Skip to content

docs(prose-tests): the record catches up with day two - #588

Merged
leeovery merged 1 commit into
mainfrom
docs/prose-tests-day-two
Jul 27, 2026
Merged

docs(prose-tests): the record catches up with day two#588
leeovery merged 1 commit into
mainfrom
docs/prose-tests-day-two

Conversation

@leeovery

Copy link
Copy Markdown
Owner

Summary

Three documentation surfaces brought current with the landed programme — one of them was actively wrong.

  • CLAUDE.md § Prose Tests still instructed "runs until state lands" — the rule P9 overturned (test(prose): give the eight read-only cases something that can fail #565): a handoff is as deterministic an ending as a delta, and "until state lands" is exactly what would push short entry walks into needing conversation stubs. Now states the current rule, adds calls_in_order to the invariants list, and names conduct and world: "claims" in the case anatomy — so an agent working from CLAUDE.md alone authors to the contract as it is.
  • tests/prose/README.md gains a Playing the User section (answers vs conduct, the arrival-order hazard that forbids mixing them around an interview, pin-choices-with-invariants, the P4 boundary extended to roleplay) and rewritten files[] guidance from what measuring it taught: harvest from walks not link-crawls, ratchet across runs, opened ≠ walked.
  • design/prose-tests.md gains P11 — conduct as contract, not convenience — and the day-two Log entry: the closing-turn race, writes-as-evidence, the whole-phase quick-fix case, defects five–seven with the recurring pattern named (a guard keyed on artifact content an earlier step writes is suspect), and the forty-PR landing.

Deliberately excluded: the CONVENTIONS.md guard rule — that's a convention decision awaiting Lee's call, not documentation of something already agreed.

Test plan

  • Prose deterministic gate + conventions lint: 50/50
  • Docs-only; no behaviour, no simulation impact

🤖 Generated with Claude Code

CLAUDE.md still told a case author to run until state lands — the rule
P9 overturned when a handoff proved as deterministic an ending as a
delta, and the exact instruction that would have pushed short entry
walks into needing conversation stubs. It now states the current rule,
lists calls_in_order among the invariants, and names conduct and claims
mode in the case anatomy, so an agent working from CLAUDE.md alone
authors to the contract as it is.

The prose-test README gains the two pieces day two added and one it
never had: a Playing the User section covering answers and conduct —
the arrival-order hazard that forbids mixing them around an interview,
the pin-choices-with-invariants rule, and the P4 boundary extended to
roleplay — and the files[] guidance rewritten from what was learned
measuring it: harvested from walks not link-crawls, a ratchet across
runs, opened-is-not-walked.

The design log gains P11, naming conduct as contract rather than
convenience, and the day-two entry: the closing-turn race, writes as
evidence, the whole-phase quick-fix case, defects five through seven
with the recurring pattern named — a guard keyed on artifact content an
earlier step writes is suspect — and the forty-PR landing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@leeovery
leeovery merged commit dd2a5ee into main Jul 27, 2026
@leeovery
leeovery deleted the docs/prose-tests-day-two branch July 27, 2026 18:54
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