Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

benchmarks

Brief benchmark scope for evaluating agent coding workflows with and without Contextrie.

Status

Planned. This directory defines the benchmark direction. It does not yet ship a runnable harness.

The coding agent harness stub currently lives in ../coding-agent/README.md.

What this is

benchmarks is the repo place for defining how Contextrie should be evaluated against generic agent coding workflows.

The initial comparison is the same tool in two modes:

  • baseline tool usage
  • the same tool with Contextrie managing context

Initial tools:

  • Codex
  • Claude Code
  • OpenClaw
  • OpenCode

Goal

Measure two things:

  • efficacy: task completion, output quality, and correctness
  • efficiency: token usage, latency, and context footprint

This is not meant to be a leaderboard. It is meant to show whether Contextrie helps realistic agent work, and at what cost.

Protocol

  • use the same task
  • use the same repo snapshot
  • use the same model where realistically possible
  • report model differences when parity is not possible
  • compare baseline and +contextrie as paired runs
  • compare efficiency and efficacy together

For the first benchmark shape, each task runs in four modes:

  • Codex
  • Codex + Contextrie
  • OpenCode
  • OpenCode + Contextrie

Each run should start from a fresh session and fresh repo worktree so thread history and prior file changes do not leak across runs.

For +contextrie runs, the agent prompt stays the same, but Contextrie prepares a curated context bundle first and that bundle is injected ahead of the task prompt.

Metrics to capture

  • success
  • correctness
  • quality
  • tokens
  • latency
  • context volume
  • judge rationale

Minimum capture for the first runnable harness:

  • wall-clock time from prompt submit to final diff
  • final diff or patch
  • pass or fail
  • optional token counts when the tool exposes them

Judging

Use Codex as the first judge on final diffs.

  • judge against the original task and one fixed rubric
  • blind the diffs so the judge does not know which mode produced them
  • keep judging separate from generation runs