RetroPairCapture
Collect your real agent-session logs. Sanitize copies before analysis; leave the originals untouched.
Human–AI Collaboration Retrospectives
Your agent keeps going in circles. What would you change?
Find the habits that get in the way of working with your coding agent — and concrete actions to try in your next task.
RetroPair reviews your real sessions: how you frame tasks, check results and steer the work. Get recommendations tied to examples, then check them against later sessions.
Native Claude Code · other agents via exports
Review the pair. Improve the human.REAL SESSIONS / EIGHT SKILLS / YOUR PROGRESS
YOUR SITUATION / YOUR NEXT MOVE
What do you want to change about working with AI?
Choose a situation you recognize. See what you could do differently, then explore the feedback behind the recommendation.
The change shipped. I’m less sure whether I framed, checked, and steered the work well.
See which collaboration behaviors deserve attention before your next task.
- Eight observable behaviors
- Feedback about your actions
- A next step tied to a session
Anonymized demo · real plugin output
Open full report Your report, in practice
Start with the latest feedback. Then open the axes over time to see which habits recur.
Opening the actual report…
Read-only curated demo with English translations of anonymized feedback and session excerpts. Explore an action for your next task. Choosing a scenario does not assess your skills; no log upload needed.
01 / THE BLIND SPOT
You review your code. Who reviews how you work with AI?
Diffs and tests show what changed. Your session history shows how you got there — including the habits you repeat without noticing.
- Vague tasks with no definition of done.
- A context window full of unrelated work.
- An “all done” report accepted without checking.
- Product decisions quietly handed to the agent.
- Another “fix it” with no new information.
- Changes too large to confidently roll back.
A feedback loop for the person at the keyboard.
Capture → Analyze → Coach → Improve. Repeat with your next period of work.
Analyze
Collectors extract facts and quotes. A separate judge assesses eight observable collaboration behaviors.
Coach
Get one to three concrete actions grounded in the places where your habits showed up.
Improve
Review the next period. Check earlier recommendations against new sessions and follow your own progress.
Eight skills. Observable behavior.
A retrospective on how you lead an agent. Not a test of your intelligence, seniority or prompt-writing style.
Verifying results
Define what done means. Run a meaningful check and read its output before accepting the result.
Framing the task
Give the agent the goal, relevant context and constraints before the work begins.
Owning the decisions
Stay responsible for the choices. Question tradeoffs, respond to warnings and explain a refusal.
Changing course after failure
Bring a new fact to the next attempt. Change the approach when repeating it stops helping.
Managing context
Supply relevant files, separate topics and save decisions before starting a fresh context.
Breaking down the work
Choose useful units of work, with clear boundaries and a check between steps.
Calibrating trust in the agent
Match your scrutiny to the cost of a mistake. A migration deserves more care than a variable rename.
Keeping changes reversible
Keep a return point and make risky changes in steps you can undo.
04 / THE EVIDENCE
Every score needs evidence.
A conclusion must point to a verifiable excerpt and its location in the session. The scripts check quotes against the log.
- Collectors record facts and citations, without scores.
- A separate judge assesses the collected evidence.
- No visible behavior means “insufficient data”, never an automatic low score.
- Your profile shapes the wording of feedback, never the score.
- Past scores stay out of the next assessment.
ASSESSMENT CONTRACT
“Build a login form: email and password, errors below the fields, no new dependencies.”English translation of a synthetic Russian test-session excerpt.
SYNTHETIC TEST SESSION · TURN 0
Framing the task3 / 5 · demo assessment
The first message supplies a goal, boundaries and acceptance criteria. This is the fixture Judge’s reading, not a customer result.
1. Facts + citationsCollector → Judge
2. Quote verificationOriginal log
3. No evidenceInsufficient data
One review is useful. Progress is the point.
Recommendations keep their identity across periods. Later sessions can show an action followed, partly followed, not followed, not visible, or too early to judge.
Choose a next step
For example: ask for the check and its output before accepting a change.
Look for the behavior
The judge checks new sessions recorded after the recommendation was received.
See the pattern
Follow the verdict history and your own progress. Missing evidence stays missing evidence.
Illustrative workflow, not customer outcomes.
Actual reports. Different perspectives.
Actual plugin output with fictional names. The English demo is a curated, translated selection of anonymized reviews; scores and recommendation statuses are preserved.
Demonstration data · not customer outcomes · Select a report to expand.
For your practice. Your team. Your classroom.
For individuals
Understand your own habits. In self mode, you are the reviewer and the feedback is ready immediately.
Explore the workflow →For teams
Use a shared language for constructive coaching. A reviewer approves feedback before it reaches each person.
Explore a team pilot →For educators
Review a cohort across modules, follow recommendations and approve individual feedback before sharing it.
Discuss a training pilot →LOCAL FIRST / HUMAN LED
Your sessions are yours.
Collect, sanitize and generate reports locally. Analysis uses your configured AI provider. You decide what to collect, retain and share.
Read the data boundaries →08 / AVAILABILITY
A local tool. A visible feedback loop.
MIT-licensed Claude Code plugin. Your model access powers the analysis; reports stay in your workspace.
Documentation →GitHub / MIT ↗
/plugin marketplace add Serg1kk/retropair/plugin install retropair@retropair/retropair:setup
09 / FOR TEAMS
Better AI adoption starts with better human skills.
A shared language for engineering enablement, developer coaching and AI-development training. Managed team offering — exploring early access.
Request a team pilot →Questions before you start
Does RetroPair review the code?
No. It assesses the person's observable collaboration behavior in agent-session logs. Code quality, model intelligence, speed and token efficiency are not the primary assessment target.
Which coding agents does it support?
Claude Code is the native plugin platform. Claude Code JSONL sessions are collected directly. Codex, Cursor and other tools use supported SpecStory or manual Markdown exports; there is no native Codex or Cursor plugin in version 0.2.0.
Does all processing stay offline?
No. Collection, sanitization, state and HTML report generation run locally. The analysis agents read sanitized context through your configured Claude Code model provider, under your account and its terms.
What if the logs do not show a behavior?
That axis is marked “insufficient data”. Work may have happened in another terminal or outside the session. Absence of a trace is not evidence of failure.
Can I compare developers?
RetroPair is built for personal development, not rankings. Group reports help a reviewer see coaching needs; progress is compared with the same person's earlier periods.
Is there a hosted team platform?
Not currently. Self, team and course modes exist in the local plugin. A managed team offering is being explored through early-access conversations.
What does it cost to run?
The plugin is MIT-licensed. Analysis uses your existing model access and can consume paid usage or subscription limits. Cost and duration depend on log volume, model selection and effort; no fixed per-review price is promised.
Review the pair. Improve the human.