Releases

2026-08-23

Desktop-first release with persistent tasks, reminders, profile switching, richer settings, imports, the skills marketplace and two-way visual preview editing.

• Persistent todo and outstanding-work stores
• Multi-source skills hub with parallel search
• Document generation for PPTX, XLSX and DOCX
• Desktop preview-control integration
• Tasks, notifications, profiles, imports and visual editing

2026-08-22

Approval hardening, loop automation, expanded evaluations, agent roles, workflow packs and CLI improvements.

• Write journal with rollback and permissions manager
• Loop engine with stop conditions and watchdog runs
• Expanded eval harness with a regression gate
• Advisor and agent-hub workflows
• Broader reusable skill library

2026-08-21

This minor release adds durable turn-bank recovery, a faster task harness and new business, personal, design and code-quality skill packs. No breaking changes.

• Durable version-2 turn-bank state: count-bound SHA-256 checkpoints, pending-sequence proof, compression lineage
• Strict restart recovery for stale, malformed, future and legacy state
• Task benchmark harness: median time, p90 time, token use, success rate, cost per successful task
• Built-in skill count increased to 181
• New skill packs: business assistant, personal assistant, game theory, mental models, AI-native UI, anti-slop TypeScript
• Memory-write results now require success=true and a clean staged flag

2026-08-06

This release upgrades the harness. It adds an eval gate on steer paths, tool-call metrics, a config-gated self-critique pass, a context-budget governor UI and flake dashboard ingestion. A new xavani tokens credential vault covers every product. The release pins five dependencies and completes the reliability backlog.

• xavani tokens: one credential vault (0600, doctor-validated) for every product
• Eval gate: golden steer-path evals block behaviour regressions in CI
• Tool-call metrics: per-session JSONL/CSV of every tool call (latency, success, retries)
• Self-critique pass: bounded final-answer review, config-gated (harness.self_critique)
• Context-budget UI: warn at 85%, block at 95% in /usage and the status bar
• Complete reliability backlog. Every planned update is in this release.
• Pinned deps: pydantic-settings, jsonschema, diskcache, structlog, orjson

2026-08-05

This patch release improves steer reliability end-to-end. It adds Python 3.14 compatibility and CI hardening (Windows footgun, Nix hash). It delivers the first part of the update programme.

• /steer verified end-to-end (TUI → gateway → agent drain)
• Python 3.14 memory-executor fix (6 tests restored)
• Crash forensics, memory/disk watchdog, per-session cost CSV
• Homebrew refresh automation, Docker healthcheck, maintenance window
• Full suite passed: 0 failures, 0 skipped

2026-08-05

This is the first public release. All internal development history was consolidated into this single release. Versioning now starts cleanly at 0.1.0. The complete update programme is implemented, tested and verified across seven pillars: Reliability & Correctness (20), Intelligence & Reasoning (15), Operability & Developer Experience (20), Safety & Guardrails (15), Observability & Debugging (10), Distribution & Ecosystem (10), Autonomy & Proactivity (10).

• Quantum Decision Cortex, the Oracle, and the Always-On Companion (deterministic, zero-LLM at the core)
• Zero-telemetry local agent gateway: MCP, memory layer, 30+ providers, cron, webhooks, messaging gateways
• Security stack: egress policy, sandbox hardening, Bandit/Gitleaks/Semgrep/Trivy CI
• 169+ skills, 90+ tools, research-guideline enforcement, R10 deterministic-first invariant