[alexiocassanifm]

What changes for a participant

Every step of the five handouts now says what to type in the terminal, what to paste in Claude Code and what to expect. Where the lesson is a design choice (the loop's stop, its interval, where the review writes), a Variant prompt changes that one thing so the point still lands. The merge policy in exercise 5 stays a team discussion, because the decision is the exercise.

Windows runs natively: py workshop …, Git for Windows (Claude Code runs its shell through Git Bash), no jq, export or \ line continuation in any handout.

Behaviour changes

• workshop events <file> replaces the two jq pipelines of exercise 1 and also reads a UTF-16 log (a Windows PowerShell 5.1 > redirect).
• workshop hook / log-tool replace the hand-written hook. The hook goes into the gitignored .claude/settings.local.json, calls the interpreter by absolute path, and reset removes it.
• .claude/settings.json (new, committed) pre-approves the test, git and read/comment gh commands. Measured: without it, the exercise 1 headless run had every test command it tried denied. It does not pre-approve merges, deletes or raw gh api calls, which still ask.
• Claude Code ignores that file until the folder is trusted, and /goal is not available until then either, so README gains a trust step and doctor checks it (parent folders included, since trusting a folder trusts what is inside it).
• /goal starts working as soon as it is set, so exercise 3 run 1 is one paste.
• A fixed-interval /loop cannot stop itself, but a self-paced one can, so Loop A is self-paced. The fixed-interval version is now the "forgotten stop" variant.

⚠️ Before merging: the starting branches must be rebuilt in the same sitting

reset replays the template's main → exN-start diff. If this merges while ex2-start, ex3-*-start and ex4-start are still the old builds, reset ex2/ex3/ex4 rolls this PR back on every participant's copy. After merging, run the FACILITATOR recipe (solutions merges main, build_branches.py, verify.sh, force-push the four starting branches). The companion PR into solutions adds a verify.sh check that fails in exactly that case: rehearsed, it is green on rebuilt branches and fails on the stale ones.

Tested and untested

• Tested: exercise 1 end to end with a real claude -p run (0 denials, hook log matching the stream), events on UTF-8 and UTF-16, reset removing the hook, pytest, verify.sh ALL VERIFIED on rebuilt branches.
• Not run live: the ex2–ex5 prompts. Their wording follows the current docs for /loop, /goal and routines.
• Not tested on Windows at all. Known risk: exercise 3 run 2. The plugin's loop gate calls python3 from Git Bash, and a python.org or winget Python only answers to python/py. doctor checks it on Windows, and README says so.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UMGFBjF5h1nKda8qUaoRVV