Tags

v0.1.1

v0.1.1 — packaging & CI fixes

- Release pipeline on Node 24 actions (docker/* + ghaction-github-runtime v4)
  ahead of the 2026-06-16 GitHub Node 24 default.
- Appliance image on Node 24 LTS (clears mail-mcp / html-to-text EBADENGINE).
- Documented the Caddy on-demand internal-CA TLS posture for the LAN appliance.

Built on v0.1.0 (first tabula-rasa release). Multi-arch ghcr image + host
binaries published via the GHA-cached release workflow.

v0.1.0

v0.1.0 — first tabula-rasa release

First packaged release of the tabula-rasa rewrite. Phase 17 packaging &
distribution: fat multi-arch (amd64+arm64) appliance image, ordered compose
topology, Caddy TLS front + token gate, keyless boot, aura doctor, installer,
and socketless network backups.

Live UAT sign-off (17-UAT.md): 12 acceptance legs verified against the live
stack; AC7 (offline recipes) and AC11 (LAN reachability) fixed this cycle.
5 legs remain for operator real-hardware sign-off (clean-VM install,
ghcr publish, appliance reboot, arm64, native-Linux auto-restart).

pre-rewrite-2026-05-27

Aura state before tabula-rasa rewrite

34027 LOC src + ~30k LOC test.
Layers: telegram bot, wiki .md + git, FTS5 + Qdrant + memoryindex hybrid retrieval,
internal/wiki graph_index+ppr+godnodes, OCR/markitdown/whisper ingest pipeline,
12-action search tool + recall_* + wiki_*, web dashboard + auth,
swarm primitives, MCP loader, skill loader, tray.

Decision 2026-05-27: target-shaped rewrite, 4 core components:
 1. Agent loop
 2. KV cache (provider-aware: DeepSeek auto, Anthropic ephemeral)
 3. Sandbox (python+shell, container-isolated)
 4. Swarm with parallel-agents + talk

Neo4j via mcp-neo4j-cypher MCP stdio (no Go adapter).
See spike at D:/tmp/aura-neo4j-spike-2026-05-27/ + CLAUDE_AGENT_INGEST.md.

tools-pre-rewrite-2026-05-24

v0.3.2

v0.3.2 — fix whisper arm64 build crash; 4 images published

v0.3.1

v0.3.1 — 4-image GHCR release for fast first-install (~5-10 min cold vs 45-60 min)

v0.3.0

v0.3.0 — Phase-WIKI-FIX + Phase-TOOL + Step 1.LAT closure

Three back-to-back milestones (174 commits since v0.2.0) restructuring
Aura's substrate retrieval, agent loop latency, and tool surface.

- Hard MaxIterations cap + step X/N pacing hint (US-LAT-01)
- Single-shot prompt bias in agent overlays (US-LAT-02)
- `text_response` as terminal tool — answer-as-argument (US-LAT-03)
- Full EN restructure of agent overlays + MaxIterations cap to 20 (US-LAT-04)
- Unified `search` tool — collapse 4 lookup tools into 1 (US-LAT-05)
- Stream-time parallel tool dispatch (US-LAT-06)
- Provider prompt caching for static prefix (US-LAT-07)
- Inline wiki TOC in system prompt — nanobot MEMORY pattern (US-LAT-08)

Direct retrieval bench 2026-05-22 surfaced 10 substrate bugs. After fix:
FTS hit rate 0/20 → 20/20, score discrimination restored, p95 ≤75ms.

- FTS5 mirror auto-sync on wiki write/delete (US-WIKI-FIX-01)
- Embed cache key includes output_dim (US-WIKI-FIX-02)
- Rebuild skip-and-continue on per-page embed error (US-WIKI-FIX-03)
- Admin `/api/wiki/reindex` + boot dim-mismatch auto-detect (US-WIKI-FIX-04)
- Filter system-category pages from default search (US-WIKI-FIX-05)
- Dedupe wiki search results by slug (US-WIKI-FIX-06)
- settings.updated_at default to CURRENT_TIMESTAMP (US-WIKI-FIX-07)
- Embed n_ctx per-slot math + boot smoke (US-WIKI-FIX-08)

Killed the tool RAG entirely per user mandate "togli il rag dei tool è
una cagata; alla fine basta il search tool come hai te". Removed ~4000
LOC of dead/duplicated/over-engineered tool infrastructure.

- KILL tool RAG entirely — all tools always-on in manifest (US-TOOL-01)
- Delete 18 unregistered orphan Tool types — ~1500 LOC dead (US-TOOL-02)
- Wire text_response + drop search_memory zombie (US-TOOL-03)
- Description audit — EN-only + 200char cap + structural markers (US-TOOL-04)
- Collapse schedule_task/list/cancel → `task(mode=)` (US-TOOL-05)
- Collapse create_xlsx/docx/pdf → `create_document(format=)` (US-TOOL-06)
- Align defaults with unified tool surfaces (US-TOOL-07)
- MCP description override sidecar — `tools-overrides.json` (US-TOOL-08)
- MCP supervisor — list_changed + circuit breaker + reconnect (US-TOOL-09)
- Dim-mismatch auto-detect + rebuild for compact collection (US-TOOL-10)

- ~-4500 LOC net (cleanup) vs ~+1700 LOC new substrate
- 22 → ~15 tools in catalog after kitchen-sink collapses
- 0 unused symbols, 6 TODO total in production code
- `aura_tool_search_v2` Qdrant collection deleted (no longer needed)
- All Anthropic / OpenAI-compat LLM endpoints gain `prompt_cache_key` support

- `EMBEDDING_OUTPUT_DIM` setting now persisted in SQLite; restart on change
  triggers automatic Qdrant collection rebuild
- `search_memory` tool removed (use unified `search` from US-LAT-05)
- `tool_search` tool removed (manifest is always full)
- `aura_tool_search_v2` Qdrant collection no longer created at boot
- MCP supervisor adds per-server circuit breaker — flapping MCP servers
  will be auto-isolated after 3 consecutive failures

v0.2.0

Aura v0.2.0 — Production-hardened, audit-clean, fully documented

Major milestone after a 2026-05-17 marathon (32+ commits to master)
covering four sequential phases. v0.0.1 was a placeholder; this is the
first meaningful release.

- **Phase-T** — Production rollout: compose.prod.yaml overlay, git SHA
  baked from link-time, daily backup-verify cron, dashboards over
  authz_decisions + tool_attempts, docs/RUNBOOK.md
- **Phase-Z** — Bounded PRD wrap-up: content_hash drift detection,
  eager freshness invalidation, degraded_read annotation, SQLite
  concurrency hardening, source conversion fixtures (PDF/DOCX/XLSX/PPTX/MD)
- **Phase-CLEANUP1+2** — Code-health sweep: god-files split (app.go
  1151→537, hub.go 760→456, identity/store.go 875→233), dead code purge,
  duplication collapsed (stringx, llm, httputil helpers extracted)
- **4 audit rounds** — security / concurrency-leaks / error-logging /
  code-health, all clean. golangci-lint integrated (80→2 issues, both
  third-party in web/node_modules)
- **README.md** consolidated with 4 mermaid diagrams: current state,
  roadmap, multimodal target, plugin-layout target

Five channels (Telegram + Web + Cron + Swarm + Silent) → unified Chat Hub
→ single Agent Loop with capability-gated tool registry, typed memory
(Storage vs Memory per docs/MEMORY-VS-STORAGE.md), subagent dispatch with
write_proposal review queue, audit trail for every authz decision and
tool attempt.

- Phase-FIX (4 stories) — kill the bad runtime UX fallback strings
- Phase-MM (9 stories, 3 waves) — multimodal core (audio + image, IN + OUT)
- Phase-U (4-6 stories) — plugin layout (option 2: clean shape, no
  marketplace; single-author scope stays preserved)
- Phase 8 — multi-agent substrate (planner + critic + DAG) anchored to a
  concrete workload (market research candidate)

See prd.md §7.4 and docs/phase-mm-synthesis-2026-05-17.md for the full plan.

v0.0.1

v0.0.1: versioning reset

Major changes since last build:
- Tool RAG hybrid backend is now the default (was fts) — fixes tool-flood on Italian queries
- Pyodide sandbox runtime removed; replaced by process runner with Python deps baked into image
- Settings UI cut from ~62 rows to the 12 a user actually configures
- Dockerfile installs requests, pandas, openpyxl, python-calamine, etc. so execute_code is useful out of the box
- Dead config fields removed (SpeculativeSearchTimeoutMS, MistralOCRIncludeImages, MaxToolIterations, OCREnabled, SandboxRuntimeMode/URL/Dir)