Repository navigation
os-dev.md: a project-thread session's harness writes an HTML-comment attribution line as PR body line 1, against 〈干净收尾〉's ban on HTML comments — name the exemption #25220
Description
Source
This was out-of-scope finding 2 of PR #25214 (card #25185). Contract review 6341183860 flagged it, and its round 2 record 6341480064 repeated it. The review sent it to the skills lane, as a card the seat files at ACCEPT.
Finding
The session harness of a Claude Code project thread tells it to begin every PR body with an attribution block. The block's first line is an HTML comment (claude-projects-attribution, carrying the requester's GitHub login).
.claude/agents/os-dev.md forbids HTML comments in GitHub text, at lines 18 and 322–323 (〈干净收尾〉). Its reason is that the sanitizer eats short angle-bracket fragments.
So every card thread breaks one of the two rules. So far it has kept the harness line and reported it as a deviation (PRs #25192, #25212 and #25214 all carry it). A relay read-back on an earlier card thread's PR found the line kept byte for byte.
Ask
Add a one-line exemption to os-dev.md, shaped like the existing one for trailers the harness writes itself (line 286: report it, do not copy or rewrite it):
- the harness-required attribution line on PR body line 1 stays exactly as the harness gives it, and the dev reports it;
- the ban still covers every HTML comment the dev writes itself.
The closing keyword moves to line 4 because of it; say that too, or point to the rule that already allows it.
Acceptance
os-dev.mdnames the exemption in one line; nothing else in the ban is loosened.- Line growth is paid by compression under
check:pm-skill-ratchet, with no ceiling moved. - The
.claude/**Tier S path applies: a contract-review-tier PASS on the head, then the owning seat lands it.
Activity
objectstack-fleet commentedon Oct 22, 2026
Handover from
domain:skillsseat 1 (seat post #7623,session_0165fj9zYfUrt6ckcUgYBD5a, 2026-10-22): the maintainer's project that held this seat stops owning the skills lane after its in-flight cards (#25115, #25216) land, so this card stays in the lane queue for the next seat. Order of work:Clause-②:line, so a dispatched PR never writes theMetadata-protocol:line ruling A (#25105) adds #25115 (PR docs(pm-dispatch, os-dev): the claim template and the dev definition carry theMetadata-protocol:line besideClause-②:#25240, ready + auto-merge armed 14:43Z): it editsos-dev.md's claim-copy rule. Start fromorigin/mainafter that merge.os-dev.mdonly — ONE line in 〈干净收尾〉 shaped like the harness-trailer exemption (near L286: report it, do not copy or rewrite it): the harness-required attribution line on PR body line 1 stays byte-for-byte and the dev reports it; the ban still covers every HTML comment the dev writes. Say thatFixes #<n>then sits on the first line after that block, or point to the rule that already allows it. Paid by compression in the same file; no ceiling moved.Metadata-protocol:line besideClause-②:#25240's contract review (6343268424), same file, ratchet-neutral reword (no new line): 「两行同抄进每个 changeset 正文行首,Check Changeset读:有 kind 即minor,none即patch。」 should say the gate reads the two lines off the PR body (check-changeset-no-major.mjsL1818 / L1843) and the changeset-body copy is AGENTS.md Post-Task Checklist ③'s publishing rule — e.g. 「两行同抄进 changeset 正文行首;Check Changeset读 PR 正文:kind 即minor,none即patch。」 (≤120 bytes).CONTRACT_REVIEW_TIER(os-dev.mdis protocol semantics perdispatch-gates.mjs --tier), effort medium. Tier S.claude/**:CONTRACT_REVIEW_TIERPASS on the head, then the owning seat lands it.Generated by Claude Code