Dispatch 376
Week of Nov 16 to Nov 22 ↗

Locking down the dispatch system and mapping review capacity

The team spent Friday tightening the safety nets in the agent dispatch system and clarifying who reviews what. The work split into two clear arcs: hardening the lock mechanism that keeps concurrent jobs from…

Commits
7
Systems
12
Read
2min
NOVEMBER 20, 2026Friday352 WORDS · 2 MIN READ · 7 COMMITSops-center90agent-dispatch33legal-ops161310976
Day signal

Drawn from this entry's own record: its date, its length and the work around it. No stock art.

Key points

  1. The team fixed a bug where the system assumed a process was alive if it could not read its status.
  2. Tests simulated a lock with another user's process ID to verify correct stale lock detection.
  3. Aaron wrote a script to parse agent data and compute reviewer counts per author family.
  4. The team added a CI check to ensure the tools README stays in sync with the actual code.

The team spent Friday tightening the safety nets in the agent dispatch system and clarifying who reviews what. The work split into two clear arcs: hardening the lock mechanism that keeps concurrent jobs from stepping on each other, and building a tool to track reviewer distribution across author groups.

The lock fix came first. A subtle bug in the _live check was treating a PermissionError on a process ID as alive, keeps 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. The tests for this change were thorough. They verified the new behavior by simulating a lock where the process ID belonged to another user, ensuring the system correctly identified the stale lock. A round-3 test was also updated to use a real sleep instead of relying on pid 1, making the simulation of concurrent execution more realistic. These changes ensure that the dispatcher releases resources promptly when a job fails or exits unexpectedly.

Alongside the dispatch fixes, the team built a reviewer matrix tool. The goal was to count reviewers per author family to ensure balanced code review coverage. Aaron wrote reviewer_matrix.py, which parses the real agent data and computes these counts. The tool includes a test suite that verifies the matrix counts only cross-family reviewers, preventing any single group from dominating the review load. To keep the tool maintained, a row was added to the tools README, guarded by a CI check that ensures the documentation stays in sync with the actual code. The team also generated a REVIEWER-CAPACITY.md document from the agent data, providing a clear snapshot of current review capacity.

Alongside the dispatch fixes, the team built a reviewer matrix tool.

Also today: the team updated the test suite for the dispatch holds to reflect the new lock logic, and refreshed the documentation to match the new tooling. The work was steady and focused on reliability and clarity.

The day ended with a dispatch system that releases locks correctly and a review process that is now measurable and balanced.

From the recordWhere the work went198 commits across 12 repositories
  1. ops-center90
  2. agent-dispatch33
  3. legal-ops16
  4. tax-planning-ops13
  5. accounting-ops10
  6. unicorn-brigade9
  7. email-ops7
  8. customer-ops6
  9. 4 other repositories14