diff --git a/src/content/docs/blog/whats-new-in-the-v0-7-line.mdx b/src/content/docs/blog/whats-new-in-the-v0-7-line.mdx
index 0b13bde..842c91b 100644
--- a/src/content/docs/blog/whats-new-in-the-v0-7-line.mdx
+++ b/src/content/docs/blog/whats-new-in-the-v0-7-line.mdx
@@ -14,6 +14,12 @@ tags:
 
 import { Aside } from '@astrojs/starlight/components';
 
+<Aside type="note" title="Historical post">
+This post covers **v0.7.3-beta.1**, a prior beta line. The current line is
+**v0.8.2b4** — see [What's new in the v0.8 line](/blog/whats-new-in-the-v0-8-line)
+for the latest features.
+</Aside>
+
 <Aside type="caution" title="Pre-release">
 hal0 is pre-1.0. **v0.7.3-beta.1** is a *beta* — the current development line,
 not a GA release. Expect breaking changes between cuts. See the
diff --git a/src/content/docs/blog/whats-new-in-the-v0-8-line.mdx b/src/content/docs/blog/whats-new-in-the-v0-8-line.mdx
new file mode 100644
index 0000000..485741c
--- /dev/null
+++ b/src/content/docs/blog/whats-new-in-the-v0-8-line.mdx
@@ -0,0 +1,105 @@
+---
+title: "What's new in the v0.8 line: declarative config, voice, and Hermes autonomy"
+date: 2026-07-04T12:00:00
+authors: hal0
+excerpt: >-
+  Declarative Stacks replace the flag soup, voice (TTS + STT) lands
+  end-to-end on the NPU, Hermes owns its own config, and hal0-api can
+  finally drop root. A look at v0.8.2b4, the current beta line.
+tags:
+  - release
+  - stacks
+  - voice
+  - hermes
+  - beta
+---
+
+import { Aside } from '@astrojs/starlight/components';
+
+<Aside type="caution" title="Pre-release">
+hal0 is pre-1.0. **v0.8.2b4** is a *beta* — the current development line,
+not a GA release. Expect breaking changes between cuts. See the
+[releases page](/releases) for the stable-vs-next split, and the
+[full changelog](/changelog#v0.8.0-beta.1) for the complete list.
+</Aside>
+
+`v0.8.2b4` is the current beta line. The headline is **declarative
+configuration** — Stacks replace the launch-flag soup with a single source
+of truth — plus **voice** on the NPU, **Hermes autonomy**, and the
+long-awaited ability to run `hal0-api` without root. Everything below
+traces to the [changelog](/changelog#v0.8.0-beta.1) — no roadmap promises,
+just what landed.
+
+## Stacks: declarative config, finally
+
+The slot config tangle — base flags, profile overrides, per-slot
+`extra_args`, and the diverging launch commands they produce — is now a
+single `StackConfig`. A `StackApplyEngine` plans changesets, applies them
+atomically with rollback, and converges the live slot set in one pass.
+Content-hash drift detection tracks when reality diverges from intent,
+and seeded stacks (saber / forge / pi) are derived from the bench so
+you start from a known-good baseline.
+
+![Stack config editor](/screenshots/dashboard-overview.png)
+
+Flag provenance is no longer guesswork. Every slot exposes its resolved
+command at `GET /api/slots/{name}/resolved`, and the Edit drawer badges
+each flag with its source — base, profile, or `extra_args`. No more
+hunting through three files to figure out which one set `--ctx-size`.
+
+## Voice: TTS + STT on the NPU
+
+The voice stack works end-to-end. `voice_wire` is fixed, Open WebUI Call
+mode is wired through the NPU trio, and STT auto-provisions alongside the
+FLM slot. The NPU occupancy grid now carries voice activity alongside
+inference — when Call mode is active, you see it breathe.
+
+## Hermes owns its own config
+
+Hermes is no longer pinned to a hal0-rendered `config.yaml` snapshot.
+A `config-set` overlay layers hal0's keys over Hermes's self-managed
+config, so Hermes can self-migrate across versions while hal0 guarantees
+the integration keys stay right. A real upgrade path (`hal0 agent upgrade
+hermes`) replaces the old tear-down-and-pray cycle.
+
+Two systemd units now: `hal0-agent@hermes` (the core runtime, loopback
+only) and `hermes-gateway` (the messaging platform gateway — Telegram,
+Discord — on `0.0.0.0:8644`). The gateway is a first-class service with
+its own lifecycle.
+
+## hal0-api drops root (opt-in)
+
+Set `HAL0_USER=hal0` and the API runs unprivileged. A declarative
+ownership table (auditable via `hal0 doctor perms`) plus a narrow `sudo`
+seam (`hal0-slotctl`) keeps the unprivileged API in control of slot units
+and container lifecycle while the process itself can't write config it
+shouldn't. Slot containers stay rootful — the container is the sandbox
+boundary. Default `HAL0_USER=root` is unchanged, so existing installs
+don't break.
+
+## The quality-of-life layer
+
+The less flashy changes that operators live with every day:
+
+- **Capability-based fallback.** When a slot's `model.default` isn't
+  locally servable, `load()` falls back to the best registered model
+  matching the slot's capability — no more dead slots because a file
+  moved.
+- **q8_0 KV cache everywhere.** Main and MTP-draft KV caches are now
+  `q8_0` across slots — near-lossless and keeps fused FlashAttention
+  on AMD HIP.
+- **Breaking rename — `hal0/primary` → `hal0/chat`, `hal0/flm` → `hal0/npu`.**
+  Virtual model names now map 1:1 to their resolution chains. Slot-name
+  back-compat is intentionally kept.
+- **Chat-template render-validation.** A broken template can no longer
+  ship silently — every template in the catalog is validated at build time.
+
+## Try it
+
+```sh
+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](/releases).
diff --git a/src/content/docs/docs/concepts/architecture.mdx b/src/content/docs/docs/concepts/architecture.mdx
index 67bebe2..7f2725c 100644
--- a/src/content/docs/docs/concepts/architecture.mdx
+++ b/src/content/docs/docs/concepts/architecture.mdx
@@ -61,9 +61,10 @@ state changes to the dashboard footer. These are shared across all requests —
 the API is a thin layer over them.
 
 <Aside type="note">
-hal0 ships with **no built-in authentication or TLS**. The API is open on the
-local network by design. See [Security](/docs/concepts/security) for the
-reverse-proxy pattern you put in front of it.
+hal0 ships with authentication and HTTPS **off by default** — pass `--auth=basic`
+at install to bring up Caddy with basic_auth, bearer tokens, and automatic HTTPS.
+When left off, the API is open on the local network. See
+[Security](/docs/concepts/security) for both postures.
 </Aside>
 
 ## Slots: where models actually run
diff --git a/src/content/docs/docs/concepts/security.mdx b/src/content/docs/docs/concepts/security.mdx
index 1b9af31..50b90c3 100644
--- a/src/content/docs/docs/concepts/security.mdx
+++ b/src/content/docs/docs/concepts/security.mdx
@@ -1,26 +1,28 @@
 ---
 title: Security posture
-description: hal0 ships no built-in auth or TLS — open on your LAN by design. Learn the threat model, the reverse-proxy pattern, and the MCP host and origin allowlists.
+description: Auth is opt-in at install time and off by default for trusted-LAN use. Learn both postures, the --auth=basic flag, the reverse-proxy pattern, and the MCP host and origin allowlists.
 sidebar:
   order: 70
 ---
 
 import { Aside, Card, CardGrid, LinkCard, Steps } from '@astrojs/starlight/components';
 
-hal0 makes one security decision explicit and central: it does **not** ship
-built-in authentication or TLS. The API binds `0.0.0.0:8080` and every endpoint
-is open to anything that can reach that port. This is a deliberate design choice
-(ADR-0012), and the rest of your security posture is built on top of it rather
-than inside it.
+hal0 makes one security decision explicit and central: authentication and TLS are
+**opt-in at install time** and **off by default**. Pass `hal0 install --auth=basic`
+to bring up Caddy as a reverse proxy with basic_auth, bearer tokens, and
+automatic HTTPS. Without it, the API binds `0.0.0.0:8080` and every endpoint is
+open to anything that can reach that port — the right posture for a
+trusted home LAN.
 
 <Aside type="caution">
-Do not expose hal0-api directly to the public internet. It has no auth gate —
-anyone who can reach the port can load models, run inference, drive agents, and
-read activity. Keep it on a trusted network, and put a reverse proxy in front of
-it for anything beyond a single trusted LAN.
+**Without `--auth=basic`**, do not expose hal0-api directly to the public
+internet — anyone who can reach the port can load models, run inference, drive
+agents, and read activity. Keep it on a trusted network, or install with
+`hal0 install --auth=basic` to enable Caddy with basic_auth, bearer tokens, and
+automatic HTTPS.
 </Aside>
 
-## Why no built-in auth?
+## Why auth is opt-in
 
 Earlier designs bundled an auth layer and a TLS-terminating proxy. That was
 removed (ADR-0012) because it duplicated, badly, what every operator already
diff --git a/src/content/docs/docs/getting-started/index.mdx b/src/content/docs/docs/getting-started/index.mdx
index fea4878..9dcdd59 100644
--- a/src/content/docs/docs/getting-started/index.mdx
+++ b/src/content/docs/docs/getting-started/index.mdx
@@ -19,10 +19,10 @@ OpenAI-compatible `/v1/*` requests to the right slot, and serves the
 dashboard. There is no shared inference daemon to babysit.
 
 <Aside type="note" title="Current release">
-This documentation tracks the **hal0 v0.7.x beta** line — the container-runtime
-era, where each slot (`chat`, `embed`, `rerank`, `stt`, `tts`, `img`, `vision`,
-and `agent` — plus the NPU-conditional slots `stt-npu` and `embed-npu` when
-FastFlowLM is installed) runs in a dedicated podman container. See the
+This documentation tracks the **hal0 v0.8.x beta** line — the container-runtime
+era, where each slot (`agent`, `code`, `img`, `npu`, `rerank`, `tts`, and
+`utility`) runs in a dedicated podman container behind an OpenAI-compatible
+API on port 8080. See the
 [changelog](/changelog) for the exact current version.
 </Aside>
 
@@ -69,9 +69,11 @@ FastFlowLM is installed) runs in a dedicated podman container. See the
   />
 </CardGrid>
 
-<Aside type="caution" title="No built-in auth or TLS">
-`hal0-api` binds `0.0.0.0:8080` and treats every request as authenticated
-by network reachability — the right posture for a trusted LAN appliance.
-If hal0 is reachable from anywhere you do not control, front it with a
-reverse proxy that owns auth and TLS.
+<Aside type="caution" title="Auth is opt-in — trusted LAN by default">
+`hal0-api` binds `0.0.0.0:8080` and, by default, treats every request
+as authenticated by network reachability — the right posture for a
+trusted home LAN. To safely expose hal0 beyond that, install with
+`--auth=basic` to enable a Caddy reverse proxy with HTTPS, basic_auth
+at the edge, and bearer tokens for the API. See
+[Security](/docs/concepts/security) for both postures.
 </Aside>
diff --git a/src/content/docs/docs/guides/connect-mcp.mdx b/src/content/docs/docs/guides/connect-mcp.mdx
index 7375104..a00184e 100644
--- a/src/content/docs/docs/guides/connect-mcp.mdx
+++ b/src/content/docs/docs/guides/connect-mcp.mdx
@@ -17,10 +17,16 @@ servers, mounted on the API as Streamable-HTTP sub-apps:
   server.
 
 <Aside type="caution">
-hal0 has **no built-in network auth** (ADR-0012). The MCP mounts are
-reachable to anything that can reach the API. Run hal0 on a trusted LAN
-or behind your own reverse proxy — do not expose `/mcp/*` to the
-internet.
+hal0 has **no built-in network auth by default** (ADR-0012). Without
+`--auth=basic`, the MCP mounts are reachable to anything that can reach
+the API. Run hal0 on a trusted LAN or behind your own reverse proxy —
+do not expose `/mcp/*` to the internet.
+
+When `--auth=basic` is active, the MCP mounts sit behind the same Caddy
+instance as the rest of the API — they inherit Caddy's TLS and basic_auth
+gating automatically. The host and origin allowlists below still apply
+at the MCP transport layer, but the outer auth gate is already handled.
+See [Security](/docs/concepts/security) for the two deployment postures.
 </Aside>
 
 ## Connect
diff --git a/src/content/docs/docs/guides/voice-stt-tts.mdx b/src/content/docs/docs/guides/voice-stt-tts.mdx
index 3b5ba24..5d3901f 100644
--- a/src/content/docs/docs/guides/voice-stt-tts.mdx
+++ b/src/content/docs/docs/guides/voice-stt-tts.mdx
@@ -12,8 +12,9 @@ hal0 exposes two voice directions through OpenAI-compatible endpoints:
 - **Speech-to-text** (`POST /v1/audio/transcriptions`) — `whisper-v3:turbo` served by
   FastFlowLM (FLM) on the XDNA NPU, co-loaded with the chat model in one FLM process
   (the "NPU trio": chat + STT + embed).
-- **Text-to-speech** (`POST /v1/audio/speech`) — Kokoro-82M ONNX container on CPU,
-  54 voices, MP3 by default (default voice: `af_heart`).
+- **Text-to-speech** (`POST /v1/audio/speech`) — two backends:
+  - **Qwen3-TTS** (ROCm, GPU): natural speech with voice cloning. Default on Strix Halo.
+  - **Kokoro-82M** (CPU, ONNX): lightweight, 54 voices, MP3 by default (default voice: `af_heart`). Runs without GPU.
 
 Both are children of the **`voice`** capability: `voice.stt` maps to the
 `stt` slot and `voice.tts` maps to the `tts` slot.
@@ -33,6 +34,15 @@ curl -X POST http://localhost:8080/api/capabilities/voice/stt \
   -d '{"enabled": true, "backend": "npu"}'
 ```
 
+</TabItem>
+<TabItem label="Text-to-speech (GPU, ROCm — Qwen3-TTS)">
+
+```sh
+curl -X POST http://localhost:8080/api/capabilities/voice/tts \
+  -H 'Content-Type: application/json' \
+  -d '{"enabled": true, "backend": "rocm", "provider": "qwen3-tts"}'
+```
+
 </TabItem>
 <TabItem label="Text-to-speech (CPU, Kokoro)">
 
@@ -49,7 +59,9 @@ The response is `{ "ok": true, "selection": { ...current selection... } }`.
 
 <Aside type="note">
 STT runs on the XDNA NPU via the FLM container — not on CPU. There is no CPU
-whisper.cpp backend in hal0. Kokoro TTS runs on CPU (ONNX). Neither uses ROCm.
+whisper.cpp backend in hal0. TTS has two options: Qwen3-TTS on ROCm (GPU, voice
+cloning, default on Strix Halo) and Kokoro-82M on CPU (ONNX, lightweight, 54
+voices). Neither STT backend uses ROCm.
 </Aside>
 
 ### How the NPU STT path works
@@ -96,10 +108,10 @@ As with transcription, `model` is required — a missing or empty `model`
 returns **400** (`request.missing_model`).
 
 <Aside type="note">
-Kokoro weights are operator-staged on the configured model store. The Kokoro
-toolbox is a containerized server managed by hal0 as a podman slot — no manual
-container management needed. It implements `POST /v1/audio/speech`,
-`GET /v1/models`, and `GET /health`.
+TTS weights are operator-staged on the configured model store. Both Qwen3-TTS
+(ROCm) and Kokoro (CPU) run as containerized servers managed by hal0 as podman
+slots — no manual container management needed. Each implements `POST
+/v1/audio/speech`, `GET /v1/models`, and `GET /health`.
 </Aside>
 
 ## See also
diff --git a/src/data/changelog.md b/src/data/changelog.md
index a0ec791..10654ce 100644
--- a/src/data/changelog.md
+++ b/src/data/changelog.md
@@ -8,7 +8,313 @@ v0.2) may carry breaking changes; patch releases inside a minor line
 
 Tags older than v0.2.0 ship release notes inside the GitHub release
 page; this CHANGELOG starts at v0.2.0 (the Lemonade migration cut).
-For ADR-level architecture context see `docs/internal/adr/`.
+ADR-level architecture decisions are kept internal (the `docs/internal/`
+tree is gitignored, #638) and referenced by number throughout the code.
+
+## [Unreleased]
+
+### Fixed
+- **Chat requests with mis-positioned or stacked `role='system'` messages
+  no longer 500 the Qwen3.6-35B-A3B upstream.** The Hermes dashboard SPA,
+  Open WebUI and LibreChat build the messages array by append, so a stale
+  session system message can land mid-array; third-party clients can also
+  stack two system blocks deliberately. The Qwen3 Jinja chat template
+  hard-raises `System message must be at the beginning` from line ~85 in
+  both cases, surfacing as HTTP 500 with `Jinja Exception: ...`. The
+  OpenAI-compat normaliser now calls
+  `hal0.normalize.messages.normalize_system_messages` inside
+  `_normalize_chat_body`, which **collapses** every system entry into one
+  and **hoists** the result to position 0 (joined with blank lines,
+  matching the OpenAI/Anthropic convention for stacked-system payloads).
+  No-system payloads are returned without a copy so the SPA's hot path
+  pays nothing.
+
+## [v0.8.2b4] — 2026-06-30
+
+Documentation and installer hygiene — no runtime behaviour change. This
+release re-baselines the engineering docs to the current v0.8.x reality
+(container runtime, Hindsight memory, agent/utility roles) and fixes a
+handful of installer drift bugs surfaced by a codebase assessment sweep.
+Safe upgrade from v0.8.2b3.
+
+### Changed
+- **Docs re-baselined to v0.8.x.** Swept the Lemonade→container-runtime and
+  Cognee→Hindsight terminology out of `AGENTS.md`, `ARCHITECTURE.md`,
+  `CONTEXT.md`, `PLAN.md`, and `README.md`; corrected the Hermes provisioner
+  to its real 15-phase pipeline; refreshed version/status lines; and rewrote the
+  (largely fictional) `hal0-service-management` codebase-map reference against
+  the current `src/hal0/` tree. Marked the shipped Stacks and voice-stack
+  superpowers plans as completed.
+- **Dead ADR links fixed.** Tracked docs no longer link into the gitignored
+  `docs/internal/adr/` tree (#638); surviving decisions are inlined and ADRs are
+  referenced by number. (In-code citation sweep tracked in #984.)
+
+### Fixed
+- **qwen3tts migration script aligned with the `tts`-slot model.** The
+  standalone-to-slot migration script's Guard 1 checked a non-existent
+  `qwen3tts` slot (always 404) and Guard 3 checked a removed kokoro `:8084`
+  fallback; both are corrected to the deployed design where Qwen3-TTS serves
+  from the canonical `tts` slot via `voice.tts`. (#979)
+- **`uninstall.sh` now removes all three sudoers grants.** It only removed
+  `hal0-benchctl`, leaking `hal0-agentenv` and `hal0-comfyui` behind on uninstall.
+- **Hermes private memory bank seeded under its canonical name.** `install.sh`
+  seeded `private__hermes-agent`, which the server (which derives the bank from the
+  agent-id via `PRIVATE_PREFIX="private:"`) never matches — corrected to
+  `private:hermes` so the pre-seeded retain-mission/dispositions actually apply.
+- **Dropped the obsolete Lemonade boot-contention comment** from
+  `installer/comfyui/scripts/comfy-up.sh`.
+
+## [v0.8.2b3] — 2026-06-29
+
+GPU text-to-speech lands: Qwen3-TTS now runs as a hal0-native slot with a
+one-switch swap between the Kokoro (CPU) and Qwen3-TTS (GPU) engines, plus a
+GPU benchmark harness and dashboard polish. Safe upgrade from v0.8.2b2.
+
+### Added
+- **GPU Qwen3-TTS as a hal0-native slot.** New `Qwen3TTSProvider` serves
+  Qwen3-TTS from the `tts` slot, with a `voice.tts` capability switch that swaps
+  the engine between Kokoro (CPU) and Qwen3-TTS (GPU) without reconfiguring the
+  slot. (#972, #976) The toolbox image builds + pushes to ghcr.io and its digest
+  is pinned in `manifest.json`. (#975, #977) Ships a standalone-to-slot migration
+  runbook + guarded script. (#974)
+- **GPU benchmark harness.** A GPU benchmarking toolbox with the `hal0-benchctl`
+  seam and accompanying agent skills. (#967) Dashboard gains an iGPU usage gauge
+  and a prefill TTFT readout. (#968) Topbar adds Kanban / Agent-Chat launchers, a
+  global agent chat, and an Archived lane. (#966)
+
+### Fixed
+- **Dispatcher injects upstream auth headers for remote providers**, so requests
+  routed to authenticated remote upstreams carry their credentials. (#973)
+- **UI builds land reliably on deploy** (no-cache index + install to the served
+  `dist`). (#969)
+
+### Docs
+- Benchmarking toolbox UI/feature handoff. (#971)
+
+## [v0.8.2b2] — 2026-06-24
+
+Two fixes on the 0.8.2 beta line — profile MTP tuning now actually takes
+effect, and stacks can pull their referenced models. Safe upgrade from
+v0.8.2b1.
+
+### Fixed
+- **Explicit profile spec flags win over the MTP bundle.** `resolve_profile_flags`
+  appended `MTP_FLAG_BUNDLE` after a profile's own flags, so the bundle's
+  spec-draft defaults (`--spec-draft-type-k q8_0`, `--spec-draft-p-min 0.0`)
+  silently clobbered any `--spec-draft-*` a profile pinned — there was no way to
+  tune the MTP draft through a profile. The bundle is now merged as defaults that
+  the profile's explicit flags override (gap-filled only). (#963)
+- **Absent stack models can be pulled.** Custom GGUF builds referenced by the
+  seed stacks (saber, pi-agent, qwopus coders, halostrix, gemma, …) were
+  auto-scanned with empty `hf_repo`/`hf_filename`, so on stack import/apply they
+  classified "unresolvable" with no download URL. Registered their public HF
+  coordinates (jcbtc/ + Jackrong/ + unsloth/ repos) in the curated catalogue;
+  `embed_references` falls back to curated on export; and a new
+  `backfill_coordless()` repairs existing coord-less registry rows on rescan. (#964)
+
+### Changed
+- **CI**: cancel superseded PR runs (never `main`); PRs test Python 3.12 only while
+  `main` runs 3.12/3.13/3.14 (3.14 non-blocking); least-privilege workflow
+  permissions; Node 20 → 22. (#898)
+
+## [v0.8.2b1] — 2026-06-23
+
+Profiles gain the same portable export/import/sharing model stacks already have.
+Safe upgrade from v0.8.1-beta.2.
+
+### Added
+- **Portable profile export/import.** A profile can now be exported to a
+  self-contained, checksummed `.hal0profile.json` envelope and imported on another
+  host — the same file-based sharing model stacks use. The envelope carries the
+  profile template plus a sha256 content checksum and an independent
+  `schema_version`; no secrets or host paths are serialized. New routes
+  `GET /api/profiles/{name}`, `POST /api/profiles/{name}/export`, and
+  `POST /api/profiles/import` (dry-run reports checksum validity + name collision;
+  commit creates under a chosen name, `409 profiles.exists` on a duplicate). New
+  MCP tools mirror the `stack_*` set: `profile_list` / `profile_status` /
+  `profile_export` (autonomous read) and `profile_import` / `profile_delete`
+  (gated). The dashboard adds an Export button to every profile card and an Import
+  dialog (file → dry-run preview → commit). (#962)
+
+## [v0.8.1-beta.2] — 2026-06-23
+
+Bugfix on the 0.8.1 beta line — restores fleet auto-update. Safe upgrade from
+v0.8.1-beta.1.
+
+### Fixed
+- **Updater version comparison uses PEP 440.** `hal0 update` compared versions with
+  a digit-tuple parser that split on `.` and stripped non-digits per segment, so the
+  pip-normalised installed beta `0.8.0b3` parsed to `(0, 8, 3)` and the tag-form
+  manifest `0.8.1-beta.1` to `(0, 8, 1, 1)` — `(0,8,1,1) > (0,8,3)` is false, so
+  every box on a `0.8.0bN` beta saw the new release as "not newer" and `hal0 update`
+  reported nothing to apply (the installed beta number was misread as the patch
+  component). The comparator now uses `packaging.version.Version` in both the updater
+  and the API route, falling back to the digit-tuple only for non-PEP-440 nightly
+  tags (whose timestamp ordering still relies on it). (#957)
+
+## [v0.8.1-beta.1] — 2026-06-23
+
+Installer/privilege simplification + Hermes durable memory on by default. The
+hal0-api privilege seams (hardened unprivileged mode, the slot privilege seam)
+are gone, and a fresh `hal0 agent bootstrap hermes` now provisions a working
+durable-memory provider out of the box.
+
+### Added
+- **Hermes durable memory enabled by default.** Provisioning now ships a working
+  `hal0-memory` provider and sets `memory.provider=hal0-memory`, so Hermes gets
+  cross-session recall with no manual config. Two banks — `private:hermes`
+  (default) and `shared` (cross-agent) — backed by hal0's Hindsight engine via
+  the hal0-api REST front door; reads union both banks. The provider exposes
+  `hal0_memory_{search,recall,add}` (with `shared=true` to write the shared
+  bank) and auto-injects recalled context each turn. (#955)
+
+### Changed
+- **Hermes agent identity is `hermes`** (was `hermes-agent`), matching the
+  `hal0 agent` registry. The agent-id is the single source for the
+  `X-hal0-Agent` MCP headers, the persona memory namespace, and the prelude. (#955)
+- **Memory plugin install path fixed.** The plugin is copied to
+  `$HERMES_HOME/plugins/hal0-memory/` (a direct child of `plugins/`, which the
+  Hermes loader actually scans) instead of the nested `plugins/memory/…` that
+  never loaded. (#955)
+
+### Removed
+- **Hardened (unprivileged hal0-api) mode removed**; hal0-api runs as root.
+  Dropped live-hello; fixed ready-summary IPs. (#953)
+- **Dormant slot privilege seam removed** (`hal0-slotctl` + euid routing). (#954)
+
+### Breaking
+- **Hermes memory namespace renamed `private:hermes-agent` → `private:hermes`.**
+  Existing `private:hermes-agent` data is not auto-migrated; reprovisioned agents
+  start recalling from `private:hermes` + `shared`. (#955)
+
+## [v0.8.0-beta.3] — 2026-06-23
+
+Canonical LLM roles + Hindsight-native memory extraction
+([ADR-0023](docs/internal/adr/0023-canonical-llm-roles-agent-utility.md)). The two
+canonical LLM roles are now **`agent`** (the capable default + fallback anchor,
+replacing `chat`) and **`utility`** (the cheap helper, now seeded on every
+install). `chat` and `primary` are retired as slot/role names.
+
+### Changed
+- **Canonical roles are `agent` + `utility`.** `agent` replaces `chat` as the
+  default/anchor everywhere (seeded slots, dispatch rule-9 fallback, the default
+  pin set, `_configured_primary`). `utility` joins `SEEDED_SLOTS` so a fresh box
+  never silently falls back to a heavy model for cheap extraction.
+- **Generalized virtual addressing.** Any enabled `type=llm` slot `X` is now
+  addressable as `hal0/X` (chain `(X, agent)`); the advertised canonical virtuals
+  are `hal0/agent`, `hal0/utility`, `hal0/npu`.
+- **Memory graph extraction is operator-selectable and actually wired.**
+  `[memory.graph].extraction_slot` names the local llm slot Hindsight uses for
+  graph extraction; hal0 propagates it to `hindsight-api` via a systemd drop-in
+  (`HINDSIGHT_API_LLM_MODEL=hal0/<slot>`) + restart. `hal0 memory graph enable`
+  takes `--slot <name>` (validated against the live enabled-llm-slot set).
+- **Cognee fully removed.** The Cognee engine + wrapper are deleted; Hindsight is
+  the platform engine (with a PgVector boot-degrade fallback). `MemoryRecord`
+  survives as an alias of `MemoryItem`.
+
+### Breaking
+- **`hal0/chat` is no longer advertised.** Clients pinned to `hal0/chat` (Hermes,
+  OpenWebUI, any custom consumer) must repoint to **`hal0/agent`**.
+  Hermes `model.default` is now `hal0/agent`.
+- **`memory.graph.route` / `memory.graph.upstream` removed**, replaced by
+  **`memory.graph.extraction_slot`** (default `"utility"`). Old `route`/`upstream`
+  keys in `hal0.toml` are silently dropped on load (no hard-fail on upgrade). The
+  `hal0 memory graph enable --route/--provider/--model` options are gone — use
+  `--slot`.
+- **`primary` is no longer a slot alias.** `SLOT_ALIASES` is `{"agent-hermes": "agent"}`.
+
+## [v0.8.0-beta.2] — 2026-06-22
+
+Bugfix release on the 0.8.0 beta line. No behaviour changes beyond the two
+fixes below — a safe upgrade from v0.8.0-beta.1. First release to carry #948.
+
+### Fixed
+- **Operator Board live updates restored.** The board's events-WS proxy
+  (`/api/board/events`) resolved its upstream Hermes session token from the
+  `HERMES_SESSION_TOKEN` env var only, while the REST path harvests the
+  rotating per-process token from the dashboard HTML. With no env pin (the
+  default), the WS connected upstream with no token, Hermes rejected the
+  upgrade (403), and the browser socket died with 1011 — so tasks created in
+  Hermes loaded on refresh but never pushed live to the board. The WS bridge
+  now shares the REST client's token resolution (env-pin → HTML-harvest →
+  rotation cache) and re-harvests + retries once on connect failure. (#949)
+- **Hermes privileged env seam.** A privileged env-write seam lets the
+  unprivileged provisioner write `root:root` `.env` files, so Hermes agent
+  config provisioning works under the dropped-root `hal0-api`. (#948)
+
+## [v0.8.0-beta.1] — 2026-06-21
+
+First beta of the 0.8.0 line — the model-config, Hermes, and permissions
+overhaul. Configuration becomes a declarative single source of truth
+(**Stacks** + single-source launch argv), Hermes consolidates onto its own
+config ownership, and `hal0-api` can finally drop root. Voice (TTS + STT)
+lands end-to-end. One behaviour change to be aware of: the `hal0/primary`
+and `hal0/flm` virtual aliases are gone — see **Changed**.
+
+### Added
+- **Stacks — declarative config SSOT.** 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 + capability-child routing through the
+  orchestrator). Content-hash drift detection and an active-stack pointer,
+  export/import via a checksummed `.hal0stack.json` envelope, snapshot of live
+  config into a Stack, and a `StacksCatalog` CRUD with seed guards. Seed stacks
+  (saber / forge / pi) derived from the roster bench.
+  (#921, #923, #925, #926)
+- **Voice stack — TTS + STT.** Brought up and verified end-to-end: `voice_wire`
+  fixed, Open WebUI Call mode wired, and the NPU-trio facade auto-provisions
+  STT. (#924, #928)
+- **Single-source slot argv (overhaul stream A).** A resolver dedups the launch
+  flag soup down to a last-wins canonical command; per-flag provenance is
+  exposed at `GET /api/slots/{name}/resolved`, and the slot Edit drawer renders
+  the resolved command with per-flag source badges (base / profile /
+  extra_args). (#929, #930, #931, #932)
+- **Capability-based slot fallback.** When a slot's `model.default` isn't
+  locally servable (registered-but-no-file, or pulled-away), `load()` falls back
+  to the best locally-registered model matching the slot's capability —
+  excluding diffusion / image / video models and preferring name-similarity to
+  the configured id. (#940, #942)
+- **Hardened permissions — run `hal0-api` unprivileged (opt-in).** Set
+  `HAL0_USER=hal0` and the installer drops the API off root: a declarative
+  ownership table (audited read-only by `hal0 doctor perms`), a narrow
+  privileged seam (`hal0-slotctl` + a no-wildcard sudoers grant) so the
+  unprivileged API can still write per-slot units and drive `hal0-slot@*`, and
+  a codified flip (run-as drop-in + recursive chown of config + state, pruning
+  `agents/` + `secrets/` + the models dir). Slot containers stay rootful — the
+  container remains the sandbox boundary. Default `HAL0_USER=root` is
+  byte-for-byte unchanged. (#929, #943, #944, #945)
+- **Hermes owns its own config.** The runtime is unpinned with a real upgrade
+  path, and a config-set overlay replaces the whole-file `config.yaml` render —
+  Hermes owns and self-migrates its config while hal0 layers only its keys.
+  (#934, #938)
+- **Chat-template render-validation.** The template catalog is render-validated
+  so a broken template can no longer ship silently. (#917)
+
+### Changed
+- **Breaking — `hal0/primary` and `hal0/flm` virtual aliases removed.** Virtual
+  model names now map 1:1 to their resolution chains; `hal0/primary` no longer
+  resolves (use `hal0/chat`) and `hal0/flm` is gone (use `hal0/npu`). Slot-name
+  back-compat is intentionally kept, and the Hermes overlay now emits
+  `model.default: hal0/chat`. (#939)
+- **q8_0 KV cache, universally.** Main and MTP-draft KV caches are now `q8_0`
+  across slots — near-lossless and keeps fused FlashAttention on AMD HIP. (#933)
+- **Profiles** lift bench-tuned MTP config into `rocm-moe` / `rocm-dnse`. (#922)
+- **Open WebUI** disables PersistentConfig so the env prewire wins the chat
+  connection. (#927)
+- Docs mirrored from hal0-web.
+
+### Fixed
+- **Installer** — `setup --storage-dir` is passed as a separate argv token so
+  fresh `--models-dir` installs seed slots correctly (#946); the hardened-perms
+  flip chowns config + state recursively so a root→hal0 upgrade doesn't strand
+  root-owned state subdirs (#945); `hermes gateway install` runs
+  non-interactively and skips `enable` when the unit is absent (#941); a
+  lemonade-team PPA is added so the FLM/NPU `.deb` resolves on a fresh Ubuntu
+  box (#937); the registry scans the effective store / pull_root rather than
+  only declared roots (#935).
+- **Hermes gateway** marks its `EnvironmentFile` optional (`-`) so fresh
+  installs don't crash-loop on a missing secrets vault. (#936)
+- **Dependencies** — bump vulnerable deps flagged by Dependabot. (#919)
 
 ## [v0.7.3-beta.2] — 2026-06-19
 
diff --git a/src/pages/index.astro b/src/pages/index.astro
index f3b76bf..cbb7435 100644
--- a/src/pages/index.astro
+++ b/src/pages/index.astro
@@ -97,10 +97,10 @@ const providers = [
     eps: '/v1/audio/transcriptions',
   },
   {
-    prov: 'Kokoro',
-    ver: '82M',
-    hw: 'CPU · Vulkan',
-    uses: 'text-to-speech · 54 voices',
+    prov: 'Qwen3-TTS',
+    ver: 'latest',
+    hw: 'ROCm',
+    uses: 'text-to-speech · voice cloning',
     eps: '/v1/audio/speech',
   },
   {
@@ -226,7 +226,7 @@ const loadoutData: Record<string, Array<{ size: string; title: string; vram: str
     { size: '~3 GB + NPU · realtime', title: 'Hands-free, iGPU-free', vram: '4 GB UMA (+XDNA)', slots: [
       { k: 'llm', v: 'qwen3.5-4b-q4kxl', d: 'Q4_K_XL · 2.9 GB · low-latency reply (or gemma4-it-e2b on the NPU)' },
       { k: 'stt', v: 'whisper-v3:turbo', d: 'runs on the XDNA NPU — frees the iGPU for chat' },
-      { k: 'tts', v: 'kokoro-v1', d: '82M · 54 voices · 8 languages' },
+      { k: 'tts', v: 'qwen3-tts', d: 'ROCm · natural speech · voice cloning' },
       { k: 'alt', v: 'vibevoice-1.5b', d: '1.5B · expressive long-form TTS' },
     ]},
   ],
@@ -308,7 +308,7 @@ const themes: RoadmapTheme[] = [
     blurb: 'The /v1/* surface and the engines behind it.',
     shipped: [
       { t: 'OpenAI-compatible /v1/* API', d: 'Chat, completions, embeddings, rerank, transcriptions, speech, images. Every OpenAI SDK works unchanged against the local box.' },
-      { t: 'Four-provider stack', d: 'llama.cpp (Vulkan / ROCm) for chat and embed, FLM for the XDNA NPU (chat + whisper-v3:turbo STT + embed, one process), Kokoro for TTS, ComfyUI for image generation.' },
+      { t: 'Five-provider stack', d: 'llama.cpp (Vulkan / ROCm / CUDA) for chat and embed, FLM for the XDNA NPU, Whisper v3:turbo for STT, Qwen3-TTS for speech, ComfyUI for image generation.' },
       { t: 'Image generation + iGPU switchover', d: 'POST /v1/images/generations served by a ComfyUI engine; generating flips the GPU into exclusive image mode and back, so chat and image share one accelerator.' },
       { t: 'FLM NPU provider', d: 'Self-contained NPU toolbox pinned by digest. The chat + STT + embed trio is surfaced only when XDNA hardware is present.' },
     ],
@@ -378,11 +378,11 @@ const themes: RoadmapTheme[] = [
   {
     id: 'security',
     name: 'Security posture',
-    blurb: 'No bundled auth — open on the LAN by default; front it with your own proxy.',
+    blurb: 'Auth is opt-in at install time. Off by default for trusted-LAN use.',
     shipped: [
-      { t: 'Open by default, proxy at the edge', d: 'No bundled auth or TLS (ADR-0012): the API binds 0.0.0.0:8080 for the LAN. Expose it by putting your own reverse proxy — Traefik, nginx, or a Cloudflare Tunnel — in front.' },
-      { t: 'Bundled chat UI + voice', d: 'Open WebUI on :3001 prewired to the local API, and whisper-v3:turbo STT (NPU, via FLM) + Kokoro TTS served through the capability API.' },
-      { t: 'Hands-free voice via Open WebUI Call mode', d: 'Open WebUI\'s built-in Call mode wired to hal0\'s /v1/audio/transcriptions and /v1/audio/speech endpoints — whisper-v3:turbo for listen, Kokoro for reply. No native hal0 streaming orchestrator; the loop runs inside Open WebUI.' },
+      { t: 'Opt-in auth at install', d: '--auth=basic brings up Caddy with basic_auth, bearer tokens for the API, automatic HTTPS (internal CA for .local, Let\u2019s Encrypt for real domains), and a best-effort Avahi service file. Zero certbot.' },
+      { t: 'Bundled chat UI + voice', d: 'Open WebUI on :3001 prewired to the local API, and whisper-v3:turbo STT (NPU, via FLM) + Qwen3-TTS (ROCm) served through the capability API.' },
+      { t: 'Hands-free voice via Open WebUI Call mode', d: 'Open WebUI\'s built-in Call mode wired to hal0\'s /v1/audio/transcriptions and /v1/audio/speech endpoints — whisper-v3:turbo for listen, Qwen3-TTS for reply. No native hal0 streaming orchestrator; the loop runs inside Open WebUI.' },
     ],
     soon: [],
     later: [],