHE` (one shared,
+portable cache for downloaded modules) are wired up automatically as
+[scratch caches](https://hephbuild.github.io/docs/concepts/scratch) —
+nothing to configure. A host-set `GOMODCACHE` is deliberately not passed
+through, so the shared cache is what every build sees. Both show up in
+`heph tool scratch ls`, are droppable with `heph tool scratch rm` if one
+goes bad, and are auditable with `heph run --no-scratch`.
+
+Share the same `GOCACHE` a module's own targets use from a hand-written
+target with `heph.go.gocache_addr()`:
+
+```python title="BUILD"
+target(
+    name    = "custom_build",
+    driver  = "bash",
+    scratch = [heph.go.gocache_addr()],   # goos, goarch, gotool, tags, goexperiment,
+    run     = "go build -o $OUT ./cmd/tool",  # gcflags, ldflags, race — match the variant
+    out     = "tool",
+)
+```
+
+`gocache_addr()` resolves the module from the *calling* BUILD file's nearest
+`go.mod`. Pass the same factors (`goos`/`goarch`/`gotool`/`tags`/…) the
+variant you're building against declares, or the target warms a different
+cache than the one it meant to share.
+
 ## Linting and formatting
 
 A Go module gets `lint-check`, `lint`, `format-check`, and `format` targets the
diff --git a/website/docs/concepts/scratch.md b/website/docs/concepts/scratch.md
new file mode 100644
index 0000000..b4576a4
--- /dev/null
+++ b/website/docs/concepts/scratch.md
@@ -0,0 +1,185 @@
+---
+title: "Scratch caches"
+sidebar_position: 7
+description: A named, mutable cache directory a target keeps between runs — for a compiler cache, a download cache, or any tool that maintains its own cache.
+---
+
+# Scratch caches
+
+A **scratch** is a directory a target declares, keeps between runs, and
+shares with every other target that references it. It is the one thing
+inside a [sandbox](/docs/concepts/sandbox) that is neither an input nor an
+output — mutable, and never hashed.
+
+## The contract
+
+> A target's outputs must be identical whether its scratch directories are
+> warm, cold, or absent. Losing one is always a slowdown, never a wrong
+> answer.
+
+heph gives every other part of a target's inputs and outputs the hermetic
+treatment: declared in, declared out, sandbox thrown away. A scratch is the
+deliberate exception, for a tool that already maintains its own
+content-addressed cache — a compiler cache, a package download cache, a
+registry blob store. Because it never enters the
+[input hash](/docs/concepts/caching), it can never invalidate anything:
+changing a scratch's settings rebuilds nothing, and deleting one costs time
+and nothing else.
+
+Check the contract on your own target:
+
+```bash title="terminal"
+heph inspect hashout //build:compile
+heph --no-scratch inspect hashout //build:compile
+```
+
+Same hash both times, or the target depends on carried-over state and is
+broken.
+
+## Declaring one
+
+A scratch is declared like any other target, with `driver = "scratch"`. It
+builds nothing — it only describes a cache.
+
+```python title="BUILD"
+target(
+    name    = "gocache",
+    driver  = "scratch",
+    path    = ".cache/go-build",   # optional; omit for env-var-only
+    env     = "GOCACHE",           # defaults to SCRATCH_<NAME>
+    access  = "shared",            # "exclusive" (default) | "shared"
+    version = "",                  # what the contents depend on, beyond the addr
+    remote  = False,               # may travel through the remote cache
+)
+
+target(
+    name    = "build",
+    driver  = "bash",
+    scratch = ["//build:gocache"],
+    run     = "go build -o $OUT ./...",
+    out     = "bin",
+)
+```
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `path` | `string` | unset | Where the directory mounts inside a consuming target's sandbox, relative to its cwd. Optional — most tools find their cache through an environment variable, and omitting `path` is the safer shape: with nothing mounted, no output can collect the directory and no dependency can be shadowed by it. |
+| `env` | `string` | `SCRATCH_<NAME>` | Environment variable a consumer reads the directory's absolute path from. Point it at the tool's own variable (`GOCACHE`, `CCACHE_DIR`) so nothing else needs wiring. |
+| `access` | `string` | `"exclusive"` | `"exclusive"` — one consumer at a time, enforced across separate `heph` processes too; the safe default for a tool with no stated concurrency story. `"shared"` — concurrent consumers allowed, only for a cache that's safe under concurrent access *by construction* (content-addressed and self-verifying, the way Go's build cache is). |
+| `version` | `string` | `""` | Everything the contents depend on, beyond the address — the whole of it. Two declarations at the same address share a directory if and only if they also agree on `version`. heph never guesses: state what the cache depends on, or leave it empty for a cache that's portable everywhere. See [Portability with `version`](#portability-with-version). |
+| `remote` | `bool` | `false` | Whether this cache may be pulled from a remote cache automatically, and published to it with `heph tool scratch push`. See [Sharing through the remote cache](#sharing-through-the-remote-cache). |
+| `max_size` | `string` | unset | A size cap, e.g. `"10GiB"`. Past it, the whole cache is dropped and starts again — heph can't tell which of a foreign tool's entries are hot, so it doesn't try to trim. |
+
+Reference a declaration from a consuming target with `scratch = [...]` — a
+plain list of addresses, the same shape as `deps`. A target may reference
+several; every target referencing the same declaration gets the same
+directory.
+
+## Choosing `access`
+
+`access` is an assertion about the tool, not a wish — get it wrong and
+`"shared"` corrupts the cache, or `"exclusive"` serializes work that didn't
+need to be. Go's build cache and module cache are safe under concurrent
+writers because they're content-addressed: an entry either matches its key
+or is not used, the same property `go build -p N` already relies on. Most
+tools don't document that, so `"exclusive"` — one consumer at a time,
+enforced with a lock held for the whole run — is the default.
+
+## Portability with `version`
+
+`version` is opaque to heph — it isn't parsed, and heph contributes nothing
+of its own to it (no host OS, no architecture) unless the BUILD file puts it
+there:
+
+```python title="BUILD"
+version = heph.core.os() + "/" + heph.core.arch()   # host-specific
+version = goos + "/" + goarch + "/" + go_version     # target-specific
+version = ""                                         # portable (the default)
+```
+
+The default is empty, which is the *less* safe direction, deliberately — a
+narrow default that guesses wrong (keying on the host for contents that
+actually depend on the target) is worse than an obviously-too-broad one.
+Changing `version` yields a fresh, empty cache without touching anything
+else.
+
+## Auditing with `--no-scratch`
+
+`--no-scratch` runs against a throwaway, empty cache instead of the stored
+one — proving the contract rather than assuming it. It implies `--force`:
+since a scratch never enters the cache key, a cache hit would otherwise just
+replay the answer the audit exists to check. The stored cache itself is
+never touched.
+
+```bash title="terminal"
+heph run --no-scratch //build:compile
+```
+
+A target that reads an unset scratch variable under `set -u` fails outright
+rather than running cold — write `${GOCACHE:-}` if the audit should report
+on the target instead of on the shell.
+
+## Sharing through the remote cache
+
+With a [remote cache](/docs/reference/configuration#caches--remote-shared-caches)
+configured, `remote = True` lets a scratch's contents travel between
+machines. A build pulls automatically on a cold cache; nothing is ever
+published as a side effect of building — that's always an explicit step, so
+CI runs it last:
+
+```bash title="terminal"
+heph tool scratch push --all --producer "$CI_RUN_ID"   # publish every remote = True cache
+heph tool scratch pull --all                            # warm a machine ahead of time
+```
+
+Which lineage a run reads and writes is controlled by `scratch.scope` in
+`.hephconfig` — see
+[`scratch` — branch-lineage policy](/docs/reference/configuration#scratch--branch-lineage-policy).
+By default every run shares one lineage; scope it per branch so work on one
+branch doesn't overwrite another's cache, and a fresh branch starts from its
+base instead of from nothing.
+
+## Inspecting and reclaiming
+
+```bash title="terminal"
+heph tool scratch ls                       # every cache: address, access, size, lineages present
+heph tool scratch head //build:gocache     # why a build would be warm or cold
+heph tool scratch path //build:gocache     # the on-disk directory
+heph tool scratch rm //build:gocache       # drop one; always safe
+heph tool scratch rm --all                 # drop every scratch cache
+```
+
+`heph tool scratch head` is the one worth remembering: it prints every
+candidate lineage a build would consult, local and remote, in order, and
+which one wins — the answer to "why did my branch start cold?", a question
+the directory itself can't answer since the interesting part is what
+*wasn't* found.
+
+Scratch caches also come under `heph tool gc`:
+
+```bash title="terminal"
+heph tool gc --scratch-max-size 50GiB --scratch-max-age-days 30
+```
+
+## A broad output beside a mounted scratch
+
+A target's own outputs must never come from inside its scratch — that would
+make the artifact's bytes depend on unhashed, mutable state. Give a mounted
+scratch (one with `path` set) a narrow `out` pattern in the same directory:
+a glob broad enough to reach the mount (`out = "**/*"`, say) fails when heph
+packs the result, naming the mount path in the error.
+
+An unmounted scratch (no `path`, the env-var-only form) sidesteps this
+entirely — there's nothing in the sandbox tree for an output to reach, which
+is why it's the shape to reach for first when a tool is happy to be told its
+cache directory through an environment variable alone.
+
+## Who uses one already
+
+The [Go plugin](/docs/plugins/go#build-and-module-caches) shares one
+`GOCACHE` per module and per build variant, and one portable `GOMODCACHE`
+for downloaded modules, both without any configuration.
+
+The [OCI plugin](/docs/plugins/oci#oci_pull)'s `oci_pull` shares one
+registry blob store across every pull in the workspace — two images sharing
+a base layer download it once, not once per pull.
diff --git a/website/docs/plugins/go.md b/website/docs/plugins/go.md
index 5034552..a09263e 100644
--- a/website/docs/plugins/go.md
+++ b/website/docs/plugins/go.md
@@ -245,6 +245,45 @@ options:
 `cctool` is resolved only when a race build that needs cgo actually runs — an
 ordinary build, and a darwin race build, never touch it.
 
+## Build and module caches
+
+Every Go list, compile, and third-party download step shares two caches
+automatically — nothing to configure:
+
+- **`GOCACHE`** — one per Go module (the directory holding `go.mod`) and per
+  [build variant](#build-variants), shared by that module's `go list` and
+  `go tool compile` steps alike.
+- **`GOMODCACHE`** — one shared, portable module-download cache for the
+  whole workspace. heph owns it outright: a host-set `GOMODCACHE` is not
+  passed through, so the shared cache is what every build sees.
+
+Both are [scratch caches](/docs/concepts/scratch) — visible in
+`heph tool scratch ls`, droppable with `heph tool scratch rm` if one goes
+bad, and auditable with `heph run --no-scratch`. Point a hand-written target
+at the same `GOCACHE` a module's own targets already share with
+`heph.go.gocache_addr()`:
+
+```python title="BUILD"
+target(
+    name    = "custom_build",
+    driver  = "bash",
+    scratch = [heph.go.gocache_addr()],
+    run     = "go build -o $OUT ./cmd/tool",
+    out     = "tool",
+)
+```
+
+| Argument | Default | Meaning |
+|----------|---------|---------|
+| `goos` / `goarch` | this machine's | Target platform of the cache to share. |
+| `gotool` | `"host"` | Toolchain selector — `"host"`, a pinned version, or a target address. Match whatever the variant you're building against declares. |
+| `tags` / `goexperiment` / `gcflags` / `ldflags` | `[]` | The rest of the variant's factors — pass what the variant uses, or the target warms a different cache than the one it meant to share. |
+| `race` | `false` | Address the race-detector variant of the cache. |
+
+`gocache_addr()` resolves the module from the *calling* BUILD file's nearest
+`go.mod`, so a target gets the same cache the Go driver's own targets in
+that module already use.
+
 ## Linting and formatting
 
 A Go module gets four extra targets the moment it has a `.golangci.yml` (or
diff --git a/website/docs/plugins/oci.md b/website/docs/plugins/oci.md
index 42f6aba..7b55bf0 100644
--- a/website/docs/plugins/oci.md
+++ b/website/docs/plugins/oci.md
@@ -341,6 +341,20 @@ WARN oci_pull: "alpine:3.20" currently resolves to
 See [Reproducibility](/docs/concepts/reproducibility) for the same principle
 applied elsewhere in heph.
 
+#### Shared blob cache
+
+Every `oci_pull` target draws downloaded registry blobs from one shared
+[scratch cache](/docs/concepts/scratch), keyed by nothing but its own
+address — blobs are content-addressed by digest, so one store holds `amd64`
+and `arm64` layers side by side. Two images sharing a base layer download it
+once, not once per pull target. Nothing to configure; inspect or reclaim it
+like any other scratch:
+
+```bash title="terminal"
+heph tool scratch ls          # the blob store shows up alongside your own caches
+heph tool scratch rm --all    # drop it if it ever needs a clean slate
+```
+
 ### `oci_push`
 
 Pushes an image archive to a registry. An **action**, not an artifact — it
diff --git a/website/docs/reference/configuration.md b/website/docs/reference/configuration.md
index 1eda6e0..915e175 100644
--- a/website/docs/reference/configuration.md
+++ b/website/docs/reference/configuration.md
@@ -286,6 +286,29 @@ Run `heph tool cache measure-latency` to force a fresh latency measurement and
 print per-cache round-trip times. See the [Remote cache guide](/docs/guides/remote-cache)
 for a complete walkthrough.
 
+## `scratch` — branch-lineage policy
+
+Controls which [scratch cache](/docs/concepts/scratch) lineage a run reads
+and writes. A scratch's *contents* are declared by its target; this is about
+which **copy** of them a given run sees.
+
+```yaml title=".hephconfig"
+scratch:
+  scope: "${git:branch}"
+  restoreScopes: ["master"]
+  seedOnFork: true   # default
+```
+
+| Key | Type | Default | Description |
+|-----|------|---------|-------------|
+| `scope` | string | `""` (one shared lineage) | The lineage this run **writes** to. `${git:branch}` resolves to the current branch, so a developer gets per-branch caches without wiring anything; CI overrides it with whatever names the branch there. |
+| `restoreScopes` | string[] | `[]` | Lineages this run may **read** from when its own scope has nothing yet, tried in order — `["master"]` is the common case. Never written to: a branch can read its base's cache but never overwrite it. |
+| `seedOnFork` | bool | `true` | Copy the first warm fallback lineage into a cold scope on its first build there, instead of starting from nothing. Turn off only for a large cache on a filesystem without reflink support, where the copy itself gets expensive. |
+
+See [Scratch caches](/docs/concepts/scratch) for the caches themselves, and
+`heph run --no-scratch` for auditing that a target doesn't secretly depend on
+one.
+
 ## `fuse` — sandbox overlay
 
 Controls whether sandboxes are assembled with a FUSE overlay instead of copying
diff --git a/website/sidebars.ts b/website/sidebars.ts
index 7bcf2d6..0ab6361 100644
--- a/website/sidebars.ts
+++ b/website/sidebars.ts
@@ -22,6 +22,7 @@ const sidebars: SidebarsConfig = {
         'concepts/dependencies',
         'concepts/reproducibility',
         'concepts/caching',
+        'concepts/scratch',
         'concepts/sandbox',
         'concepts/codegen',
         'concepts/approval',