2026‑W43: Colima Runtime Unification, Live Terminal Streaming, and Robust Task Cancellation and Recovery

01

Full-history review and additions

Across 367 commits, releases 1.106.2 through 1.115.1 shipped. Unifying macOS runtime under Colima, web agent live terminal streaming view, eight-stage verification and delivery pipeline, API-key NULL-as-inactive security floor (S5), tool-call taint boundary propagation across subagents, immediate cancellation signal propagation with first-turn abort continuity, and watchdog stale pending detection with resolved alerts landed.

All reachable history of main 6dfd605a was rechecked on 2026-10-24. Counts use committer dates from Monday 00:00 through Sunday 23:59:59 KST, including merges, releases and distinct parallel-branch commits. Commit counts are not feature counts. Tests and operational measurements below are historical source records, not evaluations rerun for this refresh.

Snapshot checked October 24 at 19:30 KST. General tags range from 1.106.2 to 1.115.1, including macOS Colima runtime unification, agent live terminal view, eight-stage verification pipeline, security floor (S5), cancellation propagation, session continuity, and watchdog recovery.

02

What changed

• The macOS OpenMake runtime was unified exclusively under Colima with hardened sandbox isolation. Docker execution on macOS was consolidated onto Colima with protection guards for existing data directories. Playwright sandbox environments are partitioned, ffprobe is bundled, and write permissions for Chromium state in readonly sandboxes are guaranteed. Chat media files persist as account-owned artifacts, authorized file serving operates from hidden install roots (.openmake), and legacy staging tunnels were removed (6c65f324, a1514c1d, 2afe7db2, 1f0e938c, fd753a08, 3ee96f91).
• The web UI now provides a Live Terminal View for running agent tasks. Output logs and browser execution screens stream in real time, auto-following new output only when the user is scrolled near the bottom. An output transmission cleanup deadline prevents late frames from overwriting current state, while terminal frames are retained even for short-lived browser executions. External QA and monitor contracts were separated, and the Apache-2.0 license was formalized (7052ac4a, 505d9873, a4e9ec25, 839cc41d, 8ad13aed, d1d058f4, 39573b6e).
• An eight-stage verification and delivery pipeline was deployed. CI execution optimizes by running only scenarios matching declared changed feature sets, linking staging checks directly into verified release automation. The new 'omk dev sync' CLI command syncs workspace dependencies and environments before 'dev up', and package-lock workspace versions are validated during release gates (f92e5ac8, 09046fd8, f909046f, 40dcd864, 708841cf).
• Security floors and authorization boundaries were significantly hardened. API key authentication treats NULL is_active rows strictly as inactive (S5) to eliminate potential authentication bypasses. Chat subagent delegations carry tool-call taint boundaries, non-zero approval floors remain enforced under chat policy 'none', and legacy host tools strictly require approval gates. Control environment overrides for MCP git tools and manual/git-ingest servers are rejected, and per-IP scrape rate limiting was placed on /metrics (75022aaa, f60a2143, fdc2ea3c, 50521c29, 04f3da65, 520d9df2, eca79b40, d5128ef2).
• Immediate cancellation propagation and first-turn abort session continuity are guaranteed. Cancellation signals propagate immediately through chat research, MCP meta calls, and all pending tool boundary hops, halting abandoned tool execution branches and stripping error outputs from research evidence. First-turn aborts isolate error retention strictly to the input session, preserving conversation session continuity for follow-up turns (67b60213, a47448c6, cc900fbb, 53d71747, 989c76d3, 97fd72b0, 22ecc46e).
• Watchdog reliability and automated recovery mechanisms were overhauled. The watchdog detects stale pending tasks without active live leases and sends exactly one resolved alert per completed recovery instead of repeating notifications. Parked model waits are tracked through DB aggregates, unleased continuation successors recover on boot against a fixed startup cutoff, and prepared successors commit atomically with lease ownership (db52999d, bae3797c, de723bf3, 4dc00ec9, a1138495, 913b7923, 590f867d, 1d797785).