hal0 v0.9.8-nightly.20260720084155

Nightly v0.9.8-nightly.20260720084155 — changes since v0.9.8-nightly.20260713085026:

• feat: define 1.0 workload-oriented seeded profiles (#1324) (e316360)
• feat(brain): first-class hal0-brain module + /api/brain/chat primary route (#1327) (a314dcb)
• Merge remote-tracking branch 'origin/main' into rework/p3-brain (8a46421)
• docs(board): sync inc-3 + FLAGS-own rows to merged state (#1326) (3ddd9f5)
• feat(slots): §7 slot-purity — fold chat_template into model; sunset the slot tier (#1325) (9bde4b7)
• feat(mcp): autogenerate admin route-map from live routes (route-id keyed, deny-by-default) (#1323) (a894188)
• feat(api): typed request bodies for models/slots command routes (#1322) (28f497c)
• Merge pull request #1321 from Hal0ai/rework/descar (56e06ae)
• fix(dashboard): stop slot cards flickering bare→enriched (#1320) (cf55333)
• docs(board): P3-runtime-db inc0-3 — bilingual id/name slot layer + seed split-brain fix (83374a5)
• feat(slots): bilingual name-or-id on-disk slot layer + identity-keyed seed (P3-runtime-db inc0-3) (c8820cf)
• docs: install & migration guide (styled) + link from getting-started + v1.0.0 (R5) changelog (69dc413)
• fix(dashboard): stop slot cards flickering bare→enriched on poll (4f309b2)
• docs(board): SLOT-B id-flip root-caused — id-keyed read-path is unbuilt (= P3-runtime-db lane) (534b1e8)
• docs(deploy): SLOT-B id-flip rehearsed on 143 — BLOCKED on runtime-version sequencing (cafcf99)
• docs(board): Phase 4 deploy hardening — m3-renderer + migrator-ctx code fixes (9d23450)
• fix(migrate): drop launch-shadowed context_size from slot-flags fold + divergence key (6f8266d)
• fix(gpu): derive GPU group-add GIDs from device-node owner, not group name (0b4e678)
• docs(deploy): R5 deploy-window record — §2 flag regression fixed (3 slots), rocm/vulkan bench (da9f41c)
• chore(descar): sync with main f2db4d6 (post-R5-collapse) (2a067c2)
• chore(descar): unfreeze — descar→main R5 promotion complete (main f2db4d6) (66563c4)
• Collapse rework/descar → main: R5 (Phase 3 complete — memory/Hermes, SWEEP, FLAGS-own centerpiece, deploy-validation) (f2db4d6)
• chore(descar): FREEZE marker — descar→main promotion in progress, hold pushes (1490739)
• Merge remote-tracking branch 'origin/rework/descar' into claude/hal0-rework-handoff-o4fcws (ea938bd)
• style: ruff format agent_commands.py (CI format gate) (cbc0fb7)
• style(cli): ruff format agent_commands.py (fix-forward wave format-check) (d1b5af3)
• docs(board): Phase 3 wave-4 (deploy fix-forward) — quadlet-restart, installer-gates, cli-hermes (0d0af38)
• fix(cli): agent status --json output + hermes ownership-drift doctor check (0aa0ca0)
• fix(installer): accurate keyring-quota diagnostic + require render group in GPU preflight (6d154bf)
• fix(quadlet): emit StartLimit* in [Unit] so slot restart-limiting applies (205d526)
• docs(handoff): R5 Phase 4 handoff — deploy window → v1.0.0 release cut (12a8c89)
• docs(deploy): R5 install-validation record — 471c365 on halo150/143 (8d15b04)
• fix(tests): migrate unit-rerender test to FLAGS-own semantics (471c365)
• docs(board): Phase 3 wave-3 merged — FLAGS-own centerpiece + migrator + v1.0.0 re-target (8903a43)
• chore(sunset): re-target deprecated-surface stamps v0.10.0 → v1.0.0 (0a0bc6e)
• feat(flags): flags own by models — strip slot flag/device/template surface + copy-on-stamp profiles + migrator (d4253f8)
• docs(board): Phase 3 wave-2 merged — reconcile-wire, sunset-ratchet, persona-flag (b486d7e)
• feat(agent): add explicit persona-overwrite flag to install hermes (5bf5a74)
• chore(sunset): ratchet deprecated flags/aliases/fields for next-release removal (48a9b48)
• fix(ports): wire reconcile_listeners into /api/ports so the reconcile pass runs (cdb9ef8)
• docs(board): Phase 3 wave-1 merged — mem→hindsight, brain-lane→lifespan, drift-watch+hermes-bump, SWEEP dead-code (8ffc990)
• refactor(boot): relocate 5 brain-lane steps from install into api lifespan (577ee46)
• feat(memory): rename hal0_memory_* → hindsight_{recall,retain,reflect} + local_external config (4c2e5f1)
• test(drift): add drift-watch fixtures + docs(runbook): hermes-bump procedure (23c9242)
• refactor(sweep): delete 4 grep-confirmed-dead symbols (7e06260)
• docs(board): Phase 2 wave-2 merged — P3-routers inc3 pull-extraction (dc0f5cf)
• refactor(models): extract pull/update-pull orchestration to pull_jobs service (15b9d34)
• docs(board): Phase 2 wave-1 merged — lifespan-split + settings-seam (4e711d5)
• fix(settings): wire the 2 placeholder pages onto real typed hooks (0b2417e)
• refactor(api): phase-split boot lifespan into typed, observable stages (a9b1ea9)
• docs(board): Phase 1 COMPLETE — wave-3 merged (dup+mock, UI-API-2, CONTRACTS) (59f6983)
• refactor(ui): useAuthExposure uses ENDPOINTS.authExposure const (no literal dup) (67d1dde)
• feat(ui): wire DuplicateModelDialog to POST /api/models/{id}/duplicate + realign mock fixtures (c156f81)
• fix(ui): wire ExposureTable to live GET /api/auth/exposure (UI-API-2) (55d1da4)
• docs(ui): reconcile endpoints.ts/CONTRACTS.md self-contradictions (SWEEP §2) (7a81af9)
• docs(board): Phase 1 wave-2 merged — missing-routes, NAMES-stale, slot-logs-redact (3a852df)
• feat(api): implement 3 declared-but-missing GET routes (0941df5)
• fix(ui): NAMES-stale — role-based slot selection over name literals (8a95b8c)
• fix(slots): redact secrets in /api/slots/{name}/logs/stream (bc2954e)
• docs(board): Phase 1 wave-1 merged — MCP-sync, CLI verbs, UI endpoints, VERS-flash, api-logs-redact (43ea03d)
• fix(logs): correct false claim that slot logs share journalctl_sse() (d1c78d7)
• fix(logs): redact bearer/client_id/_KEY secrets in /api/logs stream (2b5662d)
• feat(mcp): hand-add §4.3 tier-a admin catalog tools + reclassify memory reads (645aa49)
• refactor(ui): consolidate dead endpoint constants (SWEEP §2) (024147f)
• feat(cli): wire the missing R3/R4 verbs — auth, board, ports, slot rename, model default/update/pull-cancel (§5.2) (21153e4)
• fix(ui): kill VERS-flash — reconcile UI version to backend, build-time inject (dbf9a63)
• docs(board): Phase 0 γ fallout fixed + descar reconciled with main (b5ac6b0)
• Merge remote-tracking branch 'origin/rework/descar' into rework/descar (6466994)
• Merge remote-tracking branch 'origin/main' into rework/descar (24cd248)
• merge: combine both activity-log de-flakes — their toHaveCSS + mock warming, my emitAndAwait for the emit-loss family (abb6db2)
• fix(e2e): de-flake activity-log — emits raced the pane's remount/reconnect gap (9d362d2)
• docs(rework): drive-2 fable handoff + SWEEP inventory (overlooked-surface backlog) (4ee5b07)
• fix(ui): warm the mockFixtures import + de-race ActivityLog bounded-scroll spec (979f58e)
• docs: R5 sync assessment + graphify knowledge graph + drive-2 handoff (#1317) (f2d82d3)
• fix(e2e): memory Recall/Reflect specs page.route their POSTs — mock substitution is GET-only now (68181bc)
• docs(board): Phase 0 (Wave 0) landed — 6 correctness/security lanes ✔ (278b32a)
• merge(phase0): thread service bearer into provisioned Hermes MCP + memory calls (1c01dba)
• fix(hermes): thread service bearer into _mcp_memory_call (auth-on 401) (f1af296)
• merge(phase0): ship hal0.target, model-store PermRow, uninstaller sync (1bb9bf7)
• fix(install): ship hal0.target, add model-store PermRow, sync uninstaller to R3/R4 (b9c0b0e)
• merge(phase0): gate prod mock fallback by method+env, lazy-import fixtures (54ccd58)
• fix(hermes): thread service bearer into provisioned MCP headers (auth-on 401 fix) (23cb35c)
• fix(ui): gate prod mock fallback by method + env, lazy-import fixtures (268f96a)
• merge(phase0): CLI auth-bypass fixes + deprecated-verb help texts (d55aa36)
• fix(cli): route streaming verbs through authed helper; fix deprecated-verb help texts (1dedecd)
• merge(phase0): MCP PATCH support + _REST_MAP route-sync test (b878ff9)
• fix(mcp): support PATCH in _call_rest + pin _REST_MAP to real routes (65fe49f)
• merge(phase0): MCP client_id leak fix — hashed label, not raw bearer (5d7cd28)
• fix(mcp): derive client_id label from hash, stop stamping raw bearer into journald (54a19c8)
• docs(rework): land R5 endgame handoff + sync assessment onto descar (9b05dee)
• docs(graphify): commit curated knowledge-graph artifacts (report + wiki + analysis) (2ab6ae6)
• docs(handoff): route 8 graphify structural findings into R5 doc surfaces (cb16294)
• docs(rework): handoff — route merge/meld/fix follow-ups into the R5 docs (4dbd1ed)
• docs(board): UI-API-1 → ✔ (PR #1316 merged to descar 4f828fe) (270a35a)
• docs(board): UI-API-1 row points at PR #1316 (built) + descar→main rebase note (bec70ce)
• docs(board): P4-docs/P4-rules/§21.11 golden-paths → ✔; UI-API-1 audit finding recorded (4a2ee61)
• test(golden-paths): complete 15-scenario coverage map + deploy runbook (6aadb8d)
• docs(P4-rules): add anti-scar rules 10-11 (bit-twice CI lessons) (a26fdef)
• docs(P4-docs): collapse CONTEXT/AGENTS stubs into ARCHITECTURE.md (6e3c168)
• feat(models): UI-API-1 — close the managed-arg gap (launch + create + /validate) (#1316) (4f828fe)
• feat(models): HF update-check basename fallback + a11y + e2e coverage (#1315) (462b28f)
• Collapse rework/descar → main: HP-realtime inc-1 + default consumers + docs cleanup (b70b981)
• docs(board): HP-realtime inc-1 ✔ merged — voice endpoint live on descar (1af40b0)
• merge rework/realtime-inc1: WS /v1/realtime — server VAD + both LLM legs + MCP recipe (HP-realtime inc-1) (65df5cd)
• docs(board): default-consumers + docs-stragglers ✔ merged (93be1ab)
• test(realtime): event-contract suite + voice guide (HP-realtime inc-1) (7a2a820)
• feat(realtime): WS /v1/realtime engine core (HP-realtime inc-1) (62d2358)
• merge rework/default-consumers: fallback prefers the type's default model; create modal preselects it (57fee5b)
• feat(ui): preselect a slot's type default model on create (9aaf9b6)
• feat(registry): fallback_local_model ...

hal0 v0.9.8

Turnstone lands as a second heavyweight bundled agent alongside Hermes and
becomes a first-class companion service; the unified hal0-rocmfpx runner
becomes the default image for AMD GPUs (with an automatic slot migration on
update); and a memory security fix stops one agent deleting another's private
memories.

Highlights

• Turnstone — a second heavyweight bundled agent joins Hermes: a native turnstone-server on loopback :9129, installed into its own managed PyPI venv, coexisting with Hermes via relaxed single-pick (#1299).
• Turnstone is a first-class companion service — it shows in the Services pane and the Overview health card with start/stop/restart controls, next to Hermes/Hindsight/OpenWebUI.
• hal0-rocmfpx is now the universal default runner for AMD GPUs — one unified image (Vulkan/RADV + HIP) replaces the per-lane toolboxes; CUDA and CPU-only lanes keep their lean images (#1297).
• Memory: private-visibility is now enforced on delete — in unified-bank mode one agent could delete another agent's visibility:private memory by id; delete now applies the same fail-closed ACL as read/search/list.

Migrations

• hal0 update automatically re-pins existing AMD-GPU slots from the old amd-strix-halo-toolboxes images to the unified hal0-rocmfpx runner (no-op on CUDA/CPU lanes) (#1297).

Added

• Turnstone bundled agent — provisioning pipeline (managed PyPI venv, JWT-secret generation, model automap) plus hal0 provider/memory wiring, coexisting with Hermes (#1299).
• Turnstone companion-service registration — a ServiceDef + systemd health probe, so turnstone appears in /api/services (Services pane, full lifecycle actions) and /api/services/health (Overview card + sidebar status).

Changed

• hal0-rocmfpx as the default AMD-GPU image — basic seed profiles defer to a manifest-driven resolver that returns the unified runner for AMD lanes, the CUDA image for NVIDIA, and the lean toolbox for CPU-only (#1297).
• PyPI distribution is published under the name hal0ai (the import package and hal0/hal0-agent console scripts are unchanged) (#1298).

Fixed

• Memory delete ACL — delete enforces the visibility:private owner check in unified-bank mode, so an agent can no longer delete another agent's private memory by (guessable) document id; unresolved ids are withheld fail-closed (#1302).

hal0 v0.9.8-nightly.20260713085026

Nightly v0.9.8-nightly.20260713085026 — changes since v0.9.3-nightly.20260708080920:

• release: v0.9.8 (#1304) (402a772)
• feat(images): rocmfpx as universal default + hal0 update slot migration (#1297) (aff99e5)
• feat(agents): add turnstone as a heavyweight bundled agent (coexists with hermes) (#1299) (aac5514)
• build(pypi): publish under distribution name "hal0ai" (#1298) (40d11d3)
• release: v0.9.7.3 (#1296) (dd888d2)
• claude/hal0 hardening honcho2 (#1295) (4338739)
• fix(installer): harden Honcho memory standup (F24/F26/F27/F28) (#1294) (c7c82a2)
• fix(installer): init hc_cur_tag before use (unbound var aborted Honcho standup) (#1293) (979e29e)
• fix(preflight): container smoke runs image bare, not ... true (distroless hello has no /bin/true → false-fail aborted install) (#1292) (7f41684)
• fix(installer): harden install.sh + preflight.sh for environment robustness (#1291) (26ef24e)
• fix(agents): hermes single-pick bypass + opencode installer uninstall/PATH gaps (#1290) (0c54ca6)
• fix(container): docker/podman runtime portability (--replace, runtime probe, OpenWebUI skip reason) (#1289) (e94011d)
• fix(pymisc): honcho remediation opt-in path + guarded first-run sentinel write (#1288) (06582c8)
• fix(comfyui): FHS-aware fetch script + workflow asset resolution for wheel installs (#1286) (e1da0f4)
• fix(uninstall): sudo env forwarding + Honcho stack teardown (#1287) (4f49f26)
• fix(agents): FHS-aware installer path + crash-safe agent switch (#1285) (e3852f4)
• fix(doctor): FHS-aware preflight.sh resolution + skip own-port false failure (#1284) (01a7cb9)
• fix(memory): honcho render-env tolerates missing hal0-honcho unit (opt-in) (#1283) (628e8f2)
• test(coverage): entry-point smoke tests for five 0%-coverage modules (#1282) (84a21e7)
• docs: sweep stale ADR/Cognee refs, deprecate gttsize, document fresh-box journey (#1280) (ba72db2)
• feat(memory): dry-run item-count preview on unconfirmed bank delete (#1273) (6ea9d2e)
• fix(comfyui): ESRGAN source, hf-CLI compat, curated model orchestration (#1277) (b590ff2)
• test(hermeticity): isolate 19 tests that leaked live-host state (#1281) (81e450c)
• docs(adr-0023): scrub stale hal0/chat default refs → hal0/agent (#1276) (0da9f6e)
• docs(installer): loud LAN-only warning on the 0.0.0.0:8080 bind (#1275) (9b0162d)
• upstream controls + registry/UI: close #1148 #1150 #1157 + #1256 audit (#1279) (8a054d6)
• fix(cli+update): editable-install refusal, footgun gates, command consolidation (#1274) (baf386d)
• fix(slots): errored-slot restart recovery + drift/image_status false positives (#1278) (ba5a4f5)
• ci(agents): add opencode lane to the nightly agent-shim smoke (#1272) (53aae72)
• fix(slots): auto-recover a slot wedged in WARMING (stale-anchor watchdog) (#1269) (e9639de)
• fix(honcho-migrate): retry transient embed 5xx and surface the cause (#1268) (0f46fcc)
• feat(agents): add opencode as a bundled agent (hal0 provider + hindsight memory) (#1271) (3ab82e8)
• fix(mcp): give each gated tool call a unique approval id (#1270) (f3e1f02)
• fix(profiles): re-tune seed flags, add Vulkan embed/rerank lanes, harden seed guards (#1266) (a2c3a85)
• feat(agents): provision hal0-brain as a first-class profile, not just a persona (#1258) (ca016b1)
• fix(board-chat): reliable steward tool-calling — text-parse fallback + tool_model routing (#1265) (ae1740c)
• feat(memory): re-land Honcho provider UI on current main (supersedes closed #1253) (#1267) (e36ed5e)
• refactor(mcp): share per-tool param schemas between board_chat and admin MCP (#1264) (1380e39)
• fix(bench): resume eval runs across defers instead of restarting + duplicating records (#1261) (3abdc71)
• fix(hermes): restore uv Python provisioning reverted by #1257 + make UV_PYTHON_INSTALL_DIR traversable (#1259) (0a7de78)
• fix(ui): don't render pi-mark with empty src when agent art is missing (#1263) (e4fedc6)
• fix(memory-cli): tolerate operation_id key in ops retry + scope migrate-unify retag to transferred docs (#1262) (b5f184f)
• fix(memory): enforce visibility:private on read so private memories aren't cross-agent readable (#1260) (ad4c58d)
• feat(cli): hal0 memory bank/ops/mm/recall + migrate unify (rebased onto main) (#1257) (00bf2e1)
• fix(memory): consolidate drifted Hermes memory plugin + stray-bank identity fix (#1245) (a73a2c2)
• feat(memory): unified-bank model with server-side tagging (#1244) (b80911f)
• feat(memory): self-hosted Honcho v3 as a per-agent memory provider (unified memory) (#1243) (05efc59)
• feat(agents): pi-coder agent — provisioning, hal0 provider/memory plugins, live Pi dashboard card (#1254) (f2f7c13)
• feat(bench): queue dropdown with lane, tool-eval, and tune options (#1255) (ae7ebcf)
• release: v0.9.7.2 (fd9f2d5)
• feat(hermes): provision a uv-managed Python when no system 3.11-3.13 exists (#1252) (29a4887)
• fix(hermes): require a 3.11-3.13 venv interpreter + floor hermes-agent at 0.16.0 (#1251) (2c91d3d)
• release: v0.9.7.1 (b51062d)
• test(e2e): stub /api/upstreams so the γ connections-v3 suite stops blanking (#1242) (bda3e9a)
• feat(memory): retry-failed-extractions button, consolidate fix, mental-model delete (#1235) (4b92283)
• feat(memory): replace HAL0_MEMORY_ENABLED env var with [memory].enabled + CLI toggle (#1240) (2a0750e)
• fix(hermes): auto-provision on fresh installs (unclaimed home + self-foreign gateway) (#1239) (ff370a1)
• feat(setup): keyboard-driven guided-setup TUI (arrow keys + clean pickers) (#1237) (0ed4d8b)
• feat(upstreams): add MiniMax and DeepSeek providers + fix filter spacing (#1236) (ee87dbc)
• release: v0.9.7 (baef535)
• docs(agents): add the hal0-brain steward section — how it works, why it runs on a 1B model, MiniCPM5-1B recommendation (#1234) (26b8526)
• fix(release): sign with Sigstore bundle so client verify survives cert expiry (#1159) — HOLD, needs test-release validation (#1189) (c0d5538)
• docs(release): v0.9.7 changelog + README updates (#1233) (d78486e)
• fix(api): stop model pulls from blocking a graceful hal0-api restart (#1225) (#1232) (9a14f5e)
• fix(npu): finalize trio routing aliases in /v1 models route (#1231) (71723b6)
• feat(seed): agent + brain slot seeds (#1230) (825290b)
• feat(flm): NPU FLM naming migration (#1229) (2533bd1)
• feat(upstreams): upstream model controls — reactive CRUD, model filters, enabled kill-switch, CLI + MCP + dashboard (#1228) (9af8b23)
• feat(brain-chat): Agents/Brain settings section + slot override for the steward chat (#1223) (96d117b)
• feat(brain): tool-use hardening — pause-on-approval, tool-history replay, arg schemas, global port registry (#1222) (6ac481d)
• feat(brain-chat): [brain_chat] guardrails — kill switch, read-only mode, config-backed loop knobs (#1221) (3c55bfb)
• feat(agents): safe capture of existing hermes installs — --adopt, fatal claim abort, foreign-gateway preflight, ownership reconcile (#1220) (02e86b0)
• fix(test): stop static slot seeding from polluting every test's zero-slot baseline (#1219) (2995b72)
• feat(ui): slots-page polish batch — headings, NPU activity tint, logs drawer, image-gen header + hal0 update owui (#1216) (7e213d2)
• feat(ui): add a Documentation button to the topbar (#1213) (513c56c)
• feat(install): seed static slot TOMLs on hal0-api startup, not just fresh install (#1218) (7190919)
• feat(install): seed agent + brain slots; agent replaces chat as the LLM anchor (#1217) (cdffb22)
• feat(mcp): enforce per-persona tool policy on the Brain's admin surface (#1215) (37102c2)
• feat(ui): model updates in the notification bell + always-visible Check updates (#1190) (fb78a75)
• feat(doctor): FLM/migration/profile audits + --force delete for seeded slots (#1214) (160d71c)
• fix(slots): canon FLM-trio shadow slots to flm-stt/flm-embed + startup reconcile (#1210) (9efa6b1)
• chore: address #1182 review follow-ups — dev bench state wiring, docstring, README de-rot (#1184) (9db49af)
• feat(agents): ship hal0-brain persona on every install path; retire coder seed (#1204) (df53b65)
• fix(flm): land host pulls in the resolved store + robust 0-byte progress (#1211) (16a9ccc)
• test(e2e): follow container controls into the ComfyUI card header (#1212) (591ed7a)
• feat(mcp): expand hal0-admin to the full platform surface for the Brain agent (#1208) (423fbd6)
• feat(ui): polish image-gen header + size Local endpoints heading (#1209) (34c5b8d)
• feat(ui): unify slots-page headings and slot-status indicators (#1205) (5ca97c4)
• feat(ui): full-height slot logs drawer + surface raw logs from the Logs page dropdown (#1207) (998d62a)
• feat(dash): topbar tidy — move Kanban to Services + brain icon for Agent Chat (#1206) (fb56946)
• feat(dash): unify Needs Attention + bell into one notifications source (#1195) (f94e454)
• fix(dash): real update-banner actions + NPU grid colours by the primary FLM slot (#1194) (42ba168)
• feat(scripts): push-dev.sh — push-based inner-loop deploy to an editable box (#1193) (9134a51)
• fix(flm): FLM model pulls die instantly in prod — uvloop rejects user/group spawn kwargs (#1192) (7afc52f)
• Merge branch 'docs/sync-v0.9.5.x' (36dfb5b)
• docs: sync docs with v0.9.5–v0.9.5.2 + add docs agent team (3853429)
• release: v0.9.6.1 (#1191) (b4260d1)
• fix(setup): treat apply-selections 409 as a recoverable no-op (#1158) (#1187) (a0875c6)
• fix(ui): add no-undef lint guard for dash JSX (#1170) (#1188) (d3aef5c)
• fix(ui): consolidate --warn color token to single source (#1156) (#1186) (3dac085)
• fix(flm): probe the slot's assigned model in verify_inference, not models[0] (#1171) (#1185) (57babc4)
• test(npu): align stale NPU occupancy γ-specs with the reworked card (#1183) (879ee52)
• fix: install/setup/update audit follow-ups — manifest pins on prod, honest NPU setup, doc + harness catch-up (#1182) (9e27b91)
• feat(chat): retarget agent-chat slide-out to the hal0-brain profile (#1180) (fa082b7)
• feat(models): HF update check + per-row indicator ...

hal0 v0.9.7.3

A robustness release. Fresh installs now adapt to the host — provisioning their
own Python and Node, tolerating podman or docker, and surviving hardened
umasks and non-root operation — and the self-hosted Honcho memory stack stands
up cleanly alongside Hindsight. Plus the pi-coder and opencode bundled agents,
the unified per-agent memory model, and a large batch of install/setup/agent
fixes surfaced by end-to-end reinstall testing on a clean box.

Highlights

• Installs self-heal across environments — auto-provision Python 3.12 and Node 20 LTS when missing, tolerate podman or docker, and survive hardened/root umasks (#1291, #1289).
• hal0 doctor and bundled-agent installs work on packaged installs — FHS-aware resolution fixes the "packaged without scripts" and "could not locate preflight.sh" failures on non-editable installs (#1284, #1285).
• Crash-safe agent switching — a failed agent install --switch no longer bricks the running agent; it verifies the target first and rolls back on failure (#1285).
• Self-hosted Honcho as a per-agent memory provider — unified memory, swappable and migratable with Hindsight per agent, with a clean opt-in standup (#1243, #1294, #1295).
• pi-coder and opencode join Hermes as bundled, single-pick agents (#1254, #1271).

Added

• Node.js LTS auto-provisioning in the installer (the dashboard build and pi-coder/opencode all need npm) plus a curated qwen3-embedding-0-6b model for the memory pipeline (#1291, #1294).
• Unified per-agent memory — self-hosted Honcho v3 provider and a unified-bank model with server-side tagging, plus hal0 memory bank/ops/mm/recall and migrate unify CLI (#1243, #1244, #1257).
• pi-coder agent (provisioning, hal0 provider/memory plugins, live dashboard card) and the opencode bundled agent (#1254, #1271).
• hal0-brain as a first-class profile (#1258), upstream controls CLI and UI (#1279), and a bench queue dropdown with lane/tool-eval/tune options (#1255).

Fixed

• Installer/preflight robustness — Python floor raised to 3.12 with auto-install, hardened-umask permissions, a container-runtime smoke test that no longer false-fails, and Node/disk/graphroot preflight gaps (#1291, #1292).
• Honcho standup — unbound-var abort, pgvector embedding-dim reconcile, compose-provider and migration ordering, AppArmor-in-LXC, and full --purge teardown (#1293, #1294, #1295, #1287).
• docker/podman portability — runtime-appropriate slot units and a PATH-based runtime probe (#1289).
• Slots — errored-slot restart recovery and the WARMING watchdog, plus drift/image_status false positives (#1278, #1269).
• Memory — private-visibility enforcement on read, migrate-unify retag scoping, and Hermes plugin/bank identity consolidation (#1260, #1262, #1245).
• CLI — editable-install hal0 update refusal, footgun confirmation gates, and command consolidation (#1274).

Changed

• comfyui, doctor, and agent helpers resolve bundled scripts and assets FHS-aware so they work on packaged (non-editable) installs (#1286, #1284, #1285).
• Docs — ADR/Cognee→Hindsight sweep, hal0/chat→hal0/agent alias, fresh-box first-run journey, and a loud LAN-only bind warning (#1280, #1276, #1275).

hal0 v0.9.7.2

Hotfix for a fresh-install blocker on Python-3.14-only distros (Ubuntu 26.04,
the Strix Halo target): the Hermes venv was built on 3.14, where pip silently
resolves the broken hermes-agent 0.15.2 wheel, and the gateway crash-looped on
ModuleNotFoundError: No module named 'hermes_cli.dashboard_auth'.

Highlights

• Fresh installs now work on Ubuntu 26.04 / Python 3.14 — when the system has no Python 3.11-3.13, hal0 provisions a uv-managed 3.13 for the Hermes venv automatically; zero manual steps (#1250, #1252).

Fixed

• Hermes venv could be built on Python 3.14, where every supported hermes-agent wheel is filtered out (requires-python <3.14) and pip silently installs the broken 0.15.2 build. The interpreter resolver now accepts only 3.11-3.13, probes python3.13/python3.12/python3.11 newest-first (hosts with only 3.12/3.13 on PATH were previously ignored), and preflight fails early with an actionable message instead of a crash-loop three phases later (#1248, #1251).
• hermes-agent version floor raised to 0.16.0 — the 0.15.2 wheel imports hermes_cli.dashboard_auth but ships without it; it can no longer be selected on any interpreter (#1247, #1251).
• Broken venvs self-heal: a venv previously built on an unsupported interpreter is detected from its on-disk layout and rebuilt during hal0 agent bootstrap hermes --repair — no manual removal needed (#1251).

Added

• uv-managed Python fallback for the Hermes venv: system Python always wins; when nothing qualifies and uv is on PATH, the install phase runs uv python install 3.13 into /var/lib/hal0/python (kept out of root's home so the hal0 service user can execute it) and builds the venv on the managed interpreter (#1250, #1252).

hal0 v0.9.7.1

Hotfix for a fresh-install blocker — Hermes never auto-provisioned on a clean
curl | bash install — bundled with the dashboard, provider, and memory
improvements merged since 0.9.7.

Highlights

• Hermes now auto-provisions on a fresh install — no more manual --adopt/--repair; all bootstrap phases complete and the gateway comes up managed, idle until a bot token is added (#1239).
• MiniMax and DeepSeek are now one-click options in the upstream provider catalog (#1236).
• Keyboard-driven hal0 setup — the guided setup TUI gets real arrow-key navigation and a clearer model picker (#1237).
• Memory subsystem toggle — hal0 memory enable / disable / status replace the old install-time env flag (#1240).

Added

• MiniMax + DeepSeek upstream catalog entries (OpenAI-compatible, bearer auth) — selectable in Slots ▸ Endpoints ▸ Add upstream with prefilled URL and auth (#1236).
• Arrow-key navigation in the guided-setup TUI: ↑↓/jk move, space toggles apps/agents, enter selects; scaffold/skip are clean navigable rows; numbered entry still works over a pipe or in CI (#1237).
• hal0 memory enable / disable / status commands, plus the [memory].enabled config flag (#1240).
• Retry failed graph extractions — a Memory-tab button that re-runs Hindsight's failed extraction operations, plus mental-model delete (#1235).

Fixed

• Hermes fresh-install provisioning (#1239, closes #1238). Two chained bugs aborted provisioning on every clean install: the api-lifespan seed populated HERMES_HOME before the bootstrap could claim it ("unclaimed HERMES_HOME"), and the gateway installer started a unit that hal0 then flagged as its own "foreign" gateway. The lifespan now stamps .hal0-managed before seeding and pre-writes the gateway secrets drop-in before the unit starts.
• Model-filters spacing in the Endpoints upstream pane — the filter inputs no longer sit flush against the panel border (#1236).
• Graph-extraction consolidation reliability in the Memory tab (#1235).

Changed

• The memory subsystem is now gated by [memory].enabled in hal0.toml (default on) instead of the installer-written HAL0_MEMORY_ENABLED env var; toggle it with hal0 memory enable / disable (#1240).

Migrations

• If you had disabled memory with HAL0_MEMORY_ENABLED=0, that env var is now ignored (memory defaults on) — run hal0 memory disable to keep it off (#1240).

hal0 v0.9.7

The steward release. The dashboard's agent chat graduates from a
side-panel into a real control surface for the whole platform, external
LLM providers get a first-class management surface, and the FLM/NPU
stack settles onto canonical names with an automatic migration.

Highlights

• hal0-brain steward. The top-bar agent chat now drives every platform surface: it runs the full 74-tool hal0-admin MCP catalog under a per-persona tool policy, pauses turns on gated tools for inline approve/deny, replays tool history across turns, and renders reasoning + tool cards inline (#1208, #1215, #1221, #1222, #1223).
• Upstream model controls. A full management surface for external providers (OpenRouter, Anthropic, OpenAI, Google AI Studio, Ollama, custom) — reactive CRUD, per-upstream model filters, an enabled kill-switch, and CLI + MCP + dashboard parity (#1228).
• Graceful restarts keep your downloads. Model pulls no longer block a clean hal0-api shutdown, so restarting mid-download no longer trips the 90s SIGKILL that was killing in-flight pulls (#1225).
• FLM / NPU canonicalization. The NPU trio's shadow slots settle onto flm-stt / flm-embed, /v1/models alias routing is finalized, and a naming migration + hal0 doctor audits move existing installs onto the new scheme (#1210, #1229, #1231, #1214).
• Agent is the new anchor. Seeded agent and brain slots replace chat as the default LLM anchor, and are seeded on every startup — not just fresh installs (#1204, #1217, #1218, #1230).
• hal0 update verification actually works again. Release signing now dual-emits a Sigstore bundle (with an embedded Rekor timestamp) so cosign verification survives the short-lived Fulcio cert's expiry — curl … | bash installs and in-app updates were failing verification 100% of the time on v0.9.2/0.9.3/0.9.5. Fully back-compatible with already-deployed clients (#1159).

Added

• hal0-brain steward / agent chat as a control surface. The dashboard's top-bar agent chat becomes a first-class operator for the whole platform:
  • Full hal0-admin platform surface (#1208). The admin MCP catalog grew from 28 to 74 tools, replacing the slide-out's old hardcoded ~25-tool list, so the Brain can drive every surface end-to-end:
    • Models — model_inspect (read an HF repo before pulling), register / add-from-path, metadata edit (PUT), in-place HF re-pull (model_update), pull status/cancel, scan (+preview), catalogue, update-check, and store show/set/migrate. Pulls always land in the operator's configured [models].store (re-read per call; the tool descriptions state the contract).
    • Slots — load / unload / edit-config (PUT) / metrics / capacity / logs, on top of create / delete / restart / swap.
    • Stacks — create / update / export / snapshot, joining apply / import / delete.
    • Profiles — create / update, joining import / export / delete (author a profile straight from a model card).
    • Settings & platform — settings_get / schema / apply_plan / reload, upstream_list, and benchmark runs/status/queue reads plus gated enqueue/control.
    • Two new guards make catalog drift impossible to reintroduce: import-time catalog validation (classification ↔ REST-map ↔ annotations ↔ path-args must cohere) and a build_server registration-completeness check.
  • Per-persona tool policy (#1215). The persona TOML's tools_allowed + [persona.approval] tables — previously decorative on the sidebar path — become an enforced server-side overlay (admin.ToolPolicy, fnmatch globs over tool names): tools_allowed hides tools from the surface entirely; require_approval tightens an autonomous tool behind the approval queue; auto_approve / default_policy=auto-approve grants standing approval to gated tools; default_policy=never refuses gated calls outright. Precedence is hide > tighten > loosen > server verdict. A POLICY_NO_LOOSEN floor means model/slot/stack/profile_delete, bulk memory_delete, config_write, and provider_credential_write can never be loosened by a persona edit; denials are typed (mcp.tool_not_allowed / mcp.gated_tool_refused) with audit rows.
  • Tool-use hardening (#1222) — fixes for four failure modes seen live:
    • Runaway generation — every round now carries max_tokens (default 4096, payload-overridable); an uncapped completion against a slow local slot used to burn ~25k tokens and the 300s transport window, killing the turn before the first tool call.
    • Invisible approvals — gated calls no longer park silently on the queue; the loop emits an approval_required SSE frame and pauses (with keepalive pings) until the operator approves/denies, then streams a second tool_result so the same turn continues (timeout falls back to the pending result).
    • Guessed argument names — high-traffic tools ship real parameter schemas + explicit descriptions (no more model_inspect without hf_repo → 400, or model_pull with model_id='org/repo' → 405); path args containing / are rejected with an actionable hint.
    • Round budget — _MAX_ROUNDS raised 8 → 90 (with per-round caps it's now a runaway backstop, not a working limit multi-step sessions kept hitting).
    • UI — useBoardChat.send() now replays full tool history (calls + results) in the outgoing conversation; previously it rebuilt from user/assistant text only, so after one tool chain the model saw no tool calls in its own history and learned to skip tools and hallucinate results. Adds inline approval cards, a Stop button, an auto-approve toggle, and new-session.
    • A global port-claim registry — the single authority for which slot owns which port — lands alongside in the same PR.
  • [brain_chat] server-side guardrails (#1221), enforced independently of the persona TOML (a persona edit can loosen the persona, never these): enabled is a hard kill switch (the endpoint refuses every turn — no LLM call, no board request); read_only lets reads through but refuses every mutating/admin-write tool at the single _dispatch_tool chokepoint (unknown tools fail closed); max_rounds / completion_timeout_s move the loop budget + per-round transport timeout out of module constants into config, with the schema as the single source of truth.
  • Agents / Brain settings + slot override (#1223). A dashboard Settings → "Agents / Brain" section surfaces [brain_chat] (enabled + read-only toggles, max_rounds / completion_timeout_s inputs) plus a slot-override picker built from the live slots list — point the steward at any slot (e.g. hal0/npu to run it on the NPU chat slot). Model precedence is explicit: per-request model > [brain_chat] override > persona/default; the rows carry a live badge (brain_chat.* is apply-plan immediate).
• Upstream model controls — full management surface for external LLM providers (OpenRouter, Anthropic, OpenAI, Google AI Studio, Ollama, custom):
  • Reactive CRUD: POST/PATCH/DELETE /api/upstreams (create prefills from the provider catalog via catalog_id); upstreams.toml stays canonical — every write rewrites it atomically before touching the running registry.
  • hal0 upstream CLI group (list/show/create/update/delete/test/set-credentials); create --catalog openrouter --api-key wires a provider end-to-end in one command.
  • MCP admin tools upstream_create/upstream_update/upstream_delete (gated) + upstream_test.
  • Dashboard Upstream providers panel (Slots → Endpoints / Connections): add-from-catalog form, write-only key entry, test-connection with latency + model count, enable/advertise toggles, filter editor with live preview, delete-with-confirm.
  • Per-upstream model_filters (models allowlist + include/exclude globs, exclude wins) curating /v1/models and /api/models advertising — dispatch stays unfiltered so hidden models remain addressable by name (per the 2026-07-06 upstream-model-filters spec).
  • enabled kill-switch on every upstream: false removes it from dispatch routing and the model catalog while retaining config + credentials.
  • auth_key_present in upstream serializations — whether the declared env-var actually holds a key (drives the dashboard auth badge), distinct from auth_configured.
• Install / lifecycle:
  • Safe capture of an existing Hermes install — --adopt, a fatal claim abort, foreign-gateway preflight, and ownership reconcile so hal0 can take over an already-running gateway without clobbering it (#1220).
  • Static slot TOMLs plus the agent + brain slots are now seeded on every hal0-api startup, not only on fresh install (#1217, #1218, #1230).
• hal0 doctor grows FLM / migration / profile audits and a --force delete for seeded slots (#1214).
• Dashboard / UI — slots-page polish: unified headings + status indicators, an NPU activity tint, a full-height slot-logs drawer surfaced from the Logs page, an image-gen header pass, and a Documentation button in the topbar (#1205, #1207, #1209, #1213, #1216).
• scripts/push-dev.sh — a push-based inner-loop deploy onto an editable box (#1193).

Changed

• The default LLM anchor is the seeded agent slot; the coder seed is retired and hal0/chat resolves to hal0/agent, on every install path (#1204, #1217).
• The NPU trio's shadow slots are canonicalized to flm-stt / flm-embed, with a startup reconcile that migrates legacy names (#1210).
• FLM host pulls land in the operator's resolved model store (#1211).

Fixed

• Model pulls no longer block a graceful hal0-api restart — the shutdown path drains pulls instead of being SIGKILLed at the 90s deadline, which was killing in-flight downloads (#1225).
• FLM model pulls that died instantly in production — uvloop rejects the user/group spawn kwargs the pull worker passed (#1192).
• FLM host pulls now land ...

hal0 v0.9.6.1

Added

• hal0-brain agent profile: a third seeded persona (alongside hermes and coder) that stewards the platform from the dashboard's agent-chat slide-out — its own memory namespace (private:hal0-brain), a hal0-heavy system prompt (slot lifecycle, model setup, benchmarking), and the dedicated brain slot (hal0/brain) as the default model.
• Board/agent chat streams a {type:"thinking"} SSE frame: explicit reasoning_content and inline <think>…</think> blocks are split out of the reply and rendered as a folded "thinking" section instead of raw tags.

Changed

• The top-bar agent chat now embodies the hal0-brain profile: it runs on hal0/brain (falls back to the agent slot via the resolver chain), and an operator-edited hal0-brain persona TOML overrides its system prompt/model without a code change.
• Agent-chat suggestion chips are now platform-steward starters ("Help me create a new slot", "Download and set up a model", "Benchmark the model on a slot", "How's the hardware doing?").
• Agent-chat replies render markdown (fences, lists, headings, bold/italic/inline code, links); tool calls render as structured cards with args, live status, and a folded result — replacing the raw → tool({json}) text rows.

Fixed

• FLM NPU slots no longer wedge in warming forever: the warm→ready inference sentinel now probes the slot's assigned model instead of models[0] from FLM's full catalogue. Probing an arbitrary other model forced FLM to reload the wrong weights onto its single NPU context mid-gate and deadlocked the load (#1171).
• hal0 setup no longer aborts with a raw HTTPStatusError traceback when apply-selections returns 409 Conflict: the setup CLI now treats a 409 (install already applied / a concurrent apply in flight) as a recoverable no-op with a clean message, and still raises on genuine errors (#1158).
• Unified the dashboard warming-state color to a single canonical --warn: #f2792b token in dashboard.css, removing the divergent local redefinitions in engine-panes.css/overhaul.css so every warming indicator renders the same orange-yellow (#1156, #1155).
• Added an ESLint no-undef guard (enforced in CI) over the dash .jsx prototype so an undefined identifier like the one behind the Create Slot crash can't ship again (#1170).

hal0 v0.9.6

Highlights

• FLM NPU trio: the edit-slot drawer's Chat/ASR/Embed toggles now drive the running flm serve process, with a full-catalogue chat model picker that downloads on demand.
• FLM can run embed- or STT-primary (chat disabled) — a modality-aware readiness gate promotes the slot instead of wedging on a chat probe.
• Real FastFlowLM v0.9.44 toolbox image (ghcr.io/hal0ai/hal0-toolbox-flm:0.9.44), rebuilt from the actual v0.9.44 binary.
• NPU occupancy cards now glow purple when a slot is running (they used to read flat green regardless of state).
• hal0-bench is now in-tree: the hal0.bench engine, /api/benchmarks, and a Benchmarks dashboard page.

Added

• NPU trio drawer wiring: ASR/Embed on-off toggles + a Chat model picker that lists the full FLM catalogue and pulls a not-yet-downloaded model on select (auto-applies on completion).
• FLMProvider.verify_embed — a one-shot /v1/embeddings readiness sentinel used when a slot serves embeddings without chat.
• hal0-bench in-tree port: hal0.bench engine, /api/benchmarks routes, and the Benchmarks dashboard tab (roster / runs / evals / run-queue).

Changed

• FLM toolbox pinned to 0.9.44 across manifest.json, the flm seed profile, and the capabilities catalog (contains FastFlowLM binary v0.9.44).
• The warm→ready gate for FLM slots picks its sentinel by served modality: chat → /v1/chat/completions, chat-off+embed → /v1/embeddings, ASR-only → /v1/models liveness.
• /api/slots/flm/models returns the full FLM catalogue (installed + downloadable) with accurate installed flags, container-exec first with a host-probe fallback.

Fixed

• Chat toggle now actually gates the container: container_spec no longer passes the positional chat tag when [npu].chat=false (it was cosmetic on the container path).
• Editing an NPU modality no longer clobbers [model].default/context_size — the drawer sends the model as a nested [model] table so the backend merge preserves sibling keys.
• /api/npu/occupancy no longer 500s while the FLM slot is offline (the degraded single-tenant fallback's zip(..., strict=True) mismatched slots_out vs flm_slots).
• NPU cards read a clear purple "running" glow for up/resident slots and dim for offline; the coresident STT/embed sub-cards reflect the anchor's [npu] toggles instead of the legacy shadow-slot enabled flag.

hal0 v0.9.5.2

Fixed

• Models → Downloads tab: React invariant 310 ("Rendered more hooks than during the previous render"). The Downloads pane in the Models view called useStateM(false) inside a jobs.map((j) => { … }) callback. React counts hook calls per render, so the moment the downloads list changed length (a pull started, finished, failed, or was cleared) the hook count differed across renders and the entire <ModelsView> crashed behind the v3 error boundary — masking the rest of the dashboard. Fix: extract the per-row logic into a DownloadRow component that owns its own cancelling state, so the hook lives at row-component top level and its count is stable across renders. Pinned by ui/src/dash/__tests__/react-hooks-order.test.mjs, a new AST-based static check that walks every JSX/JS file under src/dash/ and flags any useXxx (incl. the useStateM/useEffectM/etc. aliases used in this dashboard) called inside an array-iteration callback (.map/.forEach/.filter/.reduce/.some/.every/.flatMap/.find/.findIndex).