Administrators can choose foreground or durable background execution for each agent independently of its skills. The default remains foreground, and current permissions still apply.
Edit a skill and its agent assignments as one operation. Skills can be available without being inserted into every prompt; they do not grant tool permissions.
Read current, permitted administrative facts during verification without exposing secret values.
Prepare an optional private business-data guard that stays off until explicitly configured, checks current source access and limits publication to the permitted audience.
Fixed
A verified background handoff or normal stage transition keeps its working status instead of displaying a false failure. Genuine and uncertain failures retain their existing safeguards.
Saved results, delegated work and resumed tasks retain their original actor, source and audience, with current access checks before use or delivery.
Preserve current direct/group message context, original file handles, duplicate intake protection and atomic local-user permission cleanup.
Changed
Prepare coordinated minor-version metadata for the official modules. Published catalog versions and signatures remain unchanged until actual artifacts are built, reviewed and promoted.
A history of notable AiHummer changes. The current line is v1.5.x. Each
entry below carries its date and the components it touched; the
upgrade guide explains how to move up.
[!NOTE]
This page lists changes that matter to users and administrators. The exact
build version is visible in the Admin UI and in the
API explorer (Swagger).
v1.4.5 — 2026-10-09
Security
Go 1.26.9 includes security fixes for the standard library.
Updated golang.org/x/net to 0.60.0 for HTTP/2 security fixes.
v1.4.4 — 2026-10-08
Fixed
OpenDesign and VoiceStudio preparation shows the current stage and downloaded bytes during installation, with clear failure stages and reasons that do not expose addresses, paths or secrets.
Preparing the studio cache no longer waits for an unrelated service-manager check; interruption cancels downloads and removes incomplete temporary preparation files.
v1.4.0 — 2026-10-02
Large saved attachments can be read in chunks and searched across extracted text. Excel keeps sheets, cell addresses and numbers; formulas are not executed. For a scanned PDF, the model sees the first four pages and discloses a partial view of longer files.
Migrate from OpenClaw in “Import”: upload an export, review the preview and apply your selection. A backup is made before applying; results and rollback are in the history. Warnings and skipped items need your decision; migration does not promise support for every third-party plugin.
Update Core in “System & settings” and modules in “Plugins”. Versions are independent: app and its module 1.1.6, Telegram 1.0.3, MAX 1.0.2. After updating, check your agent, attachments and channel on your own installation.
v1.0 — baseline
The baseline capabilities of the v1.0 line, as a reference point for later
changes.
A finished product, not a framework. A single self-hosted service combines
the control-plane and the turn engine; a typical install is one service plus
PostgreSQL.
Host-native, no Docker. Deployed as a release tarball managed by systemd;
no containers or orchestrators.
Multi-agent orchestration with sub-agents and reasoning strategies
(plan-steps, reflect, debate, best-of-N, self-heal).
First-class agents with personas, a structured prompt and a per-agent
model.
Long-term memory (Einstein) — facts extracted as reviewable claims and
promoted to memory; recall is wrapped in a data-fence against prompt injection.
Knowledge / RAG with citations, plus deep_research for multi-step
reports.
Channels — Telegram, Bitrix24, SIP telephony and the mobile and desktop
app, extended by installable connectors in later releases.
OpenAI-compatible API — POST /v1/chat/completions with SSE streaming.
No mandatory paid models — runs on free/local models and a Codex/ChatGPT
subscription transport; per-tenant BYOK is optional.
Security as a baseline property — an encrypted credential vault, enterprise
SSO (SAML / LDAP / SCIM / OIDC), Postgres Row-Level Security for tenant
isolation, human-in-the-loop approval gates and audit.
A plugin marketplace with one-click, host-native install.
## Core 1.5.0 — 2026-10-25
### Added
- Administrators can choose foreground or durable background execution for each agent independently of its skills. The default remains foreground, and current permissions still apply.
- Edit a skill and its agent assignments as one operation. Skills can be available without being inserted into every prompt; they do not grant tool permissions.
- Read current, permitted administrative facts during verification without exposing secret values.
- Prepare an optional private business-data guard that stays off until explicitly configured, checks current source access and limits publication to the permitted audience.
### Fixed
- A verified background handoff or normal stage transition keeps its working status instead of displaying a false failure. Genuine and uncertain failures retain their existing safeguards.
- Saved results, delegated work and resumed tasks retain their original actor, source and audience, with current access checks before use or delivery.
- Preserve current direct/group message context, original file handles, duplicate intake protection and atomic local-user permission cleanup.
### Changed
- Prepare coordinated minor-version metadata for the official modules. Published catalog versions and signatures remain unchanged until actual artifacts are built, reviewed and promoted.
A history of notable AiHummer changes. The current line is **v1.5.x**. Each
entry below carries its date and the components it touched; the
[upgrade guide](/en/v1.5/operations/upgrade-policy/) explains how to move up.
> [!NOTE]
> This page lists changes that matter to users and administrators. The exact
> build version is visible in the Admin UI and in the
> [API explorer (Swagger)](/api/).
## v1.4.5 — 2026-10-09
**Security**
- Go 1.26.9 includes security fixes for the standard library.
- Updated golang.org/x/net to 0.60.0 for HTTP/2 security fixes.
## v1.4.4 — 2026-10-08
**Fixed**
- OpenDesign and VoiceStudio preparation shows the current stage and downloaded bytes during installation, with clear failure stages and reasons that do not expose addresses, paths or secrets.
- Preparing the studio cache no longer waits for an unrelated service-manager check; interruption cancels downloads and removes incomplete temporary preparation files.
## v1.4.0 — 2026-10-02
Large saved attachments can be read in chunks and searched across extracted text. Excel keeps sheets, cell addresses and numbers; formulas are not executed. For a scanned PDF, the model sees the first four pages and discloses a partial view of longer files.
Migrate from OpenClaw in “Import”: upload an export, review the preview and apply your selection. A backup is made before applying; results and rollback are in the history. Warnings and skipped items need your decision; migration does not promise support for every third-party plugin.
Update Core in “System & settings” and modules in “Plugins”. Versions are independent: app and its module 1.1.6, Telegram 1.0.3, MAX 1.0.2. After updating, check your agent, attachments and channel on your own installation.
## v1.0 — baseline
The baseline capabilities of the v1.0 line, as a reference point for later
changes.
- **A finished product, not a framework.** A single self-hosted service combines
the control-plane and the turn engine; a typical install is one service plus
PostgreSQL.
- **Host-native, no Docker.** Deployed as a release tarball managed by systemd;
no containers or orchestrators.
- **Multi-agent orchestration** with sub-agents and reasoning strategies
(plan-steps, reflect, debate, best-of-N, self-heal).
- **First-class agents** with personas, a structured prompt and a per-agent
model.
- **Long-term memory (Einstein)** — facts extracted as reviewable claims and
promoted to memory; recall is wrapped in a data-fence against prompt injection.
- **Knowledge / RAG** with citations, plus `deep_research` for multi-step
reports.
- **Channels** — Telegram, Bitrix24, SIP telephony and the mobile and desktop
app, extended by installable connectors in later releases.
- **OpenAI-compatible API** — `POST /v1/chat/completions` with SSE streaming.
- **No mandatory paid models** — runs on free/local models and a Codex/ChatGPT
subscription transport; per-tenant BYOK is optional.
- **Security as a baseline property** — an encrypted credential vault, enterprise
SSO (SAML / LDAP / SCIM / OIDC), Postgres Row-Level Security for tenant
isolation, human-in-the-loop approval gates and audit.
- **A plugin marketplace** with one-click, host-native install.
