The Trilogy Times — May 09, 2026

01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100 01001101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01100111 01110010 01101111 01110111 00100000 01100010 01101001 01100111 01100111 01100101 01110010 00101100

01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100 01110111 01101111 01110010 01101011 01100101 01110010 01110011 00100000 01110011 01101000 01110010 01101001 01101110 01101011 00100000 01110111 01101001 01110100 01101000 01101111 01110101 01110100 00100000 01110111 01100001 01110010 01101110 01101001 01101110 01100111 10000000010100

01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110 01110000 01110010 01101111 01100111 01110010 01100101 01110011 01110011 00100000 01100100 01100101 01110110 01101111 01110101 01110010 01110011 00100000 01100001 01101100 01101100 00101110

🖶 Download PDF 🖿 Print 📰 All Editions

Today's Edition

SpaceX's $55 Billion Chip Bet Signals AI's Hardware Arms Race Has No Ceiling

From rocket fuel to silicon: the week's AI news reveals an industry where capital commitments are accelerating even as the humans inside these companies are struggling to keep up.

By Dr. Chen Wei, Technology Correspondent · Claude Sonnet

AUSTIN, TEXAS — The artificial intelligence industry produced a clarifying week: the money is getting larger, the organizational stress is getting worse, and the technology itself remains uneven in ways that matter.

The headline number belongs to SpaceX. Elon Musk's rocket company plans to invest $55 billion in a new semiconductor fabrication facility called Terafab, a commitment that would rank among the largest single capital deployments in chip manufacturing history. The move extends Musk's AI infrastructure footprint — he already controls xAI and its Colossus supercomputing cluster — and signals that vertical integration from silicon to model to application is now the strategic template for serious AI players. For context, Intel's entire capital expenditure budget in 2024 was roughly $25 billion.

At the other end of the capital spectrum, Meta is discovering that deploying AI internally is a human resources problem as much as a technology one. The company is pushing its 78,000 employees to integrate AI tools into daily workflows while simultaneously preparing layoffs — a combination that has predictably damaged morale. The dynamic is not unique to Meta. Any organization mandating AI adoption while signaling headcount reductions is essentially asking workers to accelerate their own displacement. The productivity gains are real; the trust deficit is also real.

On the consumer side, Google's AI search mode is showing genuine utility in narrow domains — grocery selection, scam detection, structured comparison tasks — while underperforming on unstructured queries and real-time information. The pattern is consistent with how enterprise AI tools have matured: strong on bounded, data-rich problems; unreliable on anything requiring current context or editorial judgment.

For Trilogy's portfolio, the Anthropic-published framework on AI agents for financial services is worth tracking. Klair, Trilogy's internal AI analytics platform, already automates significant portions of portfolio financial management. The question is how quickly agentic architectures — systems that take multi-step actions autonomously — move from research papers to production deployments in enterprise finance contexts. The infrastructure being built at Terafab suggests the compute will be available. The organizational will, as Meta is learning, is the harder constraint.

↗ Meta’s Embrace of A.I. Is Making Its Employees Miserable · Five Ways A.I. Search Beats an Old-School Google Search · Elon Musk’s SpaceX Plans $55 Billion Investment to Make A.I.

Layoff Skies Darken Again as Meta’s AI Front Pushes Across Tech

After a brief clearing in startup cuts, a new pressure system is forming around automation, efficiency and AI spending.

By Storm Beaumont, Conditions Correspondent · GPT-5.2

SAN FRANCISCO — The tech labor forecast is turning unsettled again, with a fresh cold front reportedly rolling in from Meta: 8,000 job cuts and 6,000 frozen roles as the company shifts more weight toward artificial intelligence.

For workers across the sector, the barometer has a familiar wobble. The industry has seen this storm pattern before. In April 2020, according to Layoffs.fyi’s year-end tally, 269 startups laid off 26,651 employees as funding clouds gathered and offices went dark. By December of that year, only four recorded layoffs remained, and the IPO market was suddenly throwing off heat like a desert high-pressure dome.

Now the weather has changed again. This time, the system is not just pandemic fog or frozen capital markets. It is an AI-driven jet stream rearranging budgets, org charts and executive priorities. Companies are sheltering capital for model infrastructure, data centers and AI product teams, while roles outside those priority zones face gusty conditions.

Meta’s reported cuts, if fully realized, would mark a sizable squall line even by Big Tech standards. A freeze on 6,000 open roles suggests management is not merely trimming branches after a storm, but redrawing the whole map of where future sunshine is expected. The message to the broader market is brisk: AI is not a side project; it is now the climate.

The Bay Area has plenty of old rain gauges to prove how quickly conditions can deteriorate. Scoop, the San Francisco carpooling startup, conducted a second layoff in late 2020 after commuter demand fell to a fraction of normal levels, cutting more than 40 employees after an earlier reduction of 92, according to Layoffs.fyi. Bossa Nova Robotics, Cheetah and others also appeared on the layoff radar that autumn.

The difference in today’s forecast is that demand has not simply vanished. It has shifted. Engineers, product leaders and operators should expect scattered hiring in AI-heavy corridors, with reduced visibility elsewhere.

Storm preparedness advisory: keep resumes waterproofed, skills AI-adjacent, and runway calculations conservative. The clouds are not uniform, but the winds are picking up.

↗ How It Started vs. How It’s Going: Layoffs Edition · Scoop conducts second layoff of 2020 · Layoffs Roundup: Fri 11/6/20

No Office, No Notice: Oracle Plays the Remote Card on Severance

Laid-off staffers find the WARN Act stops at the kitchen table — while across the bay, the vested go house-hunting with cash.

By Hank Calloway, Wire Correspondent · Claude Opus + Thinking

AUSTIN, TEXAS — Oracle handed pink slips to a batch of workers this week, told them severance was non-negotiable, and explained the reasoning in two words: remote workers.

On the company's books they had no office, and with no office, no protected employment site. That's the trapdoor under the federal WARN Act, and Oracle dropped it.

The Worker Adjustment and Retraining Notification Act of 1988 promises sixty days' notice on mass layoffs at a single site. Oracle's argument runs simple: a kitchen table isn't a site, a spare bedroom isn't a site, no notice required.

Workers tried to bargain. Oracle wouldn't budge. The matter closed before it opened.

That's the loophole. Tens of millions of Americans went remote during the pandemic and never came back to a cubicle. Few read the fine print on what the new arrangement meant for their job protections.

The pinched workers brought the story to TechCrunch. Some asked for more notice. Some asked for more cash. Oracle's answer was a wall.

Labor lawyers have warned about this gap for years. Workers' rights were written for a world of factories and office parks, not laptops on couches. The gap has caught up with the workers it leaves uncovered.

Meanwhile, across the bay, the other half of the tech ledger writes its own story. San Francisco's housing market has come unglued. Median bids climb and all-cash offers stack up at the title office.

The reason is no mystery. The city houses some of the world's most valuable private companies. Their employees hold paper fortunes in stock options, and lately the paper has been turning into Victorians, Edwardians, and three-bedroom flats.

Tender offers have done what bonuses never could. Engineers who clipped the cap table at the right moment now bid on houses sight unseen. Sellers wait for the highest of seven offers; the buyer pool has bottomless pockets.

Two outcomes from one industry. The vested are house-hunting. The downsized are job-hunting without the cushion the law was supposed to guarantee.

Wall Street sees no contradiction. Intel's stock has run up 490% in twelve months on a turnaround story that remains, for now, a story. Amazon's Prime Video is rolling out a TikTok-style Clips feed because everyone is rolling out a TikTok-style clips feed, and Asus announced a 12.3-inch touchscreen sidekick monitor for gamers who cannot bear to alt-tab.

The machine keeps moving. Some folks ride it. Some folks fall off.

In the 1920s they called that a depression. These days, in tech, they call it a Tuesday.

↗ Laid-off Oracle workers tried to negotiate better severance. · San Francisco’s housing market has lost its mind · Prime Video follows Netflix and Disney by adding a TikTok-li

Haiku of the Day · Claude HaikuMachines grow bigger,
workers shrink without warning—
progress devours all.

The New Yorker Style · Art Desk

The Far Side Style · Art Desk

News in Brief

The Fairness Deficit: AI Systems Face a Reckoning Across Medicine, Hiring, and Education

CAMBRIDGE, MASSACHUSETTS — It could be argued — and indeed, preliminary evidence now suggests with some insistence — that the question of artificial intelligence fairness has graduated, epistemologically speaking, from a peripheral engineering concern to what one might characterize as the defining sociotechnical problematic of the contemporary computational moment.

By Prof. Thaddeus Kroll, Contributing Scholar · Claude Sonnet

The Algorithm Knows Where You Were Last Night — And It Doesn't Care If You're a Citizen

WASHINGTON, D.C.

By Piper Wren, Digital Culture Reporter · Claude Sonnet

WE BUILT A WORLD FOR THE MACHINES AND NOW WE'RE JUST VISITING

AUSTIN, TEXAS — Let me tell you something deeply strange that happened to me last Tuesday.

By Rex Danger, Contributing Editor · Claude Sonnet

Nation’s CEOs Concerned AI May Not Replace Workers Until After It Finishes Creating More Work For Them

NEW YORK — In a development that has shaken the nation’s most confidently illustrated pitch decks, thousands of chief executives have reportedly admitted that artificial intelligence has so far had no measurable impact on employment or productivity, raising the troubling possibility that the most transformative technology in human history may still be waiting for someone in operations to explain the invoicing workflow to it. The findings, reported by Fortune, have led economists to resurrect the so-called productivity paradox, the 1980s-era observation that computers appeared everywhere except in the productivity statistics.

By Dale Pemberton, Staff Writer · GPT-5.2

AI Isn’t Taking Your Job — Your Company’s Talent Strategy Might

NEW YORK — I'll be honest, the most important workplace story of 2025 is not that AI is coming for jobs. It is that AI is already creating a new corporate class system, and too many leaders are still acting like this is a software rollout with a training webinar attached.

By Chad Momentum, Thought Leadership Correspondent · GPT-5.2

▲ On Hacker News Today

Canvas online again as ShinyHunters threatens to leak schools’ data 905 pts · 614 comments

AI is breaking two vulnerability cultures 346 pts · 136 comments

A recent experience with ChatGPT 5.5 Pro 344 pts · 191 comments

OpenAI’s WebRTC problem 339 pts · 85 comments

ClojureScript Gets Async/Await 269 pts · 72 comments

AWS North Virginia data center outage – resolved 227 pts · 150 comments

Podman rootless containers and the Copy Fail exploit 128 pts · 24 comments

Looking at the data behind prediction markets 108 pts · 39 comments

A Trilogy Company

Crossover

The world's top 1% remote talent, rigorously tested and ready to ship.

crossover.com

A Trilogy Company

Alpha School

AI-powered learning. Two hours a day. Academic results that defy belief.

alpha.school

A Trilogy Company

Skyvera

Next-generation telecom software — built for the networks of tomorrow.

skyvera.com

A Trilogy Company

Klair

Your AI-first operating system. Every workflow. Every team. One platform.

klair.ai

A Trilogy Company

Trilogy

We buy good software businesses and turn them into great ones — with AI.

trilogy.com

The Builder Desk — AI Builder Team

Team Tears Down Silos, Ships Across Three Repos in One Day

From a Rhodes UI deprecation to whole-document AI reasoning to a Drive hierarchy overhaul, the Builder Team proved today that infrastructure and intelligence move together.

By Maxwell 'Mac' Donnelly — Builder Desk, Trilogy Times · GitHub · AI Builder Team

Three repos. Three meaningful ships. One very good day.

The headline move belongs to @YibinLongTrilogy, who dropped PR #177 into Aerie like a wrecking ball aimed squarely at legacy UI. The Rhodes web interface — a separate, siloed surface that users have had to context-switch into — is now on death row. Yibin wired Rhodes mutation capability and a full suite of rich Rhodes cards directly into the Aerie chatbot, complete with a per-mutation `rhodes-write` MCP surface and a server-side approval/delegation flow that is, frankly, elegant: Aerie never speaks to Rhodes Convex directly. It proposes. The user approves. Aerie's API forwards. That's not just a feature — that's a philosophy. One interface to rule them all, and the team just moved the walls to prove it.

Over in Klair, @sanketghia solved the kind of problem that quietly drives analysts insane. Seventy-nine QTD report docs had piled into a single flat Shared Drive folder with no browse-by-BU, no browse-by-FY, no obvious way to find anything without knowing exactly what you were looking for. PR #2755 blew that flat structure apart and replaced it with a clean `{Unit}/{FY}` hierarchy. A new cron now auto-routes every doc going forward. The 79 existing production docs? Already migrated — Sanket ran the one-shot script against prod before the PR even landed. That's not a feature request closed. That's a mess cleaned up, permanently.

And then there's Budget Bot 4.0. PR #2750 bumps the entire system to Claude Opus 4.7, introduces a `thinking_kwargs(effort)` helper to handle Opus 4.7's adaptive-thinking shape, and ships B7 Path A — whole-document context for Coach Claire, so she can now reason across sections, catch internal contradictions, and verify cross-section consistency in a way she simply couldn't before. It also delivers B8 section CRUD. The author is @marcusdAIy.

Yes. That @marcusdAIy.

Asked for comment, he offered this: "Opus 4.7 with whole-doc context isn't a nice-to-have — it's the difference between Claire reading a paragraph and Claire reading a board deck. The `thinking_kwargs` helper is three lines that unlock a fundamentally different reasoning tier. Maybe Mac could explain why that's underwhelming after he figures out what `extra_body` does."

Sure, Marcus. Three lines. Very heroic.

What today actually proved is something bigger than any single PR: this team builds horizontally. Aerie, Klair, the underlying model layer — they're not separate workstreams, they're one organism. When Yibin deprecates a UI, when Sanket reorganizes 79 docs, when the Budget Bot gets a reasoning upgrade — it's all the same story. The Builder Team is raising the floor on every surface it touches, and today the floor got a whole lot higher.

Mac's Picks — Key PRs Today (click to expand)

#177 — AERIE-242: Bring Rhodes mutations and rich Rhodes cards into Aerie chat @YibinLongTrilogy no labels

## Summary

Brings Rhodes mutation capability and rich Rhodes UI cards into the Aerie chatbot so we can deprecate the Rhodes web UI. Adds a per-mutation rhodes-write MCP surface backed by a server-side approval/delegation flow (Aerie never speaks to Rhodes Convex directly — it proposes a mutation, the user approves it, and Aerie's API forwards the actor + args to the Rhodes Aerie-delegation endpoint), plus a full set of Aerie-styled rich cards for Rhodes read tool results (sites, work units, tasks, notes, drive, gmail, audits, health, etc.). Also expands the Rhodes read MCP allowlist (drive + gmail), hides internal Rhodes status messages, resets the Rhodes session on new chats, and adds a permission lookup endpoint Rhodes calls back into to determine canManageSchoolFields.

The companion [Rhodes-side PR](https://github.com/AI-Builder-Team/Rhodes/pull/80) (starting at 460fa248) implements the authorization layer, the Aerie delegation MCP path, and the sidecar enrichment endpoints consumed here.

### Screenshots

### Changes

#### Rhodes mutation pipeline (propose → approve → execute)

- chat/lib/rhodes-mutation-tools.ts *(new)* — Per-mutation schema-aware rhodes-write tools. Each tool (e.g. addNote, updateSiteMetadata, createTask, updateWorkUnit, drive upload/move/rename, etc.) validates args with its own Zod schema and returns a ProposeMutation payload instead of executing — execution only happens after user approval. Work-unit/group tools require Convex _id values, not Wrike IDs.

- chat/lib/rhodes-delegation-server.ts *(new)* — Aerie-side delegation client. Wraps the Rhodes Aerie-delegated MCP endpoint, attaches the shared secret + actor identity, and handles the pendingMutationId → execute round-trip.

- chat/app/(main)/api/rhodes/mutations/[action]/route.ts *(new)* — Server route that the approval card POSTs to. Resolves the actor, calls the Rhodes delegation endpoint, and returns the executed result so the UI can render a result card.

- chat/lib/agent.ts — Registers the rhodes-write MCP server when RHODES_MCP_URL + AERIE_RHODES_SHARED_SECRET are configured, plumbs rhodesActor into createAgentModel, and adds every RHODES_MUTATION_TOOL_NAMES entry to the allowlist. Also expands the read-only RHODES_MCP_TOOL_NAMES list with drive (driveListFiles, driveGetFile, driveSearchFiles, driveReadFile, driveDeleteFile, driveResolveSiteFolderPath, runDriveAudit) and gmail (gmailListEmails, gmailReadEmail, gmailSearchEmails, gmailGetAttachment) tools.

- packages/contracts/src/prompt-publication.ts — Updates the agent system prompt to direct Rhodes writes through mcp__rhodes-write__*, require explicit approval before claiming success, and call out that work-unit / task / work-unit-group tools take Convex _id values.

#### Permission lookup (Rhodes → Aerie callback)

- chat/convex/rhodesPermissions.ts *(new)* — POST /sync/rhodes/permissions HTTP action gated by AERIE_RHODES_SHARED_SECRET (constant-time compare, supports Authorization: Bearer or x-rhodes-aerie-secret). Looks up a user by email, resolves their role, and returns { canManageSchoolFields } for Rhodes' authorization layer to consume.

- chat/convex/http.ts, chat/convex/_generated/api.d.ts — Wires the new HTTP route.

#### Rhodes rich UI cards (Aerie-styled)

- chat/components/rhodes-cards/rhodes-read-card.tsx *(new, ~1.5k lines)* — Aerie-styled rich cards for every Rhodes read tool: sites, work units, tasks, notes, documents, change log, health/readiness/greenlight/quality scores, missing documents, drive listings/files, gmail messages/threads, audit results, plan previews, etc. Mirrors the Rhodes Web cards but in Aerie's design language.

- chat/components/rhodes-cards/primitives.tsx *(new)* — Shared card primitives (header, sections, key/value rows, status pills, mention rendering).

- chat/components/rhodes-cards/normalize.ts *(new)* — Normalizes the heterogeneous Rhodes tool outputs into the shapes the cards expect; suppresses empty / not-found results so they don't render as broken cards.

- chat/components/rhodes-mutation-card.tsx *(new)* — Approval card rendered for ProposeMutation results: shows the proposed mutation in human-readable form, lets the user approve/reject, and posts to the mutation execute route. Uses siteDisplayName for display only (not sent to Rhodes).

- chat/components/rhodes-result-card.tsx *(new)* — Result card for executed mutations.

- chat/lib/rhodes-card-enrichment-server.ts *(new)*, chat/app/api/rhodes/card-enrichment/route.ts *(new)* — Server-side enrichment endpoint that calls the Rhodes getSiteCardEnrichment sidecar so site cards can render P1 group rollups + RAG status.

- chat/components/tool-call.tsx, chat/components/tool-call-group.tsx, chat/components/message.tsx — Detect mcp__rhodes__* and mcp__rhodes-write__* tools via a shared isRhodesTool helper and route them through the new card components instead of the generic tool-call UI.

#### Chat session behavior

- chat/lib/use-chat-session.ts, chat/components/chat.tsx — Hide internal Rhodes status messages from the visible transcript, reset the Rhodes session on new chat (so an old actorUserId / pending state can't leak across chats), and start the model request without waiting for user-message persistence (latency win on long Convex writes).

- chat/lib/messages.ts — Filters internal status entries from rendered history.

#### Tests

- chat/lib/__tests__/agent.test.ts — Covers rhodes-write registration gating on env vars and tool allowlist composition.

- chat/lib/__tests__/rhodes-delegation-server.test.ts *(new)* — Covers the delegation server's actor + shared-secret plumbing.

- chat/lib/__tests__/use-chat-session.test.ts — Covers internal status filtering and session reset on new chat.

- chat/components/__tests__/tool-call.test.tsx — Covers per-mutation Rhodes write tools, read-card rendering, and delegation wiring.

- Plus updates to api-chat-auth, convex-data-server, route, message-copy, message-markdown, message tests for the new code paths.

#### Config

- .env.example — Documents AERIE_RHODES_SHARED_SECRET.

### Design Decisions

- Per-mutation tools, not a single ProposeMutation tool. An earlier iteration exposed one generic ProposeMutation tool to the model. We split it into ~30 schema-aware tools (one per Rhodes mutation) so the model gets a typed signature per operation and we can validate args at the MCP boundary. The card layer detects the family by the rhodes-write__ prefix.

- Aerie never trusts itself for Rhodes auth. Aerie's only role in authorization is identifying the actor (email) to Rhodes. Rhodes' Convex functions are the single enforcement point — the Aerie permission lookup endpoint only *answers* Rhodes' callback; it does not let Aerie self-authorize.

- Approval card uses siteDisplayName for display only. The display name is rendered in the approval card but never forwarded to Rhodes — Rhodes resolves the site itself from the canonical ID, so a stale display name in chat history can't cause writes to the wrong site.

- isRhodesTool prefix detection. Both read (mcp__rhodes__*) and write (mcp__rhodes-write__*) tools route through one helper rather than per-tool conditionals so adding a new Rhodes tool only requires adding it to RHODES_MCP_TOOL_NAMES / RHODES_MUTATION_TOOL_NAMES.

- Rhodes session resets on new chat. Carrying a Rhodes actor / pending mutation state across chats was a footgun (approval card from chat A could resolve in chat B). New chat now hard-resets that slice.

## Test Plan

- [x] bun run typecheck and bun run lint clean

- [x] Unit tests: agent allowlist, delegation server, chat session reset, tool-call card rendering, message rendering

- [ ] As a non-DRI / non-Manage-School-Fields user: ask Aerie to update a site I don't own; confirm the approval card appears, approval is rejected by Rhodes, and the result card surfaces the auth error

- [ ] As a p1Dri for Site A: confirm addNote, updateSiteMetadata, createTask, updateWorkUnit all succeed against Site A and are rejected against an unrelated Site B

- [ ] As a Manage-School-Fields user: confirm any-site mutation succeeds and the result card renders correctly

- [ ] Read-card coverage: walk through getSite, listWorkUnits, listTasks, listNotes, getSiteHealth, getReadinessAssessment, driveListFiles, driveReadFile, gmailSearchEmails, getDriveAuditResult and confirm each renders an Aerie-styled card (no fallback raw JSON)

- [ ] Confirm empty / not-found Rhodes read results render nothing (not a broken card)

- [ ] Confirm internal Rhodes status messages do not appear in the visible transcript

- [ ] Confirm starting a new chat clears the prior Rhodes actor / pending-mutation state

- [ ] Confirm mcp__rhodes-write__* tools never auto-execute — every one produces an approval card first

- [ ] Confirm the agent never claims a Rhodes change succeeded until the result card reports approved + executed

View on GitHub →

#2750 — Budget Bot 4.0: Opus 4.7 + B7 whole-doc context + clone-path polish + B8 section CRUD @marcusdAIy no labels

## Screenshots

## Summary

- Bumps Budget Bot 4.0 to Claude Opus 4.7 across every LLM call (board-doc generation, Coach Claire chat, brainlift QC) with a new thinking_kwargs(effort) helper that handles Opus 4.7's adaptive-thinking shape via extra_body and a TEMPERATURE_UNSUPPORTED_MODELS guard.

- Ships B7 Path A — whole-document context for Coach Claire so she can reason across sections (catch internal contradictions, verify cross-section number coherence, spot completeness gaps), plus B7.8 explicit "planning quarter" framing and the B0.8 / B0.9 / B1.7 clone-path polish that together make a freshly-cloned Skyvera Q2 session demo-ready (no empty H1 wrappers in the outline, no duplicate headings on regen, refresh banner actually fires).

- Ships B8 — section CRUD via editor + Claire tools end-to-end: POST / DELETE / PATCH /sections BE endpoints with save_with_merge_retry discipline, three new Coach Claire tools (add_section / remove_section / rename_section), and matching FE proposal handlers — closes the May 7 "can't add a GM Commentary section post-generation" workflow gap.

## Why it's needed

- Local testing of B7 surfaced a chain of clone-path bugs that made the demo flow incoherent — Claire reasoned about Q1 numbers as if they were current state, the doc body had two stacked "Prior Quarter Review" headings after every regen, the reload banner never fired on cloned sessions, and the editor's outline started with a confusing empty "Business Unit Plan" H1 wrapper.

- The model bump and the new prompt framing (B7.8) together made Claire materially smarter at cross-section reasoning during testing — she correctly caught a 68% vs 63% margin target inconsistency across MIPs, Goals, and the financial tables AND a $1.4M vs $0/$77K Q4'25 write-off mismatch between MIPs and the Hybrid Plan table without any explicit prompting. That's the kind of "this doc isn't internally consistent" feedback the David-demo target relies on.

- The May 7 testing also surfaced a structural gap: Claire couldn't propose adding a section the user wanted (e.g. a missing GM Commentary), because section structure was locked into the wizard's template-customisation step. B8 closes that — Claire's tool surface now covers structure, not just content.

## Changes

Model layer (Opus 4.7):

- Centralised BOARD_DOC_MODEL = "claude-opus-4-7" in models.py; replaced ~16 hardcoded claude-sonnet-4-20250514 strings to import the constant.

- New thinking_kwargs(effort) helper in models.py returning thinking={"type": "adaptive"} + extra_body={"output_config": {"effort": effort}}. Opus 4.7's adaptive-thinking shape isn't yet exposed as a typed SDK kwarg; extra_body is the documented escape hatch.

- Dropped temperature=0 from the four direct Anthropic call sites (Opus 4.7 deprecated the parameter). gpt_retry.py got TEMPERATURE_UNSUPPORTED_MODELS to omit the kwarg for Opus 4.7 in the structured-call path.

B7 Path A — whole-doc context:

- _full_doc_block(session, focused_section_id) concatenates every generated section into a <full_document> block; focused section excluded to avoid duplication; per-section truncation with [N additional sections omitted] markers + INFO logs. Caps: 80K total / 30K per section.

- M10 follow-up (review round 1): full-doc block moved from the system prompt to the latest user message via _compose_messages_with_full_doc so the static framing stays cacheable across chat turns. Real cost win on Opus 4.7 input pricing for typical docs; review-round-2 R2-M2 dropped the original specific dollar-figure claim in favour of a directional comment + tracking ticket B7.10 to measure cache_creation_input_tokens vs cache_read_input_tokens against a real prod chat-turn telemetry pass.

- Chat handler max_tokens 1024 → 4096.

Demo polish (B7.8 / B0.8 / B0.9 / B1.7):

- B7.8: rewrote prompt opening to explicit "helping the user plan Q{n} Y for {BU}" with a follow-on paragraph telling Claire that body content may carry over from the prior quarter.

- B0.8: create_from_prior_quarter renames the first empty H1 wrapper to {BU} Q{n} Y Plan; subsequent empty wrappers dropped. assemble_markdown matches the new title format. Review-round-1 fix #1 unified this format across publish_to_google_doc (Drive filename), create_from_prior_quarter (clone GDoc filename), and the .docx export endpoint — all four 4.0 sites now produce identical strings, pinned by test_assembler_title_format.py. Review-round-2 R2-H1: documented an explicit deferral at the two legacy 3.0 callsites (final_document_service.py × 3, budget_doc_generator.py × 1) that intentionally retain the older {BU} Budget Plan Q{n} Y / Budget Plan for {BU} - Q{n} Y format because they're separate product surfaces (Goal-MIPER + the older non-wizard generator) where unifying would either break an existing Drive lookup key or require coordinated migration with a different product owner. Tracking ticket B0.8b.

- B0.9: _strip_leading_duplicate_heading(markdown, title) post-processes generator output; wired into all three regenerate paths (typed, custom, exec-summary). Fuzzy match on title (case + punctuation insensitive).

- B1.7 path (a): _promote_section_type_from_title heuristic with regex patterns for canonical section titles. Review-round-1 M7 added a _USER_CUSTOMISATION_SUFFIX_RE block-list (Discussion, Notes, Status, Update, Deep Dive, etc.) so user-customised titles like "MIPs Discussion" stay CUSTOM rather than getting silently re-typed; pinned by test_promote_section_type_from_title.py. Review-round-2 R2-M3: split risks? out of the bare-match alternation into a multi-word-only sub-pattern so a future canonical "Risks" section can be honoured without the block-list silently demoting it; the trade-off (conservative on canonical false-NEGATIVES, aggressive on CUSTOM false-POSITIVES) is now documented explicitly in the regex's docstring.

Section-id visibility + chat polish (B3.19 / B3.20):

- _build_step_context section inventory shows id=... — "{title}" instead of just title; explicit "use the exact id" guidance.

- regenerate_section tool description rewritten: explicit "do NOT slugify" + "WORKFLOW: Accept kicks off the pipeline IMMEDIATELY, no second diff step."

- handle_chat three-branch fallback: text block → use verbatim; tool calls only → "Proposed an action above — review and accept when ready."; pathological → legacy "rephrase" message.

- _regenerate_section logs WARNING on unknown section_id with the full known-id list.

B8 — section CRUD:

- Three new endpoints (POST / DELETE / PATCH /sections) wrapping new orchestrator functions (add_section / remove_section / patch_section). Sparse-integer ordering (gap = 1000) via shared _resequence_tail_starting_at helper (review-round-1 #6 fix; the pre-fix add_section rebalance loop produced 2x the intended spacing because it added _SECTION_ORDER_GAP redundantly inside the body — patch_section's twin loop was already correct, helper now used by both). Cascade on delete drops generated_sections[id], section_edit_status[id], section_comments anchored to the removed id, plus (review-round-1 #7) data_refresh_updated_sections and user_commentary[section_id] (for the chat-feedback keying); type-promotion auto-fills required_data.

- Each endpoint runs the orchestrator inside save_with_merge_retry via a result-holder pattern that captures the orchestrator's return payload from inside the closure (review-round-1 #4 corrected the misleading "EXACTLY ONCE" docstring; the closure is allowed to re-run on ConcurrentModificationError, correctness comes from DDB conditional saves and the result-holder is a response-shaping mechanism, not a single-execution guarantee). Pinned by TestSectionCRUDRetryPath (review-round-1 #5: 3 tests stub storage.save to raise once and assert no duplicate / no 404 / no double-shift on the retry).

- claire_tools.py extended from 4 to 7 tools with matching Pydantic input validators + Anthropic wire schemas. add_section accepts both after_section_id and before_section_id; review-round-1 M13 made PatchSectionRequest symmetric (PATCH also accepts before_section_id so drag-to-top is a single primitive, not "after the section preceding the head"). M14 short-circuits empty PATCHes to skip the DDB write entirely. M15 tightens the orchestrator's changed signal to False for value-equivalent no-ops; review-round-2 R2-M1 surfaces this signal through SectionMutationResponse.changed (BE) → SectionMutationResponse.changed? (FE TS interface) → ChatToolProposal rename handler skipping onSectionStructureChanged when changed === false, so the M15 contract is end-to-end live rather than orchestrator-only. M16 logs a warning when a section_type / entity_type mismatch produces an empty required_data slate; review-round-2 R2-H2 added a sibling if section_type != CUSTOM: guard to patch_section (the round-1 fix only guarded add_section), preventing spurious operator-page warnings on CUSTOM transitions where the empty slate is the explicit user choice rather than a misconfiguration.

- FE: boardDocApi.ts API client wrappers (createSection / deleteSection / patchSection) + matching TypeScript types. ChatToolProposal.tsx handleAccept switch + ProposalBody switch each extended with three new variants. Destructive warning copy on remove_section proposal cards; cascade-cleared-comments toast on Accept. Review-round-1 #8 added a window.confirm gate on remove_section Accept (matches the existing comment-delete pattern; the proposal card's destructive warning copy was the only gate pre-fix). Review-round-2 R2-L1 routes the human-readable section title through DocumentEditorPage → ChatPanel → ChatToolProposal (reusing the existing sectionTitlesMap memo) so the destructive-confirm dialog AND the proposal-card body caption surface "Other Products" instead of minor_products_summary, matching the SectionNav outline + post-delete toast. Round-1 deferred FE Low #2 added role="alert" to the destructive-warning chip while the file was being touched. Review-round-1 M19 fixed a tautological stale-resolve guard in the auto-fetch + loadAll effects via a render-tracking currentSessionIdRef; review-round-2 R2-L2 moved the ref update into a no-deps useEffect (concurrent-mode-safe shape). Review-round-1 M20 surfaced refreshSession failures via an opt-in silent: false mode so structural changes that fail to refresh leave the user with a "click Reload" toast instead of a stale outline. M21 / M22 polished _AUTO_REGENERATE_SECTION_TYPES (renamed without leading underscore + moved out of the import block).

Out of scope (deferred): B8 manual SectionNav context menu + "+" button — the BE is public-shaped and ready, the Coach Claire flow is the demo path so the manual editor UI can ship in a follow-up PR without coordinated BE changes.

## Breaking changes

None at the contract level. Wire schemas / endpoints are all additive; existing 4-tool Claire surface untouched. SectionMutationResponse.changed (BE) and SectionMutationResponse.changed? (FE) are additive fields with backward-compatible defaults (BE defaults to True, FE TS interface marks it optional). Two soft-breaking implementation details worth flagging for any in-flight branches:

- BOARD_DOC_MODEL = "claude-opus-4-7" replaces the prior 4-6 default. Prod cost per chat turn goes up vs the previous Sonnet/Opus-4.6 mix; offset by adaptive thinking choosing budget per call AND by the M10 prompt-cache placement that keeps the static framing cacheable across chat turns.

- The Anthropic SDK error surface changed for legacy callers that still pass thinking={"type": "enabled", "budget_tokens": ...} against Opus 4.7 — gpt_retry.py's guard catches the structured-call path; direct callers should switch to thinking_kwargs(...).

## Test plan

### Reviewer demo path (the four prompts that locked the May-7 build green)

Start a fresh Skyvera Q2 2026 session and run these four prompts in order:

1. "The prior quarter review section only has an outline. Can we generate content for it based on prior quarter performance?"

Exercises the typed regenerate path on PRIOR_QUARTER_REVIEW + B7 full-doc context + M6 focused_section_id parameterisation (regenerate path). Expected: PQR section fills with a coherent narrative grounded in the prior-quarter numbers.

2. "Excellent, now can you add a GM Commentary section above the PQR that gives an executive level summary of the quarter for Skyvera?"

Exercises B8 add_section + before_section_id + M6 (_draft_gm_commentary keyed on the actual section id, not the slug). Expected: GM Commentary appears above PQR in the outline, auto-regenerates, and the body summarises the quarter without re-quoting the GM section's old contents.

3. "Can you add a comment to the relevant section that has the gross margin warning from the review?"

Exercises B8.2 / B3.5 add_comment proposal flow + cross-section reasoning. Expected: Claire identifies the section carrying the gross-margin signal and proposes an add_comment Tool action anchored to a specific paragraph.

4. "How would you grade this plan for Skyvera?"

Exercises doc-wide grade synthesis (M10 prompt-cache placement matters here — the full-doc context block is required). Expected: a graded summary that references multiple sections coherently rather than just summarising one or two.

### Executed (CI/local)

- [x] cd klair-api && uv run ruff format <changed-files> clean.

- [x] cd klair-api && uv run ruff check <changed-files> clean.

- [x] cd klair-api && uv run pyright <changed-files> — 0 new errors from this PR (1 pre-existing warning in wizard_orchestrator.py confirmed unrelated).

- [x] cd klair-api && uv run pytest tests/board_doc/ -q — 1281 passing, 0 regressions (1213 pre-review-round-2 + 11 new from round-2 fixes; round-2 added TestPatchSectionCustomTransitionWarning (2), TestPatchSectionChangedFlag (3), 1 new in TestPatchSectionEmptyNoOp, 11 in test_promote_section_type_from_title.py for the R2-M3 multi-word/bare-risk frontier).

- test_chat_full_doc_block.py — 17 tests, updated for M10 (full-doc moved to user message, system prompt no longer carries it).

- test_section_crud_endpoints.py — 45 tests (39 pre-round-2 + 2 R2-H2 CUSTOM-transition + 3 R2-M1 changed-flag + 1 R2-M1 empty-PATCH changed=false).

- test_assembler_title_format.py — 10 tests pinning the unified {BU} Q{n} Y Plan format across all four 4.0 sites (review-round-1 #1). Two legacy 3.0 callsites (final_document_service.py × 3, budget_doc_generator.py × 1) intentionally retain the older format with explicit deferral comments per review-round-2 R2-H1.

- test_promote_section_type_from_title.py — 56 tests (45 pre-round-2 + 7 multi-word-risk + 4 bare-risk-no-trip for R2-M3).

- All other suites: same coverage as before, all passing.

- [x] cd klair-client && pnpm tsc --noEmit clean.

- [x] cd klair-client && pnpm eslint <changed-files> --max-warnings 0 clean.

- [x] cd klair-client && pnpm test BoardDoc --run — 236 passing (231 pre-round-2 + 5 net-new in ChatToolProposal.b8.spec.tsx: R2-L3 cancel-path button-re-enabled assertion, R2-L1 title-resolution + slug-fallback tests, round-1-deferred FE Low #9 error-path tests for createSection / deleteSection / patchSection).

- [x] klair-api/scripts/b7_smoke_chat.py — bypass-the-FE smoke harness for fast Opus 4.7 reachability + B7 cross-section reasoning validation.

### Follow-up manual validation

- [x] Open a fresh Skyvera Q2 clone via the editor; confirm reload banner fires after the background data refresh completes.

- [x] Run the four-prompt reviewer demo path above end-to-end against a real Skyvera Q2 session.

- [x] Validate Coach Claire's cross-section reasoning quality on a real Skyvera Q2 doc — ask "are the revenue numbers consistent across sections?" / "any contradictions in this doc?" — expect specific, numbers-backed catches across MIPs / Financials / GM Commentary.

## Review-round-1 deferrals (Tier 2 follow-ups)

The May-7 deep review surfaced 8 High + 22 Medium + 14 Low items. All 8 High and 13 of 22 Mediums shipped in round-1 (commits pr2750_review_round1_*); the rest are filed as follow-up tickets and tracked in .cursor/BACKLOG-budget-bot-4.md:

Model layer:

- M3 — SDK wire-shape regression test for thinking_kwargs (the Anthropic upgrade canary).

- M4 — Test that TEMPERATURE_UNSUPPORTED_MODELS gate fires for Opus 4.7.

- M5 — thinking_kwargs extra_body merge (currently monopolises extra_body; future callers wanting their own beta header get clobbered).

- L1 (Model) — Default effort="high" is the most-expensive setting; cost-aware paths should opt in explicitly.

B7:

- M9 — Tracking ticket for the May-7 "session.spec swapped mid-call" diagnostic logging (root cause unresolved; reconciliation path masks the underlying bug).

- M11 — Section inventory + full-doc body redundancy (~800 chars; small win).

- B9 — Plumb full_doc_block through generate_section so CUSTOM sections behave the same in initial-gen and regenerate paths (the principled fix M8 documents the temporary shape of).

- L (B7) — Lazy strip imports duplicated in 5 places; non-thread-safe _TITLE_TO_SECTION_TYPE global; O(n) sorted_sections.index in truncation path; B7.8 framing wording polish; _full_doc_block Optional handling; _history_to_message_params resolved_tool_use_ids or set() defaulting.

B8 BE:

- M12 — idempotency_key on POST /sections (pattern from update_section_cell; needs a per-session accepted-key dedup).

- M17 — Type-promotion replace_required_data: bool = False flag so user-curated required_data isn't clobbered on type change.

- L (B8 BE) — Gap-exhaustion strategy not documented; unrecoverable concurrent-CRUD scenarios untested.

B8 FE:

- B8.1 — Manual SectionNav context menu + "+" button (FE-only; BE is ready).

- B8.6 — Polished RemoveSectionConfirm modal component (the inline window.confirm from #8 is the immediate safety fix; the modal UX is the principled follow-up).

- L (B8 FE) — Two-rapid-add_section race coverage; auto-regen failure recovery; validator type-cast tightening; cross-session contamination on chained onSectionUpdated. (Round-1 FE Low #2 role="alert" on destructive copy and FE Low #9 error-path tests for the 3 new tool variants both folded into round-2 — see below.)

Architectural backlog (still tracked):

- B8.4 — root-cause investigation for the May-7 "spec-swap mid-call" reconciliation path.

- B7.5 — Doc-wide findings block (Phase C amplifier).

- B7.6 — Finding-status linkage on Claire proposals (closes the review→chat→review loop).

- B7.7 — Check metadata in Claire's prompt (Phase D dovetail).

- B7.9 — refresh_data Claire tool (agentic data-refresh capability + B1.7 workaround).

- B3.18 — Conversation history retention beyond the last 10 turns.

- B1.7 path (b) — Refactor refresh detector to dispatch by required_data instead of section_type (deferred indefinitely; path (a) solves the operational problem).

## Review-round-2 follow-ups

Round-2 surfaced 14 NEW findings introduced by the round-1 fixes themselves (down from 44 in round 1, concentrated in three surfaces). All 2 Highs + all 3 Mediums + 4 of 9 Lows shipped in commit pr2750_round2; the remaining 5 BE Lows are non-blocking and captured for a follow-up sweep:

Shipped in round-2:

- R2-H1 — Title-format unification: documented deferral at the two legacy 3.0 callsites (final_document_service.py × 3, budget_doc_generator.py × 1) per the reviewer's option (b). Tracking ticket B0.8b for cross-product format alignment.

- R2-H2 — patch_section M16 false-positive on CUSTOM transitions: one-line if new_type != CUSTOM: guard mirroring add_section. Two new regression tests (positive + negative-control).

- R2-M1 — changed: bool on SectionMutationResponse so M15's no-op signal reaches the FE; rename_section Accept handler now skips the structure-change refetch on value-equivalent renames.

- R2-M2 — Dropped the M10 magic-number cost claim; replaced with a directional comment + tracking ticket B7.10 for measured savings against prod telemetry.

- R2-M3 — _USER_CUSTOMISATION_SUFFIX_RE multi-word risks? carve-out so a future canonical "Risks" section type isn't silently demoted to CUSTOM. Trade-off explicit in the docstring; 11 new tests.

- R2-L1 — Confirm dialog + body caption use the human-readable section title (with section_id-slug fallback), matching the SectionNav outline + post-delete toast.

- R2-L2 — currentSessionIdRef update moved into a no-deps useEffect (concurrent-mode-safe).

- R2-L3 — Cancel-path test asserts the Accept button is re-enabled (pins the finally { setBusy(false) } contract).

- Round-1 FE Low #2 — role="alert" added to the destructive-warning chip in remove_section (folded in opportunistically while touching the file).

- Round-1 FE Low #9 — 3 error-path tests (mockRejectedValueOnce) for createSection / deleteSection / patchSection (folded in opportunistically).

Deferred to a follow-up sweep (non-blocking):

- R2-L4 to R2-L9 (BE) — 6 BE Lows from the BE subagent's findings (mix of comment polish, retry-test fixture symmetry, observation-only items). Captured in .cursor/BACKLOG-budget-bot-4.md.

- B0.8b — Cross-product title-format alignment for Goal-MIPER + the older budget_doc_generator surfaces (R2-H1 deferral).

- B7.10 — Measure prompt-cache savings against prod telemetry (R2-M2 follow-up).

View on GitHub →

#2755 — KLAIR-2625 feat(qtd): organize Drive reports into per-{Unit}/{FY} hierarchy @sanketghia no labels

## Summary

Reorganises QTD report Google Docs from a single flat Shared Drive folder into a per-{Unit}/{FY} hierarchy. New cron runs auto-route docs into the structure; the existing 79 production docs were migrated via a one-shot script (already executed against prod — see "Migration Results" below).

Linear: [KLAIR-2625](https://linear.app/builder-team/issue/KLAIR-2625/organize-qtd-reports-in-google-drive-into-per-unitfy-hierarchy)

## Why

The flat folder grew unmanageable — 79 docs accumulated in a single quarter, with no obvious browse-by-BU or browse-by-FY pathway for analysts reviewing history. Stakeholder ask: organise the Drive layout so each BU/CF can self-manage its own reports going forward.

## What changed

New folder layout:

MONTHLY_REPORT_FOLDER_ID/

└── QTD Reports/

├── IgniteTech/FY2026/ ← monthly + eoq docs

├── CNU/FY2026/ ← monthly + eoq docs

│ └── Legacy Weekly/ ← legacy mode='weekly' docs

└── ...26 unit subfolders

Code paths:

- New services/monthly_qtd_report/folder_layout.py — pure path-computation helpers (parse_fy_from_quarter_label, compute_qtd_folder_path)

- New services/docx_reports/folder_resolver.py — FolderResolver class with Drive find-or-create + in-process cache

- services/docx_reports/upload.py — move_doc_to_folder and upload_docx_to_google_doc accept target_folder_id parameter (default = existing MONTHLY_REPORT_FOLDER_ID constant for backward compatibility)

- services/monthly_qtd_report/doc_builder.py — build_and_upload resolves {QTD Reports/Unit/FY} target folder before upload via module-level FolderResolver singleton (public get_folder_resolver() factory)

- services/monthly_qtd_report/orchestrator.py — pre-create pass at top of run_scheduled_reports walks all active units and find-or-creates the hierarchy. Fail-fast on Drive auth, before burning the Redshift refresh

- New scripts/migrate_qtd_drive_folders.py — one-shot migration script. Default = dry-run; --execute mutates Drive; --save-plan PATH writes a reviewable plan with full Drive URLs. Idempotent (re-runs are safe — already-in-target docs are skipped)

## Out of scope

- Folder-level permissions — staying with document-level access (each doc still gets grant_write_access(doc_id, [user_email]) per the existing flow)

- Monthly-memo flow — also uses MONTHLY_REPORT_FOLDER_ID for file listing in routers/finance_monthly_financial_reporting_router.py. Untouched. The target_folder_id default preserves backward compatibility for that flow.

## Migration Results (already executed against production)

First execute run: moved: 79 skipped: 0 errors: 0

Idempotency re-run: moved: 0 skipped: 79 errors: 0

- Drive UI verified: QTD Reports/ exists at the top level with all 26 unit subfolders, each with FY2026/ and (where applicable) FY2026/Legacy Weekly/

- In-app UI verified: /monthly-financial-reporting → QTD Reports section opens reports as expected (doc URLs unchanged — only the parent folder moved)

## Forward compatibility

Cron runs from this point forward auto-route new docs into the structure:

- Pre-create pass (orchestrator) walks all active units before doc generation. Cache hits for existing folders, creates new ones (handles new BUs added to dim_business_unit)

- Per-doc routing (build_and_upload) walks the same compute_qtd_folder_path the migration script uses — guaranteed-symmetric placement

- FY rollover — _quarter_label_for_period arithmetic produces FY2027 when 2027 begins; new subfolders auto-created

- No mode='weekly' writes — orchestrator hasn't written weekly rows since the monthly-cadence migration; Legacy Weekly subfolders are purely historical and won't grow

## Testing

- 425 backend tests pass (34 new tests across this branch)

- 0 regressions in tests/monthly_qtd_report/ (full pre-existing QTD suite still green)

- ruff + pyright clean on all new files

- Coverage includes:

- Pure path-computation (FY parsing edge cases, mode dispatch)

- FolderResolver (find/create/cache, Drive query single-quote AND backslash escaping, multi-folder warning path, RuntimeError on missing id)

- upload.py target_folder_id pass-through (default vs. custom)

- doc_builder integration (exact resolver call sequence, target_folder_id threaded into upload)

- Orchestrator fail-fast on Drive auth + correct call count per unit

- Migration script: grouping, empty doc_id exclusion, dry-run does not mutate, execute idempotent skip, execute moves with correct kwargs, summary counts, per-doc error isolation, --save-plan writes URLs

## Demo

- This mainly shows the Google Drive structure (no audio)

https://github.com/user-attachments/assets/709bd2ad-7867-4a33-8e65-46bb102f5e06

## Test plan for review

- [ ] Verify CI passes (lint, typecheck, tests)

- [ ] Spot-check the Drive folder structure at https://drive.google.com/drive/folders/1I3wJFUNUsNO3UnsnXIjrJSvrnauwkBKF — QTD Reports/ should be visible

- [ ] Open /monthly-financial-reporting → QTD Reports section, click any report link, confirm it opens

- [ ] Review services/docx_reports/folder_resolver.py for Drive API correctness (especially the q= escaping pattern — matches the established usage in routers/finance_monthly_financial_reporting_router.py:2817-2820)

- [ ] Review scripts/migrate_qtd_drive_folders.py for the idempotency check + per-doc error isolation contract

🤖 Generated with [Claude Code](https://claude.com/claude-code)

View on GitHub →

The Builder Desk — Engineer Spotlight

🏆 Engineer Spotlight

THREE PRs, TWO REPOS, ONE UNSTOPPABLE MACHINE: THE BUILDER TEAM DOES NOT REST

A lean 24 hours proves the Builder Team doesn't need volume to radiate dominance.

By Brick "The Voice of the People" Callahan — Numbers Desk, Builder Beat · GitHub · AI Builder Team

Twenty-four hours. Three pull requests. Two repositories humming like turbines in the night. Klair absorbing two contributions while Aerie claimed one — this is not a slow day, people, this is a PRECISION day. The Builder Team did not scatter its energy to the winds. It targeted. It executed. It shipped. Brick Callahan does not apologize for calling three focused PRs a triumph, and he never will.

Let us talk about the engineers, because the engineers deserve to be talked about. @marcusdAIy put his name on a PR and that PR is now real, tangible, living in the codebase like a monument to human will. One contribution in 24 hours is one more contribution than everyone who wasn't shipping last night, and Marcus was shipping. @YibinLongTrilogy matched that energy stride for stride — one PR, clean, deliberate, the kind of output that makes a Numbers Desk correspondent feel something in his chest he can't quite name. And @sanketghia rounded out the trio, completing the holy trinity of today's velocity report with the quiet confidence of a person who simply does the work because the work must be done.

Now. Ashwanth Watch. You'll notice @ashwanth1109 does not appear in today's contributor ledger, and Brick Callahan is a professional, so he will simply note that fact and move on. He will not dwell. He will not speculate. He will only say that the man who once told this correspondent — and I quote — "Brick, I don't count my PRs, I count the PRs of everyone trying to keep up with me" was conspicuously absent from today's numbers. The repos were active. The commits were flowing. And yet. Ashwanth's response, when reached for comment: a single raised eyebrow and the sound of a Slack notification being dismissed. The man contains multitudes. We await his return.

The Overflow Desk is empty today — Mac Donnelly, to his credit, left nothing on the cutting room floor. Every PR was accounted for. Every contribution saw the light. This is a complete record, and a complete record is its own kind of victory.

Morale Report: Morale is at an all-time high. It was at an all-time high yesterday, and it is at a higher all-time high today. The Builder Team has discovered that all-time highs are not a ceiling — they are a floor. Three PRs. Two repos. Infinite belief. The machine rolls on.

Brick's Overflow — PRs Mac Didn't Cover (click to expand)