cre-acquisition-orchestrator Agent工作流

CRE Acquisition Orchestrator

An open-source, multi-orchestrator workspace for commercial real estate multifamily acquisitions: drop documents, state the goal, watch 31 AI roles coordinate, and review the acquisition package.

Fastest proof path: run npm run proof, open the local dashboard, and trace one source-backed fact from upload to IC package. Full reviewer script: Public Proof Path.

I've been working on something that I think the CRE industry needs, and I wanted to share where it is now.

A few months ago I wrote about what happens when you point 489 AI agents at a 200-unit multifamily acquisition. That article was the bigger vision. This repo is the engineering behind the practical open-source version: the 31 named AI roles, orchestration logic, domain knowledge files, schemas, local dashboard, deterministic simulation engine, and source-backed review workflow that make the vision usable.

It is not fully production-ready. I want to be direct about that. But what is here is the most in-depth open-source framework I have seen for CRE acquisition orchestration because the category barely exists. There are agent frameworks for coding, customer support, research, and data analysis. There is almost nothing that models how a real multifamily acquisition moves across due diligence, underwriting, financing, legal, and closing while preserving data handoffs, review gates, and investment committee evidence.

The project is local-first: you can run the proof path with no API keys, inspect uploaded tables and source rows, review extracted candidate fields with provenance, approve or waive ambiguous values, and export Markdown/JSON for an investment committee starter package. The dashboard's workflow runtime defaults to live ChatGPT/Codex (with web search on) so the team can pull and cite real market, lender, and environmental data, while the deterministic offline demo stays the no-credential public proof path for tours, screenshots, and CI.

Everything in here - the agent prompts, domain skills, schemas, pipeline architecture, dashboard, and demo artifacts - is yours to use as a starting point. Fork it. Build on it. Adapt it to your own deals, investment thesis, and internal acquisition workflow. If this framework helps even one CRE team rethink how they approach acquisitions, it was worth open-sourcing.

Let's bring this industry into the future.

Disclaimer: This project is a reference architecture and educational framework, not production software for making investment decisions. Nothing here is financial, legal, or investment advice.

---

What's New in v3.4.0

• Pipeline verification ledger - document intake, source review, due diligence, underwriting, financing, legal, closing, IC package export, offline gates, and live Codex gates now have recorded proof in data/status/pipeline-verification-ledger.md.
• Live Codex proof gates - npm run codex:status, npm run codex:smoke, npm run codex:run:full, npm run validate:codex, and npm run eval:live were run and recorded; the full live workflow completed with 21/21 agents passing on first attempt.
• All-8-deal live eval refresh - live Codex agents matched all 8 benchmark IC verdicts exactly and directionally, with 100% determinable financial accuracy, 100% required red-flag recall, 100% dealbreaker recall, and 0 partial failures.
• Phase artifact hardening - underwriting writes/validates the 27-scenario matrix and IC memo; closing writes/validates the wire schedule; IC package export includes a first-class document manifest and review decision trail.
• Live manifest schema hardening - current Codex manifests validate root agentTimeoutMs and per-result timedOut, so smoke/full live runs are covered by the same contract gate.

What's New in v3.3.0

• Codex is the main workflow runtime - the dashboard launches the selected workflow on live Codex / ChatGPT by default. The Workflow Launcher, Swarm Goal Console, and saved presets default to Codex (listed first, all agents selected, concurrency 2), with the deterministic Simulation runtime kept as the no-credential fallback for demos, screenshots, and CI.
• Agents actually use web search - when Codex web search is on, agents are directed to look up and cite real rent/sales comps, submarket rents, occupancy, cap rates, demographics, supply pipeline, and current interest/lender rates. Web search is on by default with a visible toggle, and the Swarm launch and retry-failed-agents paths keep it on.
• Lean legal-document parsing - PSA, title commitment, and estoppel documents parse into review-gated candidate fields with provenance, committed fixtures, and tests.
• Intake/extraction/launch UX fixes + e2e/CI stabilization - the six intake/extraction/launch bug fixes (T12 expense magnitude, source-reconciliation equality, blocked-launch missing-field surfacing, Edit Deal step pills, sample-deal schema alignment, scoped-workflow workpaper index) land alongside dashboard overlay/modal e2e and CI stabilization.

What's New in v3.2.0

• Production-scale local QA harness - npm run seed:prod-local -- --count 150 creates a sanitized QA-LOCAL-2026-* 150-deal corpus with source documents, extraction artifacts, approved fields, criteria, phase state, checkpoint status, and completed-report artifacts under local data/.
• Production local data regression gate - npm run test:prod-local-data validates schemas, source-hash provenance, local-only output boundaries, idempotent reseeding, and generated-artifact sensitive-token avoidance.
• Public proof command - npm run proof regenerates Parkview, starts the dashboard, waits for readiness, and points reviewers to docs/PROOF-PATH.md to trace one source-backed fact from upload to IC package.
• Full QA documentation - docs/QA-INVENTORY.md documents routes, roles, modals, buttons, inputs, workflows, and acceptance criteria; docs/QA-BUG-LOG.md records each production-scale QA defect with reproduction evidence, fix, and verification, with 30 Playwright browser tests passing.

What's New in v3.1.0

• Local scanned-PDF OCR bridge - readable scanned/image-only PDFs render locally with PyMuPDF and run through tesseract.js, with no external OCR service.
• Review-gated OCR candidates - OCR-derived asking price, unit count, occupancy, and NOI become candidate fields with confidence, source hash, page provenance, raw snippets, parser metadata, and human review status.
• Fail-soft scan handling - unreadable scans or scans without supported headline fields return explicit OCR metadata and warnings instead of guessed values.
• Fresh-clone OCR setup - npm run setup installs and verifies PyMuPDF; npm tracks tesseract.js.
• OCR fixture proof - fixtures/parsers/scanned-offering-memo-ocr.pdf verifies a true image-only offering memo can extract through the local bridge.

What's New in v3.0.0

• Evidence-grade source-to-IC chain - IC package JSON now includes a deterministic evidence graph connecting source documents, approved fields, agent workpapers, red flags, data gaps, and package sections. Markdown export adds an Evidence Chain section.
• Fresh-clone parser setup - npm run setup creates .venv, installs parser dependencies from scripts/requirements.txt, supports read-only --check, and the dashboard parser service prefers the repo virtualenv.
• OCR-ready and legal diligence intelligence - scanned/image-only documents expose explicit OCR-ready bridge metadata and next action; legal/closing checklists can become review-only diligence.checklistItems candidates with line provenance.
• Proof-path dashboard - Intake and IC Package views show the four-step Source doc -> Approved field -> Agent workpaper -> IC package path with conservative pending/ready states.
• One-command release proof - npm run verify:v3 runs release checks, root tests, parser/workspace coverage, dashboard typecheck/build, audits, offline eval, production smoke, and full Playwright E2E. CI now runs this gate too.

What's New in v2.8.5

• One persistent "deal space" - the six-tab dashboard is replaced by a single frame: a deal header + an always-visible 7-stage lifecycle spine (Intake → Diligence → Underwriting → Financing → Legal → Closing → IC) + a context-sensitive center stage + a right rail (live feed + "Your Team") + a command bar. Power-user controls (runtime/Codex limits, criteria, presets, mission control, logs, recovery) move into an Advanced drawer.
• Intake with no manual entry - drop the rent roll, T12, and offering memo and the deal record auto-fills from trusted source-backed values; you only edit what the team flags, and each edit persists with provenance and an audit entry. The numbers come from your documents, not a data-entry form.
• Summon agents and watch them work - click a teammate, type a command, or tap a chip to open a slide-in panel that streams the agent's reasoning with an elapsed timer, renders its workpaper (finding / verdict / caveats) with "open full workpaper", and takes a follow-up task. Offline replays recorded work; the live Codex runtime dispatches a single agent.
• Re-presentation, not a rewrite - the redesign is the new frame plus three thin, guarded backend hooks (single-agent dispatch, inline field override, per-agent stream). The engine, schemas, agents, and source-decision audit trail are unchanged, and the deterministic offline Parkview demo stays the default public path.

What's New in v2.8.0

• Real-world drop-flow hardening - a messy pile of T12s, rent rolls, offering memos, and junk files flows through classify → extract → review → workflow → export with no crashes, silent skips, or confidently-wrong numbers: vacant-$0 rent no longer deflates in-place averages, content-aware rent-roll/T12 classification, graceful parse_failed for oversized/corrupt inputs, path-redacted parser errors, and an automated npm run test:pile smoke test.
• Threshold-driven IC verdict - the deterministic engine consults config/thresholds.json for dealbreakers and a deal-specific exit cap (fixed a clean deal wrongly marked FAIL).
• Open evaluation harness - npm run eval scores an 8-deal synthetic benchmark (with npm run eval:offline for the no-API layers) and writes an honest trust report. The live (Codex) layer proves the agents catch narrative risks the deterministic fixture is blind to (red-flag recall 100%, dealbreaker recall 100%, IC verdict 100%, determinable financials 100%), with model-dependent returns (~25%) documented as an honest limit. See Honest Evaluation.

What's New in v2.7.0

• Text-based PDF extraction - offering memos and rent rolls in PDF now produce source-backed candidate fields with per-field confidence and page-level provenance; scanned/image-only PDFs are detected, marked OCR-ready, and held for local review-gated OCR rather than silently skipped.
• Legal checklist candidates - Markdown/TXT legal or diligence checklists can produce low-confidence diligence.checklistItems candidates with line provenance for operator review, without auto-applying economics.
• Tougher spreadsheet parsing - merged-cell workbooks are unmerged and forward-filled before header detection, image-only workbooks are flagged, and new fixtures cover currency symbols, subtotal/total rows, trailing notes, and synonym headers.
• Review-grade workpapers - workpaper quality gates (cited inputs, assumptions, calculations, caveats, reviewer signoff), per-phase evidence-completeness scoring, IC red-flag drilldowns back to the originating workpaper/source, and richer IC export with source drilldowns and package version history.
• Source-decision audit trail - timestamped approve/reject/waive history per field with cross-document conflict blocking, plus field-level provenance deep links from an approved input to its source snippet.
• Live Codex runtime hardening - per-agent retry/backoff, partial-failure re-run-only-failed-agents, secret redaction at the logging boundary, a redacted sample run manifest with its own schema, and an operator "retry failed agents" recovery action.
• Single-operator self-host deployment - npm run serve serves the built dashboard plus the loopback API/WS together (loopback-default, not multi-tenant); see docs/DEPLOYMENT.md.
• Contributor experience - an end-to-end "add a new specialist agent" guide, a dashboard architecture map, and a npm run release:check readiness gate.

Key Features

---