Skip to content

Load one instruction source in both harnesses - #43

Merged
raghubetina merged 1 commit into
mainfrom
claude/agent-instructions-both-harnesses
Sep 30, 2026
Merged

raghubetina merged 1 commit into
mainfrom
claude/agent-instructions-both-harnesses

Conversation

@raghubetina

Copy link
Copy Markdown
Contributor

Summary

  • Instruction loading.
    • Add CLAUDE.md containing @AGENTS.md, so Claude Code and Codex load the same instructions.
    • A documentation test keeps it a bare import.
    • The exact package allowlist keeps it out of the npm tarball.
  • AGENTS.md gains:
    • the check command and its cost;
    • a sibling-repo map;
    • when to get an independent review through cross-review (codex-review from Claude Code, $claude-review from Codex).
  • Review scope. Review covers what machine consumers depend on: the accepted API-contract range and Plan formats, command names and flags, handled error values, and exit statuses. It excludes human-readable detail text and progress lines.
  • docs/README.md notes that CLAUDE.md only imports AGENTS.md.

Part of Step 3 of the 2026-09-29 agent-documentation audit; the matching firstdraft change is firstdraft/firstdraft#910.

Docs: updated AGENTS.md and docs/README.md.

Validation

  • npm ci --ignore-scripts && npm run check: pass.
  • AGENTS.md is 1,552 of 2,048 bytes.

Review

cross-review codex-review, new session over origin/main..HEAD, with firstdraft's docs/review-focus.md as focus text. Verdict: Approve, no findings (session 26de7e04-ec1b-4079-a3d8-15dcea2f1ab0).

Codex reads AGENTS.md. Claude Code reads CLAUDE.md, and only some of
its versions and settings also read AGENTS.md. A CLAUDE.md that only
imports AGENTS.md makes both harnesses load the same text. A
documentation test keeps it a bare import, and the exact package
allowlist keeps it out of the npm tarball.

AGENTS.md also gains facts that fresh sessions had to rediscover: the
check command and its cost, which sibling repository owns the Service
API, the Plan format, and Skill packaging, and when to get an
independent review through cross-review from either harness.

That review covers what machine consumers such as the Skill depend
on: the accepted API-contract range and Plan formats, command names
and flags, handled error values, and exit statuses. Human-readable
detail text and progress lines are left out, because docs/errors.md
tells consumers not to branch on them. Naming all output and errors
would cover most CLI changes. In firstdraft, a trigger that broad
matched 188 of 250 PRs, and 97 of those recorded no review.
@raghubetina
raghubetina marked this pull request as ready for review September 30, 2026 05:35
@raghubetina
raghubetina merged commit bdc9aa0 into main Sep 30, 2026
4 checks passed
@raghubetina
raghubetina deleted the claude/agent-instructions-both-harnesses branch September 30, 2026 05:35
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