docs(agent): add harness onboarding and upgrade guide by yaoyu-33 · Pull Request #3917 · NVIDIA-NeMo/Gym

yaoyu-33 · 2026-10-02T01:18:05Z

What does this PR do?

Add a contributor guide for integrating or updating harnesses, linked from Integrate Existing Agents. It covers independent benchmark/agent composition, explicit host and sandbox execution modes, installation, request and error semantics, cleanup, and observability. Both Mermaid diagrams identify the protocols: HTTP/JSON between FastAPI Gym services, provider-specific execution and file transfer, subprocess/signals, and local JSON files. The architecture diagram includes model requests from both host and sandbox harnesses.

The experiment walkthrough provides four composable YAML files for model/provider, benchmark, agent, and EnvironmentServer. New experiments use single_agent_turn with benchmark-owned taskset declarations and environment_server_routes; gym eval run collates native task_id/task_input rows, starts servers, collects one task, and shuts down. Stable instance names support a Hermes → Pi swap and a separate experimental Terminal Bench 2.1 benchmark swap. The old flat-row adapter is retained only as a migration note. The examples target current-main native adapters and require no pair-specific presets or conversion script.

The lifecycle comparison and sequence diagram make Resources Server responsible for sandbox selection, reuse, and destruction. The TB2.1 session implementation at #3902 (d4f9f1aa2) verifies in the original sandbox and leaves destruction to final Resources close. SWE-Pro exports the patch, releases the original sandbox, and creates/stops a fresh verifier per attempt. The scope is the normal borrowed-session path; the TB2.1 claim is pinned to its implementation revision rather than assumed available in every checkout.

The sandbox guidance follows the Hermes #3961 revision a309dfb9a and its review comments:

Shared execution helpers live in nemo_gym/agent_utils/; general file transport stays in nemo_gym/sandbox/utils.py.
SandboxSession uploads the supervisor, accepts stage_activation and collect hooks, and manages cleanup, capture, and release. SandboxCommand contains only argv and the installed supervisor interpreter.
The guide documents session control files, shared timeout calculation, adapter-local optional runtime metadata, and crash-cleanup requirements. Hermes now defaults an owned sandbox's TTL to installation timeout + runner timeout + 600 seconds, preserving explicit values (including null) and borrowed-sandbox ownership. Every supervised harness requires Python 3 in the sandbox, including standalone CLI binaries.
Process control uses provider exec, shell commands, signals, and files. Model requests originate from the harness. The current Hermes path uses Chat Completions while the Gym Agent exposes a Responses API contract; the supervisor does not forward API requests.

Pi #3626 (d91469b81) and Codex #3640 (5151a3a74) now use the same agent_utils session API, updated hooks and fields, and adapter-local runtime metadata. The guide removes the stale migration warning while retaining host-only startup and per-pairing runtime validation requirements. Shared installation and provider mounts remain follow-ups; the page distinguishes them from the actual Hermes installation prerequisites.

This consolidates guidance from #3961 (comment) and the subsequent reviews. Tracks GYM-1011. The observability section links the conformance tooling from #3766 and distinguishes local evidence probes from remote sandbox lifecycle validation.

Updates after `0c72cec9d`

feat(terminal-bench): support Pi, Codex, OpenCode and OpenClaw #3902 second review (d8a306135): follows the second #3902 review as resolved in ab9f3cc27. On timeout or stop the supervisor records already adopted children, stops the harness group, then drains newly adopted children across sessions; natural exit only drains the group, and the remaining limits are stated. Adds rules to keep every harness process in the supervised group (OPENCLAW_NO_RESPAWN=1), make harness hooks fail closed with a readiness receipt (OpenCode task-environment.ready), and bound downloaded logs; notes the Terminal Bench 2.1 close path without automatic retry; adds the AGENTS.md runtime-install rule; and expands termination and test-hygiene validation. feat(terminal-bench): support Pi, Codex, OpenCode and OpenClaw #3902 links now pin ab9f3cc27 (line ranges re-checked).
Integration path and capabilities (14c537bab): adds a starting point that compares Remote Agent, an Agent Server adapter, and a custom Environment Server; a capability table (evaluation, observability, environment tools and sandboxes, training) with the adapter work and validation each requires; and a split between what an adapter writes and what Gym provides. Adds guidance on routing every policy-generating model call (including summarization, compaction, and subagent calls) through the rollout-prefixed Model Server, echoing Model Server cookies, explicit handling of configuration and request fields, runtime profiles, lossless output conversion, tool source selection, and per-session isolation. Facts were checked against main: Remote Agent has no model_server; Pi and Codex reject unsupported request fields; SimpleAgent echoes Model Server cookies; the vLLM Model Server routes by rollout ID or session cookie; the DeepAgents adapter sets the model profile used for summarization.
feat(terminal-bench): support Pi, Codex, OpenCode and OpenClaw #3902 termination contract (6ae477488): aligns the guide with the #3902 review as resolved in c2a6dec84: process-group termination with detached task services preserved, retryable missing evidence versus a final negative receipt, task HOME/XDG_* preserved for tool shells, Terminal Bench 2.1 verification caching and close fencing, and the per-callback cleanup_timeout_seconds. A Note states that main keeps the earlier policy until feat(terminal-bench): support Pi, Codex, OpenCode and OpenClaw #3902 lands, so this PR should merge after feat(terminal-bench): support Pi, Codex, OpenCode and OpenClaw #3902. Also documents agent_utils/uv_bootstrap.py, which Hermes uses on main.
Contracts and structure (f094bf957): adds a terminology table, explicit Agent Server contracts with the Environment Server, the task sandbox, and tool grants, an adapter configuration table, and a step-by-step SandboxSession wiring guide. Links Hermes, Pi, and Codex to main now that agent_utils is merged there.
Review follow-ups from @lafrancef: verification-policy selection criterion and GYM_HOST discovery (ce4d98e2b).

Validation

Final documentation commit: 0c72cec9d9ce6e0966693b353ee18c431b9b503a.

At d8a306135:

python -m pytest tests/unit_tests/test_fern_docs_links.py — 23 passed.
pre-commit run --all-files — all hooks passed.
fern check — 0 errors; redirect validation skipped without Fern authentication. Full-page Fern preview remains unverified.
git diff --check passed; MDX scan found balanced <Note>/<Warning> blocks and code fences, and no unescaped </{ in prose.

Earlier validation at 0c72cec9d:

python tests/unit_tests/test_fern_docs_links.py — 18 passed.
python -m pre_commit run --all-files — all seven applicable hooks passed.
FERN_NO_VERSION_REDIRECTION=true FERN_DISABLE_TELEMETRY=true node <cached-fern-cli>/cli.cjs check --warnings — 0 errors; redirect validation skipped without local Fern authentication. Full-page Fern preview remains unverified.
git diff --check and git diff --cached --check — passed; commit has DCO sign-off.
Extracted the six final YAML blocks and validated all four SWE-Pro/Terminal Bench × Hermes/Pi compositions against isolated main 30f63c2bca0ffcc79dc0a110b2ee23b5bd2b4c82. Used offline GlobalConfigDictParser, dataset discovery, actual TrainDataProcessor.run with strict task-data validation and local synthetic fixtures, collector preprocessing, taskset routing, and SingleAgentTurnRequest validation. Confirmed source JSONL preservation, stable references, native input conversion, rejection of flat input/missing routes, and Terminal Bench pairing rejection/explicit opt-in. No model or sandbox was launched.
Parsed all three displayed Gym commands through the real CLI dispatcher, normally and with the experimental override: six checks passed. Verified gym eval run selects the end-to-end entry point without an agent selector or flat --input.
Lifecycle source review: TB2.1 feat(terminal-bench): support Pi, Codex, OpenCode and OpenClaw #3902 at d4f9f1aa209b5c20f946a781c391d493a7d821b4; SWE-Pro at 7f16ec541534b95dde258c449be09a461b33e8d4, with the extraction/verifier cleanup block unchanged at current main above. Checked EnvironmentServer close-before-verify and final Resources close ordering.
The sequence diagram is unchanged from 90c0a2386, where it rendered with the installed Mermaid CLI and existing Chrome 147 and both branches were visually inspected. Standalone diagram validation does not establish full-page rendering.

Earlier implementation evidence: targeted Hermes lifecycle/ownership tests passed at the separately recorded a309dfb9a and 2093960be revisions (6 and 258 tests respectively; 13 Linux-only skips in the latter). This historical evidence does not qualify the new example pairings.

Rollouts

N/A: documentation only; no runtime behavior changed. Real model/task rollouts and Linux process cleanup remain part of each harness implementation's validation.

Compatibility and benchmark impact

No runtime or benchmark-result changes. The guide preserves sessionless host execution, rejects implicit host fallback for episode sessions, documents the three adapters' shared API, and identifies remaining installation and provider-validation work.

Checklist

Read the contributing guidelines.
Focused documentation change.
Documentation checks and full pre-commit pass.
Documentation added and linked.
New source-file headers and model rollouts are N/A for this docs-only change.
Commits include DCO sign-off.

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

copy-pr-bot · 2026-10-02T01:18:09Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

…boarding Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

lafrancef

LGTM for observability/conformance section, two optional suggestions for other sections.

…o codex/docs-harness-onboarding Keep the verification-policy guidance and GYM_HOST discovery from ce4d98e, placed in the restructured sections. Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

…boarding Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

…ntract Address the lifecycle findings from the NVIDIA-NeMo#3902 review (NVIDIA-NeMo#3902 (review)) as resolved in c2a6dec: - The supervisor stops and reaps the harness process group; explicitly detached task services survive until the sandbox owner tears down. - Separate retryable missing evidence from a final negative receipt, which releases a borrowed connection without making the run gradable. - Keep the task HOME and XDG directories for tool shells. - Describe Terminal Bench 2.1 verification caching, close fencing, and the per-callback cleanup timeout without keep_sandbox_for_close. - Add termination, service, and task-environment validation checks. Also document uv_bootstrap.py, which Hermes now uses to pick a uv build for the sandbox platform. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Help users choose between Remote Agent, an Agent Server adapter, and a custom Environment Server before writing code, and show which Gym capability each level of adapter work enables. Split the work into what an adapter writes and what Gym already provides. Document routing every policy-generating model call, including auxiliary middleware calls, through the rollout-prefixed Model Server; Model Server cookies; explicit handling of configuration and request fields; runtime profiles; lossless output conversion; tool source selection; and per-session isolation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Follow the second NVIDIA-NeMo#3902 review as resolved in ab9f3cc: - On timeout or stop, the supervisor records already adopted children, stops the harness group, then drains newly adopted children across sessions. Natural exit only drains the group. State the remaining limits and that harnesses start each tool command in its own session. - Keep every harness process in the supervised group; disable self-respawn, as with OPENCLAW_NO_RESPAWN=1. - Make harness hooks fail closed with a readiness receipt, and bound downloaded logs. - Note the Terminal Bench 2.1 close path that has no automatic retry. - Add the AGENTS.md runtime-install rule for sandbox and host paths. - Expand termination validation and test hygiene. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Cross-checking the Stirrup sandbox adapter in NVIDIA-NeMo#4178 against this guide found gaps the guide did not state: - Return 4xx for deterministic seed and request errors, since the Environment Server retries 5xx and re-seeds the sandbox. - Size the episode timeout to cover seed, install, run, close and verification, and use admission to queue before the deadline starts. - Handle SIGTERM in the sandbox runner and write partial output. - Serialize non-string grant tool results instead of validating them as arguments. - Constrain transitive dependencies of in-sandbox installs and document package-index egress. - Keep direct_http session cookies, which reach non-tool routes, out of files the model can read. Also align the termination section with NVIDIA-NeMo#4614: preserved children include unfinished background jobs, services must be adopted before the stop, and descendants of a preserved process are not drained. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

The execution-mode table listed borrowed, agent-owned, and host paths as equal options. State that a borrowed sandbox is the default, limit agent-owned sandboxes to benchmarks without a task sandbox or setups without a shared sandbox service (starting with Docker), and mark the Local provider and host path as unisolated, for development only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

"Need only evaluation scores?" read as if other paths did not produce scores. Name the missing capabilities in the decision node and lead the Remote Agent limits with "Scores only" followed by each "No ...". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

… current adapters - Repin the Terminal Bench 2.1 and process_supervisor links to the NVIDIA-NeMo#3902 merge commit 8bd25f3 and drop the pre-merge hedges. - Name OpenCode and OpenClaw alongside Hermes, Pi, and Codex as agent_utils adapters, and add their tool-source rows. - Note where the reference adapters and Resources Servers do not follow the guide's rules yet: Hermes seed errors surface as HTTP 500, Hermes pins its version in requirements.txt, and Terminal Bench 2.1 and SWE-bench Pro return a cached reward of 0 on verification infrastructure failures. - Describe agent close as about five per-step timeouts, and size cleanup_timeout_seconds accordingly. - Set default_episode_timeout_seconds in the example so it follows the episode-timeout rule, and say to raise the admission settings with --concurrency. - Say that Terminal Bench 2.1 rows supply each task's image, add --require-sandbox to the pairing check, and clarify the session cookie, model_endpoint(), checkpoint opt-in, and the log-trimming example. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

yaoyu-33 · 2026-10-16T05:08:24Z

/ok to test 3619379

github-actions · 2026-10-16T05:11:16Z

🌿 Preview your docs: https://nvidia-preview-pr-3917-3619379d53a3.docs.buildwithfern.com/nemo/gym

Describe the Environment Server, the seed/activate/close/verify/cleanup episode flow, and Resources Server sandbox ownership on the Architecture page, and link the harness guide to it. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

…NVIDIA-NeMo#4711) * docs(architecture): describe Environment Server episode orchestration Add the Environment Server to the architecture page: a table row, an episode lifecycle (seed, agent loop, close and verify, clean up) with a sequence diagram, a server-type section, a sandbox-ownership note and a sandbox-transport note. Split from NVIDIA-NeMo#3917 so the concept change can be reviewed on its own; the harness guide and agent-server pages follow. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> * docs(architecture): scope episode orchestration to native Environment Servers Most shipped configurations still run through the legacy_agent relay, where the agent seeds and verifies itself. Scope the ordered steps, the component table row, the Environment Server section, and the sandbox-ownership note to native Environment Servers, describe the legacy_agent path, and stop saying that agent harnesses orchestrate rollouts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> --------- Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

* docs(agent): add harness onboarding and upgrade guide Add a contributor guide for integrating or updating an agent harness that runs in a task sandbox: independent benchmark and agent composition, the episode lifecycle and sandbox lifetime policies, the Agent Server contracts, configuration, a worked experiment with a Hermes to Pi and a SWE-bench Pro to Terminal Bench 2.1 swap, the agent_utils session helpers, input and error mapping, retry and close safety, and per-pairing validation. The page content is the guide from NVIDIA-NeMo#3917 at 3955a97, unchanged. It is split out so the architecture change and the agent-server page structure can be reviewed separately. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> * docs(agent): bring the harness guide up to date with main - Step 3 resolves the time budget with resolve_agent_timeout(), runs under the adapter's concurrency semaphore, and names the session lookup helpers; list agent_utils/task_timeout.py, and note that Pi accepts metadata.agent_timeout_sec (NVIDIA-NeMo#4653). - Pass --allow-unsupported-pairing only to gym eval run, which checks the pairing at dispatch; prepare and resolve do not accept it. - Set cleanup_timeout_seconds: 360 in the run.yaml example so Pi's close fits, and note that the benchmark sandbox ttl_s is a separate limit. - Close the provider when AsyncSandbox.connect() fails in the borrower snippet. - Correct the OpenCode version pin, the Hermes 500 cases, the migration note, and the SWE-bench Pro empty-patch path; add the requirements.txt row with the sandbox extra. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> --------- Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

docs(agent): add harness onboarding and upgrade guide

10375f3

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

yaoyu-33 added docs Documentation-only updates or documentation debt area:agent Agent harnesses and Responses API agent behavior labels Oct 2, 2026

yaoyu-33 added 5 commits October 1, 2026 18:24

docs(agent): trim harness guide closing guidance

f792932

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): illustrate the sandboxed harness lifecycle

ab0dfd4

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Merge remote-tracking branch 'origin/main' into codex/docs-harness-on…

0e35b2a

…boarding Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): add observability conformance validation

734d150

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): generalize the session support note

222719c

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

yaoyu-33 marked this pull request as ready for review October 5, 2026 18:37

yaoyu-33 added complexity:low Localized change in one scope with a small, straightforward review surface needs-review PR is ready for code review and waiting on a reviewer labels Oct 5, 2026

docs(agent): explain sandbox integration and execution modes

cfde229

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

athitten mentioned this pull request Oct 6, 2026

feat(gdpval): run Codex on GDPval task sandboxes athitten/Gym#1

Draft

7 tasks done

docs(agent): align sandbox guide with shared session lifecycle

8c8e389

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

athitten mentioned this pull request Oct 6, 2026

fix(gdpval): run Pi sandbox sessions end to end and score them with multistage ELO athitten/Gym#2

Draft

7 tasks done

docs(agent): align harness guide with supervisor ownership changes

897e8f2

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

github-actions Bot added the sla:triage-overdue Review assignment is over the one-business-day SLA label Oct 6, 2026

github-actions Bot mentioned this pull request Oct 6, 2026

[Tracker] Pull request handoff SLA #3402

Open

yaoyu-33 added 5 commits October 6, 2026 13:36

docs(agent): label harness protocols and sync ready implementations

33038b2

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs: clarify harness sandbox ownership and runtime mounts

d40a554

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): align onboarding with individual P0 checks

b1853b6

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): clarify task sandbox lifetime during verification

8bd000f

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): distinguish masked local provider failures

4f2d228

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

linear-code Bot mentioned this pull request Oct 12, 2026

Add harness onboarding and update guide #4477

Closed

yaoyu-33 added 3 commits October 12, 2026 16:54

docs(agent): show experiment composition and benchmark swaps

31c93bc

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): contrast Resources-owned sandbox lifetimes

90c0a23

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): use native tasksets for onboarding experiments

0c72cec

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

lafrancef reviewed Oct 13, 2026

View reviewed changes

Comment thread fern/versions/latest/pages/agent-server/add-or-update-a-harness.mdx Resolved

Comment thread fern/versions/latest/pages/agent-server/add-or-update-a-harness.mdx Resolved

github-actions Bot removed the sla:triage-overdue Review assignment is over the one-business-day SLA label Oct 13, 2026

yaoyu-33 and others added 14 commits October 13, 2026 13:02

Merge remote-tracking branch 'fork/codex/docs-harness-onboarding' int…

1b6f558

…o codex/docs-harness-onboarding Keep the verification-policy guidance and GYM_HOST discovery from ce4d98e, placed in the restructured sections. Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Merge remote-tracking branch 'origin/main' into codex/docs-harness-on…

593b29b

…boarding Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): clarify sandbox uv probe fallback

93e2e9f

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): clarify initialization receipts and subreaper scope

dbf7fd5

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): qualify detached shell service lifetime

077787e

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): refresh landed lifecycle and Pi pairing guidance

e0c2b2d

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

docs(agent): assign activation retries to the adapter

a014ba1

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

nemo-automation-bot Bot requested a review from a team October 16, 2026 05:09

copy-pr-bot Bot temporarily deployed to public October 16, 2026 05:09 Inactive

copy-pr-bot Bot temporarily deployed to public October 16, 2026 05:11 Inactive

ananthsub requested a review from cwing-nvidia October 16, 2026 17:35

yaoyu-33 marked this pull request as draft October 17, 2026 15:42

This was referenced Oct 17, 2026

docs(architecture): describe Environment Server episode orchestration #4711

Merged

docs(agent): add harness onboarding and upgrade guide #4712

Merged

docs(agent): organize agent-server pages by reader #4713

Merged

yaoyu-33 closed this Oct 23, 2026

Navigation Menu

Repository navigation

docs(agent): add harness onboarding and upgrade guide - #3917

docs(agent): add harness onboarding and upgrade guide#3917
yaoyu-33 wants to merge 35 commits into
NVIDIA-NeMo:mainfrom
yaoyu-33:codex/docs-harness-onboarding

yaoyu-33 commented Oct 2, 2026 •

edited

Loading

Uh oh!

copy-pr-bot Bot commented Oct 2, 2026

Uh oh!

lafrancef left a comment

Uh oh!

Uh oh!

Uh oh!

yaoyu-33 commented Oct 16, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

yaoyu-33 commented Oct 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Updates after 0c72cec9d

Validation

Rollouts

Compatibility and benchmark impact

Checklist

Uh oh!

copy-pr-bot Bot commented Oct 2, 2026

Uh oh!

lafrancef left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

yaoyu-33 commented Oct 16, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

yaoyu-33 commented Oct 2, 2026 •

edited

Loading

Updates after `0c72cec9d`