Bound GET /posts/notifications/:addr to the viewer's activity in a
configurable block window (NOTIFICATION_BLOCK_WINDOW, default 25000):
read the viewer's posts from addrPostHeights, prefix-scan postLikes and
postChildren for those posts only, and read follows from the new
followeeHeights index. The global likes, postChildren, and follows stores
are no longer full-scanned, and pagination.total counts only in-window
notifications.
Add the indexer write of a followeeHeights entry on every follow and
unfollow, the DB followee-index backfill utility and CLI, the
followeeheight /level route, the notification window config, unit,
property, and acceptance coverage, and developer docs.
By coder.
Bound GET /posts/notifications/:addr to the viewer's activity in a
configurable block window instead of full-scanning likes, postChildren,
and follows. Adds the followeeHeights index spec, its backfill spec, and
the notifications-query-performance spec.
By specifier.
- No production or test changes were required: buildTopicsTable passes
language mutation 2/2, dry4javascript reports no candidates, CRAP is 1.0 at
100% coverage, and soft Gherkin mutation kills all 16 mutations.
- Commit the tool-written mutate4javascript manifest and the gherkin-mutator
stamp/manifest only.
By architect.
- Add client property tests for buildTopicsTable: one row per topic in input
order, four-cell alignment with the header order, link percent-encoding
round trips, and post/follower count conservation over random topic lists
that also omit count/last-seen fields.
- No production change: the coder's view model already scans at CRAP 1.0
with zero dry4javascript candidates and only 2 mutation sites. The JSX
Topics component stays a thin, tool-excluded shell over the testable model.
Verification: client 447 unit / 94 property / 34 acceptance + lint + build;
indexer 113 unit / 12 property / 7 acceptance + lint; db 392 unit / 57
property / 16 acceptance + lint.
By refactorer.
Replace the topics page flex list with a react-bootstrap Table driven by a
testable topics-table view model: a labeled header row, aligned cells for
name, relative time, post count, and follower count, feed links on each row,
and a horizontal-scroll wrapper for narrow screens.
By coder.
Require the topics page to render a react-bootstrap Table with aligned
columns, a header row, linked rows, and horizontal scroll on narrow screens.
By specifier.
- Add the canonical per-component verification records (indexer, db, client)
pinned to the review commit 3f05488, and the architect review summary.
By architect.
- Add DB unit tests covering a summary missing postCount, a post missing
blockHeight in listTopics/getTopicPostTxids ordering, listTopics' default
limit/offset, and topicRecencyKey's missing-height fallback. These kill the
remaining language-mutation survivors in the changed DB sources.
- Refresh the mutate4javascript manifests for the changed sources and commit
the gherkin-mutator manifests/stamps written into the three topic-metadata
features.
- No production behavior changes; the review adds tests and tool-written
metadata only.
By architect.
- Replace the ternary follower delta in recordTopicFollow with
Number(isActive) - Number(wasActive), dropping its cyclomatic complexity
from 8 to 6 (CRAP 8.0 -> 6.0). No behavior change.
- Extend the indexer topic property test to conserve lastSeen (newest post
time) and followerCount (each address's final active follow state)
alongside postCount/lastHeight, with varying post seen times.
- Extend the DB backfill and topic-query property tests to conserve lastSeen
and followerCount; generated follows now include unfollows.
- Add a client topic-metadata property test covering relativeTime bucket
classification, pluralization, monotonicity, and getLastSeenLabel
consistency, and cover ensureTopicRoom's existing-summary path plus
recordTopicFollow's legacy-follower-count fallback.
- Share one in-memory DB double per component (indexer test/support/
memory-db.js, DB test/support/level-double.js) and table-drive the
parallel topic-follow/topic-message/topic-query cases, clearing every
dry4javascript candidate in the changed files.
Verification: indexer 113 unit / 12 property / 7 acceptance; db 387 unit /
57 property / 16 acceptance; client 441 unit / 90 property / 33 acceptance
+ build; lint clean. CRAP max 6.0; mutation scans under 100 sites/file.
By refactorer.
Add lastSeen and followerCount to topicSummaries across the indexer, DB,
and client. The indexer now tracks the newest post seen time, updates
follower counts idempotently on follow/unfollow, and preserves post
metadata. The DB exposes lastSeen and followerCount from GET /topics and
rebuilds both from the rooms store in the backfill. The client renders the
four-column topics page with relative-time labels.
By coder.
Define four-column topics page behavior across client, indexer, and DB:
topic name, time since most recent post, post count, and follower count.
By specifier.
A historical rooms record with an empty room name made the backfill call
put('') and abort with "key cannot be an empty String". Skip records with no
room name so the one-time backfill can complete; such rooms have no listing
entry to write.
By specifier.
Capture three time-costing lessons from this session in the architect
process notes so future sessions read them at startup: restore a
manifest stripped by a killed mutate4javascript run with
--update-manifest, gauge gherkin-mutator progress from the mutations
directory, and commit review changes before running verify.sh so the
record's git_sha matches the review commit.
By architect.
Architectural review of the topic recency indexes and pagination:
- Indexer topic index maintenance lives in its own topic-indexing.js,
separate from the action handlers; the backfill library takes injected
LevelDB handles so its logic stays testable.
- DB listTopics reads counts from topicSummaries and paginates the
topicRecency index without iterating the rooms store; the client reuses
PaginatedPage.
- Kill four topic-indexing mutation survivors with focused unit tests
covering the height fallbacks, plus exact-key/isNotFound and max-size
boundary tests.
- One topic-message survivor is an intrinsic equivalent: handlePost reads
the text at push index 1 and ignores push index 0.
Includes tool-written mutate4javascript and acceptance-mutation manifests.
By architect.
The DB soft Gherkin mutation re-ran the whole feature for every mutant;
with a fresh LevelDB world per example, a 48-mutation feature exceeded
900s. The runner-worker now diffs the mutated IR against the base feature
at <work>/base/feature.json (derived from the mutation path, since
job.work_dir may be mutation-specific) and runs only the changed
scenario/example. Safe because each scenario uses an isolated world.
The same feature now completes in ~74s.
By architect.
Run the 15 generated DB acceptance files with bounded concurrency
(default min(4, files), ACCEPTANCE_CONCURRENCY to override) instead of
strictly sequentially. Generation still runs first and sequentially.
Each generated file uses its own tmp/acceptance/level-* world, so the
pool is isolated. Full DB acceptance drops from ~430-790s to ~96s.
Also record the orphaned-mocha mutation-hang finding and remediation in
the architect process notes.
By architect.
- Extract roomFromEntry/applyPost from backfill collectSummaries so its
cyclomatic complexity drops from 7 to 4 (CRAP 7.0 -> 4.0), and extract
the shared roomRange key bounds from getTopicPostTxids/listRoomFollowers.
- Add indexer property tests over arbitrary post/follow/unfollow
interleavings: postCount conservation, newest-height recency shape,
replay idempotence, and inverted-key ordering.
- Add DB property tests for the topic index backfill: conservation,
recency shape, stale-record removal, and inverted-key ordering.
- Cover the topic action error paths (invalid push counts, oversized
topic message), the topic-query room fallbacks, and the client
openTopic/topicFeedPath navigation.
By refactorer.
- Indexer maintains topicSummaries and topicRecency for topic messages and
follows, idempotently, moving each room's recency record to its newest post.
- DB adds the two index stores and serves GET /topics from them with
limit/offset pagination, without iterating the rooms store. Adds
util/room/backfill-topic-indexes.js to build the indexes idempotently.
- Client topics page loads 50 per page with Previous/Next.
By coder.
Add Gherkin for efficient GET /topics ordering by most recent post, the
topicSummaries and topicRecency index stores, an idempotent backfill, and
client topics-page pagination.
By specifier.
Import the browser-safe buffer module in the multi-push adapter so the
CRA production bundle no longer reads the Node global Buffer, add
regression tests for the falsy-wallet and idempotent-attach paths, share
the script-encoding test helper, and de-duplicate the UTF-8 push
acceptance assertions. Refresh the language mutation manifests and the
soft acceptance-mutation manifest.
By architect.
Share a single UTF-8 encoding helper across hex, poll-create, and
topic-post instead of repeating new TextEncoder().encode(), drop the
redundant Uint8Array branch in the push normalizer, rename the
txid-action push builder parameter from buildPayload to buildPushes to
match what it returns, and factor the db repair unit-test setup into a
repairedFixture helper.
Add property tests for the multi-push adapter covering prefix/ordering,
field-byte conservation, buffer round trips, array expansion, single
field delegation, and idempotent attach, plus a unit test for the
unsupported-wallet guard.
By refactorer.
Reply, topic message, add-poll-option, poll-vote, and create-poll now
broadcast each protocol field as its own OP_RETURN script push instead of
combining them into one push. The indexer requires three pushes for reply,
topic message, and poll children, and accepts four for create-poll; the
combined form was rejected or misparsed.
hex.js builds the txid and text as separate pushes, and a new
memo-multipush adapter teaches minimal-slp-wallet's single-push sendOpReturn
to expand an array of fields into separate pushes. Acceptance step handlers
assert the individual pushes, and unit/property tests cover the new shape.
By coder.
Add Gherkin pinning that reply, topic message, add-poll-option, poll
vote, and create-poll broadcast each protocol field as its own OP_RETURN
push, so the indexer and memo.cash can parse them.
By specifier.
Refresh the mutate4javascript and Gherkin acceptance-mutation manifests,
kill the memo-like/memo-reply boundary survivors, and extract the shared
FakeDb/makeLevel double used by the txid repair unit and property tests.
By architect.
Share the txid-and-text payload builder between hex and reply, extract the
repair store loop from repairTxidEncoding to cut CRAP from 13 to 5, correct
the poll-services property test to the little-endian wire order, and add
property tests for wire round trips and repair idempotence/conservation.
By refactorer.
Client likes, replies, poll options, and poll votes now embed the
referenced txid in little-endian wire order, matching the indexer's
byte reversal. Add a psf-memo-db repair library and CLI that rewrites
byte-reversed references in likes, postParents, pollOptions, and
pollVotes and rebuilds the postLikes and postChildren indexes, leaving
correct and unknown references untouched and staying idempotent.
By coder.
Add client Gherkin pinning little-endian wire order for every action that
embeds a parent txid (like, reply, poll option, poll vote), and a
psf-memo-db Gherkin for a utility that repairs existing byte-reversed
references and rebuilds the affected indexes.
By specifier.
Add unit coverage for the LikeTipPage initial state, deps seeding, the
dust-limit boundary, and empty/null tip parsing, and assert exact
validation/broadcast error messages. Add the NewPostPage initial
result-modal state assertion. Refresh the mutate4javascript footer
manifests for the touched services and the soft Gherkin
acceptance-mutation manifest for the like-result feature.
By architect.
Deduplicate the new like-result acceptance steps behind one render helper,
and extract the bch.loping.net transaction link shared by the New Post result
modal, the post options menu, and the like/tip result into a block-explorer
service. Cover the LikeTipPage open/parse/handler failure branches and add
property tests for the explorer URL, the LikeResult markup, and the result
modal state machine.
CRAP is at or below 5 with full coverage on the changed services; every
changed/new plain source file scans at 21 or fewer mutation sites. The
existing mutation manifests are left for the mutation tool to refresh.
By refactorer.
Keep the like/tip modal open after a successful like and show a broadcast
result instead of closing: the success message, the like transaction id, and
a link to that transaction on bch.loping.net that opens in a new tab. The
like count and filled heart update immediately, and dismissing the result
closes the modal. The result state lives on the testable LikeTipPage
controller; a server-renderable LikeResult component backs the modal and the
acceptance adapter, with acceptance step handlers for the new wording.
By coder.
Move the outside-click decision into the pure post-options service so the
effect stays a thin adapter and the previously uncovered `&&` mutation is
unit testable. Add unit tests for the outside-click predicate and for the
default no-focus open menu, killing the two surviving `-1` default-focus
mutations.
Stamp the mutate4javascript manifests for both modules and the soft
acceptance-mutation manifest for post-options-menu.
By architect.
Move the keyboard shortcut mapping into the pure post-options service so the
React component is a thin adapter and the Escape/ArrowDown decisions are unit
and property testable. DRY the duplicated post-options acceptance steps behind
two helpers, and add property tests for the explorer URL, menu-item shape,
state transitions, key command, and rendered markup.
The createElement post-options component and service stay scannable by
mutate4javascript; the JSX call sites (post-feed-item, profile) are unchanged
except for rendering the shared menu.
By refactorer.
Add a shared PostOptionsMenu component, used by the feed/thread post card and
the profile post card, with a pure post-options service for the block explorer
link and the open/close/focus state. Its first item links to the post
transaction on bch.loping.net in a new tab; the menu opens on the button and
closes on the button again, an outside click, or Escape, with ArrowDown moving
focus to the first item.
By coder.
Written review report and machine-readable verification record for the
raised recent-feed total scan cap. Verification passed 4/4 (unit 358,
property 48, acceptance all 12 suites, lint). Mutation killed 37/37 in
post-query.js; CRAP max 6.0; DRY clean for the changed files.
By architect.
Architectural review of the raised recent-feed total scan cap (10 -> 500)
and the refactorer's property tests. No structural change was warranted:
the cap stays a private constant in PostQuery, the new property tests use
in-memory store doubles with no IO, and dry4javascript found no duplicate
candidate in the changed source or the new property test. Mutation killed
all 37 sites in post-query.js.
This commit carries only the tool-written manifests: the mutate4javascript
manifest in post-query.js and the gherkin-mutator acceptance stamps in the
two touched feature files.
By architect.
Cover the total-scan cap and bounded raw scan of
PostQuery.scanRecentPostTxidsAndCount over seeded random corpora: capped
total, newest-first pagination conservation, the raw-scan bound, the 500
default cap, and reply exclusion.
By refactorer.
GET /posts/recent now caps its total scan at 500 eligible top-level posts
instead of 10, so corpora smaller than the cap report an exact total while
larger corpora stay bounded. Added unit coverage for the exact-total-under-cap
and default-cap cases, plus the many-top-level-posts acceptance fixture.
By coder.
Commit the review summary and the client verification record (git_sha
19e57be, the review commit) alongside the review. All five client commands
pass.
By architect.
Merged the refactorer handoff (e77e174). The architecture holds: the image
helpers isImageUrl/imageAltText are pure post-links functions, failed-images
is a pure state transition, and PostImage/PostContent stay presentational
with no IO. Added a unit test for isImageUrl non-string input to kill the
last surviving language mutant; mutation and acceptance manifests are
tool-written.
By architect.
Extract the pure failed-image set transition into a testable service,
DRY the acceptance no-link/no-image assertions, and add property tests
for image URL detection, alt text, and image rendering.
By refactorer.
Recognize common image file extensions in post URLs (query string and
fragment ignored), render them as inline images inside new-tab anchors
with the filename as alt text, and fall back to a plain link when the
image fails to load. Non-image URLs keep the existing plain-link
behavior.
By coder.
Add Gherkin for inline image rendering in posts: extension-based image URL
detection (.jpg/.jpeg/.png/.gif/.webp/.bmp, query and fragment ignored),
anchor-wrapped inline <img> with filename alt text, non-image URLs stay
plain links, and an image load failure falls back to a plain link.
By specifier.
Mark the feature done in the backlog, add gotchas about stale architect
verification records and surviving link-example prose mutations, and
refresh the standing briefing HEAD/next-action lines.
By specifier.
The architect's record named e17c515, which predated the review commit
b63792f. Re-ran client verification on the merged b63792f; all 5 commands
pass.
By specifier.
Harden the post link parser after the refactorer handoff:
- add start-of-text and leading-token boundary unit tests and drop a
redundant range guard, killing the isBareDomainBoundary survivors
- replace the mutable React key counter with React.Children.toArray and
assert the embedded iframe is fullscreen, killing both PostContent
mutation survivors
- record the client verification pass and the review summary
By architect.
Render http(s) URLs and bare domains in post text as anchors that open in a
new tab. Explicit URLs keep their scheme; bare domains get an https href while
their visible text stays as written. Embeddable YouTube links keep their
embedded-player behavior.
By coder.
Add Gherkin acceptance spec for auto-linking http(s) URLs and bare domains
in post text, opening in a new tab, while preserving YouTube embeds.
By specifier.
The process changes were only in Git history and a briefing paragraph. Add a
consolidated what/why record with commit hashes and verification evidence, and
point to it from the feature backlog so the next specifier discovers it.
By specifier.
Unrelated tracked changes could ride along with a handoff. Refuse to send when
git status --porcelain is non-empty, list the offending files, and keep the
draft so it can be fixed and resent. SWARMFORGE_ALLOW_DIRTY=1 overrides.
By specifier.
Add state.sh to print each role branch HEAD and the main-checkout status.
Update the briefing to use the single tmp/aps checkout, note daemon
self-healing, and point merge verification at verify.sh and the architect's
verification record.
By specifier.
The default differential run can select fewer covered sites than exist after a
function is added or removed, forcing a manual --mutate-all rerun. Add
mutate-file.sh (auto-reruns when Selected < Covered) and clean-builds.sh
(removes worker-copy bloat), and document them plus the intrinsic upper-bound
Gherkin survivors in the architect cheat-sheet.
By specifier.
Verification commands were transcribed in several places and the same full
suite was re-run by every role. Add verify.sh/verify.mjs as the single
per-component verification sequence emitting a machine-readable record
(git_sha, per-command exit/duration/summary). The architect records it for
each touched component; the specifier trusts a matching-SHA pass and only
re-runs the merged feature's acceptance test.
By specifier.
Every acceptance run re-parsed and regenerated all features even when only
one changed. Record feature_hash and aps_commit in the per-feature metadata
and skip parse/generate when both match, and drop generated tests for
features that no longer exist.
By specifier.
There were up to four APS copies (tmp/aps, tmp/aps-spec, tmp/aps-fresh, and a
component-local client copy) and the briefing, wrappers, and runners disagreed
on which to use. Add ensure-aps.sh (ensure/update/reclone) as the one canonical
tmp/aps checkout, point the gherkin-parser wrapper, architect startup, and all
three acceptance runners at it, and update the architect cheat-sheet.
By specifier.
Handoffs sit undelivered when handoffd is not running, and recovery was a
manual step. Add ensure_handoff_daemon.sh and call it from swarm_handoff.sh
and ready_for_next.sh so the daemon is restarted whenever it is missing.
By specifier.
Each psf-memo-db acceptance scenario opened an isolated LevelDB directory
but close() only closed the handles, so tmp/acceptance grew without bound
(found at 961 dirs / 293 MB). Delete the scenario dir in close() and sweep
tmp/acceptance at the start of the acceptance run.
By specifier.
The architect's terminal handoff went only to coder and refactorer, so the
specifier never received a merge notification and a human had to bridge the
gap. Require an end-of-chain git_handoff to the specifier for every completed
task or batch, including manifest- and metadata-only churn.
By specifier.
Merge hardening verified on master: psf-memo-db unit (357), property (44),
acceptance (11 suites, including thread-query-performance), and lint all
pass. Record the completion and add a gotcha about upper-bound performance
assertions surviving Gherkin mutation.
By specifier.
Encapsulate the postChildren prefix scan in PostQuery.listChildTxids and have
GetPostThread depend only on the PostQuery interface (listChildTxids,
getPostOrNull). countRepliesForTxids now reuses listChildTxids. Remove the dead
whole-database buildLikeCountMap that this task made obsolete. Cover the
prefix-scan guards and cycle back-edge, realign the thread property tests to the
adapter seam, and refresh mutation/Gherkin-acceptance manifests.
By architect.
Scope the thread endpoint to per-thread reads:
- Count likes for only the thread's txids via countLikesForTxids instead of
the whole-database buildLikeCountMap scan.
- Prefix-scan postChildren by parent txid in loadChildTxids instead of walking
the entire store for every node.
- Add unit tests pinning bounded postChildren reads and thread-scoped likes.
By coder.
Bound /posts/:txid/thread to per-thread reads: prefix-scan postChildren
by parent txid and scope like counts to the thread's txids instead of
the global postLikes scan. Record both implementation parts in the
feature backlog.
By specifier.
Chris Troutner