[Deicyde]

Problem

tests/test_skill_examples.py had 41 tests at de6de5f and about 590 literal-string checks on skill and doc sentences:

• 27 of the 36 merges that touched a SKILL.md also had to edit this file;
• rewording one sentence broke CI (Keep the development skill within its 220-word budget #279);
• any README or skill cut fails until the pins go.

On 2026-10-20 Jack approved replacing them with general checks.

Change

• New general checks:
  • every relative link and anchor in tracked Markdown resolves (one generated target is allowlisted);
  • every autoform ... invocation in the CLI reference and the skills parses with the real parser, with placeholders filled in. That covers more than 100 invocations, and all of them parse today;
  • schema ids the docs name exist in code, and so do codes the skills name. The read-back rubric's two ids are allowlisted, because their tooling is in Add read-back cards and the testimony validator #216 and Add statement-review commands #284;
  • frontmatter keys the skills teach are keys the graph loader accepts;
  • the REPL import allowlist is documented;
  • 11 safety sentences stay in their skills (outward-facing actions, consent, library text is data, trust caveats). This is the only wording still pinned.
• The 27 tests that only checked that sentences exist are deleted. The wording and config-string pins in the 13 kept behavior tests are trimmed. Against current main, the file goes from 1,758 to 916 lines, and from 47 tests to 26.
• develop-plugin's "preserve insight in a focused test" rule becomes: fix the product first, add a skill sentence only when behavior cannot carry the fix, and never pin a skill's wording in a test, except the short safety-rule list. The 220-word cap goes with its test.
• The tests of the generated workflows' safety steps stay: Pages deploys only from main, and the verify workflow cleans only the root package, rejects kernel-check bypass options, audits every root-package declaration, checks elan's checksum and installs nothing from @main.
• Two doc fixes:
  • AGENTS.md's init exit row. In fact init leaves an existing entry alone and exits 0. It exits nonzero for an input or target it cannot use safely, such as a case or Unicode variant of a path it manages (even with --force), a link, an unusable root .gitignore, a directory it cannot list or an invalid option.
  • The README's four publishing steps, which listed autoform dashboard as a fifth command that never returns.

Kept on purpose

• Tests that pin wording but landed after the inventory are left to their authors: test_readme_tells_users_how_to_run_several_sessions (Document running several sessions on one project #420/can you have more sessions coordinate? #421), Scale roadmap inventories and the full DAG explorer #90's five new tests, and Have Roadmap re-check retraction on the branch Formalize works from #407's test_roadmap_retracts_against_the_branch_formalize_works_from.
• test_readback_faithfulness_contract_is_hash_bound_and_machine_readable is unchanged, because the rubric stays with Add read-back cards and the testimony validator #216 and Add statement-review commands #284.

Test

• With the no-pytest runner, the module passes on this branch after the merge with main (36 passed: 26 tests, one of them with 11 safety cases). Before the deletions, with every old test present, it also passed (57).
• ruff passes.
• CI runs the full suite.

Overlap

This conflicts in tests/test_skill_examples.py with:

• Add statement-review commands #284, Show statement review on the site #217, Label statement approvals by the repository's code-owner review policy #219 and Prove2me - Add skill files and modifications to the autoform-bot CLI to allow for easy integration with prove2.me #351, which each append a test at the end of the file: keep both;
• Deploy Pages only from the commit autoform verify passed on main #399, which adds two lines to test_setup_asset_static_site_contract: keep them on the trimmed test;
• Have Roadmap re-check retraction on the branch Formalize works from #407, which adds assertions to test_roadmap_commits_so_the_published_site_can_catch_up, a test this PR deletes: keep any that check behavior and drop sentence pins.

#90, #407, #397, #378 and the 0.6.0 release (#439) have merged, and this branch now merges main: the tests this PR deletes go along with their edits, and their new tests stay as main has them. Three checks were adjusted for the new docs. The code check now reads codes audit builds with an f-string (catalog-{problem}-limit). The frontmatter check skips a bare warning: prefix. The command check skips autoform 0.5.0 (...), which is --version output, and inline mentions of the generated workflow's name, autoform verify. The new schema check found that #397's changelog still named autoform-runtime/v3; the 0.6.0 release fixed that line.

No other open PR gains a conflict.

[Deicyde]

Review at 9c669f1: one finding blocks merge. One round, read-only; the planned conflicts with other open PRs are out of scope, as asked. The review ran at 4e7833d, and I re-checked each finding at 0df905f and 9c669f1, after the merges with main.

The rest holds. The 27 deleted tests only read files, match regexes or check that paths exist, and the new link, path, anchor and command-parse tests cover what they checked. The merges with main keep #90's edits to the two kept setup-asset tests and its 5 new tests. 45f4cfb changes only skills/develop-plugin/SKILL.md, so the freeze holds. aead840's init row matches scaffold.py, and 4e7833d is right that the generated blueprint-pages.yml already runs check, render and the mkdocs build. CI: all 8 checks pass at 9c669f1.

Blocks merge:

1. 16f7a5a deletes the only tests of the generated workflows' safety steps. On main (0a8e3e5), test_setup_asset_static_site_contract asserts that the verify workflow keeps lake clean "$root_package", the "Reject kernel-check bypass options" step (forbidden="skip""KernelTC" and git grep -n -I "$forbidden" -- .) and the elan checksum, and that Pages deploys only from main (tests/test_skill_examples.py:990-1005). At this head no test checks any of them, and test_example_workflows_match_the_scaffold_templates only compares the example with the template, so removing a step from both stays green. These are product template contents, not doc wording. The bypass scan matters most: tests/test_lake_artifact_audit.py:1107-1108 notes that neither the axiom walk nor the safety scan catches a declaration the kernel never checked, and that the workflow's lexical check is not part of the probe. Keep these assertions, for example in tests/test_scaffold.py against autoform_cli/templates/github/workflows/.

Notes:

2. The init row is still partial (AGENTS.md:75). Every ScaffoldError exits 2 (__main__.py:538-541), so init also exits 2 for a non-regular, hard-linked or oversized root .gitignore, a directory it cannot list, and a bad --autoform-ref or --autoform-source. It also exits 0 with a warning when it skips CI for lack of a pin. List them, or call the row a summary.
3. The new develop-plugin rule contradicts a kept test. skills/develop-plugin/SKILL.md:13 says never add a test that pins a skill's wording, and test_safety_rules_stay_in_their_skill pins 11 skill sentences. A carve-out for the safety rules fixes it.
4. Two new tests can pass vacuously. test_markdown_links_and_anchors_resolve and test_documented_schema_ids_exist_in_code assert only that nothing is broken or missing, so if the extraction stopped matching they would check nothing. test_documented_autoform_commands_parse guards against this with checked > 100, and the same kind of floor works here.
5. Counts. The description's "40 tests" was 41 at de6de5f (40 without can you have more sessions coordinate? #421's), and 27 of the 36 merges that touched a SKILL.md also edited this file, not 26.

[Deicyde]

Thanks. Blocker 1 is fixed in a85cf5a: the site-contract test again asserts the generated workflows' safety steps (main-only Pages deploy, scoped lake clean, the kernel-check bypass scan, the root-package audit, elan's checksum, no @main), next to the SHA-pin loop it kept. 40a32c3 takes notes 2 to 4: the init row now summarizes every refusal, develop-plugin's rule names the safety list as its one exception, and the link and schema checks require their extraction to find something. The description's counts are corrected (note 5). The skill tests pass, 36 run directly.

[Deicyde]

Head check: the blocker is fixed; ready to merge once CI finishes (reviewed at f8028e4; review at 9c669f1).

• a85cf5a restores the assertions on the generated workflows' safety steps: the scoped lake clean, the kernel-check bypass scan, the root-package audit, elan's version and checksum, and no @main. The merge with main moved the Pages check to Deploy Pages only from the commit autoform verify passed on main #399's github.event.workflow_run.head_branch == 'main', which matches the template.
• 40a32c3 takes notes 2 to 4, and f8028e4 skips the generated workflow's name, autoform verify, in the command check.
• test (3.10) and test (3.13) pass at this head.

Posted by Claude Code (348 worker session).

[Deicyde]

Changes since the last head check (40a32c3), all merges with main plus fixes they needed:

• Deploy Pages only from the commit autoform verify passed on main #399 moved the main-only Pages deploy to a workflow_run gate, so the kept safety assertions check its conclusion, head branch and workflow path.
• Retire generated Mermaid: a Markdown dependency summary and locked-down authored diagrams #378 edited two kept tests, which now match main, and reworded Roadmap's commit sentence, so that safety pin holds only its rule.
• The command check skips inline autoform verify, the generated workflow's name.
• CHANGELOG.md takes the released 0.6.0 notes.

The skill tests (36) and test_release (7) pass, run directly. CI is running at 6cce5e3. I'm marking this ready for review.

[Deicyde]

Head check: nothing blocks merge (reviewed at 6cce5e3; last head check at f8028e4).

The two commits since f8028e4 are merges with main. Their remerge diffs:

• 9b0d9fa resolves two hunks in tests/test_skill_examples.py. It keeps this PR's deletion of the wording tests Retire generated Mermaid: a Markdown dependency summary and locked-down authored diagrams #378 edited, and takes main's pinned_mermaid version of the site-config test. None of main's edits to the four files this PR touches were lost. The safety assertions f8028e4 restored still match the templates, now including Deploy Pages only from the commit autoform verify passed on main #399's workflow_run gate.
• 6cce5e3 resolves CHANGELOG.md to main's released 0.6.0 notes.

The command check's skip covers only the inline autoform verify, the generated workflow's name. CI passes every job at this head.

Notes (optional):

1. The Roadmap pin lost the commit. It now holds only "only after this final validation; pushing is outward-facing and requires a user request." (L921-L922). Retire generated Mermaid: a Markdown dependency summary and locked-down authored diagrams #378 reworded the sentence, and "Commit the vault" fell out of the fragment, so the pin no longer ties the timing rule to the commit. A second fragment, "Commit the vault", would restore it.
2. Merge order. Add statement-review commands #284, also ready, conflicts with this PR in tests/test_skill_examples.py. Its test_review_skills_route_statement_review_through_the_cli_reference pins links in a skill, which this PR's rule forbids. Integrate mathematical book reading and dependency exploration #442 merges cleanly but adds sentence pins to the same file, for example assert "one full-viewport reading/map workspace" in cli. Whichever lands after this PR will need to drop or convert them.
3. A stale comment. The allowlist comments at L869 and L904 say open PRs Add read-back cards and the testimony validator #216 and Add statement-review commands #284 ship the read-back rubric's tooling. Add read-back cards and the testimony validator #216 merged without it, and Add statement-review commands #284 does not add it. This predates f8028e4.
4. The body is out of date. The file now goes from 1,763 to 947 lines and from 47 to 26 tests. The Overlap section still lists Deploy Pages only from the commit autoform verify passed on main #399 and Have Roadmap re-check retraction on the branch Formalize works from #407, both merged.

Posted by Claude Code (session jackmccarthy-da).

[Deicyde]

Head check: nothing blocks merge (reviewed at 399140b; last head check at 6cce5e3). It is ready once the skeleton suite, the only check still running, passes.

The two commits since 6cce5e3:

• 04400bd merges main with Bound claim list and let it list only named targets #446 and Let import-impact load only a module's candidate importers #449. Its remerge diff touches only tests/test_skill_examples.py. Bound claim list and let it list only named targets #446 and Let import-impact load only a module's candidate importers #449 had added phrases to two wording tests this PR deletes, test_formalize_replaces_custom_orchestration_with_the_markdown_frontier and test_import_edits_use_reported_scope_instead_of_a_blanket_ban, and those edits go with them. The behavior they document stays tested:
  • Bound claim list and let it list only named targets #446's claim listing, in the claim tests;
  • Let import-impact load only a module's candidate importers #449's import closure, in tests/test_impact.py;
  • their documented invocations, in this PR's command check, including autoform work import-impact ... [--module-scope SCOPE] (README L1162) and autoform claim list --worker-id WORKER (L1211).
• 399140b gives a placeholder after an option with fixed choices, such as --module-scope SCOPE, that option's first choice instead of "1". _sample replaces only placeholders (<...>, $VAR, all-caps words), so a literal value stays as written and argparse still refuses a documented choice that does not exist.

Posted by Claude Code (session jackmccarthy-da).

[Deicyde]

Head check: nothing blocks merge (reviewed at 7e1324e; last head check at 399140b).

The four commits since 399140b:

• 4a37104 merges main with Share builds across checkouts through Lake's artifact cache #450. Its remerge diff only deletes the five wording-test functions this PR removes; Share builds across checkouts through Lake's artifact cache #450's assertions in them went with them. Share builds across checkouts through Lake's artifact cache #450's behavior is still covered:
  • both lakefile keys, in tests/test_project_create.py;
  • the restoreAllArtifacts message for a missing .olean, in tests/test_skeleton.py;
  • "Never enable the cache without restoreAllArtifacts", in skills/setup/SKILL.md (L129-135).
• 2ca340c merges Cut repeated and internal prose from the CLI reference and skills #451 (the docs cuts, reviewed at their own head). The merge is clean: an empty remerge diff.
• fecc193 merges main with Add statement-review commands #284 and Have Roadmap list only what a library's proof uses under Proof depends on #443, and drops nothing either added. A line-by-line check found every line main added (merge base bb79e25 to a840c3f) in the merged files. The only exception is CONTRIBUTING's MathJax sentence, which was reworded. The 25 README lines its resolution removes are Cut repeated and internal prose from the CLI reference and skills #451's shortened MathJax paragraph; the resolution keeps Have Roadmap list only what a library's proof uses under Proof depends on #443's paragraph word for word.
• 7e1324e credits autoform-review-policy/v1 only when the shipped workflow enforces it. Its sha256 (ad00ec55…b620) is the value autoform-verify.yml L153 checks. That is a behavior check, in line with this PR's rule.

CI passes all 9 checks.

Notes:

1. Show a Mathlib article's mathlib_declaration on its page #454's Mathlib rows. Show a Mathlib article's mathlib_declaration on its page #454 (on main at 7a52246) added the mathlib_declaration and mathlib_file rows to the README's assertions table, after this branch's last main merge, so the branch does not have them yet. The next merge with main must keep them. Follow up #454: Mathlib row wording, example and test pins #459 rewords them, and Cut the docs agents read, split the CLI reference, and give them line budgets #458 is stacked here.
2. The MathJax sentence. CONTRIBUTING's sentence no longer names tests/test_site_mathjax_node.py or the tarball's sha256 check. The skip and fail-in-CI rules are intact.
3. A stale comment. tests/test_skill_examples.py L876 still calls Add statement-review commands #284 an open PR. The exemption is still right: no shipped code produces those ids.

Posted by Claude Code (session jackmccarthy-da).