Skip to content

REVIEW.md

Latest commit

 

History

History
72 lines (61 loc) · 3.66 KB

REVIEW.md

File metadata and controls

72 lines (61 loc) · 3.66 KB
Raw

Maintainer run-through record

2026-09-13 — content rewrite: CLI course → Codex desktop app course

  • Folder renamed hands-on-with-openai-codex-cli → hands-on-with-openai-codex; course_id (course-codex-cli) and claim salt preserved so existing claims and codes remain valid.
  • COURSE.md/COURSE_cn.md and all ten lesson files rewritten against the Codex app as documented at developers.openai.com/codex (thread model, approval modes, AGENTS.md instruction chain, diff review). CLI-specific teaching was removed; TASK.md, tests/, scenario/, starter/, solution/ are byte-identical to the Claude Code course core (shared-core group report-tool, enforced by tools/verify_courses.py).
  • Not done in this pass: a live teaching run-through inside the Codex desktop app. No Codex account/app session was used; the content is authored from official documentation and the previously verified task core. Teaching quality in the real app remains pending below.

2026-09-12 — contract verification (mechanical)

  • Environment: macOS (arm64), Python 3.13; repository clean checkout on branch feat/0.0.3-courses-and-radar.
  • The code core is the reviewed Claude Code course core reused per the 0.0.3 plan §2.2 ("reuse C1 skins"): same starter/solution/tests contract.
  • Commands and results:
    • python verify.py starter --expect-failure — exit 0; all seven expected failure names reproduced.
    • python verify.py solution — exit 0; 9/9 tests pass.
    • All three skins exercised end to end by the shared suite.

Pending before this rewrite can be called taught-in-tool

  • One full agent-taught run-through inside the Codex desktop app ("start lesson 1" through the Lesson 5 checkpoint), recording observed deviations from COURSE.md here; the recording doubles as demo-video source material.
  • Whether a free ChatGPT tier lasts the whole course is unverified.

2026-09-13 — 0.0.8 rework: lesson 1 is now tool → Skill → agent fetch (FP-822)

  • L01 (EN+ZH) rewritten around the fixed order: get the tool running → install the FlyPython Skill (network access included) → let the agent fetch this course's files via the files endpoint. All "download this folder" wording removed from COURSE.md and lessons.
  • The learner's check command across lessons is now the bare python verify.py (FP-820): per-checkpoint status + claim codes, non-zero exit while gated checkpoints are open. solution is no longer presented as a completion standard.
  • Tool-specific Skill install and network steps are authored from the same official documentation as the course; not yet exercised inside the tool — the pending live run-through below still stands.

Deviation log

2026-09-13 — agent solvability run (challenge mode, not taught mode)

  • Agent: Devin (SWE-2 Max), CLI session on macOS arm64, system Python 3.14 (stdlib only — this course needs no third-party deps).
  • Method: implemented starter/report_tool.py from TASK.md and the test expectations only; did not read or copy solution/.
  • Result: python verify.py starter exits 0 (all nine tests); verify.py progress printed all five claim codes (l03/l04 [passed], l01/l02/l05 [attest]).
  • Observed deviations: none — contract landed on the first pass. This course shares its task core (TASK/tests/scenario) with hands-on-python-with-claude-code, and the same implementation passes both suites.
  • Limitations: challenge-mode run (TASK.md → tests), not the taught COURSE.md walkthrough — teaching quality remains separately pending.
  • Starter restored to the deliberately-unfinished state after the run (--expect-failure verified again).