- Add Clear(ctx, sessionKey) to ContextManager interface
- Implement Clear for legacy (JSONL) and seahorse (DB + JSONL)
- Add Engine.ClearSession + Store.ClearConversation
- Fix FTS5 DELETE trigger syntax in schema (was using wrong
external-content FTS5 syntax; now uses standard DELETE FROM)
- Fix ClearSession to skip sessions never ingested (was creating
blank conversations record via GetOrCreateConversation)
- Simplify summary_parents DELETE into single OR statement
- Add TestStoreClearConversation unit test
When the message tool sent to a different chat (e.g., a group), the
agent's final response to the originating chat was incorrectly skipped
because HasSentInRound() was a simple bool that didn't distinguish
targets. Replace with HasSentTo(channel, chatID) that tracks all
send targets per round and only suppresses when the target matches.
Fixes cross-conversation message causing "Processing..." to hang.
* feat(hooks): add respond action for tool execution bypass
Add a new HookActionRespond that allows hooks to return tool results directly, skipping actual tool execution. This enables plugin tool injection, caching, and mocking capabilities.
- Add HookActionRespond constant and support in HookManager
- Extend ToolCallHookRequest with HookResult field
- Implement respond action handling in process hooks and agent loop
- Add comprehensive tests for respond and deny_tool actions
- Update documentation with hook actions table and examples
* docs(hooks): add JSON-RPC protocol and plugin tool injection documentation
Add comprehensive documentation for hook JSON-RPC protocol and plugin tool injection capabilities:
- Add "Hook Actions" section to README.zh.md explaining respond action for tool execution bypass
- Create hook-json-protocol.md/.zh.md detailing JSON-RPC 2.0 protocol for all hook methods
- Create plugin-tool-injection.md/.zh.md with complete examples for external tool implementation
- Document how hooks can inject tool definitions and return results via respond action
- Include Python and Go examples for weather query plugin implementation
* feat(agent): emit tool events and feedback for hook results
Add ToolExecStart event emission and tool feedback for hook results to ensure consistent behavior between normal tool execution and hook bypass scenarios. This maintains parity in event tracking and user feedback when tools are executed via hooks.
* style(agent): format whitespace in hook structs and constants
Remove trailing whitespace and standardize spacing in JSON struct tags, constants, and test data for improved code consistency.
* feat(hooks): add media support for plugin tool injection
Extend the hook respond action to support media file handling:
- Add `media` field for returning images and files from hooks
- Add `response_handled` field to control turn completion behavior
- When response_handled=true, media is automatically delivered to user
- When response_handled=false, media is passed to LLM for vision requests
This enables plugins to directly return generated images, downloaded
files, and other media content either to users or for LLM analysis.
* docs(hooks): document security implications of respond action
Add security boundary documentation explaining that the respond action
bypasses ApproveTool checks, allowing hooks to return results for any
tool without approval. Include recommendations for secure hook
implementation and code comments marking the security considerations.
Changes:
- Add "Security Boundaries" section to plugin-tool-injection docs
- Document bypass of approval checks and associated risks
- Provide security recommendations and example code
- Add inline security comments in hooks.go and loop.go
* refactor(agent): improve completeness of tool result cloning and hook processing
Extend cloneToolResult to properly copy ArtifactTags and Messages fields,
ensuring deep copies of all ToolResult data. Consolidate event emission
and user message handling to match the normal tool execution flow.
* fix(agent): align hook respond path with normal tool execution flow
The hook respond code path was missing several critical behaviors that
existed in normal tool execution:
- Add logging for tool calls with arguments preview
- Add is_tool_call metadata to user-facing messages
- Handle attachment delivery failures by setting error state and
notifying LLM
- Set ResponseHandled=false when using bus for media delivery
- Check for steering messages and graceful interrupts after tool
execution, skipping remaining tools when appropriate
- Poll for SubTurn results that arrived during tool execution
This ensures consistent behavior between hook-responded tool calls and
normally executed tool calls.
* test(agent): add tests for hook respond media error handling
Add comprehensive tests for the hook respond code path when media
delivery fails. Tests cover error media channel scenarios and verify
proper error state handling.
Also document that AfterTool is not called when using respond action,
as it provides the final answer directly (design decision).
* fix: use per-candidate provider for model_fallbacks
Each fallback model now uses its own api_base and api_key from
model_list instead of inheriting the primary model's provider config.
Previously, a single LLMProvider was created from the primary model's
ModelConfig and reused for all fallback candidates — only the model ID
string was swapped. This caused all fallback requests to be routed to
the primary provider's endpoint, making cross-provider fallback chains
non-functional (e.g., OpenRouter primary with Gemini fallback would
send the Gemini request to OpenRouter's API).
Fix: pre-create a per-candidate LLMProvider at agent initialization
time by looking up each candidate's ModelConfig from model_list. The
fallback run closure now selects the correct provider per candidate
via CandidateProviders map, falling back to agent.Provider when no
override is found.
Fixes#2140
Made-with: Cursor
test: add test for instance.go
fix: fix test
refactor: optimize
fix: fix Golang lint issues
chore: comment cleanup
* refactor: use resolvedModelConfig() instead of buildModelIndex()
* fix
* feat(seahorse): implement short-term memory engine of seahorse
Add pkg/seahorse/ module implementing a SQLite-backed DAG-based summary
hierarchy for context management, ported from lossless-claw's LCM design:
- types.go + short_constants.go: core types (Message, Summary, Conversation,
ContextItem) and configuration constants (fanout, token targets, thresholds)
- migration.go: idempotent DB schema with FTS5 trigram tokenizer for CJK
- store.go: full SQLite CRUD (conversations, messages, summaries DAG,
context_items with ordinal gap numbering, FTS5 search)
- short_engine.go: Engine lifecycle (NewEngine, Ingest, Assemble, Compact),
session pattern filtering (ignore/stateless glob→regex compilation),
per-session mutex via sync.Map
- short_assembler.go: budget-aware context assembly with fresh tail protection
(32 messages), oldest-first eviction, summary XML formatting, RebuildContextItems
- short_compaction.go: leaf compaction (messages→summary) and condensed
compaction (summaries→higher-level summary), 3-level LLM escalation,
CompactUntilUnder for emergency overflow
- short_retrieval.go: lookupByID, FTS5/LIKE search, recursive expand with
token cap
- context_seahorse.go: agent.ContextManager adapter, registered as "seahorse",
provider↔seahorse message type conversion (ToolCalls, tool_result)
* fix(seahorse): correct 3 adapter bugs in context management
- TokenCount: use full message (Content+ToolCalls+Media) instead of Content-only
- Empty Content: rebuild Content from tool_result Parts when stored empty
- Duplicate summaries: summaries only in Summary field, not in History messages
- Grep: fix SearchResult.Snippet→Content for summaries
- Schema: fix FTS5 SQL uses VIRTUAL TABLE not TEMP TABLE
- TestFTS5SQLConstants: verify FTS5 SQL syntax correctness
- Test: fix flaky TestCompactLeaf
* fix(agent): ingest steering messages into seahorse SQLite
Steering messages were only persisted to session JSONL but not ingested
into seahorse SQLite, causing them to be missing from context assembly.
Added `ts.ingestMessage(turnCtx, al, pm)` call in the steering message
injection block alongside the existing JSONL persistence.
Test: TestSeahorseSteeringMessageIngested verifies steering messages
appear in seahorse SQLite DB after being processed.
* fix(seahorse): address 3 blocking bugs from code review
- Fix resequenceContextItemsTx scan error handling (store.go:850)
Changed `return err` to `return scanErr` to properly propagate scan errors
instead of returning nil (which silently corrupts data)
- Fix sql.NullString for INTEGER column (store.go:847)
Changed `mid` from sql.NullString to sql.NullInt64 since message_id
is INTEGER in schema. Removed unnecessary strconv.ParseInt call.
- Fix compactCondensed fallback deleting non-candidate items
Added ReplaceContextItemsWithSummary method for per-item deletion
when candidates are not contiguous in ordinal space.
Optimized to use range deletion when candidates are consecutive.
* fix(seahorse): pass Budget to Compact for correct condensed threshold
Issue #4 from PR review: When Budget was not passed to seahorse.Compact,
it defaulted to `tokensBefore * 0.75`, making `tokensBefore > budget`
always true and causing condensed compaction to trigger unnecessarily.
Changes:
- context_seahorse.go: Forward Budget from CompactRequest to CompactInput
- loop.go: Pass Budget (ContextWindow) in all 3 Compact calls
- Add test verifying condensed is skipped when tokens < threshold
- Fix lint issues in store.go and store_test.go
* fix(seahorse): add mutex for assembler lazy initialization
Issue #5 from PR review: The check-then-create pattern for e.assembler
was a data race when multiple goroutines called Assemble() concurrently:
if e.assembler == nil {
e.assembler = &Assembler{...}
}
Changes:
- Add assemblerMu sync.Mutex to Engine struct
- Add initAssemblerOnce() using double-checked locking (same pattern as initCompactionOnce)
- Add TestAssemblerLazyInitRace to verify thread-safety
* fix(seahorse): handle non-consecutive depths in selectShallowestCondensationCandidate
Issue #8 from PR review: the loop iterated depth 0, 1, 2... assuming
consecutive keys, but break when key was missing caused deeper depths
to never be checked.
Fix: collect all existing depth keys, sort, then iterate in order.
* fix(seahorse): wrap DeleteMessagesAfterID and appendContextItems in transactions
- DeleteMessagesAfterID: wrap all DELETE operations in a transaction for
atomicity, remove redundant manual FTS delete (handled by trigger)
- appendContextItems: use transaction to fix read-then-write race condition
- Add GetMaxOrdinalTx and resolveItemTokenCountTx for transaction-scoped queries
- Remove unused resolveItemTokenCount function
Fixes PR review issues 6 and 7.
* fix(seahorse): derive readable content from Parts and cap CompactUntilUnder iterations
- Derive readable content from MessageParts in AddMessageWithParts so
FTS5 indexing and summary formatting can access tool call information
- formatMessagesForSummary and truncateSummary now fall back to Parts
when Content is empty, fixing blank summaries for Part-based messages
- Add MaxCompactIterations (20) to prevent CompactUntilUnder infinite
loops; exceeded iterations are logged as warnings
* feat: add load_image tool for local file vision
* fix: address load_image PR review feedback
- Exclude load_image from sub-agent tools via Unregister after Clone,
since RunToolLoop does not call resolveMediaRefs
- Add ToolRegistry.Unregister() method
- Fix scope collision: use channel:chatID instead of filename
- Add channel/chatID context resolution matching send_file pattern
- Add comment explaining iteration > 1 guard on resolveMediaRefs
- Remove emoji from ForUser for consistency with send_file
- Add load_image_test.go
* feat: enable load_image for subagents via MediaResolver in RunToolLoop
Instead of removing load_image from sub-agent tools (28f69e71), inject a
MediaResolver into the legacy RunToolLoop fallback path so media:// refs
are resolved to base64 before each LLM call — matching the main agent
loop behavior.
- Add MediaResolver field to ToolLoopConfig and call it on iteration > 1
- Add SubagentManager.SetMediaResolver() and wire it through runTask
- Remove ToolRegistry.Unregister() (no longer needed)
- Restore load_image in sub-agent tool set (revert Clone+Unregister)
- Add TestSubagentManager_SetMediaResolver_StoresResolver
* refactor(load_image): remove prompt parameter from tool schema
* test(tools): add success-path test for LoadImageTool
Add TestLoadImage_SuccessPath that creates a real PNG file with valid
magic bytes, calls Execute with WithToolContext, and verifies:
- result.IsError == false
- ToolResult.Media contains a media:// ref
- ToolResult.ForLLM contains the [image: marker
- media ref is resolvable in the store
Add explanatory comment in loop.go for why Media and ArtifactTags
coexist on non-ResponseHandled tool results (e.g. load_image).
* fix: preallocate slice in tests and add ResponseHandled guard in toolloop
Fix prealloc linter failure in load_image_test.go.
Prevent double-resolving media by checking ResponseHandled in toolloop.go.
* Register TTS tool if provider is available
---------
Co-authored-by: Reusu <admin@yumao.name>
Co-authored-by: 美電球 <hoshina@evaz.org>
- Add `reaction` tool that reacts to a message (defaults to current inbound message via context)
- Extend `message` tool with optional `reply_to_message_id` parameter
- Introduce `WithToolInboundContext` to inject inbound message IDs into tool execution context
- Surface `MessageID` and `ReplyToMessageID` in `processOptions` for tool-surface consumption
Refs #2137
When the message tool sent to a different chat (e.g., a group), the
agent's final response to the originating chat was incorrectly skipped
because HasSentInRound() was a simple bool that didn't distinguish
targets. Replace with HasSentTo(channel, chatID) that tracks all
send targets per round and only suppresses when the target matches.
Fixes cross-conversation message causing "Processing..." to hang.
* fix(cron): publish agent response to outbound bus for cron-triggered jobs
When a cron job triggers agent execution via ProcessDirectWithChannel,
the agent response was silently discarded — the code assumed AgentLoop
would auto-publish it, but SendResponse is false on this path.
Delegate to PublishResponseIfNeeded (exported from AgentLoop) so the
response reaches the originating channel (e.g. Telegram) only when the
message tool did not already deliver content in the same round.
Also adds a "directive" message type to CronPayload, allowing cron jobs
to instruct the agent to execute a task rather than echo static text.
* fix(cron): add type validation and directive test coverage
Address reviewer blocking feedback:
1. Server-side whitelist for `type` parameter — the `enum` in
Parameters() is only an LLM schema hint; any string was persisted.
Now `addJob` rejects values other than "message" and "directive".
2. Comprehensive test coverage for the directive code path:
- directive adds prompt prefix to ProcessDirectWithChannel
- deliver=true + directive routes through agent (not direct publish)
- directive prompt content, sessionKey, channel, chatID are correct
- invalid type is rejected; valid types ("", "message", "directive") pass
- deliver=true message type goes directly to bus (regression)
- agent error path does not trigger publish (regression)
Also merge the two UpdateJob calls in addJob into one to avoid
redundant disk I/O (non-blocking suggestion from review).
* fix(cron): remove omitempty from CronPayload.Type for consistent JSON
Empty string and "message" are semantically equivalent defaults;
always serializing the field avoids asymmetric JSON output.
* test(cron): remove redundant test, strengthen error path coverage
- Remove ExecuteJobDirectivePassesCorrectContent: its assertions on
sessionKey/channel/chatID duplicate ExecuteJobPublishesAgentResponse;
its prompt check duplicates DirectiveAddsPromptPrefix.
- Strengthen DirectiveAddsPromptPrefix with exact prompt match and
publish response assertion.
- Fix ReturnsErrorWithoutPublish: set non-empty stub response so the
test verifies the error branch early-return, not the response==""
guard.
* fix(ci): satisfy golines and gosmopolitan in cron code
Cytown