ARIA 0.1.0-alpha1007

ARIA 0.1.0-alpha1007

This candidate combines reliable background Agent Jobs, resilient MCP integration, a privacy-conscious Activity Journal, bilingual administration and owner-focused statistics.

Alpha1007 adds the final owner-review polish: concrete health checks, readable cost axes, correct container resource sizes, authoritative routing-index cleanup, full-width assistant tables, terminal Job controls and a compact chat actions menu.

See docs/release/public-release-summary-alpha1007.md and CHANGELOG.md for the complete scope and limitations.

Upgrade

Back up ARIA first. Keep the current data/config mounts and Qdrant volume, and replace only the image with fischermanch/aria:0.1.0-alpha.1007. Review MCP trust settings after upgrading.

ARIA 0.1.0-alpha983

ARIA 0.1.0-alpha983

Alpha983 is a focused public hotfix for the document-inventory regression in Alpha982.

Fixed

• Native document inventory now returns one structured row per imported document from the authoritative MemorySkill inventory metadata, including document name and collection.
• Inventory explicitly requests up to 200 documents instead of silently stopping at the MemorySkill default of 12.
• Document search and answer results retain their source document names and collections when source metadata is available.
• Tool guidance now distinguishes complete inventory questions from content search: inventory says which documents exist, while search finds excerpts inside documents and does not prove completeness.

No document data was lost in Alpha982. Imported files and their chunks remained stored; only the native Tool's inventory presentation discarded the correct listing and could then fall back to an incomplete content search.

Upgrade

Back up the ARIA config/data mounts or volumes and the Qdrant volume, then replace the ARIA image with fischermanch/aria:0.1.0-alpha.983. No schema migration or data rewrite is required for this hotfix. Existing Alpha604 and Alpha982 installations are covered by the isolated upgrade gate.

The immutable image tag is fischermanch/aria:0.1.0-alpha.983; moving alpha and latest tags are updated only after explicit release approval.

ARIA 0.1.0-alpha982

ARIA 0.1.0-alpha.982

This is the first public ARIA release candidate since 0.1.0-alpha604. The browser interface remains familiar—only the GUI still looks much like before—but the application underneath has been rebuilt around modular ownership, a native Tool agent and explicit safety boundaries.

About the version numbers: ARIA's version numbers deliberately count development iterations. Alpha604 to Alpha982 are 378 iterations of building ARIA together with AI — the number is kept transparent on purpose and will keep counting, past 1000 if need be.

Highlights

• Modular core: chat, connections, recipes, memory, documents, release handling and UI composition now have explicit module owners instead of sharing a monolithic runtime truth.
• Self-learning memory: repeated personal preferences can become an explicit, user-approved memory suggestion; casual statements are not silently stored.
• Self-learning recipes: recurring actionable sequences can become reviewable inactive recipes and supported cross-connection copies.
• MCP agent and background jobs: enabled MCP servers join ARIA's native Tool catalog; long jobs support pause, resume, correction, cancellation, budget extension, confirmations and Tool-image vision.
• Memory cleanup boundary: core memories remain, while obsolete learning collections can be removed explicitly from the Maintenance page.

Added

• HTTP/SSE MCP servers with admin configuration, masked headers, discovery status, reconnect, per-server call timeout, trust control and confirmation previews.
• Persistent user-scoped Agent Jobs with progress, token/cost details and idempotent controls.
• Native preference and recipe recurrence, rolling Tool-image vision and conservative output honesty guards.
• An isolated strict-fake browser test harness for release validation.

Changed

• Public installs now run aria, aria-updater and qdrant; provider-native web capabilities replace the old web-search sidecar.
• Configuration, connection targets, personal claims, recipes and runtime actions pass through explicit module contracts and authority checks.
• Model profiles may omit temperature; long agent tasks can use larger output/token/time budgets without tying the browser request to the full run.

Removed

• SearXNG and Valkey from the supported fresh-install stack.
• Dead legacy-learning workers, review UI and data writers.
• Obsolete monolithic compatibility modules whose live owners moved into modules.

Fixed

• MCP discovery, sessions, confirmation turns and failures no longer block or crash unrelated chat turns.
• Job pause/completion events are one-shot, confirmation state survives resume, and long vision jobs keep a rolling four-image live window.
• Memory capture and recurrence are more reliable across phrasing while retaining explicit user authority.

Security

• Untrusted mutating MCP/native Tools use the confirmation kernel with frozen arguments and a one-shot ledger claim.
• Job and maintenance controls are authenticated, CSRF-protected and owner-scoped.
• Output guards refuse fabricated resource contents and action-completion claims without same-turn Tool evidence.
• MCP headers, local config, secrets, runtime data and developer state are excluded or redacted by the release hygiene gate.

Upgrade Notes from Alpha604

1. Back up the ARIA config/data mounts or volumes and Qdrant volume before replacing the image.
2. Use the same persistent mounts with Alpha982. The automated upgrade test verifies that users, auth, connections, chat history, personal facts/preferences, memories and recipes remain readable; legacy learning collections are intentionally retained until an administrator opts in to cleanup.
3. Log in as each user and open Memories → Maintenance → Clean up old learning collections. Only that user's obsolete learning candidate/evaluation/event/hint/reflection collections are deleted; memories, facts, preferences, documents and recipe experience are untouched.
4. Old SearXNG and Valkey containers are not removed by ARIA. After backup and verification, remove them manually as described in docs/release/alpha981-upgrade-note.md.
5. Models that reject temperature may leave it empty. For long Agent/Blender jobs, try 16000 maximum output tokens and 300 seconds model timeout, then tune for model and cost; set MCP call timeout per server.
6. LiteLLM proxy compatibility no longer sends tool_choice=none. See docs/setup/mcp-and-blender.md for MCP and Blender setup.

Known Limitations

• Vision on Tool-result images currently requires a Claude-family model.
• AI 3D generation through a Blender MCP add-on requires the user's own paid generator keys.
• The Blender MCP bridge runs on the user's machine and must be reachable from the ARIA container.
• Budgets, latency and cost depend on the chosen provider/model.
• The strict E2E harness is a development tool and not a production service.

Images and tags

• Release title: ARIA 0.1.0-alpha982
• Git tag: v0.1.0-alpha.982
• Immutable Docker tag: fischermanch/aria:0.1.0-alpha.982
• Moving Docker tags after approval: fischermanch/aria:alpha and fischermanch/aria:latest

What to test after upgrade

• both existing user logins and one normal chat turn
• Memories browser, one recall and legacy-learning cleanup
• one stored recipe preview/execution confirmation
• one configured MCP server, including a mutating Tool confirmation
• /stats, /help, /product-info and /updates

ARIA 0.1.0-alpha604

ARIA 0.1.0-alpha604

Public alpha release for the post-alpha511 agentic stabilization line.

Highlights

• Consolidates the LLM-first authority-chain work from the internal alpha512-alpha604 line.
• Adds source-bound Evidence/Answerability contracts and final-answer guardrails for safer agentic responses.
• Hardens RuntimeOutcome follow-ups so package-update, disk-capacity, and related follow-up questions stay bound to observed runtime records.
• Hardens multi-target SSH planning and execution contracts, including full-kind target expansion and stale missing-field pruning for finalized actions.
• Improves the Memory Browser structure/read model so notes and chunks are shown from real Qdrant rows instead of a limited semantic sample.
• Improves WebSearch source authority handling, weak-source caution, and fast-answer routing.
• Adds chat/debug export support and help/UI refinements.

Validation

• Server/Agentic validation confirmed disk-capacity, diagnostic, and package-update prompts over all expected SSH targets.
• SSH inventory stayed config-bound, scoped follow-ups stayed narrowed, and unknown explicit SSH refs failed closed.
• Medication package-insert inventory returned the expected document set.
• Notes retrieval found the expected sample note.
• Release gates passed: focused Pipeline, AriaTurn/OperatorTrace/Authority/Learning, Memory/LLM-input/SSH-scope/Inventory, Release Hygiene, Compileall, Docker CLI, Docker health, and image-layer privacy checks.

Docker

• docker pull fischermanch/aria:0.1.0-alpha.604
• docker pull fischermanch/aria:alpha
• Docker Hub digest for both tags: sha256:54a87a94226ff3073ea7f5f64835b328096e9e49cea56931f6634832fbf04add

Notes

• This is an alpha/prerelease build.
• Exact long-note recall can still summarize selectively; retrieval is working, but final answer condensation may omit lower-priority bullets.

ARIA 0.1.0-alpha511

ARIA 0.1.0-alpha511

This is the public release note source for the first public release after 0.1.0-alpha437. It documents the user-facing rollup that has been tested internally up to 0.1.0-alpha511.

Public tag: v0.1.0-alpha.511.

Summary

0.1.0-alpha511 is a larger public alpha release focused on making ARIA feel more like a controlled personal operator: better chat flow, clearer confirmations, richer Memory/Notes workspaces, stronger source discipline, and more observable agentic runtime decisions.

The release keeps ARIA's core safety boundary: LLMs help with meaning, planning, summarization, and review, while policy, guardrails, runtime execution, validation, and confirmations remain controlled system layers.

Highlights

• Chat Prompt Queue: keep writing while ARIA is busy, then edit, remove, or reorder waiting prompts before sequential execution.
• Pending Confirmations stay visible in the queue with run, plan-again, and discard actions.
• Graphical Memory browser: inspect Qdrant/Memory structure, documents, entries, chunks, semantic proximity, fullscreen, and touch/iOS interaction.
• Notes workspace: move notes between folders from cards or editor, use a dense list view, and bulk-move selected notes.
• Auto-memory and agentic learning UI: clearer user-facing controls and separation between durable memory, review artifacts, and technical maintenance.
• Generated navigation: the header, account menu, Settings, Admin, Memory, Recipes, and Connections use a shared navigation registry with calmer hubs and mobile wrapping.
• Agentic Operator Trace: Details can show understanding, context, draft/policy, runtime, result, summary, and review-only learning phases.
• Document inventory and corpus checks are more source-bound and less likely to answer from unrelated snippets.
• Web freshness answers prefer official/vendor/compare/release sources for current product and version questions.
• Multi-target SSH/runtime paths keep full-fleet scope for prompts such as all Linux servers when no true subgroup is named.

Added

• Chat Prompt Queue with per-tab waiting prompts and sequential execution.
• Visible queue item for confirmation-required actions.
• Graphical Memory browser for structure drilldown and semantic proximity.
• Dedicated Auto-memory & learning page.
• Dedicated manual Memory creation page.
• Notes list view with multi-select bulk move.
• Agentic Operator Trace detail lines for runtime observability.
• Agentic Runtime Result Contract for multi-target runtime summaries.
• Help pages for Chat & Queue, Notes, and Agentic Operator.
• Help page for Navigation & Menus, covering Extended view, the account dropdown, header navigation, Settings/Admin groups, Memory, Recipes, and Connections.

Changed

• Recipes navigation now separates the Recipes hub, saved recipes, new/templates, learned recipes, and admin maintenance more clearly.
• Settings/Admin/Memory/Recipes/Connections navigation is generated from a shared registry instead of scattered template links.
• Admin mode is presented as Extended view and moved closer to the account/settings context.
• Connections and Settings hubs are quieter and avoid operational status-card clutter.
• Web Search source selection is stricter for current/latest product questions when official sources are available.
• Document inventory questions load document metadata evidence instead of relying only on top semantic chunks.
• LLM-first routing and bounded planning remain the preferred path for meaning, while deterministic layers handle safety and runtime contracts.

Fixed

• Current public product questions can override accidental local Docs-only matches and route to Web Search when the user did not ask for local documents.
• Current/latest product-line questions can add official manufacturer/store/compare sources even when only one product family is named.
• Rumor, deal, news, leak, and future-model sources are demoted for current availability claims.
• Clear built-in actions no longer become ambiguous against stored recipes when no LLM client is configured.
• Reconstructed SSH follow-ups keep the current target context instead of being preempted by generic stored recipe arbitration.
• All-scope SSH target prompts keep the full fleet when only OS/runtime words such as Linux are present.
• HTTP API status/health drafts can stay on their local health-path contract without unnecessary recipe arbitration.
• Server disk-capacity prompts can use the SSH runtime fast path instead of falling into unrelated document search.
• Expired or invalid action confirmations now give a recovery hint instead of a generic failure.
• Memory upload CSRF tests and related route assertions were aligned with the current session contract.

Security / Safety

• No guardrail bypass was added for queued prompts or pending confirmations.
• Queue execution remains sequential; ARIA does not run multiple tool/runtime requests in parallel from the prompt queue.
• Confirmation-required actions still require explicit user action.
• Learned recipes and Experience Memory remain review/context material until deliberately promoted.
• Normal managed updates should recreate only the aria service and keep Qdrant, SearXNG, Valkey, and persistent volumes untouched.

Known Limitations

• ARIA remains alpha software intended for controlled LAN/VPN/homelab environments.
• Prompt Queue is per browser tab and not persisted across reloads.
• The current queue edit flow uses the browser prompt dialog.
• Full multi-user ownership/RBAC/sharing for Recipes, Connections, and Memories is not complete.
• Pricing remains a local estimate based on usage logs and known model prices.

Upgrade Notes

• Managed installs should use /updates or the managed update script.
• Fixed-tag installs can use the versioned Docker image once the release is actually published.
• A hard browser refresh is recommended because the release includes UI, CSS, and navigation changes.
• Do not remove Qdrant, SearXNG, Valkey, or persistent volumes during upgrade.

What To Test After Upgrade

• /health returns ok.
• /help opens and includes Chat & Queue, Memory, Notes, Recipes, Agentic Operator, Pricing, and Security.
• Submit a second chat prompt while ARIA is still answering; reorder, edit, and remove queued prompts.
• Plan a confirmation-required action and verify the visible queue item can run, plan again, or discard.
• Open /memories, drill into a document collection, and test fullscreen/touch navigation if available.
• Create, move, and bulk-move Notes under /notes.
• Ask a local document inventory question and verify source-bound document names.
• Ask a current product question and verify official sources are preferred.
• Run one read-only SSH/API/RSS workflow and inspect Details for operator/runtime trace lines.
• Check mobile Safari/iPhone width for chat, queue, help, and Memory browser usability.

Docker Images

• fischermanch/aria:0.1.0-alpha.511
• fischermanch/aria:alpha

v0.1.0-alpha.437

ARIA 0.1.0-alpha437

This public alpha release promotes the internally tested document-corpus and runtime-performance work to the public Docker/GitHub line.

Highlights

• Uploaded-document questions are safer for corpus-wide checks. When a user asks whether a term or substance appears in any uploaded document, ARIA keeps the corpus scope and scans the selected document store before answering.
• Document inventory questions now load the document store inventory instead of relying on a few semantic chunk hits.
• Runtime paths are leaner for common operational tasks, with less avoidable Meta-Catalog and follow-up overhead while keeping confirmations and guardrails intact.
• Web Search handling is more robust around SearXNG timeouts and official supplemental queries.

Fixed

• Preserved the original user prompt as a scope signal when the Meta-Catalog narrows a document question to a single promising document.
• Passed explicit document target collections through the document surface loader, recipe runtime, and Memory skill for exhaustive scans.
• Prevented unrelated document or recipe-experience records from entering source-bound document corpus answers.
• Skipped stale package-update follow-up context when the next prompt is clearly a document question.
• Kept primary Web Search results when best-effort official supplemental searches time out.

Changed

• Runtime outcome follow-up handling was moved into a focused resolver module as part of the ongoing pipeline cleanup.
• Chat feedback learning is now queued through the Learning Worker instead of being awaited in the web request.
• Clear multi-target runtime tasks can use a bounded capability-draft fast path before the full Meta-Catalog step, while execution still goes through the existing safety and confirmation path.

Architecture Notes

• This release does not add a deterministic router for user meaning.
• The document fixes are evidence-policy and context-contract work: the LLM-selected document surface still chooses the context, and ARIA then proves coverage before a source-bound answer can claim presence or absence.

Upgrade Notes

• Normal managed installs should use /updates or ./aria-stack.sh update.
• Fixed-tag installs can use fischermanch/aria:0.1.0-alpha.437.
• Normal updates should recreate only the aria service and keep Qdrant, SearXNG, Valkey, and persistent volumes untouched.

Docker Images

• fischermanch/aria:0.1.0-alpha.437
• fischermanch/aria:alpha

ARIA 0.1.0-alpha433

ARIA 0.1.0-alpha433

This public alpha hotfix finishes the uploaded-document corpus scan answer path.

Fixed

• Exhaustive document corpus scans now keep structured per-term evidence, including per-term hit counts and matched/unmatched terms.
• If the primary searched term has zero hits but supporting context terms have hits, ARIA now answers from the exhaustive scan coverage instead of treating the context-term hits as proof of a positive match.
• If the answer composer returns an empty or invalid response for this source-bound scan case, ARIA now falls back to a deterministic evidence summary from the completed corpus scan.

Architecture Notes

• This does not add a deterministic router. The Meta-Catalog still selects the document surface and corpus scope; the deterministic part is only the presentation of already-loaded scan evidence.
• The fix is designed for questions such as "is this term present in any uploaded document?" where semantic top-k retrieval is not sufficient proof.

Upgrade Notes

• Normal managed installs should use /updates or ./aria-stack.sh update.
• Fixed-tag installs can use fischermanch/aria:0.1.0-alpha.433.
• Normal updates should recreate only the aria service and keep Qdrant, SearXNG, Valkey, and persistent volumes untouched.

Docker Images

• fischermanch/aria:0.1.0-alpha.433
• fischermanch/aria:alpha

ARIA 0.1.0-alpha432

ARIA 0.1.0-alpha432

This public alpha hotfix tightens corpus-wide uploaded-document questions.

Fixed

• If the Meta-Catalog selects the structured document-corpus scope (local|docs|documents), ARIA now requires a document corpus scan even when the plan labels the context depth as shallow.
• This prevents "is this term in any uploaded document?" questions from being answered from one semantically retrieved document excerpt while claiming coverage over the whole uploaded-document set.

Architecture Notes

• This is still evidence policy, not a deterministic meaning router. The Meta-Catalog/LLM chooses the docs surface and the document-corpus scope; Memory then proves coverage for that selected scope.
• The previous named-collection fix from 0.1.0-alpha431 remains in place, so custom uploaded-document collections selected by document guides/metadata are included in the corpus scan.

Upgrade Notes

• Normal managed installs should use /updates or ./aria-stack.sh update.
• Fixed-tag installs can use fischermanch/aria:0.1.0-alpha.432.
• Normal updates should recreate only the aria service and keep Qdrant, SearXNG, Valkey, and persistent volumes untouched.

Docker Images

• fischermanch/aria:0.1.0-alpha.432
• fischermanch/aria:alpha

ARIA 0.1.0-alpha431

ARIA 0.1.0-alpha431

This public alpha hotfix completes the deep uploaded-document corpus-scan path introduced in 0.1.0-alpha430.

Fixed

• Deep document corpus scans now work for explicitly selected named upload collections, not only plain per-user aria_docs_<user> collections.
• When the Meta-Catalog/document guides select a custom document collection, the corpus scan treats that selected collection as the source-bound scope and scans its document chunks before allowing a presence/absence answer.
• If a pre-guide corpus scan cannot find an accessible corpus, ARIA now retries the scan after document-guide selection instead of falling back to semantic top-k snippets only.

Architecture Notes

• This remains an evidence-policy fix, not a deterministic router. The Meta-Catalog still chooses the docs/deep context contract; document guides or metadata select the document collection; the Memory skill then proves source-bound corpus coverage inside that selected scope.
• The context ledger now exposes whether a deep document turn requested document_corpus_scan=true.

Upgrade Notes

• Normal managed installs should use /updates or ./aria-stack.sh update.
• Fixed-tag installs can use fischermanch/aria:0.1.0-alpha.431.
• Normal updates should recreate only the aria service and keep Qdrant, SearXNG, Valkey, and persistent volumes untouched.

Docker Images

• fischermanch/aria:0.1.0-alpha.431
• fischermanch/aria:alpha

ARIA 0.1.0-alpha429

ARIA 0.1.0-alpha429

This public alpha hotfix tightens source-bound document search behavior for uploaded document collections.

Fixed

• Negative document-search answers now require explicit document-corpus coverage.
• If a docs-only recall only retrieves semantic chunks that do not contain the requested evidence terms, ARIA performs a literal scan over the selected uploaded document chunks before answering.
• The scan reports how many documents and chunks were checked, per-term match counts, and the scanned document sources.
• This prevents "not found" answers from being based only on a semantic top-k miss.

Notes

• This is not a deterministic router. It is a source-bound evidence policy inside document recall.
• Normal semantic document recall remains in place for content questions.
• No data migration is required.
• Existing Qdrant, Valkey, SearXNG, and user-data volumes can remain in place.

Verification

• Focused MemorySkill document-corpus scan tests for both missing and present literals
• Docs/Memory/Meta-Catalog focused tests
• Release/package/i18n checks
• Python compile checks and git diff --check