Skip to content

release: v0.61.0 — shape jobs (itemShape), check-list look, Result fold, Add back - #79

Merged
hamr0 merged 71 commits into
mainfrom
feat/shape-jobs
Nov 19, 2026
Merged

hamr0 merged 71 commits into
mainfrom
feat/shape-jobs

Conversation

@hamr0

@hamr0 hamr0 commented Nov 19, 2026

Copy link
Copy Markdown
Owner

v0.61.0.

  • Shape jobs (bareguard itemShape): a ~ rule gets a code truth check first; only a rule left Not checked takes a shape field; real bound required, negating rules never shaped, contested rules stay with the human.
  • Check-list look on the card, Audit and draft readout ([you]/[agent]/[default], truth/shape/file, CHECKED / NOT CHECKED / SUGGESTED).
  • Result tab folds every shape line to one row per item, values shown.
  • Add back on a shape row you dropped; Revise keeps a Drop.
  • .md report heading rule copied from fwd v0.36.0; single-file Source error text; Audit Line blocks fix.
  • README: hire framing, "the factory that builds the factory", rubric covers sheets and lists.

Branch review at 43a3ca7: ready, no blockers, stages 1-3 ran, fail-first 25/25. Full suite 6173/6171/0/2 on 381b95c.

🤖 Generated with Claude Code

hamr0 and others added 30 commits November 18, 2026 21:00
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…R1-R5)

The one menu (boolean/equals included), the name rules (reserved words
are read from bareguard at runtime, off and said so on the pinned 0.36),
and buildShapeLine/buildShapes: menu check, per-field grounding (R1),
Suggested split (R2), per-value oneOf (R3), item count only from the
person's words (R4), equals only on a rule field (R5).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
plainStepGap gets an itemShape branch: item reds read 'item N: field ...'
(1-based), whole-field gaps in plain words, redCount counts item reds.
Written against bareguard's documented red grammar; end-to-end through
gradeStep waits for the bareguard release (clause 5).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
Plan card rows (one per field plus the default row), per-row results from
itemShape gaps, the report's '#### Item N' blocks and flattenStep's
pass-through of a shape answer, all picked by the line's signed class
and never by the key 'results'. The part's real build is clause 5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…k (clauses 4 and 3)

CheckBlock.setShape: [you] rule rows carry their field, [agent] field
rows (job-derived on, no-word Suggested, Drop/Add), one [default] row,
line-class exclusion beside the sheet's, the truth of the items Not
checked. 'Rules or shape could both fit' is one counted reading-style ask
through questionsOfBlock/applyChoices. The part is built from the pure
fields; validating and signing it is clause 5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
… (clause 4)

sheetFinalEdit also serves shape [agent] field rows; the Drop is kept for
a Revise in the record's shapeDropped (readShapeDropped), passed through
authorjob. setShape takes {asks} so a caller with nobody to ask puts no
ask on the line. The menu line text moves to shapeoffer (one prompt file).
End to end through the signing gates waits for clause 5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…lauses 1 and 2 call sites)

The row-edit reply gets an optional itemShape property (shapeSuggestProp),
THE ITEM SHAPE block carries the rules and the one kind menu
(boolean/equals included), authorClose strips it from the declaration and
builds it with buildShapes into rows (setShape). OFF unless this bareguard
carries itemShape (shapeAvailable): the turn is byte-identical today.

Failure: run mwmmcjxo (research-1): a list-of-items job's per-item rules (title, status one of in progress/closed/finished, summary under 200 characters) could not be checked; two went to unbuildable fact checks with confusing words, one Not checked
Addresses: a job whose answer is a list of items with named fields has no way to be offered a shape
Corrects: nothing existing; the offer is off on the pinned bareguard

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…mr 2a)

runLookWithShape (the look door default) runs the page map or sheet look,
then ONE shape call (shape-map): the AI only suggests; authorCloseForJob
builds it with buildShapes before askRounds, so 'rules or shape could both
fit' is one counted ask. The row-edit turn proposes no shape any more
(shapeoffer, the itemShape property and the authorflow wiring are removed,
one path only). OFF unless this bareguard carries itemShape.

Failure: run mwmmcjxo (research-1): a list-of-items job's per-item rules (title, status one of in progress/closed/finished, summary under 200 characters) could not be checked; two went to unbuildable fact checks with confusing words, one Not checked
Addresses: the shape of a list-of-items answer was proposed after the asks, so its ask could not be put to the person
Corrects: b99ed8f2's row-edit offer (removed in this commit)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ail)

An unreadable, failed or capped shape look leaves the look PRESENT with
no shapes (the rules take today's path), its spend stays booked, and the
reason is a draft-log line: 'the shape look did not complete: <reason>;
going on without a shape'. The sheet look is unchanged.

Failure: run mwmmcjxo (research-1): a list-of-items job's per-item rules (title, status one of in progress/closed/finished, summary under 200 characters) could not be checked; two went to unbuildable fact checks with confusing words, one Not checked
Addresses: a failed optional shape look would have stopped a whole draft
Corrects: a02b145d's stop on an unreadable shape look

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
Rows read 'name · sentence' in the person's own form (under 200
characters), one code-owned heading above a shape line's rows, the
[default] row is drawn only for an item count, and Not checked says
whether each one is true. Same renderJobPlan for the card and the Job
tab; Suggested shape rows reach the final card's Add.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
Clean npm ci of package.json + lock in a scratch copy resolves 0.37.0 with
no peer error; typecheck clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
SHAPE_KINDS, nameMaxLength, nameForbiddenChars and reservedNames are read
from rubricVocabulary.rules.itemShape (0.37), one place; the reserved-word
refusal (wrong-value included) is on; the shape look is on by default.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
PART_CLASSES gets shape, rubricStepShapeReds checks it (names, kinds,
count, and bareguard's own createRubric over the itemShape check) and
refuses a shape beside a fact or sheet part; SIGNED_KINDS knows the shape
row; buildStepSpec adds the itemShape check over results; lineBriefs tells
the worker the file contract, each field in plain words and one code-built
example item; the block validates the part before its rows are built.
End to end: a scripted tracker job drafts, signs, and a worker answer with
item 3 status done and item 4 summary of 236 characters reds exactly those
two in plain words.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…port)

hamr signed it 2026-11-18 with two fixes: item names are plain text, and
an Audit row's red reason sits on its own line under the row.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
A shape line's section is the shape-checks line, a markdown table (items as
rows, fields as columns, a mark and the short value per cell), the Not
checked line and a Failed list when an item failed; the earlier '#### Item
N' blocks are gone. One view model (src/shaperesult.js) serves the report
and the Result tab grid.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…signed mock

Result tab: a shape line is a grid, items as rows (plain-text names) and
the signed fields as columns, a mark and a short value per cell, stats,
the plain errors and the Not checked line. Audit: one row per field with
the reason on its own line (shape rows only), the heading, Suggested and
the Not checked row. A red run's item reds are read back from the close's
plain gap lines (shapeItemOfPlain, round-tripped for every reason); a
final close records them (items) beside its rows.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…d list-line floor words

A red run records only the close's plain gap lines, so a shape line's rows
are read from them through the same row mapper (rowResultsFromRecords);
the Audit keeps the heading and Suggested; the floor row of a shape line
says it lists items. The grid's item name gets the button's row height.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
A field's own rule saying under / fewer than / less than / below N signs
the exclusive form (a count N-1; a number or money lt N) and at most / up
to / no more than N signs N, whatever the AI sent. The wording reader is
checkpick.js's (underNumbers, atMostNumbers), the same law as under N
bullets. Seen in a scripted screenshot (not a run): 'under 200 characters'
was signed 200, so a 200-character summary passed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
… loop's call as it was

The worked example for a url field is bareguard's own (rubricVocabulary
itemShape example), not a literal in src; resultLineCard keeps its three
arguments and the shape card reads the tab's own detail and data.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ield)

The floor row's listLine flag was set to false on every row of every job,
which changed the archived non-sheet cases' rows (the pre-S4 golden). Only
a shape line's floor row carries it now.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
matchRule tries today's truth check first; only a rule that ends with none
and no open ask takes the shape field proposed for it (the strongest check
wins, no ask). A line can hold a fact part and a shape part (the shape
rows live only on a line where a rule took its field, so a page job's rows
stay byte-identical). The merged itemShape is closed and lists quote and
values (one optional field per signed fact), no claim; the count stays
with the fact part. Every shape row shows [shape] and has Drop, the
person's own ~ rows too (the rule goes Not checked; Revise brings it
back). Pinned on real archived answers: an extra key in values is reported
once by itemShape, the real dell answer is green with the shape check, a
bad field plus a made-up quote give exactly two gaps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ld both fit

Rules first (hamr 2026-11-19): the look's answer form drops alsoFitsRules;
code decides which rule takes a shape field after its truth check.

Failure: run mwmmcjxo (research-1): a list-of-items job's per-item rules (title, status one of in progress/closed/finished, summary under 200 characters) could not be checked; two went to unbuildable fact checks with confusing words, one Not checked
Addresses: the shape look's answer carried a question (alsoFitsRules) that rules-first code never asks
Corrects: the shape look's answer form from the earlier offer commits

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
A line whose rules took shape fields keeps its $cited default: the step
holds the shape part and the cited part, the file is text + cites +
results, the brief says text line AND list line, and the floor row says
the line carries text. Without it a page job lost its cite default to the
shape (found by the round-2 scope harness's dry run).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
hamr0 and others added 28 commits November 19, 2026 10:29
…OUND_WHY, not the generic no-rule reason"

hamr ruled no extra code path for a hidden reason text. The prompt-rewrite CHANGELOG bullet (0c38f64) is kept; only the sentence about the reason fix is dropped. NO_BOUND_WHY stays in src/shapebuild.js.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ind word, CHECKED / NOT CHECKED / SUGGESTED, reasons behind why)

hamr signed mock A (2026-11-19). One block per job line; rows read [you|agent|default] then truth, shape or file; the old "Shape checks", "other checks" and "Always checked" groups are gone, their rows join CHECKED as [default] file. The model (jobplan.js) now carries each check's kind and the Not checked rows split into row and why (ncRows). Tests pinning the old headings/markup updated; checklist-look pins the new order and tags.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…r job line, tags, kind words, result at the right)

hamr signed mock (2026-11-19), Audit half. auditChecksOf now also returns lines (one block per job line with its CHECKED and NOT CHECKED rows) and jobNotChecked; every row carries its kind. The page draws Line N + text, CHECKED, NOT CHECKED with its why, the result word at the right of a row and a failed row's reason indented under it. The flat checks/notChecked/always lists the Result tab reads are unchanged. Suggested is card-only per the mock.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ths and why marker after the screenshot

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
… line's output-file floor check is kind file

Failure: live run mwndngl5 Audit showed only the guards under Every line; the Line 1 block was dropped from the run detail.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ted; no fact checks it, it stays a shape candidate

Failure: run mwndngl5 (panel session smwndm6b9w6ee): the look tied a Notes: fact to rule 3 and also said in cannot that the page cannot show it; code ignored the cannot, so rule 3 never reached the shape look.
Addresses: exact line+rule cannot entries now remove the rule tie in applyPageRules (the one owner of ties); a whole-line cannot contests nothing.
Corrects: nothing in a prompt register changed; the rule stays Not checked or takes a shape field.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…nt "line 1: " prefix (card and Audit, one owner in finishPlan)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ECKED/NOT CHECKED/SUGGESTED groups, tag then kind, EVERY LINE block), read from the card's own plan

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…bs, next item, POC first

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…ne's own text (copied in rule from fwd v0.36.0 proseCheckedOf; a note-only, quote-only or Not-checked-rule line reads not checked)

Failure: a facts line with text whose only check was the default non-empty note read '### The answer' in the .md report (fwd found the same on a sheet line).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…t not checked (matches fwd v0.36.0 proseCheckedOf: only a Not checked `~` rule row does)

Failure: a text line with a built maxWords still read "The AI's answer (not checked)" because every text line carries the `$` prose row.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…w per item that opens to its fields (signed mock A2); 1-4 fields stay the grid

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…old); the fold drops the separate errors list (its reasons sit on the opened items)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…id, SHAPE_FOLD_MIN, the fold flag, the column hints and the grid's errors list are removed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
… racing phase snapshot (CI run 44538689435)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJLHoQhAAAY2ZFmhYWuL5B
…gression)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… Add back, Revise keeps a Drop

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ebox for the close, approvals, headless hire)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… evidence)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s and lists, schedules; fix rerun money and fixed check-type claims

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…raming)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ld, Add back

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@hamr0
hamr0 merged commit fc56e1f into main Nov 19, 2026
2 checks passed
@hamr0
hamr0 deleted the feat/shape-jobs branch November 19, 2026 16:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant