Operate Engineering Loop

Own a tracked Operate change from issue to a review-ready draft PR. Return to the engineer only after the implementation, independent reviews, required PR checks, and Copilot review loop are complete.

Scope is the Operate pod at webapp/client/apps/orchestration-cluster-webapp/src/operate/, plus the route files in src/routes/ and cross-pod primitives in src/shared/ that the change requires.

Read frontend-unit-test for browser-mode unit tests and frontend-integration-test for Playwright tests. For the app's directory layout, pod areas and shared-code boundaries, see docs/monorepo-docs/frontend/orchestration-cluster-webapp.md. Operate's own conventions — routing, data fetching, state, styling, testing — are in operate-frontend.

Execution authorization

An explicit engineering-loop invocation, directly or through frontend-operate-migrator, authorizes local edits, branch creation, commits, pushes, draft PR creation, independent agent reviews, Copilot review requests, and handling review threads. This applies to execution, not analysis or planning. It does not authorize marking the PR ready, merging it, or changing unrelated code. Later user instructions replace the relevant permission; never carry permissions over from another invocation.

Human/team PR reviewer and assignee selection belongs to the user. Do not request or assign them unless the user explicitly names them.

This skill stops at a review-ready draft PR by default. When updating an existing PR, preserve its current state unless the user instructs otherwise. Send progress only at the end.

Inputs

Required:

• A camunda/camunda issue number or URL.

Optional:

• A concrete implementation objective.
• Backport targets.
• Validation constraints.

If the issue does not define the implementation scope, ask once before proceeding. For bug fixes, use provided backport targets or ask before opening the PR.

Operating Rules

• Follow scoped instructions, module docs, and applicable skills.
• Track branch creation, implementation, validation, independent review, publication, and Copilot review in todos.
• Keep an internal convergence ledger for the acceptance matrix, review findings, Copilot threads, gate results, and Git/PR state. This is task state, not a repository artifact: never commit it.
• Never cache issue or PR state in a file. Query GitHub live every time it matters.
• Preserve unrelated worktree changes. Never implement or push on main.
• Rebase onto origin/main; never create a merge commit.
• Treat review findings as claims to verify. Fix valid findings at their root and rebut invalid findings with evidence.
• Never count an inaccessible or failed review as approval.
• Follow repository commit and PR conventions and keep the engineer as the sole commit author.

Validation loop

Run from webapp/client. Loop on the cheapest failing tier and graduate only when it is green. npm run lint combines Prettier, ESLint, and Knip; npm run prettier:format fixes formatting.

Tier	Required gates	When	Max
edit	npm run lint:prettier; npm run lint:eslint; npm run typecheck -w @camunda/orchestration-cluster-webapp	After meaningful edits	5
component	Edit tier; npm run test:unit -w @camunda/orchestration-cluster-webapp; npm run build -w @camunda/orchestration-cluster-webapp; npm run lint:knip	When a component is complete	5
PR	npm run test:integration -w @camunda/orchestration-cluster-webapp; npm run test:a11y -w @camunda/orchestration-cluster-webapp; visual CI	After opening the draft PR	3

Never regenerate visual snapshots locally. Stop at a tier's limit and report the exact blocker; an iteration that repeats the same failure and fix without progress counts double.

Phase 1: Prepare

Inspect GitHub authentication, issue hierarchy (including paginated subissues and blockedBy prerequisites), linked work, branch, worktree, and recent main CI. Select an unblocked implementation issue if given a tracker. Create <issue-number>-<short-scope> from origin/main, then run the component tier on the untouched branch.

Phase 2: Implement and Validate

Read the surrounding code and history, then build an acceptance matrix covering observable behavior, API calls, URL state, loading/empty/error/forbidden states, permissions, tenancy, and accessibility. Record intentional omissions. Make the smallest complete change, add behavior-level coverage, and when practical prove the regression test fails without the fix. Run the edit tier after meaningful edits and the component tier when a component is complete.

Do not suppress failures, weaken assertions without rationale, or update snapshots without inspecting the result. Do not call a failure pre-existing without evidence.

Phase 3: Independent Review

Once the component tier is green, run at least two independent review agents with different perspectives:

1. A frontend domain specialist, reviewing the change against Operate's conventions in operate-frontend.
2. A separate high-confidence code reviewer, such as the code-review agent.

These are review perspectives, not skills under .claude/skills/. Select whichever independent review agents the harness provides and give them these two briefs; do not invoke a nonexistent skill.

Provide the issue, acceptance matrix, exact diff, rationale, and validation evidence. If a reviewer cannot access the worktree, provide the patch or replace that review.

Verify every finding, address valid issues, strengthen coverage where needed, and rerun the cheapest affected tier. Re-review the final diff until no actionable findings remain. Do not implement speculative suggestions or expand the issue's scope. Each review has a maximum of 3 iterations; an iteration repeating the same finding with no new evidence counts double.

Phase 4: Publish the Draft PR

Immediately before committing:

git fetch origin main
git rebase --autostash origin/main

Rerun the edit and component tiers after rebasing, commit only in-scope files, and do not add AI co-author trailers.

Push with an explicit refspec so a branch initially configured from origin/main cannot target main:

git push -u origin HEAD:refs/heads/<branch-name>

For new work, open a draft PR with gh pr create --draft, a conventional-commit title, and the completed repository PR template. For an existing PR, update it rather than creating another.

Keep the body concise: a short why and what under ## Description, the template's ## Checklist with its applicable wording (delete irrelevant options as instructed), and ## Related issues with the correct issue reference or checked no-issue opt-out. Do not invent replacement checklists or repeat scope sections. Include only brief genuine blockers or intentional behavior differences that matter to reviewers; keep detailed diagnostics in check logs or the session ledger, not long test inventories, exhaustive gate evidence, internal ledgers, or agent narration in the PR body.

Use closes #<issue> only when fully resolving it; use relates to #<issue> when the PR is one of several for that issue. A partial PR must not close an issue whose remaining work is deferred.

Confirm the PR contains the latest pushed commit.

Phase 5: Copilot Review Loop

Read and follow references/copilot-review.md. Repeat until the latest review satisfies the completion gate below. Stop only for an external blocker or repeated invalid feedback already answered with evidence.

Phase 6: CI Convergence

Run the PR tier once the draft PR is open. Diagnose a failing check with ci-fix-failure; this skill owns applying the valid fix, validating it locally, and pushing it. Rerun only verified transient failures. Refresh CI after every push; the Copilot reference owns review-thread handling.

The PR tier's 3-iteration budget in the tiered validation loop covers this Copilot/CI convergence — it is one budget, not an additional one. At the cap, report the exact blocker instead of retrying blindly or claiming completion.

Completion Gate

Before returning to the engineer, verify:

• the worktree is clean and the latest local commit is pushed to the PR
• all gates in all three tiers are green on the latest pushed SHA
• all independent review findings are addressed
• the latest Copilot review recommends approval and adds no comments, with no Copilot review pending
• no review threads or workflow todos remain unresolved
• the commit has the engineer as sole author, and the PR is correctly linked and in the state required by the execution authorization

Lead the final response with the PR number, then concisely state the delivered behavior, number of review iterations, and any blocker. Routine green checks and test counts are implicit.

Every recurring failure mode becomes a rule, not a one-off fix: encode it in this skill so it cannot recur.