[ZhengGong-amd]

• Description: five standalone commits on top of review-pr: block on a premise that does not reproduce, a failure that survives, or a fix that patches the symptom #2227; please Rebase and merge.

  1. Ask what to delete (references/subtract.md).
     • Step 5 asks whether each added test is needed: is it a repeat, a plumbing pin, or one contract re-pinned at every layer?
     • Step 6's build-it-again also covers the mechanisms a fix patches, and prefers deleting or changing code over adding a guard.
     • Dead ends already in the touched code go on an optional, non-blocking Trim in this PR: list.
  2. Fewer words. SKILL.md goes from 365 to 286 lines (−14% words) and keeps everything review-pr: block on a premise that does not reproduce, a failure that survives, or a fix that patches the symptom #2227 added. An independent sentence-by-sentence audit found nothing dropped. Docs only.
  3. Risk and decisions.
     • risk.py grades the changed paths low, medium or high into risk.txt. Its only data is references/tiers.md: Tier 1 is high, Tier 2 medium, and a new risk table grades the rest.
     • The card's Risk: may only rise above that grade, and then says (raised from …).
     • Medium and high cards list Decisions for a human:, naming every operator-visible change and every high-risk path.
     • check-card.sh holds a card to the whole card contract. CI and local runs both call it.
     • pr-review-publish.sh keeps only the CI provenance checks. It re-grades risk from the PR's files on GitHub, so the review job cannot lower the grade.
     • Local runs post card.md only after the check passes, and never approve.
     • This is shadow mode: the bot still never approves.
  4. Renames. Both graders also grade a renamed file's old path: rename from in the diff, and previous_filename from GitHub. Moving a file out of a high-risk location no longer lowers the floor.
  5. Installed example skills. examples/*/SKILL.md, which pyproject.toml installs as agent skills, is graded medium instead of low.

  On the last 400 merged PRs, risk.py grades 177 high, 219 medium and 4 low.

• Linked issue(s): none

• Tests: added/updated? commands run? (what to test; if a test was replaced, where its assertions live now):

  • New tests:
    • test_review_pr_risk.py: every table row matches a tracked file; per-path grades, including an installed example skill; a PR takes the grade of its riskiest path. It runs risk.py in-process, so diff coverage counts it (100%).
    • test_review_pr_check_card.py: the risk floor, the raise marker, and decision coverage, including directories.
  • Extended tests:
    • test_pr_review_publish.py: the trim list, and risk graded from GitHub rather than from risk.txt.
    • test_review_pr_fetch.py: risk.txt, including a renamed file.
  • Ran:
    • These four files at each commit: 37, 37, 61, 62 and 62 passed.
    • pytest scripts/tests: 769 passed, 0 failed.
    • ruff, shellcheck and actionlint.

• Size/complexity triggers crossed: split, or why the shape is right: none. 0 production Python lines (size/XS). New functions are at most cyclomatic 7 and cognitive 11.

• If this simplifies or refactors: which mechanism was removed, which contract was preserved, and the tests that show it: card validation moves verbatim from pr-review-publish.sh into check-card.sh, so CI and local runs share one check. The existing publish tests still pass with their messages unchanged.

• Observable effect: what an operator will now see, and what the old behaviour cost them (if none, why the change is unobservable):

  • Every card carries Risk: and, unless the risk is low, Decisions for a human:. A card may also carry Trim in this PR:.
  • The bot posts "no verdict applied" when a card lacks these, puts its Risk below its paths' grade, or leaves a high-risk path unnamed. This includes a path a rename moved away from.
  • Before this change, a human approving after a clean card was not told what needed their judgement.

• Breaking changes: yes/no (details if yes): no. The bot reviews with the default branch's skill, so its own cards comply once this merges.

• PR addresses single concern: yes/no (details if no): no. Five standalone commits, kept together for two reasons. The risk commit needs the line budget the wording commit frees. The last two commits correct the risk commit's input and table.

• Root cause is upstream (Magpie/TraceLens/GEAK/IntelliKit/AgentKernelArena), ticket filed: n/a

[github-actions]

Code metrics gate: PASSED

Judged: the 14 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.

Dimension	Limit	New	Worse	Improved	Over the limit	vs base
Cyclomatic complexity	fails > 20	0	0	0	0	+0
Cognitive complexity	fails > 30	0	0	0	34	+0
Function length (lines)	fails > 50	0	0	0	465	+0
Nested block depth	fails > 4	0	0	0	37	+0
Parameter count	fails > 5	0	0	0	3	+0
Module length (lines)	warns > 800, fails > 1200	0	0	0	42	+0
Duplicated lines per file	fails > 0	0	0	0	83	+0
Dead code (vulture)	fails > 0	0	0	0	2	+0
Directory size (entries)	fails > 50	0	0	0	2	+0

Check	Findings	Effect
Comments	0
Broad catch without a reason	0
Broad catch that records nothing	0
English only	0
Production code imports test code	0
Repeated literals	0
A1 Private name of another package	0
A8 Flag argument	0
A9 Tuple return	0
A10 Raw vocabulary comparison	0
A11 Hidden global write	0
A12 Pure forwarder	0
A12 Single-implementation abstraction	0

Verdict computed by the gate scripts in /home/runner/work/_temp/code-metrics-gate, not by this tree's copy.

_{Thresholds and sources: [tool.hyperloom.code_metrics] in pyproject.toml. Tools: complexipy 8.0.1, jscpd 5.4.1, radon 6.0.1, ruff 0.16.2, vulture 2.16.}

[chaojhou]

PR #2243 -- [Review] grade PR risk, list decisions for a human, flag deletions, check every card before posting

What it does: Changes the review-pr skill.

• risk.py grades a PR's changed paths low/medium/high into risk.txt, with references/tiers.md as its only data.
• Cards gain Risk:, Decisions for a human: (required at medium and high) and an optional Trim in this PR: list.
• Card validation moves from pr-review-publish.sh into check-card.sh, shared by CI and local runs. Publish re-grades risk from GitHub's file list, so the review job cannot lower the grade.
• references/subtract.md asks which tests and mechanisms need to exist, and SKILL.md is condensed.

First review at 711e857. Since then #2227 (d5f1bde) has landed on main, and this PR now conflicts with it in SKILL.md and fetch.sh (re-review.md and rules.md need the same rebase).

Blocking issues: 3

1. [V5] coverage (Python 3.10), a required check, fails at this head [verified]
   Problem: diff-cover reports .claude/skills/review-pr/risk.py (60.0%): Missing lines 32,67-69,73-79,83-97 against the 80% floor (job 122131982627). Those are Risk.__str__, is_test, grade and main, the code that does the grading.
   Impact: the PR cannot merge while a required check is red.
   Action: make the tests' runs of these functions count toward coverage, or add the missing cases, then confirm the diff-coverage job passes.

2. [free:stale-requirements-row] references/tiers.md grades requirements/*, which main has deleted, so the PR's own test fails once merged [verified]
   Problem: .claude/skills/review-pr/references/tiers.md:85 lists requirements/* in the high row "dependency pins and gate limits (rule X7)". #2237 (606476a, on main since this PR's merge base 9010db3) deleted requirements/app.in and requirements/app.lock, so no tracked path matches the pattern. scripts/tests/test_review_pr_risk.py:28-38 (test_every_row_of_the_tables_names_a_tracked_file) fails on the merge of this head into main. CI passed only because it ran on the old base.
   Impact: merging turns main's test run red. The body's claim that every table row matches a tracked file is false against the branch it merges into.
   Action: after rebasing, drop requirements/* from that row, or point it at what now holds the pins (pyproject.toml already sits there), and re-run the test on the rebased tree.

3. [free:rename-graded-by-new-path] A rename is graded only by its destination, so moving a high-risk file out of a high-risk location lowers the floor [verified]
   Problem: fetch.sh's numstat awk takes the b/ path of diff --git (.claude/skills/review-pr/fetch.sh:117). A pure rename has no ---/+++ lines, so only the new path reaches files.txt. pr-review-publish.sh:44 lists --jq '.[].filename' and drops previous_filename. risk.py therefore never sees the source path of a rename. For example, git mv src/hyperloom/orchestrator/loop/coordinator.py src/hyperloom/orchestrator/core/coordinator.py, or moving .github/workflows/lint.yml under docs/, grades medium or low on both graders. check-card.sh then accepts a card that never names the Tier 1 file or the removed CI workflow among the decisions for a human.
   Impact: the case the floor exists for, a high-risk path removed or moved, is graded below its risk on both graders.
   Action: grade both ends of a rename. Emit the old path in fetch.sh (from rename from or the a/ side), and use --jq '.[] | .filename, (.previous_filename // empty)' in publish. Add a rename case to the tests.

Checked:

• risk.py, tiers.md (53 rules) against git ls-tree at head and on origin/main, and fetch.sh's numstat awk.
• check-card.sh against the checks removed from pr-review-publish.sh (moved unchanged apart from comments, plus the new Skill/Model checks).
• pr-review.yml wiring: an issue_comment trigger, and publish runs main's scripts with the write token, not PR-head code.
• A sentence-by-sentence audit of SKILL.md at 9010db3 against head and the references it links: no instruction dropped; factual counts only.
• git merge-tree against #2227 and against main after d5f1bde.

Ran:

• risk.py on the last 20 merges and on edge inputs (CRLF, empty, overlapping rows, tests-only, renames).
• check-card.sh under Linux bash on accepted and rejected cards: risk floor, raise marker, decision coverage including directories, CRLF.
• The PR's risk test on the merge tree with origin/main: the stale-row failure.
• Two independent skeptics per finding; two candidates were dropped (#2227 overlap as non-blocking, a Git-Bash-only hang).

Skill: 56bad557b6ff | Model: claude-opus-5-5 | Base: 9010db3 | Head: 711e857

[chaojhou]

PR #2243 -- [Review] grade PR risk, list decisions for a human, flag deletions, check every card before posting

What it does: Changes the review-pr skill and its CI publisher.

• risk.py grades a PR's changed paths, including a rename's old path, low, medium or high into risk.txt. Its only data is references/tiers.md: Tier 1 is high, Tier 2 medium, and a new risk table grades the rest.
• Cards gain Risk:, Decisions for a human: (required at medium and high) and an optional Trim in this PR: list.
• Card validation moves from pr-review-publish.sh into check-card.sh, which CI and local runs share. The publisher re-grades risk from GitHub's file list.
• references/subtract.md asks which tests and mechanisms need to exist, and SKILL.md is condensed.

Re-reviewed at 10f10a2, rebased onto 0198aa7. All three earlier findings are fixed. Two new ones come from the risk table.

Blocking issues: 2

1. [free:risk-table-row-stale-on-main] The PR's own tracked-file test fails once it is merged, because a Tier 2 row names a file main has since moved [verified]
   Problem: .claude/skills/review-pr/references/tiers.md:26 names kernelforge/config.py. #2282 (05e67da, on main since this PR's merge base) moved it to src/kernelforge/config/__init__.py, and that move's script did not rewrite .claude/. scripts/tests/test_review_pr_risk.py::test_every_row_of_the_tables_names_a_tracked_file, which this PR adds, fails on the trial merge with origin/main 3abe484: assert ['*/kernelforge/config.py'] == []. CI passed only because it ran on the older base. This is the same shape as the earlier requirements/* finding.
   Impact: merging turns main's test run red, and changes to src/kernelforge/config/ grade only through the src/* row.
   Action: rebase onto current main, change the row to kernelforge/config/, and re-run the test on the rebased tree.

2. [free:shipped-skills-graded-low] The risk table grades the demo skills the wheel installs as "prose and metadata nothing executes" [verified]
   Problem: the low row examples/* (.claude/skills/review-pr/references/tiers.md:93) covers four files that pyproject.toml:229-243 installs as agent skills under .agents/, .claude/ and .cursor/skills, chained from /hyperloom-setup:

   • examples/hyperloom-qwen3-8b-3h/SKILL.md
   • examples/hyperloom-qwen3-14b-fp8-12h/SKILL.md
   • examples/hyperloom-qwen3-14b-fp8-12h-forge/SKILL.md
   • examples/hyperloom-custom-advanced/SKILL.md

   The same table grades that class medium: src/* "the prompts and skills agents load at runtime" (line 92), and examples/skills/* (line 91). risk.py on examples/hyperloom-qwen3-8b-3h/SKILL.md prints risk: low with "prose and metadata nothing executes".
   Impact: a PR that changes the instructions an installed agent skill follows gets a low floor. check-card.sh and the publisher then accept a card with no Decisions for a human:, which is the case the floor exists to prevent, and the row's own stated reason is false for these files.
   Action: grade the shipped example skills medium, for example examples/*/SKILL.md in a medium row above the examples/* low row. Add one of them to the per-path cases in test_review_pr_risk.py.

Checked:

• Earlier findings:
  • V5 fixed: coverage (Python 3.10) and (3.11) pass at this head.
  • free:stale-requirements-row fixed: requirements/* is gone from tiers.md.
  • free:rename-graded-by-new-path fixed: fetch.sh:133 adds rename from paths, and pr-review-publish.sh:44 lists .previous_filename.
• Every check main's pr-review-publish.sh makes is still made, by check-card.sh or the publisher, and none is weaker. The publish job runs only default-branch scripts, including risk.py and tiers.md from its sparse checkout, and grades risk from GitHub's paginated file list.
• SKILL.md and re-review.md on main compared sentence by sentence with the merged text: no instruction is dropped or changed in meaning. SKILL.md is 305 lines against its 365 budget.
• Every artifact and argument the new text names matches what the scripts produce and read.
• rules.md changes one prose sentence of V2, and names no new command.

Ran:

• The skill as it is on main (272461950d5f), against this PR.
• test_review_pr_risk.py in a worktree of the trial merge with origin/main 3abe484: 1 failed (finding 1), 5 passed.
• risk.py on edge inputs (empty, CRLF, tests-only, workflows, pyproject.toml, renames, docs-only), on all tracked files and on the last 15 commits on main.
• check-card.sh under Git Bash on hand-made cards: risk floor, raise marker, decision coverage, trim list.
• git merge-tree against origin/main: clean.
• CI at 10f10a2: green except readthedocs, which this PR's paths cannot affect.
• Three area readers, with two skeptics on each finding: two candidates were dropped.
• SKIPPED: the shell-script tests (test_pr_review_publish.py, test_review_pr_check_card.py, test_review_pr_fetch.py) locally, because the harness passes Windows paths to bash on this host (exit 127, on main too). They pass in CI at the head, and no commit on main since the merge base touches those scripts.

Skill: 272461950d5f | Model: claude-opus-5-5 | Prior: V5 fixed, free:stale-requirements-row fixed, free:rename-graded-by-new-path fixed | Base: 0198aa7 | Head: 10f10a2

Superseded by the re-review at 70e4fdd: both findings fixed.

[chaojhou]

PR #2243 -- [Review] grade PR risk, list decisions for a human, flag deletions, check every card before posting

What it does: Changes the review-pr skill and its CI publisher.

• risk.py grades a PR's changed paths, including a rename's old path, low, medium or high into risk.txt. Its only data is references/tiers.md: Tier 1 is high, Tier 2 medium, and a new risk table grades the rest.
• Cards gain Risk:, Decisions for a human: (required at medium and high) and an optional Trim in this PR: list.
• Card validation moves from pr-review-publish.sh into check-card.sh, which CI and local runs share. The publisher re-grades risk from GitHub's file list.
• references/subtract.md asks which tests and mechanisms need to exist, and SKILL.md is condensed.

Re-reviewed at 70e4fdd, rebased onto 356c7a2. Both findings from the last card are fixed.

Blocking issues: none

Checked:

• git range-diff of the five commits against the reviewed 10f10a2: commits 1, 2 and 4 are unchanged. Commit 3 changes only the tiers.md:26 row, and commit 5 is new. The rest of the diff is as the last card cleared it:
  • every check main's publisher makes is kept, and none is weaker;
  • the publish job runs only default-branch scripts;
  • no SKILL.md or re-review.md instruction is dropped.
• free:risk-table-row-stale-on-main: fixed. references/tiers.md:26 now names kernelforge/config/, so src/kernelforge/config/__init__.py grades medium as Tier 2.
• free:shipped-skills-graded-low: fixed. examples/*/SKILL.md is in the medium row (tiers.md:91) above the examples/* low row. risk.py grades examples/hyperloom-qwen3-8b-3h/SKILL.md and examples/hyperloom-custom-advanced/SKILL.md medium and examples/README.md low. test_review_pr_risk.py adds the first of these as a per-path case.
• main's skill changes since the last base are its own tiers.md Tier 2 rows (refactor(executors): move warm patches, bringup records, accuracy stops, launch and round guards out of baseline.py #2294, refactor(phases): move GEAK dispatch, GEAK promotion and GEMM tuning out of kernel.py into mixins #2295), which this head carries.

Ran:

• scripts/tests/test_review_pr_risk.py in a worktree of the head, which is based on the latest main: 6 passed, including test_every_row_of_the_tables_names_a_tracked_file.
• risk.py on the example-skill, kernelforge/config/ and README paths above.
• git merge-tree against origin/main 356c7a2: clean.
• CI at 70e4fdd: 34 pass, 6 skipped.

Skill: 272461950d5f | Model: claude-opus-5-5 | Prior: free:risk-table-row-stale-on-main fixed, free:shipped-skills-graded-low fixed | Base: 356c7a2 | Head: 70e4fdd

LGTM

[chaojhou]

Approved per the review above (user consent 2026-10-29).