[pascalandy]

Outside the plan: Pascal assigned #423, which holds #422 in full as its point 1, and the % bug as its point 6. This pull request stacks on feat/420-intake-proofs of #421, which is open and not merged, at 2f813aa. #421 stacks on #418. It closes #422 and #423.

In round 5 of the day-1 trials, the loop held, and the replay scored 35, 36, and 36 of 36. But no build reached "Done when" in 5 rounds. The judge's cause: authors and reviewers find one variant of a family of cases at a time, and each narrow fix lets the next review find a sibling. Some gaps only the SDK can close: a grade can hide in free form, a first failing run is checked only for its form, and a new finding with no state passes beside None.. And ncly panics at startup in a folder whose path holds %.

What changes

• Free-form grades (template: some free-form grades still pass finish, and a forged first failing run passes #422.1, template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.1). just finish now fails on Grade: block, on warn-level, and on a cell such as | R1-F1 | really a warn |. Ordinary prose and tables with the words block and warn still pass, such as "Large files block the terminal" or the cell "the header block". finish-gate.md has a new part, "The limit of the parser", which lists the exact forms. The reviewer moves any other grade to the bullet Severity.
• First failing run again (template: some free-form grades still pass finish, and a forged first failing run passes #422.2, template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.1).
  • The record names the recipe just scenario <name>, a scenario of the Scenario commit. Its assertion holds a line of that scenario. just finish checks both.
  • just signoff runs just finish. Then it checks out the Scenario commit in a new scratch worktree and runs that scenario alone with go test. The run must fail on the line that the record names.
  • Signoff builds the arguments from the name of the scenario. It never runs the text of the record. A name that is a path, an invented command, a compile error, a scenario that passes, and a failure on another line all sign nothing.
  • The receipt records the run under first_run.
  • just scenario <name> is a new recipe of the template. The test kit skips coverage only when -run selects the scenario that NCLY_TEST_SCENARIO names.
• Case matrix (template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.2). spec.md holds ## Case matrix: one row for each family of cases, with the spec IDs of the tests that pass and of the tests that refuse. just finish fails when a family names no test or no passing input, or when a line of Failures or Acceptance is in no family. The skill asks for a pair of tests for each refusal and each branch that classifies an input. The first reviewer checks and runs the matrix before any repair. A repair is fixed only with a pair of tests for its whole family.
• A block that a reader needs (template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.3). Before the first review, acceptance.sh makes a file whose block uses a value that is not the default and holds private text. An independent reader proves both results. A handler that keeps the raw block must fail. The examples use text files only.
• Reads of the intake (template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.4). Each settled fact links the part read, at a fixed revision. Otherwise it is marked inference. The reviewer of round 1 refuses a fact that only cites a source.
• A finding with no state (template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.5). A finding of the last round with no state is open. just finish fails with its ID when ## Open lines omits it. The message says that an advise can be declined instead.
• % in the working folder (template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.6). The schema compiler resolved schema.json against the working folder as a file URL. The schemas now have URNs, also in the test kit and the tests of ncly.
• Scratch folders (template: one case matrix before the first review, replay the first failing run, and deliver #422 #423.7). Each experiment uses a new scratch folder, and no recursive delete runs inside sh -c or a loop. The tests and just acceptance need no root, no mount, and no system service.
• Moves. Five moves keep files under just limits, with no change of behavior: the answer schema check, the schema compiler, the severity forms and their proofs, the coverage report, and two sections of the skill (replace-a-file.md, error-mapping.md).

Each new rule has a row in AGENTS/rule-index.md, with its check or "Prose only".

Decisions made without Pascal

• One allowed recipe. The allowed list holds one form, just scenario <name>. Signoff runs go test itself and does not use the justfile of the Scenario commit, so a changed recipe forges nothing.
• Same assertion means same line. The run must fail on a line of the scenario that the record holds. The reviewer still judges whether that line proves the missing behavior.
• The matrix check reads the references only. Finish cannot know which generic cases apply. It checks that each family names a test, that each family names a passing input, and that each line of Failures and Acceptance has a family. It cannot prove that the matrix is complete or correct, or that the extension is ready. The reviewer checks that the matrix is complete. finish-gate.md, the skill, the readiness review, and the rule index say so.
• No claim about the trials. This pull request does not claim that the trials succeeded. No build of the day-1 trials reached an earned "Done when".
• The parser limit is narrow on purpose. A grade in other words, such as "a block for me", passes finish. Rounds 1 and 2 of the review found false positives on ordinary table cells and on "block level". So the cell rule needs a finding ID in its row, and the level form needs a hyphen. "block-level" still fails.
• The assertion line is exempt from the severity forms. It copies a line of the scenario. Only the first one is exempt, and finish reads that one.
• Earlier builds are judged with the new no-state rule. A ledger of the unmerged template: limit a build to 5 rounds, keep a ledger of findings, and close the gaps of round 3 #417 to ✨ feat: template: finish proves the first failing run and one state of open work, and release runs finish again #421 stack whose passing build ended with a stateless advise now fails. No release holds the ledger, so I accepted this.
• verify-nicely without the keychain. The agent skipped ncly doctor with no domain, the feature auth, and the template step Doctor. They probe the keychain, which this task must not touch. The refusal step used a text file with a NUL byte instead of the bytes of the doc.
• Stack. I ran just signoff on this head only. just signoff-layers would also sign the heads of ✨ feat: template: a build stops at 5 rounds, and a ledger keeps every finding #418 and ✨ feat: template: finish proves the first failing run and one state of open work, and release runs finish again #421, which this task does not touch.

Checklist

• 1. Spec: finish-gate.md, extension-repository.md, generic-cases.md, and test-kit.md cover each check and each rule
• 2. Scenario: each new case failed first, see "Proved by"
• 3. Code: those cases pass in just template and just check
• 4. Text: n/a. The messages of just finish and just signoff stay in code, as every message of sdk/extensionrepo does
• 5. Checks: just check passes
• 6. Verify: an agent that did not write the change ran verify-nicely on template and on ncly --version in a folder whose path holds %. Every step that ran passed, see "Proved by"
• 7. Docs: n/a for the backlog, because the card is outside the plan. README.md shows no change for a user. The skills of the template teach each new rule
• 8. Sign off: just signoff passes on this head

Proved by

• Fails first, then passes. A probe called the parsers of sdk/extensionrepo at the base 2f813aa and at this head:

  == 2f813aa8 (base)
  grade "| R1-F1 | really a warn |"                              -> 0 problem(s)
  grade "Grade: block. The name is a block for me."              -> 0 problem(s)
  grade "This issue is warn-level and should block the release." -> 0 problem(s)
  grade "Large files block the terminal, and inspect may warn."  -> 0 problem(s)
  no-state finding of round 2 beside None.                       -> []
  first run "echo invented"                                      -> []
  first run "just scenario inspect"                              -> []
  == this head
  grade "| R1-F1 | really a warn |"                              -> 1 problem(s)
  grade "Grade: block. The name is a block for me."              -> 1 problem(s)
  grade "This issue is warn-level and should block the release." -> 1 problem(s)
  grade "Large files block the terminal, and inspect may warn."  -> 0 problem(s)
  no-state finding of round 2 beside None.                       -> ["...the finding R2-F1 of this round has no state, so it is open..."]
  first run "echo invented"                                      -> ["...names the command \"echo invented\", which is not just scenario <name>..."]
  first run "just scenario inspect"                              -> []
  

• %. The new lines of testdata/script/version.txtar failed before the fix with panic: ... invalid URL escape "%/s" at line 27, and pass after it.

• The replay. scripts/template-signoff-test.sh runs the rendered just signoff on real commits. It refuses echo invented, a Scenario commit that does not compile, a scenario that passes at its commit, and a record of another line. It signs a correct repository and prints signoff: the first scenario fails again at <sha>.

• The other cases. The template tests prove the paired grades and the prose control, the case matrix (each refusal and the passing state), the stateless finding beside None. and then with its open line, and just scenario on a scenario and on a path.

• Checks. just template passed at 4fb1d87. just check passed at f3fe78c. At 20be50c, one timed scenario of agent, agent_run_cancel, timed out while another job loaded the host. It then passed 3 of 3 runs alone with just scenario agent/agent_run_cancel --count 3. just signoff ran just check again on this head, see below.

• verify-nicely. An agent that wrote none of the change ran it on 75c627e. ncly --version names the build. Every step of the feature template that ran passed. The rendered spec.md holds ## Case matrix at line 58, and just --list shows scenario name. just finish names each Fill: line, also the one of the matrix, and the missing verdict. In work 100%, ncly --version and ncly describe --json exit 0 with no panic.

• Review. Three rounds by an agent that wrote none of the code, one reviewer each, then a review of the last fix. Round 1: 1 block, 4 warn, 6 advise. Round 2: 1 warn, 2 advise. Round 3: Ready, 1 advise. Review of the last fix: Ready, with 1 test that it suggested, now added. Two advise items were declined: an assertion that pastes many lines still needs a real failure on a line it names, and the rule for earlier builds above.

• Sign off. just signoff ran just check on 75c627e in a detached scratch worktree, then printed Signed off on 75c627ebb49498b4ab1728017dcf45f1843cdb54. The status signoff of this head is green.

by Opus-5.5 via Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Repository UI
• Review profile: ASSERTIVE
• Plan: Advanced
• Run ID: 2874cd6a-42df-430a-b757-3f7383567fc2

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts