mirror of
https://github.com/memohai/Memoh.git
synced 2026-04-25 07:00:48 +09:00
fa48f59d3819e41592664e25ce5a340bd55791ea
32 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 lijianlin
|
0c2527453b | fix(schedule): use anyOf for nullable max_calls | ||
|
Fodesu
|
3945bd913d | test(agent): satisfy lint for generate loop tests | ||
|
Fodesu
|
33461d7ac1 | fix(agent): stop generate loop aborts correctly | ||
|
Fodesu
|
59147b255d | fix(agent): guard tool loop state against concurrent tool execution | ||
|
EYHN
|
c4114227e5 | feat(agent): add background task execution and notifications (#365) | ||
|
KasuganoSora
|
a40207ab6d | feat: Misskey channel adapter, agent reliability hardening & stream error resilience (#359) | ||
|
Acbox Liu
|
7a21fd5f07 | feat: ui message (#357) | ||
|
BBQ
|
d3bf6bc90a |
fix(channel,attachment): channel quality refactor & attachment pipeline fixes (#349)
* feat(channel): add DingTalk channel adapter - Add DingTalk channel adapter (`internal/channel/adapters/dingtalk/`) using dingtalk-stream-sdk-go, supporting inbound message receiving and outbound text/markdown reply - Register DingTalk adapter in cmd/agent and cmd/memoh - Add go.mod dependency: github.com/memohai/dingtalk-stream-sdk-go - Add Dingtalk and Wecom SVG icons and Vue components to @memohai/icon - Refactor existing icon components to remove redundant inline wrappers - Add `channelTypeDisplayName` util for consistent channel label resolution - Add DingTalk/WeCom i18n entries (en/zh) for types and typesShort - Extend channel-icon, bot-channels, channel-settings-panel to support dingtalk/wecom - Use channelTypeDisplayName in profile page to replace ad-hoc i18n lookup * fix(channel,attachment): channel quality refactor & attachment pipeline fixes Channel module: - Fix RemoveAdapter not cleaning connectionMeta (stale status leak) - Fix preparedAttachmentTypeFromMime misclassifying image/gif - Fix sleepWithContext time.After goroutine/timer leak - Export IsDataURL/IsHTTPURL/IsDataPath, dedup across packages - Cache OutboundPolicy in managerOutboundStream to avoid repeated lookups - Split OutboundAttachmentStore: extract ContainerAttachmentIngester interface - Add ManagerOption funcs (WithInboundQueueSize, WithInboundWorkers, WithRefreshInterval) - Add thread-safety docs on OutboundStream / managerOutboundStream - Add debug logs on successful send/edit paths - Expand outbound_prepare_test.go with 21 new cases - Convert no-receiver adapter helpers to package-level funcs; drop unused params DingTalk adapter: - Implement AttachmentResolver: download inbound media via /v1.0/robot/messageFiles/download - Fix pure-image inbound messages failing due to missing resolver Attachment pipeline: - Fix images invisible to LLM in pipeline (DCP) path: inject InlineImages into last user message when cfg.Query is empty - Fix public_url fallback: skip direct URL-to-LLM when ContentHash is set, always prefer inlined persisted asset - Inject path: carry ImageParts through agent.InjectMessage; inline persisted attachments in resolver inject goroutine so mid-stream images reach the model - Fix ResolveMime for images: prefer content-sniffed MIME over platform-declared MIME (fixes Feishu sending image/png header for actual JPEG content → API 400) |
||
|
Acbox Liu
|
8d5c38f0e5 |
refactor: unify providers and models tables (#338)
* refactor: unify providers and models tables
- Rename `llm_providers` → `providers`, `llm_provider_oauth_tokens` → `provider_oauth_tokens`
- Remove `tts_providers` and `tts_models` tables; speech models now live in the unified `models` table with `type = 'speech'`
- Replace top-level `api_key`/`base_url` columns with a JSONB `config` field on `providers`
- Rename `llm_provider_id` → `provider_id` across all references
- Add `edge-speech` client type and `conf/providers/edge.yaml` default provider
- Create new read-only speech endpoints (`/speech-providers`, `/speech-models`) backed by filtered views of the unified tables
- Remove old TTS CRUD handlers; simplify speech page to read-only + test
- Update registry loader to skip malformed YAML files instead of failing entirely
- Fix YAML quoting for model names containing colons in openrouter.yaml
- Regenerate sqlc, swagger, and TypeScript SDK
* fix: exclude speech providers from providers list endpoint
ListProviders now filters out client_type matching '%-speech' so Edge
and future speech providers no longer appear on the Providers page.
ListSpeechProviders uses the same pattern match instead of hard-coding
'edge-speech'.
* fix: use explicit client_type list instead of LIKE pattern
Replace '%-speech' pattern with explicit IN ('edge-speech') for both
ListProviders (exclusion) and ListSpeechProviders (inclusion). New
speech client types must be added to both queries.
* fix: use EXECUTE for dynamic SQL in migrations referencing old schema
PL/pgSQL pre-validates column/table references in static SQL statements
inside DO blocks before evaluating IF/RETURN guards. This caused
migrations 0010-0061 to fail on fresh databases where the canonical
schema uses `providers`/`provider_id` instead of `llm_providers`/
`llm_provider_id`.
Wrap all SQL that references potentially non-existent old schema objects
(llm_providers, llm_provider_id, tts_providers, tts_models, etc.) in
EXECUTE strings so they are only parsed at runtime when actually reached.
* fix: revert canonical schema to use llm_providers for migration compatibility
The CI migrations workflow (up → down → up) failed because 0061 down
renames `providers` back to `llm_providers`, but 0001 down only dropped
`providers` — leaving `llm_providers` as a remnant. On the second
migrate up, 0010 found the stale `llm_providers` and tried to reference
`models.llm_provider_id` which no longer existed.
Revert 0001 canonical schema to use original names (llm_providers,
tts_providers, tts_models) so incremental migrations work naturally and
0061 handles the final rename. Remove EXECUTE wrappers and unnecessary
guards from migrations that now always operate on llm_providers.
* fix: icons
* fix: sync canonical schema with 0061 migration to fix sqlc column mismatch
0001_init.up.sql still used old names (llm_providers, llm_provider_id)
and included dropped tts_providers/tts_models tables. sqlc could not
parse the PL/pgSQL EXECUTE in migration 0061, so generated code retained
stale columns (input_modalities, supports_reasoning) causing runtime
"column does not exist" errors when adding models.
- Update 0001_init.up.sql to current schema (providers, provider_id,
no tts tables, add provider_oauth_tokens)
- Use ALTER TABLE IF EXISTS in 0010/0041/0042 for backward compat
- Regenerate sqlc
* fix: guard all legacy migrations against fresh schema for CI compat
On fresh databases, 0001_init.up.sql creates providers/provider_id
(not llm_providers/llm_provider_id). Migrations 0013, 0041, 0046, 0047
referenced the old names without guards, causing CI migration failures.
- 0013: check llm_provider_id column exists before adding old constraint
- 0041: check llm_providers table exists before backfill/constraint DDL
- 0046: wrap CREATE TABLE in DO block with llm_providers existence check
- 0047: use ALTER TABLE IF EXISTS + DO block guard
|
||
|
Acbox Liu
|
43c4153938 |
feat: introduce DCP pipeline layer for unified context assembly (#329)
* refactor: introduce DCP pipeline layer for unified context assembly
Introduce a Deterministic Context Pipeline (DCP) inspired by Cahciua,
providing event-driven context assembly for LLM conversations.
- Add `internal/pipeline/` package with Canonical Event types, Projection
(reduce), Rendering (XML RC), Pipeline manager, and EventStore persistence
- Change user message format from YAML front-matter to XML `<message>` tags
with self-contained attributes (sender, channel, conversation, type)
- Merge CLI/Web dual API into single `/local/` endpoint, remove CLI handler
- Add `bot_session_events` table for event persistence and cold-start replay
- Add `discuss` session type (reserved for future Cahciua-style mode)
- Wire pipeline into HandleInbound: adapt → persist → push on every message
- Lazy cold-start replay: load events from DB on first session access
* feat: implement discuss mode with reactive driver and probe gate
Add discuss session mode where the bot autonomously decides when to speak
in group chats via tool-gated output (send tool only, no direct text reply).
- Add discuss driver (per-session goroutine, RC watch, step loop via
agent.Generate, TR persistence, late-binding prompt with mention hints)
- Add system_discuss.md prompt template ("text = inner monologue, send = speak")
- Add context composition (MergeContext, ComposeContext, TrimContext) for
RC + assistant/tool message interleaving by timestamp
- Add probe gate: when discuss_probe_model_id is set, cheap model pre-filters
group messages; no tool calls = silence, tool calls = activate primary
- Add /new [chat|discuss] command: explicit mode selection, defaults to
discuss in groups, chat in DMs, chat-only for WebUI
- Add ResolveRunConfig on flow.Resolver for discuss driver to reuse
model/tools/system-prompt resolution without reimplementing
- Fix send tool for discuss mode: same-conversation sends now go through
SendDirect (channel adapter) instead of the local emitter shortcut
- Add target attribute to XML message format (reply_target for routing)
- Add discuss_probe_model_id to bots table settings
- Remove pipeline compaction (SetCompactCursor) — reuse existing compaction.Service
- Persist full SDK messages (including tool calls) in discuss mode
* refactor: unify DCP event layer, fix persistence and local channel
- Fix bot_session_events dedup index to include event_kind so that
message + edit events for the same external_message_id coexist.
- Change CreateSessionEvent from :one to :exec so ON CONFLICT DO NOTHING
does not produce spurious errors on duplicate delivery.
- Move ACL evaluation before event ingest; denied messages no longer
enter bot_session_events or the in-memory pipeline.
- Let chat mode consume RenderedContext from the DCP pipeline when
available, sharing the same event-driven context assembly as discuss.
- Collapse local WebSocket handler to route through HandleInbound
instead of directly calling StreamChatWS, eliminating the dual
business entry point.
- Extract buildBaseRunConfig shared builder so resolve() and
ResolveRunConfig() no longer duplicate model/credentials/skills setup.
- Add StoreRound to RunConfigResolver interface so discuss driver
persists assistant output with full metadata, usage, and memory
extraction (same quality as chat mode).
- Fix discuss driver context: use context.Background() instead of the
short-lived HTTP request context that was getting cancelled.
- Fix model ID passed to StoreRound: return database UUID from
ResolveRunConfig instead of SDK model name.
- Remove dead CLIAdapter/CLIType and update legacy web/cli references
in tests and comments.
* fix: stop idle discuss goroutines after 10min timeout
Discuss session goroutines were never cleaned up when a session became
inactive (e.g. after /new). Add a 10-minute idle timer that auto-exits
the goroutine and removes it from the sessions map when no new RC
arrives.
* refactor: pipeline details — event types, structured reply, display content
- Remove [User sent N attachments] placeholder text from buildInboundQuery;
attachment info is now expressed via pipeline <attachment> tags.
- Unify in-reply-to as structured ReplyRef (Sender/Preview fields) across
Telegram, Discord, Feishu, and Matrix adapters instead of prepending
[Reply to ...] text into the message body. Remove now-unused
buildTelegramQuotedText, buildDiscordQuotedText, buildMatrixQuotedText.
- Make AdaptInbound return CanonicalEvent interface and dispatch to
adaptMessage/adaptEdit/adaptService based on metadata["event_type"].
- Add event_id column to bot_history_messages (migration 0059) so user
messages can reference their canonical pipeline event.
- PersistEvent now returns the event UUID; HandleInbound passes it through
to both persistPassiveMessage and ChatRequest.EventID for storeRound.
- Add FillDisplayContent to message service: extracts plain text from
event_data for clean frontend display.
- Frontend extractMessageText prefers display_content when available,
falling back to legacy strip logic for old messages.
- Fix: always generate headerifiedQuery for storage even when usePipeline
is true, so user messages are persisted via storeRound in chat mode.
* fix: use json.Marshal for pipeline context content serialization
The manual string escaping in buildMessagesFromPipeline only handled
double quotes but not newlines, backslashes, and other JSON special
characters, producing invalid json.RawMessage values. The LLM then
received empty/malformed context and complained about having no history.
* fix: restore WebSocket handler to use StreamChatWS directly
The previous refactoring replaced the WS handler with HandleInbound +
RouteHub subscription, which broke streaming because RouteHub events
use a different format (channel.StreamEvent) than what the frontend
expects (flow.WSStreamEvent with text_delta, tool_call_start, etc.).
Restore the original direct StreamChatWS call path so WebUI streaming
works again. The WS handler now matches the pre-refactoring behavior
while all other changes (pipeline, ACL, event types, etc.) are kept.
* feat: store display_text directly in bot_history_messages
Instead of computing display content at API response time by querying
bot_session_events via event_id, store the raw user text in a dedicated
display_text column at write time. This works for all paths including
the WebSocket handler which does not go through the pipeline/event layer.
- Migration 0060: add display_text TEXT column
- PersistInput gains DisplayText; filled from trimmedText (passive) and
req.Query (storeRound)
- toMessageFields reads display_text into DisplayContent
- Remove FillDisplayContent runtime query and ListSessionEventsByEventID
- Frontend already prefers display_content when available (no change)
* fix: display_text should contain raw user text, not XML-wrapped query
req.Query gets overwritten to headerifiedQuery (with XML <message> tags)
before storeRound runs. Add RawQuery field to ChatRequest to preserve
the original user text, and use it for display_text in storeMessages.
* fix(web): show discuss sessions
* refactor: introduce DCP pipeline layer for unified context assembly
Introduce a Deterministic Context Pipeline (DCP) inspired by Cahciua,
providing event-driven context assembly for LLM conversations.
- Add `internal/pipeline/` package with Canonical Event types, Projection
(reduce), Rendering (XML RC), Pipeline manager, and EventStore persistence
- Change user message format from YAML front-matter to XML `<message>` tags
with self-contained attributes (sender, channel, conversation, type)
- Merge CLI/Web dual API into single `/local/` endpoint, remove CLI handler
- Add `bot_session_events` table for event persistence and cold-start replay
- Add `discuss` session type (reserved for future Cahciua-style mode)
- Wire pipeline into HandleInbound: adapt → persist → push on every message
- Lazy cold-start replay: load events from DB on first session access
* feat: implement discuss mode with reactive driver and probe gate
Add discuss session mode where the bot autonomously decides when to speak
in group chats via tool-gated output (send tool only, no direct text reply).
- Add discuss driver (per-session goroutine, RC watch, step loop via
agent.Generate, TR persistence, late-binding prompt with mention hints)
- Add system_discuss.md prompt template ("text = inner monologue, send = speak")
- Add context composition (MergeContext, ComposeContext, TrimContext) for
RC + assistant/tool message interleaving by timestamp
- Add probe gate: when discuss_probe_model_id is set, cheap model pre-filters
group messages; no tool calls = silence, tool calls = activate primary
- Add /new [chat|discuss] command: explicit mode selection, defaults to
discuss in groups, chat in DMs, chat-only for WebUI
- Add ResolveRunConfig on flow.Resolver for discuss driver to reuse
model/tools/system-prompt resolution without reimplementing
- Fix send tool for discuss mode: same-conversation sends now go through
SendDirect (channel adapter) instead of the local emitter shortcut
- Add target attribute to XML message format (reply_target for routing)
- Add discuss_probe_model_id to bots table settings
- Remove pipeline compaction (SetCompactCursor) — reuse existing compaction.Service
- Persist full SDK messages (including tool calls) in discuss mode
* refactor: unify DCP event layer, fix persistence and local channel
- Fix bot_session_events dedup index to include event_kind so that
message + edit events for the same external_message_id coexist.
- Change CreateSessionEvent from :one to :exec so ON CONFLICT DO NOTHING
does not produce spurious errors on duplicate delivery.
- Move ACL evaluation before event ingest; denied messages no longer
enter bot_session_events or the in-memory pipeline.
- Let chat mode consume RenderedContext from the DCP pipeline when
available, sharing the same event-driven context assembly as discuss.
- Collapse local WebSocket handler to route through HandleInbound
instead of directly calling StreamChatWS, eliminating the dual
business entry point.
- Extract buildBaseRunConfig shared builder so resolve() and
ResolveRunConfig() no longer duplicate model/credentials/skills setup.
- Add StoreRound to RunConfigResolver interface so discuss driver
persists assistant output with full metadata, usage, and memory
extraction (same quality as chat mode).
- Fix discuss driver context: use context.Background() instead of the
short-lived HTTP request context that was getting cancelled.
- Fix model ID passed to StoreRound: return database UUID from
ResolveRunConfig instead of SDK model name.
- Remove dead CLIAdapter/CLIType and update legacy web/cli references
in tests and comments.
* fix: stop idle discuss goroutines after 10min timeout
Discuss session goroutines were never cleaned up when a session became
inactive (e.g. after /new). Add a 10-minute idle timer that auto-exits
the goroutine and removes it from the sessions map when no new RC
arrives.
* refactor: pipeline details — event types, structured reply, display content
- Remove [User sent N attachments] placeholder text from buildInboundQuery;
attachment info is now expressed via pipeline <attachment> tags.
- Unify in-reply-to as structured ReplyRef (Sender/Preview fields) across
Telegram, Discord, Feishu, and Matrix adapters instead of prepending
[Reply to ...] text into the message body. Remove now-unused
buildTelegramQuotedText, buildDiscordQuotedText, buildMatrixQuotedText.
- Make AdaptInbound return CanonicalEvent interface and dispatch to
adaptMessage/adaptEdit/adaptService based on metadata["event_type"].
- Add event_id column to bot_history_messages (migration 0059) so user
messages can reference their canonical pipeline event.
- PersistEvent now returns the event UUID; HandleInbound passes it through
to both persistPassiveMessage and ChatRequest.EventID for storeRound.
- Add FillDisplayContent to message service: extracts plain text from
event_data for clean frontend display.
- Frontend extractMessageText prefers display_content when available,
falling back to legacy strip logic for old messages.
- Fix: always generate headerifiedQuery for storage even when usePipeline
is true, so user messages are persisted via storeRound in chat mode.
* fix: use json.Marshal for pipeline context content serialization
The manual string escaping in buildMessagesFromPipeline only handled
double quotes but not newlines, backslashes, and other JSON special
characters, producing invalid json.RawMessage values. The LLM then
received empty/malformed context and complained about having no history.
* fix: restore WebSocket handler to use StreamChatWS directly
The previous refactoring replaced the WS handler with HandleInbound +
RouteHub subscription, which broke streaming because RouteHub events
use a different format (channel.StreamEvent) than what the frontend
expects (flow.WSStreamEvent with text_delta, tool_call_start, etc.).
Restore the original direct StreamChatWS call path so WebUI streaming
works again. The WS handler now matches the pre-refactoring behavior
while all other changes (pipeline, ACL, event types, etc.) are kept.
* feat: store display_text directly in bot_history_messages
Instead of computing display content at API response time by querying
bot_session_events via event_id, store the raw user text in a dedicated
display_text column at write time. This works for all paths including
the WebSocket handler which does not go through the pipeline/event layer.
- Migration 0060: add display_text TEXT column
- PersistInput gains DisplayText; filled from trimmedText (passive) and
req.Query (storeRound)
- toMessageFields reads display_text into DisplayContent
- Remove FillDisplayContent runtime query and ListSessionEventsByEventID
- Frontend already prefers display_content when available (no change)
* fix: display_text should contain raw user text, not XML-wrapped query
req.Query gets overwritten to headerifiedQuery (with XML <message> tags)
before storeRound runs. Add RawQuery field to ChatRequest to preserve
the original user text, and use it for display_text in storeMessages.
* fix(web): show discuss sessions
* chore(feishu): change discuss output to stream card
* fix(channel): unify discuss/chat send path and card markdown delivery
* feat(discuss): switch to stream execution with RouteHub broadcasting
* refactor(pipeline): remove context trimming from ComposeContext
The pipeline path should not trim context by token budget — the
upstream IC/RC already bounds the event window. Remove TrimContext,
FindWorkingWindowCursor, EstimateTokens, FormatLastProcessedMs (all
unused or only used for trimming), the maxTokens parameter from
ComposeContext, and MaxContextTokens from DiscussSessionConfig.
---------
Co-authored-by: 晨苒 <16112591+chen-ran@users.noreply.github.com>
|
||
|
Lakr
|
daed345908 |
feat(browser): add remote Playwright session support (Tier 2) (#325)
Add native Playwright WebSocket sessions alongside the existing curated browser tools. Agents can now create remote sessions that expose a full Playwright API over WebSocket, enabling advanced use cases like HttpOnly cookie injection, storage state management, and route interception. Key changes: - Per-bot isolated browser processes (launchServer via Node child process) - New session module with create/close/status/heartbeat endpoints - New browser_remote_session agent tool (Go) - Storage state export/import on existing browser contexts - Bot ID plumbing through context creation for process isolation - Inflight deduplication to prevent duplicate browser launches - Session janitor for automatic expiry cleanup |
||
|
Ringo.Typowriter
|
09c523f0b8 | refactor(agent): merge read_media tool into read tool (#326) | ||
|
Acbox Liu
|
5cfbaa40e2 |
refactor(agent): replace XML tag extraction with tool-based send/react/speak (#330)
* refactor(agent): replace XML tag extraction with tool-based send/react/speak Remove the <attachments>, <reactions>, and <speech> XML tag extraction system from the agent streaming pipeline. Instead, the send/react/speak tools now handle both same-conversation and cross-conversation delivery: - send: omit target to deliver attachments in the current conversation; specify target for cross-channel messaging - react: omit target to react in the current conversation - speak: omit target to speak in the current conversation Backend changes: - Add StreamEmitter callback to tools.SessionContext so tools can push attachment/reaction/speech events directly into the agent stream - Wire emitter in agent.go for both streaming and non-streaming paths - Remove StreamTagExtractor, DefaultTagResolvers, emitTagEvents, and delete internal/agent/tags.go entirely - Remove StripAgentTags calls from assistant_output.go - Add IsSameConversation detection in messaging executor; same-conv sends pass raw paths through the emitter for downstream ingestion - Auto-resolve relative paths (e.g. "IDENTITY.md" -> "/data/IDENTITY.md") - Add Metadata propagation through the full attachment chain (tools.Attachment -> agent.FileAttachment -> parseAttachmentDelta) - Update system_chat.md and _contacts.md prompts Frontend changes (apps/web): - Hide send/react/speak tool_call blocks when result indicates delivered to current conversation - Defer attachment_delta blocks to end of message (flush on stream completion) for consistent positioning with DB-loaded history * fix(agent): speak tool emits synthesized audio directly as voice attachment Instead of emitting speech_delta (which requires downstream re-synthesis), the speak tool now emits the already-synthesized audio as an attachment_delta with voice type. This avoids double TTS synthesis and eliminates dependency on ttsService being configured on the inbound processor. Also fixes speak on WebUI where ReplyTarget is empty (same fix as send). |
||
|
Acbox
|
a9a9f7e955 |
feat: add image generation model and generate_image agent tool
Bots can now be configured with an image generation model (must have image-output compatibility). When set, the agent exposes a generate_image tool that calls the model via Twilight AI SDK, saves the result to the bot container filesystem, and returns the file path. - Add image_model_id column to bots table (migration 0053) - Update settings SQL queries, service, and types - New ImageGenProvider tool provider in internal/agent/tools/ - Wire provider in both cmd/agent and cmd/memoh entry points - Add image model selector to frontend bot settings with compat filtering - Regenerate swagger, SDK types, and sqlc code |
||
|
Acbox
|
a73bac05fe |
fix(agent): skip tool injection for models without tool-call capability
Models that lack the "tool-call" compatibility flag now run without tools, preventing provider errors when the model does not support function calling. |
||
|
Acbox
|
a31995424c |
feat: add per-route message dispatch modes (inject/parallel/queue)
Introduce three inbound message handling modes for channel adapters: - inject (default, /btw): when a route has an active agent stream, inject the new user message into the running stream via the SDK's PrepareStep hook between tool rounds. The message is interleaved at the correct position in the persisted round. - parallel (/now): start a new agent stream immediately, running concurrently with any existing stream (preserves current behavior). - queue (/next): enqueue the message and process it after the current stream completes. Key components: - RouteDispatcher: per-route state management with inject channel, task queue, and active-stream tracking. - PrepareStep integration: drains inject channel between tool rounds, records insertion position via InjectedRecorder for correct persistence ordering. - interleaveInjectedMessages: inserts injected user messages at their actual injection position within the persisted message round. - Parallel mode isolation: /now streams do not interact with the dispatcher, preventing them from clearing another stream's active state. |
||
|
Acbox
|
bb14bcb3bc |
refactor: move skills directory from .skills to skills and enrich prompt
- Change skills storage path from `/data/.skills` to `/data/skills` - Add usage instructions and directory location to the Skills section in the system prompt |
||
|
Acbox
|
fc2b603018 |
fix(agent): skip tools for models without tool-call capability and parse image output
- Add SupportsToolCall to RunConfig; only inject tools into SDK when set - Update twilight-ai to 497ad09 which adds SSE scanner 10MB buffer (fixes token-too-long on large image payloads) and parses the images array from OpenAI-compatible chat completions into StreamFilePart |
||
|
Acbox
|
c1e6e0cc7a |
feat(agent): add pagination and smart collapsing to container list tool
Large directories like node_modules/.venv could return thousands of entries, wasting tokens and causing timeouts. Add offset/limit pagination to ListDir RPC and collapse heavy subdirectories (>50 items) into summaries in recursive mode. Collapsing runs at the bridge layer before pagination so the page window reflects the collapsed view. |
||
|
Acbox
|
f1dd30a388 |
fix: strip agent tags from IM/WebUI output and fix attachment display after refresh
Three independent bugs fixed: 1. IM channels were sending raw <attachments>/<reactions>/<speech> tag blocks alongside file attachments. Now ExtractAssistantOutputs strips these tags before building the outbound channel message. 2. WebUI rendered these tags as markdown after page refresh. Now extractMessageText strips agent tags for non-user messages. 3. WebUI lost attachment blocks after refresh because convertMessagesToChats did not call buildAssetBlocks when merging assistant messages into a pending tool-call group. Also made LinkOutboundAssets session-aware so assets are linked to the correct assistant message. |
||
|
Yiming Qi
|
64378d29ed |
feat: openai codex support (#292)
* feat(web): add provider oauth management ui * feat: add OAuth callback support on port 1455 * feat: enhance reasoning effort options and support for OpenAI Codex OAuth * feat: update twilight-ai dependency to v0.3.4 * refactor: promote openai-codex to first-class client_type, remove auth_type Replace the previous openai-responses + metadata auth_type=openai-codex-oauth combo with a dedicated openai-codex client_type. OAuth requirement is now determined solely by client_type, eliminating the auth_type concept from the LLM provider domain entirely. - Add openai-codex to DB CHECK constraint (migration 0047) with data migration - Add ClientTypeOpenAICodex constant and dedicated SDK/probe branches - Remove AuthType from SDKModelConfig, ModelCredentials, TriggerConfig, etc. - Simplify supportsOAuth to check client_type == openai-codex - Add conf/providers/codex.yaml preset with Codex catalog models - Frontend: replace auth_type selector with client_type-driven OAuth UI --------- Co-authored-by: Acbox <acbox0328@gmail.com> |
||
|
Acbox
|
da2e999ce3 |
feat: searchable timezone select & bot timezone priority
- Add reusable TimezoneSelect component with search and UTC offset labels - Replace plain Select with searchable TimezoneSelect in profile settings, bot settings, and browser context settings - Move bot timezone setting from header dialog into bot settings tab - Resolve timezone with bot > user > system priority for all LLM-facing time formatting (user message header, system prompt, heartbeat, tools, memory extraction) - Format tool output timestamps (history, contacts) in resolved timezone |
||
|
Acbox
|
65b2797626 |
refactor: unify SDK model factories into internal/models
Move CreateModel, BuildReasoningOptions, ReasoningBudgetTokens and related types from internal/agent to internal/models as NewSDKChatModel, SDKModelConfig, etc. This eliminates duplicate ClientType constants and centralises all Twilight AI SDK instance creation in a single package. NewSDKEmbeddingModel now accepts a clientType parameter and dispatches to the native Google embedding provider for google-generative-ai, instead of always using the OpenAI-compatible endpoint. |
||
|
Yiming Qi
|
03ba13e7e5 | feat: add timezone support for schedule and user runtime (#282) | ||
|
Acbox
|
5963e787a9 |
fix(agent): preserve inline tags in message history
Stop stripping <attachments>, <reactions>, and <speech> tags from assistant messages so the LLM retains full context across turns. |
||
|
Acbox
|
e9c9ed5ab1 |
fix(agent): route native images into user message for vision models
Images sent by users were silently dropped when the model supported vision: routeAttachmentsByCapability classified them as "Native", but extractFileRefPaths only collected "Fallback" (tool_file_ref) paths, so the image data URL was computed and then discarded — the model saw neither the image nor its container path. - Add InlineImages field to RunConfig to carry native image data - Replace extractFileRefPaths with extractAttachmentPaths that collects paths from both Native (FallbackPath) and Fallback attachments so the YAML header always lists every attachment - Add extractNativeImageParts to extract inline image data URLs - Pass InlineImages as sdk.ImagePart in prepareRunConfig so the LLM receives the actual image content alongside the text query |
||
|
Ran
|
93097d50b2 | refactor(memory): replace sdk to twilight | ||
|
Acbox Liu
|
b3a39ad93d |
refactor: replace persistent subagents with ephemeral spawn tool (#280)
* refactor: replace persistent subagents with ephemeral spawn tool (#subagent) - Drop subagents table, remove all persistent subagent infrastructure - Add 'subagent' session type with parent_session_id on bot_sessions - Rewrite subagent tool as single 'spawn' tool with parallel execution - Create system_subagent.md prompt, add _subagent.md include for chat - Limit subagent tools to file, exec, web_search, web_fetch only - Merge subagent token usage into parent chat session in reporting - Remove frontend subagent management page, update chat UI for spawn - Fix UTF-8 truncation in session title, fix query not passed to agent * refactor: remove history message page |
||
|
Acbox Liu
|
b88ca96064 |
refactor: provider & models (#277)
* refactor: move client_type to provider, replace model fields with config JSONB - Move `client_type` from `models` to `llm_providers` table - Add `icon` field to `llm_providers` - Replace `dimensions`, `input_modalities`, `supports_reasoning` on `models` with a single `config` JSONB column containing `dimensions`, `compatibilities` (vision, tool-call, image-output, reasoning), and `context_window` - Auto-imported models default to vision + tool-call + reasoning - Update all backend consumers (agent, flow resolver, handlers, memory) - Regenerate sqlc, swagger, and TypeScript SDK - Update frontend forms, display, and i18n for new schema * ui: show provider icon avatar in sidebar and detail header, remove icon input * feat: add built-in provider registry with YAML definitions and enable toggle - Add `enable` column to llm_providers (default true, backward-compatible) - Create internal/registry package to load YAML provider/model definitions on startup and upsert into database (new providers disabled by default) - Add conf/providers/ with OpenAI, Anthropic, Google YAML definitions - Add RegistryConfig to TOML config (providers_dir, default conf/providers) - Model listing APIs and conversation flow now filter by enabled providers - Frontend: enable switch in provider form, green status dot in sidebar, enabled providers sorted to top * fix: make 0041 migration idempotent for fresh databases Guard data migration steps with column-existence checks so the migration succeeds on databases created from the updated init schema. |
||
|
Acbox Liu
|
7d7d0e4b51 |
refactor: introduce multi-session chat support (#session) (#267)
* refactor: introduce multi-session chat support (#session)
Replace the single-context-per-bot model with multiple chat sessions.
Database:
- Add bot_sessions table (route_id, channel_type, title, metadata, soft delete)
- Migrate bot_history_messages from (route_id, channel_type) to session_id
- Add active_session_id to bot_channel_routes
- Migration 0036 handles data migration from existing messages
Backend:
- New internal/session service for session CRUD
- Update message service/types to use session_id instead of route_id
- Update conversation flow (resolver, history, store) for session context
- Channel inbound auto-creates/retrieves active session via SessionEnsurer
- New REST endpoints: /bots/:bot_id/sessions (CRUD)
- WebSocket and message handlers accept optional session_id
- Wire session service into FX dependency graph (agent + memoh)
Frontend:
- Refactor chat store: sessions replaces chats, sessionId replaces chatId
- Session-aware message loading, sending, and pagination
- WebSocket sends include session_id
- New session sidebar component with select/delete
- Chat area header shows active session title + new session button
- API layer updated: fetchSessions, createSession, deleteSession
- i18n strings for session management (en + zh)
SDK:
- Regenerated TypeScript SDK and Swagger docs with session endpoints
* fix: update tests for session refactoring (RouteID → SessionID)
Remove references to removed RouteID and Platform fields from
PersistInput/Message in channel_test.go and service_integration_test.go.
* fix: restore accidentally deleted SDK files and guard migration 0032
- Restore packages/sdk/src/container-stream.ts and extra/index.ts that
were accidentally removed during SDK regeneration
- Wrap migration 0032 route_id index creation in a column existence check
to avoid failure on fresh databases where 0001_init.up.sql no longer
has route_id
* fix: guard migration 0036 data steps for fresh databases
Wrap steps 3-7 (which reference route_id/channel_type on
bot_history_messages) in a column existence check so the migration
is safe on fresh databases where 0001_init.up.sql already reflects
the final schema without those columns.
* feat: add title model setting and auto-generate session titles on user input
- Add title_model_id to bots table (migration 0037) and bot settings API
- Implement async title generation triggered at user message time (not after
assistant response) for faster title availability
- Publish session_title_updated events via SSE event hub for real-time
frontend updates without page refresh
- Fix SSE message event parsing: use direct JSON.parse instead of
normalizeStreamEvent which silently dropped non-chat-stream event types
- Add title model selector in bot settings UI with i18n support
* fix: session-scoped message filtering and URL-based chat routing
- Filter realtime SSE messages by session_id to prevent cross-session
message leakage after page refresh
- Add /chat/:sessionId? route with bidirectional URL ↔ store sync
- Visiting /chat shows a clean state with no bot or session pre-selected
- Visiting /chat/:sessionId loads the specific session directly
- Session switches from sidebar automatically update the URL
- Fix stale RouteID field in dedupe test (removed during session refactor)
* fix: skip cross-channel stream events to prevent session leakage
The bot-level web stream pushes events from all channels (Telegram,
Discord, etc.) without session_id context. Previously these were
rendered inline in the current chat view regardless of session.
Now cross-channel events are ignored in handleLocalStreamEvent;
persisted messages arrive via the SSE message events stream with
proper session_id filtering through appendRealtimeMessage.
* feat: show IM avatars and platform badges on session sidebar
- Add sender_avatar_url to route metadata from identity resolution
- Resolve group avatar and handle via directory adapter for group chats
- JOIN bot_channel_routes in ListSessionsByBot to return route metadata
- Display avatar with ChannelBadge on IM session items (group avatar
for groups, sender avatar for private chats)
- Show @groupname or @username as session sub-label
* fix: clean up RunConfig unused fields, fix skill system and copy bug
- Remove unused RunConfig fields: Tools, Channels, CurrentChannel,
ActiveContextTime
- Remove unused SessionContext fields: DisplayName, ConversationType
- Fix EnabledSkillNames copy bug: make([]string, 0, n) + copy copies
zero elements; changed to make([]string, n)
- Fix prepareRunConfig dead code: remove no-op loop over
CurrentPlatform runes; compute supportsImageInput from model's
InputModalities
- Fix EnabledSkills always nil in system prompt: resolve enabled skill
entries from EnabledSkillNames + Skills
- Fix use_skill tool returning empty response: now returns full skill
content (description + instructions) so LLM gets it in the same turn
- Skip use_skill tool registration when no skills are available
- Conditionally render Skills section in system prompt (hidden when
no skills exist)
* feat: add session type field and bind sessions to heartbeat/schedule executions
- Add `type` column to `bot_sessions` (chat | heartbeat | schedule)
- Add `session_id` to `bot_heartbeat_logs` for per-execution session tracking
- Create `schedule_logs` table binding schedule_id + session_id
- Heartbeat and schedule runs now create independent sessions and persist
agent messages via storeRound, enabling full conversation replay
- Add schedule logs API endpoints (list by bot, list by schedule, delete)
- Update Triggerer interfaces to return TriggerResult with status/usage/model
* refactor: modular system prompts per session type (chat/heartbeat/schedule)
Split the monolithic system.md into three type-specific system prompts
with shared fragments via {{include:_xxx}} syntax, so each session type
gets a focused prompt without irrelevant instructions.
* fix: prevent message duplication after task completion
message_created events from Persist() had an empty platform field because
toMessageFromCreate() didn't extract it from the session. This caused
appendRealtimeMessage to fail the platform === 'web' guard, and
hasMessageWithId to fail because local IDs differ from server UUIDs,
resulting in all messages being appended as duplicates.
- Extract platform from metadata in toMessageFromCreate so published events
carry the correct value
- Pass channel_type: 'web' when creating sessions from the web frontend so
List queries return the correct platform via the session JOIN
* fix: use per-message usage from SDK instead of misaligned step-level usages
Previously, token usage was stored via a separate per-step usages array
that didn't align with messages (off-by-one from prepending user message,
step count != message count). This caused:
- User messages incorrectly receiving usage data
- Usage values shifted across messages in multi-step rounds
- Last assistant message getting the accumulated total instead of its own step usage
- InputTokenDetails/OutputTokenDetails lost during manual accumulation
Now each sdk.Message carries its own per-step Usage (set by the SDK in
buildStepMessages), which is extracted in sdkMessagesToModelMessages and
stored directly via ModelMessage.Usage. The storeRound/storeMessages path
no longer needs external usage/usages parameters.
Also fixes the totalUsage accumulation in runStream to include all detail
fields (InputTokenDetails, OutputTokenDetails).
* feat: add /new slash command to create a new active session from IM channels
Users in Telegram/Discord/Feishu can now send /new to start a fresh
conversation, resetting the session context for the current chat thread.
The command resolves the channel route, creates a new session, sets it as
the active session on the route, and replies with a confirmation message.
* feat: distinguish heartbeat and schedule sessions with dedicated icons in sidebar
Heartbeat sessions show a heart-pulse icon (rose), schedule sessions
show a clock icon (amber), and both display a type label beneath the
session title.
* refactor: remove enabledSkills system prompt injection, keep sorted skill listing
use_skill now returns skill content directly as tool output, so there is
no need to inject enabled skill body text into the system prompt. Remove
the entire enabledSkills tracking chain (RunConfig.EnabledSkillNames,
StreamEvent.Skills, GenerateResult.Skills, ChatRequest/Response.Skills,
enableSkill closures in runStream/runGenerate, prepareRunConfig matching).
Keep a lightweight skills listing (name + description only) in the system
prompt so the model knows which skills are available. Sort entries by name
to guarantee deterministic ordering and maximize KV cache reuse.
* refactor: remove inbox system, persist passive messages directly to history
Replace the bot_inbox table and service with direct writes to
bot_history_messages for group conversations where the bot is not
@mentioned. Trigger-path messages continue to be persisted after the
agent responds (unchanged).
- Drop bot_inbox table and max_inbox_items column (migration 0039)
- Delete internal/inbox/, handlers/inbox.go, command/inbox.go,
agent/tools/inbox.go and the MCP message provider
- Add persistPassiveMessage() in channel inbound to write user
messages into the active session immediately
- Rewrite ListObservedConversationsByChannelIdentity to query
bot_history_messages + bot_sessions instead of bot_inbox
- Extract shared send/react logic into internal/messaging/executor.go;
agent/tools/message.go is now a thin SDK adapter
- Clean up all inbox references from agent prompts, flow resolver,
email trigger, settings, commands, DI wiring, and frontend
- Regenerate sqlc, swagger, and SDK
* feat: add list_sessions and search_messages agent tools
Provide agents with the ability to query session metadata and search
message history across all sessions. search_messages supports filtering
by time range, keyword (JSONB-aware ILIKE), session, contact, and role,
with a default 7-day lookback when no start_time is given.
* feat: inject last_heartbeat time and improve heartbeat search guidance
Query the previous heartbeat's started_at timestamp and pass it through
TriggerPayload into the heartbeat prompt template. Update system prompt
and HEARTBEAT.md checklist to guide agents to use search_messages with
start_time=last_heartbeat for efficient cross-session message review.
* fix: pass BridgeProvider to FSClient and store full heartbeat prompt
FSClient was always created with nil provider, causing all container
file reads (IDENTITY.md, SOUL.md, MEMORY.md, HEARTBEAT.md, etc.) to
silently return empty strings. Expose Agent.BridgeProvider() and wire
it into Resolver. Also fix heartbeat trigger to store the full prompt
template as the user message instead of the literal "heartbeat" string.
* feat: add line numbers to container file read output
Move line-number formatting from the bridge gRPC server to the agent
tool layer so that the raw content stored and transmitted via gRPC
remains clean, while the read_file tool output includes numbered lines
for easier reference by the agent.
* chore(deps): update twilight-ai to v0.3.2
* fix: lint, test
|
||
|
Ringo.Typowriter
|
ad08f335eb | feat(agent): restore read_media in pure Go (#257) | ||
|
Acbox Liu
|
1680316c7f |
refactor(agent): remove agent gateway instead of twilight sdk (#264)
* refactor(agent): replace TypeScript agent gateway with in-process Go agent using twilight-ai SDK
- Remove apps/agent (Bun/Elysia gateway), packages/agent (@memoh/agent),
internal/bun runtime manager, and all embedded agent/bun assets
- Add internal/agent package powered by twilight-ai SDK for LLM calls,
tool execution, streaming, sential logic, tag extraction, and prompts
- Integrate ToolGatewayService in-process for both built-in and user MCP
tools, eliminating HTTP round-trips to the old gateway
- Update resolver to convert between sdk.Message and ModelMessage at the
boundary (resolver_messages.go), keeping agent package free of
persistence concerns
- Prepend user message before storeRound since SDK only returns output
messages (assistant + tool)
- Clean up all Docker configs, TOML configs, nginx proxy, Dockerfile.agent,
and Go config structs related to the removed agent gateway
- Update cmd/agent and cmd/memoh entry points with setter-based
ToolGateway injection to avoid FX dependency cycles
* fix(web): move form declaration before computed properties that reference it
The `form` reactive object was declared after computed properties like
`selectedMemoryProvider` and `isSelectedMemoryProviderPersisted` that
reference it, causing a TDZ ReferenceError during setup.
* fix: prevent UTF-8 character corruption in streaming text output
StreamTagExtractor.Push() used byte-level string slicing to hold back
buffer tails for tag detection, which could split multi-byte UTF-8
characters. After json.Marshal replaced invalid bytes with U+FFFD,
the corruption became permanent — causing garbled CJK characters (�)
in agent responses.
Add safeUTF8SplitIndex() to back up split points to valid character
boundaries. Also fix byte-level truncation in command/formatter.go
and command/fs.go to use rune-aware slicing.
* fix: add agent error logging and fix Gemini tool schema validation
- Log agent stream errors in both SSE and WebSocket paths with bot/model context
- Fix send tool `attachments` parameter: empty `items` schema rejected by
Google Gemini API (INVALID_ARGUMENT), now specifies `{"type": "string"}`
- Upgrade twilight-ai to d898f0b (includes raw body in API error messages)
* chore(ci): remove agent gateway from Docker build and release pipelines
Agent gateway has been replaced by in-process Go agent; remove the
obsolete Docker image matrix entry, Bun/UPX CI steps, and agent-binary
build logic from the release script.
* fix: preserve attachment filename, metadata, and container path through persistence
- Add `name` column to `bot_history_message_assets` (migration 0034) to
persist original filenames across page refreshes.
- Add `metadata` JSONB column (migration 0035) to store source_path,
source_url, and other context alongside each asset.
- Update SQL queries, sqlc-generated code, and all Go types (MessageAsset,
AssetRef, OutboundAssetRef, FileAttachment) to carry name and metadata
through the full lifecycle.
- Extract filenames from path/URL in AttachmentsResolver before clearing
raw paths; enrich streaming event metadata with name, source_path, and
source_url in both the WebSocket and channel inbound ingestion paths.
- Implement `LinkAssets` on message service and `LinkOutboundAssets` on
flow resolver so WebSocket-streamed bot attachments are persisted to the
correct assistant message after streaming completes.
- Frontend: update MessageAsset type with metadata field, pass metadata
through to attachment items, and reorder attachment-block.vue template
so container files (identified by metadata.source_path) open in the
sidebar file manager instead of triggering a download.
* refactor(agent): decouple built-in tools from MCP, load via ToolProvider interface
Migrate all 13 built-in tool providers from internal/mcp/providers/ to
internal/agent/tools/ using the twilight-ai sdk.Tool structure. The agent
now loads tools through a ToolProvider interface instead of the MCP
ToolGatewayService, which is simplified to only manage external federation
sources. This enables selective tool loading and removes the coupling
between business tools and the MCP protocol layer.
* refactor(flow): split monolithic resolver.go into focused modules
Break the 1959-line resolver.go into 12 files organized by concern:
- resolver.go: core orchestration (Resolver struct, resolve, Chat, prepareRunConfig)
- resolver_stream.go: streaming (StreamChat, StreamChatWS, tryStoreStream)
- resolver_trigger.go: schedule/heartbeat triggers
- resolver_attachments.go: attachment routing, inlining, encoding
- resolver_history.go: message loading, deduplication, token trimming
- resolver_store.go: persistence (storeRound, storeMessages, asset linking)
- resolver_memory.go: memory provider integration
- resolver_model_selection.go: model selection and candidate matching
- resolver_identity.go: display name and channel identity resolution
- resolver_settings.go: bot settings, loop detection, inbox
- user_header.go: YAML front-matter formatting
- resolver_util.go: shared utilities (sanitize, normalize, dedup, UUID)
* fix(agent): enable Anthropic extended thinking by passing ReasoningConfig to provider
Anthropic's thinking requires WithThinking() at provider creation time,
unlike OpenAI which uses per-request ReasoningEffort. The config was
never wired through, so Claude models could not trigger thinking.
* refactor(agent): extract prompts into embedded markdown templates
Move inline prompt strings from prompt.go into separate .md files under
internal/agent/prompts/, using {{key}} placeholders and a simple render
engine. Remove obsolete SystemPromptParams fields (Language,
MaxContextLoadTime, Channels, CurrentChannel) and their call-site usage.
* fix: lint
|