Efficiency Conversion Loop

ui/efficiency conversion loop

This skill drives the workflow around a conversion once the test itself is written. Test authoring is a separate skill (efficiency-test-authoring). The five steps:

convert → file bug (blocking meta 2030727) → annotate @Converted + commit (with bug #) → track in Jira → submit for review

What runs where

• The agent (sandbox) does the conversion, static checks, and drives the bridge; it creates Jira items with tools/jiratool.py (preferred — it works headless), falling back to the Atlassian connector only if that tool is unavailable. It cannot reach Bugzilla or Phabricator directly.
• The host bridge (effwatch) runs the git + Bugzilla actions on the engineer's machine and returns results. It never pushes and never submits.
• The engineer runs effwatch, keeps a device attached, and runs the final moz-phab submit.

The eff* tools live in the testops-tools repo under tae-conversion/tools/ (see tae-conversion/README.md for setup). Clone that repo and start effwatch before using this loop.

Prerequisites (one-time)

1. effwatch running on your machine (from testops-tools/tae-conversion/tools/), device/emulator attached.
2. A Bugzilla API key available to effbug — set it once in ~/.zshenv (export BUGZILLA_API_KEY=…) or a tae-conversion/tools/.eff.env file (gitignored). Never paste it into chat.
3. moz-phab installed and authenticated to phabricator.services.mozilla.com.

Project-specific IDs — edit these for your team

This skill is written for the Fenix smoke-conversion campaign. If you're running it elsewhere, these are the only values you need to change:

The loop

1. Convert + pre-flight

Convert the legacy test onto ui/efficiency (see efficiency-test-authoring). Run effcheck.py (static pre-flight) then build/run via the bridge (effverify / effwatch) until green, or "good enough + notes."

Before you file anything: run the parity audit from the tae-test-review skill. A conversion that dropped a legacy assertion still goes green, so a passing run is not the gate — the legacy-vs-port diff is. Land-blocking findings are cheaper to fix now than after the bug and commit exist.

2. File the Bugzilla bug (agent → bridge → effbug)

Drop conversion-runs/_queue/<id>.request.json:

{ "bug": "create",
  "summary": "[efficiency] Convert <Test>.<method> to ui/efficiency",
  "comment": "<what was ported; parity notes>",
  "why": "<one-line rationale>",
  "kind": "conversion",            // conversion | enablement | tooling — picks the description footer
  "testrail": "<id(s)>",
  "template_bug": "2057054",        // clones product/component/version → Firefox for Android :: UI Tests
  "blocks": [2030727],              // REQUIRED for conversions — the tracking meta bug (see below)
  "type": "task" }

effbug files the bug, then rewords the title to Bug NNNNN - <summary> so it matches the commit subject exactly, and self-assigns to the API-key owner. It returns the number in _bug/<id>.bug-result.json.

Check the TestRail ids before you file. Comment 0 cannot be edited through the BMO API, so a wrong id is permanent unless you post a correction. Compare each id against the line immediately above the legacy method (not a grep context window — a neighbouring test's id looks identical in kind); three of six were wrong in one sitting this way, and two filed bugs needed correcting comments. If you do need to correct one after filing:

{ "bug": "update", "ids": [2064815], "comment": "Correction to comment 0: …" }

Hang the bug off the tracking meta bug — "blocks": [2030727]. Bug 2030727 is [meta] TAE - Migrate and remove legacy tests; it tracks the campaign via its depends_on list, so each conversion bug must block it. Pass blocks at create time — that is one field, versus a second round-trip afterwards, and a conversion that never gets linked is invisible to whoever reads the meta bug for campaign status.

Scope: test-conversion bugs go on the meta; tooling/docs/harness bugs do not. The meta is specifically about migrating and removing legacy tests, which is why e.g. the effview-tool and harness-docs bugs are deliberately absent from it. If a conversion also needed harness work, the conversion bug still blocks the meta — the enablement is tracked in Jira (step 4), not by a second meta entry.

To backfill one you already filed:

{ "bug": "update", "ids": [NNNNN], "blocks": [2030727] }

To close one as a duplicate — which happens when the same test gets converted twice, see step 1 — { "bug": "update", "ids": [NNNNN], "dupe_of": MMMMM, "self_assign": false }. effbug fills in RESOLVED/DUPLICATE for you; passing resolution alone is rejected, because Bugzilla will not take DUPLICATE without dupe_of. Note there is no API for editing a bug's description: a wrong comment 0 can only be fixed by a human in the web UI, so get the mechanism right before you file. update wraps relation lists as {"add": [...]} so this appends. Never PUT a bare list to a meta bug's depends_on — Bugzilla treats that as replace and it would drop every other bug the meta tracks.

3. Commit with the real bug number (agent → bridge → effgit)

First annotate the legacy method(s) you just replaced — this goes in the SAME commit as the conversion, and it is the step most often forgotten:

@Converted(
    replacedBy = ["org.mozilla.fenix.ui.efficiency.tests.<Class>#<method>"],
    bug = NNNNN,          // the bug you filed in step 2
    since = "YYYY-MM",
    notes = "Legacy also asserted X; not carried over because …",   // only if coverage was dropped
)

The gate is green locally (step 1's effverify verdict), not landed — you cannot annotate after landing without a second bug and a second review, and every conversion in this campaign has landed the annotation alongside its replacement. replacedBy is required, one entry per replacement, and each must resolve to a real non-@Ignored @Test. Put the parity gaps from step 1 in notes — that is the auditable record of what did not carry over. Annotate the legacy method in place; do not delete it, it keeps running alongside the replacement.

Then write the commit message to conversion-runs/<batch>/msg.txt with Bug NNNNN - [efficiency] … r=isabel_rios,aaronmt and drop { "git":"commit", "message_file":"<batch>/msg.txt", "paths":[...] } — the paths list must include the legacy test file you just annotated as well as the new/changed efficiency files. (If a commit already exists with a placeholder, backfill by rewording — the loop files the bug before committing going forward, so no reword is needed. effgit's amend does not stage: send stage first, then amend.)

4. Track in Jira (agent → jiratool.py)

Separate strict conversion from the enablement it sometimes forces, so conversion effort can be measured:

• Conversion → a Sub-task labelled conversion under the Smoke-conversion campaign story MTE-5731.
• Tooling/enablement discovered during conversion → a separate Sub-task labelled enablement under Harness Hardening MTE-5715 (or Tech-Debt MTE-5688), linked ("Relates") to the conversion sub-task.
• Put the bug number, branch/commit and Phab revision in the item.

Use tools/jiratool.py (works headless; no Atlassian connector needed). Bodies come from a file, so write the description to a temp file first:

python3 jiratool.py create '<summary>' --file body.txt --parent MTE-5731 --issuetype Sub-task --label conversion
python3 jiratool.py create '<summary>' --file body.txt --parent MTE-5715 --issuetype Sub-task --label enablement
python3 jiratool.py link <enablement-key> Relates <conversion-key>
python3 jiratool.py assign <key> --me        # create does NOT self-assign; do this explicitly

create defaults to --issuetype Story and no labels, so pass both every time or the item lands as an unlabelled Story in the wrong shape. --label is repeatable. If the Atlassian MCP connector happens to be connected it also works, but do not count on it — it is absent in headless/cron runs.

An enablement sub-task is warranted whenever the conversion needed a new page object, selector twin, nav edge or moz*/mozVerify* primitive — i.e. build mode 3. Write up what the gap was and what the proper fix would be, not just the workaround you shipped.

5. Submit the finished stack (engineer)

Submitting/landing stays with the engineer. Mozilla's moz-phab has no --dry-run — it's interactive: it prints the commit list and prompts Y/n before creating anything (that's your preview). Bound the range so it can't touch already-landed base commits:

moz-phab submit --reviewer isabel_rios --reviewer aaronmt <first-new-commit>

(or python3 tae-conversion/tools/effsubmit.py --start <first-new-commit> --execute).

A cherry-picked base whose revision is CLOSED blocks the whole submit. Borrowing an unlanded commit as a base is fine for building and running, but moz-phab refuses the stack if any commit in range maps to a closed revision, and unpicking it late means reworking whatever depended on it. So before cherry-picking, diff what you actually need against main: on 2026-08-19 everything needed was already landed except a one-line helper written in the same session, so the borrowed commit was dropped and the helper inlined. And never create a file that the unlanded commit also creates — that is an add/add conflict on every rebase until it lands. Then add the testing-exception-unchanged tag in the Phabricator web UI (no moz-phab CLI flag exists for it). moz-phab keys off Differential Revision: trailers, so base commits that carry them are excluded automatically — even if they landed on autoland and aren't in your local central yet.

If you rebase a stack that is already submitted

Dropping a commit does not remove its revision from the stack graph. Abandoning the revision leaves the next one still recording it as a parent, with a diff based on the commit you dropped, so the stack renders with an abandoned revision wedged in the middle. Resubmitting the whole range is what re-parents it — submitting only the commits whose content changed leaves the stale edge in place.

Two consequences to warn the engineer about before they push: every revision gets a fresh diff, because a rebase changes every hash, and revisions that were already accepted reset to needs-review. If avoiding that churn matters more than a wording fix, leave the commit messages alone — amending one forces the upload you were trying to avoid.

After landing

Re-sync the tracker so conversion counts catch up with the @Converted markers that landed in step 3 (see tae-conversion/README.md → "Reconciling the ledger" and tae-conversion/tools/reconcile_conversion.py).

If reconcile reports a converted test with no @Converted marker, the annotation was missed in step 3 — that is a gap to backfill under a follow-up bug, not the normal path. Annotating is step 3's job.

Conventions

• Faithful-port-first: don't rewrite behavior during conversion; log quality ideas separately.
• One bug per landable unit; mirror any conversion/enablement split in the Jira items.
• Bugs + Jira items are self-assigned back to the engineer who ran the loop.