aish v0.3.12

Added

• /doctor Shell Compatibility checker: identity, startup files, startup speed, environment inheritance (proxy values redacted), secret detection (set/unset only), PTY bash compatibility probe, and alias/function/completion counts. aish doctor --json and /doctor --json emit machine-readable results.
• Live-session resource monitoring: /live_sessions shows CPU% and RSS per session; configurable alerts when a detached session exceeds CPU/RSS thresholds (pty_resource_check_interval_secs / _cpu_percent / _rss_mb). Killing a session now terminates the full worker process tree with pidfd-pinned SIGTERM then SIGKILL so recycled PIDs are not hit.
• Long-term memory: confirm before store/forget, provenance, expiry/TTL, /memory (list / verify / forget), and a category-aware TTL policy with an explicit permanent flag.
• /audit export: portable audit package (audit.md + events.jsonl + manifest.json) with --until and --session filters, SHA-256 hashes, and owner-only permissions.
• System-level security policy: /etc/aish/security_policy.yaml takes full precedence over the user-level ~/.config/aish/security_policy.yaml.

Changed

• glob tool uses pruned DFS (skips target / node_modules / … at traversal time), a 60s wall-clock budget, Ctrl+C cancellation, and spawn_blocking. Cancellation returns partial results so the agent does not retry into another cancelled walk. Absolute patterns start at the longest wildcard-free prefix instead of scanning from /.
• Session database (sessions.db) and its WAL/SHM siblings are created or tightened to owner-only 0600 on open.

Fixed

• web_fetch returns a structured failure (ok=false, http_status) on HTTP 4xx/5xx instead of treating error pages as success, caching them, and sending them to the secondary LLM.
• /doctor no longer hangs forever in the REPL: probe subprocesses are detached from the controlling TTY so interactive bash cannot SIGTTOU-stop itself. Probe stderr noise is a warning rather than a fatal baseline error, and API connectivity treats 401/403/404 as reachable (only 5xx and network errors fail that check).
• Nested-confirm buffer truncation no longer panics on a UTF-8 character boundary (for example a large CJK MOTD inside nested SSH), and the kept tail is hard-capped.
• Sub-agent fatal outcomes preserve already-completed tool results and a structured error_category, so the parent can continue without redoing finished work.
• Main session now persists AI tool-call history, so a follow-up question can see what the AI just executed instead of misattributing those commands to the user.
• Thinking-animation frames no longer swallow tracing log lines, so LLM retry warnings stay visible.
• Ctrl+C now stops the sub-agent thinking animation thread; tool-result status lines no longer overlap leftover frames.