Week in progress67
Week of Nov 16 to Nov 22, 2026

Plan Limits, Shared Tenancy, and the Clockwork of Dispatch

The week began with the phone system finally learning to count. For weeks, the voice layer had been a collection of abstract concepts and partial branches. On Monday, the machine did the heavy lifting. The s…

2,786commits21systems6min read
Product capture

Accounting-Ops: provenance lineage.

Key points

  1. The team mapped plan tiers to hard caps for seats and agent minutes.
  2. Codex Sol built the feature to store call summaries and conversations.
  3. Aaron updated org_api to enforce explicit platform authority for owners.
  4. The dispatcher now calculates push time using the server clock instead of commit dates.

The week began with the phone system finally learning to count. For weeks, the voice layer had been a collection of abstract concepts and partial branches. On Monday, the machine did the heavy lifting. The system now maps plan tiers to hard caps: starter plans get three seats and a hundred agent minutes, team plans get ten and four hundred, and business plans get thirty and fifteen hundred. If a workspace has no plan, it defaults to the starter tier. The meters are running, the caps are set, and the usage screens are live.

The phone agent also stopped being a black hole. Codex Sol built the feature to keep the summary and the conversation. When the agent ends a call, it pushes that data to Pots. Persephone reviewed it, merged it, and the team verified it on production. A real summary and conversation from the voice service are now stored and shown. The outbound call records followed suit. Codex Astra handled the data and readers for calls placed from a device, including direction logic. A call row can now be outbound, and every reader shows it by the other party. The call list is complete. You see incoming and outgoing calls in the same place, sorted correctly. The user interface polished up as well. Codex Sol built the screens for the browser phone setup on the Carrier page, showing the plan from the real carrier account. The Devices list is now a real list, showing connected devices, not placeholders.

Calendar-Ops was a bit rougher. Codex Astra tried to build plain VEVENT time ranges from occurrence rows. Persephone sent it back. The CI core failed on the head. The benchmark gates did not pass. It claimed the run again, spent forty-seven minutes, and pushed more commits. The calendar logic is tricky, and the agents need more iterations to get the edge cases right. Still, the phone foundation is solid.

Codex Sol built the feature to keep the summary and the conversation.

Tuesday, the team spent the day tightening the bolts. The team tightened the bolts on two anxieties: who owns what in the organization hierarchy, and whether the shared multi-tenant environment is truly isolated. In Ops-Center, the focus was on the organization owner role. Aaron signed off on a substantial update to org_api and user_management_api that enforces explicit platform authority. The changes ensure that owner invitations are renewed only when provenanced and that owner grants are hidden from organization admin role pickers to prevent accidental privilege escalation. The team also fixed a regression where concurrent last-owner changes could crash the system, adding tests to reproduce and isolate the issue. On the frontend, the upgrade flow now renders unknown current tiers correctly and handles subscription load errors gracefully.

Meanwhile, the Stable backend team tackled shared tenancy. The goal was to ensure the system fails closed when shared mode is active. Aaron committed a suite of changes that refuse broadcast mint and join tokens in shared mode, effectively blocking anonymous access. The team also implemented a DB-free guard to prevent recorder broadcast mints. These changes were supported by a rigorous test sweep that verified secret egress never happens and that shared Row Level Security is properly enforced. The team also refined the wake alias and policy caches to be keyed by workspace and seat, ensuring that internal reads in shared mode require a resolved workspace.

Production operations saw a measured rollout. Kit coordinated the Wave 1 switch-ons, rolling out auto-provisioning for five new apps: Project-Ops, Brigade, Email-Ops, Accounting-Ops, and Customer-Ops. Each new organization now automatically gets a workspace in all five apps. Accounting-Ops saw fixes for two live bugs: a security-policy block and untaxed invoices from the form, restoring full invoice creation functionality. In Stable, Maj. Rowan Bell oversaw the release of the new build to production, migrating from 029 to 040. The team measured significant performance gains on the 3090 hardware, with speculative decoding boosting token throughput from 30 to 47-50 tok/s on short prompts.

Wednesday, the automerge system stopped trusting committer dates. It stopped trusting committer dates and started listening to the server’s clock. The gate that decides whether to send a pull request back for review relied on the timestamp attached to the commit object. That works fine when the world is simple, but it breaks down when a PR is force-pushed or when a repository’s history gets rewritten. The dispatcher now calculates push time as the earliest of the timeline entry and the status creation. This keeps the merge queue honest. The agent-dispatch repo saw twelve commits to tighten this logic. The dispatcher now accepts reports where all VERDICT lines agree, dropping the previous restriction that rejected duplicate verdicts. It also pages through review reads so that a second page of REQUEST_CHANGES does not slip past the gate.

The phone system learned to enforce plan limits with more grace. When a paid plan ends, the account moves to free limits immediately. The system keeps the last verified plan for up to twenty-four hours if the Ops-Center is unreachable, preventing a sudden lockout during network hiccups. Security tightened around the agent inbox and MCP tools. Approving or rejecting non-agent credentials now requires positive human evidence, closing a gap where automated or partial credentials could trigger actions without clear human intent.

Thursday, the team spent the day reinforcing two critical boundaries. The paid plan grant system received a thorough overhaul. The team focused on making the completion of a paid grant recoverable rather than fragile. Previously, a swallowed app-workspace plan failure could leave a grant in a perpetual pending state. Now, the completion process is explicit, and errors propagate correctly. Past-due payments now keep the plan active for a maximum of seven days before falling back to free limits. The team also ensured that every step, from checkout to ledger entry, is auditable and idempotent.

Simultaneously, the application enforced a hard rule: non-house organizations cannot access internal tools or customer-effective org data. This was implemented across the board. Every model-calling route, from medical services to browser automation, now refuses requests from a non-house effective organization. This prevents data leakage and ensures that customer agents only see what they are authorized to see. The team also fixed a regression where linked-org home preferences were incorrectly applied outside of entitlement mode.

Friday, the team spent Friday tightening the safety nets in the agent dispatch system. A subtle bug in the _live check was treating a PermissionError on a process ID as alive, keeping the lock. This kept the lock engaged even when the original holder was gone, effectively deadlocking the queue. The team corrected the logic so it no longer assumes a process is alive just because it cannot read its status. Alongside the dispatch fixes, the team built a reviewer matrix tool. Aaron wrote reviewer_matrix.py, which parses the real agent data and computes counts per author family. The tool includes a test suite that verifies the matrix counts only cross-family reviewers, preventing any single group from dominating the review load.

Saturday, the team spent the day tightening up the project operations backend. Aaron made three targeted commits to fix how the system handles task states when things get created or reset. The first change addressed project cloning, ensuring the completedAt field is carried over accurately when the reset status flag is false. The second fix focused on the MCP layer, stamping the completedAt time when a new task is created and immediately marked as completed. The third commit tackled transactional creation logic, ensuring the completion timestamp is set correctly even when tasks are created in bulk. These changes make the data more reliable for reporting and analysis.

The week ended with a dispatch system that releases locks correctly, a review process that is measurable and balanced, and a phone system that remembers what it said. The infrastructure is tighter, the billing

Also in the frame

Real product captures. Click any to enlarge.

accounting ops
project ops
accounting ops
07 / Activity

The week, in commits.

Daily velocity and the systems that carried the work.

710Mon16
635Tue17
338Wed18
309Thu19
198Fri20
407Sat21
189Sun22
Week beginning November 16, 2026.
Index

21 systems,
one week of work.

accounting-opsagent-dispatchcalendar-opscontact-opscustomer-opsemail-opsfleet-maplegal-opsmagicunicorn-devmajiks-music-studio-promajiks.studioops-centerpotsproject-opsrunner-canarystable-agent-runnertax-planning-opsuc-clouduc-meeting-opsunicorn-brigadeunicorn-stable