Eva - Releases

A major release that transforms EVA from a RAG chat tool into a full multimodal AI assistant with vision, speech, audio, image generation, a plugin system, and deep Nextcloud integration.

Multimodal & Vision

• Image analysis in Assistant tasks. A new bounded core:analyze-images provider reads Nextcloud image files and sends them to the configured Ollama vision or OpenAI-compatible model. Native Ollama image messages are preserved and translated to OpenAI-compatible format for remote providers.
• Image generation. A bounded core:text2image provider for configured OpenAI-compatible image APIs. Generated images are stored in the user's EVA folder with returned Nextcloud file IDs.
• Sticker tool. A confirmation-gated create_sticker chat tool generates PNG stickers from text prompts and saves them to the user's EVA folder.
• OCR. A local OCR provider for image/PDF TaskProcessing inputs, reusing EVA's Tesseract safeguards and per-user file permissions.
• Multimodal chat. core:text2text:multimodal-chatwithtools supports bounded image attachments (up to 4 files, 12 MB total) with the central EVA tool policy.
• Multimodal ContextAgent. core:contextagent:multimodal-interaction with bounded, permission-checked image attachments while preserving EVA's confirmation workflow.

Audio & Speech

• Audio transcription and subtitles. core:audio2text and core:audio2text:subtitles providers using configurable OpenAI-compatible audio endpoints, with a 25 MB input limit and per-user file permissions.
• Text-to-Speech. Configurable OpenAI-compatible TTS output stored as a Nextcloud audio file.
• Audio-to-audio translation. Transcribe, translate, and synthesize with bounded provider calls.
• Audio chat. core:audio2audio:chat transcribes a voice message, answers with the configured chat model, and saves the spoken response.
• ContextAgent audio. core:contextagent:audio-interaction routes bounded voice input through the confirmation-aware EVA agent.

Provider Configuration

• Modality-specific profile fields (image_model, audio_model, tts_model, tts_voice) are now surfaced in Settings so image/audio providers can be configured without guessing JSON keys.
• Guard native TaskProcessing provider registration by the availability of the corresponding Nextcloud core task type, preventing Nextcloud 34 dashboard failures on newer optional task types.

Terminal & Tool System

• Native confirmation forms for terminal commands and command sequences with bounded interactive stdin input and per-command prompt support.
• Terminal command sequences can run up to five bounded commands in order, with redacted per-command results and timings in Agent runs.
• Terminal traces expand automatically in live chat so confirmed commands are easier to audit, with exposed execution timeouts.

Connector & Plugin System

• Generic connector discovery probes common OpenAPI and Swagger YAML paths in addition to JSON descriptions, without requiring PHP ext-yaml.
• Schema-free GraphQL connectors are discovered through safe GET probes and learned as confirmation-gated POST operations.
• OpenAPI YAML discovery works without the optional PHP ext-yaml extension through the bundled Symfony YAML parser.
• Plugin safety hardening. Mutating or destructive plugins are always confirmation-gated. Duplicate plugin tool names can no longer silently replace earlier implementations. All five execution surfaces are now supported, including RAG and task-processing proposal runs.
• Plugin metadata is exposed in Settings: risk level, execution surfaces, and confirmation requirements.
• Connector presets for Immich (X-API-Key) and Vaultwarden (Bearer) with security-focused documentation and recipes.
• Connector route learning with automatic refresh of stale discovery, validated path variables, and authenticated endpoint probing.
• Custom OpenAPI schema URLs for connectors with safe path parameter resolution.

Talk Integration

• Read and write Talk conversations via list_talk_rooms, read_talk_chat, and send_talk_message tools. Messages are posted as the signed-in user with no bot label.
• Focused Talk bots for research, calendar, mail, and files, preventing duplicate responses.
• Email summarization through the Nextcloud Mail app.

Search & Indexing

• Nextcloud Unified Search integration. EVA's indexed file knowledge is searchable from the Nextcloud search bar with live VFS permission checks and direct links back to Files.
• Direct file search now searches unindexed Office documents and reports bounded diagnostics (nodes visited, documents extracted).
• Index lock recovery prevents stale worker locks from blocking indexing.
• Partial results on folder errors: accessible folders are returned even when a nested remote folder cannot be listed.

Dashboard

• EVA starters on the Nextcloud Dashboard with focused entry points for common tasks.

Performance & Security

• Memory bounds on index traversal, connector batches, and background queue scans.
• CPU spike prevention for local embeddings on cloud chat requests.
• Session lock release during long chat requests.
• Bearer token and credential handling hardened: secrets are encrypted and never returned in settings or exports.
• CSP-compatible native Nextcloud controls throughout Settings.
• Third-party plugin output centrally bounded with credential redaction.

Bug Fixes

• Talk chat histories are now properly indexed (previously threw silently on every room).
• Long Talk messages no longer break room indexing.
• Google News redirect links are filtered to prevent consent interstitials.
• Browser renderer now correctly reports availability and searches the right directory.
• Connector credentials are preserved on blank edits and properly re-discovered on reconfiguration.
• Shared calendars are included through canonical CalDAV ACLs.
• Stale chat locks are recovered reliably.
• Settings controls remain visible after rebuild.

Documentation

• Complete external service plugin recipes for Immich and Vaultwarden with ACL, secret-handling, and test guidance.
• Terminal prompt guidance and connector quick-setup labels in both modern JSON and legacy JavaScript translation catalogs.
• App store screenshots and polished listing.

Added

• Multiple AI provider profiles: configure up to 20 OpenAI-compatible endpoints with independent IDs, names and models, then select them directly in EVA Settings. API keys remain encrypted per user and are never returned in settings or exports.
• Provider profile validation, documentation and localized settings labels.

Fixed

• Background agent runs now reschedule themselves while work remains, and the status poll wakes a missed worker tick so queued messages continue promptly.

A major release that transforms EVA from a RAG chat tool into a full multimodal AI assistant with vision, speech, audio, image generation, a plugin system, and deep Nextcloud integration.

Multimodal & Vision

• Image analysis in Assistant tasks. A new bounded core:analyze-images provider reads Nextcloud image files and sends them to the configured Ollama vision or OpenAI-compatible model. Native Ollama image messages are preserved and translated to OpenAI-compatible format for remote providers.
• Image generation. A bounded core:text2image provider for configured OpenAI-compatible image APIs. Generated images are stored in the user's EVA folder with returned Nextcloud file IDs.
• Sticker tool. A confirmation-gated create_sticker chat tool generates PNG stickers from text prompts and saves them to the user's EVA folder.
• OCR. A local OCR provider for image/PDF TaskProcessing inputs, reusing EVA's Tesseract safeguards and per-user file permissions.
• Multimodal chat. core:text2text:multimodal-chatwithtools supports bounded image attachments (up to 4 files, 12 MB total) with the central EVA tool policy.
• Multimodal ContextAgent. core:contextagent:multimodal-interaction with bounded, permission-checked image attachments while preserving EVA's confirmation workflow.

Audio & Speech

• Audio transcription and subtitles. core:audio2text and core:audio2text:subtitles providers using configurable OpenAI-compatible audio endpoints, with a 25 MB input limit and per-user file permissions.
• Text-to-Speech. Configurable OpenAI-compatible TTS output stored as a Nextcloud audio file.
• Audio-to-audio translation. Transcribe, translate, and synthesize with bounded provider calls.
• Audio chat. core:audio2audio:chat transcribes a voice message, answers with the configured chat model, and saves the spoken response.
• ContextAgent audio. core:contextagent:audio-interaction routes bounded voice input through the confirmation-aware EVA agent.

Provider Configuration

• Modality-specific profile fields (image_model, audio_model, tts_model, tts_voice) are now surfaced in Settings so image/audio providers can be configured without guessing JSON keys.
• Guard native TaskProcessing provider registration by the availability of the corresponding Nextcloud core task type, preventing Nextcloud 34 dashboard failures on newer optional task types.

Terminal & Tool System

• Native confirmation forms for terminal commands and command sequences with bounded interactive stdin input and per-command prompt support.
• Terminal command sequences can run up to five bounded commands in order, with redacted per-command results and timings in Agent runs.
• Terminal traces expand automatically in live chat so confirmed commands are easier to audit, with exposed execution timeouts.

Connector & Plugin System

• Generic connector discovery probes common OpenAPI and Swagger YAML paths in addition to JSON descriptions, without requiring PHP ext-yaml.
• Schema-free GraphQL connectors are discovered through safe GET probes and learned as confirmation-gated POST operations.
• OpenAPI YAML discovery works without the optional PHP ext-yaml extension through the bundled Symfony YAML parser.
• Plugin safety hardening. Mutating or destructive plugins are always confirmation-gated. Duplicate plugin tool names can no longer silently replace earlier implementations. All five execution surfaces are now supported, including RAG and task-processing proposal runs.
• Plugin metadata is exposed in Settings: risk level, execution surfaces, and confirmation requirements.
• Connector presets for Immich (X-API-Key) and Vaultwarden (Bearer) with security-focused documentation and recipes.
• Connector route learning with automatic refresh of stale discovery, validated path variables, and authenticated endpoint probing.
• Custom OpenAPI schema URLs for connectors with safe path parameter resolution.

Talk Integration

• Read and write Talk conversations via list_talk_rooms, read_talk_chat, and send_talk_message tools. Messages are posted as the signed-in user with no bot label.
• Focused Talk bots for research, calendar, mail, and files, preventing duplicate responses.
• Email summarization through the Nextcloud Mail app.

Search & Indexing

• Nextcloud Unified Search integration. EVA's indexed file knowledge is searchable from the Nextcloud search bar with live VFS permission checks and direct links back to Files.
• Direct file search now searches unindexed Office documents and reports bounded diagnostics (nodes visited, documents extracted).
• Index lock recovery prevents stale worker locks from blocking indexing.
• Partial results on folder errors: accessible folders are returned even when a nested remote folder cannot be listed.

Dashboard

• EVA starters on the Nextcloud Dashboard with focused entry points for common tasks.

Performance & Security

• Memory bounds on index traversal, connector batches, and background queue scans.
• CPU spike prevention for local embeddings on cloud chat requests.
• Session lock release during long chat requests.
• Bearer token and credential handling hardened: secrets are encrypted and never returned in settings or exports.
• CSP-compatible native Nextcloud controls throughout Settings.
• Third-party plugin output centrally bounded with credential redaction.

Bug Fixes

• Talk chat histories are now properly indexed (previously threw silently on every room).
• Long Talk messages no longer break room indexing.
• Google News redirect links are filtered to prevent consent interstitials.
• Browser renderer now correctly reports availability and searches the right directory.
• Connector credentials are preserved on blank edits and properly re-discovered on reconfiguration.
• Shared calendars are included through canonical CalDAV ACLs.
• Stale chat locks are recovered reliably.
• Settings controls remain visible after rebuild.

Documentation

• Complete external service plugin recipes for Immich and Vaultwarden with ACL, secret-handling, and test guidance.
• Terminal prompt guidance and connector quick-setup labels in both modern JSON and legacy JavaScript translation catalogs.
• App store screenshots and polished listing.

Added

• Multiple AI provider profiles: configure up to 20 OpenAI-compatible endpoints with independent IDs, names and models, then select them directly in EVA Settings. API keys remain encrypted per user and are never returned in settings or exports.
• Provider profile validation, documentation and localized settings labels.

Fixed

• Background agent runs now reschedule themselves while work remains, and the status poll wakes a missed worker tick so queued messages continue promptly.

A major release that transforms EVA from a RAG chat tool into a full multimodal AI assistant with vision, speech, audio, image generation, a plugin system, and deep Nextcloud integration.

Multimodal & Vision

• Image analysis in Assistant tasks. A new bounded core:analyze-images provider reads Nextcloud image files and sends them to the configured Ollama vision or OpenAI-compatible model. Native Ollama image messages are preserved and translated to OpenAI-compatible format for remote providers.
• Image generation. A bounded core:text2image provider for configured OpenAI-compatible image APIs. Generated images are stored in the user's EVA folder with returned Nextcloud file IDs.
• Sticker tool. A confirmation-gated create_sticker chat tool generates PNG stickers from text prompts and saves them to the user's EVA folder.
• OCR. A local OCR provider for image/PDF TaskProcessing inputs, reusing EVA's Tesseract safeguards and per-user file permissions.
• Multimodal chat. core:text2text:multimodal-chatwithtools supports bounded image attachments (up to 4 files, 12 MB total) with the central EVA tool policy.
• Multimodal ContextAgent. core:contextagent:multimodal-interaction with bounded, permission-checked image attachments while preserving EVA's confirmation workflow.

Audio & Speech

• Audio transcription and subtitles. core:audio2text and core:audio2text:subtitles providers using configurable OpenAI-compatible audio endpoints, with a 25 MB input limit and per-user file permissions.
• Text-to-Speech. Configurable OpenAI-compatible TTS output stored as a Nextcloud audio file.
• Audio-to-audio translation. Transcribe, translate, and synthesize with bounded provider calls.
• Audio chat. core:audio2audio:chat transcribes a voice message, answers with the configured chat model, and saves the spoken response.
• ContextAgent audio. core:contextagent:audio-interaction routes bounded voice input through the confirmation-aware EVA agent.

Provider Configuration

• Modality-specific profile fields (image_model, audio_model, tts_model, tts_voice) are now surfaced in Settings so image/audio providers can be configured without guessing JSON keys.
• Guard native TaskProcessing provider registration by the availability of the corresponding Nextcloud core task type, preventing Nextcloud 34 dashboard failures on newer optional task types.

Terminal & Tool System

• Native confirmation forms for terminal commands and command sequences with bounded interactive stdin input and per-command prompt support.
• Terminal command sequences can run up to five bounded commands in order, with redacted per-command results and timings in Agent runs.
• Terminal traces expand automatically in live chat so confirmed commands are easier to audit, with exposed execution timeouts.

Connector & Plugin System

• Generic connector discovery probes common OpenAPI and Swagger YAML paths in addition to JSON descriptions, without requiring PHP ext-yaml.
• Schema-free GraphQL connectors are discovered through safe GET probes and learned as confirmation-gated POST operations.
• OpenAPI YAML discovery works without the optional PHP ext-yaml extension through the bundled Symfony YAML parser.
• Plugin safety hardening. Mutating or destructive plugins are always confirmation-gated. Duplicate plugin tool names can no longer silently replace earlier implementations. All five execution surfaces are now supported, including RAG and task-processing proposal runs.
• Plugin metadata is exposed in Settings: risk level, execution surfaces, and confirmation requirements.
• Connector presets for Immich (X-API-Key) and Vaultwarden (Bearer) with security-focused documentation and recipes.
• Connector route learning with automatic refresh of stale discovery, validated path variables, and authenticated endpoint probing.
• Custom OpenAPI schema URLs for connectors with safe path parameter resolution.

Talk Integration

• Read and write Talk conversations via list_talk_rooms, read_talk_chat, and send_talk_message tools. Messages are posted as the signed-in user with no bot label.
• Focused Talk bots for research, calendar, mail, and files, preventing duplicate responses.
• Email summarization through the Nextcloud Mail app.

Search & Indexing

• Nextcloud Unified Search integration. EVA's indexed file knowledge is searchable from the Nextcloud search bar with live VFS permission checks and direct links back to Files.
• Direct file search now searches unindexed Office documents and reports bounded diagnostics (nodes visited, documents extracted).
• Index lock recovery prevents stale worker locks from blocking indexing.
• Partial results on folder errors: accessible folders are returned even when a nested remote folder cannot be listed.

Dashboard

• EVA starters on the Nextcloud Dashboard with focused entry points for common tasks.

Performance & Security

• Memory bounds on index traversal, connector batches, and background queue scans.
• CPU spike prevention for local embeddings on cloud chat requests.
• Session lock release during long chat requests.
• Bearer token and credential handling hardened: secrets are encrypted and never returned in settings or exports.
• CSP-compatible native Nextcloud controls throughout Settings.
• Third-party plugin output centrally bounded with credential redaction.

Bug Fixes

• Talk chat histories are now properly indexed (previously threw silently on every room).
• Long Talk messages no longer break room indexing.
• Google News redirect links are filtered to prevent consent interstitials.
• Browser renderer now correctly reports availability and searches the right directory.
• Connector credentials are preserved on blank edits and properly re-discovered on reconfiguration.
• Shared calendars are included through canonical CalDAV ACLs.
• Stale chat locks are recovered reliably.
• Settings controls remain visible after rebuild.

Documentation

• Complete external service plugin recipes for Immich and Vaultwarden with ACL, secret-handling, and test guidance.
• Terminal prompt guidance and connector quick-setup labels in both modern JSON and legacy JavaScript translation catalogs.
• App store screenshots and polished listing.

Added

• Multiple AI provider profiles: configure up to 20 OpenAI-compatible endpoints with independent IDs, names and models, then select them directly in EVA Settings. API keys remain encrypted per user and are never returned in settings or exports.
• Provider profile validation, documentation and localized settings labels.

Fixed

• Background agent runs now reschedule themselves while work remains, and the status poll wakes a missed worker tick so queued messages continue promptly.

A major release that transforms EVA from a RAG chat tool into a full multimodal AI assistant with vision, speech, audio, image generation, a plugin system, and deep Nextcloud integration.

Multimodal & Vision

• Image analysis in Assistant tasks. A new bounded core:analyze-images provider reads Nextcloud image files and sends them to the configured Ollama vision or OpenAI-compatible model. Native Ollama image messages are preserved and translated to OpenAI-compatible format for remote providers.
• Image generation. A bounded core:text2image provider for configured OpenAI-compatible image APIs. Generated images are stored in the user's EVA folder with returned Nextcloud file IDs.
• Sticker tool. A confirmation-gated create_sticker chat tool generates PNG stickers from text prompts and saves them to the user's EVA folder.
• OCR. A local OCR provider for image/PDF TaskProcessing inputs, reusing EVA's Tesseract safeguards and per-user file permissions.
• Multimodal chat. core:text2text:multimodal-chatwithtools supports bounded image attachments (up to 4 files, 12 MB total) with the central EVA tool policy.
• Multimodal ContextAgent. core:contextagent:multimodal-interaction with bounded, permission-checked image attachments while preserving EVA's confirmation workflow.

Audio & Speech

• Audio transcription and subtitles. core:audio2text and core:audio2text:subtitles providers using configurable OpenAI-compatible audio endpoints, with a 25 MB input limit and per-user file permissions.
• Text-to-Speech. Configurable OpenAI-compatible TTS output stored as a Nextcloud audio file.
• Audio-to-audio translation. Transcribe, translate, and synthesize with bounded provider calls.
• Audio chat. core:audio2audio:chat transcribes a voice message, answers with the configured chat model, and saves the spoken response.
• ContextAgent audio. core:contextagent:audio-interaction routes bounded voice input through the confirmation-aware EVA agent.

Provider Configuration

• Modality-specific profile fields (image_model, audio_model, tts_model, tts_voice) are now surfaced in Settings so image/audio providers can be configured without guessing JSON keys.
• Guard native TaskProcessing provider registration by the availability of the corresponding Nextcloud core task type, preventing Nextcloud 34 dashboard failures on newer optional task types.

Terminal & Tool System

• Native confirmation forms for terminal commands and command sequences with bounded interactive stdin input and per-command prompt support.
• Terminal command sequences can run up to five bounded commands in order, with redacted per-command results and timings in Agent runs.
• Terminal traces expand automatically in live chat so confirmed commands are easier to audit, with exposed execution timeouts.

Connector & Plugin System

• Generic connector discovery probes common OpenAPI and Swagger YAML paths in addition to JSON descriptions, without requiring PHP ext-yaml.
• Schema-free GraphQL connectors are discovered through safe GET probes and learned as confirmation-gated POST operations.
• OpenAPI YAML discovery works without the optional PHP ext-yaml extension through the bundled Symfony YAML parser.
• Plugin safety hardening. Mutating or destructive plugins are always confirmation-gated. Duplicate plugin tool names can no longer silently replace earlier implementations. All five execution surfaces are now supported, including RAG and task-processing proposal runs.
• Plugin metadata is exposed in Settings: risk level, execution surfaces, and confirmation requirements.
• Connector presets for Immich (X-API-Key) and Vaultwarden (Bearer) with security-focused documentation and recipes.
• Connector route learning with automatic refresh of stale discovery, validated path variables, and authenticated endpoint probing.
• Custom OpenAPI schema URLs for connectors with safe path parameter resolution.

Talk Integration

• Read and write Talk conversations via list_talk_rooms, read_talk_chat, and send_talk_message tools. Messages are posted as the signed-in user with no bot label.
• Focused Talk bots for research, calendar, mail, and files, preventing duplicate responses.
• Email summarization through the Nextcloud Mail app.

Search & Indexing

• Nextcloud Unified Search integration. EVA's indexed file knowledge is searchable from the Nextcloud search bar with live VFS permission checks and direct links back to Files.
• Direct file search now searches unindexed Office documents and reports bounded diagnostics (nodes visited, documents extracted).
• Index lock recovery prevents stale worker locks from blocking indexing.
• Partial results on folder errors: accessible folders are returned even when a nested remote folder cannot be listed.

Dashboard

• EVA starters on the Nextcloud Dashboard with focused entry points for common tasks.

Performance & Security

• Memory bounds on index traversal, connector batches, and background queue scans.
• CPU spike prevention for local embeddings on cloud chat requests.
• Session lock release during long chat requests.
• Bearer token and credential handling hardened: secrets are encrypted and never returned in settings or exports.
• CSP-compatible native Nextcloud controls throughout Settings.
• Third-party plugin output centrally bounded with credential redaction.

Bug Fixes

• Talk chat histories are now properly indexed (previously threw silently on every room).
• Long Talk messages no longer break room indexing.
• Google News redirect links are filtered to prevent consent interstitials.
• Browser renderer now correctly reports availability and searches the right directory.
• Connector credentials are preserved on blank edits and properly re-discovered on reconfiguration.
• Shared calendars are included through canonical CalDAV ACLs.
• Stale chat locks are recovered reliably.
• Settings controls remain visible after rebuild.

Documentation

• Complete external service plugin recipes for Immich and Vaultwarden with ACL, secret-handling, and test guidance.
• Terminal prompt guidance and connector quick-setup labels in both modern JSON and legacy JavaScript translation catalogs.
• App store screenshots and polished listing.

Added

• Multiple AI provider profiles: configure up to 20 OpenAI-compatible endpoints with independent IDs, names and models, then select them directly in EVA Settings. API keys remain encrypted per user and are never returned in settings or exports.
• Provider profile validation, documentation and localized settings labels.

Fixed

• Background agent runs now reschedule themselves while work remains, and the status poll wakes a missed worker tick so queued messages continue promptly.

A major release that transforms EVA from a RAG chat tool into a full multimodal AI assistant with vision, speech, audio, image generation, a plugin system, and deep Nextcloud integration.

Multimodal & Vision

• Image analysis in Assistant tasks. A new bounded core:analyze-images provider reads Nextcloud image files and sends them to the configured Ollama vision or OpenAI-compatible model. Native Ollama image messages are preserved and translated to OpenAI-compatible format for remote providers.
• Image generation. A bounded core:text2image provider for configured OpenAI-compatible image APIs. Generated images are stored in the user's EVA folder with returned Nextcloud file IDs.
• Sticker tool. A confirmation-gated create_sticker chat tool generates PNG stickers from text prompts and saves them to the user's EVA folder.
• OCR. A local OCR provider for image/PDF TaskProcessing inputs, reusing EVA's Tesseract safeguards and per-user file permissions.
• Multimodal chat. core:text2text:multimodal-chatwithtools supports bounded image attachments (up to 4 files, 12 MB total) with the central EVA tool policy.
• Multimodal ContextAgent. core:contextagent:multimodal-interaction with bounded, permission-checked image attachments while preserving EVA's confirmation workflow.

Audio & Speech

• Audio transcription and subtitles. core:audio2text and core:audio2text:subtitles providers using configurable OpenAI-compatible audio endpoints, with a 25 MB input limit and per-user file permissions.
• Text-to-Speech. Configurable OpenAI-compatible TTS output stored as a Nextcloud audio file.
• Audio-to-audio translation. Transcribe, translate, and synthesize with bounded provider calls.
• Audio chat. core:audio2audio:chat transcribes a voice message, answers with the configured chat model, and saves the spoken response.
• ContextAgent audio. core:contextagent:audio-interaction routes bounded voice input through the confirmation-aware EVA agent.

Provider Configuration

• Modality-specific profile fields (image_model, audio_model, tts_model, tts_voice) are now surfaced in Settings so image/audio providers can be configured without guessing JSON keys.
• Guard native TaskProcessing provider registration by the availability of the corresponding Nextcloud core task type, preventing Nextcloud 34 dashboard failures on newer optional task types.

Terminal & Tool System

• Native confirmation forms for terminal commands and command sequences with bounded interactive stdin input and per-command prompt support.
• Terminal command sequences can run up to five bounded commands in order, with redacted per-command results and timings in Agent runs.
• Terminal traces expand automatically in live chat so confirmed commands are easier to audit, with exposed execution timeouts.

Connector & Plugin System

• Generic connector discovery probes common OpenAPI and Swagger YAML paths in addition to JSON descriptions, without requiring PHP ext-yaml.
• Schema-free GraphQL connectors are discovered through safe GET probes and learned as confirmation-gated POST operations.
• OpenAPI YAML discovery works without the optional PHP ext-yaml extension through the bundled Symfony YAML parser.
• Plugin safety hardening. Mutating or destructive plugins are always confirmation-gated. Duplicate plugin tool names can no longer silently replace earlier implementations. All five execution surfaces are now supported, including RAG and task-processing proposal runs.
• Plugin metadata is exposed in Settings: risk level, execution surfaces, and confirmation requirements.
• Connector presets for Immich (X-API-Key) and Vaultwarden (Bearer) with security-focused documentation and recipes.
• Connector route learning with automatic refresh of stale discovery, validated path variables, and authenticated endpoint probing.
• Custom OpenAPI schema URLs for connectors with safe path parameter resolution.

Talk Integration

• Read and write Talk conversations via list_talk_rooms, read_talk_chat, and send_talk_message tools. Messages are posted as the signed-in user with no bot label.
• Focused Talk bots for research, calendar, mail, and files, preventing duplicate responses.
• Email summarization through the Nextcloud Mail app.

Search & Indexing

• Nextcloud Unified Search integration. EVA's indexed file knowledge is searchable from the Nextcloud search bar with live VFS permission checks and direct links back to Files.
• Direct file search now searches unindexed Office documents and reports bounded diagnostics (nodes visited, documents extracted).
• Index lock recovery prevents stale worker locks from blocking indexing.
• Partial results on folder errors: accessible folders are returned even when a nested remote folder cannot be listed.

Dashboard

• EVA starters on the Nextcloud Dashboard with focused entry points for common tasks.

Performance & Security

• Memory bounds on index traversal, connector batches, and background queue scans.
• CPU spike prevention for local embeddings on cloud chat requests.
• Session lock release during long chat requests.
• Bearer token and credential handling hardened: secrets are encrypted and never returned in settings or exports.
• CSP-compatible native Nextcloud controls throughout Settings.
• Third-party plugin output centrally bounded with credential redaction.

Bug Fixes

• Talk chat histories are now properly indexed (previously threw silently on every room).
• Long Talk messages no longer break room indexing.
• Google News redirect links are filtered to prevent consent interstitials.
• Browser renderer now correctly reports availability and searches the right directory.
• Connector credentials are preserved on blank edits and properly re-discovered on reconfiguration.
• Shared calendars are included through canonical CalDAV ACLs.
• Stale chat locks are recovered reliably.
• Settings controls remain visible after rebuild.

Documentation

• Complete external service plugin recipes for Immich and Vaultwarden with ACL, secret-handling, and test guidance.
• Terminal prompt guidance and connector quick-setup labels in both modern JSON and legacy JavaScript translation catalogs.
• App store screenshots and polished listing.

Added

• Multiple AI provider profiles: configure up to 20 OpenAI-compatible endpoints with independent IDs, names and models, then select them directly in EVA Settings. API keys remain encrypted per user and are never returned in settings or exports.
• Provider profile validation, documentation and localized settings labels.

Fixed

• Background agent runs now reschedule themselves while work remains, and the status poll wakes a missed worker tick so queued messages continue promptly.