What's new in the v0.8 line: Stacks, canonical roles, and voice end-to-end

Sat, 04 Jul 2026 12:00:00 GMT

The v0.8 line runs from v0.8.0-beta.1 (2026-06-21) through v0.8.4b1 (2026-07-04) — eleven cuts. The throughline is config as a single source of truth: Stacks, canonical roles, portable profiles, and a resolved-argv endpoint all collapse “what is this slot actually running” down to one answer instead of several that could disagree. Everything below traces to the changelog — no roadmap promises, just what shipped.

Stacks: declarative config as the SSOT

Section titled “Stacks: declarative config as the SSOT”

v0.8.0-beta.1 introduces Stacks — a StackConfig schema plus a StackApplyEngine that plan()s a Stack into a ChangeSet, apply_config()s it as an atomic commit with rollback, and converge()s the live slot set (primary-slot load/swap/skip plus capability-child routing through the orchestrator). Stacks carry content-hash drift detection and an active-stack pointer, export/import through a checksummed .hal0stack.json envelope, and a snapshot path that captures your live config back into a Stack. Three seed stacks — saber, forge, pi — ship derived from the roster bench. Later fixes in v0.8.2b2 and v0.8.3b1 round this out: absent stack models can now be pulled (curated HF coordinates were registered for the seed-stack GGUF builds), and apply flags unresolved profile/model refs and reports degraded rather than a false “clean” when a slot fails to load.

Alongside Stacks, v0.8.0-beta.1 also single-sources slot launch argv: a resolver dedups the flag soup into a last-wins canonical command, and per-flag provenance is exposed at GET /api/slots/{name}/resolved — the slot edit drawer renders the resolved command with source badges (base / profile / extra_args) so you can see exactly which layer set which flag.

Canonical roles: `agent` and `utility` replace `chat` and `primary`

Section titled “Canonical roles: agent and utility replace chat and primary”

v0.8.0-beta.3 establishes two canonical LLM roles: agent (the capable default and dispatch fallback anchor, replacing chat) and utility (a cheap helper now seeded on every install, so extraction work no longer silently falls back to a heavy model). Any enabled type=llm slot is addressable as hal0/<slot>, and the advertised canonical virtuals are hal0/agent, hal0/utility, hal0/npu.

This work also finishes removing Cognee: Hindsight is now the only memory engine, with a PgVector boot-degrade fallback.

Voice, end-to-end

Section titled “Voice, end-to-end”

v0.8.0-beta.1 brings up the voice stack for real — voice_wire fixed, Open WebUI Call mode wired, and the NPU-trio facade auto-provisioning STT. v0.8.2b3 completes the picture on the synthesis side: a new Qwen3TTSProvider serves Qwen3-TTS from the canonical tts slot, and the voice.tts capability is a single switch that swaps the engine between Kokoro-82M (CPU) and Qwen3-TTS (GPU) — no separate slot to configure, no reconfiguration step. That release also ships a GPU benchmark harness (hal0-benchctl) and a standalone-to-slot migration runbook for anyone running an older Qwen3-TTS setup outside of hal0.

Portable profiles

Section titled “Portable profiles”

v0.8.2b1 gives profiles the same sharing model Stacks already had: export to a self-contained, checksummed .hal0profile.json envelope and import it on another host. New routes (GET /api/profiles/{name}, POST /api/profiles/{name}/export, POST /api/profiles/import) and matching MCP tools (profile_export is autonomous; profile_import/profile_delete are gated) mirror the stack_* set. The dashboard adds an Export button to every profile card and an Import dialog with a dry-run preview before commit.

v0.8.3b1: the 72-finding reliability sweep

Section titled “v0.8.3b1: the 72-finding reliability sweep”

v0.8.3b1 is a large reliability and completeness release — a 72-finding platform review remediated in eight verified waves, each regression-tested and gated by CI + Playwright. The user-visible defaults change too: the dashboard now shows Power & Thermal (live GPU clock/temp/power) and Per-Slot Throughput cards by default, disabling a capability actually stops it serving, and the bge reranker is finally classified and routed as a reranker instead of chat.

Underneath:

Interrupted pulls resume via HTTP Range/If-Range instead of restarting from zero, with the on-disk prefix re-hashed so the final SHA-256 stays exact.
Disk-space preflight fails fast with a structured model.insufficient_disk error before a multi-GB pull can fill the disk.
Host-memory-pressure LRU eviction unloads idle slots under pressure.
Cross-process safety — advisory file locks around registry and capability writes, plus parent-directory fsync after atomic writes.
Every backend-supported field the API already accepted (capabilities, backends, rope_freq_base, mmproj, hf_repo/hf_filename, per-slot vision, NPU asr/embed) is now exposed in the edit drawers.

v0.8.4b1: models, logs, and memory

Section titled “v0.8.4b1: models, logs, and memory”

The current cut, v0.8.4b1, is a safe, additive follow-up. A registry model can now declare a preferred runtime profile (defaults.profile) that a slot adopts on create and on every model swap, gated on device/type compatibility. The Models view gets a dedicated Image/ComfyUI toggle, and image-gen models are correctly tagged and grouped by their on-disk category. The slot edit drawer’s context field finally seeds from the persisted context_size instead of the live runtime metric or a hardcoded 16k, so a cold reload or model swap no longer clobbers your configured context window. Logs and events are unified into one surface with real source/slot attribution, and memory gains a durable audit trail for every destructive op — bank, memory, config, document, directive, operation, and mental-model deletes are now attributable after the fact — plus a response-shape guard so upstream drift surfaces as a loud error instead of a silently blank console panel.

Hermes durable memory, on by default

Section titled “Hermes durable memory, on by default”

v0.8.1-beta.1 makes cross-session recall a zero-config default: a fresh hal0 agent bootstrap hermes now provisions a working hal0-memory provider automatically, backed by two banks — private:hermes (default) and shared (cross-agent) — through hal0’s Hindsight engine. The agent gets hal0_memory_{search,recall,add} tools and auto-injects recalled context each turn with no manual config.

Next: v0.9

Section titled “Next: v0.9”

The v0.8 beta line has been an unusually long run of consolidation — Stacks, canonical roles, voice, portable profiles, and a 72-finding reliability sweep, capped by v0.8.2b4’s docs re-baseline against that current reality. v0.9 is coming soon, and its job is to take everything landed across these betas and stabilize it into the next line. Watch the changelog and releases page for the cut when it ships.

Try it

Section titled “Try it”

Terminal window

curl -fsSL https://hal0.dev/install.sh | bash

Already running an older line? hal0 update --channel stable (or nightly), and hal0 update --rollback reverts atomically. Full notes on the releases page and the update and rollback guide.

What's new in the v0.7 line: hal0's first beta

Fri, 19 Jun 2026 12:00:00 GMT

v0.7.3-beta.1 is hal0’s first beta. The headline is the dashboard: it graduates from a status panel into a full operations console. Everything below traces to the changelog — no roadmap promises, just what landed.

The dashboard is now an ops console

Section titled “The dashboard is now an ops console”

The inference, NPU, and ComfyUI cards are unified into one profile-card style. Profiles moved into the Slots tab under an engine-style header, the memory + throughput band was lifted above the tabs, and a live-journal footer now carries runtime and service-health groups. Slot cards came out of the accordion so they read at a glance.

ComfyUI image generation, first-class

Section titled “ComfyUI image generation, first-class”

A full ComfyUI generation engine is wired into the platform — model store, capability picker, installer, and a V2 Image-Gen pane. Generation flips the GPU into exclusive image mode via the iGPU switchover, and an Inference-tab dot tracks live state so you always know what the accelerator is doing.

NPU occupancy and the Operator Board

Section titled “NPU occupancy and the Operator Board”

The XDNA NPU gets a living occupancy grid with per-slot accents — it breathes with real activity instead of hiding behind a static picker. Alongside it, the Operator Board is a hal0-skinned kanban wired to the bundled Hermes agent (/api/board/*), with a live agent-chat drawer and working task creation.

Honesty under the hood

Section titled “Honesty under the hood”

The less flashy changes matter most for an operator’s tool:

Slot health is honest. A slot is marked ready only once its real /health passes — not on a systemd snapshot that lies about readiness.
Per-slot context is derived, never silently inheriting llama-server’s 4096-token default; the edit-drawer default is 16k.
Hardware reports the live GTT total, not a stale cached probe value — so the unified-memory pool you see is the pool you have.
Disabled-but-running slots are surfaced instead of vanishing.

Try it

Section titled “Try it”

Terminal window

curl -fsSL https://hal0.dev/install.sh | bash

Already running an older line? hal0 update --channel stable (or nightly), and hal0 update --rollback reverts atomically. Full notes on the releases page.

hal0 | Blog

What's new in the v0.8 line: Stacks, canonical roles, and voice end-to-end

Stacks: declarative config as the SSOT

Canonical roles: agent and utility replace chat and primary

Voice, end-to-end

Portable profiles

v0.8.3b1: the 72-finding reliability sweep

v0.8.4b1: models, logs, and memory

Hermes durable memory, on by default

Next: v0.9

Try it

What's new in the v0.7 line: hal0's first beta

The dashboard is now an ops console

ComfyUI image generation, first-class

NPU occupancy and the Operator Board

Honesty under the hood

Try it

Canonical roles: `agent` and `utility` replace `chat` and `primary`