- **Deep-audit remediation: security, correctness, and multi-instance robustness.** Permission gate: awk programs that shell out via `print | "cmd"` / `getline` are now classified as mutating AND destructive (previously only `system(`/`>` and only mutating), and `xargs -L N` no longer hides its payload from classification. Glob resolves symlinks before its workspace-boundary check (an in-workspace symlink could previously list outside content); progress-note titles are ANSI-sanitized like every other transcript field. Grep content lines are parsed with unambiguous field separators, so paths like `utf-8-codec.py` are no longer mangled with `-n=false` and sensitive-file attribution is exact. Multi-line edits on CRLF files work again (LF-joined old strings are CRLF-translated when needed). `/import` preserves paths byte-for-byte (only a standalone leading/trailing `--force` is treated as the flag). Post-compaction file reminders include `--add-dir` files. Double-interrupt can no longer orphan the interruption-marker write (unanswered tool_calls). Background web replay falls back to full history (not empty) when the watermark stat fails, and a malformed Agent resume id returns a clean "Agent not found". OAuth: login fails loud when the token response lacks a `refresh_token`; a refresh response without `expires_in` carries the previous lifetime forward instead of refreshing every tick; the device-id file can no longer be read empty mid-creation. A failed `theme="auto"` background probe can be retried by re-selecting auto via `/theme`. Multi-instance: sessions now take a per-session writer lock (a second `pythinker -r <id>`/web worker on the same session is refused instead of interleaving turns), the shared `pythinker.json` index uses a locked read-modify-write (no more lost work-dir registrations), JSONL appenders repair torn final lines after a crash, forks materialize atomically, project-memory mutations abort on read failure instead of wiping the file, the journal is capped at 100 recaps, inbox approve/reject claims candidates atomically, and recall re-arms when another instance writes new memory. Subagents: a failed summary continuation no longer discards a completed agent's work, hallucinated subagent types fail fast with the valid-type list (before any RunAgents child launches), background failures carry an `Agent ID:` + resume hint, and a crash inside the runner's own error handling is logged instead of silently lost.
0 commit comments