orangu v0.14.1

Measure every agent session. Keep the changes that work.

Improve your agent harness and your setup.

Better results with fewer tokens. orangu measures every agent session, proposes changes that fit how you work, and keeps the ones that work.

See a sample report

Local only. No model in the measurement. No network calls.

127.0.0.1:53211/_orangu/…/#start?zoom=machine Sample
orangu
This machine Sample 11 sessions · 224 tool calls · 12 tool errors

1 cell is 1 session1 dot is 1 tool callAn orange dot is a tool error

Sessions and agentsSample11 sessions, newest first
  1. Add rate limiting to the checkout API5.83M tokens checkout-api · ended 40m ago · orchestrator · 3 agents
    • general-purposesurvey svc-orders316k tokens
    • general-purposesurvey svc-catalog98.2k tokens
    • general-purposesurvey svc-payments120k tokens
  2. Rotate the staging database credentials667k tokensplatform-infra · 3d ago
  3. Add retries to the payment webhook handler495k tokenscheckout-api · 3d ago
  4. Write the ADR for idempotency keys190k tokenscheckout-api · 5d ago

7 more sessions

One sessionSamplelatest · ended 40m ago

Add rate limiting to the checkout API

checkout-api · 5a91c73e

1 PR · 1 commit · 8 files changed · 3 of 9 test runs failed

Turns6
Tokens5.83M
Active26m 5s
Agents3
4 files re-read within one context (10 redundant reads, 362.8k tokens)save ~363k tokens

Improvement: Read each file once in a context. Keep notes on what it holds.

claude "/orangu:improve sg_0123456789ab"
fig. 1 The app that npx orangu opens, on the sample sessions.

How it works

Three steps. One command each.

Step 1

See

Run npx orangu. It reads every session on your disk, each agent tree included, and it measures with no model. It shows one session, your sessions and their agents, and the whole machine, with the first findings.

npx orangu
Sample run~/workspace/checkout-api
$ npx orangu

orangu  Add rate limiting to the checkout API
        latest · 5a91c73e · 6 turns · 5.83M tokens · 26m 5s active
 
  1 PR · 1 commit · 8 files changed · 3 of 9 test runs failed
 
  finding  4 files re-read within one context (10 redundant reads, 362.8k
           tokens)
           Improvement: Read each file once in a context. Keep notes on what it
           holds. To find a part of it again, use Grep with line ranges. For a
           large file, read only the lines that you need (offset and limit).
  next     claude "/orangu:improve sg_0123456789ab"
  plugin   /plugin marketplace add NissanOhana/orangu
           /plugin install orangu    (once, inside Claude Code)
 
  app http://127.0.0.1:53211/_orangu/…/
      11 sessions on this machine · 0 running now
      See one session, your sessions and their agents, or the machine.
 
  The first run of /orangu:improve asks about your role and your work.
  Press ctrl-c to stop the app.

Step 2

Improve

In Claude Code, run /orangu:improve. With no input, 1 prompt runs the whole chain.

  1. Measure. orangu reads your sessions with no model.
  2. Investigate. 2 analysts on Opus look for patterns, at the highest effort.
  3. Interview. If orangu does not know your work yet, it asks about your role. You can skip it.
  4. Suggest. The brain proposes changes to skills, agents, effort, models and scripts.

You approve each change by its id. Then orangu checks it on your later sessions, beyond chance.

/orangu:improve
Claude CodeSample
> /orangu:improve
● Bash(mktemp -d)
● Bash(orangu harness --cwd '<dir>' --out '<tmp>/harness.json')
● Bash(orangu repo '<dir>' --out '<tmp>/aggregate.json')
● Bash(orangu evidence '<tmp>/aggregate.json' --scope repo --quiet
       > '<tmp>/evidence.json')
● Agent(harness-pm-analyst)      6 focus areas
● Agent(harness-devex-analyst)   6 focus areas
  Working
Proposalsg_0123456789abContext and cacheSample
Every request starts from a 28.4k-token baseline (system prompt, tools, CLAUDE.md, skills)save ~1.41M tokens

This pattern shows in all 11 sessions.

Improvement: Trim CLAUDE.md to the rules that bind. Move long docs into files that the agent reads only when it needs them.

Approve sg_0123456789abNot now

Step 3

Tune

Run /orangu:interview at any time to change your answers. The interview asks about your role and how you work. orangu keeps your answers, and the brain ranks each later proposal with them.

/orangu:interview

Which role fits your work best?

I write and ship code.I lead a team or design the system.I work with data or models.I do not write code.

How do you run agents most of the time?

One session, and I watch it.Some sessions at the same time.One orchestrator session with subagents.Long runs with no person present.

What must get better first in your next sessions?

Better results.Less of my time.Fewer tokens.

Your profile Sample

Role
I write and ship code.
Process
One orchestrator session with subagents.
Priority
Fewer tokens.

The brain ranks the rules that save tokens first. It checks agent fan-out, failed agents and the model of each agent first.

Step 1 of 3

What the brain finds

Better results with fewer tokens.

  • Skills

    A skill that never runs, a skill that loads too much text, or a procedure that needs a skill.

  • Agents

    Agents that fail, go too deep or wait. The change that keeps the work and stops the waste.

  • Effort and model

    Deep thinking on simple work. A model that does not fit the task.

  • CLIs and scripts

    The same commands again and again. A search loop that one script can do.

  • Context and cache

    A cache that breaks on each turn. Tools and long instructions that fill the context.

  • Rules that hold

    A rule in CLAUDE.md that the agent breaks. A hook or a deny rule that makes it hold.

Checked, not claimed

Each change is checked on your later sessions.

  1. No model in the measurement.

    Plain code counts, compares and ranks.

  2. Beyond chance.

    orangu marks a change verified only when your later sessions beat the earlier sessions beyond chance.

  3. You approve each change.

    orangu never changes your harness by itself. A change to the whole machine stays a proposal.

Quality, time and tokens stay separate. orangu never makes one score.

Where you work

Use the brain where you work.

Claude Code

Add the plugin once. Then /orangu:improve runs the whole chain in 1 prompt.

/plugin marketplace add NissanOhana/orangu
/plugin install orangu
/orangu:improve

claude.ai, Claude Desktop and the Claude API

Run npx orangu export. Upload the zip once as a custom skill. Then add your evidence to a chat.

npx orangu export

Codex

Add the orangu plugin. Then run $orangu-improve.

$orangu-improve
orangu, the orangutan mascot

Start with the sessions you already have.

Supported today: Claude Code, Cowork and Desktop sessions on your machine.