docs: per-component SWE-bench results + agent preset#5
Merged
Conversation
added 2 commits
July 12, 2026 17:16
Full per-component eval on SWE-bench (Claude Code, claude-sonnet-4-6) through the eval-containers gateway: zero task-reward regression across every component; mask offloads the growing file-read transcript for ~27% content-token savings with no reward loss, the biggest lever on agentic traffic. Adds docs/RESULTS.md (table + methodology + why the inert components are inert on this traffic) and a mask-centric 'agent' preset [format, dedup, failed_run, mask, extract, cacheinject]. Assisted-By: Claude (Anthropic AI) Signed-off-by: Osher-Elhadad <Osher.Elhadad@ibm.com>
…mples All-task tables (savings%, reward, agent wall-clock across 10 tasks × 13 configs), per-component summary with measured compute time (mask ~2ms, failed_run/extract ~6ms), and real before->after message rewrites captured via CONTEXT_GURU_DUMP on sphinx-7910 (mask 606->27, failed_run 423->33, extract relevance projection). Notes the marker-overhead per-message follow-up for the offloaders. Assisted-By: Claude (Anthropic AI) Signed-off-by: Osher-Elhadad <Osher.Elhadad@ibm.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up to the context-guru redesign (#4).
What
docs/RESULTS.md— per-component evaluation on SWE-bench (Claude Code,claude-sonnet-4-6) through the eval-containers gateway: the savings table, methodology/caveats, and why the inert components are inert on this traffic.agentpreset (config.go) —[format, dedup, failed_run, mask, extract, cacheinject], tuned for long agentic sessions.components.mdpresets,README.mdresults link).Result
On the 6 baseline-resolved tasks,
maskcut ~27% of content tokens with zero task-reward loss (extract11%,balanced12%,failed_run6%); every component preserved reward 6/6. Claude Code's context is dominated by line-numbered file reads re-sent each turn —mask(drop old tool outputs) attacks that directly. The components that show 0% are genuinely inert on this traffic (no JSON/commands/fenced code), not buggy.