Weekly development log 2026‑W37: Compare models, then route work by capability

01

What changed

W37 moved from guarding individual model settings to making model choice and multimodal execution visible. Compare mode put two answers side by side; Agent Tasks exposed parallel progress; and the model layer became a Planner-led capability path for text, image, vision, audio, video, and web work. From September 7 through 13, main added 86 commits and produced 35 tagged releases from 1.48.0 through 1.62.3.

• Compare mode runs two models against the same question in a split view, then gained a per-lane Thinking toggle. It is a product comparison surface, not a benchmark claim (#807, #810).
• Agent Tasks now show parallel-agent groups as live state. The same line corrected unresolved $ref tool arguments, empty-array loops, final-turn time budgets, partial output retention, and sub-agent timeout retries (#804, #813, #815–#817).
• Skills preserve stored indentation and triggers, warn when the injection budget is exceeded, and exclude persona skills from the load_skill catalog. When eligible skills exceed the aggregate cap, the model selects which to load; that is bounded by the configured cap rather than an unrestricted prompt expansion (#822–#825).
• Model and input boundaries were tightened alongside that work. The NVIDIA NIM catalog was refreshed to 22 free endpoints, whitespace is preserved in long code and secret fields, and reasoning-only turns are no longer promoted to answers while Korean jamo input is recognized as Korean (#820, #824, #826).
• macOS now distributes the native Companion path only; Electron update defaults, publishing, and bridge behavior were removed. Companion approval waits surface as native notifications, and authentication no longer drops a message that arrives before completion (#828, #830, #831).
• Capability routing replaced the separate modality-model axis. A Planner can choose parallel image, vision, audio, video, web, and text work before synthesis; follow-ups made video jobs resumable, retained completed work, and tightened storage, ownership, cost, and execution boundaries. The 1.61.0 settings UI exposes six capability groups, while installation gained named instances, public URL setup, and database backup/restore (#832–#852).
• The final stretch restored bottom auto-scroll after mobile responses and updated OpenAI SDK 7, MCP protocol negotiation, and structured-output fallback. iOS gained orchestrator progress, media, sub-agent, model-comparison, and capability-assignment surfaces, while a missing server-filesystem seed was repaired for new installations (7bb2db1, #854–#856).
• Live checks then exposed web push registration, slash-command expansion persistence, developer-document placeholders, Quickstart URL substitution, duplicate 404 messages, and three blockers that prevented Deep Research from completing. Missing scroll containers on three workspace pages were fixed next, and a security re-review closed seven findings, including a local-bridge symlink write escape and a skill-export IDOR, plus three vulnerable dependencies. The week closed on 1.62.3 (#858, #860–#864).

02

How this week was reconstructed

No matching local Claude Code project transcript was recovered for this period. This entry therefore describes only what the Git history can prove.

We audited the available Claude Code main sessions, their proven child lineages, and the repository history. Session notes explain intent and investigation; Git remains the authority for code that actually landed.

03

Team identity and project roles

OpenMake Team is the community identity; openmake_llm is the software project; OpenMake is the product name. Non-developer maker riskpw leads openmake_llm through vibe coding, and professional developer rocky supports its development.