Files
PaperClipAI/docker/daytona-runner/README.md
T
6874fed05c feat(adapters): refresh model catalogs and sandbox CLI pins (Oct 23 audit) (#17809)
<!-- entire-trail-link-start -->
https://entire.com/gh/paperclipai/-/trails/26/changes/paperclip/26
<!-- entire-trail-link-end -->

## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - Agents run on coding harnesses. Each adapter has a model catalog,
and sandboxes contain pinned CLIs.
> - Upstream providers released new CLI versions and new models. OpenAI
retired some Codex models for ChatGPT sign-in.
> - A stale catalog hides valid models. A misleading error makes
operators upgrade a CLI that is already current.
> - This pull request adds verified OpenCode model IDs, names Codex
ChatGPT retirements in the environment Test, and refreshes the sandbox
CLI pins.
> - The benefit is that operators can select current models and get the
correct reason when a retired model fails.

## Linked Issues or Issue Description

Refs #16708 (adds Claude Haiku 5.5 to `claude_local`; this PR does not
duplicate it).
Refs #17245 (moves the Codex runtime to 0.163.0; this PR does not change
Codex pins).

**What happened?**

OpenAI retired `gpt-5.5` (October 14, 2026) and `gpt-5.4` /
`gpt-5.4-mini` (August 31, 2026) from Codex with ChatGPT sign-in. A
saved agent with one of these models and ChatGPT sign-in now fails. The
legacy environment Test says that the plan, the rollout, or an old CLI
can be the cause. The Paperclip Runner Test gives generic account and
runtime advice. The real cause is the retirement.

**Expected behavior**

Both Tests give the retirement date and the documented replacement.
API-key users keep the models.

**Proposed change**

Record the verified retirements. Add the new verified OpenCode IDs.
Refresh the sandbox CLI pins with verified digests. Record the audit.

## What Changed

- Codex: add a table of models retired from Codex with ChatGPT sign-in
(`gpt-5.5` → `gpt-6.1-sol`, or `gpt-6-luna` on Free and Go; `gpt-5.4` →
`gpt-6-sol`; `gpt-5.4-mini` → `gpt-6-luna`). The shared
`codexChatGptRetirementHint` builds the hint. The legacy CLI Test
(`codexHelloProbeModelRejectionCheck`) and the Paperclip Runner Test
(`testNativeRunnerAuthentication`) both use it when the backend rejects
one of these models. For other failures, the Paperclip Runner Test hint
stays the same. The picker keeps the models for API-key users.
- OpenCode: add `anthropic/claude-haiku-5-5` and `google/gemini-4-argon`
to the remote fallback catalog. The Paperclip Runner OpenCode picker
also uses this catalog.
- Evaluation image sandbox layer (`docker/daytona-runner`):

  | CLI | Before | After |
  | --- | --- | --- |
  | Grok CLI | 1.0.46 | 1.0.70 |
  | Gemini CLI | 0.62.0 | 0.65.0 |
  | Kimi Code | 2.1.1 | 2.3.3 |
  | Cursor CLI | 2026.10.01-e373342 | 2026.10.22-27dcf5f |
  | GitHub CLI | 2.102.0 | 2.103.3 |

- Add `doc/adapter-model-audit-2026-10-23.md` with sources,
compatibility findings, and deferred items.

No model in this PR needs a newer Codex, Claude, or OpenCode runtime
than the current pins. Claude Haiku 5.5 needs Claude Code 2.1.293 or
later; #16708 records that floor, and this PR does not add it.

## Verification

- `vitest run src/index.test.ts src/server/test.remote.test.ts
src/server/acp.cli-version.test.ts` in `packages/adapters/codex-local`:
41 passed.
- `vitest run src/services/native-runtime/setup-readiness.test.ts` in
`server`: 98 passed. These include the two new Paperclip Runner Test
cases: a retired model shows the retirement hint, and a model that is
not retired keeps the generic hint. The retirement case fails on the
earlier code. Run with the host's empty `ANTHROPIC_BASE_URL` unset; see
the baseline note in the audit doc.
- `vitest run` in `packages/adapters/opencode-local`: 87 passed.
- `vitest run src/__tests__/adapter-models.test.ts` in `server`: 27
passed.
- `tsc --noEmit` in `packages/adapters/codex-local` and
`packages/adapters/opencode-local`: passed.
- Cursor archive SHA-256
`85296944e4600a83c4b764f578d1502bf9c416db200d33a593e8f3ecabfdc294` is
computed from the versioned Linux x64 tarball. The extracted binary
reports `2026.10.22-27dcf5f`. On October 26, 2026,
`https://cursor.com/install` still resolves this version.
- GitHub CLI SHA-256
`4c806978ffce1cec70853302a56d70ac6da4f19bae7adca4e1a8ab00633dd91d`
matches `gh_2.103.3_checksums.txt`.
- Grok CLI 1.0.70 and Kimi Code 2.3.3 installed from npm report those
versions. The flags that Paperclip uses (`-m`, `--effort`, `grok
models`; `-m`, `-p`, `--output-format`, `acp`) are present in `--help`.
- `git merge-tree` with #16708 (head `3a06740e9`) and #17245 (head
`41f9c611f`): this PR shares no files with either PR. The two PRs can
merge in any order.
- Not run in the authoring environment: the full `pnpm -r typecheck` and
`pnpm build` (no Rust toolchain), the full server `tsc` (memory limit),
and the image build. CI runs these.

## Risks

- Low risk for catalogs: the IDs are additive suggestions. No saved
selection, default, or spending setting changes.
- The Paperclip Runner Test changes only the hint text, and only for a
ChatGPT rejection of a retired model. Check codes and pass/fail results
do not change.
- Sandbox CLI pins:
- Kimi Code 2.2.0 and 2.3.0 remove the `kaos` MCP executor and require
qualified plugin skill names. Paperclip uses neither. Kimi Code 2.3.2
stops automatic swarms in `kimi -p` and ACP sessions.
- Gemini CLI 0.65.0 makes workspace settings read-only in untrusted
folders. Paperclip writes home settings and sets
`GEMINI_CLI_TRUST_WORKSPACE` for remote runs.
- xAI publishes no Grok CLI changelog. Grok 1.0.70 behavior changes are
verified only against the CLI reference and `--help`.
- The pins apply only after a new image is built and promoted. The fleet
sandbox image definition must get the same pins.
- Rollback: revert this PR.

## Model Used

- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, with tool use (shell,
web research, repository search).

## 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 Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-10-26 16:32:13 -07:00

10 KiB
Raw Blame History

Paperclip Daytona runner image

Ordinary provider packs now include pinned Cursor assets for Linux x64; no --candidate-providers=cursor flag is needed. Cursor admission is enabled after the production readiness gates passed. The public CLI installation command is paperclipai runtime setup cursor; it runs explicitly on the execution host and is never an npm installation hook.

This image is the Paperclip Cloud fleet sandbox image plus a source-built paperclip-runnerd and immutable provider pack. The pack contains Node 24.11, OpenCode 1.18.34, the compiled OpenCode proxy, ACPX 0.13.1 sidecar, qualified ACP agents, and the production lockfile. Its manifest digests each executable bridge and records the runner source revision as provenance. Its interface declaration allows compatible app releases to reuse it without uploading or installing a pack.

The fleet pins are intentionally copied from paperclip-cloud/fleet-sandbox-image/Dockerfile. Update both definitions together until the fleet base is published as a stable image that this Dockerfile can extend directly.

Harness versions

The October 2026 refreshes pin Codex 0.160.0, Claude Agent SDK 0.3.286 (Claude Code 2.1.286), and OpenCode 1.18.34 in the shared provider pack. Claude Code 2.1.280 is the minimum for Opus 5.5 and 2.1.284 is the minimum for Sonnet 5.5; the pinned runtime also supports Fable 5.1. Codex 0.160.0 ships bundled metadata for the current GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna model IDs. The sandbox layer pins Grok CLI 1.0.70, Gemini CLI 0.65.0, Kimi Code 2.3.3, Cursor CLI 2026.10.22-27dcf5f, and GitHub CLI 2.103.3 (October 2026 refresh). The native Cursor provider uses its separately verified distribution, pinned to 2026.09.26-dd393fe; it does not use the sandbox layer’s global Cursor executable. Grok CLI 1.0.70 supports the current Grok 4.7 model family. Hermes stays at 0.19.0, the newest release on PyPI.

Keep the patched ACP bridge versions separate from their CLI runtime pins. Their executable digests do not change when only the runtime dependency changes. Refresh the runtime executable digests from integrity-verified npm release archives for every supported platform, and keep the native runner, provider manifest, and remote controller version checks aligned.

Build and verify

Run pnpm --filter @paperclipai/paperclip-runner test:opencode:qualification after installing dependencies to exercise the actual pinned OpenCode executable. It checks health/version, session creation and retrieval, SSE messages, an async prompt, and session deletion against a loopback mock provider. It uses an isolated home, starts no paid model request, and retires its process group. Set PAPERCLIP_TEST_OPENCODE_BINARY to the materialized Linux executable when qualifying an assembled provider pack.

The fleet image is currently amd64-only because the pinned Cursor and GitHub CLI checksums cover amd64.

content_id="$(pnpm --silent test:e2e:runner:image-id)"
docker buildx build \
  --platform linux/amd64 \
  --build-arg PAPERCLIP_RUNNER_CONTENT_ID="${content_id}" \
  --build-arg PAPERCLIP_RUNNER_SOURCE_REVISION="$(git rev-parse HEAD)" \
  --tag "paperclip-daytona-runner:e2e-content-${content_id}" \
  --load \
  --file docker/daytona-runner/Dockerfile \
  .

docker run --rm --platform linux/amd64 \
  --entrypoint paperclip-runnerd \
  "paperclip-daytona-runner:e2e-content-${content_id}" \
  --build-metadata

The metadata must advertise dial_ws_loopback, dial_wss, and listen_ws. The explicit entrypoint is needed only for this local probe because Daytona's base image uses its own long-running sandbox entrypoint.

test:e2e:runner:image-id hashes the audited Docker build dependency closure, target platform, the immutable Dockerfile syntax-frontend digest, and every immutable FROM reference. It fails before the paid workflow can build when the frontend or a base is not pinned to a sha256 digest. When updating the syntax version, resolve and review its registry digest and update both values in the first Dockerfile line. Git commits that do not change those inputs reuse the same content tag.

PAPERCLIP_RUNNER_SOURCE_REVISION remains the full Git SHA that built the first published copy and is retained as provenance rather than cache identity.

Use in Paperclip

Publish the image to a registry Daytona can pull, or use the environment editor's Configure image flow to produce a Daytona snapshot. Set the environment image to that immutable tag or snapshot. Paperclip probes the sandbox user's PATH for paperclip-runnerd and codex and checks /opt/paperclip-runner/provider-pack for OpenCode and ACPX. It uses the pack when its selected provider interface and platform are compatible. App source revisions and unrelated provider pins do not force an update. Startup reads the small manifest in the existing runner metadata query. It does not scan the pack, run a version command, or transfer files on this path. Otherwise it stages the pack configured by PAPERCLIP_RUNNER_REMOTE_PROVIDER_PACK_PATH. Remote OpenCode and ACPX never fall back to host-local processes.

src/live/provider-pack-contract.json in the Runner package declares the OpenCode and ACPX interface revisions. Change a revision only for an incompatible runner/pack boundary change, and qualify the new pack before publishing it. Dependency versions and Git SHAs remain build provenance. ACPX retains its selected agent's qualified profile checks.

The image build makes the pack root-owned and non-writable to the execution user. Fast admission requires protected ancestors and manifest, with no writable path alias. Task-owned packs and older packs without an interface declaration keep full content verification. An attached session retains the exact pack and daemon identity admitted at launch. Publish one new image with the interface declaration to enable reuse across later compatible app releases.

Do not promote paperclip-runner-e2e-20260826-v2 for OpenCode or ACPX. Build a new immutable image or snapshot from a clean committed revision and pass that full Git SHA as PAPERCLIP_RUNNER_SOURCE_REVISION.

Do not bake provider credentials, Paperclip bootstrap tickets, or Daytona preview tokens into this image. They remain per-run secret material.

The provider-pack build pins the official Linux x64 Node 24.21.0 image by manifest digest. Its bundled Undici is 7.29.1, which fixes GHSA-3wwx-pv8p-q78v. Pi separately verifies its private Node executable and nested npm dependency; changing the outer interpreter does not replace either provider-owned pin.

Provider CLI updates are manifest-only changes: repository CI owns the root lockfile. Resolve the complete workspace manifest graph in the build context before invoking Docker, matching CI when a source commit precedes the lockfile bot. The trusted workflow supplies this resolved lockfile as an immutable artifact. The complete resolved lockfile must match PAPERCLIP_RUNNER_LOCK_SHA256 before package installation or lifecycle execution. Review and refresh that digest with source dependency changes; registry-time resolution drift fails closed. The Product E2E workflow resolves one lockfile before the image build. It verifies the downloaded artifact, then passes that artifact's SHA-256 as the PAPERCLIP_RUNNER_LOCK_SHA256 build argument. The Dockerfile checks the resolved lock against this value before installation. The fixed Dockerfile default is for standalone builds; it must not replace a campaign's verified lock digest. Refresh the default from the clean tracked lockfile using the exact pnpm install --resolution-only --ignore-scripts --no-frozen-lockfile command, and verify a second resolution preserves the digest. A lockfile left by a filtered or incremental install can retain stale importer patch identities. Refresh exact runtime versions and qualification digests together; do not download dependencies when a task starts.

Candidate ACP qualification assets

Provider branches can build their pinned assets with node packages/paperclip-runner/scripts/build-provider-pack.mjs /absolute/pack --candidate-providers=cursor (or copilot or pi). The source revision must include the named provider's builder. Assets are installed at build time under provider-assets/<provider>/<platform>-<architecture>. The pack manifest binds each complete asset tree. Runtime admission separately checks the provider's source-owned closure pins and copies a verified launch snapshot. A pack with candidate assets does not qualify or enable that provider.

For an isolated Linux x64 Daytona qualification image, pass --build-arg PAPERCLIP_RUNNER_CANDIDATE_PROVIDERS=cursor with the normal build arguments. Compute its content ID with the same selection: pnpm --silent test:e2e:runner:image-id --candidate-providers=cursor. Candidate assets and the default pack have distinct image identities. Never reuse the default image's content ID for a candidate build. Use each provider branch's recorded version and explicit model. Keep credentials out of images. Paid qualification requires bound provider and Daytona credentials, inspectable spend, and the shared $100 ceiling recorded in the capability report.

The pack builder tests its copied Node interpreter after relocation. Use a standalone Node distribution if the host interpreter depends on a package manager's private shared libraries. Pi additionally pins its complete portable interpreter and npm dependency graph. Refresh exact runtime versions and qualification digests together; never download dependencies when a task starts. Grok's additive native ACP profile keeps its qualified 1.0.13 executable at the verified sandbox prerequisite path. It does not replace the legacy adapter's grok command on PATH.

Native Grok is an image prerequisite at /opt/paperclip/providers/grok/1.0.13/grok. Its checksum-verified provisioning is separate from the provider pack, which ships only the built-in launcher. Public npm installation never downloads this binary.