feat(news): completeness measurement — feed health w/ silent zeros, coverage ledger + provenance, GDELT recall benchmark (#4920) by koala73 · Pull Request #4927 · koala73/worldmonitor

koala73 · 2026-07-05T20:01:14Z

Closes #4920 (Bet 2 of the strategic news-pipeline review). Do not auto-merge — held for review. ⚠️ Stacked on #4924 (the recall benchmark matches titles via scripts/shared/story-identity.js); merge #4924 first — this PR's diff will shrink to its own commits after that.

Three layers of "did we miss something?"

(a) Feed health + silent zeros — validate-rss-feeds.mjs now validates BOTH feed universes (client config and the server digest catalog _feeds.ts — the list the digest actually ingests, which was previously unmeasured; 276 unique server URLs extracted with faithful gn() URL reconstruction). With Upstash secrets in the daily GH Actions run it publishes news:feed-health:v1 with per-feed status + consecutive-empty streaks; Google News wrappers empty across ≥2 consecutive daily runs are flagged as silent zeros — the "Google ranking shift quietly blanked a source" failure mode.

(b) Coverage ledger + provenance — every silent drop gate now counts: per-feed cap (previously fully invisible), per-category cap, freshness floor → news:coverage-ledger:v1:{variant}:{lang} per digest build; insights selection gates (admissibility/source-cap/overflow) via a backward-compatible selectTopStories stats param. The World Brief payload carries provenance and the panel renders "Compiled from N stories across M sources" (all locales fanned out via translate-locales; {{token}} convention).

(c) External recall benchmark — daily, same workflow, zero Railway slots: GDELT top articles across 4 broad verticals matched against ingested digest titles using the #4919 shared similarity → news:recall-benchmark:v1 with recall % and the missed headlines sorted by lowest similarity (actionable coverage holes, not just a number).

Both keys registered in api/seed-health.js + api/health.js (age-above-coverage precedence respected).

Operator action

Add UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN as GitHub Actions secrets to activate publishing — without them both publishers log a skip and the workflow behaves exactly as before.

Verification

tests/completeness-measurement.test.mjs (14): silent-zero streak/reset/wrapper-only semantics, recall math incl. edit-variant matching + null-not-NaN, selection drop stats, server-catalog extraction (250+ feeds), wiring assertions.
Full battery: clustering (16), server-handlers (24), story-identity (30), typecheck + typecheck:api + biome + docs-stats clean.

https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

mintlify · 2026-07-05T20:01:17Z

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
WorldMonitor	🟢 Ready	View Preview	Jul 5, 2026, 8:02 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

vercel · 2026-07-05T20:01:20Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
worldmonitor	Ready	Preview, Comment	Jul 6, 2026 10:57am

greptile-apps · 2026-07-05T20:09:35Z

Greptile Summary

This PR adds three completeness-measurement layers to the news pipeline: (a) feed-health with silent-zero detection for Google News wrappers across consecutive empty runs, (b) a coverage ledger tracking per-feed/per-category/freshness-floor drop counts, and (c) a daily GDELT recall benchmark computing what fraction of top external stories the digest carries. It also consolidates story identity across the pipeline by replacing three divergent similarity implementations with a single dual-view lexical vector module.

Coverage-ledger implementation is missing from list-feed-digest.ts: the test "digest counts every drop gate and publishes the ledger" asserts four source-text patterns (droppedFeedCap, ledgerDrops.perCategoryCap, ledgerDrops.freshnessFloor, news:coverage-ledger:v1) that do not exist in the file \u2014 these four assertions will fail when the test suite runs.
Feed-health + GDELT recall benchmark are well-implemented pure modules with matching unit tests, and the GH Actions workflow correctly threads Upstash secrets to both publishers.
selectTopStories drop-stat accounting: overflowDropped is calculated as admissible.length - selected.length - sourceCapDropped, which misclassifies source-capped items that occur after the maxCount break as overflow; totals balance but per-bucket attribution can mislead.

Confidence Score: 3/5

The feed-health, GDELT recall, and story-identity unification work correctly and are well-tested, but the coverage-ledger feature described in the PR is entirely absent from list-feed-digest.ts, causing the test suite to fail on four source-text assertions.

The coverage-ledger instrumentation (droppedFeedCap, ledgerDrops, news:coverage-ledger:v1 publish) is called out in the PR description as delivered but is not present in list-feed-digest.ts or anywhere else in the diff. The test suite ships with four assertions that check for exact source-text patterns from that missing implementation; those assertions will all fail.

server/worldmonitor/news/v1/list-feed-digest.ts needs the coverage-ledger instrumentation before the test suite can pass. tests/completeness-measurement.test.mjs lines 153-158 will fail until that code lands.

Important Files Changed

Filename	Overview
tests/completeness-measurement.test.mjs	New 14-test suite for completeness measurement; the coverage-ledger describe block asserts source-text patterns in list-feed-digest.ts that do not exist — those 4 assertions will fail.
server/worldmonitor/news/v1/list-feed-digest.ts	Switches from exact sha256 title hashing to fuzzy assignStoryIdentity for corroboration and story tracking; per-hash dedup guards prevent mentionCount inflation. Coverage-ledger instrumentation described in the PR is absent.
scripts/_clustering.mjs	Removes local Jaccard clustering and delegates to shared story-identity; adds optional stats param to selectTopStories for drop counts with a minor overflowDropped attribution issue.
server/worldmonitor/news/v1/dedup.mjs	Replaces word-overlap dedup with cosine similarity via shared story-identity; new assignStoryIdentity function clusters items and assigns publishedAt-anchored canonical hashes.
shared/story-identity.js	New central story-identity module: dual-view FNV-1a hash lexical vectors, cosine similarity with containment rescue, union-find clustering; replaces three divergent similarity implementations.
scripts/_feed-health.mjs	Pure feed-health payload builder: classifies Google News wrappers, maintains consecutive-empty streaks, surfaces silent-zero detection after SILENT_ZERO_THRESHOLD=2 runs.
scripts/_recall-benchmark-core.mjs	Pure recall computation: matches external GDELT titles against digest titles using cosine similarity; unvectorizable titles excluded from denominator; null-not-NaN for empty sets.
scripts/seed-recall-benchmark.mjs	New daily GH Actions script: fetches GDELT top articles across 4 verticals, computes recall against the digest, publishes to Redis. Exits 0 on all upstream failures.
scripts/validate-rss-feeds.mjs	Extended to validate the server digest catalog in addition to the client config; adds publishFeedHealth post-run; module now importable by tests via the direct-run guard.
scripts/seed-insights.mjs	Adds provenance object (storiesConsidered, sourcesConsidered, selectionDrops) to the insights payload via selectTopStories stats param; backward-compatible.
src/components/InsightsPanel.ts	Adds renderProvenance() rendering the Compiled from N stories across M sources stamp; gracefully absent on pre-rollout payloads.
.github/workflows/feed-validation.yml	Adds seed-recall-benchmark.mjs step and Upstash secrets env; timeout bumped from 15 to 25 min to accommodate the new GDELT fetch step.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[feed-validation GH Actions daily] --> B[validate-rss-feeds.mjs client + server catalog]
    A --> C[seed-recall-benchmark.mjs]
    B --> D{Upstash creds?}
    D -- yes --> E[publishFeedHealth buildFeedHealthPayload]
    E --> F[news:feed-health:v1 silent-zero streaks]
    D -- no --> G[skip silently]
    C --> H[fetchGdeltJson 4 verticals x 25 articles]
    H --> I[computeRecall cosineSimilarity vs digest]
    I --> J[news:recall-benchmark:v1 recall% + missed headlines]
    K[seed-insights.mjs selectTopStories stats] --> L[provenance storiesConsidered sourcesConsidered]
    L --> M[InsightsPanel Compiled from N stories across M sources]
    N[list-feed-digest.ts assignStoryIdentity] --> O[clusterTexts union-find fuzzy hash]
    O --> P[canonical titleHash per cluster]
    P --> Q[writeStoryTracking mentionCount once per hash]
    R[coverage-ledger news:coverage-ledger:v1] -.->|MISSING| N

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[feed-validation GH Actions daily] --> B[validate-rss-feeds.mjs client + server catalog]
    A --> C[seed-recall-benchmark.mjs]
    B --> D{Upstash creds?}
    D -- yes --> E[publishFeedHealth buildFeedHealthPayload]
    E --> F[news:feed-health:v1 silent-zero streaks]
    D -- no --> G[skip silently]
    C --> H[fetchGdeltJson 4 verticals x 25 articles]
    H --> I[computeRecall cosineSimilarity vs digest]
    I --> J[news:recall-benchmark:v1 recall% + missed headlines]
    K[seed-insights.mjs selectTopStories stats] --> L[provenance storiesConsidered sourcesConsidered]
    L --> M[InsightsPanel Compiled from N stories across M sources]
    N[list-feed-digest.ts assignStoryIdentity] --> O[clusterTexts union-find fuzzy hash]
    O --> P[canonical titleHash per cluster]
    P --> Q[writeStoryTracking mentionCount once per hash]
    R[coverage-ledger news:coverage-ledger:v1] -.->|MISSING| N

_{Reviews (1): Last reviewed commit: "feat(news): completeness measurement — f..." | Re-trigger Greptile}

greptile-apps · 2026-07-05T20:09:38Z

+    const src = readSrc('server/worldmonitor/news/v1/list-feed-digest.ts');
+    assert.match(src, /droppedFeedCap = Math\.max\(0, matches\.length - ITEMS_PER_FEED\)/);
+    assert.match(src, /ledgerDrops\.perCategoryCap \+= Math\.max\(0, items\.length - MAX_ITEMS_PER_CATEGORY\)/);
+    assert.match(src, /ledgerDrops\.freshnessFloor = droppedStaleTotal/);
+    assert.match(src, /news:coverage-ledger:v1/);
+  });


Coverage-ledger test asserts unimplemented code

The four assert.match calls inside this test check for patterns (droppedFeedCap, ledgerDrops.perCategoryCap, ledgerDrops.freshnessFloor, news:coverage-ledger:v1) that are not present anywhere in server/worldmonitor/news/v1/list-feed-digest.ts. Running node --test tests/completeness-measurement.test.mjs will fail immediately on all four assertions.

The PR description claims "(b) Coverage ledger + provenance — every silent drop gate now counts: per-feed cap, per-category cap, freshness floor → news:coverage-ledger:v1:{variant}:{lang} per digest build" but the list-feed-digest.ts diff in this PR contains none of the instrumentation or the Redis publish for that key. The ledger wiring in list-feed-digest.ts appears to be the missing implementation piece.

greptile-apps · 2026-07-05T20:09:39Z

+    stats.sourceCapDropped = sourceCapDropped;
+    stats.overflowDropped = Math.max(0, admissible.length - selected.length - sourceCapDropped);
  }


overflowDropped misclassifies source-capped items after the break

After selected.length >= maxCount triggers a break, the remaining items in admissible are never visited. The formula admissible.length - selected.length - sourceCapDropped counts ALL unvisited items as overflow — but any of those trailing items that would have hit the source cap are instead counted as overflow rather than source-capped. The grand total is always correct, but sourceCapDropped will be systematically under-counted and overflowDropped over-counted when the selection limit binds before the source-cap list is exhausted.

…uble-quoted names, parse-cache bump, bounded external fields (#4927 review round) Applies the multi-agent + cross-model review findings: - th/tr/zh compiledFrom carried single-brace placeholders (fan-out ran before the {{token}} convention fix) — repaired; fa.json was missing the keys entirely AND absent from translate-locales' LOCALES list (pre-existing gap the parity test caught) — keys added, fa added to the translator roster - extractServerFeeds skipped double-quoted names ("Tom's Hardware") — regex accepts both quote styles; extraction test pins it (cross-model) - rss:feed parse cache bumped v5→v6: warm v5 rows lack droppedFeedCap and would undercount the coverage ledger for their whole TTL (cross-model) - coverage-ledger writes clamped to known variants + well-formed langs (request-supplied values could inflate the keyspace — adversarial) - GDELT titles AND urls bounded before publishing; urls http(s)-only - direct-run guards use pathToFileURL (Windows/symlink-safe); both publishers share scripts/_upstash-rest.mjs instead of hand-rolled fetch helpers; publishFeedHealth exported with a no-creds contract - .insights-provenance CSS added (was an unstyled div); gn()/gnLocale replica drift guard test; overflowDropped + orchestration-seam tests Battery: completeness (21), locale-completeness (25), clustering (16), story-identity (30), both typechecks, biome clean. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

koala73 · 2026-07-05T20:15:36Z

Multi-agent code review — applied at `f98e584`

Four reviewer personas (correctness, adversarial+security, testing, maintainability+standards) plus the independent Codex cross-model pass. All actionable findings applied; held for manual merge (stacked on #4924).

Applied

Sev	Finding	Reviewers	Fix
P1	th/tr/zh `compiledFrom` carried single-brace placeholders (fan-out predated the convention fix) — i18next would render them literally	correctness(100) + maint(100) + codex	braces repaired in all three
P1	`fa.json` missing the new keys — locale-parity test red; root cause: fa absent from translate-locales' roster (pre-existing)	testing(100)	Persian keys added; fa added to the translator LOCALES list
P1	`extractServerFeeds` skipped double-quoted names ("Tom's Hardware")	codex(100)	regex accepts both quote styles; extraction test pins it — server catalog now 277
P2	Warm `rss:feed:v5` cache rows lack `droppedFeedCap` → ledger undercounts for the TTL	codex + correctness probe	parse-cache bumped v5→v6
P2	Request-supplied variant/lang could inflate `news:coverage-ledger` keyspace	adversarial(75)	writes clamped to known variants + `^[a-z]{2}$` langs
P2	GDELT titles/urls (untrusted) published unbounded	adversarial + codex	title 200ch, url 300ch + http(s)-only
P2	`.insights-provenance` rendered unstyled	maint(100)	CSS added beside the stats block
P2	Duplicated Upstash REST helpers; gn()/gnLocale replica had no drift guard	maint(75)	shared `scripts/_upstash-rest.mjs`; source-parity drift-guard test
P2/P3	Direct-run guards via string-built file:// (Windows/symlinks); untested orchestration seams; overflowDropped untested	codex + testing	`pathToFileURL` guards; publishFeedHealth/gdeltUrl/unwrapEnvelope exported + tested; overflow fixture test

Verified directly during review: npm run test:feeds:ci still executes main() after the import-guard change (679 merged feeds: 557 client + 277 server).

Noted, not changed

Source-textual wiring tests remain alongside (not instead of) behavioral ones — the pure logic (payload builder, recall math, selection stats) is behaviorally tested; the wiring assertions guard integration points that would need a full serverless harness.
overflow-vs-sourceCap attribution depends on gate ordering — semantics now pinned by the fixture test.

Battery: completeness (21), locale-completeness (25), clustering (16), story-identity (30), both typechecks, biome.

…uble-quoted names, parse-cache bump, bounded external fields (#4927 review round) Applies the multi-agent + cross-model review findings: - th/tr/zh compiledFrom carried single-brace placeholders (fan-out ran before the {{token}} convention fix) — repaired; fa.json was missing the keys entirely AND absent from translate-locales' LOCALES list (pre-existing gap the parity test caught) — keys added, fa added to the translator roster - extractServerFeeds skipped double-quoted names ("Tom's Hardware") — regex accepts both quote styles; extraction test pins it (cross-model) - rss:feed parse cache bumped v5→v6: warm v5 rows lack droppedFeedCap and would undercount the coverage ledger for their whole TTL (cross-model) - coverage-ledger writes clamped to known variants + well-formed langs (request-supplied values could inflate the keyspace — adversarial) - GDELT titles AND urls bounded before publishing; urls http(s)-only - direct-run guards use pathToFileURL (Windows/symlink-safe); both publishers share scripts/_upstash-rest.mjs instead of hand-rolled fetch helpers; publishFeedHealth exported with a no-creds contract - .insights-provenance CSS added (was an unstyled div); gn()/gnLocale replica drift guard test; overflowDropped + orchestration-seam tests Battery: completeness (21), locale-completeness (25), clustering (16), story-identity (30), both typechecks, biome clean. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…d fails (#4927 external review) A transient Redis error during the previous-payload GET previously reset every consecutive-empty streak (previous=null -> fresh streaks). Read failure now skips the publish entirely; key-absent (first run) still starts fresh. Regression test included. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

koala73 · 2026-07-06T04:53:04Z

External review items attributed here via the stacked diff — both landed at 31d7a04: silent-zero streaks are preserved when the previous-state read fails (publish skipped, regression-tested), and the recall benchmark's fast-fail budget landed on #4929's branch. Still held for manual merge.

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…uble-quoted names, parse-cache bump, bounded external fields (#4927 review round) Applies the multi-agent + cross-model review findings: - th/tr/zh compiledFrom carried single-brace placeholders (fan-out ran before the {{token}} convention fix) — repaired; fa.json was missing the keys entirely AND absent from translate-locales' LOCALES list (pre-existing gap the parity test caught) — keys added, fa added to the translator roster - extractServerFeeds skipped double-quoted names ("Tom's Hardware") — regex accepts both quote styles; extraction test pins it (cross-model) - rss:feed parse cache bumped v5→v6: warm v5 rows lack droppedFeedCap and would undercount the coverage ledger for their whole TTL (cross-model) - coverage-ledger writes clamped to known variants + well-formed langs (request-supplied values could inflate the keyspace — adversarial) - GDELT titles AND urls bounded before publishing; urls http(s)-only - direct-run guards use pathToFileURL (Windows/symlink-safe); both publishers share scripts/_upstash-rest.mjs instead of hand-rolled fetch helpers; publishFeedHealth exported with a no-creds contract - .insights-provenance CSS added (was an unstyled div); gn()/gnLocale replica drift guard test; overflowDropped + orchestration-seam tests Battery: completeness (21), locale-completeness (25), clustering (16), story-identity (30), both typechecks, biome clean. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…d fails (#4927 external review) A transient Redis error during the previous-payload GET previously reset every consecutive-empty streak (previous=null -> fresh streaks). Read failure now skips the publish entirely; key-absent (first run) still starts fresh. Regression test included. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

… attribution, publish-after-guardrails (#4927 re-review) All six re-review findings applied: P1 — never-activated publishing no longer poisons health: the two keys join ON_DEMAND_KEYS (missing data = soft EMPTY_ON_DEMAND) and their seed-meta entries are activationGated in api/seed-health.js ('pending-activation', healthy) until the first publish lands — then normal 2880-min staleness enforcement applies. The workflow also emits a visible ::warning annotation on runs where the UPSTASH secrets are absent, so green scheduled runs can't silently mask the gap. P2 — the coverage-ledger write is awaited (.catch non-fatal): the Edge response could previously finish before the fire-and-forget side write landed. P2 — selectTopStories classifies EVERY admissible candidate instead of breaking at maxCount: cap-first attribution (a candidate whose source already hit the per-source cap is a sourceCap drop even when selection is full; only genuinely rankable candidates count as overflow), with a same-source-overflow regression asserting exact one-of attribution. P2/P3 — StoryMeta.source_count proto doc now describes what runtime ships (current-cycle batch corroboration, default 1) instead of 'cached from Redis Set'; generated artifacts regenerated. P3 — _upstash-rest.mjs sends a User-Agent (CF-WAF convention: never a bare fetch UA); feed-health publishing moved AFTER the config-drift hard-fail so a guardrail-failing run cannot refresh health metadata first. Battery: completeness 23, clustering 16, mcp-parity 11, seed-health 2, typecheck:api, docs-stats. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

koala73 · 2026-07-06T08:29:31Z

Re-review adjudication — all 6 findings validated, fixed at `d1c6298`

Sev	Finding	Fix
P1	Missing secrets → green workflow + missing/stale monitored health keys	Activation-gating on both surfaces: the two keys join `ON_DEMAND_KEYS` in `api/health.js` (missing data reads soft `EMPTY_ON_DEMAND`, not `EMPTY`), and their seed-meta entries carry `activationGated: true` in `api/seed-health.js` (`pending-activation`, healthy, until the first publish — then the 2880-min staleness rule enforces normally). The workflow additionally emits a visible `::warning` annotation when the secrets are absent, so a green scheduled run can never silently mask the gap.
P2	Coverage-ledger write fire-and-forget on Edge	Awaited with non-fatal `.catch`, matching the story-tracking write's pattern.
P2	Overflow/source-cap attribution breaks at `maxCount`	Loop classifies every admissible candidate, cap-first: a candidate whose source already hit the per-source cap is a `sourceCap` drop even when selection is full; only genuinely rankable candidates count as overflow. Same-source-overflow regression asserts exact one-of attribution across all four buckets.
P2/P3	`StoryMeta.source_count` doc says "cached from Redis Set"	Proto comment now describes runtime truth (current-cycle batch corroboration, default 1; the Redis set feeds importance scoring, not this field); generated artifacts regenerated via `make generate`.
P3	`_upstash-rest.mjs` missing User-Agent	`worldmonitor-ops/1.0` UA added (repo CF-WAF convention).
P3	Feed-health published before the config-drift hard-fail	Publish moved after the guardrail exit so a drift-failing run cannot refresh health metadata to fresh/OK first; ordering rationale documented in-code.

Battery: completeness 23, clustering 16, mcp-parity 11, seed-health 2, typecheck:api, docs-stats. Cascade-rebased through #4928 → #4929. Still held for manual merge.

…re FIRST publish (#4927 re-review P1) Activation was inferred from seed-meta existence, which carries a 7d TTL: a publisher that ran once and died alarmed as stale only until the meta expired, then reverted to healthy pending-activation — violating the stated contract, and /api/health softened the keys permanently via ON_DEMAND_KEYS. "Has ever published" is now a durable marker (seed-activated:news:*, SET with NO TTL on every successful publish): - api/health.js pipelines EXISTS on the markers; a marker-bearing key leaves ON_DEMAND softening permanently — missing data/meta reads EMPTY/STALE_SEED like any other monitored key (marker read errors degrade to not-activated, never to a false alarm) - api/seed-health.js entries carry activationKey; 'pending-activation' is returned ONLY while the marker is absent — with the marker present, missing meta falls through to 'missing' (alarming) - both publishers SET their marker alongside every publish Regressions: classifyKey lifecycle (pending -> EMPTY_ON_DEMAND; activated-then-dead -> EMPTY) + publisher/marker wiring pins including the no-TTL assertion. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

koala73 · 2026-07-06T08:56:33Z

Re-review round 2 — the remaining P1 fixed at $(git rev-parse --short feat/completeness-4920)

Validated in full. Activation was inferred from seed-meta existence (7-day TTL), so "has ever published" evaporated with the meta: a publisher that ran once and died alarmed only until expiry, then reverted to healthy pending-activation — and /api/health softened the keys permanently via ON_DEMAND_KEYS. Exactly as you framed it: the stated contract was violated on both surfaces.

Fix (durable marker, as suggested): publishers SET seed-activated:news:{feed-health,recall-benchmark} 1 with no TTL alongside every publish.

api/health.js pipelines EXISTS on the markers; a marker-bearing key leaves ON_DEMAND softening permanently — missing data/meta then reads EMPTY/STALE_SEED like any monitored key. Marker read errors degrade to not-activated (soft), never to a false alarm.
api/seed-health.js entries carry activationKey; pending-activation is returned only while the marker is absent — with it present, missing meta falls through to missing (alarming).

Regressions: classifyKey lifecycle (never-activated → EMPTY_ON_DEMAND; activated-then-dead → EMPTY) plus wiring pins on both publishers including a no-TTL assertion on the marker SET.

Batteries: completeness 25, seed-health 2, mcp-parity 11, full tiered pre-push (both typechecks, seed tests) green. Cascade-rebased through #4928 → #4929. Held for manual merge.

…marker probe (#4927) The harnesses assert every pipeline command is GET; the activation markers add EXISTS probes on seed-activated:* keys, which tripped the mock assertion and 503'd every request in these suites. Both mocks now accept EXISTS (marker-namespace-pinned, absent by default). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…etection, coverage ledger + provenance, GDELT recall benchmark (#4920) Turns "unmistakably complete" from a feeling into numbers, at three layers: (a) Feed health. validate-rss-feeds now also validates the SERVER digest catalog (_feeds.ts — the list buildDigest actually ingests; previously only the client config was measured) and, when Upstash secrets are present, publishes per-feed status + consecutive-empty streaks to news:feed-health:v1 daily. Google News wrappers that deliver nothing across consecutive runs are flagged as SILENT ZEROS — the failure mode where Google's ranking shift quietly blanks a source. (b) Coverage ledger + provenance. Every silent drop gate now counts: per-feed cap, per-category cap, freshness floor (digest → news:coverage-ledger:v1:{variant}:{lang}), and the insights selection gates (admissibility / source cap / overflow via selectTopStories stats). The World Brief payload carries provenance and InsightsPanel renders the stamp: "Compiled from N stories across M sources" (i18n fanned out to all locales). (c) External recall benchmark. Daily job (same GH Actions workflow, no Railway slot) pulls GDELT's top articles across four broad verticals, matches each against ingested digest titles using the shared story-identity similarity (#4919), and publishes recall % + the missed headlines to news:recall-benchmark:v1 — the first ground-truth answer to "did we miss a story?", with misses actionable by lowest similarity. Both published keys are registered in api/seed-health.js and api/health.js (daily cadence, 2-day alarm). Publishers skip silently without credentials, so local/PR runs are unchanged. Operator action to activate publishing: add UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN as GitHub Actions secrets. Closes #4920 🤖 Generated with Claude Code Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…uble-quoted names, parse-cache bump, bounded external fields (#4927 review round) Applies the multi-agent + cross-model review findings: - th/tr/zh compiledFrom carried single-brace placeholders (fan-out ran before the {{token}} convention fix) — repaired; fa.json was missing the keys entirely AND absent from translate-locales' LOCALES list (pre-existing gap the parity test caught) — keys added, fa added to the translator roster - extractServerFeeds skipped double-quoted names ("Tom's Hardware") — regex accepts both quote styles; extraction test pins it (cross-model) - rss:feed parse cache bumped v5→v6: warm v5 rows lack droppedFeedCap and would undercount the coverage ledger for their whole TTL (cross-model) - coverage-ledger writes clamped to known variants + well-formed langs (request-supplied values could inflate the keyspace — adversarial) - GDELT titles AND urls bounded before publishing; urls http(s)-only - direct-run guards use pathToFileURL (Windows/symlink-safe); both publishers share scripts/_upstash-rest.mjs instead of hand-rolled fetch helpers; publishFeedHealth exported with a no-creds contract - .insights-provenance CSS added (was an unstyled div); gn()/gnLocale replica drift guard test; overflowDropped + orchestration-seam tests Battery: completeness (21), locale-completeness (25), clustering (16), story-identity (30), both typechecks, biome clean. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…trap-hydrated), MCP exclusions, rss:feed v6 pin update (#4920) The stacked CI runs caught three registry gates the local scoped battery missed: (1) the new ops keys had landed in BOOTSTRAP_KEYS, which would hydrate feed-health/recall payloads into every page load — moved to STANDALONE_KEYS (health-monitored, on-demand); (2) both keys needed documented EXCLUDED_FROM_MCP entries (ops surfaces, not queryable news slices); (3) the rss:feed cache-prefix pin test still asserted v5 — updated for the #4920 v6 bump with the invalidation rationale. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…d fails (#4927 external review) A transient Redis error during the previous-payload GET previously reset every consecutive-empty streak (previous=null -> fresh streaks). Read failure now skips the publish entirely; key-absent (first run) still starts fresh. Regression test included. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

… attribution, publish-after-guardrails (#4927 re-review) All six re-review findings applied: P1 — never-activated publishing no longer poisons health: the two keys join ON_DEMAND_KEYS (missing data = soft EMPTY_ON_DEMAND) and their seed-meta entries are activationGated in api/seed-health.js ('pending-activation', healthy) until the first publish lands — then normal 2880-min staleness enforcement applies. The workflow also emits a visible ::warning annotation on runs where the UPSTASH secrets are absent, so green scheduled runs can't silently mask the gap. P2 — the coverage-ledger write is awaited (.catch non-fatal): the Edge response could previously finish before the fire-and-forget side write landed. P2 — selectTopStories classifies EVERY admissible candidate instead of breaking at maxCount: cap-first attribution (a candidate whose source already hit the per-source cap is a sourceCap drop even when selection is full; only genuinely rankable candidates count as overflow), with a same-source-overflow regression asserting exact one-of attribution. P2/P3 — StoryMeta.source_count proto doc now describes what runtime ships (current-cycle batch corroboration, default 1) instead of 'cached from Redis Set'; generated artifacts regenerated. P3 — _upstash-rest.mjs sends a User-Agent (CF-WAF convention: never a bare fetch UA); feed-health publishing moved AFTER the config-drift hard-fail so a guardrail-failing run cannot refresh health metadata first. Battery: completeness 23, clustering 16, mcp-parity 11, seed-health 2, typecheck:api, docs-stats. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…re FIRST publish (#4927 re-review P1) Activation was inferred from seed-meta existence, which carries a 7d TTL: a publisher that ran once and died alarmed as stale only until the meta expired, then reverted to healthy pending-activation — violating the stated contract, and /api/health softened the keys permanently via ON_DEMAND_KEYS. "Has ever published" is now a durable marker (seed-activated:news:*, SET with NO TTL on every successful publish): - api/health.js pipelines EXISTS on the markers; a marker-bearing key leaves ON_DEMAND softening permanently — missing data/meta reads EMPTY/STALE_SEED like any other monitored key (marker read errors degrade to not-activated, never to a false alarm) - api/seed-health.js entries carry activationKey; 'pending-activation' is returned ONLY while the marker is absent — with the marker present, missing meta falls through to 'missing' (alarming) - both publishers SET their marker alongside every publish Regressions: classifyKey lifecycle (pending -> EMPTY_ON_DEMAND; activated-then-dead -> EMPTY) + publisher/marker wiring pins including the no-TTL assertion. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…marker probe (#4927) The harnesses assert every pipeline command is GET; the activation markers add EXISTS probes on seed-activated:* keys, which tripped the mock assertion and 503'd every request in these suites. Both mocks now accept EXISTS (marker-namespace-pinned, absent by default). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…laced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

…nce-demotion seam (#4922) (#4929) * feat(markets): wire markets into the news spine — macro-print actuals, earnings in the daily brief, finance-demotion seam (#4922) The no-proto slice of the markets↔news epic (ticker tagging on the wire and watchlist story alerts need a NewsItem proto field — tracked in #4922 as the follow-up slice): (b) Macro actuals. The economic calendar shipped every event with actual/estimate/previous empty forever — the print itself was never captured. The seeder now pulls FRED series observations (release ids and series ids are different namespaces; mapped per event with the right transform: CPI/PCE/Retail index→MoM %, NFP level→ΔK, GDP headline % direct), publishes a recentPrints block, and fills actual/previous on print-day events. The proto already carries the fields — values flow to the client with zero schema change. (c) Earnings reach a brief for the first time. Daily Market Brief gains earningsContext (recent beats/misses + upcoming count) collected via the earnings-calendar RPC with an 8s budget and rendered into the LLM context — symbols + direction only, stable per reporting date, so the #4914 summary-cache identity does not churn. (f) Finance-demotion seam. scoreImportance/selectTopStories accept {demoteFinance:false} — the ×0.35 demotion stays default for the geopolitical World Brief and is now switchable for a finance-focused ranking surface. Honest note: the only live consumer runs the full variant today; this is the seam a finance-variant insights run plugs into. Tests: print transforms (incl. FRED '.' markers), fill semantics (no-overwrite, future-stays-empty), demotion ×0.35 default vs neutral, earnings prompt block, wiring. 85 tests across the affected battery. Refs #4922 (remaining slices tracked there) 🤖 Generated with Claude Code Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP * fix(review): markets wiring hardening — deliverable earnings lookback, print adjacency guards, unitless values, pure transform (#4929 review round) Applies the multi-agent + cross-model review findings: - the "recent beats/misses" block was undeliverable as wired: the earnings seed window started at today, so past reporters never existed in the payload (correctness + codex, cross-corroborated) — seed window widened to today-7d - computePrintValues hardened per codex P1: a '.' LEADING observation no longer falls back to an older row presented as current, and pct_mom/diff_k require date-adjacent periods (45-day tolerance) so a mid-series gap cannot mislabel a 2-month change as month-over-month - print values are now UNITLESS — event.unit already renders '%'/'K' and embedding it double-rendered in the calendar (codex conf-100) - recentPrints consumer scope documented in-code: the RPC reconstructs its proto response without it (Redis/ops-side until the proto gains a field — epic follow-up); the user-visible win flows through the event actual/previous fields the proto already carries - earnings transform extracted to pure buildEarningsBriefContext (this PR's own standard, per review) with direct tests; earnings prompt block covered by a capture test asserting content AND cache-identity stability across rebuilds; selectTopStories opts pass-through exercised end-to-end; EVENT_SERIES↔FRED_RELEASES name lockstep pinned by test - verified: fredFetchJson error paths never embed the keyed URL 45 tests across the affected battery; typecheck + biome clean. Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP * fix(review): release-day stale-print guard, benchmark fast-fail, misplaced-opts shift (#4929 external review) - observationMatchesRelease: a filled actual must belong to the period the release reports (monthly diff==1; quarterly 1..5 for the advance/second/third GDP estimates) — pre-print on release day the latest FRED observation is the PRIOR period and was presented as today's number (P1); regression pins the July-15-CPI/May-obs case - recall benchmark uses fast-fail GDELT limits (10s timeout, 1 retry, 1 proxy attempt) under a shared 4-minute deadline — production retry defaults across 4 sequential queries could eat the 25-min workflow - selectTopStories auto-shifts an opts-shaped object passed in the stats slot (positional-API hazard), tested Also validated-not-changed: cross-cycle canonical adoption is the documented #4925 follow-up; seed-meta recordCount=ok is intended job-health semantics (streak preservation landed on the #4927 branch). Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

… — unblock feed-health activation (#4970) (#4971) Three server-catalog feeds (Pentagon, The National, Truth Social) had hosts absent from the RSS allowlist, tripping the CONFIG_DRIFT hard-fail in validate-rss-feeds.mjs --ci (exit 1) which — per the #4927 ordering — blocks feed-health/recall-benchmark publishing before it runs. Pre-existing drift (DoD rebranded defense.gov→war.gov; the other two were added to the catalog without allowlisting), surfaced by activating the publish path. Adds the three hosts to all mirrors: shared/rss-allowed-domains.json (source of truth) + its byte-identical scripts/shared mirror + the api/_rss-allowed-domains.js copy the validator and Edge SSRF guard actually read + vite.config.ts RSS_PROXY_ALLOWED_DOMAINS (dev proxy). The .cjs re-requires the JSON (no edit). Verified: isAllowedDomain() now returns true for www.war.gov, www.thenationalnews.com, trumpstruth.org (exact/bare/www rule); shared and api lists are byte-equal (339); edge-functions + scripts-shared mirror guards green (228). feed-validation.yml does not run on PRs (SSRF surface) — post-merge workflow dispatch is the end-to-end proof. Closes #4970 Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

mintlify Bot deployed to staging - docs July 5, 2026 20:02 View deployment

vercel Bot deployed to Preview July 5, 2026 20:04 View deployment

greptile-apps Bot reviewed Jul 5, 2026

View reviewed changes

koala73 added a commit that referenced this pull request Jul 5, 2026

fix(review): include the digest-side hardening — ledger keyspace clam…

f98e584

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

vercel Bot deployed to Preview July 5, 2026 20:22 View deployment

This was referenced Jul 5, 2026

feat(brief): the brief contract — top-8 synthesis, verified citations, enforce-mode gate, shared grounding spine (#4921) #4928

Merged

feat(markets): macro-print actuals, earnings in the daily brief, finance-demotion seam (#4922) #4929

Merged

vercel Bot deployed to Preview July 5, 2026 21:04 View deployment

koala73 force-pushed the feat/completeness-4920 branch from 35e09c9 to 31d7a04 Compare July 6, 2026 04:44

koala73 added a commit that referenced this pull request Jul 6, 2026

fix(review): include the digest-side hardening — ledger keyspace clam…

6bcb9e3

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

mintlify Bot deployed to staging - docs July 6, 2026 04:45 View deployment

vercel Bot deployed to Preview July 6, 2026 04:48 View deployment

koala73 added a commit that referenced this pull request Jul 6, 2026

fix(review): include the digest-side hardening — ledger keyspace clam…

ba5142a

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

koala73 force-pushed the feat/completeness-4920 branch from 31d7a04 to e471d51 Compare July 6, 2026 05:14

mintlify Bot deployed to staging - docs July 6, 2026 05:16 View deployment

koala73 mentioned this pull request Jul 6, 2026

feat(news): unified story-identity service — one similarity definition for clustering, dedup, corroboration (#4919) #4924

Merged

vercel Bot deployed to Preview July 6, 2026 05:18 View deployment

mintlify Bot deployed to staging - docs July 6, 2026 08:29 View deployment

vercel Bot deployed to Preview July 6, 2026 08:31 View deployment

vercel Bot deployed to Preview July 6, 2026 08:59 View deployment

koala73 added 8 commits July 6, 2026 14:53

fix(review): include the digest-side hardening — ledger keyspace clam…

7c7274e

…p + rss parse-cache v6 bump (#4927 review round, missed pathspec) Claude-Session: https://claude.ai/code/session_01WmhkGjZrb9RbV7pV3muFxP

koala73 force-pushed the feat/completeness-4920 branch from fc059a3 to e6dd8ed Compare July 6, 2026 10:54

mintlify Bot deployed to staging - docs July 6, 2026 10:55 View deployment

vercel Bot deployed to Preview July 6, 2026 10:57 View deployment

koala73 merged commit 5cf7b0f into main Jul 6, 2026
25 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(news): completeness measurement — feed health w/ silent zeros, coverage ledger + provenance, GDELT recall benchmark (#4920)#4927

feat(news): completeness measurement — feed health w/ silent zeros, coverage ledger + provenance, GDELT recall benchmark (#4920)#4927
koala73 merged 8 commits into
mainfrom
feat/completeness-4920

koala73 commented Jul 5, 2026

Uh oh!

mintlify Bot commented Jul 5, 2026 •

edited

Loading

Uh oh!

vercel Bot commented Jul 5, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot commented Jul 5, 2026

Uh oh!

greptile-apps Bot Jul 5, 2026

Uh oh!

greptile-apps Bot Jul 5, 2026

Uh oh!

koala73 commented Jul 5, 2026

Uh oh!

koala73 commented Jul 6, 2026

Uh oh!

koala73 commented Jul 6, 2026

Uh oh!

koala73 commented Jul 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

koala73 commented Jul 5, 2026

Three layers of "did we miss something?"

Operator action

Verification

Uh oh!

mintlify Bot commented Jul 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vercel Bot commented Jul 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

greptile-apps Bot commented Jul 5, 2026

Greptile Summary

Confidence Score: 3/5

Important Files Changed

Flowchart

Uh oh!

greptile-apps Bot Jul 5, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Jul 5, 2026

Choose a reason for hiding this comment

Uh oh!

koala73 commented Jul 5, 2026

Multi-agent code review — applied at f98e584

Applied

Noted, not changed

Uh oh!

koala73 commented Jul 6, 2026

Uh oh!

koala73 commented Jul 6, 2026

Re-review adjudication — all 6 findings validated, fixed at d1c6298

Uh oh!

koala73 commented Jul 6, 2026

Re-review round 2 — the remaining P1 fixed at $(git rev-parse --short feat/completeness-4920)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

mintlify Bot commented Jul 5, 2026 •

edited

Loading

vercel Bot commented Jul 5, 2026 •

edited

Loading

Multi-agent code review — applied at `f98e584`

Re-review adjudication — all 6 findings validated, fixed at `d1c6298`