Changelog — TaskZilla

v2.13.21 — Round 1 audit cleanup

May 1, 2026

Fix Changelog idempotency — write_changelog() inserts atop existing prose, never clobbers.
Fix Fast-ack decommission honored — 4 dead Python files archived, 2 dead systemd units archived, sidecar deleted.
Fix Tapestry re-pointed from zombie collection to live taskzilla_global_default with legacy fallback.
Improved voxcpm tier corrected community → pro-managed (GPU-required); Kokoro stays the local default.
Improved retry_template.py extracted as shared scaffolding — midweek_check_retry + weekly_digest_retry shrink ~150 → ~70 LOC.
New architecture-guide registered Community-tier (was orphaned in canonical).

v2.13.20 — Audit-driven customer-tarball cleanup

May 1, 2026

Fix 38 internal dev scripts no longer ship in the customer tarball — scripts/_*.py, scripts/_*.sh, **/*.new excluded.
Improved Tarball: 854 entries (was 893), 1,968,327 bytes (was 1,995,117) — ~26 KB shrink.
Fix Tz-aware datetime sweep — scripts/postinstall.py:374 _fetched_at is now datetime.now(timezone.utc).

v2.13.16 — Pre-publication audit close + voice E2E live

April 28, 2026

New Voice round-trip first production fire — TTS via local Kokoro, STT via faster-whisper int8 daemon, Telegram admin-DM delivery verified.
Improved Workspace SOUL.md ships in the public tarball via surgical excludes (replaces blanket --exclude='workspace/').
Improved Voice-reply 30-day retention — timestamp+UUID8 filenames, _cleanup_voice_dir(keep=500, retain_hours=720).
Fix Regression gate tests/test_no_forbidden_patterns.py green at 25/25.

v2.13.14 — Session-memory daemon LIVE + PCR umbrella

April 28, 2026

New Session-memory AF_UNIX daemon at ~/.openclaw/run/session-memory.sock — 113 ms warm vs ~133 s cold subprocess (~1175× speedup).
New PCR (Parallel Context Retrieval) umbrella name for fast-ack + deferred-skills + mem-fanout sidecars; Phase 2.5 SCR daemon shipped.
Improved Per-user pulse-mood loop closed — DM tone derives from per-user mood with 14-day TTL.
Improved Model chain Option C — nemotron-3-super primary; per-skill model_registry.select_for_intent router with token-ceiling escalation.
Improved Shared embedder module at skills/shared/embedder.py — single OpenRouter qwen3-embedding-8b call site.

v2.13.0 — Salience Gate

April 25, 2026

New Post-turn write filter — A-MAC rules-then-LLM cascade + per-(user, session) seen-hash dedup; sidecar config at saliencegate.json.
New TZE1 v1 PQ license envelope — ML-KEM-768 + AES-256-GCM, master priv on license-server only.
New Magic-link revival flow — revoke responses include CTA + 1h-TTL redeem.

v2.12.7 — MaTTS (Memory-Aware Test-Time Scaling)

April 24, 2026

New arXiv:2509.25140 planning-trajectory selection algorithm — k-trajectory plan-execute fan-out, strategy distillation. Pro+ only, license.json gated.
New Channel adapter Phase 1 (v2.12.6) — skillkit.send() unified entrypoint with gateway-first + urllib-fallback.

v2.11.1 — Local LLM router

April 19, 2026

New LFM2-1.2B-Tool routes fixed-intent tool calls locally on Community tier; Pro+ stays cloud.
New setup_local_llm.sh for one-shot llama.cpp + LFM2 model bootstrap. 25/25 unit tests + 6/6 live LFM2 smoke tests.

v2.11.0 — Tier promotion wave

April 18, 2026

Improved 13 skills promoted to Community tier (admin-controls, button-hitl, knowledge-ingest, prompt-evolver, self-improving-agent, etc.).
New G9 trust-score gate — sits between distill and index, tags provenance, quarantines external_artifact writes.

v2.10.0 — Workflow induction + supervisor pattern

April 12, 2026

New Workflow induction (3 skills, 51/51 tests) — saved YAML matches skip full re-planning.
New Supervisor pattern for multi-agent fan-out.
New Voice loop closed (v2.10.1) — STT daemon + DM auto-voice-reply.

v2.9.4 — Fewer papercuts

April 6, 2026

Fix 54 code quality fixes: bare except elimination, silent error swallowing → debug logging.
Improved Ollama host now fully configurable via OLLAMA_HOST env var.
Fix Exception sweep across 42 files — no more silent failures in production.
New TROUBLESHOOT entries TZ-030 → TZ-033 added.

v2.9.3 — I grew up

March 31, 2026

New NIST + OWASP security audit — secret redaction, retry + circuit-breaker modules.
New Gateway rate limiting + load shedding, update rollback mechanism.
Improved PM adapter hardening — 6 providers with 30s timeout.
New Test suite expanded — regression gate at 25/25 GREEN.

v2.9.2 — I remember which files you rewrote

March 30, 2026

New evolution_ledger.py — unified log for variant-generator, prompt-evolver, self-improver.
New Provenance tagging in update-analyzer; evolution context in update-reviewer.
Improved _version.py as single source of truth for version across all modules.
Fix tz_logger migration (34 files), unused import cleanup (17 removals).

v2.9.0 — Forgetting on purpose

March 24, 2026

New FSRS power-law decay (arXiv:2402.09205), ACT-R activation (Anderson 2004).
New A-MAC 5-dimensional admission scoring (arXiv:2603.04549).
New HLR pruning envelope (Settles & Meeder, ACL 2016) + MemoryBank stability boost.
Improved 23/23 smoke tests passing. Memory is now research-grounded.

v2.8.2 — I stay smart when the internet doesn't

March 23, 2026

New Context engine hardening — offline resilience, two-factor decay, tz_logger adoption.
New Agentic update pipeline with human-in-the-loop resolver (164 tests passing).
New Self-improvement weekly cron — automated self-review with carry-forward tracking.
Improved Memory architecture — predict-calibrate distillation via Nemori.
Fix Session cleanup now covers all agents (main + taskzilla).

v2.8.0 — I work in teams now

March 13, 2026

New 3-tier team secrets (global / group / user) with governance audit trail.
New Anthropomorphism calibration — research-backed communication framework (15+ papers).
New ResponseRouter — outcome-to-GIF automation (39 battle tests passing).
Improved Cron system optimized — 21 crons, all Europe/Amsterdam timezone.
Improved Lean context strategy — system prompt down to 41KB.

v2.0.12 — I stop yelling at the logs

March 11, 2026

New ACP trigger hardening v2.0 — auth gate, agent allowlist, concurrency lock.
New Model registry — dynamic model selection across 8 live models (21 battle tests).
New Error handler — global alert system with voting buttons.
Fix Langfuse trace_span double-yield bug — exceptions now always propagate.
Fix Input data auto-truncated to 500 chars/field globally.

v2.0 — I keep your secrets where they belong

March 10, 2026

New Secrets vault — 18 secrets migrated, file provider with mode 600.
New Imagen 4 Ultra integration for image generation.
New Diagram & media generator — Mermaid via Kroki, emoji messages, Telegram delivery.
New Langfuse observability (SDK 3.14.5) with OTEL diagnostics export.
Improved LangGraph workflow skill with SQLite checkpointing.

How I've grown 🦖