tools: [] sent on automatic-compaction requests → HTTP 400 from OpenAI-compatible backends (vLLM); compaction never succeeds #194

Summary

When a session is run against an OpenAI-compatible backend (--api-format openai-completions) and the conversation contains tool calls, the automatic context compaction request is sent with "tools": []. vLLM rejects that with HTTP 400, so compaction never succeeds and is re-attempted on every following agent step. The main agent requests (which carry the real tool list) keep working, so the failure is easy to miss — but the context is never compacted and keeps growing until the model's window is exhausted.

Environment

• minimax-code: 0.4.12 (code references below are against main @ 81b666a, 2026-09-19)
• Provider: custom, --api-format openai-completions
• Backend: LiteLLM proxy (v1.101.0) in front of vLLM (OpenAI-compatible /v1/chat/completions)

Observed

Error returned to the client:

400 - {"error":{"message":"`tools` must not be an empty array. Either provide at least one tool or omit the field entirely. (parameter=tools)","type":"BadRequestError","param":"tools","code":400}}

Request pattern once the context passes the compaction threshold: every successful main request (with tools) is accompanied by failing requests carrying roughly the same conversation minus the tool definitions (~30k tokens smaller) — i.e. the compaction checkpoint call. In one session this produced 41 failed requests in 4 minutes while the main turns continued normally. The compaction never completes.

Note that this is not vLLM being strict for its own sake: the OpenAI API itself rejects "tools": [] as well, so any provider that mirrors OpenAI validation will fail here.

Root cause

1. The compaction request is built without a tools field:
   packages/local-runtime-v2/src/service/turn-system/compaction/execution/checkpoint-provider.ts L174–187, buildCheckpointRequest() returns { systemPrompt, messages: [...projected, control] }. The projected messages still contain the assistant tool calls / tool results of the history.

2. The vendored pi-ai provider then injects an empty array whenever the history has tool calls but no tools are supplied:
   third_party/pi-mono/packages/ai/src/providers/openai-completions.ts L548–550

   } else if (hasToolHistory(context.messages)) {
       // Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results
       params.tools = [];
   }

   (hasToolHistory is at L48.) This branch is a workaround for Anthropic-behind-a-proxy that came in upstream via earendil-works/pi [Bug]: POST /api/file/save cannot create new files — existing-path resolver 404s before the write #149 / [Bug]: Bash permission deny rules can be skipped via transparent wrappers while allow rules match the stripped form #150. Upstream later fixed the --no-tools case in #3649 / #3650 but deliberately kept this hasToolHistory branch, so it is still present in the vendored v0.79.1 and on upstream main.

The combination (1) + (2) means every compaction call against an OpenAI-compatible backend carries "tools": [].

Reproduction

1. mcode provider add --name local --base-url http://<vllm-or-proxy>/v1 --api-format openai-completions --model <model> --api-key-env KEY --use
2. Work in a session that uses tools until the context reaches the automatic-compaction trigger (a few hundred k tokens of tool output gets there quickly).
3. Watch the backend log: each agent step now logs one 200 (main request) and a 400 with the message above (compaction request).

Minimal backend-side check, independent of mcode:

curl -s http://<vllm>/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"<model>","messages":[{"role":"user","content":"hi"}],"tools":[]}'
# -> 400 "`tools` must not be an empty array"

Suggested fix

Any of these would do; (a) is the smallest and matches the repo's vendor-patch workflow (third_party/pi-mono/MINIMAX_CHANGES.md):

• (a) In the vendored provider, drop the else if (hasToolHistory(...)) branch, or gate it so it only fires for Anthropic-style compat (e.g. compat.cacheControlFormat === "anthropic"). Upstream's openai-completions-empty-tools.test.ts covers the surrounding cases and would need its third case inverted.
• (b) In transformLocalTurnPayload (local-turn-payload-transform.ts), which already runs for both main and compaction requests: if (Array.isArray(payload.tools) && payload.tools.length === 0) delete payload.tools;
• (c) Pass the tool list through in buildCheckpointRequest — the compactor already receives tools in its input — though (a)/(b) are cleaner since the checkpoint call is not supposed to call tools.

Workaround

--api-format anthropic-messages or --api-format openai-responses: both provider paths only set tools when the list is non-empty, so the compaction request goes through.