MyClaudeCode

A mini coding agent used as an instrument to put numbers on unmeasured coding-agent advice.

Starting · join earlyAI Engineering

3 of 10 tasks done

1 · Purpose

MyClaudeCode builds a small coding agent to understand how tools like Claude Code work from the inside. Official guidance already explains how to set these tools up, but it does not measure the mechanisms it describes. This project runs experiments to find the numbers: when instruction files get too big, when subagents pay off, what compaction forgets, and where answer quality starts to fall. It is for engineers using coding agents on large codebases who want evidence, not folklore.

For: Engineers and team leads who use coding agents on large codebases.

2 · Business benefits

Lower token spendKnowing which always-loaded layer dominates the baseline lets a team cut it first.
Fewer lost instructionsA catalogue of what compaction forgets tells teams what to back up or restate.
Evidence-based settingsThresholds for when to clear or compact come from measurement instead of guesses.
Reuse before buildingA reuse ledger shows which existing open-source tools already cover each need.

3 · Status and how to join

Starting · join early. Early days: the shape is still being decided, so joiners now have the most say. Lead: Basant Choudhary.

Good first tasks this week:

  • Run the existing usage tools over your own sessions and share the baseline split.
  • Write the trace format for the mini agent's first version.
  • Add one checked entry to the reuse ledger.

The code is not public yet; joiners get access.

4 · Requirements

What it must do

How well it must do it

5 · Architecture, very high level

Agent loopCalls the model, runs tools andrecords every step.Tools and editorRead, search and edit filesunder simple invariants.Context assemblerBuilds what the model sees, inbudgeted stages.Transcript probesSplit the baseline and examinewhat compaction removed.Experiment runnerRepeats each experiment andrecords the numbers.Governance kitStatus line, warning hook andbackup for real coding agents.tool callsmeasurementsevidence-based
Components and what each is responsible for, with no tool names: the tools come in the next section.

Also: Context assembler → Agent loop: budgeted context · Experiment runner → Agent loop: forced scenarios

6 · Tech stack, per platform

ComponentOpen stackAzureAWS
Agent loopHand-written Python, OpenAI-compatible client on GroqAzure AI Foundry modelsAmazon Bedrock models
Tools and editorPython——
Context assemblerPython——
Transcript probesPython over local transcripts, plus ccusage and cc-analyzer——
Experiment runnerPython scripts——
Governance kitClaude Code status line and hooks——

Bold with ✓ = built and running today. Everything else is the planned or possible binding: the components stay the same, only the tools change.

7 · What needs to be done

Reposition and design 3 of 3 done

TaskStatusSizeSkill
Check whether the planned guide already existsdoneSresearch
Decisions, glossary, reuse ledger and experiment listdoneMarchitecture
Setup advisor and droppable skill specdoneStechnical writing

Measure the baseline 0 of 2 done

TaskStatusSizeSkill
Run existing usage tools over past sessions and read off baselinestodoSanalysis
Baseline splitter and compaction forensics probestodoMPython

Mini agent 0 of 3 done

TaskStatusSizeSkill
Loop with a tracetodoMPython
Tools, invariants and a deliberately naive editortodoMPython
Context assembly calibrated against the real tooltodoMLLM engineering

Experiments and kit 0 of 2 done

TaskStatusSizeSkill
Compaction taxonomy and subagent break-eventodoLexperiment design
Quality degradation threshold: cliff or slidetodoLexperiment design

Size: S = an evening or two, M = about a week part-time, L = a few weeks.

8 · Roles

RoleHow manyHours / weekWhat they do
Agent builder13-5Builds the mini agent and probes.
Experiment designer12-3Designs repeatable experiments and checks the numbers.

9 · What you'll learn

Level-up questions

10 · Links