Files
PaperClipAI/doc/architecture/paperclip-runner-compatibility.md
6c36c07a4f feat(adapters): add GPT-6.1 Sol and refresh shared coding harness pins (#14942)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - Agents run through coding-agent adapters and the native runner. Both
use the same installed provider CLIs, model catalogs, and reasoning
controls.
> - OpenAI released GPT-6.1 Sol (`gpt-6.1-sol`) in Codex. Anthropic
released Claude Sonnet 5.5. The static Codex, Bedrock, and OpenCode
catalogs do not list these IDs.
> - The shared provider pack pins Codex 0.156.0 and OpenCode 1.18.32.
The evaluation image pins older Grok, Gemini, Kimi, Cursor, and GitHub
CLI releases. Codex 0.156.0 has no bundled metadata for GPT-6.1 Sol.
> - A model entry without a current harness, or a harness pin without
its runner integrity checks, fails at run time.
> - This pull request adds the verified model IDs and moves the harness
pins, executable digests, controller checks, and image pins together.
> - The benefit is that operators can select the current models, and the
native and local adapters share one current CLI installation.

## Linked Issues or Issue Description

Refs #13829 and #13838 (the September 22, 2026 model and harness
refresh). Related pull requests: #14993 (merged October 5, 2026,
superseding #14816) added the direct Claude Sonnet 5.5 entry and
refreshed the Claude runtime to Agent SDK 0.3.286 / Claude Code 2.1.286.
This pull request does not change the Claude runtime or the direct
Claude model list; it keeps the #14993 pins and adds only the Bedrock
Sonnet 5.5 ID. After #14993 merged, this branch was rebased onto
`master` (October 5, 2026). The six overlapping pin regions
(`docker/daytona-runner/Dockerfile`, `docker/daytona-runner/README.md`,
`package.json`, `pnpm-workspace.yaml`,
`packages/adapters/claude-local/src/index.test.ts`,
`packages/paperclip-runner/src/backends/native-backend-factory.test.ts`)
were resolved by keeping this pull request's Codex 0.160.0 and OpenCode
1.18.34 pins next to #14993's Claude 0.3.286 / 2.1.286 pins, taking the
union of the Sonnet 5.5 model IDs in the Claude test, and merging both
README paragraphs. The Sonnet 5.5 effort and CLI-gate lines in the
Claude adapter were identical in both pull requests and merged without a
diff. #14917 and #14918 reordered the Claude and Codex model lists
earlier; the new entries sit where those ordering rules put them.

Sources checked on 2026-10-02:

- [OpenAI Codex models](https://learn.chatgpt.com/docs/models): GPT-6.1
Sol uses `gpt-6.1-sol`, supports reasoning efforts from Light to Ultra,
and has Standard and Fast modes at launch. The page also records that
`gpt-5.4` and `gpt-5.4-mini` retired from Codex with ChatGPT sign-in on
August 31, 2026, and that `gpt-5.5` retires on October 14, 2026. Neither
retirement applies to the OpenAI API.
- [Codex CLI releases](https://github.com/openai/codex/releases) 0.157.0
through 0.160.0. The bundled model metadata in the 0.160.0 Linux binary
contains `gpt-6.1-sol`.
- [Claude Sonnet
5.5](https://platform.claude.com/docs/en/models/sonnet-5-5/overview):
Bedrock ID `anthropic.claude-sonnet-5-5`, released September 28, 2026.
- [OpenCode releases](https://github.com/anomalyco/opencode/releases)
1.18.33 and 1.18.34 (fixes only). The OpenCode model registry lists both
added provider-qualified IDs.
- npm `latest` tags for `@xai-official/grok` 1.0.46,
`@google/gemini-cli` 0.62.0, and `@moonshot-ai/kimi-code` 2.1.1.
[xAI](https://docs.x.ai/docs/models),
[Google](https://ai.google.dev/gemini-api/docs/models), and
[Kimi](https://www.kimi.com/code/docs/en/kimi-code/models.html) list no
newer coding models.
- Cursor CLI 2026.10.01-e373342 is the version the official installer
resolves. The pinned digest is the SHA-256 of the versioned Linux x64
archive.
- [GitHub CLI 2.102.0](https://github.com/cli/cli/releases/tag/v2.102.0)
(security fixes). The pinned digest matches the release `checksums.txt`.

## What Changed

- Codex adapter: add `gpt-6.1-sol` to the model list, the Fast mode
list, and the Ultra effort set. It is the first entry: #14918 orders the
list newest version first, and its description notes the ChatGPT app
lists GPT-6.1 Sol first. Update the adapter documentation text.
- Claude adapter: add `us.anthropic.claude-sonnet-5-5` (Bedrock Sonnet
5.5) to the Bedrock catalog in the newest-Sonnet slot after Opus 5.5;
`us.anthropic.claude-sonnet-5` moves into the older-Sonnet group,
matching what `sortClaudeModels` from #14917 produces at runtime. Any
Sonnet 5.5 ID (direct or Bedrock-qualified) now gets the documented
`xhigh` and `max` efforts and requires Claude Code 2.1.284 or later on
the CLI lane (the Claude Code changelog entry for 2.1.284 adds
`claude-sonnet-5-5`). These two lines are identical to the ones #14993
merged, so the branch carries no diff for them.
- OpenCode adapter: add `openai/gpt-6.1-sol` and
`anthropic/claude-sonnet-5-5` to the static fallback catalog.
- Codex runtime pin 0.156.0 → 0.160.0 in the root and workspace
overrides, the runner package, the Codex ACP package patch, the
qualified ACPX profiles, the Linux x64 executable digest, the Rust
provider backend and its tests, the provider-pack manifest pins, the
remote controller pins, the sandbox npm install spec, and the opt-in
qualification scripts.
- Remote Codex compatibility window: upper bound 0.157.0 → 0.161.0. The
minimum stays at 0.149.0.
- OpenCode runtime pin 1.18.32 → 1.18.34 in the runner package, the
materialization script, the server and Rust qualified versions, the eval
and live-session labels, fixtures, and the configuration label.
- Evaluation image (`docker/daytona-runner/Dockerfile`): Grok CLI
1.0.46, Gemini CLI 0.62.0, Kimi Code 2.1.1, Cursor CLI
2026.10.01-e373342 with its digest, GitHub CLI 2.102.0 with its digest,
Codex and OpenCode version probes, and the refreshed lockfile digest.
The Claude Code 2.1.286 probe comes from #14993 and is unchanged here.
- `pnpm-lock.yaml` is not part of this pull request. The repository's
pull request gate rejects lockfile edits, and the refresh bot
regenerates the lockfile on master (the same flow #13838 used). The
Dockerfile `PAPERCLIP_RUNNER_LOCK_SHA256` default is the digest of the
lockfile that `pnpm install --resolution-only --ignore-scripts
--no-frozen-lockfile` (the refresh workflow's command) produces for the
combined pins on the rebased branch (`e1856797…`); that lockfile differs
from master only in the `@openai/codex` 0.160.0 platform packages, the
`@anthropic-ai/claude-agent-sdk` 0.3.286 override that #14993 introduced
(the open refresh-bot pull request #14872 carries that part),
`opencode-ai` 1.18.34 with its Linux x64 baseline, and the `codex-acp`
patch hash.
- Documentation: runner README, runner compatibility doc, environment
variable example, and a new `doc/adapter-model-audit-2026-10-02.md` with
sources and deferred items.
- Tests: Codex adapter catalog, server adapter models, Codex
compatibility window, native session executor pins, runner package
contract, OpenCode materialization, and UI effort options.

Unchanged on purpose: Claude Agent SDK 0.3.286 / Claude Code 2.1.286
(already on `master` from #14993), ACP bridges (`acpx` 0.13.1,
`claude-agent-acp` 0.73.0, `codex-acp` 1.6.2; newer upstream releases
need a separate qualification), the native Grok runtime 1.0.13, Pi
0.84.2 / 0.87.1 (the Pi 1.0 runner stack covers it), and Hermes 0.19.0
(current). `gpt-5.4` and `gpt-5.4-mini` stay in the picker because the
OpenAI API still serves them.

## Verification

Run on Linux x64 with Node 25.9.0 and pnpm 9.15.4 after `pnpm install
--no-frozen-lockfile` (the refreshed lockfile stays local; see above).
The results below were re-run on the rebased head (October 5, 2026) for
the suites the conflict resolution touches; the other rows are from the
original run and are covered by CI on every push:

- Rebased head: `packages/adapters/codex-local` 482 passed;
`packages/adapters/claude-local` 340 passed, 4 failed (`execute.remote`,
`test.probe`, `execute.acp-fallback`, `acp` spawn/env-hardening cases
that fail identically on unchanged `master` in this host environment);
`server` adapter-models + codex-runtime-compatibility +
native-session-executor + adapter-registry 607 passed, 1 failed (the
same adapter-registry override-pause case as before, also failing on
`master` here); `packages/paperclip-runner` native-backend-factory +
qualified-profiles 36 passed; `ui` codex-reasoning-effort +
config-fields + model-utils 19 passed. Rust, full typecheck, build, and
the Docker image are left to CI as before.

- `vitest run` in `packages/adapters/codex-local`: 13 passed. `vitest
run` in `packages/adapters/claude-local` (whole package, including the
new Sonnet 5.5 gate and effort tests): see the latest CI run and the
comment below. `vitest run` in `packages/adapters/opencode-local`: 48
passed, 1 failed (`runtime-config.test.ts` reads the host
`PAPERCLIP_OPENCODE_PROVIDERS` variable; it fails the same way on the
unchanged base).
- `vitest run src/__tests__/adapter-models.test.ts
src/services/native-runtime/codex-runtime-compatibility.test.ts
src/__tests__/adapter-registry.test.ts` in `server`: 84 passed, 1 failed
(`adapter-registry.test.ts` override pause test; it fails the same way
on the unchanged base).
- `vitest run` in `ui` for `codex-reasoning-effort`,
`agent-setup-fields`, `config-fields`, and `ComposerRunSettingsPicker`:
25 passed.
- `node --test test/acpx-codex-package-contract.test.mjs
scripts/materialize-opencode-binary.test.mjs
scripts/runner-protocol-eval-campaign.test.mjs` in
`packages/paperclip-runner`: 23 passed. The package contract test
verifies the installed Codex ACP executable digest and the 0.160.0 patch
pin.
- `vitest run src/drivers/acpx src/backends src/drivers/opencode
src/live/live-session.test.ts` in `packages/paperclip-runner`: 626
passed, 5 failed, 1 skipped. The 5 failures
(`installation-integrity.test.ts` `/proc/self/fd` module loading and one
OpenCode answer-selection test) also fail on the unchanged base under
Node 25; Linux CI runs Node 24.
- `pnpm run test:opencode:qualification` in `packages/paperclip-runner`
against the installed OpenCode 1.18.34 executable: passed.
- `codex --version` from the installed pack prints `codex-cli 0.160.0`.
The Linux x64 executable digest `12eb3e81…652aad` was computed from the
`@openai/codex@0.160.0-linux-x64` archive after checking its registry
`dist.integrity`.
- `pnpm check:token-gates`: all gates clean.
- `pnpm run typecheck:typescript` in `packages/paperclip-runner`:
passed. Package typechecks ran one at a time; see the comment below for
the server and UI results.

Not run here, and needed from CI:

- Rust tests and `pnpm -r typecheck` / `pnpm build` for the server (no
`cargo` in this environment; the server typecheck prepares the runner
vendor build).
- The Docker evaluation image build and the real-binary Codex startup
and session-resume probes (no Docker; the probes need the compiled
`paperclip-runnerd`). The trusted CI runner workflow covers them.
- Authenticated inference with any new model. This change is metadata
and startup validation only.

## Risks

- Codex 0.160.0 changes the bundled model catalog and app-server
behaviour (authoritative provider catalogs, incremental running-turn
tracking). The patched `codex-acp` 1.6.2 bridge is unchanged and
declares `^0.148.0`; it worked with 0.156.0 under the same override. If
CI probes show a protocol change, the pin can return to 0.156.0 by
reverting this pull request.
- The compatibility window upper bound moves to `<0.161.0`. Remote
images with Codex 0.157 to 0.160 become accepted. Older images stay
accepted down to 0.149.0.
- Until the refresh bot lands the regenerated lockfile on master, the
Dockerfile lockfile digest default does not match the committed
lockfile. The trusted CI workflow computes the digest from its own
resolution at build time, so this affects only a local build that passes
no digest.
- Existing saved model selections and effort settings are not changed.
Agents on `gpt-5.4` or `gpt-5.5` with ChatGPT sign-in need a model
change before the OpenAI retirement dates; that is documented, not
enforced.
- Rollout order: deploy the controller and runner from this change
before promoting a sandbox image that carries these pins. Older
controllers reject the new provider-pack pins.

## Model Used

- Claude Fable 5.1 (Anthropic, model ID `claude-fable-5-1`), 1M context
window, adaptive thinking, tool use. The model ran as a Paperclip agent
through the Claude Code harness, performed the web research, edited the
code, and ran the tests listed above.

## 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: Bender (Fable) <noreply@paperclip.ing>
Co-authored-by: devinfoley <139239+devinfoley@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-05 13:59:08 -07:00

12 KiB

Paperclip Runner compatibility and rollout

Purpose

This document defines compatibility rules for introducing the experimental Paperclip Runner. These rules are acceptance criteria for each implementation change. They are not a migration plan for existing adapters.

Compatibility invariants

  1. Existing adapter selection is authoritative. A direct adapter stays direct.
  2. A non-runner run must not start runnerd or open PRP.
  3. A non-runner run must not create native result, finalization, status-decision, or runner-transport records.
  4. A non-runner run must not invoke native status arbitration.
  5. Direct adapters keep their current transcript, interaction, cancellation, result, and finalization behavior.
  6. Runner-only UI controls depend on persisted runtime facts and are absent from direct runs.
  7. The rollout flag controls fresh runner selection and fresh runner starts.
  8. A flag change does not rewrite an agent profile or a persisted run choice.
  9. Persisted native data remains readable after the flag is disabled.
  10. Recovery may finish an already persisted native run while fresh native starts remain blocked.

Runtime selection

The server resolves and persists the runtime once, before provider launch.

Persisted runtime Adapter Flag Result
none Any direct adapter off or on Use the existing direct path.
none paperclip_runner with any qualified provider off Reject the fresh start with a stable rollout-disabled error.
none paperclip_runner with a qualified provider on Use PRP v1 and the provider's persisted runnerd backend.
none paperclip_runner with an incomplete or unqualified profile on Reject the profile before runnerd starts.
direct Any changed later Keep the persisted direct path.
native Any changed later Keep the persisted native path for read, cancel, recovery, and finalization.

The server must not fall back from a selected paperclip_runner start to codex_local. A configuration or rollout error must be visible. Silent fallback would hide the runtime that executed the task.

Remote Codex version window

Remote native Codex runs accept stable CLI versions >=0.149.0 <0.161.0. The install pin is 0.160.0. A sandbox image can therefore use an older compatible Codex without failing startup or installing another copy.

The minimum is fixed at 0.149.0 until maintainers deliberately change it. It is not a rolling one-month support window. The official @openai/codex publication history records 0.149.0 on 2026-08-20, 0.153.4 on 2026-09-04, 0.156.0 on 2026-09-22, and 0.160.0 on 2026-10-01. See also the Codex changelog. This is an explicit compatibility window, not a claim that every Codex minor release follows a backward-compatible protocol.

  • Stable patch releases within the window are accepted. Alpha, dev, custom builds, malformed output, older releases, and the next minor release are rejected.
  • A compatible difference from the install pin produces one diagnostic per version during backend preparation, rather than stopping the task.
  • The same check applies to discovered, explicitly staged, and npm-installed remote Codex executables, including the final executable after linking.
  • For an incompatible image, the existing PAPERCLIP_RUNNER_REMOTE_CODEX_NPM_SPEC=@openai/codex@0.160.0 configuration allows installation of the pinned runtime. Without that configuration, startup reports the supported range and the remediation.
  • Runner binary contracts, required runner capabilities, artifact digests, provider-pack manifests, and runtime permission checks are unchanged. This window does not relax the separate ACP provider-pack qualification.

When qualifying newer Codex releases, review the upper bound in server/src/services/native-runtime/codex-runtime-compatibility.ts. Updating the install pin or upper bound does not raise the minimum. Raising 0.149.0 requires a separate, explicit maintainer decision and compatibility evidence. Do not derive eligibility from the current date or fetch release metadata during startup: an idle installation must not become incompatible merely because time passed.

Qualification on 2026-09-22 used the actual macOS ARM64 0.149.0 and 0.156.0 app-server binaries with a local deterministic Responses API fixture. Both passed initialization, thread creation with a dynamic tool, a tool-call/result round trip, turn completion, and thread resume. This verifies the tested protocol path; it is not a live model, Linux sandbox, or exhaustive feature test. The 2026-10-02 refresh moved the install pin and the upper bound to 0.160.0 after verifying the integrity-checked Linux x64 0.160.0 executable digest and the installed ACP package contract. Codex 0.157.0 through 0.160.0 add the GPT-6.1 Sol catalog entry and app-server pagination without a documented app-server protocol break; the Linux CI qualification workflow covers the real startup and session-resume probes for the new pin. Version-boundary and remote artifact-preparation regressions are covered by codex-runtime-compatibility.test.ts and native-session-executor.test.ts.

Direct adapter boundary

This rule applies to every built-in and plugin direct adapter. It includes:

  • codex_local;
  • claude_local;
  • opencode_local;
  • other local CLI or session adapters;
  • process and HTTP adapters;
  • gateway adapters; and
  • external adapter plugins.

Adding Paperclip Runner must not add runner imports or runner branches inside a direct adapter implementation. The heartbeat coordinator may select the explicit runner adapter at one narrow seam. All other adapters continue through their existing code.

For a flag-off codex_local heartbeat, compatibility proof must show:

  • one direct invocation;
  • the same normalized result and finalization bytes as the approved baseline;
  • zero runner processes;
  • zero PRP connections; and
  • zero native rows.

Configuration behavior

When the rollout flag is off:

  • creation UI does not offer paperclip_runner;
  • edit UI does not offer switching to paperclip_runner;
  • server creation and import reject a new paperclip_runner selection;
  • the server rejects a fresh start for an existing runner-configured agent;
  • read and export preserve an existing runner configuration;
  • unrelated edits to an existing runner-configured agent do not erase its configuration; and
  • switching that agent to a direct adapter remains allowed.

When the rollout flag is on:

  • creation, import, and edit accept paperclip_runner only with a qualified Codex, OpenCode, Claude Managed, AWS AgentCore, or Claude/Codex ACPX profile;
  • switching from a direct adapter affects only future unresolved runs; and
  • switching away from the runner affects only future unresolved runs.

Server validation is the authority in both states. Import files and API clients cannot bypass the flag or provider allowlist.

Persisted native runs

The following data remains readable independent of the current flag:

  • the persisted runtime selection and reason;
  • run, runner, session, turn, and provider identity;
  • ordered runner and provider events;
  • accepted result and evidence assessment;
  • finalization coordinator and status decision;
  • usage, cost, cancellation, and terminal details; and
  • durable interactions and final task-thread reply.

If the flag is disabled during an in-flight native run, the server may reconnect, cancel, reconcile, and finalize that same run. It must not use that recovery as authority to start a new native run.

Recovery must remain idempotent. Repeating it cannot add a second final reply, interaction, wake, status decision, or application effect.

Task-page compatibility

The task page uses one provider-neutral thread projection. Runtime facts may add runner event groups, semantic receipts, usage, and structured questions. They must not replace classic direct-adapter content.

Direct-adapter coverage must include:

  • an active run;
  • a settled run;
  • an empty transcript;
  • a pending interaction;
  • a resolved interaction; and
  • the classic interface state.

The existing composer remains usable for direct adapters. A direct run does not show reconnect, runner cancel, semantic receipt, or other runner-only controls. Final replies continue to use the existing issue-comment behavior.

Structured input compatibility

New structured questions use paperclip.question_set.v1 and the matching response contract. Provider-specific question objects remain inside their drivers.

Legacy unstructured interaction records remain readable and resolvable. A structured form fails closed when its required schema, question mode, question ID, option ID, or response value is invalid. The implementation must not silently convert malformed structured input to a legacy text prompt.

Protocol version compatibility

PRP wire versions, fixture versions, event schema versions, and typed schema discriminators are independently versioned.

  • Peers negotiate the highest common PRP version.
  • No common required version fails closed before command or provider execution.
  • An unknown required fixture or event schema version fails closed.
  • An unknown required typed discriminator fails closed.
  • Additive optional properties remain compatible only when an old v1 consumer can ignore them without changing behavior.
  • Unknown object properties must survive validation when the owning schema permits additive fields.

Breaking meaning requires a new required version. A provider error is not a protocol negotiation result.

Semantic action compatibility

Catalog generation and production authorization are separate steps.

  • A catalog entry does not authorize production use.
  • An operation without a production binding is undiscoverable.
  • An operation denied to the run is undiscoverable unless the protocol explicitly defines a safe denied receipt for that discovery mode.
  • A duplicate call with the same idempotency key and canonical input returns the original safe receipt.
  • Reusing the key with different input returns a conflict and performs no second effect.
  • Receipts redact credentials, private provider payloads, and hidden identity.

Required compatibility matrix

Each runner-related pull request updates only rows that it can execute. The complete first-wave matrix must cover:

Area Required cases
Runtime selection Every built-in direct adapter, explicit runner selection, unsupported provider, flag on, and flag off.
Direct regression Flag-off codex_local invocation count, byte-stable result/finalization, and zero native rows.
Configuration Enabled and disabled create, import, edit, read, export, and adapter switch.
Recovery Persisted native run after flag disable, reconnect, duplicate event, duplicate command, cancellation, and server restart.
Protocol security Cross-company binding, ticket replay, ticket expiry, malformed frame, unsupported version, and revoked lease.
Semantic actions Discovery denial, unbound action, duplicate call, conflicting retry, redaction, and governed action.
Task page Active, settled, empty transcript, interaction, and classic direct-adapter states.
Structured input Valid response, malformed response, stale response, duplicate response, provider loss, and cancellation.

Pull request acceptance

Every implementation pull request must:

  • build and work against its declared base;
  • keep its changed-file count below 100;
  • update manifests and exports only for implemented surfaces;
  • run the smallest relevant tests before the repository handoff gate;
  • run typecheck, tests, and build for handoff;
  • run protocol parity, Cargo, clean-consumer, migration, token, and browser gates when those areas change;
  • contain no unexplained failed, cancelled, or path-skipped verification; and
  • resolve actionable review and security findings before it is ready.

Generated files land with their source and a drift check. pnpm-lock.yaml is owned by CI and is not part of these pull requests.