feat: implement EDA long-term memory plugin and upgrade plugin system infrastructure#1870
feat: implement EDA long-term memory plugin and upgrade plugin system infrastructure#1870xiaobin83 wants to merge 25 commits into
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1215d52ef2
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f554c1bdbe
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
update: I noticed there is runtimeContextProviders added in chat-orchestrator-runtime.ts 5 days ago, which can be used for my injectMemoryContext, I'll change the implementation. I was using onBeforeCompose (added just in the place of runtimeContextProviders). |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d67784f9fb
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2c59f1d806
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
xiaobin83
force-pushed
the
feat/openviking-memory-plugin
branch
from
2c59f1d to
85dfcbf
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 85dfcbf5a0
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ce5e8a78cd
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 13a96c1d8b
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
xiaobin83
force-pushed
the
feat/openviking-memory-plugin
branch
from
e367d49 to
f046694
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f046694358
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
xiaobin83
force-pushed
the
feat/openviking-memory-plugin
branch
from
85ed762 to
d2b515e
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d2b515e307
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 93d7f73985
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 34c9e9ae56
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
xiaobin83
force-pushed
the
feat/openviking-memory-plugin
branch
from
162f5eb to
5ba6a9e
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5ba6a9e5be
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
@nekomeowww hello, could you check this implementation on host side when you have time? i noticed you are changing the plugin SDK in another branch. For this memory plugin, it needs two tool calls from host to plugin side,
The other tools are called by agent itself though xsai (memory_search, memory_save, memory_delete). So it would be nice that the new SDK could still have those two tool calls, then it wont need much change for this tool. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 608ad184c5
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…onfig system, context hooks, and example plugin - stage-tamagotchi: add PluginHost config system with schema validation, Eventa IPC, settings UI, and i18n - plugin-sdk: extend plugin-tools with injectMemoryContext and post-processing hooks - stage-ui: add plugin context provider registration for chat orchestrator - stage-tamagotchi: redesign plugin settings config UX with dirty tracking - stage-tamagotchi: fix plugin tool timing and config apply workaround - openviking-memory: add example plugin with memory CRUD via PluginHost SDK, reorganize source, and add build pipeline - docs: update plugin host design documents
…ptions Add config?: ModuleConfigEnvelope to PluginStartOptions for passing updated config to init Fix PluginHost.init() to use options.config when calling applyConfiguration, instead of always passing an empty envelope Update loadPluginByName in stage-tamagotchi host to accept and forward config to host.start Let loadEnabledPlugins build a config envelope from the already-fetched config and pass it directly, avoiding a redundant getConfig() call Let setPluginConfig pass the updated config envelope via host.reload() options so reloaded sessions receive the latest config
Reset conversationSessionId via useChatMaintenanceStore $onAction when cleanupMessages fires, so plugin memory saves start a fresh session. - Fix save conversation not working across cleanup boundaries - Reimplement injectMemoryContext as runtime context provider
…rom chat session - Wrap listPlugins() calls in injectMemoryContext and onChatTurnComplete with try-catch to soft-fail on transient errors - Derive conversationSessionId from active chat session instead of a global ref; remove cleanupMessages workaround - Remove noisy console.info logs from plugin-tools and openviking
…te on plugin config reload Revoke prior asset sessions after host.reload triggered by setPluginConfig, preventing stale 30-day TTL asset URLs/cookies from accumulating. Clear loaded state (loaded set + loadedSessionIds) and revoke asset sessions when host.reload throws, preventing inconsistent host state where a plugin appears loaded but holds a stale session id. - clearModuleAssetSessionCacheByOwnerSessionId + revokeByOwnerSessionId called after successful host.reload for the old session - On reload failure, loaded/loadedSessionIds are cleaned up and old asset sessions revoked before re-throwing the error
…rt it Filter `memory_save_conversation` invocation to only plugins that actually register the tool, eliminating warning noise and unnecessary IPC work on every chat turn. - Extract 'memory_save_conversation' string into MEMORY_SAVE_CONVERSATION_TOOL constant - Cache plugin IDs that expose this tool during refresh() from xsai tool definitions - Filter loaded+enabled plugins in registerPostProcessingHook against the cache - Clear the cache in dispose()
…ect memory_search tool invocation Remove the intermediate queryContext Eventa IPC pathway and the associated contract definition file Replace indirect queryContext calls with direct invokePluginTool calls to the memory_search tool in the renderer store Add limit parameter support to searchMemories API with default value Split pluginsWithMemoryTool into separate pluginsWithMemorySaveTool and pluginsWithMemorySearchTool sets for clearer responsibility Clean up stale mocks and imports in tests
…posing sends The chat send path blocked on plugin context lookup because performSend awaits runtime context providers, and injectMemoryContext invoked invokePluginTool sequentially for each loaded plugin. When the OpenViking backend is down, memory_search could take ~30s+ per plugin due to retry+timeout, freezing user sends before any token streams. Changes: - Replace sequential for-loop with concurrent Promise.allSettled - Add AbortSignal.timeout(3_000) per plugin invoke to fail fast - Each slow/timed-out plugin is silently skipped with a warning - Maximum total wait is bounded to 3s regardless of plugin count
- Add required field validation: reject missing or empty required fields - Add value type validation: enforce string/secret/number/boolean per schema - Preserve existing unknown key rejection with stricter type assertion - Add comprehensive unit test suite (13 test cases covering all validation paths)
…d cleanup in setPluginConfig - Split single try/catch into two: one for host.reload() failures, another for post-reload asset cookie revocation - Log cleanup errors instead of clearing host bookkeeping when reload succeeded but asset revocation fails - Prevents host bookkeeping inconsistency where a running session is marked as unloaded due to non-fatal cleanup errors
…atTurnComplete to prevent misattribution - Change ensureConversationSessionId signature to require (sessionId, generation) parameters so callers must provide captured metadata - Move sessionId/generation capture before the await listPlugins() call in onChatTurnComplete callback to lock the originating session synchronously - This ensures that a concurrent session switch during streaming does not persist the completed turn under the wrong session ID in long-term memory
…ation Replace generic Error with TypeError in host plugin config validation for more precise error semantics when type mismatches occur. - Config field string type validation - Config field number type validation - Config field boolean type validation
… in plugin config validation
The required-field guard in setPluginConfig only rejected empty strings for type === 'string', but skipped the same check for type === 'secret'. This allowed a schema field marked { required: true, type: 'secret' } to be saved as '', persisted, and reloaded as an unusable credential.
- Extended the empty-value check to cover both 'string' and 'secret' types
- Made the error message dynamic to reflect the actual field type
Eliminate a race condition in plugin-tools where a concurrent session switch during streaming could misattribute a completed turn to the wrong session. The sessionId and generation are now carried inside ChatStreamEventContext instead of read from the session store at async handler time. - Add sessionId and generation fields to ChatStreamEventContext so downstream consumers identify the originating session without external store state - Persist generation counter in ChatSessionMeta so it survives reloads; ensureGeneration and loadIndexForUser restore from persisted value - Remove useChatSessionStore dependency from plugin-tools store; read sessionId and generation from the event context instead
…er inputs in plugin settings Preserve unsaved config edits when plugin list refreshes instead of resetting all editors; merge existing editor states and only add/remove entries for new/stale plugins Reject invalid number input in saveConfig with an inline error message instead of silently omitting the value from the saved config Reload only the saved plugin's config after savePluginConfig instead of reloading all configs, preventing the refresh from discarding unsaved edits in other plugin forms Remove now-unused loadAllConfigs helper
applyRemoteSnapshot previously kept the local generation when already initialized, causing follower windows to retain stale generation values after another window resets a chat. Since memory session IDs include the generation, subsequent turns in the follower would be persisted under the pre-reset namespace, mixing old and new memory history. - Remove fallback to sessionGenerations.value[sessionId] when applying remote snapshot - Always trust snapshot.sessionMetas[sessionId]?.generation for each session in the incoming snapshot
The refactor-remote-plugins-to-plugin-host planning document has served its purpose — M1-M3 are completed and the core migration is done. Removing it to keep the .trae/documents directory clean.
…commit pattern
Redesign saveConversation to follow OpenViking's add-accumulate-commit pattern instead of blind message posting.
- Add SessionInfo interface and addSessionMessage/getSession/commitSession methods
- Add commitTokenThreshold (20k) and commitKeepRecentCount (10) config options
- Add committed flag to ConversationSaveResult
- Fix getSession to correctly parse { status, result, error } response
…urn, translate docs to English Add recallMemories method to OpenVikingClient (delegates to searchMemories) Register memory_recall tool for system-triggered context injection Rename memory_save_conversation tool to memory_save_turn Rename saveConversation to saveTurn in client interface and implementation Translate all tool titles, descriptions, and parameter docs from Chinese to English Mark system-only tools (memory_recall, memory_save_turn) with empty activation keywords Update plugin-tools.ts constants to reference memory_recall and memory_save_turn Update Readme with memory_recall, memory_save_turn, session memory processing rules, and system-triggered tool usage notes
…rategy Replace simple recallMemories delegation with a full sorting pipeline for context injection: query intent analysis, multi-boost scoring, dedup, and leaf-first selection. - Add ranking.ts module with FindResultItem/RecallQueryProfile types - Implement buildRecallQueryProfile for token extraction and intent detection - Implement rankForInjection with leafBoost, eventBoost, preferenceBoost, and lexicalOverlapBoost scoring components - Implement pickMemoriesForInjection entry: rank to sort to dedup to leaf-first - Update recallMemories to fetch 3x candidates and apply ranking pipeline - Add recallCandidateMultiplier property to OpenVikingClient interface
…on analysis Remove the 553-line `openclaw-plugin-implementation-analysis.md` document. This analysis is being superseded by the new rank-for-injection design documents.
- Return an empty ReplaceSelf context message instead of undefined when injectMemoryContext gathers no results from any plugin - This ensures ingestRuntimeContexts replaces the plugin-memory bucket with an empty entry, preventing formatContextPromptText from appending stale memories on subsequent sends
xiaobin83
force-pushed
the
feat/openviking-memory-plugin
branch
from
608ad18 to
aac374a
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: aac374aba9
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| id: 'memory_recall', | ||
| title: 'Recall Memory', | ||
| description: 'Automatically recall long-term memories relevant to the current conversation for context injection. NOTICE: This tool is triggered by the system only and should not be invoked by the AI assistant.', |
There was a problem hiding this comment.
This tool is registered through the same apis.tools.register path as user-facing tools, and the host’s listSerializedXsaiTools() serializes every registered available tool while the renderer registers all of them as active XSAI tools. With the OpenViking plugin loaded, the model can therefore see and invoke this “system only” recall tool (and the memory_save_turn tool below), which can trigger unintended memory lookups/saves outside the controlled context-injection and post-turn hooks; keep these internal tools out of the public tool registry or gate them from XSAI serialization.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
it needs to add a parameter to tools.register() function to mark the tool as internal, then skip when serialize.
since the plugin host is changing, this issue wont be addressed for now.
…improve UX Persist setPluginConfig only after a successful host.reload; restore previous config in the catch path so a bad config is never committed Surface host-side validation and reload errors in the settings UI instead of only logging to console Fall back to item.overview when building injected memory context text if abstract is absent, matching the ranking fallback pattern Output openviking-memory plugin build to a named directory and package the archive with that directory, making it discoverable by loadManifestsFrom
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 799abeefe1
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…gin config envelope Previously buildPluginConfigEnvelope returned undefined when no user overrides existed, causing host.start() to receive an empty config at startup. Plugins that relied on schema defaults in their manifest would lose those values. Now always resolves the manifest entry, extracts schema defaults, and merges them with any persisted user overrides. This makes the loading path consistent with the settings UI (getPluginConfig), which already performed this merge. - Resolve manifest entry first, return undefined only if plugin not found - Collect default values from manifest.config.schema fields - Merge defaults with user overrides (user overrides take precedence)
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8962d1af4a
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…ry plugin is active Return createEmptyMemoryContext() instead of undefined when no memory plugin is loaded, so the context registry replaces the bucket with an empty context rather than leaving stale recalled memories in prompts. Extract createEmptyMemoryContext() helper to eliminate duplication of the empty ReplaceSelf context object.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 985e71db83
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| // date anchor + per-message [HH:MM] prefixes, which is more KV-cache | ||
| // friendly and less prone to weak models echoing timestamps verbatim. | ||
| ingestRuntimeContexts() | ||
| await ingestRuntimeContexts(sendingMessage) |
There was a problem hiding this comment.
When a runtime context provider is asynchronous, this await can finish after the originating chat session has been reset or switched generations, but the provider result is ingested before the stale-generation check below runs. In the memory-provider path this means a user can send a message, clear/reset the chat during the 3s recall window, and still leave recalled memory for the abandoned turn in the global context registry, so the next prompt in the reset conversation includes stale context. Check the queued send generation before mutating deps.context (or pass an abort/guard into context ingestion) so abandoned sends cannot write runtime context.
Useful? React with 👍 / 👎.
2 participants
Description
This PR adds the openviking-memory EDA (Event-Driven Architecture) long-term memory plugin, and supplements & optimizes the project-wide plugin system infrastructure.
Core changes are as follows:
stage-tamagotchi: Implemented PluginHost config system with schema validation, Eventa IPC, settings UI and i18n; redesigned plugin config UX with dirty tracking; fixed plugin tool timing and config application bugs.
plugin-sdk: Extended plugin-tools with memory context injection and post-processing hooks.
stage-ui: Registered plugin context provider for chat orchestrator to support plugin capability docking.
openviking-memory: Implemented complete memory CRUD capabilities based on PluginHost SDK, sorted out plugin code structure, and added supporting build pipeline.
Additional Context
Please focus on the main project code modifications (stage-tamagotchi / plugin-sdk / stage-ui) during review, including plugin host configuration logic, SDK hook extension, chat orchestrator context registration and related bug fixes, to verify main process stability and capability compatibility.
Known Issues
Coming soon
Screenshots
Notice, the first two screenshots were captured from development mode. The last one was captured from app.
Open a new chat,
Reset the chat,
Restart the app,
