Skip to content

Owner question: accept these picks where your agent rule files disagree? (G5, #874) #910

Open

Description

@scuttletoby

Recommendation: accept all the picks below. Reply "Owner decision: accept all", or "accept all except D2 (…)".

Background: PR #874 (G5) writes the general agent rules once, in docs/fork/agent-rules.md. While gathering them, the agent found places where your rule files say different things about the same rule. Once you decide, the trim plan removes the extra copies, which saves about 3 KB on every agent call. These are the picks.

# Where the files disagree Pick (recommended)
D2 Fable's scope. One file says "design and architecture options, creative work, root causes Opus has not found, stuck audits". Another says "only design options and stuck audits". Use the longer one (your newest model decision, 11-13).
D2b Effort for reviewing critical code. One memory says extra-high effort; the review rules and review kit say two independent Opus reviewers at high effort. Two Opus reviewers at high effort (what the 40-review test measured). The model-guard policy file becomes the one source.
D3 What counts as critical code. One file says "data loss, deploy, security"; another says "merge path, deploy, security, data". The old "never Sonnet as the only reviewer" line is out of date, since Sonnet is no longer used. The review skill's definition is the only one; other files link to it.
D5 Format of owner questions. Each project words it a bit differently. The fork's format everywhere: one issue per question, recommendation first, "Owner decision: …" to answer. Each project only says which repo.
D8a What an agent does when the machine is too full (capacity check says NOT OK). "Re-check every few minutes", "wait and re-check", or "end the turn and retry on the next wake-up". End the turn and retry on the next wake-up. No sleeping loops: they waste usage and hold a session open.
D8b Memory needed before a writer starts. Your 11-21 decision says 1.5 GB for writers that never build Java; one older memory still says 3 GB for all writers. Your 11-21 decision.
D19 How to wait for CI. One memory says "background commands only". T3's built-in pull request watch. It is the newer mechanism, and the agent is woken when checks finish.
D0 Where the shared rules are linked from. The GitHub link to the doc on main. No local copy of the fork stays on main.
D1 Server access. Two files give different VM ranges, and one says the browser is "the" access route. The live server list decides the range. The command-line helper comes first; the browser is the fallback.

Before anything is removed, I will check that no installer writes a removed file back.

🤖 Generated with Claude Code

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions