[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7BWYE747PRTC2VT3TVD9GRC":3,"skill-same:01M7BWYE747PRTC2VT3TVD9GRC":53},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":32,"updatedAt":33,"install":34,"resources":46},"01M7BWYE747PRTC2VT3TVD9GRC","meticulous-increase-coverage",null,"Increase coverage for a Meticulous project by tracing specific under-covered files back to a real UI action in the codebase, driving that action with a real recorded browser session, and validating the improvement with a clean coverage comparison. Also opens a PR proposing .meticulousignore entries for code that structurally never executes in-browser. Use when asked to \"increase coverage\", \"find untested code\", or \"add .meticulousignore entries\" for a project.","CognitionAI\u002Fdevin-marketplace","CognitionAI","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F163493330?v=4","01M7BVSRD98NC4FG8T3EZS6V5M","Cognition","01M7BVSRDX111C2XPWJ0E6MY26",29,"MIT","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fmeticulous\u002Fskills\u002Fmeticulous-increase-coverage\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: meticulous-increase-coverage\ndescription: Increase coverage for a Meticulous project by tracing specific under-covered files back to a real UI action in the codebase, driving that action with a real recorded browser session, and validating the improvement with a clean coverage comparison. Also opens a PR proposing .meticulousignore entries for code that structurally never executes in-browser. Use when asked to \"increase coverage\", \"find untested code\", or \"add .meticulousignore entries\" for a project.\nuser-invocable: true\n---\n\n# Increase coverage for a Meticulous project\n\n> Run `meticulous-cli-update` first if you haven't already this conversation\n> (it also covers authentication and project selection).\n\n## What you deliver\n\nTwo separate outputs, both expected — neither substitutes for the other:\n\n1. **One or more test runs from newly recorded sessions that provably extend\n   coverage.** \"Provably\" means a comparison against the baseline naming the\n   files whose coverage went up, and by how much (Step 7). If a target turns\n   out not to be coverable, say so rather than padding the list.\n2. **A PR proposing `.meticulousignore` changes.** Only include paths — most\n   often whole directories — that you are _confident_ are not coverable at\n   all. Anything you merely failed to reach in one sitting does not belong\n   there; leave it out and mention it in the PR description instead.\n\n## Before you start: run this on `main`, with a clean tree\n\nEvery command below relies on the CLI resolving things from your local\ncheckout: `js-coverage` defaults to the current git HEAD, `trigger-test-run`\ndefaults to HEAD for the deployment and to the merge-base with the origin\ndefault branch for the base. On `main` with a clean tree those collapse to a\nsingle commit, which is exactly what you want — no diff, a head-only run, and\na union in Step 7 that the API will actually accept.\n\nOff `main` this breaks in ways that are tedious to unpick: union coverage is\nrejected unless every run executed the _exact same_ commit, and a PR's merge\ncommit is recomputed whenever its base branch moves, so a run triggered\nearlier against a since-advanced base no longer unions with a new one.\n\nSo: check out `main`, pull, and make sure `git status` is clean before Step 1.\n\nThen confirm there is actually a test run to work from:\n\n```bash\nmeticulous agent test-run-for-commit\n```\n\nKeep the id it prints — Step 7 falls back to it. If it reports **\"No test run\nfound for commit …\"**, stop and report that to the user; you cannot baseline\nwithout it. The most common cause is the CLI pointing at the wrong project, so\nsuggest they check with `meticulous auth get-project`. `meticulous auth\nset-project` only applies for OAuth tokens; API tokens are bound to one\nproject (so `set-project` fails), and injected credentials leave no local\ntoken to select — in those cases the fix is a different credential, not\n`set-project`.\n\n## Step 1 — Baseline coverage\n\n```bash\nmeticulous agent js-coverage --includeAllFiles --includeCoveragePercentage \\\n  > \u002Ftmp\u002Fbaseline-coverage.tsv\n```\n\nIf this reports **\"No test run found for commit …\"** (it shouldn't, if the\ncheck above passed), stop and report to the user. Do not work around it by\nbaselining against some other commit's run: Step 7's union requires your new\nrun and the baseline to have executed the same commit.\n\nThe base run's sessions often haven't all been replayed yet, which understates\nits coverage — if `js-coverage` says so, run `meticulous agent\ncomplete-base-run` (it waits by default until nothing more can be scheduled;\nit can take a while, so check back or re-run rather than assuming it hung),\nthen re-run `js-coverage`. Don't expect `unexecutedSessionCount` to always\nreach `0` — some sessions can be permanently unobtainable, and `js-coverage`\ntolerates a small share of those rather than refusing forever.\n\n## Step 2 — Separate dead code from real targets\n\nYou are looking for two different things in this file, and it helps to keep\nthem apart:\n\n- **`.meticulousignore` candidates** — files that are _uniformly_ at 0%\n  across a whole directory, which suggests they never execute in a browser at\n  all.\n- **Coverage targets** — files a real user flow could reach but no recorded\n  session happens to. These are **not only the 0% files**. A file at 12% or\n  40% usually means one path through it runs and the rest doesn't, and those\n  partial files are often the cheapest wins: the module already loads, so a\n  single extra interaction can light up a large block. Sort ascending by\n  percentage and work up from the bottom, rather than stopping at 0%. Some\n  0%\u002Flow files are gated behind a feature toggle that's off by default rather\n  than a UI path nobody's driven — check the toggle registry and the file's\n  gating condition before assuming it needs a brand-new flow, since flipping\n  the toggle on locally can turn a dead-looking file into an easy target.\n\nStart with the ignore candidates, since they shrink the list. Break the\n0%-coverage files down by top-level directory, so you are reasoning about\ngroups rather than 100s of individual files. The exact command depends on how\nthe repo is laid out — a monorepo wants the first two path segments, a\nsingle-app repo wants something deeper. For example, in a\n`packages\u002F\u003Cname>\u002F…` monorepo:\n\n```bash\n# example only — adjust the segment depth to this repo's layout\nawk -F'\\t' 'NR>1 && $2==\"0.0\" {split($1,a,\"\u002F\"); print a[1]\"\u002F\"a[2]}' \\\n  \u002Ftmp\u002Fbaseline-coverage.tsv | sort | uniq -c | sort -rn\n```\n\nA directory where **every single file** is at 0% (not just some) is a strong\nsignal it never ships to the browser (backend, CLI tooling, docs, e2e test\nharness). Those are your `.meticulousignore` candidates.\n\nThere is a second, stronger signal that doesn't depend on coverage data at\nall and catches individual dead files scattered inside an otherwise-live\ndirectory, which the uniformly-0% heuristic above misses entirely. For any\nfile sitting at 0%, grep the codebase for its actual exported symbol — not\njust its filename:\n\n```bash\n# example only — adjust the source root and extensions to this repo\ngrep -rn \"theActualExportedName\" \u003Csrc-root> --include=\"*.ts\" --include=\"*.tsx\"\n```\n\nIf the only match is the file's own definition, nothing imports it, so no\nsession — however comprehensive — can ever execute it. That is proof of\nunreachability, not an inference from silence, and it is worth checking even\nwhen you already have a directory-level rule elsewhere in this file: dead\nexports accumulate inside packages that are otherwise very much alive.\n\nSome categories are **safe exclusions almost everywhere** and are worth\nproposing without further tracing, provided the coverage data agrees they are\nuniformly 0%:\n\n- test files and their directories — `__tests__\u002F`, `__mocks__\u002F`, `*.test.*`,\n  `*.spec.*`\n- Storybook — `*.stories.*`, `__stories__\u002F`, `.storybook\u002F`\n- test\u002Fmock harness directories — `testing\u002F`, `mocks\u002F`, fixtures\n- build, lint and codegen config executed only by Node — `*.config.*`,\n  `setupTests.*`, scripts directories\n\nNote that some of these may already be outside the coverage report entirely;\ncheck the baseline before adding a rule that does nothing.\n\nBe suspicious of a directory showing 0% everywhere if you _know_ it is bundled\ninto the frontend (a shared component\u002Futils library the main app imports).\nThat pattern is more likely a source-map\u002Fpath-attribution gap than genuinely\ndead code — leave it out of the ignore list and flag it as unresolved.\n\nNow pick the coverage targets. Filter the generated\u002Fconfig noise out of the\nreal app package first, then order what's left by how little of it runs —\nkeeping the partially-covered files in, not just the 0% ones. The patterns are\nrepo-specific; inspect the actual paths in your baseline rather than copying\nthis verbatim:\n\n```bash\n# example only — derive the patterns from the paths this repo actually has\ngrep -v \"\u002F__tests__\u002F\\|\\.test\\.\\|\\.stories\\.\\|\u002Ftesting\u002F\\|\u002Fmock\" \\\n  \u002Ftmp\u002Fbaseline-coverage.tsv | sort -t$'\\t' -k2 -g > \u002Ftmp\u002Fcandidates.tsv\n```\n\nGroup the candidates by feature area rather than picking the single worst\nfiles: one recorded flow usually moves a whole cluster of related files at\nonce, so a directory sitting at 5-20% across a dozen files is a better target\nthan an isolated 0% file behind an obscure branch.\n\n## Step 3 — Trace, don't guess\n\nFor each candidate file, find its actual caller(s) — for example:\n\n```bash\n# example only — adjust the source root and extensions to this repo\ngrep -rln \"\u003CExportedThing\" \u003Csrc-root> --include=\"*.tsx\" | grep -v test\n```\n\nRead the caller. Confirm:\n\n- It's reachable via a simple, describable UI action (a specific button, a\n  specific menu item) — not buried behind a feature flag, a disabled config\n  (e.g. billing\u002FSSO toggles that are off in this environment), or a\n  conditional branch that only fires for certain object types.\n- If the target is a **hook**, check every branch that calls it. Hooks are\n  often called conditionally — one branch might route through a completely\n  different mechanism (a plain router `\u003CLink>` instead of the app's own\n  navigation hook, a side-panel open instead of a full navigate). Confirm\n  which branch your candidate action actually hits.\n\n## Step 4 — Drive the flow\n\n**Any of these three drivers records correctly** — verified against two\ndifferent apps:\n\n- **Claude in Chrome** — the only one that drives the user's own signed-in\n  Chrome profile, so an authenticated app needs no login flow. Its input path\n  is also the one that wedges (see below), so verify early and be ready to\n  switch.\n- **Playwright** and **`agent-browser`** — CDP-launched browsers, faster and\n  more scriptable, and reliable in every test here. The trade-off is a fresh\n  profile each time, so you have to sign in. `agent-browser` additionally\n  refuses a click when the target is covered by another element, naming the\n  covering node, which catches a class of silent mis-click the other two will\n  happily perform.\n\n**Set a realistic viewport** whichever you pick. The CDP browsers default to\nsomething small (around 1280px wide) where a real Chrome window is often twice\nthat. Responsive layouts render different components at different widths, so\nthe default viewport can quietly cover different code — or hide the control\nyou were aiming for.\n\nWhat matters far more than the choice of tool is the one rule below.\n\n**Only trusted events are recorded.** The recorder ignores anything\nsynthesised in page JS, so `element.click()`, assigning `input.value`, or\ndispatching your own events all drive the app convincingly and record\n_nothing_. The page looks right, the session comes back empty. Drive\neverything through the tool's real input actions.\n\nWhen something doesn't take effect, find out which half is broken before\nchanging tactics — install a capturing probe and repeat the action:\n\n```js\nwindow.__ev = [];\n[\"pointerdown\", \"click\", \"keydown\", \"change\"].forEach((t) =>\n  window.addEventListener(\n    t,\n    (e) => window.__ev.push({ t, trusted: e.isTrusted }),\n    true,\n  ),\n);\n```\n\n- nothing captured → your input never reached the page; a different selector\n  or coordinate won't help (see the wedged-extension note below)\n- captured but `trusted: false` → it reached the page but will not be\n  recorded; you are synthesising somewhere\n\n### Three ways an action records as nothing\n\nAll three look like success in the browser, which is what makes them\nexpensive — you find out from the coverage numbers, long after the fact.\n\n**Native `\u003Cselect>`s.** Setting the value through a form-fill action, or\nassigning it in JS, fires a `change` with `isTrusted: false` — the app reacts\nand the value visibly updates, so it looks like it worked, but the recorder\nignores it and the sort\u002Ffilter never happens on replay. Clicking the\n`\u003Cselect>` is no good either: that opens an OS-level popup the driver can't\nsee. What works is to focus the element and press the **first letter of the\noption's visible text** (`\"p\"` → \"Priority\"), which yields a trusted `keydown`\n_and_ a trusted `change`. Repeat the letter to cycle options sharing an\ninitial. Verify with a `change` listener reading `e.isTrusted` — the value\nupdates either way, so the value alone tells you nothing.\n\n**Modifier shortcuts.** Replay reproduces a modifier only if a _discrete_\nmodifier keydown was recorded and is still held. Some drivers send one only\nfor the base key: Claude in Chrome's `cmd+k` and `agent-browser`'s\n`press Alt+ArrowRight` both record a single keydown with the modifier flag set\nand no separate modifier press, so on replay the flag is cleared and the\nhandler body never runs — while working perfectly live. Playwright's\n`press('Alt+ArrowRight')` does record the discrete press and replays correctly\n(verified by coverage).\n\nSo: prefer the equivalent _click_ target where one exists, and if you must use\na chord, drive it with Playwright and confirm from coverage afterwards. If the\nline holding `if (… && event.metaKey)` is covered but the body is not, the\nkeydown was delivered and the condition evaluated false — that is this.\n\n**Double-clicks.** A double-click may be recorded as a single click, in which\ncase the replay never fires `onDoubleClick` and every later event in that\nsession targets UI that never opened — so coverage _drops_. Check the recorded\nevent count looks like two press\u002Frelease pairs, and treat any\ndouble-click-only feature as suspect until coverage confirms it.\n\n### Some interactions may not survive agent-driven recording at all\n\nOccasionally an interaction records cleanly and visibly works live, but the\nhandler it's meant to trigger never fires on replay — with no error, and the\naffected file sitting at _exactly_ its baseline percentage in Step 7's union.\nOne case seen: typing a value into an input inside a dropdown's own\nportal-rendered content (a filter chip, a view rename behind a \"...\" menu).\nPlain clicks in the same portal, and the identical type-then-Enter sequence\non an input in the main page tree, both replay fine — so this is specific to\nkeyboard\u002Ftext input inside a portal, not a driver issue.\n\nDon't assume a typed-value target worked just because the live interaction\ndid — check Step 7's diff. If a target only reproduces through this kind of\ninteraction and won't move, that's a shortcoming of agent-driven recording\nfor this flow: report it as unresolved and suggest a human drive that one\nflow manually, rather than continuing to pad the list with retries.\n\n### `claude-in-chrome` specifics\n\n- **Screenshots are downscaled** (~0.6x), so they are not CSS pixels. Click\n  coordinates read off the screenshot, or scale a `getBoundingClientRect()`\n  centre by `screenshotWidth \u002F window.innerWidth`. Re-derive after any resize,\n  and re-screenshot rather than reusing coordinates from an earlier page —\n  layout shifts, and a stale coordinate can land on the wrong element and\n  record an interaction you did not intend.\n- **Re-acquire a `ref` immediately before clicking it**, and never reuse one\n  across pages or tabs. Don't predict a number — `read_page` does not emit\n  them in order. Refs and coordinates are equally reliable; pick whichever is\n  convenient. This is not unique to `claude-in-chrome` — `agent-browser`'s\n  `snapshot` refs go stale the same way, and reusing one silently clicks\n  whatever now occupies that ref, not what you intended. It cost a real\n  mistake once: a stale ref landed on a table's \"Add New\" affordance and\n  created a blank record. Take a fresh snapshot immediately before every\n  click when the DOM might have changed, and treat an unexpected\n  page\u002Frecord-count change right after a click as a sign a ref just misfired,\n  not as an unrelated bug — clean up whatever it created before continuing.\n- **Input delivery wedges intermittently.** Every click and keystroke reports\n  success, reads keep working, and nothing reaches the page. Refs and\n  coordinates die together, so this is never a selector problem — and it is\n  not cleared by a new tab, a fresh navigation, waiting, or retrying. A full\n  Chrome restart helps but does not durably fix it. **Check the probe after\n  your first interaction, before driving a whole flow**, and if input isn't\n  landing switch to Playwright or `agent-browser` rather than switching\n  selector method. Both stayed reliable throughout, including on the same\n  page at the same moment that Claude in Chrome was dead.\n\n### Drive it, then verify\n\n1. Navigate to the target URL and drive the real action.\n2. Confirm it worked against the DOM (e.g.\n   `document.body.innerText.includes(...)`), not just a screenshot — a\n   tooltip appearing can look like success. Check you are still on the page\n   you think you are: an app that has quietly redirected you to a login screen\n   will absorb blind coordinate clicks into empty space and hand you a session\n   with zero events.\n3. **Close the tab.**\n\nThen sanity-check the recording, before you trigger anything. Wait ~10s after\nclosing the tab, so the session is complete, and check that it captured\nsomething:\n\n```bash\nmeticulous agent sessions --limit 10 --excludeSyntheticSessions \\\n  --includeDurationSeconds --includeNumberUserEvents \\\n  --includeNumberUrlsVisited --includeStartUrl --includeAbandonedReason\n```\n\nRead the row you just produced:\n\n- **no row at all** — nothing was uploaded yet; wait a little longer before\n  concluding the recording failed\n- **`numberUserEvents` of 0** — the recorder saw no user input. Your clicks\n  were not reaching the page, or were synthesised rather than trusted; go back\n  to the input-delivery probe above. Replaying this session is pointless.\n- **`numberUrlsVisited` of 1 when you navigated several times** — the later\n  pages did not make it into _this_ session. They either landed in their own\n  sessions (fine, collect those ids too) or were swallowed as an unreplayable\n  tail (see the hard-navigation note in Step 5).\n- **`durationSeconds` over 300** — everything past the 5-minute mark will be\n  silently trimmed on replay (Step 5). Re-record the overflowing part as its\n  own session rather than hoping it survives.\n- **populated `abandonedReason`** — the recorder gave up on the session (see\n  the 10-minute cap in Step 5); it is not worth replaying.\n- **`startUrl` that is not the page you drove** — the navigation you cared\n  about belongs to a different session than you assumed.\n\n**A session is only complete once its tab is closed.** While the tab is open\nthe row reflects only the chunks uploaded so far, so a low or zero\n`numberUserEvents` there means \"not flushed yet\", **not** \"the recording\nfailed\". A session measured for this skill read **0** events with the tab open\nand **12** once it was closed. Judging it early nearly caused a perfectly good\nrecording to be re-driven from scratch.\n\nEvents upload on a short interval (a few seconds), but anything still\nunflushed at unload is only stashed in `sessionStorage` and re-sent on a\n_later page load to the same origin_ — so a session's tail can be delayed\nuntil the next visit. Prefer navigating away over hard-closing the browser,\nand never judge a recording immediately.\n\n**Recording several targets in one sitting? Run this check after _each_ one**,\nnot just once at the end. Checking only at the end makes it impossible to\ntell which action lost a session, and you'll have to re-drive all of them\njust to find out which one needs redoing.\n\n## Step 5 — Session-time budget and close discipline\n\n- **Cloud replays cap at 5 minutes of session time.** Everything recorded\n  after that is silently trimmed from the replay — the run still succeeds and\n  reports itself accurate, so the loss is invisible unless you look for it\n  (snapshot routes stop early; far fewer allowed events than the session has\n  clicks). Don't leave this to feel: `agent sessions\n--includeDurationSeconds` gives you the number, and anything over 300 is\n  losing its tail.\n- **Your interaction pace eats this budget.** Each find\u002Fclick\u002Fverify round\n  trip is 10-60s of _recorded_ session time. Plan the flow completely before\n  opening the tab, batch your actions into as few round trips as possible\n  with short waits between them, verify from the recording afterwards rather\n  than mid-flow, and aim to stay under 4 minutes. Several short sessions beat\n  one long sweep.\n- **Direct URL navigation is a legitimate fast path** to each target page\n  and replays fine — prefer it over slow click-paths. But hard navigations\n  split the recording into multiple sessions, so collect every resulting\n  session id afterwards and pass them all to `--sessionIds` (Step 6). Mostly\n  this is fine, each piece staying under the replay cap. The trap is\n  navigating again too quickly: a page reached a second or two after the\n  previous one gets appended as the _tail_ of that session instead of\n  starting its own, and tails frequently do not replay — so the page renders\n  perfectly while you drive it and still contributes no coverage. Give each\n  page you actually care about its own dwell time (~10s) before moving on,\n  and check in Step 6 that it shows up as a `startUrl` in its own right. The\n  same caution applies to a plain `\u003Cform>` submit with no wired `onSubmit`\n  handler — it triggers a real browser reload rather than an SPA transition,\n  and can just as easily drop the just-recorded, unflushed session if you\n  navigate on immediately afterward.\n- **Recorder limits:** a 10-minute hard cap on tab-open time marks the whole\n  session \"abandoned\"; uploads flush on a 5s interval — wait ~6-8s after the\n  last interaction before closing the tab.\n\n## Step 6 — Collect the session ids and trigger the test run\n\nFirst list what you actually recorded, newest first:\n\n```bash\nmeticulous agent sessions --limit 20 --excludeSyntheticSessions \\\n  --includeDurationSeconds --includeNumberUserEvents \\\n  --includeNumberUrlsVisited --includeStartUrl\n```\n\nSkip any row with `numberUserEvents` of 0 — it will replay as nothing and\nonly dilutes the run. Note any row with `durationSeconds` over 300: it will\nreplay, but only its first five minutes, so treat coverage from its tail as\nabsent rather than assuming the whole flow ran.\n\nIdentify your sessions by recorded-at time and `startUrl`. Be careful here:\nother people — and other apps pointed at the same project — record too, so\nnever assume the newest N rows are yours. `--recordedSince` and\n`--visitedUrlFilter` are the quickest way to narrow it down when the list is\nbusy.\n\nExpect **more sessions than pages you drove**: a hard navigation usually ends\none session and starts another, so a five-page sweep can produce five ids.\nCollect all of them. A page whose URL never shows up as a `startUrl` was\nprobably swallowed as the tail of the previous session and will not replay —\nre-record it on its own if you need it covered.\n\nThen trigger:\n\n```bash\nmeticulous agent trigger-test-run --sessionIds \"\u003Cid1>,\u003Cid2>,...\"\n```\n\n## Step 7 — Compare with a union, not a raw diff\n\n```bash\nmeticulous agent js-coverage --headPlusTestRunIds \"\u003CnewRunId>\" \\\n  --includeAllFiles --includeCoveragePercentage > \u002Ftmp\u002Fcombined-coverage.tsv\n```\n\nThis unions your new run into the baseline run resolved from HEAD — the same\nrun Step 1 used — so it is baseline coverage **plus** your new sessions'\ncoverage. Commit resolution skips runs over an explicit session set, so the\nrun you just triggered won't be picked as the baseline. The union is needed\nbecause `--sessionIds` _replaced_ the selected set for that run rather than\nadding to it, so your run on its own covers far less than the baseline and a\nraw diff would read as mass regressions. Diff the union against\n`\u002Ftmp\u002Fbaseline-coverage.tsv`: you should see zero regressions, and only the\nfiles your new flow touched improve.\n\nPass only your new run — a run cannot be unioned with itself.\n\nTwo signs HEAD resolved to something other than the golden-set run: it is\n**rejected** with \"is the run being queried\", or _every_ file has dropped\n(which means you unioned into another narrow session set, not a regression).\nThe most likely cause is a pinned-session run predating the skip. Either way,\nname both sides explicitly, using the baseline id from Step 0:\n\n```bash\nmeticulous agent js-coverage --testRunIds \"\u003CbaselineRunId>,\u003CnewRunId>\" \\\n  --includeAllFiles --includeCoveragePercentage > \u002Ftmp\u002Fcombined-coverage.tsv\n```\n\nIf the union is rejected because the runs executed different commits, that is\nthe `main`\u002Fclean-tree precondition biting — see the top of this skill.\n\n## Step 8 — Verify and report\n\nFor each traced target file: did coverage move? For each presumed-dead file\n(the `.meticulousignore` candidates from Step 2): did it stay at 0% in the\nunion comparison (confirming it's genuinely unreachable)?\n\nIf a well-traced target didn't move, don't assume the recording failed —\nre-check the session first (Step 4: does it exist, did it capture user events\nand URL visits, is it abandoned, is its `startUrl` the page you drove?), then\nre-check whether the click actually goes through the file you expected\n(Step 3) rather than a sibling\u002Fparent component.\nReport honestly if a target remains unresolved; don't claim success without\nthe coverage number to back it.\n\nWhen you report the gains, be clear about what they are not yet: the test run\nproves the coverage is reachable, but the project's own coverage figure will\nonly improve once the next session selection picks these new sessions up into\nthe selected set. Until then nothing changes for recurring runs.\n\n## Step 9 — Open the `.meticulousignore` PR\n\nThe second deliverable. Branch, commit the `.meticulousignore` change, and\nopen a PR.\n\n**Only propose paths you are confident are not coverable at all.** The bar is\n\"no session could ever execute this\", not \"I didn't get to it today\". Prefer\ndirectory-level rules over long lists of individual files — a directory rule\nstays correct as files are added, whereas a file list silently goes stale. In\npractice most entries come from the safe-exclusion categories in Step 2 plus\nwhatever whole non-browser packages the baseline showed at a uniform 0%.\n\nConfirm before you commit: every path you are about to ignore stayed at 0% in\nthe Step 7 union. A path your own new sessions just covered obviously does not\nbelong in the ignore list, and that check catches it.\n\nA common structure is \"ignore everything, then un-ignore what does run in the\nbrowser\" — the pattern Meticulous's own monorepo uses:\n\n```\n# Ignore everything except packages that are executed in the\n# browser and have meaningful frontend coverage.\n\npackages\u002F*\n!packages\u002F\u003Cfrontend-app>\u002F\n!packages\u002F\u003Cfrontend-app>\u002F**\n!packages\u002F\u003Cshared-ui-lib>\u002F\n!packages\u002F\u003Cshared-ui-lib>\u002F**\n```\n\nNote that `.meticulousignore` follows gitignore semantics, so a file cannot be\nre-included once its parent directory is excluded — un-ignore the directories\n(`!some\u002Fdir\u002F**\u002F`) as well as the files.\n\nIn the PR description, give the reasoning for each rule — _why_ this code\ncannot run in a browser (it's a Node-only build script, a test harness, a\nbackend package). The baseline can only show you that something is at 0%\ntoday, which is never proof it is unreachable, so the justification has to\ncome from what the code actually is. Also call out explicitly:\n\n- anything you left **out** of the ignore list despite low coverage because\n  you suspect a source-map\u002Fattribution gap rather than dead code (Step 2)\n- anything unreachable only because of environment or feature-flag config\n  rather than structurally — that is the reviewer's judgement call, not\n  yours, so flag it instead of silently ignoring it\n\n## Reference\n\n`references\u002Fworked-example.md` runs the whole skill against a small app\n(kanban-demo), with the real coverage deltas. Worth reading for calibration:\nit has no 0% files at all, one target that gained coverage and one that\nrecorded cleanly and then failed to replay — and it shows how to tell the\ndifference from executed ranges.\n","# Increase coverage for a Meticulous project\n\n> Run `meticulous-cli-update` first if you haven't already this conversation\n> (it also covers authentication and project selection).\n\n## What you deliver\n\nTwo separate outputs, both expected — neither substitutes for the other:\n\n1. **One or more test runs from newly recorded sessions that provably extend\n   coverage.** \"Provably\" means a comparison against the baseline naming the\n   files whose coverage went up, and by how much (Step 7). If a target turns\n   out not to be coverable, say so rather than padding the list.\n2. **A PR proposing `.meticulousignore` changes.** Only include paths — most\n   often whole directories — that you are _confident_ are not coverable at\n   all. Anything you merely failed to reach in one sitting does not belong\n   there; leave it out and mention it in the PR description instead.\n\n## Before you start: run this on `main`, with a clean tree\n\nEvery command below relies on the CLI resolving things from your local\ncheckout: `js-coverage` defaults to the current git HEAD, `trigger-test-run`\ndefaults to HEAD for the deployment and to the merge-base with the origin\ndefault branch for the base. On `main` with a clean tree those collapse to a\nsingle commit, which is exactly what you want — no diff, a head-only run, and\na union in Step 7 that the API will actually accept.\n\nOff `main` this breaks in ways that are tedious to unpick: union coverage is\nrejected unless every run executed the _exact same_ commit, and a PR's merge\ncommit is recomputed whenever its base branch moves, so a run triggered\nearlier against a since-advanced base no longer unions with a new one.\n\nSo: check out `main`, pull, and make sure `git status` is clean before Step 1.\n\nThen confirm there is actually a test run to work from:\n\n```bash\nmeticulous agent test-run-for-commit\n```\n\nKeep the id it prints — Step 7 falls back to it. If it reports **\"No test run\nfound for commit …\"**, stop and report that to the user; you cannot baseline\nwithout it. The most common cause is the CLI pointing at the wrong project, so\nsuggest they check with `meticulous auth get-project`. `meticulous auth\nset-project` only applies for OAuth tokens; API tokens are bound to one\nproject (so `set-project` fails), and injected credentials leave no local\ntoken to select — in those cases the fix is a different credential, not\n`set-project`.\n\n## Step 1 — Baseline coverage\n\n```bash\nmeticulous agent js-coverage --includeAllFiles --includeCoveragePercentage \\\n  > \u002Ftmp\u002Fbaseline-coverage.tsv\n```\n\nIf this reports **\"No test run found for commit …\"** (it shouldn't, if the\ncheck above passed), stop and report to the user. Do not work around it by\nbaselining against some other commit's run: Step 7's union requires your new\nrun and the baseline to have executed the same commit.\n\nThe base run's sessions often haven't all been replayed yet, which understates\nits coverage — if `js-coverage` says so, run `meticulous agent\ncomplete-base-run` (it waits by default until nothing more can be scheduled;\nit can take a while, so check back or re-run rather than assuming it hung),\nthen re-run `js-coverage`. Don't expect `unexecutedSessionCount` to always\nreach `0` — some sessions can be permanently unobtainable, and `js-coverage`\ntolerates a small share of those rather than refusing forever.\n\n## Step 2 — Separate dead code from real targets\n\nYou are looking for two different things in this file, and it helps to keep\nthem apart:\n\n- **`.meticulousignore` candidates** — files that are _uniformly_ at 0%\n  across a whole directory, which suggests they never execute in a browser at\n  all.\n- **Coverage targets** — files a real user flow could reach but no recorded\n  session happens to. These are **not only the 0% files**. A file at 12% or\n  40% usually means one path through it runs and the rest doesn't, and those\n  partial files are often the cheapest wins: the module already loads, so a\n  single extra interaction can light up a large block. Sort ascending by\n  percentage and work up from the bottom, rather than stopping at 0%. Some\n  0%\u002Flow files are gated behind a feature toggle that's off by default rather\n  than a UI path nobody's driven — check the toggle registry and the file's\n  gating condition before assuming it needs a brand-new flow, since flipping\n  the toggle on locally can turn a dead-looking file into an easy target.\n\nStart with the ignore candidates, since they shrink the list. Break the\n0%-coverage files down by top-level directory, so you are reasoning about\ngroups rather than 100s of individual files. The exact command depends on how\nthe repo is laid out — a monorepo wants the first two path segments, a\nsingle-app repo wants something deeper. For example, in a\n`packages\u002F\u003Cname>\u002F…` monorepo:\n\n```bash\n# example only — adjust the segment depth to this repo's layout\nawk -F'\\t' 'NR>1 && $2==\"0.0\" {split($1,a,\"\u002F\"); print a[1]\"\u002F\"a[2]}' \\\n  \u002Ftmp\u002Fbaseline-coverage.tsv | sort | uniq -c | sort -rn\n```\n\nA directory where **every single file** is at 0% (not just some) is a strong\nsignal it never ships to the browser (backend, CLI tooling, docs, e2e test\nharness). Those are your `.meticulousignore` candidates.\n\nThere is a second, stronger signal that doesn't depend on coverage data at\nall and catches individual dead files scattered inside an otherwise-live\ndirectory, which the uniformly-0% heuristic above misses entirely. For any\nfile sitting at 0%, grep the codebase for its actual exported symbol — not\njust its filename:\n\n```bash\n# example only — adjust the source root and extensions to this repo\ngrep -rn \"theActualExportedName\" \u003Csrc-root> --include=\"*.ts\" --include=\"*.tsx\"\n```\n\nIf the only match is the file's own definition, nothing imports it, so no\nsession — however comprehensive — can ever execute it. That is proof of\nunreachability, not an inference from silence, and it is worth checking even\nwhen you already have a directory-level rule elsewhere in this file: dead\nexports accumulate inside packages that are otherwise very much alive.\n\nSome categories are **safe exclusions almost everywhere** and are worth\nproposing without further tracing, provided the coverage data agrees they are\nuniformly 0%:\n\n- test files and their directories — `__tests__\u002F`, `__mocks__\u002F`, `*.test.*`,\n  `*.spec.*`\n- Storybook — `*.stories.*`, `__stories__\u002F`, `.storybook\u002F`\n- test\u002Fmock harness directories — `testing\u002F`, `mocks\u002F`, fixtures\n- build, lint and codegen config executed only by Node — `*.config.*`,\n  `setupTests.*`, scripts directories\n\nNote that some of these may already be outside the coverage report entirely;\ncheck the baseline before adding a rule that does nothing.\n\nBe suspicious of a directory showing 0% everywhere if you _know_ it is bundled\ninto the frontend (a shared component\u002Futils library the main app imports).\nThat pattern is more likely a source-map\u002Fpath-attribution gap than genuinely\ndead code — leave it out of the ignore list and flag it as unresolved.\n\nNow pick the coverage targets. Filter the generated\u002Fconfig noise out of the\nreal app package first, then order what's left by how little of it runs —\nkeeping the partially-covered files in, not just the 0% ones. The patterns are\nrepo-specific; inspect the actual paths in your baseline rather than copying\nthis verbatim:\n\n```bash\n# example only — derive the patterns from the paths this repo actually has\ngrep -v \"\u002F__tests__\u002F\\|\\.test\\.\\|\\.stories\\.\\|\u002Ftesting\u002F\\|\u002Fmock\" \\\n  \u002Ftmp\u002Fbaseline-coverage.tsv | sort -t$'\\t' -k2 -g > \u002Ftmp\u002Fcandidates.tsv\n```\n\nGroup the candidates by feature area rather than picking the single worst\nfiles: one recorded flow usually moves a whole cluster of related files at\nonce, so a directory sitting at 5-20% across a dozen files is a better target\nthan an isolated 0% file behind an obscure branch.\n\n## Step 3 — Trace, don't guess\n\nFor each candidate file, find its actual caller(s) — for example:\n\n```bash\n# example only — adjust the source root and extensions to this repo\ngrep -rln \"\u003CExportedThing\" \u003Csrc-root> --include=\"*.tsx\" | grep -v test\n```\n\nRead the caller. Confirm:\n\n- It's reachable via a simple, describable UI action (a specific button, a\n  specific menu item) — not buried behind a feature flag, a disabled config\n  (e.g. billing\u002FSSO toggles that are off in this environment), or a\n  conditional branch that only fires for certain object types.\n- If the target is a **hook**, check every branch that calls it. Hooks are\n  often called conditionally — one branch might route through a completely\n  different mechanism (a plain router `\u003CLink>` instead of the app's own\n  navigation hook, a side-panel open instead of a full navigate). Confirm\n  which branch your candidate action actually hits.\n\n## Step 4 — Drive the flow\n\n**Any of these three drivers records correctly** — verified against two\ndifferent apps:\n\n- **Claude in Chrome** — the only one that drives the user's own signed-in\n  Chrome profile, so an authenticated app needs no login flow. Its input path\n  is also the one that wedges (see below), so verify early and be ready to\n  switch.\n- **Playwright** and **`agent-browser`** — CDP-launched browsers, faster and\n  more scriptable, and reliable in every test here. The trade-off is a fresh\n  profile each time, so you have to sign in. `agent-browser` additionally\n  refuses a click when the target is covered by another element, naming the\n  covering node, which catches a class of silent mis-click the other two will\n  happily perform.\n\n**Set a realistic viewport** whichever you pick. The CDP browsers default to\nsomething small (around 1280px wide) where a real Chrome window is often twice\nthat. Responsive layouts render different components at different widths, so\nthe default viewport can quietly cover different code — or hide the control\nyou were aiming for.\n\nWhat matters far more than the choice of tool is the one rule below.\n\n**Only trusted events are recorded.** The recorder ignores anything\nsynthesised in page JS, so `element.click()`, assigning `input.value`, or\ndispatching your own events all drive the app convincingly and record\n_nothing_. The page looks right, the session comes back empty. Drive\neverything through the tool's real input actions.\n\nWhen something doesn't take effect, find out which half is broken before\nchanging tactics — install a capturing probe and repeat the action:\n\n```js\nwindow.__ev = [];\n[\"pointerdown\", \"click\", \"keydown\", \"change\"].forEach((t) =>\n  window.addEventListener(\n    t,\n    (e) => window.__ev.push({ t, trusted: e.isTrusted }),\n    true,\n  ),\n);\n```\n\n- nothing captured → your input never reached the page; a different selector\n  or coordinate won't help (see the wedged-extension note below)\n- captured but `trusted: false` → it reached the page but will not be\n  recorded; you are synthesising somewhere\n\n### Three ways an action records as nothing\n\nAll three look like success in the browser, which is what makes them\nexpensive — you find out from the coverage numbers, long after the fact.\n\n**Native `\u003Cselect>`s.** Setting the value through a form-fill action, or\nassigning it in JS, fires a `change` with `isTrusted: false` — the app reacts\nand the value visibly updates, so it looks like it worked, but the recorder\nignores it and the sort\u002Ffilter never happens on replay. Clicking the\n`\u003Cselect>` is no good either: that opens an OS-level popup the driver can't\nsee. What works is to focus the element and press the **first letter of the\noption's visible text** (`\"p\"` → \"Priority\"), which yields a trusted `keydown`\n_and_ a trusted `change`. Repeat the letter to cycle options sharing an\ninitial. Verify with a `change` listener reading `e.isTrusted` — the value\nupdates either way, so the value alone tells you nothing.\n\n**Modifier shortcuts.** Replay reproduces a modifier only if a _discrete_\nmodifier keydown was recorded and is still held. Some drivers send one only\nfor the base key: Claude in Chrome's `cmd+k` and `agent-browser`'s\n`press Alt+ArrowRight` both record a single keydown with the modifier flag set\nand no separate modifier press, so on replay the flag is cleared and the\nhandler body never runs — while working perfectly live. Playwright's\n`press('Alt+ArrowRight')` does record the discrete press and replays correctly\n(verified by coverage).\n\nSo: prefer the equivalent _click_ target where one exists, and if you must use\na chord, drive it with Playwright and confirm from coverage afterwards. If the\nline holding `if (… && event.metaKey)` is covered but the body is not, the\nkeydown was delivered and the condition evaluated false — that is this.\n\n**Double-clicks.** A double-click may be recorded as a single click, in which\ncase the replay never fires `onDoubleClick` and every later event in that\nsession targets UI that never opened — so coverage _drops_. Check the recorded\nevent count looks like two press\u002Frelease pairs, and treat any\ndouble-click-only feature as suspect until coverage confirms it.\n\n### Some interactions may not survive agent-driven recording at all\n\nOccasionally an interaction records cleanly and visibly works live, but the\nhandler it's meant to trigger never fires on replay — with no error, and the\naffected file sitting at _exactly_ its baseline percentage in Step 7's union.\nOne case seen: typing a value into an input inside a dropdown's own\nportal-rendered content (a filter chip, a view rename behind a \"...\" menu).\nPlain clicks in the same portal, and the identical type-then-Enter sequence\non an input in the main page tree, both replay fine — so this is specific to\nkeyboard\u002Ftext input inside a portal, not a driver issue.\n\nDon't assume a typed-value target worked just because the live interaction\ndid — check Step 7's diff. If a target only reproduces through this kind of\ninteraction and won't move, that's a shortcoming of agent-driven recording\nfor this flow: report it as unresolved and suggest a human drive that one\nflow manually, rather than continuing to pad the list with retries.\n\n### `claude-in-chrome` specifics\n\n- **Screenshots are downscaled** (~0.6x), so they are not CSS pixels. Click\n  coordinates read off the screenshot, or scale a `getBoundingClientRect()`\n  centre by `screenshotWidth \u002F window.innerWidth`. Re-derive after any resize,\n  and re-screenshot rather than reusing coordinates from an earlier page —\n  layout shifts, and a stale coordinate can land on the wrong element and\n  record an interaction you did not intend.\n- **Re-acquire a `ref` immediately before clicking it**, and never reuse one\n  across pages or tabs. Don't predict a number — `read_page` does not emit\n  them in order. Refs and coordinates are equally reliable; pick whichever is\n  convenient. This is not unique to `claude-in-chrome` — `agent-browser`'s\n  `snapshot` refs go stale the same way, and reusing one silently clicks\n  whatever now occupies that ref, not what you intended. It cost a real\n  mistake once: a stale ref landed on a table's \"Add New\" affordance and\n  created a blank record. Take a fresh snapshot immediately before every\n  click when the DOM might have changed, and treat an unexpected\n  page\u002Frecord-count change right after a click as a sign a ref just misfired,\n  not as an unrelated bug — clean up whatever it created before continuing.\n- **Input delivery wedges intermittently.** Every click and keystroke reports\n  success, reads keep working, and nothing reaches the page. Refs and\n  coordinates die together, so this is never a selector problem — and it is\n  not cleared by a new tab, a fresh navigation, waiting, or retrying. A full\n  Chrome restart helps but does not durably fix it. **Check the probe after\n  your first interaction, before driving a whole flow**, and if input isn't\n  landing switch to Playwright or `agent-browser` rather than switching\n  selector method. Both stayed reliable throughout, including on the same\n  page at the same moment that Claude in Chrome was dead.\n\n### Drive it, then verify\n\n1. Navigate to the target URL and drive the real action.\n2. Confirm it worked against the DOM (e.g.\n   `document.body.innerText.includes(...)`), not just a screenshot — a\n   tooltip appearing can look like success. Check you are still on the page\n   you think you are: an app that has quietly redirected you to a login screen\n   will absorb blind coordinate clicks into empty space and hand you a session\n   with zero events.\n3. **Close the tab.**\n\nThen sanity-check the recording, before you trigger anything. Wait ~10s after\nclosing the tab, so the session is complete, and check that it captured\nsomething:\n\n```bash\nmeticulous agent sessions --limit 10 --excludeSyntheticSessions \\\n  --includeDurationSeconds --includeNumberUserEvents \\\n  --includeNumberUrlsVisited --includeStartUrl --includeAbandonedReason\n```\n\nRead the row you just produced:\n\n- **no row at all** — nothing was uploaded yet; wait a little longer before\n  concluding the recording failed\n- **`numberUserEvents` of 0** — the recorder saw no user input. Your clicks\n  were not reaching the page, or were synthesised rather than trusted; go back\n  to the input-delivery probe above. Replaying this session is pointless.\n- **`numberUrlsVisited` of 1 when you navigated several times** — the later\n  pages did not make it into _this_ session. They either landed in their own\n  sessions (fine, collect those ids too) or were swallowed as an unreplayable\n  tail (see the hard-navigation note in Step 5).\n- **`durationSeconds` over 300** — everything past the 5-minute mark will be\n  silently trimmed on replay (Step 5). Re-record the overflowing part as its\n  own session rather than hoping it survives.\n- **populated `abandonedReason`** — the recorder gave up on the session (see\n  the 10-minute cap in Step 5); it is not worth replaying.\n- **`startUrl` that is not the page you drove** — the navigation you cared\n  about belongs to a different session than you assumed.\n\n**A session is only complete once its tab is closed.** While the tab is open\nthe row reflects only the chunks uploaded so far, so a low or zero\n`numberUserEvents` there means \"not flushed yet\", **not** \"the recording\nfailed\". A session measured for this skill read **0** events with the tab open\nand **12** once it was closed. Judging it early nearly caused a perfectly good\nrecording to be re-driven from scratch.\n\nEvents upload on a short interval (a few seconds), but anything still\nunflushed at unload is only stashed in `sessionStorage` and re-sent on a\n_later page load to the same origin_ — so a session's tail can be delayed\nuntil the next visit. Prefer navigating away over hard-closing the browser,\nand never judge a recording immediately.\n\n**Recording several targets in one sitting? Run this check after _each_ one**,\nnot just once at the end. Checking only at the end makes it impossible to\ntell which action lost a session, and you'll have to re-drive all of them\njust to find out which one needs redoing.\n\n## Step 5 — Session-time budget and close discipline\n\n- **Cloud replays cap at 5 minutes of session time.** Everything recorded\n  after that is silently trimmed from the replay — the run still succeeds and\n  reports itself accurate, so the loss is invisible unless you look for it\n  (snapshot routes stop early; far fewer allowed events than the session has\n  clicks). Don't leave this to feel: `agent sessions\n--includeDurationSeconds` gives you the number, and anything over 300 is\n  losing its tail.\n- **Your interaction pace eats this budget.** Each find\u002Fclick\u002Fverify round\n  trip is 10-60s of _recorded_ session time. Plan the flow completely before\n  opening the tab, batch your actions into as few round trips as possible\n  with short waits between them, verify from the recording afterwards rather\n  than mid-flow, and aim to stay under 4 minutes. Several short sessions beat\n  one long sweep.\n- **Direct URL navigation is a legitimate fast path** to each target page\n  and replays fine — prefer it over slow click-paths. But hard navigations\n  split the recording into multiple sessions, so collect every resulting\n  session id afterwards and pass them all to `--sessionIds` (Step 6). Mostly\n  this is fine, each piece staying under the replay cap. The trap is\n  navigating again too quickly: a page reached a second or two after the\n  previous one gets appended as the _tail_ of that session instead of\n  starting its own, and tails frequently do not replay — so the page renders\n  perfectly while you drive it and still contributes no coverage. Give each\n  page you actually care about its own dwell time (~10s) before moving on,\n  and check in Step 6 that it shows up as a `startUrl` in its own right. The\n  same caution applies to a plain `\u003Cform>` submit with no wired `onSubmit`\n  handler — it triggers a real browser reload rather than an SPA transition,\n  and can just as easily drop the just-recorded, unflushed session if you\n  navigate on immediately afterward.\n- **Recorder limits:** a 10-minute hard cap on tab-open time marks the whole\n  session \"abandoned\"; uploads flush on a 5s interval — wait ~6-8s after the\n  last interaction before closing the tab.\n\n## Step 6 — Collect the session ids and trigger the test run\n\nFirst list what you actually recorded, newest first:\n\n```bash\nmeticulous agent sessions --limit 20 --excludeSyntheticSessions \\\n  --includeDurationSeconds --includeNumberUserEvents \\\n  --includeNumberUrlsVisited --includeStartUrl\n```\n\nSkip any row with `numberUserEvents` of 0 — it will replay as nothing and\nonly dilutes the run. Note any row with `durationSeconds` over 300: it will\nreplay, but only its first five minutes, so treat coverage from its tail as\nabsent rather than assuming the whole flow ran.\n\nIdentify your sessions by recorded-at time and `startUrl`. Be careful here:\nother people — and other apps pointed at the same project — record too, so\nnever assume the newest N rows are yours. `--recordedSince` and\n`--visitedUrlFilter` are the quickest way to narrow it down when the list is\nbusy.\n\nExpect **more sessions than pages you drove**: a hard navigation usually ends\none session and starts another, so a five-page sweep can produce five ids.\nCollect all of them. A page whose URL never shows up as a `startUrl` was\nprobably swallowed as the tail of the previous session and will not replay —\nre-record it on its own if you need it covered.\n\nThen trigger:\n\n```bash\nmeticulous agent trigger-test-run --sessionIds \"\u003Cid1>,\u003Cid2>,...\"\n```\n\n## Step 7 — Compare with a union, not a raw diff\n\n```bash\nmeticulous agent js-coverage --headPlusTestRunIds \"\u003CnewRunId>\" \\\n  --includeAllFiles --includeCoveragePercentage > \u002Ftmp\u002Fcombined-coverage.tsv\n```\n\nThis unions your new run into the baseline run resolved from HEAD — the same\nrun Step 1 used — so it is baseline coverage **plus** your new sessions'\ncoverage. Commit resolution skips runs over an explicit session set, so the\nrun you just triggered won't be picked as the baseline. The union is needed\nbecause `--sessionIds` _replaced_ the selected set for that run rather than\nadding to it, so your run on its own covers far less than the baseline and a\nraw diff would read as mass regressions. Diff the union against\n`\u002Ftmp\u002Fbaseline-coverage.tsv`: you should see zero regressions, and only the\nfiles your new flow touched improve.\n\nPass only your new run — a run cannot be unioned with itself.\n\nTwo signs HEAD resolved to something other than the golden-set run: it is\n**rejected** with \"is the run being queried\", or _every_ file has dropped\n(which means you unioned into another narrow session set, not a regression).\nThe most likely cause is a pinned-session run predating the skip. Either way,\nname both sides explicitly, using the baseline id from Step 0:\n\n```bash\nmeticulous agent js-coverage --testRunIds \"\u003CbaselineRunId>,\u003CnewRunId>\" \\\n  --includeAllFiles --includeCoveragePercentage > \u002Ftmp\u002Fcombined-coverage.tsv\n```\n\nIf the union is rejected because the runs executed different commits, that is\nthe `main`\u002Fclean-tree precondition biting — see the top of this skill.\n\n## Step 8 — Verify and report\n\nFor each traced target file: did coverage move? For each presumed-dead file\n(the `.meticulousignore` candidates from Step 2): did it stay at 0% in the\nunion comparison (confirming it's genuinely unreachable)?\n\nIf a well-traced target didn't move, don't assume the recording failed —\nre-check the session first (Step 4: does it exist, did it capture user events\nand URL visits, is it abandoned, is its `startUrl` the page you drove?), then\nre-check whether the click actually goes through the file you expected\n(Step 3) rather than a sibling\u002Fparent component.\nReport honestly if a target remains unresolved; don't claim success without\nthe coverage number to back it.\n\nWhen you report the gains, be clear about what they are not yet: the test run\nproves the coverage is reachable, but the project's own coverage figure will\nonly improve once the next session selection picks these new sessions up into\nthe selected set. Until then nothing changes for recurring runs.\n\n## Step 9 — Open the `.meticulousignore` PR\n\nThe second deliverable. Branch, commit the `.meticulousignore` change, and\nopen a PR.\n\n**Only propose paths you are confident are not coverable at all.** The bar is\n\"no session could ever execute this\", not \"I didn't get to it today\". Prefer\ndirectory-level rules over long lists of individual files — a directory rule\nstays correct as files are added, whereas a file list silently goes stale. In\npractice most entries come from the safe-exclusion categories in Step 2 plus\nwhatever whole non-browser packages the baseline showed at a uniform 0%.\n\nConfirm before you commit: every path you are about to ignore stayed at 0% in\nthe Step 7 union. A path your own new sessions just covered obviously does not\nbelong in the ignore list, and that check catches it.\n\nA common structure is \"ignore everything, then un-ignore what does run in the\nbrowser\" — the pattern Meticulous's own monorepo uses:\n\n```\n# Ignore everything except packages that are executed in the\n# browser and have meaningful frontend coverage.\n\npackages\u002F*\n!packages\u002F\u003Cfrontend-app>\u002F\n!packages\u002F\u003Cfrontend-app>\u002F**\n!packages\u002F\u003Cshared-ui-lib>\u002F\n!packages\u002F\u003Cshared-ui-lib>\u002F**\n```\n\nNote that `.meticulousignore` follows gitignore semantics, so a file cannot be\nre-included once its parent directory is excluded — un-ignore the directories\n(`!some\u002Fdir\u002F**\u002F`) as well as the files.\n\nIn the PR description, give the reasoning for each rule — _why_ this code\ncannot run in a browser (it's a Node-only build script, a test harness, a\nbackend package). The baseline can only show you that something is at 0%\ntoday, which is never proof it is unreachable, so the justification has to\ncome from what the code actually is. Also call out explicitly:\n\n- anything you left **out** of the ignore list despite low coverage because\n  you suspect a source-map\u002Fattribution gap rather than dead code (Step 2)\n- anything unreachable only because of environment or feature-flag config\n  rather than structurally — that is the reviewer's judgement call, not\n  yours, so flag it instead of silently ignoring it\n\n## Reference\n\n`references\u002Fworked-example.md` runs the whole skill against a small app\n(kanban-demo), with the real coverage deltas. Worth reading for calibration:\nit has no 0% files at all, one target that gained coverage and one that\nrecorded cleanly and then failed to replay — and it shows how to tell the\ndifference from executed ranges.\n",{"name":5,"description":7,"user-invocable":24},true,"main","9836d94e88404aabb682086405c2aa9cba7657ac","plugins\u002Fmeticulous\u002Fskills\u002Fmeticulous-increase-coverage",156,"https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace","https:\u002F\u002Fgithub.com\u002FCognitionAI","https:\u002F\u002Fcognition.ai\u002F","Makers of Devin, the AI software engineer.","2026-11-13 07:08:23",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":24},[36,39,42],{"tool":37,"command":38},"npx","npx skills add CognitionAI\u002Fdevin-marketplace --skill meticulous-increase-coverage",{"tool":40,"command":41},"bunx","bunx skills add CognitionAI\u002Fdevin-marketplace --skill meticulous-increase-coverage",{"tool":43,"command":44},"pnpm","pnpm dlx skills add CognitionAI\u002Fdevin-marketplace --skill meticulous-increase-coverage","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fmeticulous\u002Fskills\u002Fmeticulous-increase-coverage\u002FSKILL.md",[47],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"references\u002Fworked-example.md","REFERENCE","参考文档",4757,"https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fmeticulous\u002Fskills\u002Fmeticulous-increase-coverage\u002Freferences\u002Fworked-example.md",[54,60,67,73,80,87],{"skillNo":55,"name":56,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M7BYC4AJGM98P2G9JH7QS24Z","business-data","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fwindsor-ai\u002Fskills\u002Fbusiness-data\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M7BYBZ0CD4CBCPZHRZ551BNK","workflow","Creates durable, resumable workflows using Vercel's Workflow SDK. Use when building workflows that need to survive restarts, pause for external events, retry on failure, or coordinate multi-step operations over time. Triggers on mentions of \"workflow\", \"durable functions\", \"resumable\", \"workflow sdk\", \"queue\", \"event\", \"push\", \"subscribe\", or step-based orchestration.","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fvercel\u002Fskills\u002Fworkflow\u002Fupstream\u002FSKILL.md",[],[],{"skillNo":68,"name":62,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M7BYB1XSZHKHD8JZTQJ62JN0","Vercel Workflow SDK expert guidance. Use when building durable workflows, long-running tasks, API routes or agents that need pause\u002Fresume, retries, step-based execution, or crash-safe orchestration with Vercel Workflow.","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fvercel\u002Fskills\u002Fworkflow\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M7BYASGMK935J33E21N5WTW0","verification","Full-story verification — infers what the user is building, then verifies the complete flow end-to-end: browser → API → data → response. Triggers on dev server start and 'why isn't this working' signals.","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fvercel\u002Fskills\u002Fverification\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M7BY9WJERKSKEH9HK6E71GW2","vercel-storage","Vercel storage expert guidance — Blob, Global Config (formerly Edge Config), and Marketplace storage (Neon Postgres, Upstash Redis). Use when choosing, configuring, or using data storage with Vercel applications.","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fvercel\u002Fskills\u002Fvercel-storage\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M7BY9N034894YV2J9DCRZHYS","vercel-services","Configure and troubleshoot Vercel Services for multiple frontends and backends in one project. Use when composing a polyglot or multi-service application on one Vercel deployment; defining the `services` key, service-targeted rewrites, or service bindings in `vercel.json`; or running all services with `vercel dev`.","https:\u002F\u002Fgithub.com\u002FCognitionAI\u002Fdevin-marketplace\u002Fblob\u002Fmain\u002Fplugins\u002Fvercel\u002Fskills\u002Fvercel-services\u002FSKILL.md",[],[]]