Summary
A delta scan of AI coding agent releases since mid-July 2026, covering Claude Code v2.1.224–226 (Aug 8 2026), Codex CLI v0.147.0 (Aug 7 2026), and Cursor (Aug 3 2026). Beyond cataloguing new features, the scan records a key validation event: Claude Code independently fixed a bug class that the author had previously catalogued as bug class #662, demonstrating that competitor changelogs can serve as a "pre-graded validation ledger" for one's own predictions.
Key Points
- Claude Code v2.1.224–226 adds self-hosted runners, cross-session messaging (SendMessage + ListAgents), archived plugin source distribution, and gateway spend-limit warnings.
- Claude Code v2.1.225 fixed a bug in auto mode where a safety-filter refusal of the model's own permission check counted toward the consecutive-block limit — exactly the author's bug class #662 (deterministic refusals burning retry budget), fixed independently by a rival the same week.
- Codex CLI v0.147.0 introduces an
--approve-for-meflag that delegates approvals to an independent auto-review subagent, and removes--full-auto. - Codex adds support for the MCP 2026-07-28 protocol, a project-trust gate for credential use, and bearer-token redaction in replays.
- Cursor expands its permission surface with Google Workspace plugins (write access to Gmail, Drive, Calendar, Docs, Sheets) and ships an iPad app with a full PR-review surface.
Concepts
- Self-hosted runners (Claude Code) —
claude self-hosted-runnerturns a user's own machines into session hosts; scoped to Team/Enterprise plans. - SendMessage / ListAgents (Claude Code) — cross-session messaging primitives that let agent sessions message each other across machines (macOS and Linux).
- Archived plugin source — plugin distribution as zip-over-HTTPS with SHA-256 pinning for integrity verification.
- Bug class #662 — deterministic refusals that burn an agent's retry budget; the fixed behavior tells the model to move on rather than retry after a safety-filter refusal.