Replace the blocking Chat() call in process() with
chatStreamWithFallback: ChatStream first, accumulate chunks, fall back
to non-stream Chat on connect failure or empty-stream failure.
Why: the non-streaming path blocked for the ENTIRE LLM generation (up
to the 180s HTTP timeout). Reasoning models thinking 60-120s plus AUTO
chain failover regularly exceeded it -> context canceled -> full turn
wasted. With streaming the first chunk arrives in ~1-3s and any
flowing token keeps the connection alive; total generation time is no
longer bounded by an overall timeout.
Compatibility (external behavior unchanged):
- process() signature/return values unchanged
- Aggregated events (EventReasoning / EventAgentLLMChain) still fire
once per turn with full text after stream completion - existing
plugin subscribers see identical payloads as before
- New incremental events EventReasoningDelta / EventContentDelta are
additive; old subscribers ignore unknown event types
- Tool execution loop, memory pipeline, stage pipeline untouched
Streaming details:
- Tool call fragments accumulated per OpenAI streaming convention:
id/name arrive on the first fragment, arguments as raw JSON string
shards across fragments; merged and parsed once at stream end
- normalizeStreamToolCalls keeps nameless argument shards (the
non-stream normalizer drops them); ToolCall gains RawArguments to
carry shard text
- Interrupt mid-stream returns partial content instead of discarding
the whole generation
Verified end-to-end against llmsproxy: plain chat streams correctly;
curl confirms tool-call shard wire format ({" + command" + :"date"}
-> {"command":"date"}); unit tests cover shard merging and
content/reasoning accumulation.
Rewrite LuaAdaptedProvider.ChatStream to match the maturity of
llmsproxy's streaming implementation:
HTTP layer:
- Dedicated stream HTTP client with no overall timeout (SSE must not
be cut by the 180s Chat timeout); only a 30s dial timeout
- Uses applyAdapterHeaders (supports build_headers dynamic signing
hook), matching the non-streaming Chat path
Non-200 response handling:
- New TransformError Lua hook (adapter.transform_error) for per-source
protocol knowledge in error messages
- Safe fallback truncation of raw error bodies (prevents HTML dump
leakage to clients)
SSE parsing enhancements:
- parseOpenAICompatibleStreamChunkFull: handles token usage in the
final chunk (prompt_tokens/prompt, total_tokens/total dual keys),
prompt cache detail fields, and empty-string finish_reason filtering
(sensenova sends "" on every chunk)
- Replaced old SSEScanner with bufio.Scanner (larger buffer, fewer
allocations)
Stream integrity:
- errorOnlyChunk detection: holds back the first chunk to reject
degenerate streams (e.g. zen free pool's finish_reason:"network_error"
with empty content) before any byte reaches the caller
- [DONE] dedup: adapters that already emit a terminating done chunk
with the real finish_reason don't get a second reason-less done
- Clean EOF sends a final Done:true if no done was seen
Struct changes:
- StreamChunk: added FinishReason and Usage fields for callers
- LuaAdaptedProvider: added streamClient (lazy) + streamMu
Tested: curl against llmsproxy SSE confirms reasoning_content parsing
is correct (delta.reasoning_content), usage chunk handling works, and
[DONE] termination is properly emitted.
The selfInputCh previously treated ALL internal messages as memory
consolidation tasks (hardcoded _consolidation_ output channel), which
silently discarded child-agent completion notifications:
- processConsolidation never appends to conversation context, so the
parent agent could not see that its child had finished
- it also discards the LLM response without emitting to any output
channel, so nothing reached the user
- net effect: notifications vanished; parent never called child_result
Restore the intended design: each self-input message now carries a
target output channel. Only consolidation tasks (_consolidation_) go
through the no-memory path (no context write, no emit). Child
notifications carry the parent's original output channel and are
processed as normal input: appended to context, LLM sees them and can
call child_result, and the response is emitted back to the user.
Changes:
- new selfInputMsg{text, channel} type + channelConsolidation const
- selfInputCh: chan string -> chan selfInputMsg
- injectSelf (consolidation) keeps _consolidation_; new
injectSelfChannel for flagged messages
- handleSelfInput routes on msg.channel instead of hardcoding
- executeSpawnChild captures a.currentOutputChannel and passes it to
runChildTask so the notification returns to the originating channel
(falls back to "cli" when unset or consolidation)
- executeChildResultTool: remove dead double-lock/re-check block
Verified end-to-end with tmux PTY against llmsproxy:
spawn_child -> child done -> notification processed via normal path
(log shows 'input from system -> response, tools=[child_result]'),
parent agent retrieved the child result successfully.
The local llmsproxy AUTO chain tries 6+ slots across 3 tiers sequentially.
Each failed tier incurs busyWait (2s) + upstream timeout, so a full chain
exhaustion can exceed 120s. The llmsproxy logs showed 143 'context canceled'
errors for the homeagent key — the client gave up before the chain finished.
180s gives the chain enough room to complete before the client timeout fires.
Also remove stale backup files under /usr/local/bin/.
- C ABI invoke_stage 增加 result 输出参数，外部插件 stage 回调可将修改后的
StageContext(RawMessage/LLMText/FinalText/Response/ToolResults) 写回内核
- ABI 标识版本改为字符串 semver 与核心 Version 对齐(ABIVersion="0.9.0")，
C 层协商用派生整数 CABINum=900(major*100+minor)，不再使用独立数字编码
- version_min 保证 v0.8.x(800) 旧插件向后兼容可加载
- 修复工具循环 zen 兼容补位误伤首轮 system 上下文(仅尾部为 assistant/tool 时补位)
- 更新 README 项目状态说明
- Drop skill.NewManager from homed bootstrap; skills dir no longer core-managed
- Remove GetInjectedPrompt system-prompt injection (skills are not first-class)
- Delete internal/skill package, SkillAPI, webui /api/v1/skills, status skills block
- ConfigRegistry: plugin config tables now created only via RegisterDef; arbitrary
scope Set/Get no longer implicitly creates config_<name> tables (fixes stray
config_today_task table from SKILL directory name being used as a scope)
- Six-phase plan complete: webui/cli/healthcheck/pluginmgr/clawhubadapter
now interact with the kernel exclusively via internal/sdk interfaces;
all Configure() calls and package-level global injection removed
- buildSDK in internal/plugin/registry.go is the single assembly point
- Add internal/sdk/events.go exporting event types/constants
- Fix ProviderManager cooldown sharing: LuaAdaptedProvider.Name() now
returns the source name instead of lua_<adapter>, so multiple sources
sharing an adapter (single script load via shared VM AdapterCache) no
longer share failure-cooldown state
- Verified: build/vet/tests green, deployed to homeagent.service with
full plugin capability testing via local OpenAI-compatible mock
- process.go: attach resp.ReasoningContent when building assistant messages
- deepseek.lua v2.1.0: remove last_reasoning closure hack; rely on
core-provided reasoning_content in messages
- openai.lua: strip reasoning_content from messages in transform_request
(not supported by OpenAI API)
- process.go: only emit resp.Content on the first tool call per batch,
subsequent assistant messages use empty content (serialized as null)
- provider.go: MarshalJSON outputs null content when empty with tool_calls
to comply with DeepSeek/OpenAI expected format
- output.go: update parameter interface (payload/type/meta) to match
tool definitions (uncommitted from previous refactor)
- Redesign output_send__ tools: content JSON string -> structured
payload/meta/type params for LLM reliability
- executeOutputSendTool: route by type with capability check
- executeOutputSendHelp: show meta format + type enum
- Updated system prompt rules for new interface
- docToTriples: use jieba exact mode adjacent co-occurrence
- Unify vector space: Doc.Vector field, ContextToDoc vectorizer,
ReindexWithVectorizer on startup
- StaticEmbedder: pre-trained ConceptNet Numberbatch/fastText word embeddings,
auto-download with TF-IDF fallback, comma-separated multi-model paths
- CleanTemplateText: regex stripping of QQ tool call templates and noise
- textForVector: per-source vector strategy (agent→Response, user→Input,
cold_storage→both)
- Indexer.BuildContext and ExtractKeywords now clean input before vectorization
- Protect recent 10 events in Prune (regression fix: use local var not const)
- Output channels generate per-channel tools: output_send__{name} (type=output) + output_send__{name}_help
- content is JSON string transparently passed to plugin handler for routing
- EventAgentLLMChain: full LLM response forwarded after each turn for webui/logs
- sdk.New refactored to SDKConfig struct (no more 13 positional args)
- RegisterOutputChannel adds desc param for JSON format documentation
- channelDevice simplified (no Tools method), desc field added
- Child agent permission updated for output_send__ prefix
- System prompt: output gates, multi-call, long messages split
- WebUI: subscribes to EventAgentLLMChain in SSE, no output channel
- Tests updated for new naming convention
Root node items (items without category) were skipped by depth>0 check,
causing knowledge_list to always return '(空)'. Also simplified formatTree
to avoid redundant child item rendering.
- Prune context BEFORE processing (LSTM forget gate pattern)
so LLM only sees relevant context, instead of pruning after the fact
- Fix ensureTrained() to recompute all event vectors after retraining
vectorizer, fixing feature-space mismatch between stored vectors and
query vector that made relevance scoring effectively random
- Add read lock to knowledge BuildTree() (data race fix)
- Log writeIndex() errors instead of discarding them
- Fix TOCTOU race in document ContextToDoc() dedup
- Fix healthcheck timing (measure elapsed before cleanup)
- Refactor waiter CLI into separate files (state, conn, config, editor,
history, builtin) for maintainability
- Add CLI plugin API key authentication
- annotate queued and interrupt inputs with source/output channel context
- drain all pending interrupts instead of only one
- avoid double-processing by routing interrupts exclusively:
active LLM -> intercept channel, idle -> input queue
- keep interrupt source/channel in payload and system context
