Weekly Summary (Monday-morning recap, folded into Daily Digest)

What it is

Weekly Summary is a once-a-week AI-written recap of how you used Omniscio over the just-completed week, plus a short list of personalized suggestions for getting more out of the app. As of 2026-05-14 it ships as a briefing inside Daily Digest — same sidebar entry ("Briefings"), same inbox section, same triage shortcuts. The time window is Monday 00:00 → Sunday 23:59:59 local (ISO week) and the input is a single envelope blending three things: aggregate stats (sessions, messages, top projects, top features, AI spend, week-over-week deltas), an eleven-field detector payload (six privacy-preserving behavioral detectors from Workflow Coach plus four newer ones: top operator phrases, feature-usage deltas, attention backlog aging, cron-job health), and the last 90 days of suggestion dismissals so the model learns from your past "not useful" picks. The suggestions themselves come from a Map → Reduce → Verify three-step pipeline (Luna → Opus 4.7 → Luna) where Opus must self-cite which envelope fields each suggestion depends on and Luna fact-checks those claims before the row lands. (The Map and Verify model follows the global Utility AI model setting and the per-feature weeklySummaryModel override — see Settings field map below; Reduce stays on Opus 4.7.) Weekly Summary subsumes Workflow Coach — the standalone Suggestions panel and the separate weekly inbox alert are gone. In the inbox the briefing now surfaces as two cards (2026-07-20): a slim scannable recap card and a separate "N suggestions from last week" card, so the suggestions get their own labelled row (see How to use it §4).

On by default (owner decision 2026-07-27): a fresh install generates a recap every Monday at 7 AM local without opting in. On first launch it seeds only the most-recently-completed week — older missed weeks are treated as past-due and skipped (the freshness gate under How it works, 2026-08-03), so a user who missed an app update or was away for a while never gets a burst of stale, weeks-old recaps dumped into their inbox. Existing users were force-enabled once when default-on shipped (a one-time weeklySummaryDefaultOnMigrationDone flip — their config already carried an explicit false, so the new schema default couldn't reach them). It costs nothing until there's an API-key account to bill — an account on a Claude login (OAuth) can't call the Messages API, so the Monday run skips instead of erroring. That skip is no longer silent (2026-08-29): because OAuth tokens can never reach the Messages API, it fired every week forever with nothing but a log line — the skip returns before the LLM calls, so it throws nothing, never reaches the failure path, and never fed the give-up alert, leaving an on-by-default feature permanently dead and unexplained. It now raises the shared "Add an API key to turn on AI helpers" inbox card — the same deduped card the other AI helpers use, carrying the one-click Add API key button, latched once per run and self-clearing once a key exists. Turn it off any time in Settings (below); once off, it stays off — and a user who turned it off gets no card.

Where to find it

Open Briefings from the Omniscio sidebar. That one entry lists daily digests and weekly summaries together, newest first, each row carrying a header pill that says which cadence it is, with unread rows marked by the same small amber dot used elsewhere for needs-you signals. A weekly row falls into the same Active, Snoozed and Archived buckets as a daily one.

In the Inbox, the same week appears under the Briefings section as two separate rows: the recap card, and a second card titled "N suggestions from last week".

Its settings live at Settings → Email & Summaries → Daily digest. Expand the card and scroll to the Weekly Summary sub-section at the bottom of the panel, where the master toggle, the generation hour, the daily cost cap and the Run now button all sit. The standalone weekly sidebar entry, its own inbox section and its old Settings → AI Coach home are gone — those four surfaces were folded into Daily Digest.

How it behaves

How to use it

1. It's already on (change it here). Weekly Summary is on by default — there's nothing to enable. To turn it off (or back on): Settings → Email & Summaries → Daily digest (expand the card), scroll to the Weekly Summary sub-section at the bottom of the panel, and flip the master toggle. Pick a generation hour (0–23, defaults to 7 AM local — the summary lands before you check email Monday morning). Optionally tune the daily cost cap (default $0.10, range $0.05 – $5.00) — once the cap is hit Omniscio stops generating that day and resumes the next. The Daily Digest and Weekly Summary toggles are independent — you can turn one on without the other.
2. Let it run. A periodic tick (every 15 minutes) checks the clock; once your chosen hour has passed on a Monday and the just-completed week's summary hasn't been generated yet, it runs. First launch: if weeklySummaryBackfillCompleted is still false, Omniscio seeds only the most-recently-completed week — older candidate weeks are past-due (below the freshness horizon) and skipped, so a stranded or returning user never gets a backlog of old recaps. The generation is still gated by the same daily cost cap as the live path, and the flag flips in finally so a transient error doesn't loop on every relaunch (a past-due skip is not a cap-halt, so the flag still completes).
3. Read it. Click Briefings in the Omniscio sidebar. The left pane lists every briefing newest-first — daily digests AND weekly summaries appear in the same list, distinguished by a header pill on each row. Unread rows carry the same small amber dot used elsewhere for needs-you signals. Clicking opens the viewer: a stats card at the top (sessions total / active, agent vs. operator message counts, top projects with session counts, top features with usage counts, weekly AI usage (a nominal, pay-as-you-go value — see AI usage is a nominal value below), and total tokens for the week — shown as the combined input+output count with an "in / out" breakdown; every stat-card figure renders in the app's compact form, e.g. 25.4K messages · $237.6K spend · 62.2B tokens, matching the Statistics dashboard), the AI-written scannable recap in the middle (a bold headline + highlight bullets, not dense paragraphs), then the suggestion cards at the bottom. (This dedicated Briefings sidebar viewer shows the full recap + suggestions; the inbox splits the same content into the two cards described in §4.)
4. Inbox cards (TWO, since 2026-07-20). The current week surfaces as two independent rows in the unified Inbox under the Briefings section (the same section Daily Digest rows appear under): a recap card (weekly-summary integration — prose + stats, no suggestions) and a separate suggestions card (weekly-suggestions integration — titled "N suggestions from last week"). The split exists so the AI's suggestions get their own labelled, unmissable row instead of being buried at the bottom of a paragraph-heavy recap. The two are independent: dealing with one never removes the other. Click either row (or press Enter) to open its pane; right-click for the same context menu (Archive, Snooze, Unsnooze), or press S while the row is active to open the SnoozePalette. The suggestions card always reflects the latest week only — it never reaches back to older summaries, so dismissing it leaves nothing older to pop up in its place (fixing the 2026-07-20 backlog "whack-a-mole") — and self-clears when every suggestion has been dismissed (open count → 0). Both cards derive from the SAME weekly-summary row/store — no new store, no migration. Architecture + invariants: weekly-suggestions-card-contract.md.
5. Generate or archive manually. The Run now button at the bottom of Settings → Daily digest → Weekly Summary forces a recap for the most-recently-completed week (useful if you just changed sources, want to re-run after an error, or skipped a Monday). To archive a summary you've finished reading, use middle-click on the row (desktop), right-click → Archive, the E keyboard shortcut while the row is selected, or the small X button shown on mobile widths only. Archived summaries move to a collapsed Archived section at the bottom of the Briefings sidebar; nothing is hard-deleted and a 5-second Undo toast restores the row if you clicked by mistake.
6. Snooze. Right-click → Snooze opens the shared SnoozePalette ("1 hour", "Until tomorrow morning", custom date/time). The row immediately moves to the Snoozed bucket of the Briefings sidebar, vanishes from the Inbox, and decrements the unread badge. When Date.now() > snoozedUntil, the row passively re-buckets back to Active on the next render — no notification, no badge.
7. Suggestion cards. Each suggestion has a short title (e.g. "Archive your idle sessions in project X"), 1–3 sentences explaining the why + how (citing the detector evidence — counts, project name), and optionally a Build it now button (wand icon) when there's a concrete next step the renderer can carry out: create-automation opens the Automation Builder pre-filled with the suggested trigger/conditions/actions; archive-idle-sessions resolves the idle sessions in the flagged project (freshly, at click time), confirms the real count, then bulk-archives them (undoable); explore-feature opens the Settings panel at the named feature. (The old save-quick-reply action was retired 2026-07-06 — repeated phrases now get their own always-on inbox nudge; see quick-reply-nudge.md.) Suggestions never include raw text from your messages — the action payload is restricted to small structured keys. Each card also has a small X in the top-right that opens a "why are you hiding this?" picker — pick a reason (I'm already doing this, Not useful, Doesn't fit my workflow, Wrong / inaccurate, Other) and the card collapses to a one-line dismissed strip with an Undo button. Dismissal reasons feed back into the next weekly generation so the model learns to skip near-duplicates.

Where it lives now (the 2026-05-14 merge)

Weekly Summary used to have its own sidebar entry under WEEKLY SUMMARY with its own Active/Snoozed/Archived buckets, its own inbox section, and its own settings home at Settings → AI Coach → Weekly Summary. On 2026-05-14 we folded all four surfaces into Daily Digest so the two cadences (daily + weekly) share one place to read, triage, and configure:

• Sidebar entry: gone. Use Briefings (the Daily Digest sidebar entry) — weekly summaries appear in the same Active/Snoozed/Archived buckets as daily digests, sorted by date.
• Inbox section: gone. The Briefings inbox section now carries both cadences. Each row's header pill identifies which kind it is.
• Settings home: moved. Was Settings → AI Coach → Weekly Summary; now Settings → Email & Summaries → Daily digest → Weekly Summary (the sub-section at the bottom). The legacy weekly-summary SettingsSection key is preserved as an alias in LEGACY_FEATURE_SECTIONS that redirects directly to email (the Email & Summaries panel where Daily Digest now lives), so older deep links (e.g. from a saved bookmark) still land on the right panel. The redirect goes straight to email — NOT via daily-digest — because LEGACY_FEATURE_SECTIONS is a single-hop lookup, not a chain; a weekly-summary → daily-digest → email indirection would land users on a blank screen.
• Virtual project row: retired. Migration v179 soft-deletes the __weekly_summary__ row in projects — the sidebar entry stops rendering. If you'd customized the entry's pin or color, the row stays in the DB with is_deleted = 1 so it's recoverable on request.

The service, the DB tables, the IPC handlers, and the AI model calls are all unchanged. Only the renderer surface relocated.

How it works

The Monday-morning fire path

The service is /src/main/services/weekly-summary-service.ts. On app boot it calls start(), which:

1. Runs softMigrateFromWorkflowCoachOnce() — gated by weeklyCoachMigrationDone. If your previous setup had workflowCoachEnabled: true, the service flips weeklySummaryEnabled: true and copies your Workflow Coach daily cost cap (workflowCoachDailyCostCapUsd) onto weeklySummaryDailyCostCapUsd. Since default-on shipped it never writes weeklySummaryEnabled: false — a coach that was off just leaves the on-by-default value alone (clobbering it would opt a new user out). The migration only fires once.
2. Runs forceEnableWeeklySummaryOnce() — gated by weeklySummaryDefaultOnMigrationDone. Default-on (2026-07-27) can't reach existing users through the schema default because their config already holds an explicit weeklySummaryEnabled: false, so this flips it true once and marks itself done (a harmless no-op write on a new install). Runs BEFORE the enable-gate so a freshly force-enabled user boots straight into registration + backfill.
3. Gates on weeklySummaryEnabled. Off → returns immediately (no interval, no registration).
4. Registers the service with the service registry (pausable: true) so the global Pause/Resume controls apply.
5. Fires an IIFE that sequences backfillRecentWeeks() → catchUp(). Fire-and-forget; failures are logged but never propagate from start().
6. Sets a 15-minute setInterval running periodicCheck() — which compares the current local time against weeklySummaryScheduleHour and, if the just-completed week's summary hasn't been generated yet, calls generate().

Load-aware deferral (don't gather during a box storm)

The three automatic entry points — periodicCheck() (the 15-min tick), catchUp(), and backfillRecentWeeks() — check the shared whole-box load signal isBoxUnderHeavyLoad() (CPU ≥ 85% or disk hard-fault pressure, from /src/main/process/cpu-load-sampler.ts) before launching the heavy gather, and defer when the box is under heavy load. The gather spawns a short-lived DB host that scans the multi-GB cold archive plus 12 detectors, so during a whole-box storm it would otherwise (a) lose the off-thread reader's 90 s per-RPC deadline ("The database is taking too long to respond") and (b) add archive-paging fuel to the storm. Deferring is a clean skip — the natural retry (the next 15-min tick, or the next boot for catch-up/backfill) picks it up on a calm tick, and backfillRecentWeeks() returns before its completion-flag finally, so a deferred backfill retries next boot instead of being wrongly marked done. This mirrors how holdForCalmDispatchWindow (cron) and waitWhileBoxHeavyForArchiveCopy (backups) defer the same class of heavy scheduled work. The manual Run now button (generate() via IPC) is intentionally exempt — a user-initiated action always runs.

Because a box-load timeout is transient infra (not an AI failure), a gather failure while isBoxUnderHeavyLoad() is true is not counted toward the give-up streak (below). Neither is a credential or billing failure (2026-09-08): a rejected API key, an exhausted monthly AI allowance, no key left to fall back to, or a dry credit balance are things only you can fix, so they must not accumulate into a "the AI pipeline keeps failing" alarm that points nowhere — the shared isCredentialOrBillingFailure predicate excludes them, the same one the daily digest routes on (daily-digest-credential-contract.md I7). This mattered in practice: the streak counter on the owner's install had reached 62 on exactly those errors. So the "Couldn't build your weekly summary" alert fires only for a genuine AI-pipeline failure (provider outage / parse error), never a passing box storm and never a credential problem.

Generating a single week (Map → Reduce → Verify)

generate(weekStartIso?) is the canonical path. With no argument it picks the just-completed week (Mon→Sun, last full ISO week). It short-circuits in four idempotency layers:

1. weeklySummaryEnabled off → null.
2. Service paused (via the Pause/Resume controls) → null.
3. hasWeeklySummaryForWeek(weekStartIso) already true → returns the existing row (no API call).
4. Same weekStartIso already inFlight (in-process Set) → null. Guards against a second tick firing while the first is still mid-generation.

If none short-circuit, the daily cap pre-flight runs: getTodayCostByLabel('weekly-summary') ≥ weeklySummaryDailyCostCapUsd → log + return null. Otherwise the week is added to inFlight and doGenerate(bounds) proceeds:

1. Gather stats. gatherWeekRawStats(weekStartDate, weekEndDate) reads sessions, conversation_messages, feature_events, and api_cost_log to produce a RawStats object — sessions total / active counts, agent vs. operator message counts, top 5 projects by session count, top 8 features by usage count, weekly AI spend in USD, and total tokens (SUM(input_tokens) / SUM(output_tokens) over the same session set as the spend, so cost and tokens always agree). Also gathers the prior week's raw stats so the envelope can carry signed week-over-week deltas (deltaSessions, deltaMessages, deltaAiCostUsd, deltaTokens). The token fields surface on WeeklySummaryStats as optional totalInputTokens / totalOutputTokens / deltaTokens — optional so summaries stored before this shipped still parse and the renderer hides the tile when they are absent (same pattern as the commits tally).
2. No-activity skip. If both sessionsTotal === 0 and messagesAgent === 0, the function returns null without calling any LLM (a "you didn't use the app this week" suggestion would be noise, and the row would still cost API credits).
3. Run the detectors (richer envelope, detectorVersion '4'). runAllDetectors() emits an AggregatedSignals payload of hashes and counts only — never raw operator text. The single allow-listed sampleText carve-out is PII-scrubbed and capped at 200 chars per the workflow-coach hash module. Eleven detector fields ship: the original six (repeatedOutgoingClusters, manualArchive, idleSessions, duplicateApprovals, featureGaps, featureUsage) + activityLandscape (v3) + four new ones the Map → Reduce → Verify pipeline depends on: topOperatorPhrases (30-day phrase-frequency mining), featureDeltas (week-over-week usage trends with direction labels), attentionBacklog (error / needs-you / stalled aging buckets), cronJobStatus (failing-job names with redactPII() applied so the canary token is scrubbed).
4. Pick an account. AI features call Anthropic's Messages API which does NOT accept OAuth tokens. Order: active apikey account → first available apikey account → fail with "No API key account available for AI features (login accounts cannot use the Messages API directly)". CLI session spawning is unaffected.
5. Build the envelope. A single JSON envelope { weekStartIso, weekEndIso, stats, detectorSignals, priorDismissals, featureCatalog } is built once and reused across all three phases (Map sees it, Reduce sees it, Verify sees it). priorDismissals is the last 90 days of suggestion rejections (capped at 100, newest first) projected down to { suggestionTitle, reason, dismissedAt } — title + reason teach the pattern; body is intentionally omitted to keep the envelope small. featureCatalog (added 2026-08-31) is the released-feature menu Reduce may recommend from — see Feature recommendations below.
6. Map (Luna / user-configured utility model, ≤2000 tokens, 45 s timeout). First LLM call. Asks the model to extract 8–15 factual observation bullets from the envelope — no advice, no interpretation. Each bullet must cite the specific envelope field it came from in parentheses (e.g. "(topOperatorPhrases[0])"), quote numbers verbatim, and skip empty/null signals. Output is plain text, not JSON; the parser strips "- " markers and caps at 20 bullets. This step exists so Reduce has a pre-digested factual checklist instead of having to re-read the whole envelope.
7. Reduce (Opus 4.7, ≤4000 tokens, 120 s timeout). Second LLM call. Receives the envelope plus the Map bullets and produces the final JSON { summaryMd, suggestions[] }. Each suggestion MUST include a groundednessCheck: { signalsUsed[], claimsInBody[] } self-citation — signalsUsed lists dot-paths into the envelope the suggestion depends on; claimsInBody breaks every numeric/named claim out as a separate string for Verify to fact-check. The Reduce system prompt also folds in the dismissal-feedback rules: if priorDismissals shows a near-duplicate suggestion was rejected as already-doing / not-useful / not-relevant / wrong, Reduce either skips it or reframes with materially-stronger NEW evidence (full reason-code semantics are spelled out in the prompt). Code fences are stripped from the response before JSON.parse. On parse failure the row still lands with summaryMd = rawText, suggestions = [] (logged warn) so a malformed Opus response never silently swallows a week.
8. Verify (Luna / user-configured utility model, ≤1500 tokens, 45 s timeout). Third LLM call. Receives the envelope + the Reduce JSON and fact-checks each suggestion's claimsInBody against the envelope. Returns { ok, issues: [{ suggestionIndex, reason }] }. Defense-in-depth, never a hard blocker: a parse failure / shape mismatch is treated as { ok: true, issues: [] }. applyVerifyIssues() drops any suggestion whose index appears in issues[]; the summaryMd is unchanged (Verify only judges suggestion-level fabrications). Flagged drops are logged with kept N/M so you can see Verify intervention in main.log.
9. Track cost (per phase). Every messages.create is followed by trackApiCost(accountId, 'weekly-summary', model, input_tokens, output_tokens) so the dashboard total matches the Anthropic console. The label 'weekly-summary' is registered in OTHER_AI_SOURCE_LABELS so it appears in the Cost tab filters. A per-phase ModelChainEntry { phase, model, inputTokens, outputTokens, costUsd } is also accumulated for the row's modelChain column (lets you split a single week's spend into its Map / Reduce / Verify components in postmortems).
10. Insert. insertWeeklySummary({ weekStartIso, weekEndIso, stats, summaryMd, suggestions, detectorSignals, inputTokens, outputTokens, costUsd, modelChain }) writes the row. week_start_iso is UNIQUE so a parallel insert (e.g. backfill + periodicCheck racing on the same week) is caught at the DB level — the second insert fails, but the first row stands. Row-level inputTokens / outputTokens / costUsd are the sum of the three modelChain entries.
11. Emit push. emitPush(IPC.WEEKLY_SUMMARY_GENERATED, { summary: stripFullPayload(inserted) }) (lean list-entry only, full stats/suggestions/detectorSignals fetched on demand via WEEKLY_SUMMARY_GET) and IPC.WEEKLY_SUMMARY_UNREAD_CHANGED. Renderer's useWeeklySummaryPushEvents hook updates the Briefings sidebar / inbox / unread badge live — both daily and weekly use the same merged unread count.

Why three steps (vs. one Haiku call)

Up through the Workflow Coach era the suggestions came from a single Haiku call per detector type — eight short-prompt calls that often produced surface-level "save a quick reply" / "make an automation" output regardless of whether the underlying signal was meaningful. P5 retired those 8 per-kind prompts in favor of the single Map → Reduce → Verify pipeline because:

• Opus 4.7's writing is materially better at synthesizing across detector signals (citing the project name and exact counts in the body, not just the title). Per-suggestion cost is higher but call count is fixed at 3 per week, so the line item stays under $0.10/day even on backfill.
• Cross-signal correlation needs the whole envelope at once. "You sent 'please continue' 23 times AND your top project is X" is a better suggestion than either phrase frequency or top-projects alone could produce. The per-kind shape couldn't see across signals.
• Verify catches fabrication that single-pass prompts can't self-detect. When Opus invents a project name or rounds a count, the Verify pass — looking at the envelope and the JSON side-by-side — flags it. The flagged suggestion is dropped; the rest of the week ships. Without Verify, a single fabricated number in week N would erode trust in week N+1's whole list.

Dismissal memory + the why-picker

Each suggestion card has a small X in the top-right that opens a five-button "why are you hiding this?" picker ('already-doing' | 'not-useful' | 'not-relevant' | 'wrong' | 'other'). The chosen reason is recorded in weekly_summary_dismissals (one row per (weekly_summary_id, suggestion_index) — composite UNIQUE so re-picking last-write-wins). The card collapses to a one-line "Dismissed: <title> — Reason: <label>" strip with an Undo button that hard-deletes the dismissal row (no soft delete). Dismissals are populated on WEEKLY_SUMMARY_GET via the dismissals?: WeeklySummaryDismissal[] slice on WeeklySummary.

The same store of dismissals is read back by buildPriorDismissals() (last 90 days, cap 100, newest first) and folded into every future Reduce call's envelope as priorDismissals: Array<{ suggestionTitle, reason, dismissedAt }>. The Reduce system prompt's "learn from past rejections" block then steers Opus around near-duplicate topics — reason codes carry stable meanings (see the type comment in shared/types.ts):

• already-doing → skip near-duplicates (user already has it covered)
• not-useful → avoid the same category
• not-relevant → only re-raise if THIS week's evidence is materially stronger
• wrong → treat related topics with extra rigor on groundednessCheck
• other → treat as a soft not-useful

If the dismissal read fails for any reason (getRecentWeeklySummaryDismissals throws), the envelope still ships with priorDismissals: [] — dismissal memory degrades gracefully, never blocks a generation.

IPC: IPC.WEEKLY_SUMMARY_DISMISS_SUGGESTION { weeklySummaryId, suggestionIndex, suggestionTitle, suggestionBody, reason } and IPC.WEEKLY_SUMMARY_UNDISMISS_SUGGESTION { weeklySummaryId, suggestionIndex }. Both push WEEKLY_SUMMARY_UPDATED so other open viewers refresh.

For agents

Settings field map

Setting key	Default	Range	Notes
weeklySummaryEnabled	true	bool	Master toggle — on by default (owner decision 2026-07-27)
weeklySummaryScheduleHour	7	0–23	Local-time hour on Monday
weeklySummaryDailyCostCapUsd	0.1	0.05–5.00	Cumulative cap per UTC day
weeklySummaryBackfillCompleted	false	bool	Internal — flips after first backfill attempt; NOT in renderer Zod schema
weeklyCoachMigrationDone	false	bool	Internal — flips after one-time Workflow Coach migration
weeklySummaryDefaultOnMigrationDone	false	bool	Internal — flips after the one-time force-enable of existing users when default-on shipped; NOT in renderer Zod schema
weeklySummaryModel	Luna (Automatic)	model id	Per-feature utility model override — Map + Verify phases only; Luna sentinel = inherit global / auto-DeepSeek / Luna. See utility-ai-model.md.

Related

This page continues in Weekly Summary (part 2), which carries the rest of the behaviour — what the recap may recommend, the buttons a suggestion can carry, the first-launch backfill and its freshness gate, the nominal AI usage figure, the phrase exclusions, the commit tally, the cost cap and the Workflow Coach retirement — plus the stored tables, settings keys and renderer surfaces behind all of it.

• daily-digest.md — daily sibling. Same generation shape, same inbox card, same sidebar entry (Briefings) as of 2026-05-14, same Active/Snoozed/Archived buckets, same triage hotkeys.
• workflow-coach.md — retired; redirects here.
• ai-providers.md — Weekly Summary Map + Verify run on Luna (the cheap/utility model) by default, overridable via weeklySummaryModel; Reduce runs on Opus 4.7. Falls back through apikey accounts when active account is OAuth.
• inbox-overview.md — Weekly Summary row appears under Briefings (the merged Daily Digest section); J/K nav and the 6-slice mutex.