Skip to content

Facilitator: verify the exercise 1 helpers, and that reset never rolls main back - #2

Open
alexiocassanifm wants to merge 4 commits into
solutionsfrom
kit/guided-prompts-solutions
Open

alexiocassanifm wants to merge 4 commits into
solutionsfrom
kit/guided-prompts-solutions

Conversation

@alexiocassanifm

Copy link
Copy Markdown
Contributor

Companion to the handouts PR into main (its commits are included here through the merge).

  • verify.sh drives workshop events (UTF-8 and UTF-16), the hook command workshop hook writes (successes and PostToolUseFailure), and checks that nothing is left for git to commit.
  • verify.sh fails when a starting branch would roll main's files back: reset replays the old main → exN-start diff. Rehearsed in a scratch clone: green on rebuilt branches, and VERIFY FAIL: reset ex2 rolls main's files back on the current ones.
  • The guide says the branches are rebuilt in the same sitting as every change to main, and what is untested on Windows.

Order: merge the main PR, merge this one, then python3 tools/build_branches.py, verify.sh, and force-push ex2-start ex3-goal-start ex3-loop-start ex4-start.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UMGFBjF5h1nKda8qUaoRVV

alexiocassanifm and others added 4 commits September 22, 2026 22:20
Each step now says what to type in the terminal and what to paste in Claude Code,
with 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.

Windows runs natively: `py workshop`, Git for Windows for Claude Code's shell, no
jq in any handout. `workshop` gains `events` (reads the stream-json log, UTF-16
too) and `hook` / `log-tool` (the exercise 1 hook, with no jq). A committed
.claude/settings.json pre-approves the test, git and gh commands: without it, the
exercise 1 headless run was denied every test it tried. Claude Code ignores that
file until the folder is trusted, so README gains the step and doctor checks it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UMGFBjF5h1nKda8qUaoRVV
…3 named

/goal starts working as soon as it is set, so the condition carries the task and
there is no second prompt to land mid-run. The allow-list keeps the gh verbs that
read or comment: pre-approving `gh pr *` also pre-approved `gh pr merge --admin`,
the act exercise 5 says only a prompt forbids. Exercise 5's duplicate filing
happens in the same copy, the only place its search can see. On Windows, doctor
checks the `python3` the loop gate calls from Git Bash, and README says exercise 3
run 2 is untested there.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UMGFBjF5h1nKda8qUaoRVV
…s main back

verify.sh drives `events` (UTF-8 and UTF-16), the hook command `hook` writes, and
`log-tool`, and fails when a starting branch built from an older main would replay
that main's files over a copy's. Rehearsed: green on rebuilt branches, and the new
check fails on the stale ones. The guide says the branches are rebuilt in the same
sitting as every change to main, and what is untested on Windows.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UMGFBjF5h1nKda8qUaoRVV
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant