Files
openclaw/docs/tools/loop-detection.md
Vincent Koc 180e295dc6 docs(tools): rewrite loop detection, code execution, and tighten elevated/skills
Loop detection (docs/tools/loop-detection.md): substantial rewrite.
Fixed the post-compaction guard default story — the guard runs whenever
tools.loopDetection.enabled is not explicitly false, even with no
config block at all (verified in src/agents/pi-embedded-runner/run.ts
near line 800: 'enabled: resolvedLoopDetectionConfig?.enabled !==
false'). The previous doc framed it as opt-in. Added the missing
unknownToolThreshold field (default 10) sourced from
src/config/schema.help.ts, a complete fields table, and a CardGroup
related links section.

Code execution (docs/tools/code-execution.md): rewrote with
Steps-driven setup, code-verified defaults from
extensions/xai/src/code-execution-shared.ts (default model
grok-4-1-fast, default timeout 30 s, optional maxTurns), the
missing_xai_api_key structured error documented as JSON, and a
properties summary table. Replaced the trailing bullet list with a
CardGroup pointing at exec, exec-approvals, web tools, and the xAI
provider page.

Elevated (docs/tools/elevated.md): converted Related to a CardGroup
and added a Note that the bash chat command (! prefix / /bash alias)
also requires tools.elevated, sourced from
src/config/schema.help.ts:1375.

Skills config (docs/tools/skills-config.md): renamed the
'Sandboxed skills + env vars' subhead to remove the brittle '+'
character per docs/CLAUDE.md, promoted the host-only env warning to a
Warning block so the most common skill-config footgun stays visible,
and converted Related to a CardGroup including a config-reference
link.
2026-05-05 16:49:29 -07:00

155 lines
8.2 KiB
Markdown

---
summary: "How to enable and tune guardrails that detect repetitive tool-call loops"
title: "Tool-loop detection"
read_when:
- A user reports agents getting stuck repeating tool calls
- You need to tune repetitive-call protection
- You are editing agent tool/runtime policies
- You hit `compaction_loop_persisted` aborts after a context-overflow retry
---
OpenClaw has two cooperating guardrails for repetitive tool-call patterns:
1. **Loop detection** (`tools.loopDetection.enabled`) — disabled by default. Watches the rolling tool-call history for repeated patterns and unknown-tool retries.
2. **Post-compaction guard** (`tools.loopDetection.postCompactionGuard`) — enabled by default unless `tools.loopDetection.enabled` is explicitly `false`. Arms after every compaction-retry and aborts the run when the agent emits the same `(tool, args, result)` triple within the window.
Both are configured under the same `tools.loopDetection` block, but the post-compaction guard runs whenever the master switch is not explicitly off. Set `tools.loopDetection.enabled: false` to silence both surfaces.
## Why this exists
- Detect repetitive sequences that do not make progress.
- Detect high-frequency no-result loops (same tool, same inputs, repeated errors).
- Detect specific repeated-call patterns for known polling tools.
- Prevent context-overflow then compaction then same-loop cycles from running indefinitely.
## Configuration block
Global defaults, with every documented field shown:
```json5
{
tools: {
loopDetection: {
enabled: false, // master switch for the rolling-history detectors
historySize: 30,
warningThreshold: 10,
criticalThreshold: 20,
unknownToolThreshold: 10,
globalCircuitBreakerThreshold: 30,
detectors: {
genericRepeat: true,
knownPollNoProgress: true,
pingPong: true,
},
postCompactionGuard: {
windowSize: 3, // armed after compaction-retry; runs unless enabled is explicitly false
},
},
},
}
```
Per-agent override (optional):
```json5
{
agents: {
list: [
{
id: "safe-runner",
tools: {
loopDetection: {
enabled: true,
warningThreshold: 8,
criticalThreshold: 16,
},
},
},
],
},
}
```
### Field behavior
| Field | Default | Effect |
| -------------------------------- | ------- | ------------------------------------------------------------------------------------------------------------------------------- |
| `enabled` | `false` | Master switch for the rolling-history detectors. Setting `false` also disables the post-compaction guard. |
| `historySize` | `30` | Number of recent tool calls kept for analysis. |
| `warningThreshold` | `10` | Threshold before a pattern is classified as warning-only. |
| `criticalThreshold` | `20` | Threshold for blocking repetitive loop patterns. |
| `unknownToolThreshold` | `10` | Block repeated calls to the same unavailable tool after this many misses. |
| `globalCircuitBreakerThreshold` | `30` | Global no-progress breaker threshold across all detectors. |
| `detectors.genericRepeat` | `true` | Detects repeated same-tool + same-params patterns. |
| `detectors.knownPollNoProgress` | `true` | Detects known polling-like patterns with no state change. |
| `detectors.pingPong` | `true` | Detects alternating ping-pong patterns. |
| `postCompactionGuard.windowSize` | `3` | Number of post-compaction tool calls during which the guard stays armed and the count of identical triples that aborts the run. |
For `exec`, no-progress checks compare stable command outcomes and ignore volatile runtime metadata such as duration, PID, session ID, and working directory. When a run id is available, recent tool-call history is evaluated only within that run so scheduled heartbeat cycles and fresh runs do not inherit stale loop counts from earlier runs.
## Recommended setup
- For smaller models, set `enabled: true` and leave the thresholds at their defaults. Flagship models rarely need rolling-history detection and can leave the master switch at `false` while still benefiting from the post-compaction guard.
- Keep thresholds ordered as `warningThreshold < criticalThreshold < globalCircuitBreakerThreshold`.
- If false positives occur:
- Raise `warningThreshold` and/or `criticalThreshold`.
- Optionally raise `globalCircuitBreakerThreshold`.
- Disable only the specific detector causing issues (`detectors.<name>: false`).
- Reduce `historySize` for less strict historical context.
- To disable everything (including the post-compaction guard), set `tools.loopDetection.enabled: false` explicitly.
## Post-compaction guard
When the runner completes a compaction-retry after a context-overflow, it arms a short-window guard that watches the next few tool calls. If the agent emits the same `(toolName, argsHash, resultHash)` triple multiple times within the window, the guard concludes that compaction did not break the loop and aborts the run with a `compaction_loop_persisted` error.
The guard is gated by the master `tools.loopDetection.enabled` flag with one twist: it stays **enabled when the flag is unset or `true`** and only deactivates when the flag is explicitly `false`. This is intentional. The guard exists to escape compaction loops that would otherwise burn unbounded tokens, so a no-config user still gets the protection.
```json5
{
tools: {
loopDetection: {
// master switch; set false to disable the guard along with the rolling detectors
enabled: true,
postCompactionGuard: {
windowSize: 3, // default
},
},
},
}
```
- Lower `windowSize` is stricter (fewer attempts before abort).
- Higher `windowSize` gives the agent more recovery attempts.
- The guard never aborts when results are changing, only when results are byte-identical across the window.
- It is intentionally narrow: it fires only in the immediate aftermath of a compaction-retry.
<Note>
The post-compaction guard runs whenever the master flag is not explicitly `false`, even if you never wrote a `tools.loopDetection` block. To verify, look for `post-compaction guard armed for N attempts` in the gateway log immediately after a compaction event.
</Note>
## Logs and expected behavior
When a loop is detected, OpenClaw reports a loop event and either dampens or blocks the next tool-cycle depending on severity. This protects users from runaway token spend and lockups while preserving normal tool access.
- Warnings come first.
- Suppression follows when patterns persist past the warning threshold.
- Critical thresholds block the next tool-cycle and surface a clear loop-detection reason in the run record.
- The post-compaction guard emits `compaction_loop_persisted` errors with the offending tool name and identical-call count.
## Related
<CardGroup cols={2}>
<Card title="Exec approvals" href="/tools/exec-approvals" icon="shield">
Allow/deny policy for shell execution.
</Card>
<Card title="Thinking levels" href="/tools/thinking" icon="brain">
Reasoning effort levels and provider-policy interaction.
</Card>
<Card title="Sub-agents" href="/tools/subagents" icon="users">
Spawning isolated agents to bound runaway behavior.
</Card>
<Card title="Configuration reference" href="/gateway/configuration-reference" icon="gear">
Full `tools.loopDetection` schema and merging semantics.
</Card>
</CardGroup>