## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - Adapter packages are the bridge between the control plane and local agent harnesses such as Claude Code, Codex, and Gemini CLI. > - ACP support was concentrated in a separate `acpx_local` adapter, which made ACP feel like a separate agent choice instead of an execution capability of the harness adapters. > - Claude, Codex, and Gemini now have ACP-capable harnesses, so the native adapter should own ACP selection, fallback, config, transcript parsing, and environment diagnostics. > - The standalone ACPX adapter still needs a compatibility path for existing rows, but it should not be offered as an active adapter for new agents. > - This pull request moves the shared ACP runtime into `@paperclipai/acpx-engine`, wires Claude/Codex/Gemini local adapters to prefer ACP when prerequisites are available, and retires `acpx_local` to a tombstone. > - The benefit is one adapter per harness, richer ACP transcripts by default where possible, and a migration path for existing Claude/Codex ACPX agents. ## Linked Issues or Issue Description Closes #5932 — the broken default `acpx_local` Claude path is replaced by native `claude_local` ACP support, existing Claude/Codex ACPX rows migrate to native adapters, and new agents no longer choose the standalone ACPX adapter. Refs #4893 — original merged ACPX local adapter runtime that this PR replaces with native per-harness ACP engines. Refs #6590 — prior ACPX-Claude seamlessness work folded into the new native Claude ACP path. Refs #197 — related open generic ACP/Kiro adapter work; this PR does not close it because Kiro/custom generic ACP remains a separate adapter decision. Refs #7018 — related Kimi-specific `acpx_local` shell failure; this PR retires the built-in standalone adapter but does not add a native Kimi adapter. Refs #8864 — related ACPX prompt/API guidance PR; this PR moves runtime guidance into the shared/native ACP engine path instead of the old standalone adapter. Refs #8881 — related `acpx_local` POSIX shell failure from the old `acpx` pin; this PR updates ACP dependencies but does not claim custom/OMP ACP support as a first-class native adapter. Refs #8964 — related open `acpx_local` stderr cleanup PR; this PR makes the old runtime path obsolete for new agents but keeps it as a non-closing reference. Problem description: - The standalone `acpx_local` adapter duplicates Claude/Codex agent choices that already have first-class local adapters. - ACP should be an execution engine capability of each harness adapter when the underlying harness supports ACP. - Existing `acpx_local` agents should either migrate to native harness adapters or fail with an explicit retirement message instead of silently falling back to the process adapter. ## What Changed - Added `@paperclipai/acpx-engine` as the shared ACP execution, session-codec, CLI formatter, and UI parser package. - Wired `claude_local`, `codex_local`, and `gemini_local` to auto-select ACP by default when prerequisites pass, with `engine=cli` opt-out and `engine=acp` strict mode. - Added ACP config schema/UI fields, environment checks, session-codec preservation, transcript parsing, and adapter capability metadata for the native adapters. - Retired `acpx_local` to a server tombstone, removed its UI/package/runtime image surface, and added a migration for existing Claude/Codex ACPX agents. - Updated package manifests, lockfile, release tooling, docs, Kubernetes sandbox defaults, and tests. ## Verification - `corepack pnpm --filter @paperclipai/acpx-engine typecheck` - `corepack pnpm --filter @paperclipai/adapter-claude-local typecheck` - `corepack pnpm --filter @paperclipai/adapter-codex-local typecheck` - `corepack pnpm --filter @paperclipai/adapter-gemini-local typecheck` - `corepack pnpm --filter @paperclipai/acpx-engine exec vitest run` - `corepack pnpm --filter @paperclipai/adapter-claude-local exec vitest run src/server/acp.test.ts src/server/execute.acp-fallback.test.ts src/ui/build-config.test.ts` - `corepack pnpm --filter @paperclipai/adapter-codex-local exec vitest run src/server/acp.test.ts src/ui/build-config.test.ts` - `corepack pnpm --filter @paperclipai/adapter-gemini-local exec vitest run src/server/acp.test.ts src/ui/build-config.test.ts src/ui/parse-stdout.test.ts` - `corepack pnpm --filter @paperclipai/plugin-sdk ensure-build-deps && corepack pnpm --filter @paperclipai/server exec tsc --noEmit` - `corepack pnpm --filter @paperclipai/server exec vitest run src/__tests__/adapter-routes.test.ts src/__tests__/adapter-session-codecs.test.ts src/__tests__/adapter-models.test.ts` - `corepack pnpm --filter @paperclipai/ui typecheck` - `corepack pnpm --filter @paperclipai/ui exec vitest run src/adapters/metadata.test.ts src/adapters/adapter-display-registry.test.ts src/components/AgentConfigForm.test.ts src/components/AgentConfigForm.render.test.tsx src/components/transcript/RunTranscriptView.test.tsx` - `node --test scripts/bootstrap-npm-package.test.mjs scripts/release-package-map.test.mjs scripts/verify-release-registry-state.test.mjs` Note: the server typecheck script calls `pnpm` internally; this dev shell exposes pnpm through Corepack only, so I ran the two script steps manually with `corepack pnpm`. ## Risks - Migration changes existing `acpx_local` Claude/Codex agents to native adapter types and clears old ACPX task sessions/runtime state. - Custom ACP commands remain on the retired tombstone and will need a separate future adapter/plugin path. - ACP auto-selection depends on local Node and ACP server command prerequisites; remote and unsupported environments fall back to CLI unless `engine=acp` is explicit. - `@paperclipai/acpx-engine` is a new public package and needs npm trusted-publishing bootstrap before release automation can publish it. > 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 OpenAI GPT-5 via Codex coding agent. Exact hosted model build and context-window size are not exposed in this runtime. Tool use included shell execution, repository editing, GitHub CLI operations, and local test/typecheck execution. ## 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 - [x] All Paperclip CI gates are green - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing> |
||
|---|---|---|
| .. | ||
| Dockerfile.base | ||
| Dockerfile.claude | ||
| Dockerfile.codex | ||
| Dockerfile.gemini | ||
| Dockerfile.hermes | ||
| Dockerfile.opencode | ||
| Dockerfile.pi | ||
| README.md | ||
| buildx-bake.hcl | ||
README.md
Agent Runtime Image Family
Container images for running coding-agent harnesses in sandboxed environments (for example the kubernetes sandbox provider, stage 1 of the k8s contribution). Images are named agent-runtime-{harness}:{version} and published to ghcr.io/paperclipai/ by the agent-runtime-images workflow. The registry is overridable: every reference flows through the REGISTRY bake variable.
Image Lineup
agent-runtime-base: Foundation. Ubuntu 22.04 + Node 22 + git + tini + non-root user (uid 1000) + the agent shim.agent-runtime-opencode: Extends base withopencode-aiglobally installed.agent-runtime-pi: Extends base with@mariozechner/pi-coding-agent.agent-runtime-codex: Extends base with@openai/codex.agent-runtime-gemini: Extends base with@google/gemini-cliplus headless auth-mode settings.agent-runtime-claude: Extends base with@anthropic-ai/claude-code(symlinked asclaude-code).agent-runtime-hermes: Dockerfile included in the bake group, not in the default publish scope (stub until a CLI package exists).
Base Image Contents
OS & Runtime:
- Ubuntu 22.04
- Node.js 22 (via NodeSource APT repo)
- git
- tini (PID-1 init, ensures signal propagation)
- Non-root user
paperclip(uid/gid 1000)
Paperclip Binaries:
/usr/local/bin/paperclip-agent-shim: Go binary compiled fromtools/agent-shim/. Reads/run/paperclip/runtime-command.jsonandsyscall.Execs the harness CLI.
Defaults:
USER: 1000:1000 (paperclip, non-root)WORKDIR:/workspace(mount workspace volumes here)ENTRYPOINT:/usr/bin/tini --(PID-1 reaper, forwards signals)CMD:/usr/local/bin/paperclip-agent-shim
Building Locally
All targets build linux/amd64 by default (see buildx-bake.hcl). Derived images chain off the base target through bake contexts, so the literal registry in each FROM line is overridden at build time and the whole family builds in one pass without pushing intermediates.
docker buildx bake -f docker/agent-runtime/buildx-bake.hcl --load
Custom tag or registry
REGISTRY=myregistry VERSION=mytag \
docker buildx bake -f docker/agent-runtime/buildx-bake.hcl --load
Quickstart Smoke Test
Build and verify the agent-runtime-claude image runs locally:
docker buildx bake -f docker/agent-runtime/buildx-bake.hcl base claude --load
docker run --rm ghcr.io/paperclipai/agent-runtime-claude:dev claude-code --version
Agent Container (paperclip-agent-shim)
The main agent process runs as the shim (PID 1 under tini). The shim:
- Reads
/run/paperclip/runtime-command.json(path overridable via-spec), a JSON file mounted by whatever schedules the run - Parses
{ "command", "args" }: the harness CLI and arguments - Resolves the command on PATH and
syscall.Execs it, replacing itself - SIGTERM from the kubelet propagates directly to the harness (no zombie processes)
runtime-command.json Contract:
{
"command": "claude-code",
"args": ["--token", "xyz", "--workspace", "/workspace"]
}
The shim makes no assumptions about command structure; it is harness-agnostic. New harnesses swap the command/args; the base image stays the same.
Security Model
- Non-root execution: user 1000:1000, no capability grants
- PSS Restricted compatible: no privileged containers, no host mounts; works with a read-only root filesystem (writable
/workspace+/tmpmounts) - No secrets baked in: API tokens and credentials come from per-run ephemeral Secrets mounted as env vars or files
- Image signing: cosign keyless OIDC in the publish workflow
Publishing
.github/workflows/agent-runtime-images.yml builds and pushes the default scope (base, opencode, pi, codex, gemini, claude) on workflow_dispatch (with an explicit version tag) or on pushes to master touching these paths, then signs each digest with cosign keyless OIDC.