feat: route all agents through Gemini via LiteLLM proxy#17
Conversation
Replace custom Sendblue integration with chat-adapter-sendblue and @chat-adapter/telegram, both unified under Chat SDK. All platforms share the same handlers, memory, automations, and Composio tools. - Add server/bot.ts with env-driven adapter registry (registerIfConfigured) - Delete server/sendblue.ts (replaced by chat-adapter-sendblue) - Refactor server/index.ts webhook bridge to forward all headers + debug logs - Add scripts/telegram-webhook.mjs for auto-registration on dev boot - Update scripts/dev.mjs: auto-register Telegram webhook alongside Sendblue - Add README section: "Adding more chat platforms" with Slack walkthrough - Add .agents/skills/chat-sdk/SKILL.md for agent-assisted Chat SDK work Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Capture raw request body via express.json verify callback so HMAC signature verification works for Slack, GitHub, Discord, etc. - Proxy adapter responses faithfully (status + headers + raw bytes) instead of forcing JSON — fixes URL verification challenge flows - Fix double-handler: scope onNewMessage catch-all to sendblue only; all other platforms use onDirectMessage to avoid firing twice on @-mentions - Gate webhook body logging behind DEBUG_WEBHOOKS=true env var to avoid PII in production logs - Surface webhook registration failures explicitly in autoRegisterWebhook - Update dev banner to list all active platform webhooks dynamically instead of hardcoding Sendblue-only messaging Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- litellm.config.yaml: map Claude model IDs (claude-sonnet-4-6, claude-haiku-4-5-20251001, etc.) to gemini/gemini-2.5-flash so Claude Code CLI accepts the model name while LiteLLM routes to Gemini - scripts/start-proxy.sh: load .env.local stripping inline comments; run LiteLLM on port 4001, thin proxy on port 4000 - scripts/anthropic-proxy.mjs: intercepts /v1/messages/count_tokens and returns a mock 200 (LiteLLM+Gemini sends empty body → 500 bug); proxies all other requests to LiteLLM on 4001 - server/interaction-agent.ts: fix empty-reply fallback (was sending literal "(no reply)" string to iMessage); add self-description to safe-to-answer list; log unknown SDK message types for debugging - server/bot.ts: send helpful fallback when model produces no text; log elapsed time in no-reply path Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rt and enhance execution agent logic Co-authored-by: Copilot <copilot@github.com>
Greptile SummaryThis PR routes all agent requests through a LiteLLM proxy to Gemini, adds a thin Node.js proxy to work around a LiteLLM+Gemini
Confidence Score: 4/5Safe to merge after resolving the litellm.config.yaml / .env.example model-name mismatch; all other changes are straightforward. One P1 defect: documented model names (gemini-main, gpt-main) are absent from the LiteLLM config, causing immediate failures for any user who follows the .env.example instructions. No P0 issues found. litellm.config.yaml and .env.example — model name definitions must be consistent between them. Important Files Changed
Sequence DiagramsequenceDiagram
participant CC as Claude Code CLI
participant P as anthropic-proxy :4000
participant L as LiteLLM :4001
participant G as Gemini API
CC->>P: POST /v1/messages/count_tokens
P-->>CC: 200 {input_tokens: 10000} (mocked)
CC->>P: POST /v1/messages (claude-sonnet-4-6)
P->>L: forward request
L->>G: gemini/gemini-2.5-flash
G-->>L: response
L-->>P: response
P-->>CC: response
Note over CC,G: All three mapped Claude model IDs route to gemini-2.5-flash
|
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
|
|
||
| // --------------------------------------------------------------------------- |
There was a problem hiding this comment.
After the first Sendblue message, handleTurn calls thread.subscribe(). On every subsequent message that thread is now subscribed, so onSubscribedMessage fires and onNewMessage(/[\s\S]*/) (which only guards on adapter.name !== "sendblue") also fires — resulting in two calls to handleTurn, two thread.post() calls, and two Convex mutations per message.
A simple fix is to add an early-exit guard inside onSubscribedMessage since onNewMessage already covers follow-up messages for that adapter:
bot.onSubscribedMessage(async (thread, message) => {
// Sendblue messages are already handled via onNewMessage above
if (thread.adapter.name === "sendblue") return;
await handleTurn(thread, message);
});| if (req.url?.startsWith("/v1/messages/count_tokens")) { | ||
| res.writeHead(200, { "Content-Type": "application/json" }); | ||
| res.end(JSON.stringify({ input_tokens: 10000 })); | ||
| return; |
There was a problem hiding this comment.
Returning input_tokens: 10000 for every count_tokens request means Claude Code always believes the context is ~10 k tokens, regardless of actual conversation length. If a conversation approaches the real model limit, Claude Code won't get a signal to compact or warn the user — it will silently exceed the context window and see truncation or errors at inference time.
A safer sentinel is a very large value (e.g. 200_000) so Claude Code never hits the heuristic threshold, or the comment should explicitly document the chosen value and its rationale.
| set +a | ||
| fi | ||
|
|
||
| # Ensure mandatory keys are set for LiteLLM |
There was a problem hiding this comment.
Exporting OPENAI_API_KEY as an empty string (via the :-"" default) always sets the variable in the environment, even when OpenAI is not in use. LiteLLM inspects environment variable presence to decide whether to validate/route OpenAI requests; an empty string may trigger unexpected validation errors or suppress the clearer "key not configured" diagnostic. Only exporting it when the value is non-empty would be safer.
Closes #16
Summary
litellm.config.yaml: maps Claude Code's internal model IDs (claude-sonnet-4-6,claude-haiku-4-5-20251001,claude-sonnet-4-5-20250929) togemini/gemini-2.5-flashso requests route to Gemini transparentlyscripts/start-proxy.sh: fixed.env.localloading (inline comments brokexargs); now runs LiteLLM on port 4001 + thin proxy on port 4000scripts/anthropic-proxy.mjs: new thin Node.js proxy that intercepts/v1/messages/count_tokens(LiteLLM+Gemini bug: sends empty body → 500) and returns a mock 200; proxies all other requests to LiteLLMserver/interaction-agent.ts: fixed"(no reply)"literal being sent to iMessage when model produces no text; added self-description to safe-to-answer list; added unknown SDK message type loggingserver/bot.ts: empty reply now sends a helpful fallback message instead of nothingTest plan
sh scripts/start-proxy.shstarts without env errors/v1/messages/count_tokenscalls return 200 (no more 500 floods)🤖 Generated with Claude Code