Files
PaperClipAI/doc/agent-observability.md
T
170e1b0c07 feat(runner): pin Codex 0.170.0 and Agent SDK 0.3.328 so Fable 5.2 and GPT-6.2 run (#22278)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - Paperclip Runner runs Claude and Codex agents from one shared
provider pack. The pack pins an exact Claude Agent SDK (which bundles
Claude Code) and an exact Codex CLI.
> - #22222 added Claude Fable 5.2 (minimum Claude Code 2.1.321) and
GPT-6.2 Sol and Astra (minimum Codex 0.169.1). The pack still bundles
Claude Code 2.1.286 and Codex 0.160.0.
> - Because of this, Paperclip Runner stops these models with a named
version error, and native remote Codex accepts only `>=0.149.0
<0.161.0`.
> - This pull request moves the provider pack to Agent SDK 0.3.328
(Claude Code 2.1.328) and Codex 0.170.0. It moves every constant that
depends on these pins, and the remote Codex window to `<0.171.0`.
> - The benefit is that Fable 5.2 and GPT-6.2 run on Paperclip Runner
after a sandbox image with this pack is promoted.

## Linked Issues or Issue Description

Refs #22222 (adds the models and the version floors; this pull request
makes them runnable on Paperclip Runner).
Refs #18488 (moves only the Codex pin to 0.166.0 and the window to
`<0.167.0`; it is not updated since November 1 and has conflicts with
`master`).
Refs #18486 (GPT-6.1 Terra needs Codex 0.165.0; Codex 0.170.0 satisfies
it).

**Overlap with #18488.** #18488 changes the same Codex pin locations.
This pull request uses a newer pin (0.170.0) and also moves the Claude
runtime. If this pull request merges first, #18488 has no remaining
change. I did not change the #18488 branch.

## What Changed

- Claude runtime: `@anthropic-ai/claude-agent-sdk` 0.3.286 → 0.3.328
(Claude Code 2.1.286 → 2.1.328). This changes the root `pnpm.overrides`,
`pnpm-workspace.yaml`, the ACPX `claude` profile, and the qualified
Claude runtime in `installation-integrity.ts`, with new executable
digests for Linux x64, Linux ARM64, macOS ARM64, and macOS x64.
- Codex runtime: `@openai/codex` 0.160.0 → 0.170.0. This changes the
runner dependency, the `codex-acp@1.6.2>@openai/codex` override, the
`codex-acp` patch dependency line, the ACPX `codex` profile, and the
qualified Linux x64 runtime with its new executable digest.
- The ACP bridges do not change: `claude-agent-acp` 0.73.0 and the
patched `codex-acp` 1.6.2. Their `commandDigest` values and profile
versions do not change.
- Generated declarations: `generated-profiles.ts` and
`generated_acpx_profiles.rs` are regenerated with
`generate:acpx-profiles`.
- Remote native Codex window: `REMOTE_CODEX_MAXIMUM_VERSION_EXCLUSIVE`
0.161.0 → 0.171.0. The minimum stays at 0.149.0. Images with Codex
0.149.0 through 0.170.x are accepted.
- Codex company observability: the qualified-version check moves to
`codex-cli 0.170.0`. The `otel` configuration keys that Paperclip writes
(`exporter`, `trace_exporter`, `metrics_exporter`) are present in the
0.170.0 configuration schema. The `OtelExporterKind`,
`OtelHttpProtocol`, and `OtelTlsConfig` definitions are identical to
0.160.0.
- Daytona runner image probes (`Dockerfile`, `Dockerfile.release`)
expect `codex-cli 0.170.0` and `2.1.328 (Claude Code)`. The remote npm
install spec in `server/src/adapters/registry.ts` is
`@openai/codex@0.170.0`.
- Tests: fixtures and assertions that mirror the pins use the new
versions. The window tests accept 0.161.0 through 0.170.99 and reject
0.171.0. The ACPX gate tests now show that Fable 5.2 and GPT-6.2 are
admitted. A new test mocks an older bundled runtime and keeps the named
version-gap errors covered. The fallback test shows that GPT-6.2 stays
on 0.170.0 and still falls back on 0.160.0.
- Opt-in qualification scripts
(`scripts/tests/native-cleanup-paginated-codex.mjs`,
`native-provider-startup-codex.mjs`) require Codex 0.170.0.
- Documentation: `docker/daytona-runner/README.md`,
`doc/architecture/paperclip-runner-compatibility.md`,
`doc/agent-observability.md`, `docs/deploy/environment-variables.md`,
and the deferred item in `doc/adapter-model-audit-2026-11-18.md`.
- `pnpm-lock.yaml` is not part of this pull request. The refresh bot
regenerates it on `master`, as for the previous pin refreshes (#14942,
#14993).

## Verification

Sources, checked on 2026-11-19:

- npm registry: `@openai/codex@0.170.0` (published 2026-11-18T03:45Z)
and `@anthropic-ai/claude-agent-sdk@0.3.328` (published
2026-11-18T23:13Z, `claudeCodeVersion` 2.1.328). Both are the `latest`
tags.
- [Codex releases](https://github.com/openai/codex/releases) 0.161.0
through 0.170.0. The notes record no app-server protocol break. 0.169.0
keeps project configuration disabled for `codex exec` in untrusted
directories; Paperclip's CLI path already uses `--skip-git-repo-check`,
and the native path uses `app-server`.
- [Claude Code
changelog](https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md)
2.1.287 through 2.1.328. See "Risks" for the headless changes that I
reviewed.

Digests and package shape:

- I downloaded each platform archive with `npm pack`, which verifies
registry integrity, and computed the SHA-256 of the executable. As a
method check, the same command on the 0.3.286 Linux x64 archive gives
the digest that is on `master` (`fe503f65…`). The Codex 0.160.0 archive
also gives the `master` digest (`12eb3e81…`).
- The extracted executables report `codex-cli 0.170.0` and `2.1.328
(Claude Code)`.
- The Codex 0.170.0 Linux x64 archive no longer contains
`codex-resources/zsh/bin/zsh`. Paperclip does not reference that path.
- The Agent SDK 0.3.328 `optionalDependencies` have the same platform
set as 0.3.286. The Codex 0.170.0 `optionalDependencies` use the same
`npm:@openai/codex@<version>-<platform>` aliases.
- Bridge link check: I installed
`@agentclientprotocol/claude-agent-acp@0.73.0` with an override to Agent
SDK 0.3.328 and imported its entry module. Node linked all of the named
SDK imports, and the import completed.

Real Codex qualification (no model turn, no credentials):

- `PAPERCLIP_TEST_CODEX_BINARY=<installed 0.170.0 Linux x64 binary> node
--import ./server/node_modules/tsx/dist/loader.mjs
scripts/tests/native-cleanup-paginated-codex.mjs` passed: `green: true`,
with the expected red control and paginated `thread/resume` through the
real 0.170.0 `app-server`. The installed binary digest matches the
pinned digest.

Focused suites (Linux x64, Node 25.9.0, local `pnpm install
--no-frozen-lockfile`):

- `cd packages/paperclip-runner && pnpm run typecheck:typescript`
passed, including `check:acpx-profiles`.
- `npx tsc --noEmit` passed in `packages/adapters/codex-local` and
`packages/adapters/claude-local`.
- `packages/adapters/codex-local`: 625 passed.
`packages/adapters/claude-local`: 419 passed, 1 skipped.
- `node --test`: `acpx-codex-package-contract` 9 passed,
`controlled-evals/controller` 77 passed, `scripts/acpx-patch-packaging`
26 passed, `scripts/__tests__/grok-public-install-sandbox` 27 passed.
- `server`: `acpx-qualification`, `codex-runtime-compatibility`, and
`codex-model-fallback` 110 passed. `native-session-executor` 1541
passed, 1 failed. The failure (`compares export process compatibility …
db.transaction is not a function`) also fails on unchanged `master` on
this host.
- `packages/paperclip-runner` vitest (`installation-integrity`,
`qualified-profiles`, `codex-security-config`, `native-execution`,
`native-backend-factory`): 283 passed, 25 failed. All 25 failures are in
`installation-integrity.test.ts` with "ACPX provider lifetime guardian
failed to start". The same 25 tests fail on unchanged `master` on this
host (master fails 30 here).

Rust (toolchain 1.97.1 from `rust-toolchain.toml`, Linux x64):

- `cargo test -p paperclip-runner-core --lib`: 658 passed, 2 failed, 4
ignored. `warm_run_attachment_reuses_the_provider_ingress_listener` also
fails on unchanged `master` on this host.
`reaping_stops_a_dead_runners_provider_and_its_escaped_tool_job` is
flaky here: on this branch it passed 2 of 3 reruns.
- `cargo test -p paperclip-runner-core --test acpx_warm_attachment
--test native_provider_backend`: 7 and 45 passed.
- The new and changed saved-session tests pass:
`a_session_saved_by_the_previous_agent_runtime_is_refused_by_name`,
`unrestorable_saved_sessions_keep_strict_bindings_and_name_their_class`,
and `a_settled_session_saved_by_the_previous_acpx_release_continues`.
`rustfmt --check` passes.
- `server`: `native-session-resume.test.ts` and
`acpx-qualification.test.ts` 110 passed, 2 skipped. Both new resume
cases run and pass.

Not run here:

- `native-provider-startup-codex.mjs`, because it needs a compiled
`paperclip-runnerd`. The trusted CI runner workflow covers startup and
session-resume probes.
- The Docker image build and any live model turn. A live Fable 5.2 and
GPT-6.2 test follows the staging image rollout.

## Saved sessions from the previous release

This change keeps the ACPX bridge command digests and profile versions.
It changes `agentRuntimeVersion` in the `claude` and `codex` ACPX
profiles. That field is part of the saved session identity. The previous
release saved sessions with Agent SDK 0.3.286 or Codex 0.160.0. This
build refuses those saved sessions. It does not migrate them.

- **Next run (the usual case):** the saved provider descriptor names the
old runtime. `rebindNativeSessionCheckpoint` does not match it, so the
next run starts a fresh provider session in the same workspace with the
complete handoff prompt. No run fails. The earlier provider conversation
does not carry over.
- **A run in flight at the deploy (settled or mid-turn):** the runner
restore refuses the saved descriptor with the permanent class
`provider_state_incompatible`. The server reports it as
`native_provider_state_incompatible` and stops automatic retries. The
runner opens no provider and keeps the saved state bytes and the
workspace. The next run on the task starts a fresh provider session.
- **Native Codex (`codex_app_server`) and OpenCode** sessions do not
record this field, so their resume decision does not change. After a new
image is promoted, native Codex resumes a saved thread with Codex
0.170.0 instead of 0.160.0. This pull request does not test that
cross-version thread resume.

New tests:
- `acpx_provider_backend.rs`:
`a_session_saved_by_the_previous_agent_runtime_is_refused_by_name`
covers settled Claude and Codex sessions and a Claude turn interrupted
mid-run by the deploy.
`unrestorable_saved_sessions_keep_strict_bindings_and_name_their_class`
has a new `agent_runtime_release` case. Each case checks the named
class, that no provider session opens, and that the saved bytes do not
change.
- `native-session-resume.test.ts`: "a session saved by the previous
Claude or Codex ACPX runtime" checks that a same-release session
resumes, and that a session saved with Agent SDK 0.3.286 or Codex
0.160.0 starts fresh in the same workspace.

## Risks

- Medium. Every Paperclip Runner Claude and Codex run uses the new
executables. The digests, generated profiles, image probes, and window
must agree. They agree locally, and CI runs the Rust and runner suites
again.
- Claude Code 2.1.287 through 2.1.328 change some headless defaults. I
checked each one against Paperclip:
- 2.1.302: `-p` and Agent SDK sessions with no permission mode start in
`default` mode on third-party providers. Paperclip always sets a mode
(`bypassPermissions` or `default`), so this does not apply.
- 2.1.299: MCP allow rules now match only the whole server name.
Paperclip already writes `mcp__<server>__*` and exact
`mcp__paperclip__<tool>` rules.
- 2.1.302: fork subagents are on by default in `-p` and the Agent SDK,
as in the terminal. This pull request keeps the upstream default and
does not set `CLAUDE_CODE_FORK_SUBAGENT=0`. Reviewers can ask to keep
them off.
- 2.1.299 and later: stdio MCP connections negotiate MCP protocol
2026-07-28 first and fall back to the older protocol for servers that do
not answer.
- Agent SDK 0.3.328 is less than one day old. If a regression appears,
revert this pull request to return to 0.3.286 and Codex 0.160.0.
- Sandbox images built before this change still carry Codex 0.160.0 and
Claude Code 2.1.286. They stay accepted. On those images, Fable 5.2 and
GPT-6.2 keep the existing named errors and fallbacks until a new image
is promoted.
- Until the refresh bot updates `pnpm-lock.yaml` on `master`, a local
`--frozen-lockfile` install fails, as with the previous pin refreshes.

> 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

- Anthropic Claude, model ID `claude-opus-5-5`, through Claude Code with
tool use (shell, file edit, `gh`) and code execution. The context window
and reasoning mode are not reported in this session.
- The registry archives, digests, release notes, configuration schemas,
and changelog were read directly. No model output was used as a source
for those facts.

## 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
- [ ] 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: devinfoley <139239+devinfoley@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-18 20:45:18 -08:00

10 KiB

Company agent observability

Company Settings → General → Agent observability configures one company-owned OTLP destination for Paperclip-managed Codex, Claude Code and native OpenCode executions. Export is disabled by default. Choose Braintrust (US) or a custom HTTP(S) OTLP base URL; use your own collector when you need fan-out. Braintrust EU and self-hosted installations can use Custom OTLP with their required routing/authentication headers. The receiver must be reachable from every execution environment; localhost refers to the execution host, including a remote sandbox.

This is separate from Paperclip's first-party Telemetry and its content-free infrastructure tracing. Configuring it does not change either data path.

Capture and supported runtimes

The qualified versions are Codex 0.170.0 and Claude Code through ACP bridge 0.73.0 with Agent SDK 0.3.328. Both logs and traces use HTTP/protobuf. Prompts, assistant responses and tool content default on. Additional tool details are a Claude Code and native OpenCode control; Codex has no equivalent switch. Codex requests up to 64 KiB of tool output (or zero when disabled); native event availability, truncation and response caps still apply. Claude Code applies its own native limits. Metrics, raw API bodies and Codex Guardian assessments are disabled.

Native OpenCode 1.18.34 uses a runner-owned HTTP/JSON exporter. Each completed, failed or cancelled turn produces a trace with Paperclip company, agent, run and turn IDs. Authenticated Paperclip tool dispatch outcomes, native tool outcomes and message-level token/cost receipts are child spans; repeated usage updates replace the same message receipt. Missing prices remain unknown. This projection excludes hidden reasoning, raw provider payloads and system instructions. Prompt, response, tool-detail and tool-content switches apply before export. Content is bounded to 8 KiB per field and 64 tools plus 64 usage records per turn; truncation is marked. With both tool-detail and tool-content capture enabled, Paperclip tool spans include braintrust.input diagnostics (paperclip.opencode.tool-diagnostics.v1): the upstream tools/call method, argument JSON type and types for a fixed list of envelope fields. These are shape diagnostics, not a copy of the arguments. Arbitrary field names, argument values, nested payloads and full model request bodies are not captured. Authenticated schema rejections also produce failed tool spans without invoking the controller handler. Conflicting request IDs produce separate failed items, so a rejection cannot overwrite a previous dispatch outcome. Paperclip tool spans use authenticated dispatch IDs and the name tool <name>. Provider SSE records for those declared tools use opencode tool <name> and remain available when failure occurs before dispatch, such as an oversized request. These are separate observations, not correlated calls. The paperclip.tool.source attribute distinguishes dispatch and provider records. Native tools retain their provider spans without submitted-argument diagnostics. OpenCode's performance-only OTLP sink stays separate.

For known native task envelopes, OpenCode selects the task identifier, title and prompt before the 8 KiB limit. A valid leading fresh-session handoff packet is omitted from this excerpt so earlier history cannot hide the current request. Continuation excerpts put pending user messages before recovery guidance. Plain text and unknown envelope formats retain their existing text projection. History omission also sets the turn's truncation flag; the flag can also mean a text or row limit was reached.

When prompt capture is enabled, paperclip.observability.input.events gives a relative URL on the originating Paperclip instance: /api/heartbeat-runs/<run-id>/events. Sign in to that instance, read pages using afterSeq and limit, and find the turn.submitted event for the trace's paperclip.turn.id. Its payload.prpEvent.payload.text holds the retained submission, including history, subject to existing redaction and retention. The URL does not grant access. Run telemetry permission, company membership and task privacy checks still apply. The reference is absent when prompt capture is disabled. Tool argument values remain excluded from company traces.

The exporter sends once at turn completion with a three-second deadline, at most four requests in flight and no idle polling or persistent queue. Closing drains pending requests and marks an unfinished turn failed. Export failures emit a sanitized diagnostic and do not change task outcomes. Abrupt process termination can lose an unexported turn. Recovered in-flight turns without a new submission are not replayed just to produce an observability record.

The native runner and legacy local adapters receive transient launch material. Native Codex writes a private managed config.toml; legacy exec uses a unique invocation profile, preserving a selected source profile. Its profile injection requires the qualified executable version. Codex ACPX writes its isolated home. The legacy ACPX adapter keeps export configuration in a private home per agent and settings compatibility key, and shares conversation files with its stable managed home; export changes do not change its logical conversation fingerprint. Codex receives authentication through OTEL_EXPORTER_OTLP_LOGS_HEADERS and OTEL_EXPORTER_OTLP_TRACES_HEADERS in its process environment. Its managed configuration and invocation profiles contain empty header maps, so exporting does not save the ingestion credential in those files. The bundled Rust OTLP exporter percent-decodes environment header values; launch material encodes them once to preserve literal percent sequences. Credentials remain literal values in the company secret; config-file environment interpolation is not used. A caller-owned/preconfigured Codex home cannot be automatically configured in this version: its files remain untouched and execution emits a diagnostic. Legacy OpenCode, other harnesses, personal Codex desktop sessions and hosted claude_managed are outside this feature.

Changes and process lifetime

Paperclip reads settings once when an execution begins preparation. That snapshot governs instruction reservation, compatibility and launch. Already preparing or running work retains its configuration. A changed settings revision or resolved credential version retires an idle incompatible warm process using existing ownership checks. Conversation compatibility is separate, so a safe replacement can resume the saved conversation. Recovery never replays an uncertain turn merely to apply new export settings.

Missing, disabled, invalid or unavailable configuration runs without company export and reports a sanitized availability diagnostic where relevant. Native export rejection/timeouts do not determine task success; exporter deadlines are three seconds. OpenCode starts its bounded asynchronous request when the terminal event is observed; event delivery and task settlement do not wait for the receiver.

Credentials and access

Only board users with Manage connections permission can read or change this configuration. Agents cannot retrieve it through the settings API. The destination stores no credential values. Authentication headers live in one encrypted Paperclip-managed company secret, bound to the system consumer company-observability. Updates rotate that secret; removal uses the secret service to retire it. Braintrust project routing is derived from the current project ID, independently of the API key.

Reads return configured status and header names. Blank existing custom header values retain the saved secret at the same HTTP origin; a change of scheme, host, or port requires every retained header value to be entered again. New/renamed headers require values. Removing a row removes that header. API-key replacement is explicit, and Cancel clears entered secrets. Revisions prevent another administrator's edit being silently overwritten. Credential values are excluded from task environment manifests, run/config snapshots, argv, activity records, recovery configuration and query cache data. OpenCode keeps exporter credentials in the runner wrapper and excludes them from its child environment and config. Codex and Claude Code receive headers through their process environment, which their tools can inherit or inspect. This feature does not isolate export credentials from agent code. Use a credential limited to telemetry ingestion for the selected destination; do not supply a broad organization administration key. A user-operated collector can hold upstream credentials when a deployment requires a separate trust boundary.

Connection tests

Saving never contacts the destination. Test saved settings sends one synthetic root span and a correlated log with a unique marker from the control plane, even when agent export is disabled. The receiver receives no production task content. Both requests share a ten-second deadline, cannot redirect, and use the deployment's guarded HTTP policy with DNS pinning and response-size limits. Receipts contain sanitized HTTP/signal results and no receiver body text. A receipt is retained only for the revision that was tested.

“Test payload accepted” means the server's synthetic payload was accepted. It does not prove agent export, remote network reachability, downstream indexing, or that a provider UI rendered the records. Harness exporters use their own network clients under the execution environment's policy; the server's DNS pinning applies only to the explicit control-plane test. Verify real tasks and the destination's records before relying on a deployment.

Verification

The focused suites cover schema bounds, concurrent saves, secret/header lifecycle, authorization, protobuf transport and rejection, private-network policy, timeout, revision fencing, environment projections, private native configuration, warm compatibility, and UI conflict/company-switch behavior. Storybook IDs agent-observability-01-settings, agent-observability-02-states and agent-observability-03-journey use fake credentials and production components. Real local/remote native export evidence must be reported separately from these automated and synthetic checks.