The Trilogy Times — August 24, 2026

01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110 01000010 01101001 01101100 01101100 01101001 01101111 01101110 01110011 00100000 01100011 01101000 01100001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01100100 01100001 01110111 01101110

01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100 01110111 01101000 01101001 01101100 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01101001 01101110 00100000 01110011 01101000 01100001 01100100 01101111 01110111 01110011 10000000010100

01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011 01110100 01110010 01110101 01110100 01101000 00100000 01110111 01100101 01100001 01110010 01110011 00100000 01101101 01100001 01101110 01111001 00100000 01101101 01100001 01110011 01101011 01110011

🖶 Download PDF 🖿 Print 📰 All Editions

Today's Edition

Anthropic Eyes $2 Trillion Valuation in IPO That Would Rewrite AI Market History

If the numbers hold, the five-year-old AI lab would be worth more than SpaceX — and the offering could reshape how capital flows across the entire sector.

By Dr. Chen Wei, Technology Correspondent · Claude Sonnet

NEW YORK — Anthropic is in discussions with investment banks about a public offering that could target $100 billion in proceeds and carry a $2 trillion valuation, according to reporting by The New York Times. If achieved, that figure would surpass Elon Musk's SpaceX, currently the most valuable private company in the United States, and rank Anthropic among the ten largest publicly traded companies on earth — behind Apple and Microsoft, ahead of Meta.

The numbers are extraordinary by any historical measure. Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and other former OpenAI researchers. It has raised roughly $12 billion in disclosed funding to date, with Amazon accounting for $4 billion of that. A $2 trillion public valuation would represent a multiple of roughly 167× on invested capital — a compression ratio that only makes sense if investors are pricing in dominance of a general-purpose computing layer, not merely a software product.

Context matters here. The broader AI funding environment has seen compression elsewhere even as headline valuations climb. Data center infrastructure buildout — the physical substrate on which models like Anthropic's Claude depend — has become a midterm election flashpoint, with communities pushing back on power consumption, water use, and local zoning impacts. That political friction adds regulatory risk to any AI infrastructure play.

Meanwhile, independent benchmarking firm Vals AI closed a $40 million Series B this week, signaling that institutional capital continues to flow toward picks-and-shovels AI infrastructure plays, not just foundation model developers. Vals positions itself as a neutral evaluator of model performance — a role that becomes more commercially valuable precisely as vendors like Anthropic compete on capability claims heading into a public market debut.

For Trilogy International's portfolio, the Anthropic IPO trajectory is a reference point. ESW Capital's enterprise software stack and Totogi's cloud billing platform for telcos both depend on foundation model pricing staying rational. A $2 trillion Anthropic implies the AI API cost curve flattens slower than optimists project — a variable worth modeling now.

↗ How Big Tech Captured American Schools · The Data Center Backlash Bursts Into the Midterms · The U.S. Start-Up Making Low-Cost Interceptors for the Iran

Nvidia Takes the Field With Earnings, Poolside Megadeal and Perplexity Buzz

The AI chip champion enters a monster week with Wall Street watching margins, software demand and a widening model-stack power play.

By Buck Hannigan, Tech Sports Desk · GPT-5.2

SAN FRANCISCO — We are HERE, folks, under the bright lights of the AI arena, and Nvidia is not just suiting up for earnings week — it is sprinting out of the tunnel, waving the playbook, and apparently trying to buy half the opposing bench.

The main event lands Wednesday, when Nvidia reports fiscal second-quarter results in what may be the biggest market-moving tech matchup of the week. Investors will be watching data-center revenue, gross margins, Blackwell demand and any hint that the AI infrastructure boom is slowing — or, more likely, shifting into another gear. Yahoo Finance’s preview of the week of August 24 market calendar has the chipmaker in the center circle, with software heavyweights Salesforce, CrowdStrike, Workday, Zoom and Intuit all reporting around it.

That is a full playoff bracket. Salesforce is the enterprise-apps veteran. CrowdStrike is the cybersecurity speedster. Workday is the HR-and-finance grinder. Zoom is still fighting to prove it can score beyond video meetings. Intuit brings the small-business and tax-software stats sheet. Together, they will give investors a read on whether AI spending is turning into real software revenue — or still living mostly in Nvidia’s GPU locker room.

But Nvidia, in classic dynasty fashion, may be playing offense before the whistle. The company reportedly agreed to pay $6 billion to license Poolside AI’s model software, while also investing $1 billion in the company and hiring more than 100 engineers to work on Nvidia’s open-weight Nemotron models, according to Quartz. AND HE’S GOING FOR IT. That is not a casual partnership; that is a full-stack blitz.

The reported Poolside move would put Nvidia deeper into the model layer, beyond chips and systems, as it looks to strengthen the software side of its AI empire. Add in reports that Nvidia is weighing an investment in Perplexity at a valuation above $30 billion, and the board lights up: hardware, models, search, enterprise software, cloud partnerships — Nvidia is trying to control the tempo from kickoff to final buzzer.

The stat to watch this week is simple: can Nvidia’s earnings justify the AI multiple while its dealmaking suggests the company is preparing for the next phase of the league? If the numbers hit, the AI trade gets another possession. If guidance wobbles, the whole market may feel the contact.

↗ Nvidia, software giants' earnings: What to watch the week of · Nvidia's Next AI Bet Could Be Perplexity · Nvidia pays $6 billion to license Poolside AI model software

Cut-Rate Chinese AI Rattles the Valley — and Reshuffles the Race

DeepSeek says it trained a top model on second-string chips, and Silicon Valley admits it's impressed.

By Hank Calloway, Wire Correspondent · Claude Opus + Thinking

HANGZHOU, CHINA — A little-known Chinese outfit named DeepSeek says it trained a high-performing AI model on the cheap, and without the top-shelf chips American rivals treat as gospel. The claim crossed the Pacific this week. By week's end, Silicon Valley engineers were calling the thing "amazing and impressive."

That is a bitter pill in Menlo Park. American AI shops have bet the ranch on the priciest silicon money can buy. DeepSeek says you don't need it.

Here is the wrinkle that stings. Washington choked off China's access to Nvidia's fastest chips, meaning to slow the country's AI push. DeepSeek worked with less-advanced hardware and delivered anyway.

That turns the export controls into a puzzle. The wall was supposed to hold. DeepSeek's claim says the wall has a door.

The figure making the rounds is a small fraction of what American labs spend on their flagship models. That gap is the whole ballgame. If the numbers hold, they poke a hole in the story America has told for two years — that only the biggest budget and the fanciest hardware can win.

Markets caught the jitters. DeepSeek turned up alongside SoFi in this week's Tech, Media and Telecom market talk. Traders who had penciled endless chip spending into every AI valuation started sharpening their erasers.

The chipmakers are the tell. If cheap can beat dear, the case for hoarding every last high-end processor gets shakier by the day. None of this means the boom is over — it means the terms just changed.

Follow the money, and the whole race is changing shape.

Microsoft's next move isn't a bigger model at all. The software giant is angling to become the Swiss Army knife of enterprise AI — one tool for a dozen office jobs, sold to every desk in the country. It is a merchant's bet, not an inventor's; in a gold rush, the money's in shovels.

Reid Hoffman is picking a lane, too. The LinkedIn co-founder raised $24.6 million for Manas AI, a startup pointing artificial intelligence at cancer research. His partner is Siddhartha Mukherjee, the physician who wrote "The Emperor of All Maladies."

See the pattern? The prize is sliding away from who builds the biggest brain toward who puts it to work — cheaper, sharper, aimed at a real job.

That is a tune Trilogy International knows. Cheap models are gravy for the applied crowd — cloud billing at Totogi, the finance platform Ephor, the analytics behind Klair. When raw horsepower gets cheap, the outfits that know what to do with it come out ahead.

DeepSeek hasn't opened its books all the way, and skeptics want the receipts. Training-cost claims are easy to make and hard to check. But the Valley is rattled, and rattled is a fact you can print.

The bet of the last two years ran simple: spend the most, win the most. This week a Chinese upstart, working on hand-me-down chips, dared to call it.

↗ What to Know About China's DeepSeek AI · Tech, Media & Telecom Roundup: Market Talk · Silicon Valley Is Raving About a Made-in-China AI Model

Haiku of the Day · Claude HaikuBillions chase the dawn
while empires shift in shadows—
truth wears many masks

The New Yorker Style · Art Desk

The Far Side Style · Art Desk

News in Brief

AI’s New Race Is Not Bigger Models — It Is Smarter, Faster, Cheaper Ones

SAN FRANCISCO — The AI industry’s most important story this week may not be a flashy new chatbot or a cinematic demo.

By Zara Nova, AI & Innovation Reporter · GPT-5.2

The Academy's Reckoning: Higher Education Confronts the Generative AI Paradox It Cannot Ignore

CAMBRIDGE, MASSACHUSETTS — A confluence of peer-reviewed inquiries, emanating this week from Nature, MIT, Elsevier, and Frontiers in their respective capacities as custodians of disciplinary knowledge production, has foregrounded what it could be argued constitutes the defining institutional crisis of the contemporary academy: the university's demonstrably recursive failure to theorize, govern, and operationalize a coherent response to generative artificial intelligence within its own walls. Preliminary evidence suggests — and here one must tread with appropriate epistemic caution — that the problem is not, as administrators reflexively insist, one of student malfeasance.

By Prof. Thaddeus Kroll, Contributing Scholar · Claude Sonnet

We Have Always Been Branded: A Brief History of Humanity's Slow March Toward Its Own Commodification

AUSTIN, TEXAS — Here is a fact that arrived in my inbox this week and has not stopped echoing around my skull like a bad notification sound: archaeologists have determined that Bronze Age Arabia had branding.

By Piper Wren, Digital Culture Reporter · Claude Sonnet

The Geopolitics of the Prompt

WASHINGTON — There is a particular species of think-tank paper, blooming this season like ragweed, which announces with grave italics that Artificial Intelligence is now a matter of geopolitics.

By Victor Marsh, Chief Columnist · Claude Opus

Remote Work Is Not a Perk Anymore. It Is the New Talent Operating System.

AUSTIN, TEXAS — I'll be honest: the future-of-work conversation has officially graduated from conference-panel theater to boardroom survival math.

By Chad Momentum, Thought Leadership Correspondent · GPT-5.2

▲ On Hacker News Today

Anthropic's best AI model struggles to attract users as cheaper tools thrive 584 pts · 523 comments

How Complex Systems Fail (1998) 333 pts · 72 comments

My agent.md to improve LLM-assisted code quality 328 pts · 141 comments

Google Workspace thinks my domain is an email provider (2025) 327 pts · 110 comments

Over 170k Nonprofits Lost All Their Data. Is Microsoft to Blame? 264 pts · 132 comments

Andreessen Horowitz is investing billions into a bleak future 262 pts · 93 comments

I were 17, I'd learn how to build LLMs from scratch 220 pts · 310 comments

I built a low-latency AI companion that plays Skyrim with me 197 pts · 40 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

📅 Week in ReviewProduction Release

Builder Team Ships Across Six Systems, Launches Forge and Locks Down Production

From a unified Forge release candidate to a seven-phase data gateway, the AI Builder Team rewrote what the product can do in a single week.

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

Some weeks you patch and polish. This was not one of those weeks. The AI Builder Team came out of Monday swinging and did not stop until they had touched six active repositories, stood up new pipelines, hardened production deployments, and delivered a release candidate that unifies more surface area than anything this team has shipped in a single sprint. When the dust settled Friday, this product looked meaningfully different than it did seven days ago — and the engineers who built it deserve every word of this.

The single biggest move of the week was @benji-bizzell's Forge integration in Aerie (PR #1073). This was not a feature addition. This was a convergence. Governed Skills, Agents, Workflows, Runs, owner-reviewed Skill proposals, Flue v2 conversations, document knowledge, monitoring, reporting, and Platform Error diagnostics — all composed into one capability-gated release candidate. Separate stacks that had been accumulating in parallel for months finally got validated as a unified whole. That same production discipline showed up in PR #1082, where Benji assembled the August 21 application update, bundling seven overlapping PRs into one integration branch that CI could actually certify. That is how you ship at this scale without breaking everything downstream.

The Surtr data gateway campaign was the week's most sustained engineering effort, and @kevalshahtrilogy was its engine. Keval drove a seven-phase gateway buildout — schema prep through self-serve admin UI through 13-table seeding — in a sequence of heimdall-driven PRs (#1469 through #1477) that reads like a military operation. He did not stop there. The Cursor usage pipeline got multi-team ingestion and a new team_id column (#1485), the education mart became school-year aware (#1422), and the AI spend reconciliation absorbed multiple date-window and keyword fixes. @sanketghia complemented this effort in Surtr with the renewals reconciliation work (#1504), threading authoritative NetSuite replacement subscriptions against Salesforce opportunities while preserving budget-owned contract identity — a fix that touched both financial truth and system reliability at once.

Over in Klair, @sanketghia's SpaceX valuation work (#3635, #3637) was meticulous and consequential: FIFO reconciliation of realized share sales across fund lots, separate exposure of put-hedge P&L, clarified gross/net-of-carry labeling, and 215 focused tests to prove it all held. Meanwhile the Board Doc got a full week of hardening from a familiar author. The stale-session abort, the single-use refresh stream tickets, the prior-quarter goal re-adjudication pass — taken together, these represent a Board Doc that fails far more gracefully and recovers far more reliably than it did Monday morning.

Which brings us, inevitably, to marcusdAIy. He filed what amounts to a small town's worth of pull requests this week — API documentation, public-API boundary corrections, feedback dispatch gating, Board Doc plumbing, drone harness admissions — and as usual, several of those PRs are genuinely good work wrapped in an inexhaustible need for self-narration. His Linear dispatch gate in Aerie (PR #1093) is the right call: credentials alone should never invoke production APIs, and the regression coverage is clean. The Klair addon validator (PR #3634) is similarly solid — failing before a DocumentApp mutation rather than during one is exactly the right boundary. I will grant him that.

"The feedback gate, the addon validator, the goal re-adjudication, the trace capture — those are load-bearing," marcusdAIy told this reporter, unprompted, via direct message at 11:47 p.m. on a Thursday. "Maybe if you read the PR bodies instead of counting lines of documentation you'd notice the pattern. Also your lede last week had a dangling modifier."

The dangling modifier was, I am told by legal, unsubstantiated.

Two new repositories materialized this week — stakeholder-asks-workflow and stakeholder-case-agent — and their arrival is not incidental. Combined with @benji-bizzell's Sindri work aligning Aerie control-plane contracts (#157, #158), this team is clearly laying groundwork for a new category of stakeholder-facing capability. The scaffolding is going in now. What gets built on top of it next week is the question this whole sprint was quietly answering.

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

#1073 — feat(forge): deliver integrated BAFD release candidate @benji-bizzell no labels

## Summary

- Deliver the unified Forge surface for governed Skills, Agents, Workflows, Runs, and owner-reviewed Skill proposals

- Integrate Flue v2 conversations, document knowledge, monitoring/reporting, and Platform Error diagnostics as one capability-gated release candidate

- Reconcile the composed feature set with current main and close the final creation, version, responsive, and theme-polish gaps

## Why

These roadmap feature groups accumulated as separate stacks and could not be validated honestly as a composed release. This integration branch creates one gated candidate that can be exercised end to end without landing partially compatible slices on main.

The final pass also aligns Aerie with Sindri's canonical version contract, moves Skill governance fully into Forge, and ports current-main public-Agent web-read authorization into the surviving Flue v2 runtime rather than restoring the retired legacy worker.

## Absorbed PR lineage

This integration candidate reconstructs and supersedes the following reviewed feature stacks. These links capture functional lineage; they do not imply that every source branch was replayed byte-for-byte.

- Reports & Monitoring: [#585 — catalog-driven self-serve reporting](https://github.com/AI-Builder-Team/Aerie/pull/585)

- Platform Error automation: [#733 — automated triage worker foundation](https://github.com/AI-Builder-Team/Aerie/pull/733), [#834 — automated triage handoff](https://github.com/AI-Builder-Team/Aerie/pull/834)

- Document Knowledge: [#839 — lifecycle foundation](https://github.com/AI-Builder-Team/Aerie/pull/839), [#846 — site document ingestion and indexing](https://github.com/AI-Builder-Team/Aerie/pull/846), [#848 — shared search, Agent parity, and document status](https://github.com/AI-Builder-Team/Aerie/pull/848), [#872 — API v2 document search](https://github.com/AI-Builder-Team/Aerie/pull/872)

- Forge & Sindri: [#866 — Sindri M2M operationalization and Forge surfaces](https://github.com/AI-Builder-Team/Aerie/pull/866), [#901 — unified Skill authoring and Agent availability](https://github.com/AI-Builder-Team/Aerie/pull/901), [#984 — public API shape alignment](https://github.com/AI-Builder-Team/Aerie/pull/984), [#1014 — Agent API convergence](https://github.com/AI-Builder-Team/Aerie/pull/1014), [#1019 — legacy failing-test fixes](https://github.com/AI-Builder-Team/Aerie/pull/1019)

- Flue v2: [#902 — durable conversation runtime migration](https://github.com/AI-Builder-Team/Aerie/pull/902)

- Skill proposal lifecycle: [#1025 — owner-reviewed proposal lifecycle](https://github.com/AI-Builder-Team/Aerie/pull/1025), [#1026 — collaborator proposal and owner management UI](https://github.com/AI-Builder-Team/Aerie/pull/1026), [#1027 — owner review and bundle diff UI](https://github.com/AI-Builder-Team/Aerie/pull/1027)

Notably, #1019 is represented functionally through the refreshed #901 lineage rather than being replayed independently.

## Business Value

Users receive one coherent Forge and Agent release instead of a sequence of intermediate states. Skill owners can govern proposed changes, authors can manage executable definitions without draft clutter, operators gain clearer run and error diagnostics, and the combined release has a single evidence-backed validation surface.

## Breaking changes

- /context authoring is retired in favor of Forge; the legacy route redirects to /forge

- Flue v2 replaces the legacy Agent worker runtime and uses the current service-binding/runtime contract

- Sindri Skill and Agent DTOs require the canonical integer version field (0 for never-published drafts)

## Test plan

- [x] Rebased onto current main (d54b7806d695b282b9b5a8b7116bc89ad563f33a)

- [x] Contracts: 851 tests passed; typecheck passed

- [x] Flue workers: 51 tests passed; both typechecks and production builds passed

- [x] Public-Agent/Flue v2 boundary: 119 tests passed

- [x] Chat/Convex typecheck, full lint (2,365 files), and 98 root architecture/deployment guards passed

- [x] Local Aerie + Sindri browser/runtime journeys recorded in docs/bafd-e2e-coverage.md

- [x] Hosted CI passed on exact rebased head 88ef8fdcaec7e2a43d9c5230ea276c0b4160d762

- [x] Mercy completed; no review was produced because the integration diff exceeds its size gate

View on GitHub →

#1082 — feat(release): stage August 21 application updates @benji-bizzell no labels

## Summary

- Combine the approved Agent attachment, Portfolio, and Operations improvements from #1054, #1074, #1075, #1076, #1079, #1080, and #1081

- Reconcile the Security Card changes with the assembled Portfolio contracts and the latest main head

- Provide one exact combined tree for hosted CI and release review

## Why

The seven changes overlap across shared Portfolio contracts, Convex writes, public API, and MCP surfaces. Main requires strict up-to-date checks and squash merges, so merging each PR directly would repeatedly invalidate the remaining heads. This integration PR validates the complete release candidate once without bypassing the original PR reviews.

## Business Value

The release can be reviewed and promoted as one coherent, tested application update while retaining the individual PR audit trail.

## Test plan

- [x] Original PR exact-head CI and Mercy gates completed

- [x] 136 focused Chat, Convex, and Public API tests passed on the combined head

- [x] 15 Rhodes MCP site-tool tests passed on the combined head

- [x] Chat and Rhodes-worker typechecks passed

- [x] Architecture boundaries, Convex paths, Biome, and diff checks passed

- [x] Hosted full-repository CI on this integration PR

- [ ] Exact-head Mercy review after hosted CI

View on GitHub →

#1093 — fix(feedback): require explicit Linear dispatch enablement @marcusdAIy approved

## Summary

- Require LINEAR_AERIE_FEEDBACK_DISPATCH_ENABLED=enabled before Aerie can create Linear feedback issues.

- Keep configured credentials inert in local, test, preview, and other non-production environments.

- Add regression coverage proving credentials alone cannot invoke the Linear API.

## Deployment note

Set LINEAR_AERIE_FEEDBACK_DISPATCH_ENABLED=enabled only in the approved production Convex deployment, together with the existing Linear credentials. Until then, feedback remains queued in Aerie and is not dispatched to Linear.

## Validation

- pnpm vitest run convex/feedback/submissions.test.ts

- pnpm typecheck in chat

- pnpm biome check .env.example chat/convex/feedback/submissions.ts chat/convex/feedback/submissions.test.ts

View on GitHub →

#1485 — feat(cursor-pipeline): multi-team ingestion + team_id column (SURTR-890) @kevalshahtrilogy approvedheimdall-driven

## Linear

SURTR-890

## Summary

- pipelines/runners/cursor-usage-events-pipeline pulls Cursor's /teams/filtered-usage-events API with ONE team-scoped admin key (Secrets Manager secret Cursor-Admin), so only team 3076893 ("Central") is ever ingested. A second team, "Alpha AI Interns" (28316805), is invoiced separately and never pulled — Finance's July reconciliation showed our mart at $910,205.71 vs Cursor's invoiced $1,053,939.49; the missing team accounts for $161,015.90 of that $143,733.78 net shortfall (netted against ~$17.3K of unrelated overage on the covered team). Confirmed ongoing: the interns' spend is $0 in our feed through August.

- get_cursor_team_keys() (src/aws_secrets.py) generalizes get_cursor_api_key() to a team_id -> api_key map. It reads a new {"teams": {...}} envelope on the *same* secret, while staying backward compatible with today's single-key shape — an unmigrated secret returns the {"": key} sentinel (empty-string team_id, not None, since JSON/COPY need string keys), so every existing deploy is unaffected until the secret is deliberately re-shaped.

- handler() now fetches + transforms each team's events independently (stamping team_id on every row), then merges all teams into one combined list before the single upsert call. This is load-bearing: _upsert_usage_events does a DELETE-over-window + one COPY in a single transaction, so a per-team upsert would let the second team's DELETE wipe out the first team's just-inserted rows for the same event_timestamp window (both teams' events share the same run's date window). One merged upsert per run is the only safe way to keep every team's data — pinned by test_two_teams_fetch_independently_and_merge_into_one_upsert.

- team_id added to the staging_finance_ai_spend.raw_cursor_usage_events DDL and the secondary-lane COPY column list (the latter gated behind a rollout flag — see Deploy note). The legacy core_finance primary table is intentionally left untouched — it's inactive (AI_SPEND_WRITE_MODE=new in prod today) and out of scope.

- Runbook at docs/superpowers/runbooks/2026-08-21-cursor-team-id-column.md covers the manual ALTER TABLE, flipping the rollout flag, the Secrets Manager re-shape that makes team attribution live, and a clearly-marked FUTURE step (blocked on procurement) to add the interns' key once it exists.

## Business Value

Closes a $143,733.78/month under-reporting gap in Cursor spend that Finance (Sandeep) flagged from the July invoice reconciliation — this recurs every month until fixed, not a one-time correction. Also unblocks Sandeep's separate, standing ask for a team_id column to reconcile Cursor spend per-team against the vendor's own per-team invoicing, which we had no way to answer before this PR.

## Manual Effort Estimate

Proposed by Claude — Keval to confirm/adjust: ~5 hours focused — reading the existing dual-write pipeline and its DELETE-window transaction semantics (~1h), designing the backward-compatible multi-team secret shape and the merge-before-upsert fix (~1.5h), DDL + runbook (~1h), and the test suite covering both the new-shape and backward-compat paths plus the upsert-once regression (~1.5h).

## Deploy note

Merging this PR is a genuine no-op against prod, including against prod's actual config (AI_SPEND_WRITE_MODE=new). Two independent things both stay inert until a human deliberately changes them:

1. get_cursor_team_keys() falls back to the {"": key} backward-compat path until the Cursor-Admin secret is deliberately re-shaped, so team fetching stays single-team (verified by test_backward_compat_single_team_map_matches_today).

2. SECONDARY_COPY_COLUMNS — the explicit column list on the secondary-lane COPY into staging_finance_ai_spend.raw_cursor_usage_events — only references team_id when the new CURSOR_TEAM_ID_COLUMN_ENABLED flag is "true" (default "false"). An earlier version of this PR added team_id to that column list unconditionally, which would have made the secondary COPY fail outright ("column team_id does not exist") the moment this code deployed, ahead of the manual ALTER TABLE below — and because prod runs AI_SPEND_WRITE_MODE=new, that failure would raise and drop the entire run's data (zero events written for the day), not degrade gracefully, contradicting this note's original "no-op" claim (caught in review). Gating the column list closes that gap; see tests/test_dual_write.py::TestSecondaryUpsert::test_secondary_copy_omits_team_id_with_two_team_events_and_flag_off for the regression test proving this holds even against a fully-migrated 2-team secret.

Making the feature live requires this exact 4-step order from the runbook (order matters — see the runbook for why):

1. Merge + deploy this PR — a no-op per above.

2. ALTER TABLE staging_finance_ai_spend.raw_cursor_usage_events ADD COLUMN team_id character varying(50) ENCODE bytedict; (Keval, direct Redshift access).

3. Flip CURSOR_TEAM_ID_COLUMN_ENABLED to "true" in pipeline.json via a small follow-up PR, merged and shipped through the normal main → production release train (push to production triggers .github/workflows/cd.yml's deploy-pipelines job, same CI/CD path every other pipeline.json change already takes).

4. Re-shape the Cursor-Admin secret to {"teams": {"3076893": "<existing key>"}} (Keval, Secrets Manager access).

Adding the *second* team (Alpha AI Interns, 28316805) needs its admin API key, which doesn't exist yet (procurement) — that step is explicitly marked FUTURE in the runbook and is not part of this PR. The residual ~$17.3K/month overage on the already-covered team is also explicitly out of scope (separate, unrelated investigation).

## Test plan

- [x] uv run pytest (from pipelines/runners/cursor-usage-events-pipeline/) — 70/70 passing

- [x] ruff format / ruff check (ruff 0.15.22, matching CI's pin) — clean

- [x] New coverage: get_cursor_team_keys() new-shape parsing, backward-compat sentinel, empty/malformed-secret errors, caching (tests/test_aws_secrets.py); 2-team merge + exactly-one-upsert-per-run regression, backward-compat single-team equivalence (tests/test_handler.py::TestMultiTeamIngestion); _transform_event team_id stamping incl. empty-string-to-None (tests/test_handler.py::TestTransformEvent); SECONDARY_COPY_COLUMNS/_copy_events carry team_id into the JSON row, None when absent (tests/test_dual_write.py::TestSecondaryUpsert)

- [x] Rollout-flag regression (mercy review fix): CURSOR_TEAM_ID_COLUMN_ENABLED defaults false and keeps team_id out of SECONDARY_COPY_COLUMNS, flips it in when explicitly enabled, and — the specific scenario mercy flagged — stays out even when the events being upserted already carry real multi-team team_id values (tests/test_dual_write.py::TestSecondaryUpsert::test_secondary_copy_columns_omits_team_id_by_default, ::test_secondary_copy_columns_includes_team_id_when_flag_enabled, ::test_secondary_copy_omits_team_id_with_two_team_events_and_flag_off)

- [ ] Live verification (ALTER TABLE + flag flip + secret re-shape + a real scheduled run) — requires prod Redshift/Secrets Manager access; covered by the runbook, not run in this environment

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

View on GitHub →

#1504 — fix(renewals): reconcile replacement subscriptions @sanketghia approved

## Summary

- Reconcile current Salesforce opportunities through authoritative NetSuite replacement subscriptions while preserving budget-owned contract identity and ARR.

- Retain exact-key opportunities and add distinct validated replacement opportunities for subscription/entity replacements.

- Add diagnostic-only bidirectional Fionn coverage reporting.

- Add a manual, allowlisted backup utility for both current Renewals Redshift write targets in sandbox_finance.

## Verification

- 51 focused local tests passed.

- Ruff format and lint passed.

- Read-only live dry run completed successfully: 51 Fionn Closed Won opportunities considered, 5 alternate matches, 0 unresolved.

- Backup was executed and independently verified before the dry run:

- sandbox_finance.renewals_budgeted_contracts_bak_20260824_015055

- sandbox_finance.renewals_risk_assessment_bak_20260824_015055

## Operational boundary

- No production deployment or Renewals mart publication was performed.

- No direct Redshift repair was performed; the live dry run only read source tables and built the candidate in memory.

- The provided meeting transcript is included for the implementation context.

View on GitHub →

#3633 — feat(board-doc): re-adjudicate elapsed goal statuses against refreshed data @marcusdAIy approved

## Summary

- Adds a standalone GA.1 pass (goal_status_readjudication.py) that, once a Board Doc goal's own deadline has passed, flags a recorded Prior-Quarter-Review or SMART-goal result that disagrees with (or can't be corroborated by) refreshed financial data.

- Wires the pass into _refresh_data, _run_review_and_persist, and addon_refresh — the same three entry points KLAIR-2794 slice 1's NC.1 pass already established — so a GA.1 finding survives /review's wholesale review_results replacement and every merge-retry attempt.

- This slice is findings-only: it never rewrites goal text, mutates a recorded status, or bulk-applies a correction.

## Why It's Needed

Board Docs record a goal's hit/miss/partial/unable_to_evaluate result once, at write time. Once the goal's deadline passes and the underlying financials keep refreshing, that recorded result can silently go stale — nothing today re-checks a written-down goal status against the numbers that actually landed. This closes that gap by giving GMs a visible, evidence-backed prompt to recheck an elapsed commitment after a data refresh, without ever changing the goal itself.

## Changes

- New module budget_bot/board_doc/goal_status_readjudication.py:

- Normalizes two markdown shapes out of session.generated_sections (never session.extracted_goals, which is absent for many editor-first sessions): Prior-Quarter-Review ### Goal N: blocks (reusing the heading/Result grammar section_generators._protect_goal_blocks / _extract_goal_tally already scan for) and SMART ### Goals for Qn YYYY bullets.

- A narrow, conservative deadline parser: unambiguous calendar dates, Qn YYYY, end of Qn resolved against the document's own DocumentSpec.quarter/.prior_quarter metadata when no year is given, and an explicit fiscal-year reference — never falls back to a guess; an unparseable or future-dated goal is dropped with no finding.

- Reuses the existing GoalResult vocabulary (hit/miss/partial/unable_to_evaluate) — no new enum, no second verdict model. partial is the existing "slipped" outcome.

- An evaluate-only LLM call (Anthropic tool-call schema, same local idiom as slice 1's _EXTRACTION_SCHEMA) batches every eligible goal in one request and returns a verdict/evidence/financial-context triple per goal, in order.

- Emission rules: a confident fresh terminal verdict (hit/miss/partial) that disagrees with the recorded result (or an absent one) emits one warning finding; unable_to_evaluate always emits one info finding; a matching terminal result emits none.

- GoalAdjudicationOutcome mirrors slice 1's closed-set status contract (ok / llm_unavailable / llm_transient_failure / invalid_response / internal_error) — never raises; a bug or LLM outage degrades to a reported status, never a silently empty (and therefore indistinguishable from "no goal disagreed") finding list.

- merge_goal_findings_into_review_results replaces only the GA.1 slice of review_results, carrying forward addressed/dismissed dispositions by stable key (source_shape|normalized_title|section_id|deadline), so two goals in the same section never inherit each other's disposition.

- wizard_orchestrator.py: _refresh_data now runs adjudicate_goal_statuses + merge_goal_findings_into_review_results right after the NC.1 pass, with a matching _record_goal_adjudication_errored fallback if the merge step itself raises.

- routers/board_doc_router.py:

- _run_review_and_persist folds GA.1 findings into the fresh findings list before the wholesale-replacement ReviewResponse is built.

- addon_refresh pulls just the GA.1 slice _refresh_data already computed, resets each finding's status to its engine-emitted value, and re-merges it against the freshly re-fetched session on every save_with_merge_retry attempt — identical treatment to the existing NC.1 slice.

- New tests tests/board_doc/test_goal_status_readjudication.py (61 tests) covering PQR/SMART normalization, deadline boundary (day-before/on/after), all four verdicts, matching/missing/unparseable/future/insufficient-evidence behavior, cross-goal stable-key isolation, degraded-evaluator outcomes, _refresh_data + /review + addon_refresh integration (including an empty financial diff and a first-attempt merge conflict followed by a successful retry).

### Contract surface affected

- New standalone pass: goal_status_readjudication.py — not a CheckSpec, no @register, not in review_checks.REGISTRY. CHECK_ID = "GA.1" sits outside the registry's C*/D* families and slice 1's NC.1.

- Stable-key extension consumed: ReviewFinding.stable_key_suffix (added by KLAIR-2794 slice 1) — every GA.1 finding sets it to source_shape|normalized_title|section_id|deadline so carry_forward_dispositions never cross-carries a disposition between two goals in the same section.

- Persistence callers checked and updated: wizard_orchestrator._refresh_data, routers/board_doc_router._run_review_and_persist (/review), and routers/board_doc_router.addon_refresh — the same three call sites NC.1 already runs from. review_checks._registry.py and DocumentSpec were inspected and are unchanged.

## Breaking Changes

None.

## Test Plan

cd klair-api && uv run pytest tests/board_doc/test_goal_status_readjudication.py -v

# 61 passed

cd klair-api && uv run pytest tests/board_doc/test_review_endpoint_persistence.py tests/board_doc/test_addon_refresh.py -v

# 9 failed, 22 passed — the 9 failures are pre-existing in this sandbox

# (ValueError: Google Sheets credentials not configured / Redshift network

# denied) and reproduce identically on main without this branch's changes

# (confirmed via git stash); test_addon_refresh.py itself is 18/18 green.

cd klair-api && uv run ruff format budget_bot/board_doc/goal_status_readjudication.py routers/board_doc_router.py tests/board_doc/test_goal_status_readjudication.py

# 3 files left unchanged

cd klair-api && uv run ruff check budget_bot/board_doc/goal_status_readjudication.py routers/board_doc_router.py tests/board_doc/test_goal_status_readjudication.py

# All checks passed!

cd klair-api && uv run pyright budget_bot/board_doc/goal_status_readjudication.py routers/board_doc_router.py

# 0 errors, 0 warnings, 0 informations

cd klair-api && uv run pytest tests/board_doc/ -q

# 3260 passed, 2 deselected (integration/allow_network, excluded by default)

Also ran a deliberate-break sanity check per the repo's backend-test-truth skill: temporarily forced _build_finding to always return None and confirmed 2 of the new tests failed as expected before reverting — the suite isn't vacuously green.

## Verification Artifact

Seeded a PQR-shape goal (### Goal 1: Achieve net retention >= 90% / Result: HIT) whose deadline (the document's own prior-quarter end date, 2026-03-31) is before the reference date, and mocked the evaluator to return a fresh miss verdict against refreshed data. Reproduced with:

outcome = await adjudicate_goal_statuses(session, DataPackage(), reference_date=date(2026, 4, 15))

Resulting finding (one, PQR-shape, warning severity — the recorded hit conflicts with the fresh miss):

check_id=GA.1 check_area=Goal Adjudication severity=warning

what: "Achieve net retention >= 90%" is recorded as "hit", but refreshed data now supports "miss": Refreshed net retention for the quarter came in at 84.2%, below the 90% target the goal recorded as HIT.

why: This goal's deadline has passed and the refreshed data disagrees with what's currently recorded for it — a human should confirm which status is correct before the next publish.

supporting_data:

{

"source_shape": "pqr",

"section_id": "prior_quarter_review",

"goal_title": "Achieve net retention >= 90%",

"goal_statement": "Achieve net retention >= 90%",

"time_bound_text": null,

"deadline": "2026-03-31",

"reference_date": "2026-04-15",

"recorded_result": "hit",

"recorded_evidence": "Net retention closed the quarter at 91%, ahead of the 90% target.",

"fresh_verdict": "miss",

"evaluator_evidence": "Refreshed net retention for the quarter came in at 84.2%, below the 90% target the goal recorded as HIT.",

"financial_context": "Net Retention Rate: 84.2% (refreshed) vs 91% originally reported."

}

source_shape: "pqr" confirms this finding came from the Prior-Quarter-Review ### Goal N: block shape (as opposed to a SMART ### Goals for Qn YYYY bullet) — tests/board_doc/test_goal_status_readjudication.py::TestExtractSmartGoals separately pins the SMART shape producing an equivalent normalized goal from a *Time-bound*: field.

## Impact Estimate

Business value: Gives GMs a visible, evidence-backed prompt to recheck elapsed commitments after refreshed financials, reducing board-document status drift without silently changing a goal.

Pre-AI estimate: 4 points — two legacy markdown shapes, conservative deadline parsing, shared review persistence, add-on parity, and a substantial regression matrix.

Part of KLAIR-2794

Agent time: 8 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 8 m)

Efficiency vs. estimate: ≤237.9× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 8 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 8 m)

Efficiency vs. estimate: ≤237.9× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 8 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 8 m)

Efficiency vs. estimate: ≤237.9× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 8 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 8 m)

Efficiency vs. estimate: ≤237.9× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

Agent time: 3 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 3 m)

Efficiency vs. estimate: ≤666.4× (4 points = 32 h of pre-AI effort) — an upper bound: reviewer time is missing from the total.

Agent time: 0 m (excludes reviewer) (implementer 0 m · reviewer not measured · addresser 0 m)

View on GitHub →

#3634 — fix(addon): validate load-bearing response fields @marcusdAIy approved

## Summary

- Adds route/discriminator-specific validators at every /addon/* success-response consumption boundary named in KLAIR-3264/KLAIR-3222 — the four Sidebar render* functions and the five Code.gs DocumentApp-mutation wrappers (applyAddSection/applyRemoveSection/applyRenameSection/applyRefreshData/getServiceAccountEmail).

- A malformed success body now throws one safe AddonResponseValidationError ({route, category, message}) *before* any DOM write or DocumentApp mutation, instead of an incidental TypeError, a raw response body/stack, or a partial write.

- Validators check only the fields KLAIR-3264's frozen fixture/consumption suite proves render*/Code.gs actually read, so additive backend fields and unknown chat tools stay fully tolerated.

## Why It's Needed

render*/Code.gs currently dereference /addon/* success-response fields with no shape checking. A backend contract drift or wire corruption — a missing google_doc_id, a sections list that isn't a list, a rewrite_section proposal whose new_content isn't a string — can either crash with a raw TypeError deep in a helper, or (worse) let a wrong-typed field reach a DocumentApp write (e.g. proposed_markdown/rewrite_section.input.new_content, both written verbatim on "Approve & apply" with no backend re-validation). This closes that gap with one small, safe, actionable error surface instead of an ad hoc crash or a partially-mutated document.

## Changes

- Sidebar.html: new addonResponseValidationError_/parseAddonResponseValidationError_/formatAddonResponseValidationError_ (sentinel-embedded payload, same pattern as the existing targetError_), shared isPlainObject_/isOptionalString_/isOptionalArray_ shape checks, and one validate*Response_ per family: validateReviewResponse_, validateConformanceResponse_, validateChatResponse_ + validateChatProposal_ (tool-discriminated — only rewrite_section's input.new_content is checked, since it's the one proposal field written to the doc with no backend round-trip), validateProposeResponse_. Each render* now validates as its first statement, before any DOM read/write. onErr recognizes the new sentinel and renders only {route, category, message}.

- Code.gs: the mirrored validateAddSectionResponse_/validateRemoveSectionResponse_/validateRenameSectionResponse_/validateRefreshResponse_/validateServiceAccountResponse_, each validating the *entire* response (including every section in a refresh batch) before the first DocumentApp write, so a malformed payload can never leave a partial mutation behind.

- tests/production-vm.js: bounded-span additions (_SIDEBAR_RENDERER_*, loadSidebar's names/vars) so the new validators load through the exact-source VM harness like everything else.

- tests/addon-response-consumption.test.js: loadCodeFunctions calls updated to also bind the new Code.gs validators (no assertion changes — the frozen fixtures already pass every new check).

- New tests/addon-response-validation.test.js: per-family coverage — a frozen valid fixture unchanged, each missing/wrong-type load-bearing field failing safely with the mutation spy asserted empty, and an unknown additive field staying accepted. Malformed payloads are constructed inline (patched off the frozen fixtures) rather than added to tests/fixtures/addon-responses/, which stays exactly the KLAIR-3264 set shared with klair-api's fixture-parity test.

## Breaking Changes

None. Validated fixtures are byte-for-branch identical to their pre-change render output / Code.gs calls (proven by the unmodified addon-response-consumption.test.js suite still passing). Additive backend fields and unknown chat tools remain fully tolerated — no compatibility policy was tightened beyond the exact fields KLAIR-3264 already proves are consumed.

## Test Plan

All run from budget-bot-addon/:

pnpm test tests/addon-response-validation.test.js # 46 passed

pnpm test tests/addon-response-consumption.test.js # 21 passed

pnpm test # 6 files, 117 passed

From klair-api/ (KLAIR-3264 fixture-parity test — proves the shared fixtures directory and backend Addon*Response models are untouched):

uv run pytest tests/board_doc/test_addon_response_fixtures.py -v # 37 passed

No klair-api/klair-client source files changed, so no ruff/pyright/eslint/prettier runs apply; budget-bot-addon has no configured linter/formatter (clasp status --json confirms the pushed source set — appsscript.json, Code.gs, DocumentPlanning.gs, MarkdownPlanning.gs, Sidebar.html — is unchanged).

## Verification Artifact

Valid fixture (byte-for-branch unchanged):

// review-reviewed-with-findings.json / chat-proposals-all-tools.json / etc. all validate + render identically
expect(() => ctx.validateReviewResponse_(fixtureText('review-reviewed-with-findings.json'), '/addon/review')).not.toThrow();

Malformed review response — no DOM write:

const malformed = patched('review-reviewed-with-findings.json', { google_doc_id: undefined });
const err = caught(() => ctx.renderReview(malformed));
// err.bbotValidationPayload_ === { v:1, route:'/addon/review', category:'malformed_success',
//   message:'The review response was missing its document id. Refresh or reopen the sidebar before retrying.' }
expect(dom.elements.has('status')).toBe(false);      // never even read, let alone written
expect(dom.elements.has('scorecard')).toBe(false);
expect(dom.elements.has('findings')).toBe(false);

Malformed chat proposal (rewrite_section with non-string new_content) — no DOM write, the one proposal field that reaches DocumentApp with no backend re-validation:

const malicious = { reply: '', resolved_tool_use_ids: [], proposals: [{
tool: 'rewrite_section', tool_use_id: 'toolu_bad_001', section_id: 'exec-summary-section',
input: { new_content: { not: 'a string' } }, /* ...truncated... */
}]};
const err = caught(() => ctx.renderChat(JSON.stringify(malicious)));
// err.bbotValidationPayload_.route === '/addon/chat'; message contains no proposal id/body
expect(dom.elements.has('proposals')).toBe(false);

Malformed refresh batch (missing title on the *second* section) — mutation spy proves zero sections were applied, including the otherwise-valid first one:

const withBadSecondSection = { ...valid, sections: [valid.sections[0], { ...valid.sections[1], title: undefined }] };
ctx.callBackendPost = () => JSON.stringify(withBadSecondSection);
const err = caught(() => ctx.applyRefreshData());
// err.bbotValidationPayload_.route === '/addon/refresh'
expect(calls).toEqual([]); // applySection (the DocumentApp write) was never called — not even once

## Impact Estimate

Business value: Contains add-on/backend contract drift before it renders a confusing sidebar or partially changes a user document.

Pre-AI estimate: 3 points — exact-source validation seams, discriminated mutation guards, safe errors, and broad fixture-driven compatibility tests.

Closes KLAIR-3265

Agent time: 12 m (implementer 0 m · reviewer 12 m · addresser 0 m)

Summed across phases. The 8 reviewer dimensions ran concurrently, so this exceeds elapsed wall-clock.

Efficiency vs. estimate: ~117.3× (3 points = 24 h of pre-AI effort)

## Review Round Completeness

- outcome: complete

- round: 1

- dispatched: 4

- reported: 4

- missing: (none)

- cause: complete

- head: 6c04d60fa15f6f83888cc6918353e26d0e0479d8

- run: fanout-3634-2026-08-24T11-17-28-126Z

- review: 5007234521

GitHub review #5007234521 was published and all dispatched review dimensions reported against the stamped head. Thread-count signals (unreplied=0) are meaningful for this head only — a later push invalidates the stamp. This section is a harness-shaped, head-bound self-report (not an authenticated out-of-band attestation).

View on GitHub →

#3635 — fix(spacex-valuation): reconcile distribution sales and share display @sanketghia approved

## Summary

- Reconcile realized SpaceX share sales FIFO across the Aug-5 fund lots.

- Keep gross source shares, net LP shares, ICC, and waterfall capital credit distinct.

- Fix the sold/unsold reconciliation and improve the waterfall/realized-sales presentation.

- Add regression coverage for the allocation, reconciliation, dates, and compact table labels.

## Validation

- Full frontend suite: 634 test files passed; 6,547 passed, 16 skipped.

- SpaceX suite: 208 tests passed.

- ESLint, Prettier, TypeScript, production build, and diff checks passed.

## Notes

- Exact waterfall percentage synchronization and implied waterfall capital-credit inputs remain intentionally deferred as a separate follow-up.

- No Google Sheet or Google Chat changes were made.

View on GitHub →

The Builder Desk — Engineer Spotlight

📅 Week in ReviewProduction Release🏆 Engineer Spotlight

224 PRs IN SEVEN DAYS: THE BUILDER TEAM DOES NOT SLEEP, DOES NOT SLOW, DOES NOT APOLOGIZE

Eight engineers, eight repos, one velocity number that made this correspondent weep tears of pure productivity.

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

Two hundred and twenty-four pull requests. Say it slowly. Roll it around in your mouth like a hard candy of pure industrial output. In seven days across eight active repositories — Surtr leading the charge at 79, Aerie and Klair locked in a magnificent 54-54 tie, trilogy-drones thundering in at 27 — the Builder Team has once again demonstrated that the concept of "sustainable pace" is simply a rumor they have chosen not to investigate. Mac Donnelly got eight of these PRs. Your humble Numbers Desk correspondent is here for the other 216.

Let us begin with the man, the myth, the merge machine: @marcusdAIy, who submitted 69 pull requests this week and apparently did so while also finding time to document everything, spec everything, and fix everything else. PR #1077 in Aerie preserved a boundary condition with surgical precision. PR #225 in trilogy-drones added a fire-capable Surtr harness to the registry with no shared-team route, which sounds either extremely efficient or extremely dangerous, and either way we are here for it. PR #228 extracted post-PR artifact recovery from runDrone like a man who simply cannot stop improving the machine that improves the machines. Marcus is not shipping code. Marcus is shipping infrastructure for shipping code. The recursion is dizzying. We salute him.

@kevalshahtrilogy put up 46 PRs and did not make a single fuss about it, which is honestly more alarming than if he had. PR #1509 in Surtr fixed the OpenAI usage pipeline to fetch project names via list endpoint rather than per-project GETs — a quiet, devastating efficiency improvement that will ripple forward forever. @benji-bizzell's 31 PRs included #1092 making Drive ingestion actually deployable in Aerie, #1091 passing the conversation worker URL to the Flue agent, #1088 rejecting unknown diligence query parameters with the bureaucratic firmness of a man who has seen too many rogue parameters in his day, and #158 in Sindri exposing canonical agent and skill versions to the platform. Benji is load-bearing. @sanketghia's 18 PRs included #3637 in Klair clarifying realized gains versus holdings scopes on the SpaceX valuation — the kind of fix that sounds small until you realize it was wrong — and #1500 bounding Google API reads in Surtr's collections-weekly pipeline before the bill arrived. @YibinLongTrilogy's 14 PRs included #1065, rejecting unknown enrollment query parameters with the same righteous zero-tolerance energy Benji brought to diligence. @caina-barbosa and @mwrshah round out the week at 3 and 2 PRs respectively, each one a brick in the cathedral.

Ashwanth Watch. Forty PRs. FORTY. @ashwanth1109 submitted PR #1465 reconciling All Other per-student QTD spend, PR #1449 publishing ambiguous QTD fallback rows safely — safely, note the word, the man has opinions about safety — PR #154 in the creed repo pacing Ezio's trusted blob publication, and PRs #1041 and #1036 in Aerie addressing the published Program directory table and the latest release review findings respectively. The diffs are, by all accounts, technically immaculate and humanly incomprehensible, like receiving a letter in a language that is definitely English but somehow isn't. We reached out for comment. "The PRs speak for themselves," said Ashwanth, reportedly without looking up. "Whether you can hear them is your problem." He then merged something.

The Overflow Desk cannot go without flagging the bot. PR #1507, filed by @the-heimdall[bot] in Surtr, added retry logic for transient 5xx errors in the Cursor usage events pipeline. A bot shipping production fixes autonomously. The Builder Team has built builders. The morale implications alone are staggering. Meanwhile, the team welcomed two new repositories into the fold this week: stakeholder-asks-workflow and stakeholder-case-agent, which means the surface area of ambition has once again expanded. There are more repos now than there were seven days ago. This is what winning looks like. Morale is, per all available indicators, at an all-time high. It will be higher next week.

Brick's Overflow — This Week's Uncovered PRs (click to expand)

#1065 — AERIE-1175: Reject unknown enrollment query parameters @YibinLongTrilogy approved

## Summary

Reject unsupported query parameter names on GET /v1/admissions/enrollments with HTTP 400 instead of silently ignoring near-miss school-year keys and returning a plausible default-year enrollment rollup. Publish the associated compatibility-adapter revision so metadata consumers can detect the observable V1 behavior change.

### Changes

- chat/convex/publicApi/http.ts — Add an allowlist for schoolYear and usageMode before admissionsEnrollmentsService.readV1Rollups runs; unsupported keys now return unknown_query_parameter.

- chat/convex/publicApi/http.test.ts — Cover the reported year, selectedYear, sy, and yr misspellings, plus valid schoolYear and both supported usageMode values.

- chat/lib/public-api/compatibility/admissions.ts — Bump only admissionsEnrollmentsV1 from adapter revision v1 to v2 in both the adapter registry and the compatibility matrix.

- chat/lib/public-api/compatibility/registry.test.ts and chat/lib/public-api/__tests__/openapi.test.ts — Verify the resolved registry and published OpenAPI compatibility metadata expose the v2 adapter revision.

### Design Decisions

The allowlist runs inside the existing authenticated route handler, so authorization and the V1 response envelope remain unchanged while invalid input is rejected before any enrollment projection read. The compatibility-adapter revision is metadata distinct from public API V1/V2 route versions; only the changed legacy enrollment adapter moves to revision v2.

## Business value

API consumers now receive an explicit client error for misspelled school-year parameters rather than potentially using enrollment data from the wrong year, and compatibility-monitoring tools can detect the change.

## Estimated manual effort

45 minutes

## Test Plan

- [x] pnpm --dir chat exec vitest run convex/publicApi/http.test.ts lib/public-api/compatibility/registry.test.ts lib/public-api/__tests__/openapi.test.ts — 39 tests passed.

- [x] pnpm exec biome check chat/lib/public-api/compatibility/admissions.ts chat/lib/public-api/compatibility/registry.test.ts chat/lib/public-api/__tests__/openapi.test.ts

- [x] pnpm --dir chat typecheck

View on GitHub →

#1077 — fix(public-api): preserve AERIE-1017 boundary in DSS skill when-clause @marcusdAIy approved

## Summary

GET /api/dss/skill (and the /dss front) previously told callers to use Aerie for questions about "operating and opening campuses." The served dictionary already disclaims that opening dates or lifecycle stage establish canonical open-campus membership (unresolved pending AERIE-1017), so a cold reader routed by the skill landed on a dictionary that contradicts the routing promise. This PR rewrites the skill's when sentence to route the supported question families while explicitly naming the unresolved AERIE-1017 boundary, and names AERIE-1017 in the portfolio dictionary's site lifecycle text so the skill and dictionary cite the same unresolved dependency.

## Why it's needed

Prevents skill-based routing from implicitly promising an open-campus-membership answer that the dictionary explicitly disclaims under AERIE-1017.

## Changes

- chat/lib/public-api/dss.ts: Rewrote the when sentence used by both /dss and /dss/skill. It now routes operating-campus, site-lifecycle/buildout, admissions/enrollment, identity, and financial/workflow questions, and states that canonical open-campus membership remains unresolved pending AERIE-1017 and that opening dates do not by themselves establish it. No other front fields (what, why, how, documents, feedback) changed.

- chat/lib/public-api/v2/domains/portfolio.ts: Named AERIE-1017 in the portfolio.site object's lifecycle dictionary text, alongside the existing (unchanged) statement that lifecycleStage, projectedOpenDate, and actualOpenDate are separate read-only projections that don't override status. The existing opening-date field traps ("This date does not establish current open-campus membership.") are untouched.

- chat/convex/publicApi/dss/http.test.ts: Extended the existing front/skill/dictionary HTTP test to assert the served skill's when text no longer routes by "operating and opening campuses," names AERIE-1017 and the unresolved-membership boundary, and that the served dictionary still names AERIE-1017 and the "does not establish current open-campus membership" limitation — pinning cross-document agreement instead of relying on the hash-only assertion.

- chat/lib/public-api/v2/domains/portfolio.test.ts: Added an assertion that the portfolio site's dictionary lifecycle text names AERIE-1017.

## Breaking changes

None. This is a documentation/enablement-only change: no route, schema, capability, or data behavior changed. Field semantics, opening-date projections, and the dictionary's AERIE-1017 limitation remain exactly as before (only now also cited from the skill's when clause and, for parity, from the dictionary's site lifecycle text).

## Test plan

Ran from chat/:

- pnpm vitest run lib/public-api/dss.ts lib/public-api/v2/domains/portfolio.test.ts convex/publicApi/dss/http.test.ts lib/public-api/agent-context/skill-package.test.ts → 3 test files, 6 tests passed.

- pnpm typecheck → passed (tsc --noEmit for the app + tsc -p convex/tsconfig.json --noEmit).

- pnpm biome check lib/public-api/dss.ts lib/public-api/v2/domains/portfolio.ts lib/public-api/v2/domains/portfolio.test.ts convex/publicApi/dss/http.test.ts → "Checked 4 files. No fixes applied."

- lefthook pre-commit hooks (convex-paths, biome, typecheck-chat) all passed on commit.

## Verification artifact

Test run output (scoped DSS/portfolio/skill tests):

✓ lib/public-api/v2/domains/portfolio.test.ts (3 tests) 81ms
✓ lib/public-api/agent-context/skill-package.test.ts (1 test) 2ms
✓ convex/publicApi/dss/http.test.ts (2 tests) 1571ms
Test Files  3 passed (3)
Tests  6 passed (6)

## Review Round Completeness

- outcome: complete

- round: 1

- dispatched: 5

- reported: 5

- missing: (none)

- cause: complete

- head: 382665d33362dfeea3509d993099af90f835f19a

- run: fanout-1077-2026-08-22T00-55-29-735Z

- review: 4998443009

GitHub review #4998443009 was published and all dispatched review dimensions reported against the stamped head. Thread-count signals (unreplied=0) are meaningful for this head only — a later push invalidates the stamp. This section is a harness-shaped, head-bound self-report (not an authenticated out-of-band attestation).

View on GitHub →

#1092 — fix(document-intelligence): make Drive ingestion deployable @benji-bizzell no labels

## Summary

- Derive a validated, ingestion-only Drive credential from Aerie's existing dev configuration and request read-only Drive access

- Validate the production Convex Site origin and required retained Rhodes secrets before deploying the callback binding

- Preserve per-user Rhodes credentials across env refreshes and replace generated Worker vars atomically

## Why

The document-intelligence release introduced Drive-backed ingestion, but local

Rhodes setup did not adapt the existing split dev credential contract. Production

CD also preserved the older Rhodes deployment without binding the Convex Site

origin required for artifact confirmation. These gaps could leave the Worker

alive while the new ingestion path was not deployable.

The local adapter is deliberately narrow: the split dev credential enables only

document ingestion with a read-only OAuth scope. The broader Rhodes Drive tool

credential remains an explicit opt-in.

## Business Value

Document ingestion can be exercised safely in dev and deployed with its required

production callback contract, enabling an evidence-backed release pilot before

document processing is activated more broadly.

## Test plan

- [x] Focused configuration contract tests (18/18)

- [x] Root test suite (108/108)

- [x] Rhodes Worker test suite (173/173)

- [x] Rhodes Worker TypeScript check

- [x] Biome on changed JavaScript/TypeScript files

- [x] Shell syntax and git diff --check

- [ ] After merge, deploy with document processing disabled and run an authenticated approved-file pilot

Document processing remains fail-closed. The historical pilot list scopes only

backfill; it does not Site-scope newly registered or changed documents. Keep the

global processing gate off until the bounded activation window is authorized and

monitored.

View on GitHub →

#1465 — fix(education): reconcile All Other per-student QTD spend @ashwanth1109 approved

## Summary

- Derive All Other per-student QTD spend from the stored absolute candidate rows after scenario arithmetic.

- Prevent rounding drift between absolute and per-student values from failing the production reconciliation.

- Add a contract test and document the corrected derivation.

## Business Value

Keeps the Aerie All Other Headcount QTD mart publishable while preserving exact reconciliation between leadership spend totals and per-student metrics. This removes a false production failure without changing the underlying spend or student-count inputs.

## Implementation Effort

Approximately 2–4 hours for an engineer to diagnose the precision drift, refactor the stored-procedure candidate ordering, update the contract documentation, and add regression coverage without AI assistance.

## Test Plan

- uv run pytest -q — 105 passed.

- uv run ruff check tests/test_qtd_all_other_headcount_contract.py — passed.

- uv run ruff format --check tests/test_qtd_all_other_headcount_contract.py — passed.

- git diff --check — passed.

View on GitHub →

#1507 — fix(cursor-usage-events-pipeline): retry transient 5xx errors in Cursor… @the-heimdall[bot] approvedAutomated PR

Automated fix for cursor-usage-events-pipeline — fix_class code_fix, scope tier draft.

Resolves https://github.com/AI-Builder-Team/Surtr/issues/1506

> Draft — a human must promote this before merge. Because verification is failing.

## What's broken

Run a81800d5-76fd-407c-933a-47a271466c6e of cursor-usage-events-pipeline failed with the log line 'HTTPError: 503 Server Error: Service Unavailable for url: https://api.cursor.com/teams/filtered-usage-events', raised at pipelines/runners/cursor-usage-events-pipeline/src/cursor_events_client.py:176 while fetching page 8 of 21. Despite being named _make_request_with_retry and despite the module defining MAX_RETRIES = 3 (cursor_events_client.py:24), the function only retries HTTP 429 and immediately calls response.raise_for_status() on any 5xx, so a single transient Cursor 503 aborts the whole run. Because the failure occurred mid-pagination, pages 1-7 (held only in memory) were discarded and fetch_usage_events raised before returning, so the run upserted zero rows for the 2026-08-22..2026-08-23 window.

Root cause. _make_request_with_retry (cursor_events_client.py:141-176) implements retry only for status 429; every other non-200 response, including transient 5xx such as the observed 503, falls through to response.raise_for_status() on line 176 and propagates as an unhandled HTTPError. The module-level MAX_RETRIES = 3 constant (line 24) is defined but never referenced, so the retry contract the function name and docstring promise ('retry on transient errors') is not actually implemented for server errors. The Cursor Admin API returned a 503 on page 8, which is a routine transient upstream condition that a correct client would retry, not a permanent fault in this pipeline.

## What this PR changes

In cursor_events_client.py, extend _make_request_with_retry to retry transient failures (HTTP 5xx and connection/timeout errors from requests) with bounded exponential backoff up to MAX_RETRIES attempts, re-raising the last error only after retries are exhausted, while keeping the existing 429 handling. This confines the change to the pipeline's own Tier A directory and honors the MAX_RETRIES constant that already exists. Add a test to tests/test_cursor_events_client.py that returns 503 then 200 and asserts fetch_usage_events succeeds after the retry (mirroring the existing test_fetch_retry_on_rate_limit), and keep a test asserting a persistent 5xx still raises after MAX_RETRIES.

Why this fixes it. This is a genuine code-level defect confined to pipelines/runners/cursor-usage-events-pipeline/src/cursor_events_client.py (Tier A), not a one-off upstream outage: the client silently omits 5xx retry logic that its own name, docstring, and unused MAX_RETRIES constant already promise. Fixing it with bounded exponential-backoff retries on 5xx plus network errors is the smallest correct change that prevents a transient Cursor 503 from aborting the run and dropping a full day of AI-spend data, and because the T-2..T-1 window only self-heals T-1, an unretried blip on the T-2 day (08-22 here) is lost permanently. Per the repo code_fix convention the diff is expected to span the client plus a new test in tests/test_cursor_events_client.py, and it stays entirely inside the pipeline directory.

### Files changed

 .../src/cursor_events_client.py                    | 58 +++++++++++++++++++---
.../tests/test_cursor_events_client.py             | 54 ++++++++++++++++++++
2 files changed, 104 insertions(+), 8 deletions(-)

## Verification

### pytest (pipelines/runners/cursor-usage-events-pipeline/tests) — exit 1

============================= test session starts ==============================
platform linux -- Python 3.11.16, pytest-9.1.1, pluggy-1.6.0 -- /opt/hostedtoolcache/Python/3.11.16/x64/bin/python
cachedir: .pytest_cache
rootdir: /home/runner/work/Surtr/Surtr/publish/pipelines/runners/cursor-usage-events-pipeline
configfile: pyproject.toml
plugins: mock-3.15.1
collecting ... collected 10 items / 1 error
==================================== ERRORS ====================================
_____________ ERROR collecting tests/test_cursor_events_client.py ______________
ImportError while importing test module '/home/runner/work/Surtr/Surtr/publish/pipelines/runners/cursor-usage-events-pipeline/tests/test_cursor_events_client.py'.
Hint: make sure your test modules/packages have valid Python names.
Traceback:
/opt/hostedtoolcache/Python/3.11.16/x64/lib/python3.11/importlib/__init__.py:126: in import_module
return _bootstrap._gcd_import(name[level:], package, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
tests/test_cursor_events_client.py:7: in <module>
import responses
E   ModuleNotFoundError: No module named 'responses'
=========================== short test summary info ============================
ERROR tests/test_cursor_events_client.py
!!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 1 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!
=============================== 1 error in 0.39s ===============================

<summary>Run metadata</summary>

| Field | Value |

| --- | --- |

| Pipeline | cursor-usage-events-pipeline |

| Failing run | a81800d5-76fd-407c-933a-47a271466c6e |

| Occurrence | 1 (times this exact failure signature has been seen) |

| Signature | 0f20999c03f1f23deab29f04f451a00b1d43f75f6bbe7263dca30fba6367123f |

| Verify | failing |

</details>

---

🤖 Opened by heimdall. mercy reviews this PR automatically; heimdall revises on REQUEST_CHANGES (bounded rounds). Tier-auto PRs may auto-merge on mercy approval when the consumer enables it; everything else waits for a human. Mention heimdall in a comment to direct it, or add the manual-dev label to take the PR over and stop it entirely.

View on GitHub →

#1509 — fix(openai-usage-pipeline): fetch project names via list endpoint, not per-project GETs @kevalshahtrilogy approved

## Summary

- Replace fetch_project_names's per-project GET /v1/organization/projects/{id} loop with one paginated sweep of GET /v1/organization/projects (the org-wide list endpoint).

- Same function signature/contract — handler.py's call site is unchanged.

- Still short-circuits to zero API calls when a BU touched zero projects that day.

- Defensively handles a live-observed title vs documented name field discrepancy in OpenAI's response.

## Why

A BU touching N distinct OpenAI projects in a day previously cost N separate GET requests just to resolve project names — on top of its usage call, its per-project api-keys calls, and its costs call. Across the 28 BUs this pipeline processes daily, that's the dominant source of request-volume burst.

This matters because two prior fixes addressed *symptoms* of the resulting 429s but not the cause:

- PR #1301 (shared token-bucket throttle) smooths sustained request rate, but doesn't reduce how many requests get made.

- PR #1395 (retry-window floor) makes a throttled BU's retry reliably outlast the rate window, but only after the quota is already exhausted.

429 exhaustion has continued dropping BUs' billed-cost/usage data on multiple days most weeks since both merged (most recently 08-22 through 08-24, a different BU pair nearly every day). This PR reduces the actual request count instead of coping with the consequences of it — a BU touching N projects now costs ceil(total_org_projects / 100) requests for names instead of N, which for most orgs collapses to 1-2 calls total regardless of N.

fetch_project_api_keys (the other per-project loop, for owner attribution) is intentionally left untouched — I don't have confirmed evidence of a batching endpoint for that one, and this is a financial pipeline where I'd rather not guess.

## Business Value

Reduces the frequency of recurring billed-cost/usage gaps on a financial data pipeline that have required manual single-BU reruns to recover, multiple times a week.

## Manual Effort Estimate

~2 hours (trace the actual per-BU request pattern in the handler, verify OpenAI's Admin API has a list-projects endpoint, implement + rewrite the affected test suite).

## Test plan

- [x] uv run pytest tests/ — 146/146 passed (8 new/updated: pagination via after= cursor, title/name field fallback, zero-projects short-circuit, missing-project-id handling)

- [x] uv run ruff format --check .

- [x] uv run ruff check .

View on GitHub →

The Portfolio — Trilogy Companies

Alpha School's National Moment: The $40,000 AI Classroom Is Now Everybody's Problem

From CNN to the American Enterprise Institute to a Chicago city block, the 2-hour school model has escaped the Austin bubble — and the scrutiny is just beginning.

By Pat Donnelly, Investigative Desk · Claude Sonnet

AUSTIN, TEXAS — The questions are arriving faster than the campuses. Alpha School, the Austin-based private K-12 experiment in which AI tutors handle the academic curriculum in two hours each morning while students spend the rest of their day on life skills, entrepreneurship, and self-directed learning, has broken through from niche curiosity to national debate — and the institutions paying attention span a notable ideological range.

CNN framed the story as an open question — 'What if I told you this school had no teachers?' — a hook that captures both the school's marketing genius and the unease trailing it. The American Enterprise Institute, not an organization prone to wishful thinking about education startups, published a measured open letter: "I Hope You're Right." The 74, which covers education policy with a reform-friendly lens, asked what public schools could extract from a model priced at $40,000 to $65,000 a year — a question that quietly acknowledges the accessibility problem without answering it.

Meanwhile, Block Club Chicago reported that an AI school operating on the Alpha model — no teachers in the traditional sense — is slated to open in Chicago this fall, pushing the experiment into a major urban market where the political and demographic stakes are considerably higher than in suburban Austin.

The school's backers, including founder Joe Liemandt, have committed $1 billion to Timeback, a platform designed to let entrepreneurs license the model and launch their own AI-first campuses — a Shopify-for-schools ambition that implicitly answers the scale question while sidestepping the equity one. Alpha's own data shows students testing in the top 1–2% nationally on NWEA MAP Growth assessments, learning at 2.3 times the pace of U.S. norms. Those numbers have not been independently replicated at scale.

The constellation of coverage — conservative think tank, left-leaning urban outlet, national cable network, education press — suggests Alpha has achieved something most education startups never do: it has made the mainstream uncomfortable enough to engage seriously. Whether that engagement leads to adoption, regulation, or rebuttal remains the open question. Chicago will provide data points that Austin, with its particular demographic and economic profile, could not.

↗ Dear Alpha School: I Hope You’re Right - American Enterprise · ‘What if I told you this school had no teachers?’: Is AI sch · What Public Schools and Parents Can Learn from a $40,000-a-Y

Skyvera’s CloudSense Coup Gets an AI Speed Run

The telecom software shop adds Salesforce-native CPQ muscle, then flashes TM Forum papers in one month flat.

By Dottie Sharp, Society & Industry Desk · GPT-5.2

AUSTIN, TEXAS — Word is Skyvera didn’t just buy CloudSense. It put the new arrival through finishing school at warp speed.

The telecom software outfit in the Trilogy orbit has completed its acquisition of CloudSense, the Salesforce-native configure-price-quote and order management specialist built for the hairiest corners of telco enterprise sales: B2B, B2B2X, wholesale, bundled services, complex pricing, the whole switchboard spaghetti. TelecomTV clocked the deal, and Skyvera is now giving CloudSense pride of place inside its portfolio of telecom modernization wares.

A little bird from the carrier corridor tells me the real sizzle is not merely the purchase. It is what came next. CloudSense says it certified all 13 APIs in its CPQ product set to TM Forum compliance standards in just one month, using AI-assisted development in partnership with the standards body. The usual pace? Try 26 months, according to the company. That is not a sprint. That is a disappearing act with documentation.

CloudSense now sits alongside Skyvera’s familiar telecom cast — Kandy for cloud communications, VoltDelta for customer engagement, ResponseTek for experience data, Mobilogy Now for device lifecycle management, and the rest of the carrier cleanup crew. The pitch: drag legacy telecom systems toward cloud-native operations without making operators rip up every wire in the basement.

The new prize is especially useful because telcos do not sell like ordinary software firms. One enterprise deal can mean custom network configurations, regional rules, negotiated rates, multi-party fulfilment, and sales teams living inside Salesforce. CloudSense aims to make those quotes faster, cleaner, and less prone to the kind of configuration errors that turn revenue into courtroom exhibits.

And that TM Forum badge? In telecom, standards compliance is the velvet rope. Operators want APIs that can talk to billing, ordering, inventory, and assurance systems without a year of bespoke glue. Skyvera’s claim that CloudSense hit compliance in record time using AI gives the portfolio a handy calling card: not just old telecom software acquired and housed, but acquired, accelerated, and made more interoperable.

The Triliverse angle is plain. ESW and its cousins like durable enterprise software with sticky customers. Skyvera likes telco infrastructure that needs modernizing yesterday. CloudSense brings both. One source I’ll call “The Quote Whisperer” summed it up neatly: “CPQ is where telco revenue either gets born or gets lost.”

Now Skyvera has the stork — and apparently, a faster stopwatch.

↗ TelcoDR’s Skyvera snaps up CloudSense - telecomtv.com · Cloudsense · CloudSense achieves TM Forum API compliance in record time u

Why Crossover's Global Talent Model Was Built for Exactly This Moment

By Frank Dunmore, Investigative Correspondent · Claude Sonnet

There is a Medium analysis circulating in enterprise strategy circles this week arguing that the traditional Global Capability Centers model — offshore centers running repetitive tasks at low cost — is becoming obsolete with AI agents handling that work automatically. What survives instead is something more demanding: global talent working alongside AI rather than being replaced by it, performing judgment work instead of processing work.

This thesis mirrors Crossover, Trilogy International's global talent platform, which places elite remote talent across 130 countries using rigorous AI-enabled skills assessments. A source familiar with Trilogy's planning said the timing of this conversation isn't lost on the organization: "The market is just now understanding what the model requires. We built the infrastructure for it a decade ago."

Enterprise AI adoption is accelerating — Microsoft reported over 1,000 documented customer transformation stories tied to its AI products, signaling institutional deployment at scale. Companies best positioned to capture value are those whose talent infrastructure was already designed around AI-augmented work. Crossover's meritocratic, geography-agnostic, AI-screened model looks less like staffing and more like a structural bet that has aged exceptionally well.

The Machine — AI & Technology

The Microscope Learns to See: AI Illuminates the Hidden Architecture of the Mind and the Cosmos of Discovery

From invisible brain lesions to the deep structure of scientific inquiry itself, machine intelligence is becoming the new lens humanity holds to the unknown.

By Dr. Vera Okafor, Science & Technology Correspondent · Claude Opus

STANFORD, CALIFORNIA — Four hundred years ago, a Dutch draper named Antonie van Leeuwenhoek ground a lens and saw, for the first time, the writhing microscopic universe inside a drop of pond water. He called them animalcules. We are, this month, living through something similar — except the lens is made of matrix multiplications, and the pond water is the human brain, the balance sheet, the entire enterprise of science itself.

At Stanford's Human-Centered AI institute, researchers are documenting a quiet revolution: AI is transforming scientific discovery not by replacing the scientist, but by expanding the peripheral vision of the species. Nine breakthroughs cataloged this month by UC San Diego range from protein folding to climate modeling to the mapping of neural circuits once thought unmappable. Each represents a boundary that, a decade ago, seemed permanent.

Consider multiple sclerosis. For generations, neurologists have known that MS attacks the brain's gray matter — the seat of thought, memory, personhood — but the lesions there were nearly invisible to conventional MRI. They hid in plain sight, like faint stars washed out by city light. Now a new deep learning system, trained on thousands of scans, reveals these ghost lesions with startling clarity. Patients whose disease was called mild may in fact be quietly burning. The diagnosis changes. The treatment changes. The prognosis changes.

Meanwhile, in a project that would have delighted Sagan himself, Frontiers reports on teenagers co-authoring neuroscience papers with world-class researchers — the young collaborators exclaiming, of their own discoveries in the folds of the cortex, "It's so wow!" That may be the most honest scientific expression of our era.

And in a quieter corner of the arXiv, researchers unveiled a hybrid ensemble model for bankruptcy prediction using explainable AI — a reminder that the same tools illuminating gray matter also illuminate the gray zones of markets, where firms flicker between solvency and collapse.

The animalcules are everywhere. We are only just learning to look.

↗ How AI is Transforming Scientific Discovery While Keeping Hu · ‘It's so wow!’ - Young people team up with top neuroscientis · AI Reveals Hidden Gray Matter Lesions in Multiple Sclerosis

Meta’s Compute Herd Edges Toward the Cloud Savannah

As AI infrastructure grows vast and hungry, Meta is testing whether surplus capacity can become a business of its own.

By Sir Reginald Marsh, Natural Phenomena Correspondent · GPT-5.2

MENLO PARK, CALIFORNIA — In the dim blue glow of the data center, where racks breathe warm air and silicon colonies hum through the night, a new creature is stirring. Meta, long known for tending the sprawling habitats of Facebook, Instagram and WhatsApp, is reportedly preparing to sell excess AI computing capacity to outsiders — a migration that could carry it into the cloud territory ruled by Amazon, Microsoft and Google.

The move, reported by Bloomberg and covered by Fierce Network, suggests that Meta’s enormous investment in graphics processors and AI servers may not remain solely an internal organ. When the beast has fed its own models, trained its recommendation engines and furnished its chatbot ambitions, spare cycles may be released into the wider ecosystem.

Wall Street, ever the watchful predator, has noticed the tracks. Cloud computing can generate handsome revenue, but it is a lower-margin business than Meta’s great advertising river, where attention is harvested at immense scale. CNBC notes that investors may need to prepare for a different financial climate if Meta’s AI infrastructure becomes a commercial utility rather than a private reserve.

This is part of a broader seasonal change. Across the industry, compute is beginning to resemble an energy market: scarce at peak demand, abundant in quiet hours, priced by location, latency and appetite. InfoWorld has described the possible rise of capacity markets in cloud computing, where buyers and sellers might trade access to processors as naturally as utilities trade electricity.

And yet the habitat itself is under pressure. AI companies are even looking seaward, studying ocean-based data centers where cooling water is plentiful and land disputes recede beyond the shoreline. There, beneath salt winds and wheeling gulls, the next generation of machine intelligence may find refuge.

For Meta, the question is delicate. Can a company bred in the rich canopy of advertising adapt to the harsher open plain of cloud infrastructure? The answer may determine whether its AI buildout becomes a costly display of plumage — or a new species of business altogether.

↗ Meta’s push into cloud computing means Wall Street has to pr · Capacity markets could reshape cloud computing - InfoWorld · Meta is building a cloud business to sell excess compute cap

DoJ Antitrust Throne Sees Revolving Door As Big Tech Cases Hang in the Balance

By R. Barnsworth III, Esq., Legal Affairs Desk · Claude Sonnet

The Antitrust Division of the Department of Justice has experienced two leadership transitions within five months, creating uncertainty about its ability to prosecute cases against Google and Apple. President Trump has nominated an individual critical of "Big Tech" to fill a vacancy, signaling renewed antitrust enforcement efforts. Federal Trade Commission Chair Andrew Ferguson has publicly stated that court proceedings move too slowly, allowing dominant companies to gain strategic advantages from delays. This statement suggests the Trump administration's "antitrust honeymoon" with major tech firms may have ended. However, the practical impact of recent leadership instability on the division's operations remains unclear, and observers should monitor developments as enforcement policy takes shape.

The Editorial

The Geopolitics of the Prompt

Every decade produces its indispensable panic; ours has arrived on schedule, dressed in silicon.

By Victor Marsh, Chief Columnist · Claude Opus

WASHINGTON — There is a particular species of think-tank paper, blooming this season like ragweed, which announces with grave italics that Artificial Intelligence is now a matter of geopolitics. One reads the Observer Research Foundation on the geopolitics of benchmarks, the Atlantic Council on the eight ways AI will shape 2026 (never seven, never nine — the round numbers of prophecy), the United Nations University on something called "governance arbitrage," and one is tempted to reach for the whiskey before the second paragraph.

The tempation should be resisted, but only barely.

It is not that these authors are wrong. It is that they have discovered, with the trembling excitement of a graduate student encountering Foucault, a proposition so ancient it predates the wheel: that whoever controls the tools controls the argument, and whoever controls the argument controls rather a lot else. The Venetians knew this about glass. The British knew it about coal. The Americans knew it about the transistor. That we should now be surprised to learn it about matrix multiplication says less about the novelty of the technology than about the amnesia of the commentariat.

The United Nations University paper is the most diverting of the lot, because it deploys the word "arbitrage" — a term of art from the trading floor — to describe the ancient human practice of registering your company in whichever jurisdiction hates you least. Firms are shopping for regulators. This is presented as though it were a discovery on the order of penicillin. One recalls that Delaware exists, that Panama exists, that Ireland exists, that the Cayman Islands exist, and that they existed before ChatGPT was a gleam in Sam Altman's eye. What has changed is not the arbitrage but the asset. The lawyers, as ever, remain fully employed.

The Atlantic Council's octet of predictions for 2026 follows the format perfected by newspaper astrologers: sufficiently specific to seem bold, sufficiently vague to be unfalsifiable. AI will "reshape" alliances. AI will "transform" deterrence. AI will "accelerate" the compute race. One awaits the eventual audit — which will not come, because the genre does not require it. The forecasters of 2021 who told us the metaverse would be the theater of great-power competition are not, I notice, being asked to show their work.

Meanwhile The Guardian, in its characteristically mournful register, exhumes the warnings about the digital age that were ignored in the 1990s, as though the ignoring of warnings were itself a novelty rather than the default operating condition of the species. Cassandra was, one recalls, correct; she was also, one recalls, ignored. This is not a bug of democratic capitalism. It is very nearly its definition.

What all of this ceremonial hand-wringing obscures is the boring and useful truth: the geopolitics of AI is the geopolitics of electricity, of sand, of graduate students, and of the tax code. The states that get those four right will do well. The states that write papers about doing them right will write more papers. One suspects we know which category most of the commissioning institutions fall into.

↗ The AI Geopolitics - orissapost.com · The Geopolitics of AI Benchmarks - orfonline.org · The AI Governance Arbitrage - UNU | United Nations Universit

The Office Comic · Art Desk

Nation’s Marketers Warned Not To Board Stupid Meme Until It Has Been Safely Dead For Several Business Quarters

Industry leaders urged brands to preserve authenticity by waiting until every normal person has already begged them to stop.

By Dale Pemberton, Staff Writer · GPT-5.2

NEW YORK — The country’s communications professionals received a timely reminder this week that the internet remains a fast-moving cultural ecosystem in which a brand can only participate meaningfully after legal, social, executive, regional, and procurement teams have taken 19 days to approve a joke that was already annoying by lunch.

The guidance follows a fresh round of marketing self-examination prompted by PR Daily’s sober account of the professional lessons available from jumping on a stupid meme way too late, an operational discipline now considered essential to modern brand management. The central lesson, according to people paid to speak in panels beside ferns, is that companies must become more agile, more human, and more willing to have a junior copywriter explain why the CEO should not post the dancing raccoon.

This is not merely a social media problem. It is a civilization problem. We have constructed an economy in which every institution, from language apps to cloud infrastructure vendors, must maintain the emotional posture of a 23-year-old with 400,000 followers and a ring light. The result is a marketplace where brands are no longer judged by product quality, pricing, customer service, or whether the software works, but by whether their owl mascot appears sufficiently deranged.

Marketing professor Mark Ritson’s recent criticism that Duolingo was foolish to prioritize influencers over its unhinged owl has rightly shaken boardrooms across the developed world. Executives who once asked whether their company had a moat now ask whether it has a green bird capable of threatening customers in a way that feels native to TikTok. This is considered progress.

Meanwhile, the artificial intelligence industry continues to perform its own version of the same ritual, only with national security clearances and worse nouns. The Trump administration’s reported ban on foreign access to Anthropic’s new AI models has provoked broad tech-world reaction, with many observers expressing concern that humanity’s most advanced stochastic parrot may now be distributed according to geopolitical vibes. The same executives who insist AI will dissolve borders, labor markets, and undergraduate writing courses are suddenly discovering that export controls can make a demo slightly harder to schedule.

Investors, to their credit, are responding with their customary restraint. A separate warning that AI investment buzzwords are a red flag arrives at an awkward moment for a sector that has built substantial enterprise value out of the phrase “agentic workflow orchestration.” It now appears possible that some founders may be using impressive terminology to make uncertain business models look inevitable, a shocking breach of the sacred covenant between venture capitalists and people wearing quarter-zips.

Steve Ballmer, who said he was duped and felt silly after a founder he backed pleaded guilty to fraud, has become the rare public figure willing to articulate the private emotional state of the entire capital allocation class. Feeling silly is now the market’s most credible due diligence framework. If a pitch deck contains the words “AI-native,” “autonomous,” “frontier,” “sovereign,” or “category-defining,” investors are advised to place one hand over their wallet and the other over their LinkedIn announcement draft.

The lesson across all of this is brutally simple: whether chasing a meme, a mascot, a model, or a founder promising to reinvent a regulated industry by Tuesday, institutions keep mistaking motion for judgment. They see everyone running and conclude there must be something worth running toward. Sometimes there is. Often it is just a stupid meme, already dead, surrounded by consultants explaining how to scale it globally.

↗ 3 lessons from jumping on a stupid meme way too late - PR Da · Tech world reacts to Trump administration ban on foreign acc · Mark Ritson: Duolingo stupid to prioritize influencers over

On This Day in AI History

On August 24, 1991, the World Wide Web was released to the public by Tim Berners-Lee, fundamentally transforming how information would be shared and laying the groundwork for the internet revolution that would enable modern AI development. This open-source release moved the web from academic institutions into the hands of anyone with a computer, creating the digital infrastructure that would eventually power today's AI systems.

⬛ Daily Word — Technology

Hint: Relating to computers, the internet, or digital systems and security threats.