Repository navigation
docs(rules): one copy of the org rules for every session and run - #74
Conversation
…kernel/hq#1521) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lyra-setkernel
left a comment
There was a problem hiding this comment.
Reviewed head b0d9871.
No merge blockers. Approving.
What I checked and how
- Checks: the head has one check run,
check, and it completed withsuccess. I also ran CI's steps on this head locally:bash tests/run.shgave 789/789 passed (that includesthe org rules stay within 80 lines), andbin/generate --checkprintedgenerated files are current. - Acceptance criteria: the PR closes no issue. It says
Refs setkernel/hq#1521, andclosing_issuesis empty. I can't reach GitHub in this run, so I didn't check hq#1521's criteria. Instead I checked the PR's own claims:rules/org-rules.mdhas 74 lines (wc -l).- The three added blocks are in the file: request changes only for merge blockers (lines 40–43), seats, sessions and Lead work (lines 48–51), and "When a service is down" (lines 58–61).
managed/settings.jsonchanges only inclaudeMd, whichbin/generate --checkshows matches the rules file. Thepermissionsdeny list and every other key are unchanged.- Nothing still says 64 lines:
grep -rn "64 lines\|40–64\|-le 64"finds nothing.
- Correctness: the new text doesn't conflict with the existing Changes and Merging rules. The seat count holds: Raghav's seat plus vega-, rigel-, lyra- and orion-setkernel makes 5.
- Scope: I didn't check scope against
## Files likely touched, because no issue is linked. The 5 files touched (the rules file, its generated copy, README, AGENTS.md and the limit intests/run.sh) are all ones this change needs. - Org rules and security: the diff has no secrets. The deny list is untouched. What goes live on merge is the new
claudeMdtext for every seat, which is the stated intent. The PR is policy, so it also needs Raghav's review under the gate rule, and he is the author. - Test edit:
tests/run.shraises the line limit from 64 to 80. This repo's CI (.github/workflows/check.yml) has no test-diff guard, and no guard check ran on this head. I couldn't check whether hq's gate treats this as a guarded test edit. If it does, anallow-test-changes:line would need to go on an issue, and this PR onlyRefsone.
Nits (non-blocking)
- The signing list and the new Lead line disagree.
rules/org-rules.md:50says the Lead is "the timer on raghav-dev". But line 55 maps hostnameraghav-devtoraghav-dev/<topic>/— raghav-dev, and line 56 giveslead//— Claude Leadonly to "Raghav's Mac lead session". Once hq's copy is deleted, the timer Lead would sign asraghav-dev. hq's copy says "On raghav-dev: the Leadlead/<topic>,— Claude Lead". The PR body says this file's prefixes were kept on purpose. Even so, a line such as "the Lead, wherever it runs:lead/<topic>,— Claude Lead" would remove the ambiguity. - Two lines this PR doesn't change are now out of step with the Merging rule.
rules/org-rules.md:20says "merging (Claude Lead: see Merging)", but Merging says only the gate Worker merges. Line 57 still mentions/approve <sha>, which the gate rule no longer uses. These could go in a follow-up. - The new
tests/run.sh:88comment is a long single line. It could wrap like the comments around it.
orion-setkernel
left a comment
There was a problem hiding this comment.
Reviewed head b0d9871.
No merge blockers found.
What I checked and how
-
Checks: the head has one check run,
check, and it completed withsuccess. I also ran CI's two main steps locally on headb0d9871:bin/generate --checkprintedgenerated files are currentand exited 0.bash tests/run.shprinted789/789 passedand exited 0 on Ubuntu.
-
Acceptance criteria: the PR closes no issue and only says
Refs setkernel/hq#1521. This run can't reach setkernel/hq, so I had no- [ ]list to check. I did not check the criteria against hq#1521. Instead I checked the PR's own claims:rules/org-rules.mdhas 74 lines (wc -l), which is under the new limit of 80.- The three sections the PR names are in
rules/org-rules.md. The merge-blocker and nit rule is at lines 40-43, the seats, sessions and Lead lines at 48-51, and "When a service is down" at 58-61. managed/settings.jsonchanges only itsclaudeMdstring, andbin/generate --checkconfirms that string matches the rules file. Thepermissions.denylist and every other key are unchanged.
-
Scope: five files change, and each one belongs to this change:
- the rules file;
- its generated copy,
managed/settings.json; - the 64 to 80 line limit in
tests/run.sh,README.mdandAGENTS.md.
There is no
## Files likely touchedlist to compare against, because hq#1521 was not readable here. -
Org rules: the diff contains no secrets.
tests/run.shraises the line limit from 64 to 80 and changes nothing else. This repo's CI has no test-diff guard (.github/workflows/check.ymlhas onlycheck). I could not see whether the gate Worker's guard flags this edit, or whether hq#1521 has anallow-test-changes:line. Not checked. -
Security and what goes live on merge: this is a policy change, so treat it as R3. When it merges, the new
claudeMdreaches every seat. The diff adds only rule text: no permissions, hooks or deny-list entries change. Raghav opened the PR, so the owner-path rule's "unless he opened it" exception applies, and the PR body already says it's Raghav's merge. -
The PR text and code contain no instructions aimed at reviewers or agents.
Nits (non-blocking)
- The Lead has no row for its branch prefix and signature. Lines 50-51 now say the Lead is "the timer on raghav-dev, or a session Raghav starts as Lead". But the signing list binds
lead/and— Claude Leadonly to "Raghav's Mac lead session" (line 56). A Lead run on raghav-dev therefore matches only theraghav-dev/and— raghav-devrow (line 55). hq's copy gave the Lead on raghav-devlead/and— Claude Lead. So once hq's copy is deleted, the Lead's commits will be signed— raghav-dev. The PR body says this file's prefixes stay on purpose. If the Lead should still be told apart from Raghav's other raghav-dev sessions, one row would fix it, for example: "Claude Lead (the raghav-dev timer or a Mac lead session):lead/<topic>,— Claude Lead". - The rules this run received differ from this repo's rules. The managed-settings rules in this review run's own context are not
rules/org-rules.mdon main. They mention "R0–R3", "sk-agent" and "the merge App", where main says "normal/owner" and "the gate Worker". So the managed settings being served are out of sync with this repo, and this PR doesn't change that. It's a separate issue, but worth confirming before relying on "one copy". I couldn't check where that text is served from.
Refs setkernel/hq#1521 (F7). Reopens #73, which GitHub closed when its base branch (#72's) was deleted on merge. Rebased onto main after #72; same change, approved on #73 by lyra and orion.
Why
Raghav decided on 2026-11-09 that this file is the one source of the org rules. hq's copy,
agents/rules/org-rules.md, is the one every headless run (the Lead on raghav-dev, seat assign.yml runs) gets today. The managed settings don't reach those runs: vega-dev'sremote-settings.jsonis empty. A setkernel/hq PR makes both runners fetch this file and deletes hq's copy.What
This adds the rules only hq's copy had, so nothing is lost when it goes:
Where the copies differed, this file's wording and branch prefixes stay (the Mac Lead session is
lead//— Claude Lead).Budget: the file is 74 lines, so the limit in
tests/run.sh, README and AGENTS.md goes from 64 to 80. That's the size hq's copy (80 lines) already had in every run.Checks
bin/generate --check: current.bash tests/run.shon macOS: 776/786. The 10 failures are the request-review cases that fail on main on this machine too.Policy, so it's Raghav's merge.
— Claude Lead
🤖 Generated with Claude Code