# doc-wiki benchmark — repo + issue manifest # # 25 real merged PRs (8 Django + 9 Cal.com + 8 Mastodon), every fix_commit SHA # verified live via `gh api` on 2026-05-31. Each entry includes pr_url / # issue_url / loc_changed / why_chosen for review traceability — the harness # only consumes id / title / body / fix_commit / test_path; extras are ignored. # # Django caveat: django/django uses Trac (not GitHub Issues) as its canonical # tracker. issue_url points at code.djangoproject.com/ticket/NNNN; the body is # the Trac description verbatim. PRs encode the ticket via "Fixed #NNNN -- ". # # See benchmark/PLAN.md for selection criteria + methodology. repos: - id: django url: https://github.com/django/django lang: python python: "3.12" setup: # Django main requires Python 3.12+; pylibmc in test reqs needs system libmemcached # (skip via grep -v if libmemcached isn't installed on the host) - "python3.12 -m venv .venv" - ".venv/bin/pip install --quiet -e ." - "grep -v '^pylibmc' tests/requirements/py3.txt > /tmp/reqs.txt && .venv/bin/pip install --quiet -r /tmp/reqs.txt" # runtests.py expects fully-qualified dotted path: .tests.. test_cmd: "cd tests && ../.venv/bin/python runtests.py --verbosity=2 {test_path}" issues: - id: "37036" title: "Fixed #37036 -- Fixed TypeError when using defer() with FETCH_PEERS on FK fields." body: | Using a related manager (e.g. author.books) inheriting a fetch mode of FETCH_PEERS from its origin instance, and then chaining defer(), gives a TypeError when trying to fetch that instance. Root cause: DeferredAttribute.fetch_many() used .values_list(attname).in_bulk() to batch-load deferred field values for peer instances. Without flat=True, in_bulk() returns values as single-element tuples (e.g., (None,)) rather than scalar values, which were then incorrectly set on model instances via setattr. fix_commit: "820c7d3248af6afbdf3390c97b29e41ba126a421" test_path: "tests/defer/tests.py::DeferTests::test_defer_fk_fetch_mode_fetch_peers" pr_url: "https://github.com/django/django/pull/21110" issue_url: "https://code.djangoproject.com/ticket/37036" loc_changed: 30 why_chosen: "Tiny one-line fix (add flat=True) with a focused regression test on a real ORM data corruption bug." - id: "37047" title: "Fixed #37047 -- Fixed crash in Query.orderby_issubset_groupby for descending and random order_by strings." body: | #26434 caused a crash. It can be reproduced with: User.objects.values("is_staff").annotate(latest=Max("date_joined")).order_by("-latest").count() You should see the following exception: django.core.exceptions.FieldError: Cannot resolve keyword '-latest' into field. Choices are: activity_logs, date_joined, email, first_name, groups, id, is_active, is_administrator, is_staff, is_superuser, last_login, last_name, latest, logentry, module_access, password, user_permissions, username fix_commit: "a284a49153f005f2a7af087025e5112ba06cbd5f" test_path: "tests/aggregation_regress/tests.py::AggregationTests::test_count_preserve_group_by" pr_url: "https://github.com/django/django/pull/21121" issue_url: "https://code.djangoproject.com/ticket/37047" loc_changed: 25 why_chosen: "Regression caused by a recent ORM optimization — concrete crash with a clear repro." - id: "37057" title: "Fixed #37057 -- Adjusted UniqueConstraint handling of UNKNOWN condition." body: | When a UniqueConstraint has a condition that references a nullable field (e.g., condition=Q(cash_register_type=10) on a nullable PositiveSmallIntegerField), Django's UniqueConstraint.validate() incorrectly reports a constraint violation if the instance being saved has cash_register_type=None and another record matching the condition already exists. Creating a Device(cash_register_type=None) raises ValidationError because Q(condition).check() does not properly handle UNKNOWN (three-valued logic) results from SQL evaluation of nullable fields. fix_commit: "61a62be313e395ce1265132bfc99f51476fb3c95" test_path: "tests/constraints/tests.py::UniqueConstraintTests::test_validate_nullable_condition" pr_url: "https://github.com/django/django/pull/21152" issue_url: "https://code.djangoproject.com/ticket/37057" loc_changed: 29 why_chosen: "Three-valued-logic edge case in ORM constraint validation — exactly the sort of nuanced fix where having a wiki of constraint semantics should help." - id: "37016" title: "Fixed #37016 -- Avoided propagating invalid arguments from When() to Q()." body: | The Security Team just closed a report extrapolating from 279f8b9557f0fef9790822b0c38164fc9dfcab2a arguing that When() is missing the same protection we gave to filter() and friends, whereby we raise errors for the _connector and _negated arguments instead of passing them down to Q(). Add this same protection to django.db.models.expressions.When so that hostile _connector and _negated kwargs raise a TypeError instead of being silently passed to Q(). # CORRECTED 2026-06-03: prior SHA 82a2465f71... is an unrelated commit in PR 21046 # (Tim Graham's DatabaseFeatures.disallowed_simple_test_case_connection_methods). # Actual #37016 fix is 3b161e6096 (varunkasyap, Apr 2 2026, "Avoided propagating # invalid arguments from When() to Q()"). Multiple subagents independently caught # this by `git log --grep`; using their correction. fix_commit: "3b161e60964aff99eddcd2627a486d81c1836b3a" test_path: "tests/expressions_case/tests.py::CaseWhenTests::test_when_rejects_invalid_arguments" pr_url: "https://github.com/django/django/pull/21046" issue_url: "https://code.djangoproject.com/ticket/37016" loc_changed: 16 why_chosen: "Defensive security follow-up with an obvious test surface — verifies the agent can mirror an existing pattern onto a sibling API." - id: "36912" title: "Fixed #36912 -- Added connector validation to Q.create()." body: | In 98e642c69181c942d60a10ca0085d48c6b3068bb, we mitigated a SQL injection vector for user-controlled arguments to filter() and friends (CVE-2025-64459) by adding validation for the _connector argument. We deliberately avoided adding the same validation to Q.create(), because Q.create is an undocumented internal not to be used with user-controlled field names and was created specifically for the purpose of speed. The Security Team then received more than one report extrapolating from CVE-2025-64459, suggesting that Q.create was missing the same validation. Although we don't consider this a security vulnerability, we would be interested to evaluate if adding validation to Q.create to match Q.__init__ would be cheap enough to implement. fix_commit: "fb292a549371a7c011c25b1a10fe5d25c579814a" test_path: "tests/queries/test_q.py::QTests::test_connector_validation" pr_url: "https://github.com/django/django/pull/21022" issue_url: "https://code.djangoproject.com/ticket/36912" loc_changed: 17 why_chosen: "Short, scoped refactor where a shared _check_connector helper protects both Q() and Q.create()." - id: "36966" title: "Fixed #36966 -- Fixed ValueError when `query_params` and `follow` are used on test client." body: | When the test client is called with query_params and follow=True, a ValueError is raised when following the redirect. _follow_redirect populates data from the redirect URL's query string via QueryDict(url.query). Since query_params from the original request is never reset, both data and query_params end up non-empty, triggering: ValueError: query_params and data arguments are mutually exclusive. # CORRECTED 2026-06-03: prior SHA dc467fdc... was for ticket #36991, not #36966. # Verified via git log --grep "36966" by the django baseline subagent. fix_commit: "6c95af5c9d294a20d09f080d7c144e3b7362d65f" test_path: "test_client.tests.ClientTest.test_follow_redirect_with_query_params" pr_url: "https://github.com/django/django/pull/20831" issue_url: "https://code.djangoproject.com/ticket/36966" loc_changed: 8 why_chosen: "Crisp one-line bug in the test client itself with a single matching regression test — easy to score and unambiguous." - id: "37024" title: "Fixed #37024 -- Made SITE_ID system check validation use Site._meta.pk." body: | To catch a programmer mistake where the SITE_ID setting is defined with an incorrect type (e.g. SITE_ID = "1"), #31802 added a system check that only allows SITE_ID to be an int or None. This forces third-party databases with non-integer primary keys like MongoDB (which uses ObjectId) to silence this check. It would be more appropriate if this check performed the validation based on Site._meta.pk rather than with hardcoded types. Recommended implementation: use Site._meta.pk.to_python(settings.SITE_ID) to validate the SITE_ID, also checking settings.SITE_ID against Site._meta.pk.to_python(settings.SITE_ID) to catch invalid values that to_python() coerces to the correct type. fix_commit: "179aa21b6dfeb9a0560a8c2bbfcf056301fc619f" test_path: "tests/sites_tests/tests.py::SitesFrameworkTests::test_check_site_id_incorrect_type" pr_url: "https://github.com/django/django/pull/21069" issue_url: "https://code.djangoproject.com/ticket/37024" loc_changed: 52 why_chosen: "Mid-sized refactor of a system check with two distinct positive/negative test paths — exercises agent's ability to follow a maintainer-proposed design." - id: "36961" title: "Fixed #36961 -- Fixed TypeError in deprecation warnings if Django is imported by namespace." body: | Calling django_file_prefixes will fail when Django is imported as a namespace package (django.__file__ is None): TypeError: expected str, bytes or os.PathLike object, not NoneType django.__file__ can be None when Django is imported as a namespace package — the AttributeError-only guard in django_file_prefixes() doesn't catch this case. Use getattr with a default to handle both. fix_commit: "c1d8646ec219b8b90ebdd463f40e5767876658a0" test_path: "tests/deprecation/tests.py::DjangoFilePrefixesTests::test_no_file" pr_url: "https://github.com/django/django/pull/20789" issue_url: "https://code.djangoproject.com/ticket/36961" loc_changed: 12 why_chosen: "Tiny try/except to getattr swap with a trivial regression assertion — sanity-check candidate." - id: cal-com url: https://github.com/calcom/cal.com lang: typescript node: "20" setup: - "corepack enable" - "yarn install --immutable --silent" - "yarn workspace @calcom/prisma db:generate" test_cmd: "yarn test {test_path} --silent" issues: - id: "28616" title: "fix: block localhost and loopback addresses in SSRF protection" body: | Webhook subscriber URLs are validated with `z.string().url()` which accepts any valid URL structure regardless of scheme. This means `javascript:`, `ftp:`, `file:`, and other non-HTTPS schemes pass validation and can be stored as webhook endpoints. The SSRF protection layer should also reject loopback hostnames (127.0.0.1, ::1, 0.0.0.0) alongside `localhost`, since these can be used to bypass simple hostname-based blocklists. fix_commit: "ad791f8ea5baaef062d13e5aeabf9a3a0e61923e" test_path: "packages/lib/ssrfProtection.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/28622" issue_url: "https://github.com/calcom/cal.com/issues/28616" loc_changed: 27 why_chosen: "Security-defense fix with a co-located test file already exercising hostname blocklist." - id: "27988" title: "fix: V2 API returns \"Booking limits must be in ascending order\" due to unfiltered `disabled` property" body: | Unable to enable `seatsPerTimeSlot` on an existing event type via the API. Request fails with TRPC error "Booking limits must be in ascending order" regardless of payload structure. Root cause: bookingLimitsCount transformer leaks the UI-only `disabled` boolean into the interval-limits validator, breaking the ascending-order check. fix_commit: "4b247640a87e78bdfd438fc6a76ad4dceaa93042" test_path: "apps/api/v2/src/ee/event-types/event-types_2024_06_14/transformers/api-to-internal/api-to-internal.spec.ts" pr_url: "https://github.com/calcom/cal.com/pull/28035" issue_url: "https://github.com/calcom/cal.com/issues/27988" loc_changed: 34 why_chosen: "API-v2 regression with concrete payload repro and a transformer-level spec test." - id: "22319" title: "fix: refresh slots on timezone change for booker timezone restrictions" body: | The problem here is that we are relying on booker's timezone to change the slots that are available. We don't refresh the available slots on timezone change which I believe wasn't a problem till now because we did the conversion at client side and it was fine. needSlotsRefresh = timezoneChanged && restrictScheduleEnabledWithBookerTimezone fix_commit: "4c73695d3ae3dd4dbb0b364aaf406a94fa65d699" test_path: "packages/features/bookings/Booker/hooks/useStableTimezone.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/27491" issue_url: "https://github.com/calcom/cal.com/issues/22319" loc_changed: 140 why_chosen: "Cross-package timezone-handling change with a dedicated hook test — exercises Prisma/Next.js plumbing doc-wiki coverage should help with." - id: "27963" title: "fix: Correct hours-to-days conversion in convertToNewDurationType" body: | Logic error in packages/lib/convertToNewDurationType.ts. The conversion from hours to days is currently multiplying by HOURS_IN_DAY (24) instead of dividing. convertToNewDurationType("hours", "days", 24) returns 576 (24 * 24) instead of 1 (24 hours = 1 day). fix_commit: "9d4cb08c556af66e14c3ba44961b86de2650e6e6" test_path: "packages/lib/convertToNewDurationType.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/27964" issue_url: "https://github.com/calcom/cal.com/issues/27963" loc_changed: 86 why_chosen: "Pure-function arithmetic bug with a single-file fix and a co-located test — sanity-check candidate." - id: "28764" title: "fix: prevent negative wait time in rate limit error message" body: | The `checkRateLimitAndThrowError` function can generate a negative wait time in its error message when the `reset` timestamp is in the past. This results in incorrect and confusing user-facing messages like "Rate limit exceeded. Try again in -5 seconds". The wait time displayed should always be non-negative (clamp to 0 if reset is in the past). fix_commit: "3c52f5723105a9c65bb24c4fd3ef12f1d85a2485" test_path: "packages/lib/checkRateLimitAndThrowError.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/28765" issue_url: "https://github.com/calcom/cal.com/issues/28764" loc_changed: 19 why_chosen: "One-line Math.max clamp with a single new test case." - id: "28610" title: "fix: join Reply-To addresses as string for SMTP compatibility" body: | When using SendLayer as the SMTP provider for a self-hosted cal.com instance, all booking confirmation and notification emails fail with `500 5.0.0 ERROR` at the SMTP `DATA` command. Root cause: `getReplyToHeader()` returns `replyTo` as an array when there are multiple email addresses. Nodemailer serializes this as a comma-separated Reply-To header, which SendLayer's SMTP server rejects. The fix is to return a single comma-joined string instead of an array. fix_commit: "fbf6510dd85be1a4c5fad9eb88013df30f875812" test_path: "packages/lib/getReplyToHeader.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/28611" issue_url: "https://github.com/calcom/cal.com/issues/28610" loc_changed: 94 why_chosen: "Concrete provider-compat bug with a new dedicated test file (92 of 94 LOC are tests)." - id: "28034" title: "fix: handle JSON string name field from URL prefill for firstAndLastName variant" body: | When a `name` field with `firstAndLastName` variant is passed as a JSON string from URL prefill (e.g., name={"firstName":"John","lastName":"Doe"}), the preprocessing does not currently handle parsing it from a JSON string into the expected {firstName, lastName} object before the Zod schema validates it. The schema then bails with a type error. A previously-skipped test exists at packages/features/bookings/lib/getBookingResponsesSchema.test.ts:336 ("firstAndLastName variant to fullName when passed as JSON string from URL prefill") — un-skip it once the fix lands. fix_commit: "0eb2c15b393c3d8597263fa4900e34e6f1225622" test_path: "packages/features/bookings/lib/getBookingResponsesSchema.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/28039" issue_url: "https://github.com/calcom/cal.com/issues/28034" loc_changed: 136 why_chosen: "Issue explicitly names the skipped test the fix should activate — the harness can score success literally by un-skipping that test and watching it pass." - id: "20358" title: "fix: use WEBAPP_URL for booking confirmation redirects to fix localhost behind proxy" body: | After accepting a booking via the link in the e-mail I am redirected to https://localhost:3000/booking/IDredacted?error=Booking%20already%20confirmed NEXTAUTH_URL and NEXT_PUBLIC_WEBAPP_URL are set correctly. The actual acceptance is processed correctly. The booking already confirmed error comes because our E-Mail Security system while analyzing the e-mail calls the booking link. But the error should not point to localhost. fix_commit: "8238d4ffcd47e3c3a9e8b394bf5c83cadcca9157" test_path: "apps/web/app/api/link/__tests__/route.test.ts" pr_url: "https://github.com/calcom/cal.com/pull/28144" issue_url: "https://github.com/calcom/cal.com/issues/20358" loc_changed: 95 why_chosen: "Real production reverse-proxy regression — the agent must learn the WEBAPP_URL pattern." - id: "19163" title: "fix: defer email validation to after first blur on signup form" body: | Currently, when a user is creating an account for the first time, `Invalid email` is thrown even if user enters a single character and it persists until user types a valid email. Ideally the `Invalid email` should be thrown when the email doesn't follow the emailRegex AND has an error in the pattern. Expected: Validation should only trigger after the user has interacted with a field and moved focus away (onBlur), not on every keystroke. fix_commit: "cdf901fee31248856b0c4856eaa51b6904a72284" test_path: "apps/web/modules/signup-view.test.tsx" pr_url: "https://github.com/calcom/cal.com/pull/27765" issue_url: "https://github.com/calcom/cal.com/issues/19163" loc_changed: 105 why_chosen: "Long-standing UX bug closed by a one-line fix plus a dedicated 103-line test — the harness can attribute success specifically to the touched-state guard." - id: mastodon url: https://github.com/mastodon/mastodon lang: ruby ruby: "3.3" setup: - "bundle config set --local without 'development production'" - "bundle install --quiet" - "yarn install --immutable --silent" test_cmd: "bundle exec rspec --format progress {test_path}" issues: - id: "37948" title: "Fix poll expiration notification being re-triggered on implicit updates" body: | Steps to reproduce: 1. Vote in some poll 2. Receive notification that poll has ended 3. Clear up that notification 4. Receive it again 5. Clear it up 6. Receive it again 7. Infinite loop Expected: One time notification Actual: Unending notifications about the same poll ended fix_commit: "9b4a09f760616a8ced7101eaedd730852ea89410" test_path: "spec/services/activitypub/process_status_update_service_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/38078" issue_url: "https://github.com/mastodon/mastodon/issues/37948" loc_changed: 47 why_chosen: "ActivityPub-side regression with an explicit reproduction sequence and a focused service-spec test." - id: "38203" title: "Add support for FEP-2c59" body: | In split domain setups (ActivityPub documents living on another domain than the one used in nicknames / WebFinger handles) discovery of this split-domain setup and validation that both parties consent to the association can be a pain and brittle. FEP-2c59 makes all of this much simpler and more reliable by including the preferred handle in the ActivityPub actor data: "webfinger" => "user@example.com" Implement this proposed FEP-2c59 attribute on the actor side and respect it when fetching remote actors so split-domain setups are resolved against the embedded webfinger handle rather than preferredUsername alone. fix_commit: "e81a4e258c62e14b5bbff860aa1ef9f18c465b8c" test_path: "spec/services/activitypub/fetch_remote_actor_service_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/38239" issue_url: "https://github.com/mastodon/mastodon/issues/38203" loc_changed: 133 why_chosen: "Federation feature with a precise FEP spec to follow and a fetch-remote-actor service spec — feature side of the 70/30 mix." - id: "38376" title: "Fix Webfinger endpoint not handling new AP ID scheme" body: | Steps to reproduce: 1. curl -i https://example.com/.well-known/webfinger?resource=https://example.com/ap/users/123 yields 404 2. https://example.com/ap/users/123 does (still) exist 3. The same query with the legacy /users/ scheme works fine. Expected: Webfinger queries with resource= set to a /ap/users/ URL should resolve to the correct local account. Actual: 404 — the resource resolver doesn't recognize the new AP ID route as a valid account locator. fix_commit: "806e2a993a8de166b96aa046c214f10eb140362f" test_path: "spec/lib/webfinger_resource_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/38391" issue_url: "https://github.com/mastodon/mastodon/issues/38376" loc_changed: 195 why_chosen: "Federation routing bug with a clear curl repro — the harness can verify against the heavily-touched webfinger_resource_spec which the PR rewrites." - id: "38045" title: "Redirect to short account URLs when requesting HTML for one of the AP endpoints" body: | 1. Visit a user profile from a different instance (e.g. Lemmy), like https://lemmy.world/u/iurii@mastodon.world 2. Try to open the user profile on their home instance, which navigates to the AP ID of the user, like https://mastodon.world/ap/users/115973097967320630 3. Get 404 response Manually creating the profile URL works fine: https://mastodon.world/@iurii Expected: redirect to the user profile Actual: 404 The new /ap/users/ form should serve HTML by redirecting to the account's short profile URL (/@) for browsers, while continuing to serve AP JSON for ActivityPub clients. fix_commit: "1add29cf40f8a9ede5d2720adea278ce5d53806d" test_path: "spec/requests/accounts_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/38056" issue_url: "https://github.com/mastodon/mastodon/issues/38045" loc_changed: 20 why_chosen: "Small request-spec-level fix with concrete URL repro." - id: "19985" title: "Split `invite_users` permission into `invite_bypass_approval`" body: | Provide an option for administrators so that when a user signs up through a user created invite link, that new signup still has to go through the usual moderator approval process. A single user, intentionally or accidentally sharing their invite link publicly would defeat the purpose. The fix is to split the existing invite_users permission so an admin role can grant "create invite links" without also granting "bypass moderator approval", and require approval for invitee signups unless the inviter has the new invite_bypass_approval permission. fix_commit: "1ee457f2d37c2b77a7fbce246a7c72ac9f9d3056" test_path: "spec/controllers/auth/registrations_controller_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/38278" issue_url: "https://github.com/mastodon/mastodon/issues/19985" loc_changed: 101 why_chosen: "Long-requested admin feature with a small, scoped diff and a registrations-controller spec — feature side of the 70/30 mix." - id: "37754" title: "Fix processing of object updates with duplicate hashtags" body: | Steps to reproduce: 1. Have a status cached from PeerTube 2. Receive an Update for that status Expected: To process the Update Actual: Job fails with ActiveRecord::RecordNotUnique error: PG::UniqueViolation: ERROR: duplicate key value violates unique constraint "statuses_tags_pkey" DETAIL: Key (tag_id, status_id)= (586704, 115424319308598593) already exists They all appear to be Updates from a peertube instance, which sometimes emits the same hashtag multiple times in the same object. ProcessStatusUpdateService should deduplicate tags before inserting them into statuses_tags. fix_commit: "5b24f4097dc56a146f76d541608789d78712f8ee" test_path: "spec/services/activitypub/process_status_update_service_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/37756" issue_url: "https://github.com/mastodon/mastodon/issues/37754" loc_changed: 6 why_chosen: "Tiny 6-LOC fix for an ActiveRecord uniqueness violation, with the exact failing service spec already in the repo." - id: "37652" title: "Fix hashtag matching by replacing negative lookbehind with positive lookbehind" body: | Steps to reproduce: Post with this URL: https://en.wikipedia.org/wiki/Google_LLC_v._Oracle_America,_Inc.#Decision Expected: It should not be treated as a hashtag. Actual: Mastodon thinks it contains a #decision hashtag. WebUI says there is a #decision hashtag, and the #decision hashtag timeline would show that post. The hashtag regex's negative lookbehind doesn't fire for non-ASCII characters preceding #, so URL fragments get mis-classified. Replace the negative lookbehind with an explicit positive lookbehind for whitespace / start-of-string. fix_commit: "4a6d17ad7bfea59d451b64b90eb794055c965557" test_path: "spec/models/tag_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/37684" issue_url: "https://github.com/mastodon/mastodon/issues/37652" loc_changed: 8 why_chosen: "Tiny regex fix with a concrete URL repro and a model spec — extremely sharp success criterion." - id: "37373" title: "Fix serialization of context pages" body: | Load a context like https://browser.pub/https%3A%2F%2Fmastodon.social%2Fcontexts%2F14715-115822340752222312 Expected: either an unpaged collection with items as normal, or a paged collection where the first page is a typical collection page. Actual: the "first" page is embedded with a "type" of "unordered" (instead of as:Collection) and there is a "part_of" property (instead of as:partOf) and has an unstripped "next" property with a null value. The ActivityPub context serializer needs to emit proper AS2 vocab terms (Collection, partOf, conditional next) for the embedded first page rather than the snake_case Rails attribute names that currently leak through. fix_commit: "b5bc301cbd13533202a506cff14e49dfbb684376" test_path: "spec/requests/activitypub/contexts_spec.rb" pr_url: "https://github.com/mastodon/mastodon/pull/37376" issue_url: "https://github.com/mastodon/mastodon/issues/37373" loc_changed: 14 why_chosen: "AS2 vocabulary serializer bug with a precise expected/actual pair and a request spec." # Model + harness config (overridable via CLI flags on run.ts) defaults: model: claude-sonnet-4-6 max_turns: 30 per_run_cost_cap_usd: 20 atlas_max_cost_usd: 50 conditions: [baseline, with-docwiki]