[OmarAlJarrah]

Summary

#1532's review ran past seven rounds, and #1542 traced what kept it going: defects in the previous round's own fixes, cells a reader named instead of reaching, slow commands, and calls that batched nothing while the main session's context grew. This puts each of #1542's ten Direction steps into the review and shipping skills as rule text, with no new script.

• Readers (morphic-reviewing-prs/references/readers.md, new):
  • each delegated reader starts at effort xhigh, set on the call that starts it (start-readers-at-xhigh), gets a tool-call budget summed from its cells, and reaches every cell or says why it cannot (brief-a-reader-to-its-cells);
  • its brief has it batch independent calls and keep bulky output in files (reader-batches-its-calls), and bound slow commands by what they walk and build trial merges by the fast form (reader-bounds-its-commands);
  • every harness the reviewer reruns sits under one fingerprinted directory (one-harness-directory);
  • a round reading what moved that carries whole reads splits between two parallel readers (split-whole-reads-from-what-moved), and no round divides its files between readers (no-split-by-file).
• Rounds (rounds.md):
  • before each hand-off, the author reruns the plants a fix can have changed, derives each body number from a command and reruns the trial merges (author-pass-before-hand-off, cited from review-own-fixes);
  • the main session compacts only at a clean round boundary and rereads its state afterwards (compact-at-a-round-boundary);
  • the report format gains each reader's minutes and tool calls.
• Claims (claims-and-process.md#run-prescribed-commands): a prescribed command also runs with each placeholder holding the hardest value it admits.
• Shipping: a PR's scope settles before its second review round (commits-and-prs.md#scope-settled-by-the-second-round), and a trial merge is built as a worktree of a git commit-tree commit (merging-main.md#trial-merge-by-commit-tree, new) and confirmed by both its parents from the worktree that built it (merging-main.md#trial-merge-happened, which cites the build, so both trial-merge Probes reach the two through their citation of it).
• Two sites of the mechanisms these rules use: closing-issue-read-whole named a PR's issues from closingIssuesReferences, which a stacked PR leaves empty, and now names each issue a closing keyword in the body names, as the new scope rule does. verdict-outlives-unchanged-files' Probe diffed only the mutated file and the test, though its Rule also keeps the fixtures and the files between them; it now diffs all four and fails a path that matches no file, and the rule covers a commit as well as a merge, as author-pass-before-hand-off applies it to a round's fix commit.

Where this departs from the Direction as drafted:

• Steps 1 and 7 name fresh-reader-for-the-second-round and reader-owns-its-directory, at 150 and 149 of their 150 words on main, so each lands in a new rule the named one reaches by citation. Step 2 goes to brief-a-reader-to-its-cells, a rule naming what a brief holds, as its Direction allows. Step 6, which asks for one rule or an edit of reader-owns-its-directory, lands in two, reader-batches-its-calls and reader-bounds-its-commands, as one would pass 150 words.
• Step 4's "rerun the whole mutant set after any test edit" becomes a citation of verdict-outlives-unchanged-files, which already names the plants a change can alter, a test, its fixtures and the code between a plant and its test included, rather than every plant.
• Step 6's fast trial merge is written once, in trial-merge-by-commit-tree, which trial-merge-happened cites, rather than restated in each trial-merge Probe.
• Step 8's scope rule exempts a stacked child's issues that a PR landing the stack carries, which closing-keywords has it add whatever the round.
• fan-out-an-angle-to-a-reader and SKILL.md now name a split round's two readers, and fresh-reader-for-the-second-round and later-rounds-read-what-moved allow them, which all four would otherwise contradict. fan-out-an-angle-to-a-reader also briefs each reader it starts by brief-a-reader-to-its-cells, so a requested fan-out reaches the effort, batching and bounding rules too.

The reader rules move from rounds.md into readers.md, the four already there with the new ones, as rounds.md would otherwise pass its 4,000-word reference cap. Each keeps its key, and every citation follows it.

Closes #1542

Test plan

• python3.14 .claude/skills/morphic-curating-skills/scripts/skills_lint.py is clean. A citation of a moved rule planted at its old file draws references/rounds.md holds no rule reader-owns-its-directory, so the lint holds every moved citation.
• A per-rule diff against main, keyed by rule rather than by file, shows each moved rule unchanged bar the edits above and its citations' file.
• Each command an edited rule prescribes ran at the head with a hard value in its placeholders, exiting 0 where the case should pass and non-zero where it should fail:
  • the trial merge with <dir> holding a space, brackets and a plus, a conflicting merge stopping its chain, a <dir> still holding an earlier merge, of ci: minimize a fuzz input for 1s, not 10s #1385, which is stacked on ci: move the fuzz search from pull requests to a nightly job #1382, or of the same PR at an earlier head: --is-ancestor passed the first and a second-parent check the second, and the both-parents check refuses each; and a merge built from a <head> of HEAD, which the confirm accepts from the building worktree and would refuse inside <dir>;
  • the harness Probe over such a directory, with its list kept outside it, silent across reruns and naming the one file an edit changed, following a directory named through a symlink and a harness file linked from elsewhere, and failing on a missing directory;
  • the batching Probe on a reader transcript at such a path, reproducing skills: make review rounds reach every cell and need fewer rounds #1542's call counts for chore(skills): scope later review rounds, isolate their readers #1532's rounds 5 to 7 (64, 48 and 47);
  • the effort Probe on the starting session's transcript, printing xhigh for each reader this PR's review started and unset for each of chore(skills): scope later review rounds, isolate their readers #1532's seven, whose own transcripts nonetheless record xhigh;
  • verdict-outlives-unchanged-files' Probe on a glob of fixtures, a path matching nothing, a file the head deleted, and files named a[1].py and b+c d.py: the glob diffs its files, the unmatched path and the deleted file fail git ls-files --error-unmatch, and the hard names pass;
  • the scope Probe's userContentEdits(last:1) query on this PR, chore(skills): scope later review rounds, isolate their readers #1532, a stacked PR and a PR whose body was never edited: the node it returns carries the PR's creation time, and an empty list leaves the body itself;
  • gh issue view as closing-issue-read-whole runs it, and later-rounds-read-what-moved's git diff --name-only and git grep -n '#<key>'.
• Trial merges by the new trial-merge-by-commit-tree and trial-merge-happened against every open PR: 15 merge cleanly, each with both heads as parents, a clean lint and no citation of a moved rule at its old file. The other 12 conflict, each already conflicting with main and none in a file this change touches. The whole run took about 11 seconds.