v2026.7.9-2

Added

• Added prompt cache miss tracking and optional TUI notices.
• Added the agent_settled lifecycle event for extensions and RPC clients.
• Added provider arguments for /login.
• Added a before_provider_headers extension hook.
• Added named inline extension factories for SDK users.
• Exposed GPT-5.6 model metadata through the CLI model registry.

Changed

• Improved project-local config resource management.
• Aligned reload command descriptions across docs, examples, and slash commands.
• Updated model catalogs, including GitHub Copilot extended-context windows and Xiaomi token-plan catalogs.
• Warned when --session-id creates a new session.
• Tool-result truncation markers now include retrieval guidance (re-run with a narrower range); legacy markers remain parseable.

Fixed

• Counted context-visible custom messages in compaction budgeting.
• Avoided Windows context-file discovery hangs.
• Applied modelOverrides to extension-provided models.
• Fixed upstream-merge regressions that could run queued follow-up or goal-continuation prompts after aborted turns, drop context-exclusion filtering during compaction, or show stale MCP diagnostics.
• Restored exact comma-formatted token counts in the interactive footer.
• Fixed native clipboard image handling in Bun releases.
• Prevented double-selecting a fork menu entry.
• Cleared label timestamp cache state when starting new sessions.
• Normalized null message content at ingestion boundaries.
• Retried Bun socket-drop provider failures and ResourceExhausted errors.
• Used a (no tool output) placeholder for empty text-only tool results.
• Failed tool calls from length-truncated assistant messages.
• Kept TUI paste tracking correct when paste markers are deleted or the terminal is cleared.
• Tool results are no longer head/tail-truncated below the emergency context threshold; the model now sees full tool output until context genuinely runs out (session 019f45c0 regression).

Removed