Files
PaperClipAI/doc/runner-image-contract.md
T
Nicky LeachandClaude Fable 5.2 503a736195 feat(ci): read a registry build cache in the Runner image build (#21310)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The Runner is the part that executes agent work: a daemon
(`paperclip-runnerd`) and a provider pack of agent tools
> - `.github/workflows/runner-image.yml` builds the Runner alone and
publishes it from `master` as `ghcr.io/paperclipai/paperclip-runner`,
one time for each set of inputs
> - The build used no cache, so every run compiled the daemon again. The
first publication took 25m46s from push to tag
> - The toolchain and the daemon change only with the Rust sources. Most
commits that change the Runner's inputs do not touch them
> - This pull request lets the build legs read a registry build cache
for those stages. Only a new job can write the cache, and only for a
`master` push
> - A weekly run builds with no cache, so the cache cannot hide a broken
build for long
> - The benefit is about 3 minutes less on each build leg when the Rust
sources did not change, with no change to the published image

## Linked Issues or Issue Description

Refs #21230 (publication of the Runner image), #21229 and #21204 (the
build this change makes faster). No issue exists for this change.

**What existing behavior does this improve?**

The Runner image build in `.github/workflows/runner-image.yml`. It
builds the same image in less time.

**Current behavior**

Each build leg starts from an empty builder. It installs the Rust
toolchain and compiles the daemon in every run. In the first `master`
publication (run 43365670428) the amd64 leg spent 7m22s in the build
step. The toolchain and the cargo build took 3m26s of that.

**Proposed behavior**

The build legs read
`ghcr.io/paperclipai/paperclip-runner-buildcache:<arch>` for the stages
up to the compiled daemon. A new `cache` job writes that cache after a
publication. A scheduled run builds with no cache one time each week.

**Reason and benefit**

The stages before the last one read the Rust sources, the protocol files
and the toolchain pin. In the last 30 days, 438 `master` commits changed
an input of the image, and at most 181 of them changed a file those
stages read. For the other commits, and for each pull request that does
not touch those files, the build step is about 3 minutes shorter on each
leg.

**Breaking changes**

None. The `Dockerfile` and each file in the image identity are
unchanged. The published digest for a set of inputs does not change.

This is part 4 of 4 of a stack ([stack
#21309](https://github.com/paperclipai/paperclip/stack/21309)). Merge
the parts in order:

1. #21306: the resolver and the no-rebuild mode
2. #21307: the report-only job in `docker.yml` and the pull-request
check of the application image's Runner path
3. #21308: reading a retained provider pack from its Runner image
4. #21310: a registry build cache for the Runner image build (this pull
request)

This pull request uses no code from parts 1 to 3. It is stacked on them
because all four parts edit `doc/runner-image-contract.md`.

## What Changed

- The build legs pass `--cache-from
type=registry,ref=ghcr.io/paperclipai/paperclip-runner-buildcache:<arch>`
to the first build command. They do not log in. Their permissions stay
`contents: read` and `packages: read`.
- A new `cache` job writes the cache with `--cache-to
type=registry,...,mode=max` and `--output type=cacheonly`. It runs only
for a `master` push in the canonical repository, for inputs that had no
image. Its permissions are `contents: read` and `packages: write`. It
has no `id-token` and no `attestations` permission.
- The `cache` job needs the `publish` job, and no job needs the `cache`
job. A cache write cannot delay or fail a publication.
- The `cache` job builds the target `runner-artifacts-runnerd` and no
later stage. The last stage takes the commit as a build argument, so no
other commit can reuse it. The install and its dependency lifecycle
scripts are in that stage, so the job with the write token does not run
them.
- The `cache` job has one concurrency group for each architecture, a job
time limit and a step time limit.
- A `schedule` trigger starts a weekly build. It passes `--no-cache` to
the first command and reads no cache. It runs all checks and publishes
nothing. A fork's schedule runs nothing.
- `server/src/__tests__/runner-image-workflow.test.ts` pins the new
security properties: which jobs hold `packages: write`, that the build
legs have no credential and cannot write the cache, the conditions of
the `cache` job, that the `publish` job does not depend on it, and that
a scheduled run cannot publish or write the cache.
- `doc/runner-image-contract.md` has a "Build cache" section.

## Verification

Tests, on macOS (Node 26.12.0) and in a `node:24.22.3-bookworm`
container (Linux arm64, Node 24.22.3):

- `pnpm exec vitest run src/__tests__/runner-image-workflow.test.ts
src/__tests__/runner-image-identity.test.ts` in `server/`: 55 of 55
tests pass on both.
- `node --test .github/scripts/tests/*.test.mjs`: 807 of 807 tests pass
on both. The container needs `jq`.
- `pnpm verify` exits 2 for this change, because a workflow file is a
global CI input. Required CI supplies the broad coverage.
- Each new assertion fails when the workflow breaks its rule. I checked
seven such edits, for example: the `publish` job needs the `cache` job,
the `cache` job gets `id-token: write`, a build leg gets `--cache-to`.

Cache mechanics, on linux/arm64 with a local registry, BuildKit v0.34.2
(the version the hosted runners use), and an empty builder for each
build:

| Build of both variants | Time | Result |
| --- | --- | --- |
| No cache, cache ref does not exist | 260 s | An importer error in the
log, exit 0 |
| One changed TypeScript file, with the cache | 149 s | Toolchain and
cargo steps cached |
| Unchanged inputs and a new commit, with the cache | 129 s | Toolchain
and cargo steps cached |
| Each stage in the cache (a whole-build cache, which this change does
not write) | 16 s | Both OCI archives complete |
| `--no-cache` on the first command | 272 s | No cached step in the
first command. The second command reuses all of it |

- The images built from the cache pass `runner-image-layout.mjs check`
and `runner-image-oci.mjs check-layers`. The daemon has the same SHA-256
as the daemon of a cold build.
- A cache ref that does not exist, a host that refuses the connection, a
name that does not resolve, and the real
`ghcr.io/paperclipai/paperclip-runner-buildcache:arm64` (which does not
exist yet) each give an importer error in the log and a complete build
with exit 0.
- The `cache` job command on an empty builder with the cache present
takes 6 s. It downloads no layer.

Not tested before merge:

- The `cache` job. It runs only on `master`.
- The scheduled run. It starts only from the default branch.
- The cache package. It does not exist. I did not test a private
package, which answers an anonymous read with an authorization error. I
expect the same importer error and a cold build.
- A fully cached build on the hosted runners. The local test used the
same BuildKit version and a different buildx version (0.37.0, and the
runners have 0.38.1).

## Risks

- **A second job can write packages.** Before this change, only the
`publish` job held `packages: write`, and it runs no build. Now the
`cache` job holds it too, and it runs BuildKit on the repository's
`Dockerfile`. GitHub cannot limit that token to one package, so it can
also write to `ghcr.io/paperclipai/paperclip-runner`. The limits are:
the job runs only for a `master` push in the canonical repository, the
token goes only to the registry login, the build steps run inside
BuildKit without it, the job does not run the install, and the job
cannot sign. An image that it wrote has no attestation.
- **Part of a published image can come from an earlier run.** The daemon
in a published image can come from a step that an earlier `master` run
compiled and cached. The attestation names the run that published, not
the run that compiled. BuildKit reuses a step only for the same
instruction and the same file content, but a party that can write the
cache package can replace what a later build reuses. Write access to the
cache package must stay as narrow as write access to the image package.
- **A cached download is not repeated.** The toolchain installer, the
toolchain and the crates are not downloaded again until the inputs of
their step change. The weekly cold build finds a download that is gone.
It does not compare its output with the cache.
- **A registry failure after the cache import fails the leg.** A cache
that cannot be read at the start is not a failure. A layer download that
fails later is a failure. Run the job again.
- **The cache package starts private.** GitHub creates it on the first
write. A maintainer must make it public one time. Until then each build
runs cold and nothing fails.
- **The first cache write waits for an input change.** The merge of this
pull request changes no input of the image, so its `master` run builds
nothing and writes no cache.
- **A cancelled cache job marks its run as cancelled.** When three cache
writes for one architecture wait at the same time, GitHub cancels the
middle one. The image of that run is published.

> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.

## Model Used

Claude Fable 5.2 (`claude-fable-5-2`), Anthropic, through Claude Code.
Tool use and shell execution. It wrote the change, ran the local builds
and tests, and wrote this description.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-12 23:46:15 -08:00

17 KiB

Runner image contract

The Runner image is the Runner built alone, as a plain file tree with no entrypoint and no operating system. An image build copies paths out of it instead of compiling the Runner again. Nothing in this repository consumes it yet. Its name is ghcr.io/paperclipai/paperclip-runner, and one build makes two variants:

  • full, the runner-artifacts target of the Dockerfile: every path below;
  • slim, the runner-artifacts-slim target: the same without the two paths marked "full only", the provider pack's transport archives.

Layout

Every path a consumer may copy (RUNNER_IMAGE_PATHS in scripts/runner-image-layout.mjs). Every file is readable by every user.

Path Content
/runner-image.json Identity and description of this image.
/paperclip-runnerd The daemon. A stripped release binary, mode 0755.
/provider-pack/ The remote provider pack, with provider-pack.json.
/provider-pack.archive.json Full only. Receipt of the pack's transport archives.
/provider-pack.archives/ Full only. The transport archives the receipt lists.
/runner-package/package.json The Runner package manifest.
/runner-package/dist/ The Runner's complete build output. dist/bin/paperclip-runnerd is the same file as /paperclip-runnerd.
/runner-package/node_modules/.cache/paperclip-runner-build-inputs.json The build stamp the Runner's build recorded for that output.
/runner-package/scripts/provision-grok.mjs Provisions the native Grok executable. Reads platforms.json by relative path.
/runner-package/src/providers/grok/platforms.json The Grok release pins.
/runner-package/src/live/runnerd-codex-transport.ts The transport source, as a contract file. It defines RUNNER_CLIENT_VERSION.
/bundles/runnerd-codex-transport.mjs The pack's dist/live/runnerd-codex-transport.js as one module. It imports without node_modules.
/pnpm-lock.yaml The lockfile the image was installed from.

Paths under /runner-package/ keep their place inside packages/paperclip-runner. runner-image.json describes one platform, and is the same file in both variants:

  • inputsDigest (the identity below) and buildInputsDigest (the digest in the Runner's build stamp);
  • sourceRevision: the commit that built the image, also in the pack as payload.runnerSourceRevision. An image is published once for its inputs, so later commits with the same inputs keep this revision;
  • publication: true only for a build with the build argument PAPERCLIP_RUNNER_IMAGE_PUBLICATION=1. Only then does the pack carry the identity, as payload.runnerInputsDigest. Another image build that runs the stage inline builds the same Runner but is not this image;
  • platform (linux, amd64 or arm64) and libc: the glibc the daemon was compiled on (Debian bookworm, the oldest a consumer runs) and the newest glibc version it needs;
  • providerPack: the pack's digest and target;
  • paths and sha256: the table above, and the digest of every single file;
  • fullImageOnly: the names in paths that only the full image has.

Layers

  1. The pack's dependency files of 1 MiB or more: provider binaries and Node.
  2. Everything else, with runner-image.json.
  3. Full only: the transport archive of each file in layer 1.
  4. Full only: the other transport archives, the overlay archive, the receipt.

The slim image is layers 1 and 2. The full image is those same two layer blobs, then layers 3 and 4, which hold transport content and nothing else. Layers 1 and 3 have the same bytes in every build with the same dependency files, so a new image shares them with the one before it. A large file under the pack's dist is the Runner's own output: it and its archive are in layers 2 and 4. An archive is named by its file's place in the receipt, which lists the dependency files first, so their names do not depend on dist.

Identity

Both variants are named by one digest: the variant is not part of the identity. The full image has the tag inputs-<digest> and the slim image inputs-<digest>-slim. scripts/runner-image-identity.mjs computes the digest from:

  • every file under the paths the build copies from the build context (RUNNER_IMAGE_CONTEXT_INPUTS, which include pnpm-lock.yaml): path, executable bit and content, and each symbolic link by its target;
  • the whole Dockerfile: the recipe, the base image pinned by digest and the toolchain pins. Any edit to it makes a new image;
  • the two platforms, linux/amd64 and linux/arm64.

The source revision is not part of it, and neither is the text of .dockerignore: a rule changes the digest only when it changes which of those files the build receives. The build computes the digest for itself:

docker buildx build --target runner-artifacts-identity --output type=local,dest=/tmp/identity .
jq -r .digest /tmp/identity/runner-image-identity.json

The install is frozen. It fails when the lockfile is behind the manifests the build reads (the root one, packages/paperclip-runner and packages/paperclip-eval-kernel). A contributor cannot edit the lockfile, so pull-request runs resolve it first. For a local build, run pnpm install --resolution-only --ignore-scripts --no-frozen-lockfile.

Checks

CI builds each variant as an OCI archive and nothing else. node scripts/runner-image-oci.mjs unpack <oci archive> --out <tree> writes the tree of an image from its layers, so the tree CI checks is the bytes of the image. It refuses a whiteout, an absolute or .. path, a path written twice and an entry behind a link.

node scripts/runner-image-layout.mjs check <tree> --variant slim|full runs in the build and again on each unpacked tree. The variant has no default: a slim tree with the transport paths fails, and so does a full tree without them. The check fails unless:

  • every documented path exists, matches its digest in runner-image.json, and resolves inside the tree;
  • the build stamp names the inputs the identity names;
  • every artifact in provider-pack.json has its digest, the dist tree and each provider asset tree match their tree digests, and the pack names this image's revision and platform;
  • the pack names the inputs digest in a publication build and in no other. --expect-inputs-digest fails on any other build;
  • in a full image, every transport archive matches its receipt, and the archives extract to the pack. The check extracts them as a sandbox does (GNU tar with the options in provider-pack-layers.ts) and compares paths, content, link targets and all twelve permission bits, after the a+rX grant the build makes once the archives are written.

node scripts/runner-image-oci.mjs check-layers --slim <oci archive> --full <oci archive> checks "Layers" on the two archives, and that no file is in two layers.

The build also runs paperclip-runnerd --build-metadata on every base image in docker/runner-image/Dockerfile.smoke. Add a base there before a consumer uses it.

Consuming the image

Read the image only with COPY --from, and pin it by digest:

FROM ghcr.io/paperclipai/paperclip-runner@sha256:<digest> AS runner-artifacts
FROM debian:trixie-slim
COPY --from=runner-artifacts /paperclip-runnerd /usr/local/bin/paperclip-runnerd

A Dockerfile with a runner-artifacts or runner-artifacts-slim stage of its own, as in this repository, can take a built image or tree in its place with --build-context <stage>=<image or directory>. The stage is then not built.

Choosing the Runner for an image build

scripts/runner-image-resolve.mjs decides whether a build of a checkout can copy the published image. No build uses its answer yet: on a master push, the runner_image_report job of .github/workflows/docker.yml runs it and writes the answer to the job summary, and nothing reads that job. The job runs on the committed files, holds no credential (the Runner package is public, so the lookup is anonymous) and cannot fail the run. The script has two commands, so that the one that builds reads no registry credential:

node scripts/runner-image-resolve.mjs inputs --dir "$dir"
node scripts/runner-image-resolve.mjs resolve --dir "$dir"

inputs exports the Runner build context, checks the lockfile and computes the inputs digest in the build, as the publication workflow does. Run it on the committed files, before any step that rewrites pnpm-lock.yaml. resolve looks up inputs-<digest> and reports:

  • mode: published or inline, and reason;
  • image and slim_image: both variants, pinned by index digest. The slim digest is the one the full index names (see "Pairing");
  • build_contexts: one --build-context value on each line, which replace the runner-artifacts and runner-artifacts-slim stages;
  • labels: io.paperclip.runner.image and io.paperclip.runner.inputs-digest, for the image that copies the Runner.

published needs a complete publication whose full index names these inputs and both platforms. Every other answer is inline, with the last four values empty: the lockfile is behind the Runner's manifests (lock-behind), the image does not exist (absent), a publication is incomplete or broken, the full index does not match (mismatch), or the registry refused, failed or could not be reached (unknown). An inline build runs the Runner stages itself and carries neither label. The command does not wait for a publication and does not verify the attestation of the image.

Publication

.github/workflows/runner-image.yml publishes on a master push in this repository and nowhere else. Other runs build and check, and publish nothing.

  • Lockfile. A master push builds from the committed pnpm-lock.yaml, so the digest is a function of the commit. When that file is behind the Runner's manifests, the run writes a notice, publishes nothing and stays green. The commit that merges the lockfile refresh publishes the image.
  • Once. The build runs only when inputs-<digest> does not exist.
  • Verified bytes. The build legs cannot write to the registry. The publish job pushes the images they checked by digest and joins each variant into a multi-platform index under a candidate- tag. It attests both indexes, and only then writes the two inputs- tags. candidate- tags are internal: one can name an index that is not attested, and the next run overwrites it. Do not consume them.
  • Order. It writes inputs-<digest>-slim first and inputs-<digest> last, so the full tag marks a complete publication. The full index names its slim index in the annotation io.paperclip.runner.slim-digest.
  • Pairing. Resolve the full tag first. Take the slim digest from that annotation of the full index, and pull the slim image by that digest. Never resolve inputs-<digest>-slim by tag on its own: while a publication is incomplete, that tag can name the slim image of another run.
  • Recovery. A run that stops between the two tags leaves a slim tag alone. Those inputs are not published. The next run for them builds again, replaces that tag with its own slim image and then writes the full tag. A full tag whose slim tag is missing or names another image fails every run for those inputs. An operator must then delete both package versions and re-run the workflow run of a master push with those inputs.
  • Create-only. The workflow never moves a tag of a complete publication, but the registry does not enforce that. Pin the index digest and verify it: gh attestation verify oci://ghcr.io/paperclipai/paperclip-runner@sha256:<digest> --repo paperclipai/paperclip --signer-workflow paperclipai/paperclip/.github/workflows/runner-image.yml --source-ref refs/heads/master
  • First publication. GitHub creates the package private. A maintainer must make it public once.

Build cache

The build legs read a registry build cache, ghcr.io/paperclipai/paperclip-runner-buildcache:<arch>. It is a package of its own and holds no image, so the image package holds only attested images. It covers the stages that do not take the commit: the Rust toolchain and the compiled daemon. The last stage takes the source revision as a build argument, so every build runs it: the frozen install, the Runner's build and the provider pack. Of the two build commands of a leg, only the first reads the registry cache; the second reads what the first left in the job's builder.

  • Effect. BuildKit reuses a cached step only when the step and the content of every file it reads are the same. The cache changes how long a build takes, not what it builds, and the identity does not cover it.
  • Writer. Only the cache job of the workflow writes it: on a master push in this repository, after the publish job, for inputs that had no image. When the image exists, nothing was built, and the run that published it wrote the cache for those inputs. That job cannot sign, builds no further than the daemon and never runs the install. Its token reaches the registry login and no other step. It reads the cache it writes and keeps every step of the cached stages (mode=max). Every other run reads the cache without a credential and cannot write it.
  • Not on the way to an image. No job waits for the cache job. When it fails, the image is published and the next build is slower.
  • One write at a time. The job holds one concurrency slot for each architecture. GitHub keeps one waiting job in a group and cancels the one that waited before it, so the newest write waits. A write from an older commit that lands last costs time and nothing else.
  • No Actions cache. The workflow never uses it: workflows that handle pull requests can write it, and what they write must not reach a published image.
  • Trust. The daemon in a published image can come from a cached step that an earlier master run compiled. Whoever can write the cache package can decide what a later build reuses, so write access to it must stay as narrow as write access to the image package.
  • Fetches. A cached step does not run again. A download inside one (the toolchain installer, the toolchain, the crates) is not repeated until the inputs of that step change.
  • Weekly cold build. A scheduled run builds both variants once a week with no cache: it passes --no-cache and reads no registry cache. Like a manual run, it resolves the lockfile, always builds, runs every check and publishes nothing. A fork's schedule runs nothing. It fails when a download is gone or when a step only works from the cache.
  • Missing cache. A cache that does not exist or cannot be read is a line in the build log, not a failure. The build then runs every step.
  • First write. GitHub creates the cache package private. A maintainer must make it public once. Until then every build runs cold.

Reading the transport archives alone

The server needs the transport form of an older pack when a retained session names a pack that is not the server's own. A pack with payload.runnerInputsDigest came out of this image, so the server reads it from inputs-<digest> (retained-provider-pack-oci.ts). It does not count layers. It uses the rule of "Layers": the full image is the slim image, then layers that hold transport content and nothing else.

  • The full index must carry the annotations io.paperclip.runner.inputs-digest, equal to the pack's digest of inputs, org.opencontainers.image.revision, equal to the pack's runnerSourceRevision, and io.paperclip.runner.slim-digest.
  • The server reads the slim index by that digest, never by the slim tag, and checks the bytes against the digest.
  • For the pack's platform, the layers of the slim image must be the first layers of the full image, as the same blobs in the same order. The layers after them are the transport layers: at least one and at most eight.
  • The server downloads only those layers. It refuses any entry in them that is not a transport archive or the receipt, and any file that is in two of them.
  • The importer then checks every archive against the pack's manifest, as for any other source.
  • The record the server keeps of the source names the digest of inputs and every transport layer under runnerImage. Its transportLayerDigest is then the last of those layers and transportLayerBytes their total.

When the image cannot be read (a tag or the slim index does not exist, or the registry refuses a request), the server reads the application image tagged sha-<runnerSourceRevision>, as it does for a pack without runnerInputsDigest. An image that can be read and does not match fails.

Known limits

  • A build without PAPERCLIP_RUNNER_SOURCE_REVISION or PAPERCLIP_BUILD_COMMIT has no pack. It is for local use.
  • The build runs the steps of the Runner's build script that write output, not its drift checks; repository CI runs those.
  • A build that vendors /runner-package/ reuses dist only when its own Runner build-inputs digest equals the stamp's. The image computes that digest in the .dockerignore-filtered context; a checkout gives another.
  • # syntax=docker/dockerfile:1.20 is a tag. A new frontend release under it does not change the identity.
  • The archive check does not tell a directory with its own archive member from one created only as a file's parent.
  • A lockfile that the lockfile refresh cannot bring up to the Runner's manifests leaves every master run green with only a notice, for as long as it lasts. No image is published in that time, and nothing fails.