/prune — zero-LLM-cost context pruning for agentic sessions

[jamesonBradfield]

Long tool-heavy chats accumulate a lot of redundant payload: the same file read three times, a failed command that was fixed two turns ago, reads that were immediately superseded by a write to the same path. /compact handles this well, but it costs an LLM call and rewrites everything. Sometimes you just want to surgically cut the junk.

I wrote a small slash command called cc_dcp (Dynamic Context Pruning) that runs entirely on the Lua side — no LLM call, no HTTP round-trip.

Three strategies:

Strategy	What it drops
Deduplication	Older copies of identical tool+input pairs (e.g., iwe_tree called 4× with the same args)
Error purging	Tool results matching error/exception/traceback patterns that have already been dealt with
File supersede	Any file read/write on a path that was touched again in a later turn

A PROTECT_TURNS guard keeps the most recent N cycles untouched so you never prune something still in play.

Works with both adapter formats — Claude (call.name/call.input) and OpenAI/DeepSeek (call["function"].name/call["function"].arguments as JSON string).

In practice on a 35k-token Godot session: 10k tokens removed in one run, instant.

Repo + install instructions: https://github.com/jamesonBradfield/codecompanion-dcp

[olimorris]

Appreciate the share, this looks fantastic.

I will be merging #3083 this coming week which will include context editing and compaction for all http adapters.

[jamesonBradfield]

Man, thank you! I really appreciate the shoutout. I'm incredibly hyped for #3083 to get merged. The new context management and hooks are exactly what the ecosystem needs to take things to the next level. Looking forward to porting this over to the new APIs once they drop!

[jamesonBradfield]

If you want to adapt the deduplication and file-supersede logic from my DCP script for it, I'd love to formally submit it as a PR.

[jamesonBradfield]

Just a heads up, I'm kicking off a new term right now and haven't had a chance to dig into the new compaction API surface yet. If you're open to pointing me toward the right hooks once #3083 drops, I'd be happy to wire it up!

[mnifzied-create]

Great lever — the transcript-accumulation side (duplicate reads, superseded writes, dead error payloads) is real, and /compact via an LLM call is overkill for it. Zero-LLM dedup is the right call; the error-purging strategy especially — stale failed-command output is pure dead weight.

Worth pairing with the fixed half /prune doesn't touch: the tool-schema overhead re-sent every turn regardless of how clean the transcript is. It doesn't accumulate, but you pay it on every request — across 13 agents I measured, the per-turn schema tax ran a median ~547 tokens (up to 3,546 for the GitHub MCP server), and ~20% of that is non-semantic cruft ($schema, additionalProperties, auto-added title) strippable with zero behavior change.

So they're complementary: prune the transcript (yours) + trim the schemas (one-time). Data/method if useful: https://mnifzied-create.github.io/agentloop/token-tax/