Add focused guidance for READMEs, file descriptions, routine/code comments, and plain language, with a documentation entry skill selected by AGENTS.md. This gives agents one route to the writing rules for the current task.
Shorten the existing PR/commit guide and require BFB discussion only when a change could plausibly affect model results or scientific output.
All 18 skill definitions and 45 local guidance links validate; git diff --check passes.
* Require code reading and checks before finishing documentation
Route small documentation edits through the writing skills. Require agents to inspect the implementation before drafting and check facts, coverage, links, and comment syntax before finishing. Clarify that weak file headers need complete descriptions in the appropriate language form.
Validation: all 18 skill schemas, 46 local guidance links, and whitespace checks pass. The fresh-agent study uses these exact frozen guidance files.
* Clarify documentation wording and optional usage examples
Explain meaning before usage and make important ordering clear. Read the applicable language or porting standards for documentation conventions. Describe file interfaces independently of current callers, using callers as optional examples. Recommend examples in file headers when useful, or beside individual routines for collections of independent helpers.
Validation: all 18 skill schemas, 49 local guidance links, and whitespace checks pass. These refinements have not had a new fresh-agent trial.
Add Jenkins workflows and branch finalization guidance (#1396)
* Set up Jenkins API login over HTTPS
Add an interactive token login script and teach the Jenkins skill to reuse its private local credential file or direct users through setup. This lets each developer use their own Jenkins account without SSH access.
Ten focused regression checks and a live HTTPS login/save check passed. BFB is not applicable: login tooling and guidance do not affect model execution or output.
* Explain Jenkins login setup and agent use
Describe why agents use the saved login, how developers set it up, and what the public routines and failure paths do. Apply the new documentation guidance to the script, test comments and usage documentation.
Executable Python statements are unchanged. Ten focused login checks pass.
* Remove Jenkins login pytests
Remove the provisional test file from the Jenkins login change at the user’s request. Keep the login script and workflow guidance unchanged.
* Document Jenkins access and test authoring
Add a Jenkins README and a separate authoring skill for simple, reproducible pipelines and coordinated live-job changes. Keep API access in its existing skill and route both workflows from AGENTS.md.
Place the login script with the Jenkins files and update its references. Runtime statements and the credential location are unchanged. Login help, import, live HTTPS identity, skill validation and local links pass.
* Prefer relevant Jenkins evidence in test reports
Add shared guidance for choosing checks and explaining what their results establish. Lead PR reports with relevant Jenkins build links, include local checks when they add evidence, and keep unexpected failures visible.
Connect validation and change-message workflows to the shared skill. Skill metadata and local guidance links validate; only Markdown changes.
* Keep README verification notes in change reports
Remove the authoring-history sentence from the Jenkins README and clarify where the README skill should record example verification. Review the PR documentation for similar notes while retaining reader-relevant setup.
Skill metadata, local links and diff formatting checks pass.
* Group related skills under their entry points
Nest documentation guides and Jenkins test authoring while preserving skill names and workflow text. Rebase links, make shared parent guidance explicit, and use recursive skill discovery for validation.
The Codex skill loader discovers every nested child. Recursive schema, link and preservation checks pass.
* Clarify Jenkins test authoring scope and when to wait
Require a user request before changing Jenkins test definitions, while allowing relevant existing branch builds during normal validation. Align pytest integration guidance with that boundary.
Guide intermediate work toward fast meaningful checks and later review of slow builds; weigh longer waits at final validation and report pending results.
All 20 skill schemas and 99 local guidance links validate; diff formatting passes. No pipeline or runtime changes, or live Jenkins actions.
* Explain Jenkins permissions before promising site changes
Describe the permissions needed to read, run, create or edit jobs without assuming a particular account. Link the guidance from authoring and setup, and explain how to use assigned tokens or switch saved accounts.
All 20 skill schemas and 101 local guidance links pass, including the new section anchors. Diff formatting passes; this commit changes Markdown only. Prior live HTTPS checks confirmed creation, build and cancellation access, read-only configuration access, and denied configuration writes.
* Add branch finalization workflow and agent PR conventions
Define merge preparation as a full branch audit followed by current validation and a commit, push and draft-PR handoff. Pause for substantive unintended issues the user has not accepted; allow minor cleanup without interruption.
Route finalization from AGENTS and shared testing guidance. Keep agent PRs as drafts and sign PR comments posted through a user account with the agent name.
Codex discovers all 21 repo skills. All 21 schemas and 110 local guidance links validate; diff formatting passes. Only agent guidance changes.
* Clarify which builds an agent may cancel
Limit cancellation to builds and queue items launched for the current task unless the user authorizes stopping others. A shared Jenkins username does not establish ownership. Wrap the nearby setup and authoring text.
Final branch audit found no substantive blockers. Login smoke checks passed for live HTTPS authentication, private credential replacement and failure preservation without changing the user credential file. All 21 skill schemas, 110 local links and 5 heading anchors pass; Codex discovers all 21 skills.
Reset microphysics defaults when initializing another case (#1398)
Restore module and saved-local namelist defaults before reading a new case. This prevents omitted settings from inheriting the previous case's values. This isn't an issue for any way that clubb is currently run in the codebase, but would potentially cause an issue for a future use case (this is foreshadowing).
Restore missing debug printouts in JAX core (#1390)
* Restore missing debug printouts in JAX core
The port kept many failure checks but omitted their Fortran messages, making failed columns and solvers difficult to diagnose. Restore runtime messages, column headers and available field dumps with the source debug gates. Record diagnostics needing unported machinery in the skip report.
BFB is not established by the available comparisons. All 23 curated CPU JAX/Fortran cases matched within default tolerances: absolute 1e-7 and percentage 1e-7%, or 0.001% for the curated Morrison cases. Runs used four columns, standard statistics and existing durations/caps. Separately, 24 starting-tree kernel comparisons found exactly equal result/state values.
The CPU JAX suite passed 787 tests, including compiled diagnostic checks; the focused rerun passed 14. Source/diff review and git diff --check passed.
* Remove diagnostic restoration report
Remove the standalone report on intentionally skipped diagnostics. Observed Fortran differences remain documented in the affected JAX modules.
This removes documentation only, so model/output BFB is not applicable. Diff review, a tracked-file reference search and git diff --check passed. Runtime tests were not rerun for this deletion.
* Identify scalar fields in solver diagnostics
The separate scalar solves all printed the same generic field label. Restore the Fortran names sclrp2, sclrprtp and sclrpthlp so a failure identifies which variance or covariance solve failed.
Source comparison confirms the three labels match their Fortran call sites. Python syntax and whitespace checks passed. Jenkins validation will exercise the published revision.
