alchemistwu/dsh-tool-call-guard

dsh-tool-call-guard

English | 中文

Neutralize tool calls with invalid JSON arguments before they reach the wire — so one malformed model generation cannot brick an entire session.

dsh plugin add dsh-tool-call-guard

Why

Some models occasionally emit a tool call whose arguments string is not valid JSON — most often unescaped inner quotes:

{"queries": [""The Idiots" trailer youtube official"]}

Strict OpenAI-compatible servers (vLLM and friends) are asymmetric about this:

• Streaming generation path — lenient. The malformed call streams through, the harness persists it into the append-only session log, and the turn appears to succeed.
• History-replay path — strict. The next request replays the poisoned tool call, the server validates it, and rejects the entire request:

400 {"message": "Assistant tool call function.arguments must be valid JSON.",
     "type": "BadRequestError"}

From that moment, every subsequent request in the session fails. The session is bricked until its log is surgically repaired by hand. The same failure class exists across ecosystems (openai-agents-python #2061, vLLM #41122).

What it does

Intercepts the harness's llm/stream waterfall. For every assistant tool-call block whose arguments fail JSON.parse:

1. The call becomes an honest text record (wire-only) — the model sees exactly what it emitted and can re-issue a corrected call:

   [A tool call to 'web_search' was removed from history because its arguments
   were malformed JSON. Original arguments as emitted: {"queries": [""The Idiots" …]}]
   

2. The matching tool result is re-expressed as a plain user message — result content is preserved, and the conversation stays protocol-balanced (no dangling tool_calls entry, no orphan role:"tool" reply — each of those is itself a 400 on strict servers):

   [Tool Result: web_search] 10 results about The Idiots
   

The orphaned-result-as-user-message pattern follows the upstream serializer discussion (deepseek-harness #4668).

Properties

• Zero overhead for clean history — one JSON.parse per tool-call block; messages array passes through with object identity when nothing is wrong.
• Append-only log respected — the harness's durable session log is never rewritten; neutralization is applied per-request on the wire only.
• Provider-agnostic — sits on llm/stream, so every adapter (deepseek, community, custom) is covered.
• Fail-open — if the guard itself errors, the original request passes through untouched.
• Zero configuration — install and restart.

Install

dsh plugin --profile web add dsh-tool-call-guard
# or desktop:
dsh plugin --profile desktop add dsh-tool-call-guard

Or from GitHub: dsh plugin add github:alchemistwu/dsh-tool-call-guard.

The package declares a dsh.bundle patch, so the plugin self-registers on install. Restart your dsh web / DSH Desktop host, then start a new session.

Observed in production

First observed with zai-org/GLM-5.3-Flash served by vLLM 0.27 (--enable-auto-tool-choice): one web_search call with unescaped quotes around a film title streamed through fine, was persisted, and permanently poisoned the session with 400 … arguments must be valid JSON on every later turn. This plugin keeps that session alive: the poisoned entry is re-expressed per-request, the session continues working, and the model sees its own mistake in context.

Tests

npm test

Nine tests cover the exact production poison, mixed valid/invalid tool calls, protocol-balance invariants (no invalid tool_calls and no orphan role:"tool" on the wire), degenerate inputs, and per-request state reset.

License

MIT