Turn your AI history into actionable insights.

Orangu reads your local AI sessions so you don't have to guess what went right (or wrong).

Inspect: Dive deep into steps and tool calls from a single run.
Discover: Spot recurring patterns across your whole repository.
Improve: Use real evidence to build smarter, faster workflows.

$ npx orangu report
one session  →  steps, tool calls, errors, outcome evidence
repo/global  →  recurring patterns worth a larger change
improve skill  →  a structured proposal you review
scope policy →  session: apply + verify · repo: apply · global: review

Two jobs, one evidence trail

Diagnose one run. Improve repeated work.

A session explains what happened. Repo and global views show which patterns deserve a change to the way future work runs.

SESSION

Observe one session

Follow the work step by step. See tool actions, results, errors, context, subagents, time, tokens, and the evidence connected to the outcome.

Use this scope to review, explain, or diagnose a run.

REPO / GLOBAL

Improve repeated work

Find errors, re-reads, tool patterns, and harness gaps that recur across sessions. Use that evidence to choose a reviewable larger change.

repo may be applied explicitly. global is proposal-only.

The observe-to-proposal sample

Follow evidence from one step to the next run.

This illustrative flow shows what each layer owns. It is not a measured customer result.

orangu report · illustrative supported session

Usage · errors · latencySee where tool time went and what failed.

23 calls

Bashexecution9 calls1 erroravg / p95 680 ms / 3.2 s

Readfiles7 calls0 errorsavg / p95 34 ms / 61 ms

Grepsearch4 calls0 errorsavg / p95 57 ms / 86 ms

Editchanges3 calls0 errorsavg / p95 18 ms / 21 ms

Recurring errortest command exited 11 occurrence in this illustrative session

Local evidence. Counts, errors, and latency come from parsed tool calls. Recurring pattern. Repo and global scopes compare the same evidence shape across sessions.

This sessionRepoGlobal

Session: proposed → applied → verified · Repo: proposed → applied · Global: proposed

Instruction filesScripts and CLIsHooksSkills to createSkills to discoverSubagents and agentsMCP serversPluginsWorkflow and configuration

Draft proposalscript or hookproposed

Evidence. The same local setup failure precedes retries in the selected evidence.

Proposal. Add one deterministic check with explicit pass and fail output. /orangu:improve drafts the smallest reviewable change.

Explicit apply. A Claude Code or Codex skill contract requires staying within reviewed files and records a skill-authored receipt of the checks it ran. Orangu validates the receipt's shape and file-list agreement, not the diff itself. Next-run verification. For this session-scope proposal, Orangu resolves a later session from the same workspace and confirms that average failed test runs decreased without increasing tool errors.

handled by orangu:improveproposed

Local code measures and matches. An optional skill drafts. The browser only copies a handoff. A verified session comparison means its reviewed metrics passed on selected later evidence; it is not a causal or overall-quality claim.

Observe locally. Draft from evidence. Use only the lifecycle the scope supports.

Open the full sample report ↗Open the repository-level sample ↗

Change the right layer

A suggestion should name what needs to change.

Deterministic matches narrow the options. /orangu:improve drafts the smallest useful change for review; /orangu:apply performs reviewed session/repo changes only when invoked.

Instruction files

Persistent project guidance and conventions.

Scripts and CLIs

Repeatable actions with checkable output.

Hooks

Guaranteed actions at lifecycle boundaries.

Skills to create

Reusable knowledge and workflows for your setup.

Skills to discover

Existing capabilities to evaluate before installing.

Subagents and agents

Isolated or specialized work with clear ownership.

MCP servers

External tools and data the work actually needs.

Plugins

A reusable package of related extensions.

Workflow and configuration

How work is sequenced, checked, and repeated.

Built around work, not job titles

Use the same evidence at the level you need.

Detailed and plain-language views describe the same supported session. Choose the level of detail that fits the task.

Review delegated work

See the steps taken, the tools used, what completed, and which evidence supports the result.

Diagnose a rough outcome

Trace errors, retries, context pressure, and unfinished work back to the moment they appeared.

Improve a repeated workflow

Use recurring repo or global patterns to propose a script, skill, hook, instruction, tool, agent, plugin, or configuration change.

One system, clear ownership

Local evidence first. Optional interpretation on top.

No instrumentation, no account, no upload, and no telemetry are required to inspect supported sessions already on disk.

Bounded local evidence

orangu evidence accepts a supported session or current Orangu Analysis, SlimAnalysis, or Aggregate output. It redacts, bounds, matches, and ranks without a model or network request.

Optional AI interpretation

/orangu:improve explains bounded evidence, starts with catalog matches, researches missing options when useful, and writes a structured draft. It does not edit the project.

Explicit reviewed changes

/orangu:apply runs only for session/repo proposals after review. Its contract requires staying within declared files and truthfully recording local checks; Orangu validates the receipt's shape and file list. Session-only later evidence owns the current verification claim.

Raw transcriptsContain the record, but leave the reading and synthesis to you.

Local usage countersSummarize usage across time and sources.

Instrumented trace platformsCapture applications that have been wired to send traces and evaluations.

oranguConnects supported local evidence to a reviewable draft, explicit session/repo application, and session-scope later verification.

Choose the scope of the decision

One run for diagnosis. More runs for larger changes.

Scope controls the question, not the honesty of the evidence.

session

What happened here?

Inspect the run and draft one bounded proposal from one finding.

repo

What keeps happening here?

Connect recurring project patterns to example sessions and a focused change.

global

What should change in the harness?

Review patterns across supported sessions and start a separately labeled whole-harness analysis.

Start with observation

Inspect first. Add improvement when you want it.

The CLI creates the local evidence. Claude Code and Codex skills share the same scope-aware proposal, application, and session-verification contracts.

Inspect a session

No global install. Finds the latest supported session and opens its report.

$ npx orangu report

Keep the local app open

Use report, analyze, repo, global, serve, and watch without sending session data away.

$ npm i -g orangu

Add the Claude Code plugin

Install after you have seen the evidence. The report copies a chat handoff; /orangu:improve drafts and /orangu:apply applies session/repo proposals explicitly.

/plugin marketplace add NissanOhana/orangu

/plugin install orangu

One improvement workflow, explicit stages

Measurement stays in the local CLI. Session can propose, apply, and verify; repo can propose and apply; global stays proposal-only.

/orangu:improve

Turn one finding into one bounded, reviewable proposal with evidence, expected effect, risk, and a verification check. Never edits the target repository.

/orangu:apply

Apply exactly one reviewed proposal (session or repo scope) to the current repository, run its own checks, and record a receipt. Global proposals are review-only.

/orangu:analyze

Explain what happened in one session, finished or still running, from local deterministic evidence, and keep a report refreshed while a session runs.

/orangu:harness

Review what your harness declares against what your sessions actually used, across one repository or every session on the machine, and propose ranked changes.

/orangu:feedback

Send candid beta feedback about Orangu itself through a private localhost form with an exact user-reviewed GitHub preview.

Start with the run in front of you.

Inspect it locally, then decide whether the next outcome needs a proposal.

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