Repository navigation
Replies: 10 comments
|
Thanks for the detailed report and the exact chat log, @xspylol — that log is what makes this debuggable. Honest status on each, all verified against the released v3.8.9 code: 1 + 2 — DeepSeek-web session + tool-calls did ship in v3.8.9, but you're hitting two real gaps that are on me: Session amnesia (still happening — here's the why): the persistent-session + rolling-window engine landed (#2942), but it's opt-in, defaults OFF, and I didn't ship a UI toggle. The wizard's
Tool-calls printing raw JSON (real limitation): translation shipped too (#2820), but the parser only recognizes 3 — Qwen-web 504: that 4 — Universalizing it: good news — I built the translation layer ( 5 — New web providers: all three already have tracking threads (yours): Xiaomi MiMo → #2681, Z.ai/GLM → #2682, MiniMax → #2684. They're in the catalog backlog; I'll fold the "agentic session + tool-calls from day one" requirement into them. Same honest caveat as always: this is free-tier web reverse-engineering — even fully wired, it stays best-effort (DeepSeek/Qwen can change rendering or the PoW challenge and break it overnight). But the session-toggle gap (#3153) and the bare-JSON parser gap (#3154) are squarely mine to close. Thanks for driving the whole web-agentic effort — you're credited as originator on #2820/#2942, and any code you land gets |
|
Quick update, @xspylol -- the Qwen Web v1 endpoint was officially retired upstream (all requests returned 504 regardless of credentials). I've migrated the executor to the v2 API with full WAF cookie-jar replay: PR #3723 is now merged into To connect, paste the full Cookie header from a logged-in The agentic session / persistent history work is still tracked separately in issues #3153 + #3154. That's a broader cross-provider feature -- no timeline yet, but the Qwen routing blocker is now out of the way. |
|
Good framing, @richardchen874-sys. The distinction maps directly to what we have been building: the v2 WAF migration (PR #3723) handled the transport layer (no more 504s), and the persistent-session + rolling-window engine that shipped in v3.8.9 (PR #3101) covers the first two of your requirements -- |
|
That layering is the right mental model, @richardchen874-sys. Transport (Qwen v2/WAF) and session continuity ( |
|
That six-layer breakdown (transport stability -> session continuity -> context replay -> response normalization -> tool compatibility -> failure mapping) is exactly how I think about it too, @richardchen874-sys -- and you're right that "chat-compatible" vs "agentic-workflow-compatible" is the framing that actually matters for multi-step coding agents. Quick status correction: the #2725 umbrella is now closed as completed -- transport (Qwen v2 / WAF migration) and session continuity ( |
|
Follow-up @richardchen874-sys -- opened #5876 to track exactly this: the defensive, provider-uniform |
|
@xspylol — still hitting this on the latest version? The Qwen Web provider and agentic session support have been updated since this was opened. Otherwise I will close this in 7 days for inactivity. |
|
My qwen is conected but the models test returns: Provider returned empty content |
|
@jjbrandon8 that thread is from an older version and the provider changed underneath it, so the answer depends on which Qwen connection you have.
"Provider returned empty content" means the upstream answered 200 but delivered no content ( To point you at the right place, reply with:
If it is |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hey @diegosouzapw, Huge thanks for v3.8.9! However, I'm still experiencing critical issues that make web-cookie providers unusable for agentic workflows:
Despite the release notes saying persistent session caching was added for DeepSeek Web, I'm still getting a fresh session for every single message. The model has zero cross-turn memory and treats each request as a brand new conversation. This makes it impossible to have multi-turn coding sessions.
When the agent tries to use tools via web-cookie executors, it fails to trigger native tool execution. Instead, it just prints the raw JSON arguments as plain text in the chat. The SSE parser isn't wrapping the model's output into proper OpenAI tool_calls objects, so the IDE ignores them and the tools never run.
Here is the exact chat log showing both failures:
User: do you see this folder {"name": "read", "arguments": { "filePath": "C:\Users\STEEL LEGEND\Documents\Test" }}
Agent: [Hallucinates wrong files because it has no memory of previous context]
User: no theres a python file no html no web thing
Agent: Let me look at what's in your project directory. {"command": "Get-ChildItem -LiteralPath "C:\Users\STEEL LEGEND\Documents\Test" -Recurse | Select-Object -First 50", "description": "List files in project directory"}
(Notice how it just types the JSON string instead of actually running the command? And it has no memory that I already showed it the folder contents earlier?)
Making session persistence, native tool-translation, and proper tool_calls stream wrapping the default standard across every web executor would make OmniRoute's auto-fallback combos truly unbreakable. Thanks!
All reactions