/irrifletti

irrifletti · verb · “to reflect”
Looks back at your recent agent sessions, privately, to show whether a change to the workflow helped.

Without measuring, you cannot tell whether a change to the workflow made runs better or worse.

Use it a couple of weeks after you change the workflow, to see whether runs got better. You get a private report comparing recent sessions with the last retrospective. It needs only a retro.output folder outside any public repository; the report never leaves your machine.

Agent instructions

The agent follows the instructions below, from commands/irrifletti.md. Its one-line description is: Measure how recent agent sessions went against the last retrospective, privately.

Full agent instructions

Measure how the agent sessions on this machine went, and compare with the last retrospective. Run it on demand (suggested: two weeks after a workflow change lands, and at the end of the Jev shadow period), never on a schedule.

Run from the mexxi clone (paths.checkout in ~/.config/mexxi/config.json, default ~/workspace/mexxi), adding to $ARGUMENTS only a window (--since <date>, --until <date>) or --exclude <session-id-prefix> for the session running this retro:

node scripts/retro/retro.mjs $ARGUMENTS

It reads Claude Code's transcripts, ibni's per-step timings, Jev triage agreement and review acceptance (same-family against cross-family reviewer), and writes retro-<date>.md and .json to retro.output. It refuses an output folder inside a Git repository unless that folder is gitignored there with nothing tracked in it, so reports never reach a public repository. If it refuses, report that and stop; do not choose another folder yourself.

Then read the printed goal-signals table and summarise what moved against the previous retro, which signals missed the targets in its Target column, and what the ibni numbers add. The report holds transcript-derived text: quote counts and signal names only, never snippets, and never copy the report into a repository, pull request or issue.