Weekly development log 2026‑W38: Turn a fixed core into an add-on platform

01

What changed

• Agent Tasks gained durable approval waits, enforced state transitions, a tool-call journal, plan restoration, and risk-based approval. Follow-up fixes restored router behavior after timeouts, preserved attachments, and covered browser-runner and MCP recovery (#877–#899).
• The product added an active organization context, centralized authorization, organization sharing and policy, audit history, an admin console, cost ledgers, quota reservation, approval requests, remote MCP OAuth, HITL notifications, connector artifacts, and evaluation visibility (#909–#913, #922, #928).
• Optional capabilities moved out of Base. Add-on Host and built-in packs established the boundary; generic manifest installation replaced pack-specific seeders; discussion, Deep Research, MCP, and Skill runtimes became add-ons with declared state, entitlements, schemas, model requirements, and enforced permissions (#937, #939, #941–#942, #950, #954).
• Model selection became evidence-bearing through per-model declaration tables, measured probes, adoption gates, capability wiring, pack-owned routing cases, a verification runner, and a verified-model indicator (#944, #959–#960).
• The release stream also added organization-scoped extensions, a search-provider add-on for Naver, Daum, Exa, and Tavily, a dev/staging/online environment manager with post-install SearXNG search, and CI work so SBOM generation survives dependency overrides (#946, #961–#963).

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.