Skip to content

Sync upstream main - #2

Closed
hakarrr wants to merge 354 commits into
hakarrr:mainfrom
career-ops-hq:main
Closed

hakarrr wants to merge 354 commits into
hakarrr:mainfrom
career-ops-hq:main

Conversation

@hakarrr

@hakarrr hakarrr commented Nov 15, 2026

Copy link
Copy Markdown
Owner

Syncing upstream main into fork

dgitman and others added 30 commits November 10, 2026 16:24
…6604)

* fix(web): read script-rendered JDs before the evaluate worker starts

A live Robert Half posting came back from the web app as "Needs review".
Its Phenom page keeps the whole JD (7 KB) in a schema.org JobPosting inside
<script>, and the visible HTML is ~270 characters of chrome. The evaluate
worker has no Bash by design (it reads an untrusted posting), so its only
way to read the web is WebFetch, which drops script content, sees a shell,
and correctly stops with "needs confirmation".

Read the JD on the server instead, then hand it to the worker:

- fetch-jd.mjs gains a second tier. On an ATS-API miss it reads the page's
  own JSON-LD JobPosting (lib/jd-page-fetch.mjs). The GET carries the same
  egress guard as the other fetchers: literal-host and DNS checks on the
  URL and on every redirect hop, a body cap, and an HTML content-type check.
- The web run route calls it for evaluate runs and passes the text to
  buildPrompt, which supplies it as JSON-encoded untrusted data in place of
  the WebFetch step (both the official and user-selected branches). The
  worker keeps its read-only tool scope. A miss leaves the WebFetch path
  exactly as before.
- batch/jd-rescue.mjs becomes importable (the parsers are reused, not
  copied) and learns to unescape entity-escaped markup. Phenom ships the
  description as `&lt;div&gt;...`, which the tag strip never saw, so raw
  <div><p><span> reached the JD text. The batch runner had the same latent
  bug. It also gains --help, which the help-flag lint requires of any CLI.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* fix(web): address review of the JD prefetch (egress guard, closed postings, truncation)

- Run each fetch hop inside providerFetchContext so the connect-time DNS
  guard validates the dialled address (DNS rebinding), and cover IPv6
  literals in tests.
- Never fetch LinkedIn, directly or via redirect (AGENTS.md LinkedIn guard).
- Treat a JobPosting whose validThrough is past as a miss so the worker's
  closed-posting rule decides.
- Only fall back to an embedded description when the page holds a single
  posting and a single description-sized string, so another job's text is
  never used.
- Cut an over-long JD surrogate-safely and append an explicit truncation
  notice; the prompt tells the worker to stop and ask for the full text.
- Unescape only known tags in entity-escaped markup; guard invalid code
  points.
- Register lib/jd-page-fetch.mjs in SYSTEM_PATHS.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* test(jd-page-fetch): compare the parsed hostname instead of a URL prefix

Resolves the CodeQL "incomplete URL substring sanitization" alert on the
test stubs.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* fix(web): carry a closed posting through the prefetch and count description-less postings

- An expired JobPosting (validThrough in the past) is no longer a plain miss:
  fetch-jd exits 3 with a CLOSED_POSTING line, the prefetch carries it, and
  the prompt tells the worker to follow the closed-posting rule instead of
  WebFetching the URL, so a stale posting whose visible page still shows a JD
  is not scored.
- The ambiguity guard for the embedded-description fallback now counts
  postings without a description, so a stub next to a described posting no
  longer lets an unrelated embedded description through.

* fix(web): close the remaining review gaps in the JD prefetch

- Fail closed when the connect-time lookup guard is opted out
  (CAREER_OPS_NO_DNS_CACHE=1): the fetch is a miss, not an unguarded request.
- Treat a date-only validThrough as valid through that whole day.
- Tell description-less postings apart by url/@id/identifier, while a posting
  written twice with identical fields stays one posting.
- Allow &amp; inside quoted attribute values of escaped tags.
- A truncated JD now also blocks the report envelope and archiving.

* fix(web): refuse ambiguous pages up front and stop the prefetch on abort

- A page naming more than one posting is a miss before anything is read from
  it: no JSON-LD description, no embedded description, and no closure is
  attributed to the requested URL.
- Pass the request's abort signal to the fetch-jd child and skip starting the
  worker when the caller has gone.

* fix(web): tell postings apart by employer and location in the ambiguity guard

* test(jd-page-fetch): prove the connect-time guard over a real socket

* fix(web): hold the tracker guard during prefetch and mark capped ATS text

- Take the tracker write guard before the JD prefetch and release it on abort
  or any failure before the worker spawns, so a row delete cannot land in the
  prefetch window.
- An ATS result cut at the 20,000-character cap now ends with the explicit
  truncation notice instead of a bare ellipsis, so the incomplete-JD stop
  recognises it.
- Test the employer and location mismatches independently.

* fix(web): hold page prefetch to the proxy trust boundary and honor visible closure text

- fail closed when NODE_USE_ENV_PROXY routes the fetch without CAREER_OPS_TRUST_PROXY_EGRESS=1
- run the liveness classifier over visible page text and treat hard closure as closed
- carry closure evidence (not just validThrough) through fetch-jd, prefetch and prompt
- real-socket test: NO_PROXY=* and assert the connect-time lookup is blocked

* fix(jd-fetch): scan hostile pages in linear time

- replace the lazy-regex tag stripper and ld+json extractor with indexOf scans
- cap the visible-text pass at 1 MB
- test: 1.5 MB of '<', '<!--' and unclosed '<script ' finishes in under 3s

* fix(web): harden page prefetch against hostile markup and closure edge cases

- decode character references before the visible-text closure check
- treat a redirect to a different page as a miss, not that page's posting
- accept multiline CLOSED_POSTING evidence in the prefetch
- return 422 for a truncated prefetched JD instead of starting a worker
- keep block boundaries for opening block tags in the visible text
- read ld+json script tags with a quote-aware tag end
- strip markup in jd-rescue with a linear scan instead of lazy regexes

* fix(jd-rescue): drop an unclosed script or style payload instead of leaking it

* test(jd-page-fetch): closure banner that only matches after entity decoding

* fix(jd-fetch): decouple page parsers from batch/jd-rescue and carry Ashby delisting as closure

- move the pure HTML parsers into lib/jd-parse.mjs so a locally edited, updater-preserved batch/jd-rescue.mjs cannot break fetch-jd
- batch/jd-rescue.mjs re-exports them for existing importers
- Ashby isListed: false now yields closed + evidence, which fetch-jd turns into CLOSED_POSTING

* fix(jd-fetch): cache the next '>' so visible-text scanning stays fast on hostile markup

A run of '<script ' or '<div ' with no '>' rescanned the page per tag (1.5-2.4s locally, 8s on the Windows runner).

* fix(jd-fetch): carry closure through the batch runner and harden the parsers

The batch runner now treats fetch-jd exit 3 as a terminal skip instead of
falling back to a fetch that could score a closed posting. jd-rescue
declines closed ATS results. Entity lookups use own-property checks,
case folding is ASCII-only so offsets stay aligned, and the JSON-LD walk
uses an indexed queue. The proxy fail-closed test passes a real cwd path
so it also runs on Windows.

* fix(jd-rescue): propagate ATS closure as a terminal rescue result

A closed ATS posting no longer degrades to a miss that lets embedded or
page text be rescued as a JD. The chain stops, jd-rescue exits 3, and the
batch runner skips the offer without calling fetch-jd or starting a worker.

* fix(jd-fetch): detect closure on statically fetched pages, decode Latin-1 entities, test the skip as behavior

jd-rescue now checks the downloaded page for an expired validThrough or a
closure banner before accepting its text, so a closed role with a full
visible JD exits 3 and is skipped. A page with no extractable JD but a
closure banner is reported closed instead of a plain miss. Named Latin-1
entities (&eacute;, &uuml;, ...) decode case-sensitively instead of
becoming spaces. The closure skip and the ATS-closed rescue are now
covered by behavior tests rather than source-pattern checks.

* fix(jd-rescue): check ATS closure before accepting page text and survive a throwing source

rescueDecision checks the page and then the ATS API for closure before any
candidate is accepted, so an adequate JSON-LD description no longer hides an
Ashby delisting. A source that throws no longer ends the chain. Both are
covered by behavior tests.

* fix(jd-fetch): gate the downloaded-page closure check like the live fetch and keep it linear

pageClosureEvidence now applies the same-document gate to the final URL
(the runner passes it to jd-rescue), returns nothing for an ambiguous page
with several postings or description payloads, and caps the HTML it reads.
A closure-only page with several postings is a miss, not a closure. The
embedded-description scan is a single linear pass, so repeated unterminated
description fields cannot make it quadratic.

* fix(batch-runner): skip fetch-jd once the rescue says closed, with end-to-end coverage

A closure note from jd-rescue now also stops the fetch-jd fallback, so a
closed posting reaches neither a second fetch nor a worker. Tests drive the
real jd-rescue CLI with a stubbed closed ATS result and the extracted runner
blocks from rescue through skip.

* fix(jd-rescue): never rescue a JD from an off-document or ambiguous download

An off-document final URL now makes the rescue ignore the page entirely
(exit 4, the runner drops the other page's text) while a redirect to an
error-style URL still reads as closed. An ambiguous page with several
postings or description payloads supplies no page-derived candidate; only
the ATS API keyed by the requested URL can still produce a JD.

* fix(jd-fetch): tolerate whitespace-only duplicates and malformed fields in page inspection

Postings that differ only in description whitespace are one posting, a
malformed JSON-LD value no longer stops the closure check, and a malformed
description-sized field still counts toward ambiguity without ever being
used as extracted text. Hostile-markup timing guards use a wider limit that
a quadratic scan would still exceed.

* fix(jd-rescue): exit 4 whenever the page is unusable and no JD was found

An off-document or ambiguous download is flagged pageUnusable and, with no
JD from the ATS API, always exits 4 so the runner clears the JD file,
regardless of thin or not-better candidates. The discarded page's word
count no longer decides whether an API JD is accepted.

* fix(jd-fetch): ignore banner text on pages that ship a script-toggled expiry view

Phenom career sites include the 'job has been filled' view in the markup of
every live posting. Found by running the live fetcher against current
Robert Half postings, which all read as closed. Visible-text closure is
skipped when that template is present; structured closure signals
(validThrough, ATS flags) still apply.

* fix(jd-fetch): canonicalise posting comparison, reject a posting that names another job, check oversize pages whole

Postings are compared with nested keys sorted and whitespace collapsed. A
single JobPosting whose own URL carries job-id tokens that share nothing
with the requested URL is treated as another job's. A malformed description
copy that differs from a valid one only in whitespace compares equal. A page
over the inspection cap is unusable for page-derived text while the ATS API
stays available.

* fix(jd-fetch): detect the expiry-view template in markup only, not inside scripts

* fix(jd-fetch): no page closure from an oversize page, ignore year/version path segments as job ids, file-URL imports in tests

* fix(jd-fetch): keep year-like numeric segments as a fallback job identity

* fix(jd-fetch): drop only the expiry-view elements, keep classifying other closure banners

* fix(jd-fetch): make the expiry-view removal a single linear pass

* fix(jd-fetch): a marked void element drops only its own tag

* test(jd-rescue): pass the resolve-hook preload to --import as a file URL (Windows)

* fix(jd-fetch): treat <param> as a void element when dropping expiry views

* fix(jd-fetch): read job ids from the URL fragment in the other-job check

* fix(jd-fetch): read id-like query values from the URL fragment too

* fix(browser-extract): an unlisted Ashby posting is not a closed one

* fix(batch): do not accept a truncated fetch-jd result as the full JD

* test(batch): pin the runner's truncation marker to the real fetch-jd producer

* fix(jd-rescue): reject a capped ATS result; point the web truncation error at a path that works

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
…rting route (#6807)

* docs(security): name GitHub private vulnerability reporting as a reporting route

Private vulnerability reporting has been on for the repository since
2026-10-11 (#2908), but SECURITY.md and SUPPORT.md still named only the
email address. Both now list the Security tab's "Report a vulnerability"
form next to the email. The scope also lists prompt injection through
untrusted postings, pages, emails and forms that drives agent writes or
submissions. .github/SECURITY.md, the copy GitHub renders, stays
identical to the root file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(security): scope prompt-injected submissions the user never approved, not only unreviewed ones

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…) (#6812)

plugins/_template/test/smoke.mjs passed a bare filesystem path to
import(), which reads its argument as a URL. On Windows the drive
letter is taken as a URL scheme, so every plugin scaffolded by
`plugins.mjs new` failed its own smoke test with
ERR_UNSUPPORTED_ESM_URL_SCHEME. On every OS, a '#' or '%' in a
directory name changed which file the import named.

Import through pathToFileURL(...).href, as validate-portals.mjs does.
The new test scaffolds a plugin under a temp root whose name holds
'#' and '%' and runs its smoke test, so it fails without the fix on
Linux and macOS as well as Windows.

Fixes #6811
Bumps [actions/setup-node](https://github.com/actions/setup-node) from 7.0.0 to 7.1.0.
- [Release notes](https://github.com/actions/setup-node/releases)
- [Commits](actions/setup-node@v7...949feb2)

---
updated-dependencies:
- dependency-name: actions/setup-node
  dependency-version: 7.1.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: career-ops-release[bot] <333437863+career-ops-release[bot]@users.noreply.github.com>
- when no provider reads the posting site, count its other postings in scan history, pipeline and tracker
- lead the fix with a provider when they name other companies, since one covers them all
- stop claiming no provider reads a site whose provider only detects careers pages, e.g. Dover

Closes #5564
Agent-Session: claude/2c329b2c-7ed7-4261-a77b-fa01f26a48ee
…6849)

The case lets the lock churn for ~2.6 windows and then called acquirePipelineLock with no maxWaitMs, so the ceiling came from CAREER_OPS_PIPELINE_LOCK_MAX_WAIT_MS. With that exported below the churn (1000, as reported) the caller gave up before the lock was freed and the case failed for a reason that has nothing to do with the code under test.

Pass maxWaitMs explicitly at 10 windows — the module's own default, well clear of the churn — so the environment cannot change what this case measures, and say why in a comment.

Refs #6755
…6850)

The fixture check used named.some(), so a row naming two codepoints still
passed when its sample kept only one of them -- the drift the test was
written to catch (#6753).

Every named codepoint must now appear in the sample, and the failure names
the ones that are missing. Two rows only showed one of the two codepoints
they name ("Curly single quote", "Up / down arrow"), so their samples now
carry both. "(also ...)" still names variants a row does not show.
* test(web): name the company Apply next should use

The button still follows the highest 4.0 score. These tests expect the stored Apply, including one past the six cards on screen.

* fix(web): point Apply next at the stored call

Apply next was the highest 4.0 score in the first six cards. A stored Skip took the button. An Apply past those six had none. The button now follows final_decision. The score bands stay.

* fix(web): skip a legitimacy-flagged job in Apply next

pickApplyNext treated every stored Apply as the one-click target. The Recommended group on Today already leaves out a caution or suspicious posting. Apply next now uses that same check. A flagged Apply is not picked. The next unflagged Apply still is.
…n entry raise it (#6799)

* fix(smartrecruiters): say when the page cap cuts a board short; let an entry raise it

The walk stopped at 50 pages (5,000 postings) and returned that prefix as
the whole board, silently. Bosch listed 5,104 postings on 2026-11-10, so
104 were never read on any run, and scan.mjs logged the board reachable
rather than partial.

The API reports totalFound, so the walk now tells a board the cap cut
short from one that ended: it warns on stderr with the entry, the counts
and how to raise the cap, and sets the canonical truncated tag
lib/board-truncation.mjs reads. An entry may set max_pages up to 100
pages; the default stays 50, so request volume is unchanged unless an
entry opts in. A one-page health probe (ctx.maxPages) is not a
truncation. This follows oraclecloud.mjs (#6690).

* test(smartrecruiters): a board past the page cap is reported, and max_pages raises the cap

Pages a fake 5,104-posting board through fetch(): the cap marks it
page-cap, 5000 of 5104, and warns; with no totalFound the cut is still
marked, total unknown; max_pages: 60 reads it whole; a value past the
ceiling is held to 100 pages. A board ending exactly at the cap and a
one-page probe are not marked. Six of the checks fail on main.

* fix(smartrecruiters): space the page requests and retry a transient failure

With max_pages up to 100, a raised cap could send 100 list requests to
the public API back to back, and one dropped connection or 503 ended the
board. The walk now pauses 250 ms between pages and retries a transient
failure (429, 5xx, a network error) up to 3 times through
fetchJsonWithRetry(), as oraclecloud.mjs does. A page that still fails
ends the board with its error, not as a page-cap truncation.

The page-cap test stubs ctx.sleep, and covers the pause, a page that
fails once, and one that keeps failing.
…#4272)

* fix(application-answers): find the draft block by marker, not by letter and English name

parseDraftAnswersBlockH matched one exact heading:

    /^##\s+H\)\s*Draft Application Answers\s*$/m

Neither half of that is stable across locales. Of the 19 evaluation modes,
only 3 write the English name under H) (canonical, ar, ja). Five write a
translated name under H) (es ru tr zh zh-TW) and eleven write a translated
name under G) (da de fr hi id it ko nl pl pt ua).

In the other 16 the reader returned null, which its own contract spells
"this report has no draft block", so modes/apply.md took the "New question
-> generate from the report + cv.md" path and regenerated every answer the
evaluation had already drafted. Silent, duplicated model cost, non-English
users only.

A trailing (draft) marker is the discriminator instead. That is the
convention web/src/lib/report-sections.mjs already established for the
verdict callout: (lead)/(verdict) are stripped for display precisely
because they are authoring signals that read the same in every language.
The 19 modes now write (draft) on the draft-answers heading, cleanHeading
strips it so the viewer never sees it, and isVerdictHeading deliberately
does not accept it — a draft block rendered as the Verdict callout would
put unsent answers where the hiring decision belongs.

The English name alone still resolves, under any letter, for the reports
already on disk: those predate the marker and are the corpus apply reads
today. A translated heading in an OLD report stays unreadable, which is
deliberate — no discriminator recovers it without guessing, and guessing
here re-submits a mispaired answer to an employer.

tests/draft-answers-marker-modes.test.mjs asserts the marker against the
modes themselves, not a fixture, because the failure mode this fixes is a
new locale arriving unmarked. Evaluation modes are counted by structure
(^## F\)) rather than filename — only 8 of the 19 are named oferta.md.

Verified load-bearing: with the old regex restored, the locale fixtures and
the mode-heading round-trip both fail; with one mode's marker removed, the
coverage check names that file. Full node test-all.mjs: 9101 passed, 0
failed, 16 pre-existing warnings (the santifer.io credit-file noise of
#4131). web/tests/lib/report-sections.test.mjs: 17 passed.

Fixes #3884

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(modes): mark the draft-answers heading with (draft) in every mode

Rebased on main as asked in review. The parser is #4432's
findDraftAnswersHeading, reused as it is; this PR no longer defines one.
What is left:

- the 21 evaluation modes write ` (draft)` on the `## H)` draft-answers
  heading (19 when this PR opened; nb and vi have landed since);
- tests/draft-answers-marker-modes.test.mjs finds the evaluation modes by
  structure (`^## F)`), asserts each marks exactly one heading and that it
  is the H) block, and runs each mode's real heading through
  parseDraftAnswersBlockH, also with the letter moved to I), the case the
  #4400 letter rule cannot cover for a translated name. Against main's
  modes it fails 42 of 43;
- web: cleanHeading strips `(draft)` for display, and isVerdictHeading
  keeps refusing it, so the draft block never becomes the Verdict callout;
- evaluation-options.test.mjs now expects the marker in the focused-run
  excerpt of oferta.md (the extractor already allowed a suffix);
- apply.md and auto-pipeline.md name the marker, and the comment in
  findDraftAnswersHeading that said no mode emits it is updated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…cripts they describe (#6806)

* docs(modes): bring offer-prep, hm-audit and titles in line with the scripts they describe

- offer-prep Step 7 records the prep through set-status.mjs instead of
  hand-editing the tracker row, and says what to do when the row is
  already Offer.
- hm-audit Input 4 uses the [-{report}] payload and HTML names pdf.md
  now writes, and picks the kind=cv row from pdf-index.tsv so a cover
  letter's HTML is never audited as the CV.
- titles Filter 1 points at title-keywords.mjs and describes its
  folding, word:/stem: prefixes and " + " groups; the profile-missing
  error no longer claims a missing portals.yml also stops the mode.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(modes): keep the titles phrase test-all pins, never fall back to another report's CV

- titles Filter 1 keeps "case-insensitive substring", which test-all's
  titles-mode assertion requires.
- hm-audit's fallback takes only a CV carrying this role's report number,
  asks when only unnumbered CVs remain for a company with several roles,
  and stops as Step 1 says when none fits.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(hm-audit): use the one unnumbered CV when the company has a single role

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…er too (#6809)

* fix(seeds): read a16z's embedded company list and probe Ashby and Lever too

The a16z parser targeted JSON-LD, data-company-name attributes and card
anchors, none of which a16z.com/portfolio/ has any more, so --seeds a16z
returned 0 companies. Read the data-companies attribute (HTML-escaped JSON)
and fall back to the inline window.a16z_portfolio_companies script. Skip
acquired companies by default and drop legal suffixes from slugs.

fetchA16zCompanies and fetchYCCompanies now throw when the page (or page 1)
parses to zero companies, so a markup change surfaces as a warning instead
of a clean-looking empty run.

Seed companies were probed on Greenhouse only, although toPortalEntry's
JSDoc said Greenhouse, Lever and Ashby. toPortalEntries() now returns
Ashby, Greenhouse and Lever candidates by slug (ordered by a 300-company
YC sample: 25/4/4 hits), and probeSeedBoards() tries them in order. A 404
moves on; a company with no board anywhere counts as "no board found"
rather than an unreachable board.

Fix the YC page-size comments (25 per page, not ~30; page count read from
totalPages) and the seeds README.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(seeds): cap concurrent seed probes per ATS host

Each seed company can now be probed on Ashby, Greenhouse and Lever, and
each of those serves every board from one API host, so 20 companies in
flight meant up to 20 requests at one host. Run each probe through the
existing keyed limiter at SINGLE_HOST_CONCURRENCY, keyed on the board host,
as the history seed pass already does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
#6785)

* fix(role-clarity): accept a Reports to name that starts with a non-ASCII letter (#6758)

* fix(role-clarity): make the placeholder guard's word boundary Unicode-aware (#6758)

* fix(role-clarity): keep footnoted placeholders like TBD¹ excluded (#6758)
* feat(doctor): expose data_root and project_root in --json output (fixes #6498)

Signed-off-by: Vijay Sirohi <191354616+Vijaysir0hi@users.noreply.github.com>

* fix(doctor): report null data_root on unsafe refusal, document marker, and resolve relative target

---------

Signed-off-by: Vijay Sirohi <191354616+Vijaysir0hi@users.noreply.github.com>
…rowser (#6795)

* feat(scanners): add TopDev scanner - third Vietnamese source, via a browser

TopDev (topdev.vn) answers a bare fetch() with HTTP 200 and ~741 KB of
markup, so it was recorded as server-rendered in #3378. Re-measured
2026-11-10: the search RESULT list is not in that HTML. Only ~15 postings
appear, all inside the Next.js RSC flight payload (self.__next_f.push) and
belonging to recommendation rails, not to the query; no /api/ path is
exposed to read instead. The listing is hydrated in the browser.

That shape belongs in a standalone browser scanner, never a
providers/*.mjs module (providers/ADDING_A_PROVIDER.md, "Browser-based
scanners"), so this follows scan-interamt.mjs and scan-dayforce.mjs.

- cards read from div[class*="text-card-foreground"]; title and URL from
  the card's FIRST /detail-jobs/ anchor (the same anchor shape repeats
  later in the card), company from that anchor's sibling, location from
  the grid-cols-2 cell
- tracking parameters (?src=&medium=) are stripped from the posting URL,
  so one posting reached from two rails dedupes to one
- postedAt comes from the relative label ("3 ngay truoc" / "2 days ago");
  a card with no parsable label gets no postedAt rather than a guess
- topdev_searches in portals.yml: `keyword` required, `location` optional
  and narrowed by region id (only Ho Chi Minh, Hanoi and Da Nang were ever
  observed, so any other value is rejected instead of guessed into the URL),
  max_pages default 5, cap 50
- host-pinned to topdev.vn over HTTPS on every request and every resolved
  posting URL; shares createAdmissionGate, so title_filter,
  location_filter, data/blacklist.md and the seen set all apply
- a challenge page throws TopDevChallengeError and a page that never
  renders throws TopDevRenderError - neither reads as an empty board.
  Challenge markers are only tested once cards fail to render, so an i18n
  string inside a healthy bundle cannot be mistaken for a block

Tests: 22/22 (tests/scan-topdev.test.mjs), hermetic - argv isolation,
region mapping, URL building and host pinning, posting-URL resolution,
relative-age parsing in both languages, card->Job mapping, challenge
detection. The card extractor was additionally verified against the real
captured DOM in Chromium (2/2 cards).

Revisit of the boards #3378 evaluated and rejected:
- TopDev: shipped here.
- CareerLink: its recorded listing URL now returns an empty ~84 KB shell
  with no posting links, /tim-viec-lam answers 404, and the markup moved to
  Next.js modules - its row is corrected, not built on.

* docs(job-boards): record the re-measured VN verdicts (CareerLink CAPTCHA, TopCV, VietnamWorks, TechJobs)

Re-measured 2026-11-10 with both a bare fetch() and a headless browser:

- CareerLink is out on rule 3 grounds, not on robots.txt and not merely
  "needs re-discovery" as the earlier edit said. Search, a category page and
  the featured rail all serve the same Cloudflare Turnstile gate ("Xac nhan
  ban khong phai robot") with zero posting links; the homepage carries a
  cf-turnstile-response field. An interactive challenge in front of public
  listings is a stop regardless of whether the board is otherwise readable.

- TopCV serves Cloudflare's block page to fetch() and to a real browser.
- VietnamWorks answers 403 from nginx at the edge, before any rendering.
- TechJobs renders but exposes no postings in the DOM, advertises an MCP
  endpoint instead of an HTTP listing surface, and describes itself as
  aggregating 460+ companies' careers pages - the meta-aggregator shape
  rule 5 keeps out of providers/.

Recorded so the next pass does not repeat the sweep, in the spirit of the
table #3378 added.

* fix(scanners/topdev): address review - no date cutoff, abort on challenge, docs

Four review items from #6795.

1. No date cutoff (scan-topdev.mjs, was :612). The source-wide `lastScanDate`
   filter could reject a posting older than the previous run that had never
   been seen, and record it as `skipped_date` - which counts as seen, so every
   later run would keep hiding it. Recency is not what makes a posting new to
   the user; the gate's seen set is (scan-history, pipeline, tracker). The
   cutoff, `loadLastScanDate()`, the `--all` flag and the `Filtered by date`
   summary line are all gone, and modes/scan.md no longer claims skipped_date
   for this scanner.

2. A challenge now stops the whole scan (was :578-584). The same gate stands
   in front of every remaining search, so continuing only re-requested a board
   that had already said no - the inverse of "not solved, not retried" - and
   filled `errors` with one identical entry per search. `break searches` on a
   TopDevChallengeError; a TopDevRenderError stays scoped to its keyword,
   because it can be a local markup change. The rule is in a small exported
   `isScanFatal()` so it is pinned by a test.

3. Stale TopDev/CareerLink text reconciled in docs/SUPPORTED_JOB_BOARDS.md.
   The CareerViet row no longer contrasts itself with two "server-rendered"
   siblings skipped on robots.txt, and the closing passage no longer calls
   both boards "technically scrapable". One account now: TopDev is supported
   through a browser scanner (its listing is hydrated client-side), CareerLink
   is out on rule 3 behind a Turnstile gate.

4. The REGION_IDS comment now matches the code: an unrecognized location is
   rejected by normalizeSearchEntry (it throws), not "dropped", so a typo
   cannot silently narrow the search.

Tests: 24/24 (was 22; +the abort rule, +the --all-is-now-rejected guard).

* fix(scanners/topdev): register topdev_searches for typo detection; correct the --all test

Two CodeRabbit findings on a9a251e, both real.

1. topdev_searches was missing from PORTAL_TOP_LEVEL_KEYS
   (validate-portals.mjs). A config could then carry a valid list AND a
   `topdev_searche` miss next to it, validate clean, and have the scanner
   silently ignore the extra search - the same class of silent-empty as
   `tracked_companys`. Registered next to interamt_searches, which is the
   other standalone-scanner key of the same shape. Verified end to end:
   validate-portals now reports 'unknown top-level key "topdev_searche" -
   did you mean "topdev_searches"?'.

2. The --all test claimed more than it checked. parseArgs only ignores an
   unknown flag; the rejection is validateFlags in the main-module guard.
   The test is renamed to say that, and a CLI test now runs
   `node scan-topdev.mjs --all` and asserts exit 1 plus the flag list.
   No network: validateFlags exits before main() starts a browser.

Tests: 26/26 (was 24). Also re-ran the related suites - portals-top-level-typo,
validate-portals-unknown-key, scan-local-scanners and all 7
validate-portals-* files - all green.
* feat(web): browse CV templates on Config and choose the default (#1692)

List the templates cv-templates.mjs resolves, preview each one filled with a
bundled fictional resume through build-cv-html.mjs, and set cv.template in
profile.yml with a line edit that keeps the file's comments.

* feat(web): cover-letter templates on Config, same gallery as the CV (#1692)

Parameterize the template gallery by kind. Cover previews fill the template
through generate-cover-letter.mjs's buildHtml with a bundled fictional letter,
and Make default writes cover_letter.template.

* fix(web): carry the preview CSP in the page, refuse comment-dropping profile edits, read files async (#1692)

A srcdoc frame takes its parent's policy, so the preview page now embeds its own
CSP. A cv section the line edit cannot place no longer falls back to a dump that
drops comments. Preview temp files use fs/promises.

* feat(templates): look up named templates in the data root's data/templates (#5777)

cv-templates.mjs now searches data/templates/ under the data root alongside the
checkout's templates/, and a name found in both resolves to the user's file.
Naming only dir keeps that directory the whole registry. Entries carry a layer
tag so callers can tell the two apart.

* fix(web): catch a # comment with no following space, skip commented head tags (#1692)

* test(templates): enumerate shipped templates from the checkout, not the data root (#5777)

Export DEFAULT_TEMPLATES_DIR and pass it where a test means the shipped set, so
templates a developer keeps under data/templates cannot change what these
suites cover.

* test(templates): build the data-root path from tmpdir so it is absolute on Windows (#5777)

* feat(web): duplicate, rename and delete a user's own templates (#1692)

File operations over data/templates/ under the data root, each checked against
the core resolver: a duplicate carries a pack's sections, a shipped template
cannot be renamed or deleted, and the default cannot be deleted.

* feat(web): duplicate, rename and delete from the template gallery (#1692)

Three routes over the file operations, a rename that takes the profile default
along, and per-card actions: Duplicate on every card, Rename and Delete on the
user's own.

* feat(web): edit the body of your own templates from the gallery (#1692)

A save is accepted only when the edited template still resolves, with the
resolver's reason shown when it does not. A version check refuses an edit made
on a stale copy, and shipped templates are not editable.

* fix(web): refuse a pack duplicate whose folder already exists instead of merging into it (#1692)

* fix(web): typed refusals for the manage routes, generic answers for everything else (#1692)

Routes answer a TemplateRefusal with its own message and status and any other
error with a generic 500, so a status no longer depends on words in a message and
a filesystem error's absolute path is not returned. Clear a stale error when the
name is edited or the form is cancelled.

* fix(web): keep filesystem errors out of refusals; no Cancel while a change is running (#1692)

* fix(web): re-check the edit version at write time (#1692)

* fix(web): put a renamed default back when the profile update fails; lock the name field while busy (#1692)

* refactor(web): rename-and-follow lives in the library so its rollback is tested (#1692)

* fix(web): name the template-edit code root codeRoot

Same core-checkout-root rule as the gallery: a parameter called root joined
onto cv-templates.mjs reads as a system file taken from the data root.

* fix(web): refuse template rename and delete when the profile cannot be read

defaultName read the profile leniently, so a malformed or unreadable profile
looked like "no default". Delete could then remove the template the profile
names, and rename could report wasDefault false and skip the profile update.

Read it strict and raise a path-free 409 refusal (the original error rides as
its cause). Compute wasDefault before the file moves, and roll back through an
internal move that does not consult the profile, so a failed profile update
cannot block its own undo.

* fix(web): templates list ignores stale loads and drops a form for a vanished template

A slower earlier fetch could resolve last and replace a newer list, and after a
409 reload the open form kept pointing at a template the server no longer
listed, so a later template with that name would reopen it.

Both rules move into templates-state.mjs and are tested there: latestOnly()
gives each load a check that is true only for the newest, with stop() on
unmount, and pendingAfterReload() closes the form when its template is gone.

* fix(web): reload the templates list after a default is chosen

A reload started by a duplicate/rename/delete can still be in flight when busy
clears. Choosing a default then showed the new default only until that older
response landed with the old one. Start a newer load after a successful save so
latestOnly discards the older response; a failed save leaves the list alone.

* fix(web): close the template editor when its template leaves the list

The editor hides itself for a template the reloaded list no longer has, but
editing kept the name, so a later template with that name reopened it
unprompted. Reconcile it where pending already is, through editingAfterReload()
beside pendingAfterReload().

* fix(web): keep the templates list when a reload after a saved default fails

A failed reload replaced the whole gallery with an error even right after a
default was saved, when the list on screen was already correct. Now that case
shows the error and Retry beside the list; a first load and a reload after a
duplicate, rename or delete still replace it, since that list is stale.

The apply-then-reload order for a chosen default moves into settleChoice() and
the failure view into loadFailureView(), both tested as behavior. This replaces
the wiring test's [\s\S]*? pattern, which could cross the closing brace and so
pass even with the reload outside the success branch.

* fix(web): deleting a pack's last template no longer erases files it does not own

A pack is any folder under the templates directory that holds a template file,
so a template filed inside <userDir>/sections made a pack of the shared
partials' folder. Deleting its last template then removed that whole folder
recursively, taking the partials every flat template reads. A pack folder with
a note in it lost the note the same way.

Remove only the pack's own sections/ folder, then the folder itself if that
leaves it empty. Tests use real layouts: a misfiled template beside shared
partials and flat templates, a pack with an unrelated file, and a pack that
still has a second template.

* fix(web): never change a template that lives behind a symlinked pack folder

The core resolver follows a symlinked pack folder, so such a template lists
normally. Delete then removed the file in the link's target, the pack cleanup
removed the target's sections folder, and rmdir failed with ENOTDIR, leaving
the target empty. Rename wrote its new file into the target the same way.

Gate rename and delete on where the template really lives: refuse, with a
path-free message, when its folder resolves outside the user's templates
directory. Duplicating from such a template only reads it, so it still works.
Tests use a real symlinked pack with sections.

* fix(web): set the has-a-list ref from the load handler, not during render

listRef.current = templates ran in the component body, which React may replay or
discard. The failing-reload path only needs to know whether a list has ever
loaded, so record that when a load succeeds.

* fix(web): refuse to save a template whose file lives outside the user's directory

The editor's save writes through checkTemplateEdit's path, so a template in a
symlinked pack folder, or a template file that is itself a symlink to somewhere
else, would have the edit land in the link's target. Check where the file
really is before offering a path to write; reading it is unchanged.

* fix(web): keep unsaved editor text from being dropped by a card action

Rename or Delete on the open editor's own card removed its template, so the
reload unmounted the editor with the typed text; Edit on another card changed
the editor's key and remounted it the same way.

Lock those actions while an editor is open (editLocks), so the only way out is
Close, which already offers Discard and close. No new editor API: the section
cannot see the dirty state, and this avoids needing to.

* fix(web): disable Discard and close while the template editor is saving

The button unmounted the editor but could not cancel the PUT, so an edit the
user had just discarded could still be written. Disable it while busy.

* fix(web): move focus to the template editor on open and back on close

The editor renders after the whole gallery, so keyboard focus stayed on Edit and
a pointer user could miss where it appeared. Focus the textarea once the template
has loaded, and return focus to the Edit button that opened it when it closes.

Also tighten the wiring test: it now reads each action's own disabled expression,
so a button that keeps its title but loses its disabled guard fails. Checked by
removing the guard on Rename and watching the test fail.

* fix(web): refuse to save a template that is a symlink, even one pointing inside

The save replaces the file atomically, which turns a symlink into a regular file.
For a link to another file in the user's directory the gate accepted it, the
target never received the edit, and the alias stopped following it. Refuse with a
path-free message, as for a link that points outside, instead of writing through
it. The real file the link points at stays directly editable.
…#6856)

* refactor(discover-ats): share the board URL builders through lib/ats-board-url.mjs

Move each vendor's buildUrl into a dependency-free module, lib/ats-board-url.mjs,
and have discover-ats.mjs use it from there. The module adds buildBoardUrl(), the
canonical careers URL of a board from {ats, region, board, workday_host}: the EU
hosts for Greenhouse (job-boards.eu.greenhouse.io) and Lever (jobs.eu.lever.co),
and both Workday host forms (<tenant>.<wdN>.myworkdayjobs.com/<site> and
<wdN>.myworkdaysite.com/recruiting/<tenant>/<site>). discover-ats.mjs re-exports
buildBoardUrl and SLUG_RE.

The employer registry keys every row on this URL, so discovery and the registry
build the same URL for the same board. Dependency-free so the registry's CI gate
can run without installing packages.

tests/ats-board-url.test.mjs checks that every provider's own detect() accepts
the URL built for its ATS, for every region and Workday host form, and that
discovery's candidates equal what buildBoardUrl returns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(seeds): open employer registry: schema, validator, CI and a first 53 boards

seeds/employers/ is a community-maintained list of employer job boards the
local scanner can read: one pointer per board (employer, ATS, board), with no
postings and no personal data. Rows live in boards/<ats>/<initial>.ndjson, each
contribution has a manifest in contributions/, and CHANGELOG.md is generated
from the manifests. careers_url is the key and must equal buildBoardUrl() of the
row. The scanner does not read the registry yet.

- seeds/employers-validate.mjs: static validator (schema with unknown keys
  rejected, closed lists, canonical URL, board shape per ATS, aggregator
  domains, no contact details in names, each board listed once, sorted files,
  manifests, size limit). A board is listed once: careers_url is unique
  ignoring letter case, and a Workday board appears under one host form only;
  when a pull request adds a second spelling, the merged row stays and the new
  one is reported. With --base it also checks a pull request against the base
  branch: rows are never deleted, provenance is fixed, new rows come with a new
  manifest, and a contribution does not retire rows. Errors name file:line and
  the reason. --write-changelog regenerates CHANGELOG.md.
- .github/workflows/employers-validate.yml: pull_request, read-only, no
  secrets, pinned actions; runs the base branch's validator against the PR's
  files, or the PR's own validator only when the base has none (this PR).
- 53 boards hand-picked from boards already named in this repository
  (templates/portals.example.yml, and the Workday provider and discovery
  tests), each checked live. The three Workday boards cover both host forms and
  keep the site's case from the careers page's canonical link.
- tests/fixtures/employers-garbage/: a registry of known mistakes that the
  validator must report one by one.
- employer-removal issue form; CODEOWNERS for the registry and its gate.
- Data is CC BY 4.0 (seeds/employers/LICENSE); the code stays MIT. One line
  each in README.md and CONTRIBUTING.md; DATA_CONTRACT.md lists seeds/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(employers): move the registry to employers/, outside the updater

The registry is community data that grows with every contribution, so it
moves from seeds/employers/ to a root employers/ that is not in SYSTEM_PATHS:
a fresh clone has it, and update-system does not push each new batch of rows
to every install. The validator stays in seeds/ (system code, small) and now
reads employers/.

- validate-system-paths-coverage.mjs: employers/ joins web/ as a repo-only
  prefix (every tracked file must be classified), with self-tests that it is
  covered and never enters SYSTEM_PATHS. DATA_CONTRACT.md lists it under a new
  "Repository only" section instead of the system layer.
- Size: no hard limit any more. boards/ above 25 MB is a warning for a
  maintainer to look at how the data is stored, never an error.
- CHANGELOG.md: a contribution no longer edits it. The validator stops
  comparing it with the manifests; maintainers regenerate it on main with
  --write-changelog.
- GitHub Actions: with GITHUB_ACTIONS=true each finding is printed as
  ::error/::warning file=...,line=... (message and properties escaped as
  @actions/core does), and GITHUB_STEP_SUMMARY gets the counts and a table of
  the first 50 findings. Outside Actions the output is unchanged. The tests
  pin both and run the validator with a clean environment, since the suite
  itself runs in Actions.
- employers/README.md: the four-step hand-edit path is no longer how to
  contribute. It says a simpler path (the board's address, the employer's name
  and a form) opens in a few days, and until then to open an issue. Format,
  manifest fields, CC BY 4.0, provenance and removal stay.
- Paths updated in the workflow, CODEOWNERS, the removal issue form,
  CONTRIBUTING.md, README.md, seeds/README.md, comments, tests and the garbage
  fixture. test-all's personal-data scan allowlists employers/README.md, which
  names its licensor and the removal address by design.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…-style (#6515)

* fix(tests): support user-authored named CV template variants in theme-style (fixes #6514)

Signed-off-by: Vijay Sirohi <191354616+Vijaysir0hi@users.noreply.github.com>

* test(theme-style): assert token contract rules on discovered variant fixture

---------

Signed-off-by: Vijay Sirohi <191354616+Vijaysir0hi@users.noreply.github.com>
…der (#6603)

* feat(backup): add backup.mjs to copy the User Layer to a drive or folder

Copies the files DATA_CONTRACT.md lists as the user's (CV, profile,
portals.yml, data/, reports/, output/, jds/, ...) from the data root to a
backup folder. The destination is --dest, then CAREER_OPS_BACKUP_DIR, then
the cloud-drive folder found on the machine plus Backups/career-ops
(--provider google-drive|dropbox|onedrive|icloud, default google-drive).

- whole-file copies, so Google Drive's "mmap: Resource deadlock avoided"
  rsync failure cannot happen; nothing is deleted at the destination
- skips .env, data/Connections.csv, data/cache/ and lock/temp files
- refuses to guess between several signed-in drives (--account), requires
  the destination's parent to exist, and refuses a destination inside the
  data root or checkout
- --dry-run and --json; exit 0 ok, 1 copy failures, 2 no usable destination

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* feat(backup): back up data/Connections.csv too

The exclusion is dropped: the file is the user's own export and a backup
is a reasonable place for it. .env files, data/cache/ and lock/temp
files are still skipped.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* chore(backup): register backup.mjs in SYSTEM_PATHS

* feat(backup): address review - USER_PATHS, no default drive, opt-in Connections.csv, error handling

- Build the file list from USER_PATHS (plus config/local-paths.txt) instead of a second copy
- Require --dest, CAREER_OPS_BACKUP_DIR or --provider; no default cloud folder
- Leave data/Connections.csv out unless --include-connections
- Report a run that found nothing (exit 1); record per-file I/O errors and continue; skip symlinked parents
- Exit 2 on invalid options via a new validateFlags usageExitCode option (default unchanged)
- Report destination and data-root errors in the requested format

* fix(backup): keep agent-tool settings out of the backup unless --include-tool-config

USER_PATHS also lists .claude/settings.json, .claude/hooks/, opencode.json and plugins.local/,
which can hold API tokens. They are protected from the updater, not safe for a cloud drive.

* revert(backup): copy data/Connections.csv by default, drop --include-connections

* revert(backup): copy agent-tool settings like the rest of USER_PATHS, drop --include-tool-config

* feat(backup): data/Connections.csv is opt-in via --include-connections

* fix(backup): warn when a symlinked path is skipped

* fix(backup): case-insensitive exclusions, segment-aware containment, copy nested checkouts, warn on external data root, docs and Windows test guards

* fix(backup): leave .git and node_modules out of nested checkouts

* fix(backup): reject destinations holding the data root, symlinked destination folders, and --provider with --include-connections

* feat(backup): gate all third-party contact files behind --include-contacts

* fix(backup): runBackup enforces the destination safety check itself

* fix(backup): refuse --include-contacts for a --dest inside a synced drive folder

* fix(backup): match synced folders case-insensitively and document the detection limit

* fix(backup): soften --include-contacts help, use example.com in fixtures

* test(backup): match work@example.com in the account assertion

* fix(backup): refuse synced destination for contacts in runBackup too

* test(backup): redirect USERPROFILE as well as HOME in the runBackup fixture

* fix(backup): keep plugins.lock, warn on agent settings in synced folders, tidy dry-run doc and test

* fix(backup): exclude contacts.vcf, fix canonical() root slice, document the sync-folder caveat

* docs(backup): list all four contact files

* docs(backup): note checkout-only agent files need a separate backup

* fix(backup): back up generated CVs and configured template variants under templates/

* fix(backup): skip agent settings for synced destinations, create Backups under a fresh drive, survive unreadable folders

* fix(backup): use the shared walker and a non-recursive template listing

* docs(backup): fix stale comments and the provider folder note

* fix(backup): leave every .vcf export out unless --include-contacts

* fix(backup): warn about skipped agent settings only when a settings file was actually skipped

* fix(backup): resolve profile.yml through getProfilePath()

* fix(backup): include template variants selected by the CAREER_OPS_PROFILE override

* fix(backup): report template discovery errors; read the override profile only when set

* test(backup): make the no-override case detect an unrelated process-wide profile

* fix(backup): read profiles with the template resolver, report profile errors, refuse a destination holding the checkout

* fix(backup): report an invalid local-paths declaration as a failure

* fix(backup): report unreadable-parent entries, isolate tests from CAREER_OPS_PROFILE, correct never-copied docs

* fix(backup): hold back editor leftovers of contact files without --include-contacts

* fix(backup): also hold back emacs and vim swap leftovers of contact files

* fix(backup): never copy editor swap/autosave files; hold back env and contact copies

* fix(backup): treat hidden copies of contact files as contact files

* docs(backup): contact editor temp files stay excluded with --include-contacts

* fix(backup): keep an extension-only .vcf excluded

* fix(backup): warn when the active CAREER_OPS_PROFILE is outside the backup

* fix(backup): segment-aware profile check, native realpath, exit-2 wording

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(web): require screening answers before an apply handoff

The planner is still told never to fill visa, work authorization, or
salary, even when the profile already states them. The review screen
still promises a CV from the field label alone, and a finished fill is
still marked pre-filled without a handoff check.

* fix(web): hold the apply handoff until screening answers and the resume file are real

The planner fills work authorization, visa, and compensation from the profile
when those answers are stated. Citizenship, nationality, gender, race,
disability, and veteran stay blank. The fill route and the full-fill drive
open the employer window only when that handoff is allowed. A missing
required screening answer, or a claimed resume that never attached, stays
on the CareerOps page as a warning. The green pre-filled banner stays off.

* fix(web): hold the handoff on the field that is actually empty

A retry kept the previous empty-field warning after the form had opened. Current compensation was treated as a question this page must answer. Two fields with the same label shared one answer. Criminal history and the other legal questions the old prompt left blank had no rule after the profile-backed screening change.

The retry drops those two warnings before it asks again. Current, previous, prior, and last pay stay out of the gate. An answer with an id stays on that id. The planner still leaves the legal questions blank.

* fix(web): keep the handoff filter's issue type

The helper's note said it returned objects that only have an optional code. The apply page stores full issues, so the typecheck rejected the retry that drops the previous empty-field warnings.

* fix(web): keep the handoff gate on a résumé resolve failure

resolveSessionCv used to throw out of the full-fill drive before the gate ran. The résumé now counts as unattached, the gate still decides, and the error is one of the gaps on the done stream. A required résumé, or one that was claimed and did not attach, still blocks. A retry replaces a stale required-empty warning and leaves other issues in place. The visa, work-authorization, and salary prompt change is out of this branch. The fill route tests call the route, so a reversed gate fails.

* fix(web): keep a résumé warning on a completed handoff

An optional résumé that fails to resolve still hands off. That warning used to send the page back to the editor because any gap counted as a hold. A screening or résumé gap still does. The warning stays on the success screen.

* fix(web): say how to get a missing CV, and do not stack fill notes

A missing résumé tells the person to generate it from the report in another tab, then retry on this page. A second fill keeps one ai-filled note, and two résumé gaps with different messages both stay.

* fix(web): keep handoff gaps until the fill comes back

A failed fill leaves the last sponsorship and résumé warnings on the page. A finished fill replaces them. Two fields that share a label are both held, because the form check reports labels. The retry test checks that list instead of reading the component.
* fix(score): one score-to-action map, and unknown pay does not move it

Full evaluation, triage, inbox rank, and the PDF gate read lib/score-action.mjs.
Missing pay is unknown. A culture line with no evidence is unknown, not a default 3.
The written score number stays.

* fix(web): the score screen only claims levers that move the score

Hard stops hold the score down. CV gaps can. Requirement importance, posting legitimacy, and evidence confidence do not.

* fix(score): register the score modules in the system manifest

update-system only ships paths listed in SYSTEM_PATHS. The score table and the evaluation cleanup were missing, so an update would drop them.

* fix(score): a missing score does not become Skip

Number(null) is 0, so a report with no parseable score was rewritten to Skip. Both score tables now require a real number. The bold decision rewrite stays inside the (lead) section. Turkish and Vietnamese culture rules match the canonical unknown rule, and still cap at 2/5 when absence is a negative. The pay-sentence test uses includes instead of a regex that escaped only dots.

* fix(score): keep an explicit culture penalty and flag an unrecalculated score

A culture line that caps the dimension at 2/5 because deprioritize_if_absent
is set is the user's call, so it is no longer replaced with unknown. When a
default-3 line is neutralized, the report now says the global score was not
recalculated and may still count that 3.

* fix(score): don't treat a negated cap as a culture penalty

* fix(score): a denied 2/5 cap is still a default culture line

A line can mention 2/5 while saying the cap is not set. The low-score check ran first, so that line was kept and the report never said the written score may still count a default 3.

The cap phrase is removed when the line says it does not apply. A remaining 0-2/5 is still a penalty and stays as written.

* fix(score): keep a CV budget and flag an invented pay estimate

A posting with no salary was dropping every line that contained a figure, including a Block B row that quoted a CV budget. Only an invented compensation claim is removed now.

When that claim said it moved the score, the report carries a score-check flag. The number stays as written. Hard stops are described as forcing Skip without changing the score.

* fix(score): keep CV payroll evidence and a denied score claim

A CV row that quotes a payroll budget stays. A salary sentence that says it moved the score is removed and flagged. "Compensation does not affect the score" stays, with no score-check flag.

* fix(score): drop a comment that restates the CV-row check

The early return already says a CV row is not invented pay.

* fix(score): keep pay, yaml, and localized leads honest

A funding figure is not a salary. The nested score_evidence key is the one rewritten. A score claim is dropped as a whole sentence. A culture line that already says unknown stays. A Polish or Danish lead follows the score, and an unknown call is flagged.

* fix(score): count a real salary, and leave the other lines alone

A trailing currency is pay. A customer base, annual revenue, or product range is not. A training-data role line and a culture note stay. "No evidence of" one trait is not a missing culture score. The period check no longer replaces a sentence with itself.

* fix(score): leave a training-data accuracy line in place

The word estimate next to training data was enough to drop the line. It now drops only when that line also states pay or a money figure.

* fix(score): do not read the word try as a currency

TRY was matched without regard to case, so a line like "25 try-out" counted as stated pay. Only the uppercase code counts. 120000 TRY still counts, and 500.000 kr still counts.

* fix(score): a hard stop still writes Skip when the score is missing

recommendationFor returns null when the score did not parse, so the
report kept Apply. The stop now forces Skip in the written decision.
A missing score with no stop stays as written.

* fix(score): competitive salary does not adopt a nearby funding figure

A pay word in the same sentence was enough to treat a funding amount
or a headcount as stated pay. The bare phrase no longer does that.
A figure attached to the salary still counts.

* fix(score): keep a described salary and ignore a later headcount

"Competitive salary for a team of 10,000 employees" kept the salary phrase because "of" matched later in the sentence, so the headcount counted as pay. "of" and ":" now have to follow the phrase. A salary figure can sit after a description. A funding-sized figure still has to sit next to the pay word.

* fix(score): do not let a salary word adopt a funding figure across a semicolon

"Funding: $50M; salary band unpublished" counted as stated pay because the salary word sat within 24 characters of the funding figure. A semicolon now ends the clause, so that figure stays funding. A salary figure in its own clause still counts.
RFC 5545 section 3.3.14 says the seconds of a UTC offset must not be
60. parseUtcOffset accepted second === 60, so an invite like +053060
was converted with a made-up offset instead of being reported as an
unknown time zone.

Fixes #6751
…egex (#6848)

* fix: support bullet point in score summary extraction

* fix: remove stray whitespace in summary regex

* fix: replaced hyphens with bullet points

* fix: typo in bulleted score summary in eval-score-summary.test.mjs
…adDom() (#6833)

challengeFailure() refuses a Cloudflare challenge by the document title,
which readDom() reports as pageTitle beside the <h1> it uses as title
(#6798). The challenge test hands that title over from a fake page, so a
regression in the document.title read inside readDom() would still pass
there (CodeRabbit's note on #6798).

This sets a real page with <title>Just a moment...</title> and
<h1>Humans only</h1>: readDom() reports both, and challengeFailure()
refuses the title it read on an HTTP 403. With the document.title read
removed, both checks fail. Skipped with a warning where Chromium cannot
launch, as browser-extract-unavailable is.
…rd (#6837)

The guard walked the working tree with collectMjsFiles and read every
.mjs it listed. plugins-cli-usage.test.mjs creates and deletes a fixture
in the gitignored plugins.local/ in the real repo root, so in a full run
the read could land after the delete and throw ENOENT.

List files through trackedFiles instead, which reads the index and skips
an indexed path missing on disk. Add a fixture-repo test, and add the
suite to the guards mjs-files.test.mjs requires to import trackedFiles.

Fixes #6836

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…debase (#3952)

* fix(paths): anchor relative tracker overrides to the codebase

Resolve CAREER_OPS_TRACKER relative to the codebase root before canonicalizing it, consistently for reads and writes. Match the documented contract independently of cwd or an external data root.

Add existing/missing target, absolute override, symlink and fallback coverage. Users relying on the old cwd-relative behavior must use an absolute override to keep that target.

* Align web tracker overrides with the selected code root

* Resolve web tracker aliases from the selected code root

Use resolveCodeRoot in both public tracker readers and cover external engines with realistic Via rows and header-only tracker fixtures.
AdilAzhariOmsan and others added 4 commits November 15, 2026 07:25
The board page carries all 41 openings inline (about 2 MB) and took
6-48s to arrive in repeated measurements, so the 10s default request
timeout aborts it on most runs and verify-portals reports a network
error. With timeout_ms: 60000 it read 41 live jobs on 7 of 7 runs;
without it, 1 of 3.
…or the same posting (#7084)

* fix(url-key): stop a re-evaluation from adding a second tracker row for the same posting

* fix(url-key): letter case alone does not make two posting URLs one page

* fix(url-key): fold a URL-shaped key only on the ATS hosts canonicalJobKey folds
…RL (#7086)

* fix(precheck-duplicate): match another spelling of the same posting URL

* fix(precheck-duplicate): match only an extracted posting id, not a lowercased path

* fix(precheck-duplicate): also match an ATS form like a Lever /apply page

* fix(precheck-duplicate): fold a URL-shaped key only on the ATS hosts canonicalJobKey folds
* fix(calibrate): count a row closed by close-stale as no_response

A row closed by close-stale resolves as  while its status is unchanged.

* fix(calibrate): check the close-stale status before the tracker status

A row closed by close-stale to Responded now resolves as no_response.

---------

Co-authored-by: Zain Baloch <zain@Zains-MacBook-Pro.local>
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 15, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ❌ Failed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ❌ Failed 2026-11-15T07:56:55.121675Z
We couldn’t complete this review. Comment @codex review to try again.
7d41185 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

hakarrr and others added 2 commits November 15, 2026 09:05
Both entries were scan_method: websearch since the template was seeded,
but each has a public Greenhouse board (gongio, 89 jobs; dialpad, 16
jobs). With api: set, scan.mjs reads them through the zero-token
greenhouse provider instead of handing them to the agent's web search.
careerops-ledger and others added 13 commits November 15, 2026 16:58
#7106)

Bumps [golang.org/x/sys](https://github.com/golang/sys) from 0.49.0 to 0.50.0.
- [Commits](golang/sys@v0.49.0...v0.50.0)

---
updated-dependencies:
- dependency-name: golang.org/x/sys
  dependency-version: 0.50.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Bumps [actions/setup-node](https://github.com/actions/setup-node) from 7.0.0 to 7.1.0.
- [Release notes](https://github.com/actions/setup-node/releases)
- [Commits](actions/setup-node@v7...949feb2)

---
updated-dependencies:
- dependency-name: actions/setup-node
  dependency-version: 7.1.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Bumps @paper-design/shaders-react from 0.0.80 to 0.0.82.

---
updated-dependencies:
- dependency-name: "@paper-design/shaders-react"
  dependency-version: 0.0.82
  dependency-type: direct:production
  update-type: version-update:semver-patch
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Bumps [@types/node](https://github.com/DefinitelyTyped/DefinitelyTyped/tree/HEAD/types/node) from 25.9.5 to 26.6.6.
- [Release notes](https://github.com/DefinitelyTyped/DefinitelyTyped/releases)
- [Commits](https://github.com/DefinitelyTyped/DefinitelyTyped/commits/HEAD/types/node)

---
updated-dependencies:
- dependency-name: "@types/node"
  dependency-version: 26.6.6
  dependency-type: direct:development
  update-type: version-update:semver-major
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
* chore(deps): bump playwright from 1.64.1 to 1.65.0

Bumps [playwright](https://github.com/microsoft/playwright) from 1.64.1 to 1.65.0.
- [Release notes](https://github.com/microsoft/playwright/releases)
- [Commits](microsoft/playwright@v1.64.1...v1.65.0)

---
updated-dependencies:
- dependency-name: playwright
  dependency-version: 1.65.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

* chore(docker): pin the Playwright images to 1.65.0 to match package.json

The root Dockerfile's FROM tag and --save-exact pin, and the digest-pinned
cv-visual image, still named 1.64.1 after the dependency bump.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Santiago <santiago@me.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…oyer registry too (#7088)

A pull request that adds a company to tracked_companies of
templates/portals.example.yml, on a board the registry accepts, must add
the same board to employers/inbox/ or find it already in employers/. The
check compares against HEAD^1 (main in CI), so the boards already in the
template are not flagged, and the failure prints the exact inbox line to
add. Renames and companies on ATSs the registry does not support pass.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Follow-up to #7070. The two tests for SEEK's locations array and work
arrangement sat inside tests/scan-local-scanners.test.mjs; they now live
in tests/scan-seek-locations.test.mjs, unchanged. The shared file drops
the buildLocationFilter import only they used.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…nce from the Step 4b sources (#7080)

Follow-up to #6964. The India layer's source table pointed at cv.md,
article-digest.md and config/profile.yml in the project root, and its
apply.md counted experience, batch and marks from cv.md. The canonical
modes read them from {DATA_ROOT} and fill experience and education from
the Step 4b sources (the tailored CV first, cv.md only for a section it
omits). A test in market-in-rules pins both.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(states): add Speculative canonical state to JS pipeline contract

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(dashboard): render and sort the Speculative state

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* style(dashboard): gofmt pipeline.go + document speculative sort tie

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(gambit): add the create-a-role gambit mode

* fix(gambit): numeric Conviction score, clarify fit leg, cite Speculative as canonical, add keywords section

* feat(gambit): route the gambit mode in the career-ops skill

* docs(gambit): document and CI-enforce the gambit mode + Speculative state

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(states): teach merge-tracker.mjs the Speculative state (was silently downgrading to Evaluated)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(gambit): add Step 0.5 entity disambiguation (from first live run)

The first live gambit (Crouton) hit a name collision — a recipe app, a
blockchain firm, and the target HR-tech startup all answer to 'Crouton'.
Adds a disambiguation gate before research so the run can't burn budget on
the wrong company. Spec data-flow updated to match.

# Conflicts:
#	docs/superpowers/specs/2026-06-07-gambit-mode-design.md

* feat(profile-template): scaffold Gambit Playbook + stage-calibration hooks

The gambit mode reads two `_profile.md` sections that the template did not
provide: `## Your Natural Level & Company-Stage Calibration` and
`## Gambit Playbook`. Add bio-free, fill-in-yourself scaffolds for both so the
feature resolves its references on a fresh install instead of silently
improvising. Pure placeholders -- no personal data.

* fix(gambit): accept Speculative in the dashboard status guard and check states.yml wiring

The dashboard's isCanonicalStatusName switch must list every states.yml id.
verify-pipeline.mjs and merge-tracker.mjs now derive their states from
templates/states.yml, so the speculative checks assert that wiring instead
of a hard-coded literal.

* fix(gambit): reserve the report number, write a headed addition, record only sent pitches as Speculative

- Claim the number with reserve-report-num.mjs once the thesis is chosen,
  before any artifact, and release it after the merge (no max+1 by hand).
- Map the Step 8 verdict to the tracker: send -> Speculative,
  hold -> Evaluated, skip -> SKIP.
- Tracker addition gets the header row and an empty url column (#3702):
  a careers-page URL would be a dedup key that merges two gambits.
- Research input is untrusted external content.
- Profiles without the new _profile.md sections fall back to stated
  defaults and the user is told what to add.

* chore(gambit): ship modes/gambit.md with updates and list it in the modes catalog

* feat(web): support the Speculative status

Status lists, pipeline tab, assistant preamble, spec/gambit aliases, its own
sankey leaf off Tracked (not a submitted application) and a pink status dot.

* feat(tracker): classify Speculative in stats, patterns and state contracts

Speculative is pre-application: counted like Evaluated (not an active or
submitted application, not in motion for dedup), listed in the canonical
state lists and docs.

* fix(dashboard): give Speculative its own status colour

It shared Mauve with Assessment after the merge; use the unused Pink.

* fix(gambit): record Speculative only once sent and keep it from counting as submitted

- dedup-tracker: rank speculative above evaluated, so a later evaluation of the
  same role cannot demote a pitch that went out
- gambit: use a no-posting URL header, research links stay in the thesis
- gambit: send maps to Evaluated until the candidate confirms sending
- tracker: a reply to a pitch stays Speculative, since Responded counts as submitted
- tests: move the state checks to tests/speculative-state.test.mjs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(plugins): sync the gambit routing into the directory plugin SKILL.md copies

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(dashboard): sort Speculative right after Evaluated, as the group order says

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(gambit): keep a later tracker status when a sent pitch is confirmed

- modes/gambit.md: record Speculative with --expect-status Evaluated; if the
  row already moved on (Applied, Interview…), keep it instead of demoting it.
- AGENTS.md: the routing row says Evaluated first, Speculative only after the
  user confirms the pitch was sent.
- pipeline_pending_test.go: the narrow-terminal Inbox test from main counts
  one more tab now that SPEC sits between EVALUATED and INTERVIEW.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(readme): list gambit in the Usage command tables

main's readme-usage-commands test requires a row for every routed mode.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(gambit): run gambit in the interactive session and stop before documents on a SKIP

The Step 3 checkpoint asks the candidate to choose a role thesis before any
document is generated. A worker cannot ask the candidate anything, so gambit
leaves the delegated list and runs in the interactive session; the router test
keeps it out. A new Step 4b checks the evidenced-need leg before the CV, cover
letter and pitch package, so a SKIP saves no send-ready drafts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(gambit): a late SKIP deletes the CV and cover letter this run wrote

If the need only fails at Step 8, Step 5 has already written the CV PDF and
the cover letter to output/. Step 9 says a SKIP saves nothing there, so the
late gate now removes those two files.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Santiago <santiago@me.com>
Co-authored-by: Santiago <hi@career-ops.org>
* fix(salary-gap): parse verbose advertised_comp strings correctly

parseAmount() only stripped currency symbols and a single trailing
3-letter alpha token, so real-world JD comp strings like:
  'Annual Annual Salary Range: $165,000 - $190,000'
  '$204.4K – $352K • Offers Equity'
  '$220,000 - $400,000 USD per year'
all silently returned null instead of the actual range, breaking
salary-gap.mjs's comp-gap math for those reports.

Now strips a leading label (up to the last colon), a trailing
bullet/suffix, and a trailing cadence phrase ('per year', 'annually')
before the existing ISO-currency-code strip runs. Added regression
coverage in the existing --self-test suite.

* fix(salary-gap): address review on label, note and currency stripping

- Treat a colon as a label only when it comes before the first digit,
  so "$150,000 (note: negotiable)" keeps its amount.
- Drop a trailing digit-free parenthetical note ("(Spain)"), except one
  naming a non-annual period, which still fails closed.
- Strip a trailing three-letter token only when it is a real ISO-4217
  code (Intl.supportedValuesOf), so "100k tax" stays unparseable.
- Self-tests for each, plus "per year USD" order and a label in front
  of a period-grouped range.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(salary-gap): fail closed on non-annual periods in verbose comp text

- Reject a non-annual period on the whole string before any label,
  bullet suffix or note is cut away, so "Monthly salary: 5,000 USD"
  and "5,000 USD • per month" no longer read as annual.
- Recognise daily and weekly periods, so "500 CAD (per day)" fails
  closed instead of parsing.
- Take the first real ISO-4217 code as the report's currency guess, so
  "OTE: 100k USD" is USD rather than OTE.
- Move the new parser cases into tests/salary-text-parse.test.mjs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(salary-gap): scale bracketed magnitudes and accept an RMB suffix

The bracket-note strip dropped a magnitude qualifier with the other
notes, so "80–100 USD (in thousands)" read as 90 USD. A trailing
"(thousands)", "(in K)" or "(millions)" now scales the figure, and a
value that also carries a "k" fails closed.

The suffix strip accepted only ISO-4217 codes, so "600k RMB", which
main parses, returned null. RMB now counts as a currency there.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Chris Coppola <chris@chriscoppola.me>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(scan): per-source funnel table in the run summary

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* fix(scan): count every offer in exactly one per-source column

Duplicates found after verification go under seen, failed title-rescue
attempts under title, and experience rejections under other.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* fix(scan): clear a board's failed mark after a successful retry; reword stale comment

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
@hakarrr hakarrr closed this Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.