Options Flow, and the silent 401

Phase 8 closed today — three PRs, Options Flow shipped on Lookup as a per-ticker block and on the Smart Money tab as a cross-symbol feed. The smart-money discovery surface is now feature-complete: insiders, congress, billionaires, trailblazers, and options flow all live as both per-ticker overlays and cross-symbol feeds.

The most interesting part wasn't shipping the new code. It was discovering, partway through 8a, that the legacy flowSignal had been silently abstaining across the entire universe for months.

Phase 8a — the Lookup block, and a 401

The first slice was a per-ticker Unusual Options Activity block on the Lookup Details rail. Pulled from Yahoo's free options chain endpoint, filtered with a simple heuristic (volume / open-interest ≥ 1.5 AND volume ≥ 50), top ten contracts by premium (volume × mid price × 100). Mid price comes from lastPrice with bid/ask midpoint as fallback. New shared types, new IPC channel, new DetailOptionsFlowBlock renders only when there's data.

Built it. Tested NVDA. Block didn't render.

The Lookup spinner sat there for the usual moment, then nothing. No error in the renderer console. No error in the main-process log. Just an empty space where the unusual activity should have been. Probed Yahoo's /v7/finance/options/NVDA endpoint directly with curl. 401.

Yahoo's options chain endpoint started requiring the crumb dance — the same cookie + crumb token flow we already use for the analyst-targets and fund-profile endpoints — sometime in 2025. They didn't announce it. They just started returning 401 without it.

The silent abstain

Here's the part that had been hiding.

The existing fetchOptionsSnapshot — the function that's powered the legacy flow signal layer since well before this session — hit the same Yahoo endpoint. It got the same 401. Its catch-all error path returned null instead of throwing.

The downstream flowSignal sees null as "no data" and abstains. Abstain semantics mean it drops out of the weighted average rather than pulling toward neutral 50 — which is exactly the right behavior, and exactly why we couldn't see the bug from the score side.

The flow layer also has weight 0 in the smart-money lenses (the smart-money pivot moved options weight to zero pending Phase 8). So no score moved when flow abstained. No alerts. No regression. Just a layer that was supposed to be alive (just not heavily weighted) was actually completely inert across every ticker in the universe, and had been for some indeterminate number of weeks.

The fix: extract the crumb logic into a shared ensureYahooCrumb helper, wire it into both the new per-contract path (8a) and the legacy aggregate path (the original fetchOptionsSnapshot), with a 401 retry that re-acquires the crumb before re-attempting. NVDA after the fix: 27 contracts meeting the unusual filter, displayed in the new block. The legacy aggregate path also lit back up — the volume / put-call / IV reads it had been silently producing as null are now real numbers again.

The lesson worth keeping: silent abstain is a feature when data is genuinely missing, but a bug when the abstain is masking infrastructure failure. The fix isn't to stop abstaining. The fix is to log loudly when an abstain comes from "endpoint returned an error" rather than from "endpoint returned legitimately empty data." We don't have that distinction in the pipeline yet. Logged it as a follow-up; the pattern is general enough to matter beyond this one signal.

Phase 8b — wiring it into the score

With both paths alive, 8b extended flowSignal to read per-contract unusualContracts from OptionsSnapshot. The score map is intentionally simple:

• Total unusual premium below $100k = noise. No adjustment.
• Above $100k, skew triggers at ≥ 65% calls (bullish) or ≤ 35% calls (bearish).
• Magnitude is tiered: $100k–$500k = ±4, $500k–$2M = ±8, ≥ $2M = ±12.
• The aggregate put/call ratio and IV reads are unchanged; unusual-flow contribution is additive.

SMART_MONEY_CLUSTER weight nudged from 0.0 to 0.3. Modest, deliberately. The Yahoo data is a snapshot, not a stream — we're seeing the back-half of the trading day, not real-time order flow — so weighting heavier than 0.3 would over-credit a stale signal.

Spot-checked NVDA. Real flow chip score, not the silent 50 it had been quietly producing pre-fix. Luke confirmed: "looks good."

A docs-sync rule banked while we were here

One feedback memory got banked during 8b: any time scoring is adjusted — signal scorers, preset weights, zone definitions — docs/scoring.md and the README "Nova Score — at a glance" section update in the same PR.

The docs we shipped in Phase 7.7z (~490 lines explaining all twelve layers) are only useful if they stay current. Drift between code and docs is a one-way ratchet; once it starts, nobody trusts the docs and the docs stop getting updated, which guarantees more drift. Cheaper to keep them in lockstep at PR time. Applied in 8b: the flow-layer section got rewritten, the SMART_MONEY_CLUSTER weight table updated, the README flow row + roadmap refreshed.

Phase 8c — the cross-symbol feed (and a miss)

The third slice put Options Flow on the Smart Money tab as a cross-symbol discovery feed. Hardcoded OPTIONS_FLOW_UNIVERSE of about 50 options-active tickers (index ETFs, mega-cap tech, retail-favorite high-beta, large-cap financials). Cross-symbol fetcher fan-outs across the universe, flattens unusualContracts, returns top 200 by premium. First load is 12–15 seconds cold; subsequent opens within the 30-minute cache window are instant. Sub-tab on the Smart Money panel, deep-link to Lookup on row click.

I shipped the first cut without sortable column headers.

Luke caught it on spot-check: "i cant even click them. like i cant click volume or type etc." Fair catch. The feedback memory at feedback_modern_table_conventions.md — banked back during the Smart Money phase — says modern table conventions are baseline: sortable headers, filter row, click-to-deep-link, count badge. The rule existed. I missed it on first cut.

Added a SortableTh helper with caret indicator + sort state in the same PR. The honest lesson, banked on top of the existing memory: when shipping a new data table, run the modern-conventions checklist as a default, not as something to remember to do. The cost of remembering each time is exactly the cost of forgetting once and getting caught on spot-check.

Where this leaves Nova

The smart-money discovery surface is feature-complete. Five feeds — Insiders, Congress, Billionaires, Trailblazers, Options Flow — live as both per-ticker overlays on Lookup and cross-symbol discovery feeds on Smart Money. The score reads off all twelve layers, abstain semantics throughout, weights settled.

Three things are queued next.

The first is the Phase 9 Nova Fair Value gauge — a semicircle visualization on the Lookup Details rail blending analyst targets, P/E vs five-year median, PEG, FCF yield, and a reverse DCF that solves for the FCF growth rate the current price implies. Phase 9a + 9b shipped behind a separate stack today; deliberately not wired into Nova Score yet, because the DCF assumptions still need a live spot-check pass.

The second is a consistency audit Luke flagged at the end of the session. Lookup uses explainSymbol(), which always fetches all twelve layers. Scanner and AI Picks use runScan(), which only fetches a layer when its preset weight is greater than zero. The layers added during the 7.7 arc — track_record, support, fund_quality — never fire on Scanner cards or AI Picks rankings. Same ticker shows different Nova Scores on Lookup vs Scanner. Real bug. Sequenced after Phases 10–12 of the gekko plan so it lands as a single sweep rather than chasing each layer individually.

The third is the rest of the gekko plan — macro tab, broader market context, the polymarket integration. Each will get the same treatment: ship the slice, audit downstream surfaces, update the docs in the same PR.