csvstat controlled handoff sample

Scope
The historical run on 2026-10-20 used Claude Code 2.1.294 and Codex
0.153.4, directly through their CLIs in the same directory. The first
agent deliberately stopped after steps 1–2. This was not actual quota
exhaustion, not a Windows run, and not an Orkas UI end-to-end test.
The supplied final files were replayed on 2026-11-05 without model calls.
The published verification log and its screenshot show that replay only.

Repeat the build
1. Copy TASK.md and sample.csv into a new empty directory.
2. Run Claude Code there. Ask it to implement only steps 1–2, add tests,
   write PROGRESS.md (done, left, decisions and checks), and stop.
3. After it stops, run Codex in the same directory. Ask it to read both
   Markdown files, preserve the decisions, finish steps 3–4, and verify.
4. Compare the outcome with final/. handoff/PROGRESS.md is the historical
   first-stage note recovered from the archived diff, not a fresh model run.
   The task and historic notes are in English; prompts can use your language.

Orkas configuration
Install and sign in to both CLIs. Connect them as External agents.
In AI Team > agent > Project directory > Change choose the same sample
directory for sequential handoff, then start a new conversation.
Run the two stages sequentially. Leaving the model blank uses CLI defaults.
Changing a directory is local to this device. An existing conversation
retains its effective directory; start a fresh one after changing it.

Parallel work is a different route
Use separate Git worktrees for independent tasks. Initial commit, clean
working tree and unused new branch names are prerequisites:
git worktree add -b feature/claude ../feature-claude
git worktree add -b feature/codex ../feature-codex
Commit the agreed specification first. Each worktree has separate files;
same-named notes do not synchronize. Give the integrator explicit paths.
Stop both agents, review and commit each branch before merging into a clean
integration branch, resolve conflicts and run the combined checks.

Replay the supplied final result (Python 3.9+; standard library only)
cd final
python3 -m unittest -v test_csvstat
python3 csvstat.py sample.csv temp
On Windows use python (or py -3) if python3 is unavailable.
Expected successful output:
count: 3
min: 3.0
max: 21.0
mean: 14.0
python3 csvstat.py sample.csv humidity
Expected: exit status 1, column-not-found error on stderr, no traceback.

The two providers have separate login, billing and available limits.
Switching agents or importing a session does not increase either quota.
No API keys, raw conversations or machine-private paths are included.
