8 replies

Tools broken after first exec use

---

Bug:
call_auto_1
call_auto_1
/
functionsread
functionsread
— tools permanently broken after first failed exec call

Version: 2026.3.2 · Provider: nvidia-nim · Model: moonshotai/kimi-k2.5 · api: openai-completions

Gateway warning (repeats every request):

[tools] tools.profile (coding) allowlist contains unknown entries
(apply_patch, cron, image). These won't match any tool unless plugin is enabled.

[tools] tools.profile (coding) allowlist contains unknown entries
(apply_patch, cron, image). These won't match any tool unless plugin is enabled.

The break pattern — session transcript:

// Working call:
{"toolCallId":"functions.read:0","toolName":"read"} → OK

// First broken call (stopReason=toolUse but NO toolCall in content):
{"role":"assistant","stopReason":"toolUse","content":[
  {"type":"thinking","thinking":"I'll use exec to delete the file"},
  {"type":"text","text":"Ich prüfe die Existenz... [truncated mid-sentence]"}
]}
// → OpenClaw generates: {"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found"

// All subsequent calls also produce call_auto_1 — session never recovers:
{"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found" ×10+
{"toolCallId":"functionsread:0","toolName":"functionsread"} → "Tool functionsread not found"

// Working call:
{"toolCallId":"functions.read:0","toolName":"read"} → OK

// First broken call (stopReason=toolUse but NO toolCall in content):
{"role":"assistant","stopReason":"toolUse","content":[
  {"type":"thinking","thinking":"I'll use exec to delete the file"},
  {"type":"text","text":"Ich prüfe die Existenz... [truncated mid-sentence]"}
]}
// → OpenClaw generates: {"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found"

// All subsequent calls also produce call_auto_1 — session never recovers:
{"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found" ×10+
{"toolCallId":"functionsread:0","toolName":"functionsread"} → "Tool functionsread not found"

Key observation:
-

read

read

normally produces ID

functions.read:0

functions.read:0

→ resolves correctly
- When model tries

exec

exec

, NIM returns

stopReason: toolUse

stopReason: toolUse

but omits the toolCall block entirely — text cuts off mid-sentence
- OpenClaw generates

call_auto_1

call_auto_1

as fallback → session corrupted permanently
- Same behavior with all fallback models (GLM, Qwen, GPT-OSS) → suggests session corruption, not provider bug

Config:

tools.profile: full

tools.profile: full

sandbox.allow

sandbox.allow

includes exec ·

workspaceAccess: rw

workspaceAccess: rw

Question: Is

call_auto_1

call_auto_1

a known fallback ID when

stopReason=toolUse

stopReason=toolUse

but no toolCall is present in content? Does OpenClaw have handling for this case?

FotC�

Friends of the Crustacean 🦞🤝•3d ago•

8 replies

Imheretorage

Tools broken after first exec use

[tools] tools.profile (coding) allowlist contains unknown entries
(apply_patch, cron, image). These won't match any tool unless plugin is enabled.

[tools] tools.profile (coding) allowlist contains unknown entries
(apply_patch, cron, image). These won't match any tool unless plugin is enabled.

The break pattern — session transcript:

// Working call:
{"toolCallId":"functions.read:0","toolName":"read"} → OK

// First broken call (stopReason=toolUse but NO toolCall in content):
{"role":"assistant","stopReason":"toolUse","content":[
  {"type":"thinking","thinking":"I'll use exec to delete the file"},
  {"type":"text","text":"Ich prüfe die Existenz... [truncated mid-sentence]"}
]}
// → OpenClaw generates: {"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found"

// All subsequent calls also produce call_auto_1 — session never recovers:
{"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found" ×10+
{"toolCallId":"functionsread:0","toolName":"functionsread"} → "Tool functionsread not found"

// Working call:
{"toolCallId":"functions.read:0","toolName":"read"} → OK

// First broken call (stopReason=toolUse but NO toolCall in content):
{"role":"assistant","stopReason":"toolUse","content":[
  {"type":"thinking","thinking":"I'll use exec to delete the file"},
  {"type":"text","text":"Ich prüfe die Existenz... [truncated mid-sentence]"}
]}
// → OpenClaw generates: {"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found"

// All subsequent calls also produce call_auto_1 — session never recovers:
{"toolCallId":"call_auto_1","toolName":"unknown"} → "Tool not found" ×10+
{"toolCallId":"functionsread:0","toolName":"functionsread"} → "Tool functionsread not found"

Key observation:
-

read

read

normally produces ID

functions.read:0

functions.read:0

→ resolves correctly
- When model tries

exec

exec

, NIM returns

stopReason: toolUse

stopReason: toolUse

but omits the toolCall block entirely — text cuts off mid-sentence
- OpenClaw generates

call_auto_1

call_auto_1

as fallback → session corrupted permanently
- Same behavior with all fallback models (GLM, Qwen, GPT-OSS) → suggests session corruption, not provider bug

Config:

tools.profile: full

tools.profile: full

sandbox.allow

sandbox.allow

includes exec ·

workspaceAccess: rw

workspaceAccess: rw

Question: Is

call_auto_1

call_auto_1

a known fallback ID when

stopReason=toolUse

stopReason=toolUse

but no toolCall is present in content? Does OpenClaw have handling for this case?

Tools broken after first exec use

Similar Threads

Tools broken after first exec use

Similar Threads

Similar Threads

Similar Threads