TT-Wang/cortex-plugin

Six more bugs found in a second code review. Five are pre-existing (not introduced by v0.10.0) but one — --mine-all being silently broken — means our v0.10.0 migration advice to "run --rebuild-playbooks" mined nothing at all for affected users.

High severity

• --mine-all was a no-op for historical sessions. cli.py cleared the install-time gate marker, then _get_installed_at lazily recreated it with the current timestamp, re-blocking all pre-install sessions. Fixed: _get_installed_at is now read-only; _ensure_installed_at is the only creator, called explicitly by miner_daemon._run_loop on first boot. find_settled_sessions takes bypass_gate parameter, mine_all defaults it to True.
• Topic-shift silent starvation after transient Haiku failure. auto-recall.sh wrote .last-brief.json BEFORE attempting assembly, so any transient failure would leave stale keywords that suppressed future recall. Write moved to AFTER successful assembly.
• Merge TransientMiningError silently dropped + session marked complete. Now propagates out so the session finishes STATUS_FAILED and gets retried. ValueError (genuine junk) still dropped — not retryable.

Medium severity

• context_assemble dropped memories+transcripts on Haiku failure. Returned only playbook_content or "". Now returns raw materials (playbook + memories + transcript).
• smart_recall(scope_id=...) leaked cross-project memories. Haiku path loaded the global index without applying scope filter. Now filters picked_files by scope.

Low severity

• auto-recall.sh passed message as argv — ARG_MAX exposure. Large pasted prompts could hit the ~2 MB limit. Now writes stdin to tempfile, matching the pattern in session-start.sh and pre-tool-use.sh.

Tests

Six new regression tests in tests/test_v010_fixes.py — one per bug, each failing before the fix and passing after. Total suite: 79 tests, ruff clean.

User impact

If you ran --mine-all on v0.10.0 expecting historical sessions to be processed, run it again on v0.10.2 — this time it will actually work. Your vault's existing memories are not affected; the re-run just fills in what was silently skipped.

Upgrade

/plugin update memem