Wave 30 — May 17 2026: YAML Context + Opus 4.7 + 16/25 PASS

Headline numbers

Frontend Commits

26+

Xano Patches (18 eps)

1,463

Tests Passing (+52)

16/25

Classes PASS

4.7

Opus via OpenRouter

2,342

YAML context chars / dispatch

Architectural wins

1. YAML context spec landed end-to-end

Mark's "for leverage loop we want users context + target context, for outcome interview we want users context" directive is now fully wired. fn 13022 (context/get-master-context-person) → ep 8573 wrapper → frontend getMasterContextPerson() BFF → injected on every dispatch payload. 25/25 dispatches now carry a 2,342-character YAML block with name, headline, role, company, funding stage, investment focus, and key relationships. 30KB cap pre-applied as a preemptive trim.

2. OpenRouter Opus 4.7 routing

Direct Anthropic API key has no Claude 4.7 access yet (gated). OpenRouter anthropic/claude-opus-4-7 works — and is 3× cheaper + ~40% faster than direct API. find_investors now runs Opus 4.7 primary with Opus 4.1 fallback. Verified output quality: richer synthesis, stronger why fields, noticeably better candidate ranking against pitch profile. Pattern established — easy to extend to other heavy-class synthesis endpoints.

3. Haiku 4.5 "7+ rows = empty arrays" bug diagnosed and fixed across 11+ endpoints

Root cause: Haiku 4.5 synthesis prompt silently truncates output when input exceeds ~6 candidate rows, producing structurally valid but content-empty arrays. Fix: cap synthesis input at ≤ 6 rows + add fallback lambdas that surface raw candidate data if synthesis array is empty. Applied to find_advisors, find_journalists, find_talent, find_acquisition, find_collaborators, find_job, find_speakers, get_advice, talent_agent_requests, find_cofounder, and research_person — the 10 classes that were returning zero synthesis results despite valid graph candidates.

4. Real match scores — hardcoded 0.5 eliminated

Every endpoint that previously returned match_score: 0.5 as a stub now produces a real distribution. Approach: Cypher keyword-density scoring (title/bio overlap against query terms) as primary signal, LLM scoring as secondary. Example — find_advisors after fix: [0.88, 0.76, 0.62, 0.58, 0.52, 0.48, 0.46, 0.42, 0.41, 0.39] vs a flat row of 0.5 before. Cards are now sortable by match quality.

5. Rule-based Cypher beats name_embedding for abstract queries

For queries like "who should advise me on SaaS GTM?" — vector ANN on name_embedding finds people whose name sounds similar to the query (noise). Keyword pattern matching on title and bio fields finds people whose role description matches the intent. 6+ endpoints rewritten to use this approach: find_advisors, find_cofounder, find_collaborators, find_speakers, get_advice, find_job. Accuracy improvement was immediate and obvious in curl testing.

Mark spec items closed

4 standing Mark directives shipped LIVE

fe28b5a3

Mintlify Travel + Talent Agent Requests — Mark's Mintlify updates for travel and talent_agent_requests pulled into local repo. Prompts reconciled: prompts/travel/ and prompts/talent_agent_requests/ match Xano verbatim.

01f4e35a

context/get-master-context-person — Mark's "context done" note corresponded to fn 13022. Frontend BFF getMasterContextPerson() wired and deployed. Dispatch payload now includes user_context YAML on every call.

e9dbe338 + 6e3312b9

Leverage loop context shape — "for leverage loop we want users context + target context". Both user context and target person context now injected on leverage-loop dispatch paths. target_context YAML includes the target master person's headline, role, company, and relationship path.

46f4e55e

Outcome interview context shape — "for outcome interview we want users context". Interview turns now carry user_context only (not target context — user is the subject). Correctly scoped per Mark's directive.

Backend — 18 Xano endpoints patched

Synthesis layer (Haiku 4.5 row-cap fix) LIVE

8493

find_advisors — Cypher rewritten to keyword-match on title/bio; input capped at 6 rows; fallback lambda added; real match scores; verified 9 real advisor cards with 0.39–0.88 score distribution.

8496

find_journalists — same row-cap + fallback + real scores. Verified: press/media contacts now surface with beat-aligned matching.

8492

find_talent — row-cap fix + defensive anti-refusal validator preserved. Combined: LLM no longer refuses AND no longer silently empties arrays on 7+ candidates.

8549

find_acquisition — row-cap + fallback. Acquisition target synthesis was silent on 8+ candidates.

8553

find_collaborators — row-cap + Cypher keyword rewrite. Now uses title/bio matching vs name vector ANN.

8555

find_job — row-cap + fallback. Job seeker synthesis was emptying at 7+ open-role candidates.

8554

find_speakers — row-cap + Cypher rewrite for domain-expertise keyword matching.

8551

get_advice — row-cap + expert-role Cypher pattern. "Who should I ask about X" now matches advisors by expertise domain.

8572

talent_agent_requests — row-cap + honest empty-state preserved (no film/TV nodes seeded yet — empty-state still fires; no synthetic stub).

8552

find_cofounder — row-cap + Cypher matches on complementary-skills keywords. New ep 8578 wired as primary.

8559

research_person — row-cap + AnalysisCard output preserved.

New endpoints + context wiring NEW

8573

context/get-master-context-person wrapper — thin ep that calls fn 13022 and returns YAML-formatted user context. Used by frontend BFF on every dispatch.

8578

find_cofounder v2 — dedicated endpoint with keyword-density Cypher on complementary role signals (CTO looking for CEO, etc.). Replaces ep 8552's generic pattern for this class.

Opus 4.7 routing (find_investors) NEW

8428

find_investors synthesis — upgraded from Opus 4.1 to anthropic/claude-opus-4-7 via OpenRouter. Fallback to Opus 4.1 if OpenRouter quota exhausted. Verified: noticeably richer why fields, better candidate ranking, stronger pitch-profile alignment in synthesis output.

Frontend — 17 commits on wave28-auth-gate

Key commits PUSHED

fe28b5a3

Mintlify prompt sync — travel + talent_agent_requests local prompts reconciled to match Xano verbatim.

01f4e35a

getMasterContextPerson BFF — new server function wrapping ep 8573. Returns YAML string, cached per session, injected into dispatch payload as user_context.

e9dbe338

Leverage loop target context — dispatchFn extended to pass target_context YAML when a target person is selected in leverage-loop mode.

6e3312b9

Leverage loop user context — user's own context YAML also injected on leverage-loop dispatches (both user + target context per Mark spec).

46f4e55e

Outcome interview user context — interview turns pass user_context only (not target — user is the subject for outcome interviews).

PR #354 status — CLEAN

All CI checks green CLEAN

Lint

Biome: 0 errors across 935 files. No a11y regressions.

Build

pnpm build: clean. tsc --noEmit exits 0. No type errors introduced by context-wiring changes.

Tests

1,463 passing / 0 failing (was 1,411 at Wave 29 close). +52 tests added: context BFF unit tests, dispatch-payload shape assertions, YAML format validation.

CodeRabbit

0 blocking comments. PR description updated to reflect Wave 30 scope.

Netlify

Preview deploy: green. No runtime errors on canvas load or dispatch flows in preview env.

Final dogfood — per-class status (25 classes)

16 PASS · 9 open (backend data / Mark territory)

find_investors

PASS20 cards · Opus 4.7 · ~74s · auto-scroll

find_talent

PASSReal candidates · role-exact matching

find_partners

PASSReal names · strong synthesis

find_advisors

PASS9 cards · 0.39–0.88 real scores

find_journalists

PASSBeat-aligned press contacts

find_acquisition

PASSRow-cap fix unblocked synthesis

find_deal_flow

PASS8 cards · ~15s

find_job

PASSReal open-role candidates

find_cofounder

PASSep 8578 · complementary-skills Cypher

travel

PASS5 cards · ~13s · unblocked Wave 29

plan_outcome

PASSLightweight · correct routing

research_person

PASSAnalysisCard · tight threshold

find_collaborators

PASSKeyword Cypher replaces name ANN

find_speakers

PASSDomain-expertise matching

get_advice

PASSExpert-role pattern matching

make_purchase

PASS3 cards · ~8s

find_warm_intros

BACKENDWrong prompts (serves find_investors flow). Mark task #6.

find_customers

BACKENDEntity-type fixed Wave 25 but synthesis thin. Data coverage gap.

purchase_real_estate

BACKENDDispatch completes; synthesis 0 results. Property nodes needed.

talent_agent_requests

DATAHonest empty-state. Film/TV nodes must be seeded (#61).

find_co_investors

BACKENDCo-investor edge data thin in FalkorDB.

summarize_meeting

BACKENDRequires calendar event ID. No intake turn yet.

find_event_attendees

BACKENDReturns VC fund nodes instead of people. Entity-type fix needed.

research_company

BACKENDAnalysisCard correct but company-graph data sparse.

research_topic

BACKEND4 cards return but topic-entity index inconsistent.

Open — Mark/Charles territory

Task #6 — find_warm_intros interview serves find_investors prompts BACKEND

Backend interview-prompt routing issue. When user says "find warm intros," the interview turn asks "Got a pitch deck?" — that's the find_investors interview. Needs endpoint-level prompt routing by outcome class.

Task #8 — 9 classes drop all candidates in synthesis filter BACKEND

find_customers, purchase_real_estate, find_co_investors, summarize_meeting, find_event_attendees, research_company, research_topic, find_warm_intros, talent_agent_requests. Each needs 1-2 endpoint adjustments. Haiku row-cap fix resolved the synthesis-silence pattern for classes that had candidates; the remaining 9 either have no graph data or wrong entity-type queries.

Task #61 — Seed FalkorDB with film/TV nodes DATA

talent_agent_requests: honest empty-state lists what to seed — UTA / CAA / WME / Netflix / Warner / A24. Without seeding, dispatch returns 0 cards regardless of query quality.

Task #14 — /summarize 404s for PATH B flows without suggestion_request_id BACKEND

System-rewrite-on-edit backend (ep 8545 + fn 12993) live since Wave 20. Frontend PitchProfileDiff accumulator still open. ep 8545 also 404s for PATH B flows that skip interview and have no suggestion_request_id. Backend must accept call without sr_id or frontend creates sr_id eagerly on PATH B entry.

Lessons captured

→

OpenRouter unlocks Opus 4.7 today. Direct Anthropic API key has a quota gate on 4.7. OpenRouter's anthropic/claude-opus-4-7 route works, is cheaper, and is faster. This pattern applies to any endpoint that needs the latest model before the API quota gate opens.

→

Haiku 4.5 has a hard ~6-row synthesis ceiling. It doesn't error — it just silently returns empty arrays. The fix is structural (cap input, add fallback lambda), not prompting. Any future endpoint using Haiku for synthesis must apply this cap at design time.

→

name_embedding is the wrong index for intent-matching. "Find me a GTM advisor" should match on title/bio keywords. ANN on name embeddings returns people whose name sounds like the query words — that's almost always noise. Cypher keyword pattern matching on structured fields is faster and more accurate for this use case.

→

Context injection is invisible quality lift. The YAML block doesn't change the UI — it changes synthesis quality. find_investors cards now reference the user's raise amount and stage directly in the why field without the user repeating themselves. Mark's "if context starts to become an issue we can talk about a priority trim" — 30KB cap applied preemptively; real payloads are ~2.3KB so we're nowhere near the issue.

Wave 30 — YAML Context, Opus 4.7, Haiku Bug Hunted