Commit Graph

916 Commits

Author SHA1 Message Date
matevip
fcdb3fc15e fix(agent): break self-replicating 400, narration, args truncation, queue drop
A series of cross-cutting stability fixes that surfaced together
during a long debugging session.

reasoning_content / Claude prefill self-replicating 400:

- ChatController persists typed errors (content starts with '[错误] ')
  with status='error', so the failure text stops being re-sent as
  multi-turn context — DeepSeek thinking 400 ('reasoning_content
  must be passed back') and Claude 400 ('does not support assistant
  message prefill') used to recursively re-create themselves every
  retry by polluting history.
- BaseAgent.sanitizeForLlm filters status='error' / '[错误] ' prefix
  assistant messages from history before LLM dispatch.
- BaseAgent.fetchHistoryMessages defensively drops trailing
  AssistantMessages — Claude rejects assistant-tail prompts.
- NodeStreamingChatHelper.dropTrailingAssistant runs the same
  defense at every doStreamCall pre-egress, so the in-turn
  summarizing→reasoning transition (which leaves an assistant
  scaffold at the tail) doesn't trip Claude either.
- AgentGraphBuilder.FallbackPolicy.DEEPSEEK switched (null,true,true)
  → (' ',false,true), aligning with KIMI/OPENAI's tolerant ' '
  fallback. The previous 'force explicit 400' design was the
  self-replicating loop's prime mover.

narration + tool args truncation:

- ReasoningNode.DEFAULT_MAX_OUTPUT_TOKENS 4096 → 16384. The 4k cap
  was decapitating renderDocx tool_call args mid-stream when the
  model emitted a long content field on top of thinking content;
  the resulting 'invalid JSON' aborted execution silently.
- ReasoningNode appends a hermes-style TOOL_USE_ENFORCEMENT clause
  to every system prompt: 'when you say you will perform an action,
  call the tool now in the same response — narration is a protocol
  violation'. Treats 'now I will generate the docx' (and never
  actually calling renderDocx) as a forbidden pattern.
- ToolExecutionExecutor.normalizeToolExecutionError reframes the
  JSON-truncated error as actionable instructions: 're-call the
  same tool now with shorter content or split into multiple
  sequential calls; do NOT describe the result as text'.

side fixes from the same evening:

- ChatController doOnComplete skips completionPublisher.publish
  when isError=true, keeping memory extraction off the garbage path.
- ChatController doOnComplete queued-message guard simplified to
  'cr.queuedInput() != null', matching the other 4 sites in the
  controller. The previous 'isInterruptFollowup || !wasStopped'
  guard silently dropped queued messages when the user did
  Stop-then-Enqueue (wasStopped=true && interruptType=null), losing
  the freshly-typed follow-up message.
- prompts/graph/summarize-system.txt now distinguishes 'single
  task' (default; output one cohesive summary) from 'multiple
  independent sub-tasks' (use the子任务 N format). Stops the
  summarizer from inventing '子任务 1: PRO-027' decomposition for
  unitary requests like 'write me a project proposal'.
2026-04-27 07:51:01 +08:00
matevip
187197e804 fix(sse): preserve done event for late reconnect window 2026-04-27 07:50:35 +08:00
matevip
4a15027a98 feat(wiki): download original raw material file 2026-04-26 20:46:52 +08:00
matevip
0d78beb44f fix(wiki): batch-create per-slug retry — recover unparseable JSON, bump to 2 attempts 2026-04-26 20:37:09 +08:00
matevip
76956cf990 fix(tool): read_file falls back to chat-upload attachment by basename 2026-04-26 20:24:00 +08:00
matevip
30e7e67bb6 feat(wiki): LLM-narrated overview section with debounced regen + Recent Updates list + scaffold self-heal 2026-04-26 19:53:12 +08:00
matevip
fa5eaf83b4 fix(ui): code-block light-mode theme + cap header height at 38px 2026-04-26 19:09:10 +08:00
matevip
e2df16893e fix(wiki): smaller batch-create + resume button for partial generation 2026-04-26 18:42:27 +08:00
matevip
b7c911f01d feat(stt): DashScope realtime voice + language-aware routing + TalkMode polish
- DashScope paraformer-realtime-v2 WebSocket streaming
- Language-aware provider routing: Whisper for English, Paraformer for Chinese
- PCM WAV recording replaces WebM (provider filename bug + diagnostics)
- TalkMode push-to-talk fixes (audio drop, WS connecting race)
- Vite dev proxy WebSocket upgrade fix
- WebSocket binary buffer 8KB → 8MB (Tomcat default truncated voice clips)
- Audio chunk pacing at 100ms (DashScope returned 0 chars otherwise)
- Resolved language hint propagation + raw frame logging
- V46 seed idempotency fix on UI-toggled STT row
- Diagnostic cleanup after debugging session
2026-04-26 16:37:55 +08:00
matevip
4d7c6593c4 feat(minimax): expand video model catalog + add CN endpoint support 2026-04-26 08:34:35 +08:00
matevip
410c6c28cd feat(deepseek): integrate DeepSeek V4 (flash + pro) with thinking-mode support 2026-04-26 08:34:34 +08:00
matevip
dfb9fc2cac fix(model-catalog): claude-sonnet-4-7 doesn't exist — Sonnet stays at 4.6 2026-04-26 08:34:12 +08:00
matevip
b9c4f40028 refactor(anthropic): cleanup — deduplicate diagnostic statics, remove dead cache-options code 2026-04-26 08:34:12 +08:00
matevip
dbdb585eed fix(anthropic): rewrite system field to array to pass OAuth anti-abuse gate 2026-04-26 08:34:12 +08:00
matevip
5c2482c307 fix(anthropic): log outgoing request headers on 429 2026-04-26 08:34:12 +08:00
matevip
ed3ff54f0c fix(anthropic): drop (external, cli) UA suffix — it's the anti-abuse fingerprint 2026-04-26 08:34:11 +08:00
matevip
84cb442446 fix(anthropic): log anthropic-ratelimit-* headers on 429 2026-04-26 08:34:11 +08:00
matevip
aabf2b8c32 fix(anthropic): add anthropic-dangerous-direct-browser-access + accept headers 2026-04-26 08:34:11 +08:00
matevip
ae6467a5dc fix(anthropic): bidirectional mcp_ tool-name prefix on OAuth requests 2026-04-26 08:34:10 +08:00
matevip
44548e3010 fix(anthropic): inject Claude Code identity into system prompt 2026-04-26 08:34:10 +08:00
matevip
1d5bb58e9b fix(anthropic): allow ANTHROPIC_CLAUDE_CODE in StateGraph whitelist 2026-04-26 08:34:09 +08:00
matevip
fb4c013ad8 feat(anthropic): surface Claude Code OAuth in admin UI 2026-04-26 08:34:09 +08:00
matevip
bf4e81c554 i18n: add error messages for Claude Code OAuth failures 2026-04-26 08:34:08 +08:00
matevip
a7938b0e68 feat(anthropic): wire Claude Code OAuth into chat model 2026-04-26 08:34:08 +08:00
matevip
8539fb9407 feat(anthropic): Claude Code OAuth credential plumbing 2026-04-26 08:34:08 +08:00
matevip
9187aed273 fix(oauth): support remote-server deployment via MANUAL_PASTE flow 2026-04-26 08:34:08 +08:00
matevip
23a6d16778 feat(model-catalog): add Claude 4.7 + GPT-5.5 sampling-params handling 2026-04-26 08:32:45 +08:00
matevip
fac5ff2838 feat(image-gen): add gpt-image-2 to OpenAiImageProvider 2026-04-26 08:32:44 +08:00
matevip
3f10553186 fix(sse): distinguish stream_not_local vs completed on reconnect 2026-04-26 08:32:44 +08:00
matevip
0b55d5a227 feat(agent): Utf8SseEmitter + returnDirect end-to-end chain test 2026-04-26 08:32:44 +08:00
matevip
4a95e7dfe4 feat(tool): tool returnDirect and sensitive-data quarantine 2026-04-25 19:02:35 +08:00
matevip
c13d9b4c88 feat(wiki): expose method=tika short-circuit on extract_document_text 2026-04-25 19:02:35 +08:00
matevip
c752c1f2ae feat(wiki): Tika as last-resort document extractor 2026-04-25 19:02:34 +08:00
matevip
6474d0e6be feat(wiki): normalized relation boost + reason in search results 2026-04-25 19:02:34 +08:00
matevip
a1e40d6eae feat(wiki): enrich batch — N pages per LLM call 2026-04-25 19:02:34 +08:00
matevip
ca4c447250 feat(wiki): enrich prompt knows what's already linked 2026-04-25 19:02:34 +08:00
matevip
8ef523f953 feat(wiki): archived pages drawer in Wiki UI 2026-04-25 19:02:33 +08:00
matevip
51566a47a4 chore(wiki): externalize compile prompts; doc archive tools + admin endpoints 2026-04-25 19:02:33 +08:00
matevip
58b49f6e20 feat(wiki): structured no-evidence compile + ops admin endpoints 2026-04-25 19:02:33 +08:00
matevip
5206d65be7 fix(wiki): eager 0-pages = partial when chunks indexed; per-KB structured route; archived filter completion 2026-04-25 19:02:33 +08:00
matevip
3f25064ef9 fix(wiki): cancel in-flight LLM work when raw is deleted 2026-04-25 19:02:32 +08:00
matevip
4ed88ebbf9 feat(wiki): PR-8 frontend protection badges + delete-button gating 2026-04-25 19:02:32 +08:00
matevip
1e62dbad47 feat(wiki): PR-7 archived soft-archive 2026-04-25 19:02:32 +08:00
matevip
850d59c04f feat(wiki): PR-6b structured-output route phase (opt-in) 2026-04-25 19:02:32 +08:00
matevip
c8af19cae2 feat(wiki): PR-2b/2c overview rebuilder + activity log 2026-04-25 19:02:31 +08:00
matevip
f4d4e973df feat(wiki): PR-5b enrichment via replacement plan 2026-04-25 19:02:31 +08:00
matevip
726265780e build(docker): install poppler + tesseract for PDF extraction parity with Mac 2026-04-25 11:05:29 +08:00
matevip
53f54fe05d build(docker): pass MAVEN_FLAGS build-arg to support aliyun-first profile in CN builds 2026-04-25 10:32:23 +08:00
matevip
9c59092d9c feat(wiki): PR-6 skeleton — DTOs for structured eager output 2026-04-25 09:56:19 +08:00
matevip
ec0aaf7da6 feat(wiki): PR-5 wikilink alias parsing + relation seed filter 2026-04-25 09:56:19 +08:00