5 min read

Weekly development log 2026‑W41: Managing the waits, permissions, and lifetime of work on your computer · snapshot checked Oct 8

  • Weekly log
  • Vibe coding
  • Self-hosted AI
  • 2026-W41
W41 illustration of Mac and Windows Companion connections, browser control and site policies, task queues and retention. Created on 2026-10-08 from the development log; this is not a product screenshot.

MAIN SNAPSHOT / EVIDENCE

This week in OpenMake

The October 5–8 KST audit found 62 main commits, including merges. This week improves Companion endpoint discovery and updates, asks before retrying writes with unknown outcomes, and brings organization browser policies into the admin UI. Device revocation, queue wait metrics, and optional retention cleanup accompany those changes. This is a midweek snapshot pinned to 799dd001; the latest general openmake_llm release remains v1.91.0.

Period
2026-10-05 – 2026-10-08
Git commits
62
Evidence
Git history

01

What changed

  • Companion resolves connection and web endpoints from one server address. It queries public port configuration while retaining the user-provided host, and keeps the original address for proxies or older servers. This replaces fixed macOS choices and Windows port guessing. Plain HTTP to a remote computer triggers a warning (59aad238).
  • Desktop delivery received fixes: macOS build versions exceed the published build, updates are skipped in read-only locations, and DMG retries and CI artifacts were added. Windows CI tests, smoke-checks, and builds installers; the empty tray icon was fixed. Main contains macOS 0.3.1 and Windows 0.1.3, with companion-v0.1.0 through 0.1.3 in its history. The October 6 companion-v0.1.4 and 0.1.5 tags sit outside this main snapshot and are not counted as landed main changes (71ce3a7d, f90b2a0d, 27f0cc2f, 6c3c8a03, 10d27d9e, bb145a6c).
  • Retrying a write whose outcome became unknown after disconnection now requires user confirmation. Browser input and clicks are checked at action level, including split calls; a rejected write is blocked without repeatedly asking. Input tool results return actual values to reduce repeated input just to verify a change (cb151923, 7cf08f0d, cb74e449).
  • A new path parks tasks and releases execution slots while a user controls the Companion browser. Devices report control changes and returning control resumes that user's parked work. LOCAL_BRIDGE_TAKEOVER_PARK defaults to off; when enabled, its default wait limit is one hour (74ab2ac0).
  • Administrators can edit browser allow and deny lists and test policy decisions in a dedicated editor. Organizations use the same editor with merge-rule guidance; organization BROWSER_SITE_POLICY writes use the same pattern validation as global settings. These are UI and validation changes, not proof of safety across production deployments (caf188a9, 467a8a4d, b2fb3305).
  • Connected-device administration now accompanies revocation. Key deletion, deactivation, rotation, or bridge-scope removal closes connections; setting an already-past expiry or disabling or deleting the account does too. Future expiry changes update connection state, with the existing 30-second heartbeat closing expired connections (9845075d, b9fae0f2).
  • Chat task cards show queue position, failure reasons, and next actions, while resumed tasks retain approval policy. Queue metrics summarize the latest 200 waits, record a histogram and the oldest pending wait, and preserve enqueue time during boot recovery (cb151923, ebafc900, 3860a916).
  • Retention cleanup is opt-in: AGENT_TASK_RETENTION_ENABLED defaults to off. When enabled it runs at boot and every six hours, with default retention of 90 days for ended task records and associated audit logs and 30 days for server-sandbox image files. Active and parked tasks are excluded; batches are bounded and deletions logged (75357573).
  • Internal-model tool turns now detect missing first chunks after 120 seconds and idle gaps after 60 seconds; monitored calls avoid a duplicate overall call cap. Request priority is also carried in extra_body through the gateway to vLLM. A real-TCP integration test adds 100 WebSockets, user-result isolation, 500 mixed calls, and reconnection scenarios. We did not rerun these tests or measure production concurrent-user performance for this article (f67d3d33, 5cafbc9d, 5b5f1885, dcd968c1).

02

How this week was reconstructed

A midweek snapshot checked against GitHub main on October 8, 2026, at 06:39 KST. It counts 62 commits, including merges, from October 5 at 00:00 KST through the check; latest main is 799dd001 (October 6, 04:14 KST). This is not a final review of the whole week. General release v1.91.0, later main changes, and Companion tags are distinguished. Tests are source records, not operational evaluations rerun for this article.

Source evidence

Evidence

OPENMAKE · MIT OPEN SOURCE

Back to Engineering Log