HealthLog/docs/audit/v146-findings.md

HealthLog v1.4.6 — Findings & Spec

Status: 2026-05-09 — handoff from the v1.4.5 deploy session. All findings were produced by four parallel audit agents (dashboard visual, KI section, admin panel, KI insights data window) against the v1.4.5 codebase (e8f4820). A Playwright probe also visually verified the dashboard tile strip at five viewports. Verify each finding by reading the cited file:line BEFORE editing — agents occasionally reference stale state.

The single working-tree change carried over from the v1.4.5 session: src/components/charts/trend-card.tsx:85 — added w-full so the tile component fills its grid cell. Do not commit alone — fold into v1.4.6.

---

Tier 1 — Release blockers for v1.4.6

T1 — Dashboard tile fill ✅ (already in working tree)

src/components/charts/trend-card.tsx:85

The TrendCard outer div had flex h-full flex-col but no w-full. As a flex-child of a <div className="flex"> wrapper, it took content width instead of stretching to the grid cell. Visible on mobile-narrow (375px): tile widths varied 122-166px in the same row. With w-full added, the Playwright harness confirms the tiles fill their grid cells uniformly at all viewports. Marc's primary complaint was "die Kacheln füllen den Platz nicht aus" — this is exactly that.

Verify after change: Run the deep diagnose probe at multiple viewports (/tmp/v145-tile-deep-diagnose.js from prior session if still on disk; otherwise re-create — set HL_SID to a valid session and run via node /Users/marc/.claude/skills/playwright-skill/run.js $probe).

T2 — --muted-foreground equals --foreground (no hierarchy)

src/app/globals.css:97 (dark) and :134 (light)

Currently --muted-foreground: #f8f8f2; is identical to --foreground: #f8f8f2; in dark, and #1f1f1f in both for light. Every secondary string (tile labels, "Ø7d:" prefixes, units, nav inactive items) is rendered at the same brightness as primary text — visually flat.

Fix: dark → --muted-foreground: #9aa3b3; (dracula-comment). Light → --muted-foreground: #5b6273;.

Verify: Visual diff at a few pages — dashboard, settings, admin — should show clear emphasis between primary numbers and label/footer text. WCAG AA contrast must still hold (≥ 4.5:1 against the surface).

T3 — KI insights primaryRecommendation renders raw chart tokens

src/components/insights/insight-advisor-card.tsx:411

The base-system prompt (src/lib/insights/prompts/base-system.ts:67) explicitly invites the LLM to embed metric:WEIGHT etc. tokens inside primaryRecommendation. The renderer prints {insight.primaryRecommendation} raw — no token-strip, no <InlineCharts> parse. Result: when the model follows instructions, the user sees literal metric:WEIGHT in their hero recommendation.

Fix: Apply the same pattern used at lines 419-423 for summary: strip tokens for prose, parse and render via <InlineCharts>. Re-use the existing helper.

T4 — aiBaseUrl cross-provider leak (security)

src/lib/ai/provider.ts:75, 312, 355, src/app/api/user/ai-provider/route.ts:54-62

buildUserProvider and resolveProviderForTest reuse a single aiBaseUrl column across providers. If a user once configured LOCAL with http://192.168.x.x/v1 and later switches to OPENAI, the stored aiBaseUrl is still set, so OpenAIClient is constructed with baseUrl: row.aiBaseUrl ?? "https://api.openai.com/v1" — sending the user's OpenAI API key to their local server.

Fix: In the PATCH handler at route.ts:54-62, when provider changes to anything other than LOCAL, wipe aiBaseUrl. Belt-and- braces in buildUserProvider: ignore aiBaseUrl for OPENAI and ANTHROPIC (use a hardcoded default). Add a unit test for provider-switch with stale baseUrl.

T5 — /api/insights/generate 502 → Cloudflare HTML rewrite

src/app/api/insights/generate/route.ts:84-86

Same Cloudflare-rewrites-5xx footgun the test endpoint just got fixed for in v1.4.5. Parse failure returns apiError("Failed to parse AI response", 502) → Cloudflare swaps the body for an HTML error page → await res.json() in ai-section.tsx:206-215 crashes with Unexpected token '<'. The user has zero rate-limit budget left to retry (2/hour) and sees nothing useful.

Fix: Map parse errors to 422 with a readable JSON error body. Same one-liner pattern as the v1.4.5 ai/test fix (see CHANGELOG).

T6 — Admin status-card hrefs broken

src/components/admin/status-card-grid.tsx:147, 159, 177, 189, 210, 222

Hrefs to non-existent routes/anchors:

• /admin/users, /admin/audit-log — sub-routes don't exist
• /admin#integrations, /admin#monitoring, /admin#backups, /admin#maintenance — anchor IDs don't match the rendered DOM (actual: section-admin-services, section-admin-umami, etc.)

Every "Manage users" / "View backups" CTA reloads /admin to top. The status-card-grid.test.tsx:105 only asserts startsWith("/admin") — so this regression slipped through.

Fix: Point hrefs at the existing anchor IDs:

• Users → /admin#section-user-management
• Audit log → /admin#section-login-overview
• Integrations → /admin#section-admin-umami
• etc. Verify each anchor exists in the rendered DOM.

Tighten status-card-grid.test.tsx to assert each href anchor IS present in the rendered admin page DOM (use render(<AdminPage />) and assert getByText(...) for each section ID). Plain regex check is not enough.

T7 — Bug-report toggle has no effect (admin)

src/app/api/bugreport/route.ts:47-49, src/app/bugreport/page.tsx:70

The bugReportEnabled === false check lives on the legacy /api/bugreport route. The actual form on /bugreport posts to /api/feedback, which has no such gate. Result: toggling "Bug reports off" in admin does not actually hide or block the form — users still submit successfully into the local feedback inbox.

The CLAUDE.md / commit message claim ("the report button gracefully disappears") is false.

Fix: Pick one approach:

• (a) Add bugReportEnabled to /api/bugreport/status and to /api/feedback POST so the form is hidden + blocked when off. Recommended.
• (b) Update copy to clarify the toggle only governs GitHub publishing.

Pick (a). Update i18n copy. Add an integration test that toggles the flag off and asserts /bugreport returns 503 OR redirects.

T8 — "Wipe all data" deletes more than the copy says

src/app/api/admin/data/route.ts:44-47

The DELETE handler also nukes apiToken, withingsConnection, auditLog, authChallenge. The UI confirmation copy (admin.deleteAllConfirmDescription) and the success toast (admin.deletedResult) only mention measurements/medications/ intakeEvents.

Worse: tx.auditLog.deleteMany({}) runs in the same transaction — every prior security-relevant entry is gone. For a feature whose primary purpose is auditability, this is wrong.

Fix:

• Stop deleting AuditLog. Keep the audit trail. (If retention is desired, add a separate maintenance route that prunes by age.)
• Reflect the actual scope in the response toast (the API already returns apiTokens, withingsConnections, authChallenges).
• Update i18n keys admin.deleteAllConfirmDescription to enumerate everything that gets cleared.

T9 — KI per-card window 30 → 360 daily + 24 monthly (NEW SCOPE)

Affected files (one constant + one bucket-builder per file):

• src/lib/insights/general-status.ts:15 — GENERAL_STATUS_POINTS
• src/lib/insights/blood-pressure-status.ts:21 — BLOOD_PRESSURE_STATUS_POINTS + mood.slice(-30) at line 362
• src/lib/insights/pulse-status.ts:17 + line 199 mood-slice
• src/lib/insights/weight-status.ts:17 + line 233 mood-slice
• src/lib/insights/bmi-status.ts:11
• src/lib/insights/mood-status.ts:14 + line 283 (recentEntries slice 90 — already padded × 3)
• src/lib/insights/medication-compliance-status.ts:150 — rangeStart

Goal: the LLM should see ~3 years of history with newer data at higher resolution. Not raw rows — daily means for the recent window, monthly means for the older window. Marc explicitly asked for "mindestens 360 Tage" daily, plus "Monatswerte" for "2 oder 3 Jahre".

Concrete shape of the prompt payload (per metric):

{
  daily: Array<{ dayOffset: number; value: number; n: number }>; // 360 items, dayOffset 0..359 (newest first)
  monthly: Array<{ monthOffset: number; value: number; n: number }>; // 24 items, monthOffset 12..35 (12 months ago to 36 months ago)
}

So months 1-12 are sampled daily (= 360 daily buckets), months 13-36 are summarised monthly (= 24 monthly buckets). Total 384 data points per metric.

Token budget guard: Before sending, log the JSON-stringified payload size. If >50 KB, fall back to 180 daily + 24 monthly. Add this as a LOG_INSIGHTS_PAYLOAD_SIZE env var (default true) so we can monitor.

Tests: add a test per generator that asserts the payload contains both daily and monthly arrays with the expected lengths, and that no raw measurement rows leak into the prompt (the audit window is 3 years; raw mode is unchanged in features.ts).

---

NEW FEATURE — Chart bucketing for ranges > 1 year

Marc's request:

"Überlege dir bei den Charts ob wir Sachen halt nicht ab einer gewissen Sache dann stacken in den Monaten. Wenn wir jetzt bei All drücken zum Beispiel, dass man dann halt ab einer Zeit von über einem Jahr dann halt danach tagesmonats- oder monats- oder wochendurchschnitte summiert und nicht dann 30.000 Punkte anzeigt"

Spec:

The chart data adapter (in src/components/charts/ — likely health-chart.tsx or a sibling) currently renders every measurement in the selected range. For the dashboard "All" / "Alle" range filter this can be 1000s of points and is unreadable.

Add automatic bucketing based on the rendered range:

• ≤ 90 days → daily points (current behaviour)
• 91-730 days → weekly average
• > 730 days → monthly average

The bucket type must be displayed in the chart header (e.g., "Tageswerte" / "Wochendurchschnitt" / "Monatsdurchschnitt" chip in the corner of the chart) so the user knows what they are looking at.

The data points must keep their ISO timestamps so Recharts' x-axis formatting works correctly. Smoothing/averaging must skip days with n=0 rather than count them as 0.

Tests: unit-test the bucketing function with synthetic series of varying length. Visual regression: render the chart at a 2-year range and verify the chart shows ~24 monthly points, not 730 daily.

Research: the autonomous executor MUST research best practice for time-series bucket aggregation in healthcare/Recharts dashboards before implementing. Use Context7 or websearch for "recharts aggregation buckets" and similar — there's a standard pattern for this. Don't reinvent.

---

Tier 2 — Polish (also for v1.4.6)

ID	File	Change
P1	trend-card.tsx:93,109,118,130,139	Add tabular-nums to numeric spans (digits jiggle on refresh)
P2	trend-card.tsx:87,93	Bump value text-2xl → text-3xl tracking-tight, label text-sm font-medium → text-xs uppercase tracking-wide (KPI hierarchy)
P3	trend-card.tsx:85 vs health-chart.tsx:551	Match padding: tile p-3 → p-4 md:p-6 to align with chart cards
P4	trend-card.tsx:64-69	Trend-arrow color: keep flat at muted; up/down both text-foreground (neutral). Direction-as-good-or-bad is metric-specific and v1.5+ work.
P5	trend-card.tsx:103-145	Always render avg7/avg30 chips; use — when null so vertical rhythm is consistent across tiles
P6	page.tsx:362	Welcome subtitle: drop hidden sm:block, add text-muted-foreground (visible on mobile, muted)
P7	auth-shell.tsx:129	Bottom-nav buffer: pb-[calc(5rem+env(safe-area-inset-bottom,0px))] → pb-[calc(4rem+env(safe-area-inset-bottom,0px))] md:pb-0
P8	empty-state.tsx:71-77	TrendHint title default: add text-muted-foreground so the title doesn't fight chart titles after T2
P9	getting-started-checklist.tsx:299	Drop font-mono from "X von Y" / "%" — keep tabular-nums
P10	page.tsx:846-862	Medications card: rounded-lg border p-4 → rounded-xl border p-4 md:p-6 (match other chart cards)
P11	src/lib/logging/redact.ts:22	Add sk-(ant-)? regex to redactSecrets to scrub OpenAI/Anthropic keys centrally
P12	src/lib/idempotency.ts:227	Extend the hlk_ / hlr_ exclusion to also reject bodies containing "sk-" or "sk-ant-"
P13	src/app/api/insights/generate/route.ts:18-22	Move checkRateLimit BELOW the cache-return so cached hits don't burn rate-limit tokens
P14	src/lib/ai/codex-client.ts:28,34	Copy structured-error pattern from openai-client.ts:42-58 (httpStatus, bodyExcerpt, upstream fields)
P15	src/components/settings/ai-section.tsx:49	Drop gpt-5 from MODEL_PRESETS["OPENAI"] (not a released model). Drop o3-mini too unless you wire the o-series param contract (max_completion_tokens not max_tokens, no temperature override).
P16	src/components/settings/ai-section.tsx (~29 strings)	i18n: pipe German strings through t("settings.ai.…") — add keys to messages/en.json + messages/de.json
P17	src/components/admin/feedback-inbox-section.tsx:209-222	Replace bg-red-500/15 text-red-400 with bg-dracula-red/15 text-dracula-red etc. for theme consistency
P18	src/components/admin/danger-zone-section.tsx:110	Drop string-prefix-matches-success heuristic — track via mutation.isSuccess / mutation.isError
P19	src/components/admin/_shared.tsx:268-277,216-225	useSystemStatus() and useAdminSettings(): surface isError with an inline "Failed to load settings" message instead of infinite spinner
P20	src/app/api/admin/status-overview/route.ts	Promise.all → Promise.allSettled server-side so one failed probe doesn't blank the whole grid

---

Tier 3 — explicitly deferred

• One-row scroll vs. wrap — the dashboard audit recommended rolling back v1.4.5's wrap behaviour to v1.4.4's horizontal-scroll- always. Marc already chose wrap deliberately in v1.4.5 (CHANGELOG entry, commit e8f4820). Do not roll back.
• /admin/users, /admin/audit-log sub-routes — feature, not bug. Anchor-link fix in T6 is enough for v1.4.6.
• API token revoke / login block actions in admin — feature for v1.5.
• "Remove saved AI key" button — UX gap, not a bug. v1.5.
• Feedback / mood semantic colors — needs design pass.

---

Verification gates (before tagging v1.4.6)

After all fixes, the autonomous executor MUST:

1. Test suite green: pnpm test, pnpm typecheck, pnpm lint, pnpm format:check. Stop if any fails — never --no-verify.
2. Integration tests green (testcontainers): pnpm test:integration
3. Build succeeds: pnpm build (no Turbopack-specific failures)
4. Multi-agent QA review — dispatch in parallel and reconcile findings into a follow-up commit if any HIGH/CRITICAL surfaces:
   • superpowers:code-reviewer — review the v1.4.6 diff vs. main
   • Plan agent as senior-design-reviewer — visual review of dashboard, settings, admin at desktop + mobile viewports
   • Plan agent as senior-security-reviewer — review T4, T5, T7, T8, P11, P12 specifically for residual risk
   • simplify skill — run on the v1.4.6 changed files to look for accidental complexity
5. Mobile QA via Playwright — at least three viewports (375×667, 390×844, 412×915). Probe must verify:
   • No horizontal scroll on any page
   • Bottom-nav doesn't overlap content
   • Tile strip wraps cleanly with uniform widths per row
   • Touch targets ≥ 44×44 px
   • Charts render and are readable
6. Hallucination check — for every finding fixed, git diff must show a real change at the cited file:line. Audit-agent claims must be verified; if a finding's file:line doesn't match what's actually there, log it in the SUMMARY and skip rather than guess.
7. Visual regression — Playwright screenshots of dashboard at 1280×900 and 390×844, before+after, saved to /tmp/v146-tiles-{desktop,mobile}-{before,after}.png.

---

Release & deploy

1. Bump package.json to 1.4.6
2. Update CHANGELOG.md with a [1.4.6] — 2026-05-09 (or 2026-05-10 depending on completion time) block. Tone: user-facing German + English, no Claude mention, no "Audit-Agent", no internal phase names. Group by section: "Fixed — Dashboard", "Fixed — KI", "Fixed — Admin", "Improved — KI Insights", "Improved — Polish".
3. Atomic commits per phase, each Co-Authored-By Claude Opus 4.7.
4. Tag v1.4.6, push tag.
5. Wait for GHCR build (typically 10-13 min).
6. Coolify deploy: mcp__coolify-apps01__deploy with tag_or_uuid: pg8wggwogo8c4gc4ks0kk4ss, force: true.
7. Verify production:
   • Poll /api/version until 1.4.6
   • docker inspect running container — image digest must differ from current.
   • Marc's session: cmox4d6fj000101p8w9ykhcnm (still valid as of 2026-05-09; if expired, re-pull from prod DB — SELECT id FROM sessions WHERE user_id='cmlupy4tn000001rpzx1pxvz7' AND expires_at > now() ORDER BY created_at DESC LIMIT 1;)
   • Hit /api/ai/test with fake key → still returns 422
   • Hit /api/insights/generate with forceRefresh=false and a fresh user → does not consume rate-limit token (P13 verify).
8. Mobile QA Playwright probe against production.
9. Brief Marc — German, ≤ 200 words. Hard-reload reminder. Group by what changed perceptually ("Tile-Strip füllt jetzt aus", "KI bekommt jetzt 3 Jahre Historie", "Charts aggregieren bei Langzeit-Range", "Admin-Buttons springen jetzt zum richtigen Anker", "Bug-Report-Toggle wirkt jetzt"). No "Claude", no internal phase names.

---

Repo housekeeping (also in scope for the v1.4.6 marathon)

CI

• e2e workflow has been failing on every commit since at least cbded7c3. Investigate (gh run view <id> --log-failed), root- cause, fix the spec or the underlying app bug. Don't .skip() it — Marc explicitly asked for green CI.

GitHub Releases

GitHub Releases stop at v1.4.1. Backfill v1.4.2, v1.4.3, v1.4.4, v1.4.5, v1.4.6 — each release notes pulled from the corresponding CHANGELOG.md block. Tag each release --verify-tag so we don't create an orphan release.

GHCR Packages

ghcr.io/mbombeck/healthlog packages page may have stale tags from the v1.4.0/v1.4.1 deploy thrash. Don't delete v1.x.x tagged versions (some users may pin). Only delete explicitly untagged manifests if any exist.

Docs site (/Users/marc/Projects/healthlog-docs)

Astro site at https://docs.healthlog.dev. Latest content covers up to v1.2. Bring it up to v1.4.6:

• Update version selector / Latest indicator
• Add docs pages or sections for v1.3-v1.4.6 features the user would care about — read the CHANGELOG.md user-facing bullets and translate into how-to docs where appropriate
• Add the v1.4.6 release entry

Read the docs repo's CLAUDE.md/CONTRIBUTING (if present) FIRST to understand conventions. Commit + push.

Landing site (/Users/marc/Projects/healthlog-landing)

Next.js site. Latest showcase mentions v1.2. Minimal updates only — don't redesign:

• Update headline version + supported-features list if anything v1.3-v1.4.6 is landing-page-worthy
• Update screenshots if anything visibly changed (likely the dashboard tile strip)

Commit + push.

---

Marc's hard rules (verbatim)

1. Niemals --no-verify bei git commit.
2. Niemals --no-gpg-sign außer explizit angefragt.
3. Niemals force-push to main.
4. Changelogs/Release-Notes nur user-facing — kein "Claude", keine internen Phasennamen, keine Audit-Sprache.
5. Authorization gilt nur für den spezifischen Auftrag — wenn etwas Größeres aufkommt, lieber fragen als entscheiden.

For this overnight session, Marc has explicitly authorized autonomous execution of Tier 1 + Tier 2 + the new chart bucketing feature, and the multi-agent QA + simplify pass. Anything that crops up which is not in this document and is larger than a typo-level change must be deferred to v1.5 with a note in the SUMMARY rather than implemented.