[nplez1]

ELI5

Orca can now read a Jenkins build and paint its pipeline. But the first pass was too much: the
Checks pane listed every stage of a pipeline that can have hundreds, the failing stage was buried
underneath, an "Output" box only repeated the build's name, and a Jenkins permissions problem was a
dead-end sentence. This makes the pane short, draws the pipeline as a tree like Blue Ocean's, puts
the failures on top, and turns a fixable Jenkins error into a button that opens Settings.

What Changed

Before: expanding a Jenkins check listed every stage of the build — hundreds of rows for a
large pipeline — with the failed and still-running stages mixed in. The check-details tab repeated
the full stage list with no nesting, opened onto an Output box whose only content was the build's
display name, and reported a permissions problem with a sentence you could not act on. A build's
failure text was not read at all.

After:

• The pane lists only the stages still running and failed, running first, capped at eight,
  with a +N more stages line. A passing stage is no longer noise.
• The check-details tab renders the stages as a nested tree (stage → its steps) with a glyph and
  duration per node; the failed and running stages are open by default.
• A Failures block sits directly under the run header naming each failed stage and its error
  text, so "what broke" needs no scrolling.
• The Output box is gone for a stage list; the build metadata is the run summary.
• A Jenkins failure the user can fix — no server configured, a rejected token, or no permission —
  now shows an Open Jenkins settings button that deep-links to the Jenkins card in
  Settings → Integrations.

The mechanism. The pane reduction is paneCheckSteps in src/shared/check-step-visibility.ts.
The tree and failure text come from one more bounded round of reads: main follows each stage's own
wfapi/describe (_links.self.href, resolved against the build and refused unless it stays under
the configured server), bounded to 40 requests at six concurrent, failed/running first, under a
four-second budget. The mapper nests the stageFlowNodes on an optional PRCheckStep.children
and lifts the failing node's error.message onto the stage, so the flat pane list and the nested
tree read the same data. The Jenkins node id is carried so repeated stage names cannot collide as
React keys. The settings affordance is a typed JenkinsCheckDetailsError whose reason becomes a
serializable errorAction on the pane and tab state — not a callback, because that state is
persisted — cleared on every success and retry.

Why

A cap with priority, not an unbounded fan-out. A pipeline can have hundreds of stages, and the
Blue Ocean tree needs each stage's own data. wfapi exposes that one stage at a time, so a fan-out
is unavoidable; fetching failed and running stages first means the cap can only ever hide the steps
of later passing stages, never the failure that brought the user here. The four-second budget is the
real latency guard, so the request cap is only a server-load ceiling.

One nested, provider-neutral field rather than a Jenkins-only tree type. children and
errorMessage are generic enough to cross IPC, and threading a second tree through
PRCheckRunDetails and every surface would be more churn. The flat consumers explicitly do not
recurse — that is documented on the type and covered by a test — so outcomes are never double-counted.

The node id as the key. Loops and parallel branches repeat stage names, and parallel branches can
share a start time or have none; a name + startedAt key drops or duplicates rows. The id is
Jenkins' own and unique per run.

A structured error action, not message matching. failureMessage localizes at throw time, so
the reason is gone by the time a surface decides whether to offer a link. The typed error carries it
forward, and only the fixable reasons (not-configured, unauthorized, forbidden) offer the button.

Alternatives considered: depending on the Blue Ocean REST /nodes endpoint for a one-shot tree
(rejected — it needs the Blue Ocean plugin, which we cannot assume on psci-jenkins.adobe.io, and a
missing plugin would silently empty the pane); removing the Output box for every provider (rejected —
it is real check output for GitHub and GitLab, so it is dropped only for a stage list).

Linked Issue

None filed — nothing to link (same as #39). Say the word if you want one opened and referenced here.

Visual Proof

No screenshots: this needs a live Jenkins plus a running app, and capturing one here would be
dishonest. What follows is the rendered text of the components, whitespace-normalized, taken
from the component tests (jenkins-check-row-details.test.tsx,
CheckRunDetailsPanel.jenkins.test.tsx) — real renders, not hand-written.

Before — the pane's stage list was every stage, including the passing ones:

Failed jobs
  #12          failure
    ✓ Checkout            success
    ✗ Unit tests          failure
    ◌ Integration tests   pending
    – Deploy              skipped

After — only what is running and what failed, running first:

Failed jobs
  #12
    ◌ Integration tests   pending
    ✗ Unit tests          failure

After — the check-details tab, top down:

continuous-integration/jenkins/pr-merge              [Fix with AI] [Refresh]
Status: Failed   Started Jun 16, 5:00 AM   Completed Jun 16, 5:04 AM   Duration 4m 12s

Failures  1
  Test
    script returned exit code 1

Build
  Number        #12
  Triggered by  Started by user Ada
  Queued        2s
  Estimated     4m 0s
Parameters
  GITHUB_TOKEN  ••••
Commits
  abc123de      Fix the thing          Ada Lovelace

Failed jobs                                             1 / 2 steps failed
  #12
    ▾ ✗ Test                          56s
        ✗ Shell Script                56s
          script returned exit code 1
    ▸ ✓ Checkout                      4s

Testing

• pnpm tc (all projects) clean.

• pnpm run check:code-quality:changed clean (all seven scans, including design-system and React
  Doctor).

• Localization extraction / catalogs / coverage green; check:max-lines-ratchet green.

• Affected suites: full src/shared, src/main/jenkins, and the whole editor, right-sidebar
  and store directories — green (≈17,000 tests across those runs).

• New coverage worth naming: pane filtering/cap/order (6), rendered settings button and its clearing
  after a retry (2), the stage tree including a 250-stage list with 50 repeated names and no key
  collision (2), the failure summary (2), Output-box removal for a stage list (1), settings-error
  typing (2), reducer identity for a repeated patch (1), a stage detail that never answers ends at
  the budget (1), and the 40-request cap (1).

• Platforms: macOS only. No live Jenkins was exercised; every endpoint interaction is mocked.

• I manually tested these changes locally

• Automated tests added/updated, or explained why not below

AI Disclosure

Implemented by Pi (coding agent), with an independent review pass from the advisor escalation
before finalizing. That review found three real issues, all fixed: React key collisions for repeated
stage names (now keyed on the Jenkins node id), a regression where a check whose only output was a
title lost its Output box (restored for non-stage providers), and an inaccurate comment claiming the
request cap could never hide failure text.

Review

Worth a reviewer's eye:

1. The 40-stage cap. Stages past it render as rows with no steps. Failed and running stages are
   fetched first, so the failure path is unaffected; a very large pipeline will show every stage name
   but only expand the failed/running ones plus the earliest passing ones. Raising the cap is a
   one-constant change if we want fuller coverage (MAX_STAGE_DETAIL_REQUESTS).
2. The flat consumers must not recurse. visibleCheckSteps, paneCheckSteps and
   summarizeJobSteps read top-level steps only; only CheckStageTree descends into children.
   A test pins the pane's non-recursion.
3. Where the anchor lives. The settings deep link scrolls to [data-settings-section= "integrations-jenkins"], which the Jenkins card now sets. The id is a lib constant shared by the
   card and both surfaces.
4. The Output-box rule. hasOutput = (title || summary || text) && !showsStageList — the drop is
   scoped to a stage list, not to a provider name.

Agent skill upstream boundary

• Not applicable, or this change follows docs/reference/agent-skill-sharing-upstream-boundary.md and copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

• Also in this PR: a second commit splits five files that were over the oxlint max-lines budget
  on nplez1/main (pre-existing — pushed over by fork-only commits that never ran PR CI), which was
  failing the whole-repo lint gate for every PR cut from it. No behavior change; the repo's rule is
  to split, not suppress.
• Security: a server-supplied stage href is followed only if it stays under the configured server
  prefix (or the build origin when unconfigured), so the Authorization header never travels to a
  host we did not configure.
• Cross-platform: platform-neutral (URL handling plus the shared main-process HTTP port); only
  macOS was run.
• SSH/remote: unchanged — details are read from the machine running Orca.
• Mobile: no mobile surface consumes check details.
• Backwards compatibility: PRCheckStep.id/children/errorMessage and the new errorAction
  are optional, so an older client ignores them.
• Performance: the extra reads are bounded to 40 requests at six concurrent inside a four-second
  budget; the pane is capped at eight rows regardless of pipeline size.

Checklist

• This PR is small and focused
• I explained what changed and why (ELI5, the user-facing before/after, the mechanism, and why over the alternatives)
• Before/after screenshots or videos attached for UI changes, or N/A with reason
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and path/shortcut impact considered (or N/A)
• pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass (or CI will cover; local preferred)
  — locally: pnpm tc, the affected suites and every affected pnpm lint gate are green; the
  full pnpm test and pnpm build were left to CI.