93 lines
3.0 KiB
Markdown
93 lines
3.0 KiB
Markdown
---
|
||
summary: "Session pruning: opt-in tool-result trimming to reduce context bloat"
|
||
read_when:
|
||
- You want to reduce LLM context growth from tool outputs
|
||
- You are tuning agent.contextPruning
|
||
---
|
||
# Session Pruning
|
||
|
||
Session pruning trims **old tool results** from the in-memory context right before each LLM call. It is **opt-in** and does **not** rewrite the on-disk session history (`*.jsonl`).
|
||
|
||
## When it runs
|
||
- Before each LLM request (context hook).
|
||
- Only affects the messages sent to the model for that request.
|
||
|
||
## What can be pruned
|
||
- Only `toolResult` messages.
|
||
- User + assistant messages are **never** modified.
|
||
- The last `keepLastAssistants` assistant messages are protected; tool results after that cutoff are not pruned.
|
||
- If there aren’t enough assistant messages to establish the cutoff, pruning is skipped.
|
||
- Tool results containing **image blocks** are skipped (never trimmed/cleared).
|
||
|
||
## Context window estimation
|
||
Pruning uses an estimated context window (chars ≈ tokens × 4). The window size is resolved in this order:
|
||
1) Model definition `contextWindow` (from the model registry).
|
||
2) `models.providers.*.models[].contextWindow` override.
|
||
3) `agent.contextTokens`.
|
||
4) Default `200000` tokens.
|
||
|
||
## Modes
|
||
### adaptive
|
||
- If estimated context ratio ≥ `softTrimRatio`: soft-trim oversized tool results.
|
||
- If still ≥ `hardClearRatio` **and** prunable tool text ≥ `minPrunableToolChars`: hard-clear oldest eligible tool results.
|
||
|
||
### aggressive
|
||
- Always hard-clears eligible tool results before the cutoff.
|
||
- Ignores `hardClear.enabled` (always clears when eligible).
|
||
|
||
## Soft vs hard pruning
|
||
- **Soft-trim**: only for oversized tool results.
|
||
- Keeps head + tail, inserts `...`, and appends a note with the original size.
|
||
- Skips results with image blocks.
|
||
- **Hard-clear**: replaces the entire tool result with `hardClear.placeholder`.
|
||
|
||
## Tool selection
|
||
- `tools.allow` / `tools.deny` support `*` wildcards.
|
||
- Deny wins.
|
||
- Empty allow list => all tools allowed.
|
||
|
||
## Interaction with other limits
|
||
- Built-in tools already truncate their own output; session pruning is an extra layer that prevents long-running chats from accumulating too much tool output in the model context.
|
||
- Compaction is separate: compaction summarizes and persists, pruning is transient per request. See [/concepts/compaction](/concepts/compaction).
|
||
|
||
## Defaults (when enabled)
|
||
- `keepLastAssistants`: `3`
|
||
- `softTrimRatio`: `0.3`
|
||
- `hardClearRatio`: `0.5`
|
||
- `minPrunableToolChars`: `50000`
|
||
- `softTrim`: `{ maxChars: 4000, headChars: 1500, tailChars: 1500 }`
|
||
- `hardClear`: `{ enabled: true, placeholder: "[Old tool result content cleared]" }`
|
||
|
||
## Examples
|
||
Minimal (adaptive):
|
||
```json5
|
||
{
|
||
agent: {
|
||
contextPruning: { mode: "adaptive" }
|
||
}
|
||
}
|
||
```
|
||
|
||
Aggressive:
|
||
```json5
|
||
{
|
||
agent: {
|
||
contextPruning: { mode: "aggressive" }
|
||
}
|
||
}
|
||
```
|
||
|
||
Restrict pruning to specific tools:
|
||
```json5
|
||
{
|
||
agent: {
|
||
contextPruning: {
|
||
mode: "adaptive",
|
||
tools: { allow: ["bash", "read"], deny: ["*image*"] }
|
||
}
|
||
}
|
||
}
|
||
```
|
||
|
||
See config reference: [Gateway Configuration](/gateway/configuration)
|