One repo, one branch, one merge
The push-to-100% loop ran on a 20-minute cron 89bb50f9 — every wake dispatched 3-5 parallel subagents (Sonnet 4.6 default, Opus 4.7 for cross-file architecture). Each closure dual-pushed to roboulos/orbiter-sandbox AND M-Pederson/orbiter-sandbox. The frontend branch lives only on M-Pederson/orbiter-frontend.
The full 5-stage pipeline, live
Mark's May 9 architecture lock: "we're not touching the graph until the end. We use these little profile things to match-make vector scores to get from 8000 to 40 or 20. Then we start pulling graph and everything else." Every Anything Engine class follows this shape. find_investors is the reference; the other 14 inherit.
Stage 1 Interview (8420 ready). ≥5 of 6 narratives + raise_amount + company_name. Mark's hard gate.
Stage 2 Pre-filter (8426). Vector-only. Single-dim cosine ScaNN-shaped scoring against fundraising_pitch_profiles (table 710). NO graph touch. Honors excluded_uuids.
Stage 3 Headroom. Over-fetch by ~2x. Tunable per class via stress harness.
Stage 4 Graph RAG (8401). FalkorDB Cypher constrained to candidate_ids. Pulls full context only on the small filtered set.
Stage 5 Final-trim (8428). Opus 4.1 reasoning pass. SIGNAL TAXONOMY tags [portfolio] / [co-investor] / [thesis], FACTS GUARDRAIL on dataset, source attribution, freshness markers.
tool_calls: [{tool: "update_pitch_profile", params: {...}}] → 8399 applies edit at head of stack and re-embeds → real cards land with WHY paragraphs. Mark uses it. Mark feels good. Charles green-buttons.find_talent live, Mark's prompt verbatim
Mark dropped the system prompt May 6 in Slack at 12:57 — eight adaptive MCQ dimensions (archetype / wedge / scope / must-have / signal / style / comp / sourcing), per-turn JSON schema with options + freetext escape, opening move always picks archetype first. find_talent v0 deliberately runs WITHOUT the 6-dim vector pre-filter — it goes straight from interview to graph_rag with role/skills/location filters. Persona-matching V2 ships with a separate patent-pending pipeline.
Two live, thirteen scaffolded
Mark's verbatim May 6 Slack list. Six classes survived from our pre-May-6 14 (find_investors, find_talent, find_customers, find_partners, find_advisors, find_journalists). Eight deprecated (research_*, find_co_investors, find_event_attendees, find_warm_intros, summarize_meeting, plan_outcome). Nine new scaffolded May 7 with Mark's find_talent shape: deal_flow, make_purchase, purchase_real_estate, find_cofounder, find_collaborators, find_acquisition, find_speakers, find_job, get_advice.
Two of the new classes are two-sided — find_acquisition handles both acquirer-hunting-targets and target-seeking-acquirers via a mode sub-question on turn 1. find_speakers handles both event hosts seeking speakers and speakers seeking gigs. Single endpoint, branched flow.
Scaffolds shipped, prompts pending Mark
Each new class has a prompts/{slug}/interview.md skeleton modeled on Mark's verbatim May 6 find_talent shape — adaptive MCQ interview, per-turn JSON schema, opening-move archetype question, hard rules importing from prompts/_shared/voice-rules.md. The summary-format section is a [PLACEHOLDER] Mark fills with 1-2 example summaries from each class's Mintlify ontology page. Nothing went live without his content.
The cron-driven push-to-100% loop
Cron 89bb50f9 fires every 20 minutes off-minute (:07/:27/:47). On wake the orchestrator reads backlog.md, identifies the next 3-5 ready tasks (deps closed per the intersection graph in snappy-leaping-cake.md), and dispatches parallel subagents using the brief library at docs/subagent-briefs.md. Sonnet 4.6 default; Opus 4.7 reserved for chat-shell extraction (A3), 8399 toolkit reader (C4), E2E harness build (D1), and Mark-corpus + theme analysis (G1, G2, G3).
One slot is reserved every wake for Track F1 — Krisp intel sweep (pulls new Mark+Robert meetings since last sweep, surfaces verbatim quotes to docs/mark-directives-feed.md, prepends new directives to the backlog). One slot every 3rd wake is reserved for Track G3 — connect-the-dots analysis (cross-references new directives against the Mark theme corpus, surfaces drift between code and stated principles). Documentation as a deliverable: clients value it, Snappy positioning depends on it being verifiable.
Hard rules every wake honors: image generation ONLY gemini-3.1-flash-image-preview (Nano Banana 2 — Robert-locked). Real screenshots at 1440×810 to DigitalOcean Spaces. Sandbox-clone every Xano live change first; save prior XanoScript to /tmp/<id>-pre-may-7.xs; post Slack pulse with the exact revert command. Pathspec-only git add. Dual-push every commit. "Reversibility beats sign-off gates."
Henry feedback, persona matching, the 13 remaining classes
Henry feedback application (C5). Mark forwards detailed Henry Peck feedback items, F1 Krisp sweep auto-unblocks the track, fixes land per-item with verification artifacts in docs/anything-engine/henry-feedback-may-N.md.
Persona matching (G2 / Mark patent). 8 narrative dimensions per contact, auto-profiled and refreshed on cadence. find_talent V2 + leverage_loop both inherit this shape. Mark filing on the synthetic-personas matchmaking process. AlloyDB cluster + Vertex AI embeddings (Kenya delivery) ship under the same window.
13 remaining classes E2E. Each gets the same 5-stage shape with class-specific Cypher, SIGNAL TAXONOMY tags, voice rules from prompts/_shared/voice-rules.md. Mark's hands-on protocol: he edits the prompts directly via Claude CLI; the cron synchronizes per-class prompt updates to Xano via pnpm sync-prompts.