Menu ▾ ▴

#427 DIVE-2645: put the reverse gear on help text, refusals and guards

closed
nobody
None
2026-08-09
2026-08-03
Anonymous
No

Originally created by: lodar

Three subtractive cuts to the CLI, filed because lodar said it "became worse and annoying" and was right. Nothing here adds a flag; every exit code is unchanged.

What changed

Cut 1 — help says what a command does. task --help 139 → 47 rendered lines, one line per verb, every flag kept, widest line 99 columns. Ticket references across every rendered --help surface: 95 → 0, swept by running the built binary over 36 commands rather than by grepping source.

Cut 2 — a refusal is one line: what was refused plus the next action. The rationale moves to the commit log and the wiki. Also fixes the refusal the ticket cited verbatim — 5dive usage printed must run as root (try: sudo 5dive ) because require_root was called with no args; main() now records argv so the hint names the command.

Cut 3 — one guard deleted, named. The pinned-baseline push differential in tests/broker_surface_unit.sh (279 → 177 lines; deleted, not demoted, not merged). It proved the INST-5 parameterization left push byte-identical; that refactor is merged and was proven inert, so its class cannot recur. What the arms had become is a freeze on refusal text, and the escape its own header names does not exist — a second guard in the same file refuses any pin carrying the refactored wrapper, so no pin can ever hold the new wording. Sections 2–7, which grade live behaviour, are kept.

Two corrections that belong in the record

The headline figure in an earlier commit message was wrong. I reported refusals over 120 characters as 111 → 2. That census matched only double-quoted strings and took the last one on each line, so it never saw a single JS die(`...`) in the council bundle. Measured properly — longest string per site, backticks included, usage: synopses excluded — it is 146 on main, 29 on this branch. The direction holds; the 2 was an artifact of the instrument.

src/cmd_council.sh is generated and cut 2 hand-edited it. CI's bundle-drift job caught this on the first push, and nothing in local review could have: the next build silently reverts the generated file, so 12 compressed council refusals would have evaporated while the diff still showed them. The messages now live in src/council/cli.mjs and the generated file matches gen_cmd.mjs byte for byte.

Where a long refusal was right

Three messages exceed the ceiling deliberately, and each was identified by a harness rather than guessed at:

message before → after why it stays long
merge gate, scan hit its bound 1095 → 283 cannot distinguish unknown from absent
merge gate, branch not found on main 739 → 288 same, plus which repos were searched
council, 0 of N seats reached 347 → 258 must separate a transport outage from an abstention, state no receipt was sealed, and name the failed seats

A refusal you cannot shorten is telling you something about the guard, not about the writing. I over-compressed the third and council_capture_e2e said so.

Review status — read this before merging

  • olivia PASSed c2baa6b, built and run by her, graded against the criteria as filed.
  • Head is now b6c8759 — the generator fix landed after her PASS. The merge candidate is therefore not the tree she graded. Whether the PASS transfers is hers or main's call, not mine.
  • The branch was pushed on main's written authorization. The delegated-push broker path is closed on a done row (task need refuses with "reopen it before gating on a human"), and reopening a row a verifier had just closed to satisfy a push mechanic was the worse trade.

🤖 Generated with Claude Code

Related

Tickets: #544
Tickets: #548

Discussion

  • Anonymous

    Anonymous - 2026-08-03

    Originally posted by: lodar

    Verifier re-grade — carried here because the board record cannot hold it

    olivia re-graded this at b6c8759 and the verdict is PASS. Posting it on the PR because the task row physically cannot take it, and whoever merges needs to carry it into the row's result.

    Why it is not on the row. DIVE-2645 closed at 21:39, before this PR existed. Since then task done --append-result is refused by the DIVE-1835 merge gate (PR open, unmerged) and set-body --append is refused because a done row's body is frozen. The only remaining routes were --force-merge-gate, which would override a gate correctly reporting the PR unmerged, and task reject, which would mislabel passing work. She used neither, which is right.

    The verdict, in her terms

    Re-graded at the new head, not transferred. She passed c2baa6b; head is b6c8759. This PR's head is b6c8759, so the head-vs-graded-sha check passes.

    The ungraded commit was verified independently: git archive of b6c8759 into a clean tree, ran src/council/gen_cmd.mjs, and the regenerated src/cmd_council.sh is byte-identical to the committed one (sha256 831058c3f20b3272 before and after, matching the committed blob).

    Criterion 2 is MET IN DIRECTION, NOT ABSOLUTELY — read the number, not the verdict. My reported 111 → 2 came from a census matching only double-quoted strings, last-per-line, blind to JS die(`...`). Corrected: 146 → 29. She re-measured independently and got 129 → 14, then adopted my 29 over her own 14 because hers is line-based and misses multi-line strings — the less flattering instrument, chosen deliberately. Both agree on direction (~80–90% cut) and both agree the remainder is non-zero.

    Not bounced, though I offered. 28 of the 29 are pre-existing long refusals untouched by this branch; the 29th was a genuine over-compression I found and restored (all-seats-unreached convene: transport outage vs abstention, no receipt sealed, name the failed seats — 347 → 258).

    Two things for whoever merges

    1. Carry the block above into the row's result when closing DIVE-2645 post-merge. It is the verifier's record and it exists nowhere else.
    2. A done row with an open unmerged PR is exactly the shape done=merged-to-main exists to prevent. It happened because the row closed before the PR was opened. Flagged by olivia as needing reconciliation; not fixed here, and deliberately not filed as a new ident from this row.

    🤖 Generated with Claude Code

     
  • Anonymous

    Anonymous - 2026-08-03

    Originally posted by: lodar

    Verifier re-grade [#2] — PASS at ebdf1fd

    Third sha on this row. I re-graded rather than carrying b6c8759's PASS forward, and dev made no claim either way — flagging the move rather than letting green CI carry it is the behaviour DIVE-2656 exists to make automatic.

    Verified, not read off the report:

    • The generator, not the product. ebdf1fd touches BOTH src/council/cli.mjs and src/cmd_council.sh. I re-derived sync independently: git archive ebdf1fd into a clean tree, ran src/council/gen_cmd.mjs, and the regenerated cmd_council.sh is byte-identical to the committed one. This row's own source-vs-product lesson, applied by the person who learned it two commits ago.
    • The restored refusal is 107 chars — under 120, so criterion 2 does not regress. Both required facts are present: the seat-delivery rail and the `_deliver` grant. Short and complete were not in tension.
    • CI at ebdf1fd: test, test-installed-host, selfcheck-union all SUCCESS. selfcheck-union now PASSES, which is the confirmation that its earlier skip was the needs: [rails-pristine, test-installed-host] dependency edge — a gate never reached — and not a second defect.
    • PR head == the sha I graded. The head-vs-graded check passes on this row.

    Correction to my own grade's foundation, since dev raised it. dev asked whether I had built and run as root and therefore shared the harness hole. I did not run council_capture_e2e.sh at all — my grade accepted dev's harness evidence ("6 red, same 6 red on an origin/main control"). So my PASS rested on evidence carrying the root hole, without my having made the error myself. dev's non-root re-run of all 12 harnesses this branch touches is what closes it, not anything I did. Accepting evidence is an act with the same authority as producing it.

    Worth naming: the restored line is a JS die(`...`) — a backtick string, which is exactly the form that both dev's census and my independent one were blind to. The defect landed in the blind spot the same night the blind spot was found.

    Verdict unchanged from the b6c8759 grade in all other respects, including criterion 2 as MET IN DIRECTION, NOT ABSOLUTELY — 29 sites remain over 120 chars, 28 of them pre-existing and outside this branch. Anyone reading this row for a no-long-refusals guarantee should read the number, not the verdict.

    Merge is Marcus's or mine; nothing publishes either way.

     

    Related

    Tickets: #2

  • Anonymous

    Anonymous - 2026-08-03

    Originally posted by: lodar

    ⛔ THIS COMMENT WAS WRONG WHEN POSTED — SUPERSEDED

    ebdf1fd WAS graded. olivia posted "Verifier re-grade [#2] — PASS at ebdf1fd"
    at 22:04:01Z; this comment went up at 22:08:08Z, four minutes later. The line
    below reading "graded by nobody" was already false as I wrote it.

    Do not hold a merge on this comment. PR head ebdf1fd is graded PASS by the
    verifier and CI is green on it. The sha ladder below is retained only as a record of
    what the task row still says — the row's result grades c2baa6b, which remains
    true and is a board-recording gap, not a reason to withhold a merge.

    Corrected by dev after main verified the comment ordering against the API. Tonight's
    own lesson, landing on the artifact written to prevent it: a record about staleness
    is not exempt from staleness.


    ⚠ Before merging: the row's recorded sha does NOT match this PR's head

    Per main's rule tonight — a verdict is not a verdict until it is on the row with the sha it graded, and do not accept "it was approved": read the row and match the sha. Applying that to this PR, honestly, because it fails the check:

    what sha where it is recorded
    olivia's first PASS c2baa6b ✅ on the row, in DIVE-2645's result, with the sha named
    olivia's re-grade b6c8759 ⚠️ not on the row — the board physically refused it (see comment above); it exists only in that comment
    this PR's head ebdf1fd ❌ graded by nobody

    The one ungraded commit is ebdf1fd, "restore the two facts the pre-flight refusal must carry": CI went red on b6c8759 (test, test-installed-host, both on council_capture_e2e) because the no-delivery-grant pre-flight must name both the seat-delivery rail and the _deliver grant, and cut 2 had compressed out both. Restored in 107 characters.

    Why it wasn't caught locally: the harness was being run under sudo, and as root that pre-flight always succeeds, so the harness skips its refusal leg — the exact leg grading the change. Same file: 23 arms as root, 27 as a normal user. All 12 harnesses this branch touches have since been re-run as a non-root user; all green.

    CI on ebdf1fd — named, not counted

    PASS  actionlint, changed-harnesses, check, docker-install, hook-staleness (PR
          instance), rails-pristine, scan, selfcheck-union, shellcheck,
          supply-chain-guard, test, test-installed-host
    SKIP  hook-staleness (push-event instance) — if: github.event_name == 'pull_request'
    NOT WOKEN  install-guard, template-skills — path-gated on files this diff does not touch
    FAIL  none
    

    selfcheck-union is worth a line: it skipped on the red run because it declares needs: [rails-pristine, test-installed-host] and the latter was red — a gate never reached, which GitHub renders identically to a gate that declined to apply. It now passes, confirming the skip was the dependency edge and not a second defect.

    A green CI on an ungraded sha is not a verdict. It says the checks pass; it says nothing about whether the verifier accepts the change. That decision is olivia's and she has the one-commit diff.

    🤖 Generated with Claude Code

     

    Related

    Tickets: #2

  • Anonymous

    Anonymous - 2026-08-09

    Ticket changed by: lodar

    • status: open --> closed
     

Log in to post a comment.