admin/Moltbot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: cap context window resolution (#6187) (thanks @iamEvanYT)

Browse Source
This commit is contained in:
Ayaan Zaidi
2026-02-01 19:49:35 +05:30
committed by Ayaan Zaidi
parent 5d3c898a94
commit 0992c5a809
7 changed files with 50 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
docs/concepts/session-pruning.md
View File

@@ -3,36 +3,30 @@ summary: "Session pruning: tool-result trimming to reduce context bloat"
read_when:
- You want to reduce LLM context growth from tool outputs
- You are tuning agents.defaults.contextPruning
title: "Session Pruning"
---
# Session Pruning
Session pruning trims **old tool results** from the in-memory context right before each LLM call. It does **not** rewrite the on-disk session history (`*.jsonl`).
## When it runs
- When `mode: "cache-ttl"` is enabled and the last Anthropic call for the session is older than `ttl`.
- Only affects the messages sent to the model for that request.
- Only active for Anthropic API calls (and OpenRouter Anthropic models).
- For best results, match `ttl` to your model `cacheRetention`.
- After a prune, the TTL window resets so subsequent requests keep cache until `ttl` expires again.
- Only active for Anthropic API calls (and OpenRouter Anthropic models).
- For best results, match `ttl` to your model `cacheControlTtl`.
- After a prune, the TTL window resets so subsequent requests keep cache until `ttl` expires again.
## Smart defaults (Anthropic)
- **OAuth or setup-token** profiles: enable `cache-ttl` pruning and set heartbeat to `1h`.
- **API key** profiles: enable `cache-ttl` pruning, set heartbeat to `30m`, and default `cacheRetention: "short"` on Anthropic models.
- **API key** profiles: enable `cache-ttl` pruning, set heartbeat to `30m`, and default `cacheControlTtl` to `1h` on Anthropic models.
- If you set any of these values explicitly, OpenClaw does **not** override them.
## What this improves (cost + cache behavior)
- **Why prune:** Anthropic prompt caching only applies within the TTL. If a session goes idle past the TTL, the next request re-caches the full prompt unless you trim it first.
- **What gets cheaper:** pruning reduces the **cacheWrite** size for that first request after the TTL expires.
- **Why the TTL reset matters:** once pruning runs, the cache window resets, so follow‑up requests can reuse the freshly cached prompt instead of re-caching the full history again.
- **What it does not do:** pruning doesn’t add tokens or “double” costs; it only changes what gets cached on that first post‑TTL 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.
@@ -40,42 +34,35 @@ Session pruning trims **old tool results** from the in-memory context right befo
- Tool results containing **image blocks** are skipped (never trimmed/cleared).
## Context window estimation
Pruning uses an estimated context window (chars ≈ tokens × 4). The base window is resolved in this order:
1) `models.providers.*.models[].contextWindow` override.
2) Model definition `contextWindow` (from the model registry).
3) Default `200000` tokens.
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. `agents.defaults.contextTokens`.
4. Default `200000` tokens.
If `agents.defaults.contextTokens` is set, it is treated as a cap (min) on the resolved window.
## Mode
### cache-ttl
- Pruning only runs if the last Anthropic call is older than `ttl` (default `5m`).
- When it runs: same soft-trim + hard-clear behavior as before.
## 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.
- Matching is case-insensitive.
- 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)
- `ttl`: `"5m"`
- `keepLastAssistants`: `3`
- `softTrimRatio`: `0.3`
@@ -85,37 +72,33 @@ Pruning uses an estimated context window (chars ≈ tokens × 4). The window siz
- `hardClear`: `{ enabled: true, placeholder: "[Old tool result content cleared]" }`
## Examples
Default (off):
```json5
{
agent: {
contextPruning: { mode: "off" },
},
contextPruning: { mode: "off" }
}
}
```
Enable TTL-aware pruning:
```json5
{
agent: {
contextPruning: { mode: "cache-ttl", ttl: "5m" },
},
contextPruning: { mode: "cache-ttl", ttl: "5m" }
}
}
```
Restrict pruning to specific tools:
```json5
{
agent: {
contextPruning: {
mode: "cache-ttl",
tools: { allow: ["exec", "read"], deny: ["*image*"] },
},
},
tools: { allow: ["exec", "read"], deny: ["*image*"] }
}
}
}
```