bareloop 0.55.0 → 0.57.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +82 -0
- package/README.md +3 -3
- package/bareloop.context.md +69 -209
- package/package.json +2 -2
- package/src/authorflow.js +164 -340
- package/src/authoring.js +7 -127
- package/src/authorjob.js +139 -472
- package/src/authorreadout.js +24 -368
- package/src/authorrun.js +21 -215
- package/src/behaviour.js +14 -0
- package/src/blockasks.js +43 -0
- package/src/bundlerun.js +4 -1
- package/src/checkblock.js +798 -0
- package/src/checklist.js +128 -0
- package/src/checkpick.js +39 -312
- package/src/checkwords.js +3 -1
- package/src/cli.js +3 -3
- package/src/declaredclose.js +38 -108
- package/src/destpoint.js +4 -2
- package/src/draftspend.js +46 -0
- package/src/facts.js +8 -1
- package/src/index.js +3 -43
- package/src/interviewrun.js +13 -8
- package/src/job.js +20 -1
- package/src/jobblock.js +11 -46
- package/src/judged.js +13 -1229
- package/src/kinds.js +5 -347
- package/src/noteask.js +124 -0
- package/src/pagedry.js +27 -5
- package/src/pagemap.js +15 -3
- package/src/pagerules.js +18 -3
- package/src/panel/authorroutes.js +24 -6
- package/src/panel/authorsession.js +222 -177
- package/src/panel/importrun.js +0 -1
- package/src/panel/index.html +198 -26
- package/src/panel/jobplan.js +87 -33
- package/src/panel/server.js +74 -68
- package/src/panel/status.js +3 -0
- package/src/pickrecord.js +34 -55
- package/src/placerule.js +355 -0
- package/src/planrun.js +2 -48
- package/src/promptregisters.js +13 -17
- package/src/providerrows.js +0 -36
- package/src/providers.js +4 -25
- package/src/questionbox.js +104 -0
- package/src/rowedit.js +115 -0
- package/src/rubricstep.js +234 -127
- package/src/run.js +11 -16
- package/src/userrun.js +60 -112
- package/types/authorflow.d.ts +41 -197
- package/types/authoring.d.ts +1 -14
- package/types/authorjob.d.ts +35 -59
- package/types/authorreadout.d.ts +10 -88
- package/types/behaviour.d.ts +12 -0
- package/types/blockasks.d.ts +36 -0
- package/types/checkblock.d.ts +376 -0
- package/types/checklist.d.ts +134 -0
- package/types/checkpick.d.ts +8 -140
- package/types/cli.d.ts +1 -3
- package/types/declaredclose.d.ts +7 -33
- package/types/destpoint.d.ts +2 -0
- package/types/draftspend.d.ts +21 -0
- package/types/index.d.ts +2 -4
- package/types/job.d.ts +8 -0
- package/types/jobblock.d.ts +8 -38
- package/types/judged.d.ts +7 -357
- package/types/kinds.d.ts +3 -104
- package/types/noteask.d.ts +76 -0
- package/types/pagedry.d.ts +13 -0
- package/types/pagemap.d.ts +4 -0
- package/types/pagerules.d.ts +7 -0
- package/types/panel/authorsession.d.ts +47 -5
- package/types/panel/jobgrid.d.ts +8 -8
- package/types/panel/jobplan.d.ts +48 -10
- package/types/panel/server.d.ts +27 -18
- package/types/panel/status.d.ts +5 -0
- package/types/pickrecord.d.ts +20 -13
- package/types/placerule.d.ts +194 -0
- package/types/planrun.d.ts +1 -22
- package/types/providerrows.d.ts +0 -31
- package/types/providers.d.ts +3 -19
- package/types/questionbox.d.ts +87 -0
- package/types/rowedit.d.ts +32 -0
- package/types/rubricstep.d.ts +100 -48
- package/types/run.d.ts +9 -19
- package/types/userrun.d.ts +0 -1
- package/src/calibrate.js +0 -702
- package/src/cardauthor.js +0 -502
- package/src/fitcheck.js +0 -166
- package/types/calibrate.d.ts +0 -201
- package/types/cardauthor.d.ts +0 -358
- package/types/fitcheck.d.ts +0 -113
package/CHANGELOG.md
CHANGED
|
@@ -5,10 +5,92 @@ All notable changes to bareloop are documented here. Format:
|
|
|
5
5
|
[SemVer](https://semver.org/spec/v2.0.0.html). Pre-1.0: **minor** = a ladder rung or
|
|
6
6
|
feature lands, **patch** = docs, fixes, scaffolding.
|
|
7
7
|
|
|
8
|
+
## [0.57.0] - 2026-11-14
|
|
9
|
+
|
|
10
|
+
### Removed
|
|
11
|
+
|
|
12
|
+
- **BREAKING: the old judge's runtime engine is gone (hamr 2026-11-14: "i don't need old runs in anything").** `src/judged.js` keeps only `resolveJudge` / `resolveJobJudge` (the judge identity the spec stamp and the workflow key read); the locate call, the owned rulebook, `decide`, the calibration set's shape and `defaultJudgeLoop` are deleted. The `judged-floor` kind is no longer a live kind or in `KIND_CATALOGUE` (`RETIRED_KINDS` and `judged-floor-gone` go with it), `closeDecl.calibration` is an unknown field, and `runJob` / `runPlan` / `cliMain` no longer take `judgeProvider`, `judgeModel` or `judgeRates` (no `JUDGE_API_KEY` is read; `buildRunnerProviders` returns `{provider, providerFor}`). Library exports removed: `JUDGE_RULES`, `JUDGE_RULE_IDS`, `runLocate`, `decide`, `CALIBRATION_SIZE`, `validateCalibrationSet`, `expectedOf`, `JUDGE_ATTEMPTS`, `MAX_JUDGED_PATHS`. The panel Summary box no longer shows a `judge:` line. A signed spec with a `judged-floor` stage or PASS/FAIL lines now also refuses at `run-u --spec` in the plain words ("draft it fresh"), as it does at Reuse, at a scheduled fire and at the bundle door. Kept: the learnings ledger and how it reads old rows, and the spend readers of an old spine's `judge-round` records.
|
|
13
|
+
- **BREAKING: PASS/FAIL judge examples are gone (hamr 2026-11-14, rubric vs deterministic split, piece 1).** A `~ PASS:` / `~ FAIL:` line is refused at $0 in plain words ("PASS/FAIL examples are gone — write what you want as a ~ rule."), on the panel and the CLI. `jobblock.js` no longer exports `renderExamples` or `jobBlockFitsCheck`, a parsed job line has no `examples` field, the soft-green question set is The job alone (no Judge Examples answer), and the hint text and the panel's rubric placeholder no longer mention the examples. The `~` / `!` / `$` check list (with Drop and Add) covers what the examples did.
|
|
14
|
+
- **BREAKING: the old judge is gone (hamr 2026-11-14, rubric vs deterministic split, piece 2).** `src/cardauthor.js` (the judge-card compile) and `src/calibrate.js` (the 10-case calibration and its signing gate, gate 4 of `prepareSigning`) are deleted, with their library exports: `proposeJudgedArtifacts`, `signJudgedArtifacts`, `foldJudgedArtifacts`, `runCalibration`, `INJECTION_LOCATE_BATTERY`, `CASUALTY_AXES`, and `authorjob`'s `seedProofIsCalibration`. `prepareSigning` takes no `judgeLoop`/`judgeModel`/`onJudgeCost`/`calibrateFn`/`ceilingUsd`/`priorCalls` and returns no `gates.calibration`; `authorCloseForJob` takes no `judgeModel`/`signerFix`/`proposeFn`/`compileOpts` and returns no `judged`. `judged-floor` is no longer on any composer menu (`classMenu`), and a composed or signed close that still carries one is refused at $0 as `judged-floor-gone` ("The old judge is gone. Write what you want checked as a ~ rule on the check list, or draft the job fresh."). The panel and the CLI author run lose the `calibrating` step, the calibration readout and the judge-key checks. The `judged-floor` runtime engine is removed (see the BREAKING bullet above): an old signed spec with a `judged-floor` stage is refused in plain words, not run.
|
|
15
|
+
- **BREAKING: the fit check is gone (hamr 2026-11-14, piece 3).** `src/fitcheck.js` (v0.45.0, "I found N of M of your asks") is deleted: no paid fit call before the look, no Go on / Stop question, no `fit-stopped` stop, no `fit-check` drafting-log row; `authorCloseForJob` takes no `fitFn`/`fitAsk`/`onFit` and the panel session no `fitFn`. One place for all checks: the check list. "A run needs at least one check" stays, from the check list.
|
|
16
|
+
|
|
17
|
+
### Fixed
|
|
18
|
+
|
|
19
|
+
- check list: a `!` phrase carrying a source (ledger-suggest adds) was judged against every look fact with no source or the same source, on any line, because the money filter in `matchBang` had `&&` and `||` ungrouped; a phrase on a line with no money fact was wrongly Not checked "in none of its results" when another line had one. The filter now reads money facts on this line only, then the source test. Found by /branch-review on this branch; `tests/checkblock-bang-result.test.js`.
|
|
20
|
+
- check list: one rule, one Not-checked row on the confirm card (live run mwgsl7zx: after Revise the person's earlier Drop is kept and the look dropped the same `~` rule again, so the card listed it twice, "the page never shows a time right after ..." and "You chose to leave it Not checked ..."). A `~` rule the check list says Not checked (`CheckBlock.saysRule`) is said by its own row alone, with the person's choice as its reason; the look's line for that rule is no longer drawn beside it (`accepted.notChecked`, the drafter's list, is unchanged). The block itself already held one row (signed list and card); the second row was the confirm card's merge in `runConfirmTurn`. `tests/checkblock-one-row.test.js`.
|
|
21
|
+
- page rules: a label the page ALWAYS writes with a colon right after it ("Departure time: 10:24 AM.") is signed as the page writes it, whether the look gave "Departure time:" or "Departure time" (live run mwgsl7zx: after Revise the look dropped the colon, the time fact read no time and the person was asked to drop the rule). One owner, `withColon` in `src/pagerules.js`, a plain count of the page's text: every place that holds the label has a colon after it, else the word stays as given; the run's checker reads the same signed word, so drafting and run agree. Money and other value labels follow. `tests/time-label-colon.test.js` on the run's real page; the JFK-LAX map test and the golden port now expect the page's own word.
|
|
22
|
+
- **After an Agree, every view shows the replaced rule (live run mwgp0au6).** The chat message after the answers, the plan card's job lines and the saved `card.json` still read "~ not a red-eye" while the signed spec held "departure time between 05:00 and 23:59". The session's own lines now take the check list's rebuilt lines in place (one owner), `card` (what `card.json` stores) follows `state.card`, and the answers message is drawn after the rebuild.
|
|
23
|
+
- check list: one rule in, one rule out (hamr 2026-11-14, supersedes "an agreed answer is a new `~` line"; live run mwgjnrkz showed two `~` rules from one Agree). An Agree, or the person's own typed words, REPLACES the asked rule (same line, same rule number); the check list is the one owner and the job text, card box and signed rows are rebuilt from it (`jobLinesView`). Deleted: follower rows (`parent`, `matchFollow`, `followsOf`, `follow`, `pairWith`/`pairRule`, the "you agreed:" card tail), the F318 `sameRuleWords` dedupe in `answerQuestions`, and the picks record's `added` pairing. A counting rule whose own words are the check ("exactly 3 list items") is built as written with no question (`obviousCheck`), so a replaced rule never re-asks on a re-draft or Reuse. The ask text now says the answer "replaces that ~ rule". `!` and `$` rows already replace, never add beside (`onTheList`, carried choices are by row key). Test-proven; not seen live.
|
|
24
|
+
- rubric: a facts-only line that also asks for written words gets its note (live run mwgjlezb: "Write a one-sentence triage note ... quoting its title word for word" produced results and no note). At drafting, ONLY when a line is facts-only, ONE typed call (`src/noteask.js`, the destination call's provider/model/thinking-off, one retry, no repair) asks per such line whether it also wants written words; code decides: yes adds a `$` default row "line N writes a note — the note must be there" with a countable text part (brief FACTS AND TEXT, text-present floored), no changes nothing, a failed call shows a Not checked row "could not ask whether line N also wants a note". Test-proven; not seen live.
|
|
25
|
+
- rubric: the confirm card's "line N file ... must exist, be complete JSON, and carry text" lists "and carry text" only for a line whose signed close has a text-present check (`lineCarriesText`, derived from the close), not for every line.
|
|
26
|
+
- check list: a `!` phrase the AI adds (ledger-suggest) that is on the page but in none of its result rows ("Rome" in the search box and the "Track prices" toggles only) is Not checked with the look's own sentence ("... is on the page but in none of its results, so it is not checked."); it was built, offered under Suggested, signed, and the run went red every try. One owner for "what is a result row" (`resultQuotes`, `src/pagedry.js`) shared with the look's dry read; with no money fact on the line nothing is guessed. Found live by fwdloop (run walk-m16-flights run-1); the same class as the look's `phrase-no-result` (bareloop run mwg3pov1's flights checks).
|
|
27
|
+
- check list: a `~` money rule written in another currency than the page's fact ("under $150" on a euro price) is Not checked with the plain reason "the amount in your rule is in another currency than the page's prices"; it no longer falls back to the look's or the AI's reading and checks the number on the wrong currency. One spelling of the reason (`WHY_OTHER_CURRENCY`, `src/rubricstep.js`). Found live by fwdloop (run walk-m16-flights run-1, euro page); the same class as bareloop run mwg3pov1's money rows.
|
|
28
|
+
- panel: the Audit's checks fold (and the Result tab's checks box) show ONLY the run's latest close (the last `outer-close` or fix-loop `close-verdict`), not the `outer-close` precheck that runs before the fix loop; a run that went green in the fix loop no longer shows failed rows, and a stage the latest close never reached (it stops at its first red) reads grey "not reached" instead of the precheck's old result (live run mwg3pov1).
|
|
29
|
+
- panel: the Run summary "checks:" line ("N run · N passed · N failed") reads the same latest close as the Audit checks fold, through `checksOf`, not every step exit check plus the pre-fix-loop precheck; a stage the latest close never reached shows "N not reached" and counts as neither; the tries parenthesis is gone (live run mwg3pov1).
|
|
30
|
+
|
|
31
|
+
### Added
|
|
32
|
+
|
|
33
|
+
- **Tickets + rubric goes through the check list (hamr 2026-11-14, piece 5; POC runs mwg70fd7 and mwg3pov1).** A Tickets job on the Rubric check has the ticket as its one input: the ticket line is the only input and is renumbered from 1 in ONE place (`cardInputs`, `src/panel/authorsession.js`), read by the fresh card, the session, the Destination check, the schedule and Reuse (`buildReuseSpec`) alike, so a Reuse no longer dies at signing with `spec-invalid (inputs[0].n expected 1)`. The drafting-time read takes the newest ticket the job has not run (`newestUnrunTicket`) and reads it (`readTicket`), with no ticket-number box; the frozen draft input is that ticket's text, and run start re-reads the same list row. Test seams: `createSession` deps `newestUnrunTicket` and `readTicketFn`.
|
|
34
|
+
|
|
35
|
+
### Changed
|
|
36
|
+
|
|
37
|
+
- docs (hamr 2026-11-14, piece 7): `bareloop.context.md` (header, verdict classes, interview set, job block, rubric-on-files section, retired `judged-floor` note, Tickets + rubric section, exports list) and the README verification lines say what the code now does (no calibration, no judge decides, no fit check, no PASS/FAIL examples); PRD v1.134 contract lines ticked.
|
|
38
|
+
- **An old job (PASS/FAIL examples or the old judge) is not reused or bundled (hamr 2026-11-14, piece 6).** A signed spec with a `~ PASS:` / `~ FAIL:` job-line rule or a `judged-floor` stage refuses in plain words ("This job was made with PASS/FAIL examples or the old judge, which are gone. Draft it fresh.") at Reuse (the panel route and the session), at a scheduled fire (it is a Reuse), and at `bareloop run <bundle>`, before any key, repo or copy is touched. One reader: `oldJobWords` (`src/job.js`). Old runs are not repaired.
|
|
39
|
+
- **Rubric on a repo is refused at $0 (hamr 2026-11-14, piece 4).** A Deterministic check is anything a machine can check, so a repo job is Deterministic; Rubric on a repo is allowed only for a Tickets job. A rubric card whose Inputs line 1 is a repo, and a `soft-green` pick at the CLI interview (it always takes a repo), are refused before anything is made or asked: "A repo job uses a Deterministic check."
|
|
40
|
+
- panel: the Audit checks fold header is lowercase like the other part rows (hamr 2026-11-13); README walkthrough GIF regenerated for the check block.
|
|
41
|
+
- Dependency: bareguard ^0.36.0 (check helpers, ruling #82; existing rubric fingerprints and verdicts unchanged — peer-tested on 90 saved plans and 35 verdicts).
|
|
42
|
+
|
|
43
|
+
## [0.56.0] - 2026-11-13
|
|
44
|
+
|
|
45
|
+
### Added
|
|
46
|
+
|
|
47
|
+
- **Place rules: a phrase check that names a version or a date follows the input instead of being refused as week-only (place rules P1-P6, hamr go 2026-11-11; `src/placerule.js`).** Test-proven; one live paid test passed (0.54.1 to 0.55.0 swapped; a mixed heading stopped at $0). At drafting, code (never the AI) fits a place rule (the newest release heading's version or date) to a phrase fact and strips a heading piece such as `## [0.54.1]` or `2026-11-04` to its bare value(s); a version plus a date become two facts, named plainly ("version", "date"). The signed fact carries the rule, never the run's value. At each run, code alone re-reads the value before any spend: the same value passes, a new value is swapped in, and a rule that finds no value stops at $0 listing every rule in plain words. The week-only decision has one owner, the list row; the door's Accept/Rerun bind the run's own read value; one place-rules record per run goes to the spine; the Job tab shows the row. Old runs are not repaired.
|
|
48
|
+
- **The Audit lists every check the card lists, each with its result (hamr ruling A, 2026-11-13, live run mwenvde5).** The Audit tab had only the built checks; it left out the Always checked guards and the Not checked rows. It now has one fold, "Checks", with the card's own sections in bold: Checks, Not checked, Always checked. Each row carries its tag (`[you]`, `[agent]`, `[default]`) and, after the run, its result: passed or failed read from the run's own recorded close verdicts, "not checked" with its reason, or "no result recorded" when the run recorded nothing for it. The checks were also listed in the drafting table of the Audit; that second list is gone. Test-proven on the real run (tests/audit-checks-section.test.js); not seen live in the panel.
|
|
49
|
+
- **The Audit's write fence row shows the gate's own record (hamr ruling A, 2026-11-13, live run mwenvde5).** The row read "no result recorded" because the fence is enforced per write by the gate, not by a close stage. It now reads "held · N writes blocked" from the gate audit of this run (N = write/edit attempts the gate denied as outside the write scope; allowed writes and denied reads are not counted), "held · 0 writes blocked" on mwenvde5; with no gate audit for the run it says "the gate's record for this run is missing", never "held". The fold header counts "held" as passed and wears one state sign before "Checks" ([✗] if any row failed, else [·] if any row has no recorded result or the gate's record is missing, else [✓]; "not checked" rows do not move it). Test-proven on the real audit (tests/audit-checks-section.test.js); not seen live in the panel.
|
|
50
|
+
- **Live draft 3 fixes (hamr 2026-11-13, session smwen1p12tyhc; F336).** The job's own word wins over the long page name: a `!` fact whose page words are "Boston Logan International Airport" for the job's "Boston" is job-derived and checks `must say "Boston"` (only a row with no word from the job goes to Suggested). A Drop on an `[agent]` row moves it to Suggested with Add instead of removing it; Add puts it back; a Revise carries it. The "Always checked" rows carry `[default]`; the section headings (Suggested, Not checked, How the page was read, Always checked) are bold. Test-proven (the real page and page-map reply of that session, tests/card-live3.test.js); not seen live. Seen, not fixed: after a drop a new suggestion "must say 12 results returned" (a page status line) appeared under Suggested; legal, off unless added.
|
|
51
|
+
- **The confirm card has Drop and Add (hamr 2026-11-13, live drafts smwelq2l6srdb and smwelw0jrax5c).** A job-derived `[agent]` check has a Drop button; the checks the agent suggests that are not in the job sit under "Suggested · not in your job, not checked unless you add it" with an Add button, and only ones code has already proved it can build are offered. Both are instant and cost nothing: no model call, no new draft; the row goes back through the check block and the card is drawn again from the rows. The choices are kept when you Revise, and signing reads the rows. A `[you]` row has no Drop (edit the job). New route `POST /api/author/<id>/row-edit`.
|
|
52
|
+
- **A suggested check code could not build is no longer a Not checked row on the card (hamr 2026-11-13, F-b).** It used to show raw gate text ("code could not build a legal check from it: ... Nothing was signed."). It is now one plain sentence in the Audit's draft lines. A `~` rule you wrote always shows, built or Not checked.
|
|
53
|
+
- **Card tags are by row kind, written by code (hamr 2026-11-13, live drafts smwelq2l6srdb and smwelw0jrax5c).** A `~` row, and a fact built from the person's own rule, is `[you]`; a `!` row is `[agent]`; a `$` row is `[default]`. Rule-built facts used to read `[agent]`, and picks read `[you picked]`. Not checked rows carry their tag too, and the heading "Not checked (the AI's own reading)" is now "Not checked". The confirm card, the signed card, the Job tab and the Result tab share the tags.
|
|
54
|
+
- **The check list splits the agent's `!` checks into job-derived and suggested, with Drop and Add (hamr 2026-11-13, live drafts smwelq2l6srdb and smwelw0jrax5c; block level).** A `!` check whose words (or result count) are in the job line or its `~` rules is built and can be dropped; one that is not is only Suggested: not checked unless the person adds it, and offered only when code has already proved it builds. A suggestion code cannot build is hidden from the card (it is one plain sentence in the Audit's draft lines). Drop and Add are instant row edits through the block (no AI call, every one reconciled), and a Revise keeps them. The signed list gains an optional per-row `facts` (the names of the facts a row built) so every reader can tag each fact by its row. Off by default for older callers; the panel's drafting turns it on.
|
|
55
|
+
- **A `~` rule that says what is NOT needed builds no check (hamr 2026-11-13, live drafts smwelq2l6srdb and smwelw0jrax5c).** "nonstop is not mandatory" used to get the check `must say "Nonstop"`, the opposite of the rule. A rule with a negation (not, no, never, without, except...) can no longer give a phrase check: it is Not checked, "it says what is not needed, so there is nothing to check". A suggested check whose words are in such a rule is refused alone, naming the rule ("your rule 3 says it is not needed").
|
|
56
|
+
- **The panel answer door refuses an answer with no typed choice (hamr 2026-11-12, fwd M10 am32 A).** While a question is open (a look, drafter, fit-check or Destination question), a send with no `choice` is refused 400 `bad-choice` with "This page is out of date. Reload it and answer again. Nothing was saved."; nothing is logged and the question stays open. The page already posts a choice on every answer (buttons, and the box's Send as `rule` or `path`). Other chat sends and the CLI are unchanged. Test-proven (real route, real session); not seen live. fwd's sheet-look part does not apply (bareloop has no sheet look).
|
|
57
|
+
- **The job-plan legend "Done when every › check passes" shows only when the plan has a › check row (hamr 2026-11-12, bareloop's own).** The goal-only confirm card, which has no check rows yet, no longer draws it; the plan card, final card and Job tab keep it whenever they have rows. Test-proven, not seen live.
|
|
58
|
+
- **The fit-check question names the asks it found too (hamr 2026-11-12, bareloop's own).** It read "I found 3 of 4 ... Not found: X"; it now lists the found asks with their lines before the not-found list (found 0 keeps the old shape). Wording only; the prompt sent to the model is unchanged. Test-proven, not seen live.
|
|
59
|
+
- **The fit-check question and the Destination question get buttons in the panel (hamr 2026-11-12, bareloop's own; items D and E).** Test-proven (real panel sessions); not seen live. The fit check shows [Go on] [Stop] and no answer box (Go on = today's "go on", Stop = anything else stops). The Destination question shows [Yes, set it to <folder>] [No, keep <current>] and a box "Or type a different Destination:"; a typed Destination goes through the same $0 check the card's Destination uses (inside the repo, not absolute, must exist) and a refusal keeps the question open with that check's own words; the log row says `yes`, `no` or `typed`. The panel hides each question's typed-words tail (exact suffix); the stored question and the CLI keep it.
|
|
60
|
+
- **A look question's answer is a typed choice in the panel (hamr 2026-11-12, fwd M10 am29 + am30 + am31; item B).** Test-proven (real panel sessions over a real rubric job and the real page of run mve5y8fa); not seen live. The send route takes an optional `choice` (`agree`, `drop`, `rule`); no choice is today's words reading (the CLI, old pages). Agree shows only when the question carries a proposal, Drop it only on a question about one of the person's `~` rules; the box's own words are kept whole as a `~` rule even when they read "agree" or "drop it". The panel shows the question without its `Answer "agree"...` tail (exact suffix; the stored question and the CLI keep it). A bad or not-offered choice is refused in fwd's words and leaves the question open, nothing logged. The `human` drafting-log row also records the `choice`.
|
|
61
|
+
- **A question asked while drafting gets its own question box in the panel (hamr 2026-11-12, as fwdloop M10 am29; item A).** Test-proven (real panel sessions, the page's own functions on stubs); not seen live, phone width not checked. The box shows the head ("A question the plan raised (i of n)"), "About job line N", the question, a labelled one-line answer box with an alert error line, and a wide Send; while it is open it takes the chat input's place and the chat input returns when it closes. The poll rewrites the texts only when the question changes, never what is typed. The question's message waits out of the thread until it is answered. The CLI's terminal asks are unchanged.
|
|
62
|
+
- **"Under N" / "below N" bullets, items or points now build `maxListItems N-1`** (F315 follow-up; "under 1" is no bullets, "under 0" is refused). Before it built the exact `listItems`. "Under N words" and lines are unchanged. Test-proven on real bareguard; not seen live.
|
|
63
|
+
- **Adopted bareguard `^0.35.0`** (was `^0.31.0`); no existing `rubricSha` changes. Test-proven; not seen live.
|
|
64
|
+
- **"At most / up to / no more than / fewer than N bullets" and "at least / minimum / no fewer than / more than N bullets" are now directional list-item checks (F315, BG-5).** They build bareguard's `maxListItems` / `minListItems` instead of the exact `listItems` (which graded "at most 3" red at 2 bullets); "exactly N" and a bare "N bullets" stay `listItems`, and a signed spec with `listItems` is untouched. Both are pickable (`OFFERED`), with plain words, card sentence, red line and worker brief. Test-proven on real bareguard; not seen live.
|
|
65
|
+
- **A rubric job now has ONE list of checks, made once before the first try (check list, Phase A, hamr 2026-11-11).** Test-proven; not seen live. Before, the confirm AI, the page look, the person's picks and the drafter each kept their own list of what a line checks, and they drifted (paid rounds 8-10). Now, after the look and the pick path and before the confirm turn, code builds one list (`src/checklist.js`): a `~` row for each rule the person wrote, a `!` row for each fact the look kept and for up to three count or phrase checks one new call (`ledger-suggest`, its spend joins the drafting cap) suggests from the inputs, and a `$` row for each default code adds itself (every claim quotes the input on a text line of a one-input job; the numbers tie after a fact line). Code proves each `!` row against the input and drops it with a reason otherwise; a `~` rule always wins over a `!` on the same thing. The drafter is shown the list in one code-written block; after each try a part with no row goes to Not checked ("Not on the list of checks made before the draft, so you judge it at Accept."), a fact's result count rides with its fact, and the default cited part is decided after the draft settles (also after the last-try drop). A check that needs exact words holding a version or a date ("## [0.54.1] - 2026-11-04") is refused as true only for the input as it is today ("It names a version or a date from the input, true only for the input as it is today, so you judge it at Accept."). An empty list is logged as "empty list: nothing to check" and ends in the existing no-check stop. Both doors build it (the panel and the author CLI; the CLI asks nothing); there is no switch. The drafting log gets a `ledger` line per row at the build and at the end, and a `ledger-try` line per try. Rubric jobs only; a code job's prompts are unchanged.
|
|
66
|
+
- **The signed close carries the list, and the readers read it (check list, Phase B, hamr 2026-11-11).** Test-proven on 38 real saved drafts of paid rounds 9 and 10; not seen live. `closeDecl.list` holds each row's line, key, from, by, kind, state, rule, part (the part that carries a built row) and reason, and never the row's words (the words stay in the drafting log); the close validator shape-checks it, and no hashing code changed. When a close carries the list: the card's and the Job tab's Not checked read its not-checked rows (one reader, `refusedOf`), the end check is "no built row" (the mechanical floor never counts), retention is counted from its final rows, and signing refuses a close whose list and checks disagree (`listMatchesClose`). A close with no list reads as before. The card after drafting is read from the signed close alone (`cardOfClose`, the same source as the Job tab) and now shows the close's notes too; `reconcileRubricPlan`'s three-source merge, `moved` and the `rulesOfLine` twin are gone. The rubric drafter is no longer offered `covers` or numbered confirmed checks (a legacy `covers` on a saved part still validates).
|
|
67
|
+
- **The run-time worker is told the checks the person picked (R10, fwd am28 "Check 0b", ported by rewrite).** Test-proven; not seen live. A facts line that also carries picked checks (keep-both) now has one plain line per pick in its worker brief, written by code from the signed record of the picks (`closeDecl.picks`), for example `Your "text" must have exactly 3 list items; a list item is its own line starting with "1." or "- " followed by a space and its text.` or `Your "text" must have at most 80 words.`; a check that is not the person's pick is not told, and a close with no picks reads as before. A text-only line with picks is not told yet (fwd tells only its page-quoting steps; same scope here).
|
|
68
|
+
- **A rubric job's confirm turn now plans the goal only.** The turn sends one sentence; any check, gap or question the model sends is discarded by code, so the confirm card shows no check and the drafter is no longer told "the person confirmed" checks (the list is the one plan). The plan card and the Job tab mark a check code added itself `[default]`, beside `[you picked]` and `[agent]`. Phone width not checked.
|
|
69
|
+
|
|
70
|
+
- **A rubric job's checks are now built by code from ONE list; the AI only sends edits to it (check block, hamr 2026-11-12: "one flow one place, ai suggests !, code adds $ and code decides what worked/doable").** Test-proven through the real drafting loop and the real panel session; not seen live. The drafter's reply is a list of edits (add a fact, phrase or result count; read a limit for a `~` rule; pick a check for a counting rule; say a rule cannot be checked), each judged alone: a refused edit says why and costs the others nothing. Code proves every word on the saved page, builds every check, and says in plain words why a row is Not checked. The AI never writes a check or a declaration, and the prompt no longer restates the list. A Tickets job's reply keeps its optional `ticketFilter`.
|
|
71
|
+
- **A draft revises only when an edit was refused or the starting-point read failed, at most twice (check block, ruling 3).** Before, every failed gate sent the drafter back. A plan code built that fails its own gate now stops as a build fault in plain words ("the plan code built from your checks did not pass its own gate ... a fault in the build, not in your job"), and the AI is never asked to fix it. Test-proven; not seen live.
|
|
72
|
+
- **Code reads a limit from the person's own words (hamr ruling A, 2026-11-12).** When the look picked the shape for a `~` rule but gave no limit, "under $450", "at least $200", "between $100 and $300", "between 05:00 and 21:59" are read by code; a rule with two limits, a negated one, a one-sided time or a 12-hour time stays Not checked with a plain reason. Test-proven on a frozen table of 42 rules; not seen live.
|
|
73
|
+
- **The progress lines of a rubric draft say edits: "N edit(s) kept, M refused" and "the list at the end: N built, M not checked".** Test-proven.
|
|
74
|
+
|
|
75
|
+
### Fixed
|
|
76
|
+
|
|
77
|
+
- A rule like "under $450" is now checked as under, from your own words, even when the page reader wrote at most. A page-reader or AI reading that disagrees with a limit code can read in your words is refused with a plain reason; words code cannot read behave as before.
|
|
78
|
+
- **The confirm card of a rubric job now shows the checks the draft built (hamr 2026-11-12, live draft session smwe2qz5g3xrq).** It showed only the summary, the job line, "How the page was read" and "Always checked": none of the rules, facts or the result count the list had built, and no Not checked row. It read the confirm turn's own plan (the goal only) instead of the list's rows. It now draws each row under its job line, and each Not checked row with its reason, through the same readers as the signed card, so the two cards cannot differ. The model's own plan is still the goal only and what is accepted is unchanged. Test-proven (real page map, real session, fixture of the live rows); not seen live. The short "Your answers are in" bubble after a question still shows the lines and your picks only (the full card follows it).
|
|
79
|
+
- **A fact the look tied to a `~` rule carries that rule by code when it has a bound (hamr 2026-11-12: "checks that are checkable and agreed should be part of the plan and should be verified").** On a job with "~ under $450" the look ties the price fact to rule 1 and drops its list row ("a ~ rule wins"); the drafter then put `rules` on the fact or nowhere, the fact was refused "not on the list" and the rule went unchecked (a draft stopped after 3 refused tries; another run went green with the rule not checked). Now the part carries the rule number by code, and the fact's bound is the drafter's own `rule`, else the look's reading, else (a phrase fact) its own fixed words; with no bound the rule stays not-checked. Test-proven on the saved session shape (smwdiy4k3bk0v); not yet seen in a fresh live draft.
|
|
80
|
+
- **The Job tab now names every Not checked row of the check list, and shows the one the list dropped for naming a version or a date (F318, first live run mwc2u3yh).** Test-proven on a copy of that run's own signed spec and draft log; the real screen is re-checked by hamr. A phrase row the plan did not build read ` — the plan did not build it` with no name; it now reads `phrase "### Fixed" — the plan did not build it` (the words come from the drafting log; with no log, a CLI or imported run, from the row's key: `phrase "### fixed"`). A row dropped because it names a version or a date used to be missing from the tab; it now shows under its line with its reason, as the card preview of the build spec says. Other dropped rows (a `~` rule wins, words not in the input) stay out.
|
|
81
|
+
- **The drafting plan card's Not checked now names a phrase row the same way the Job tab does (F318).** Test-proven; not seen live. It read `line 1: rule phrase — the plan did not build it`; it now reads `line 1: phrase "### Fixed" — the plan did not build it` (one owner, `listRowName`; the drafting log's words, else the row's key).
|
|
82
|
+
- **Answering "agree" to a reading that is the rule's own words no longer adds a second identical `~` line (F318).** Test-proven (real panel session); the pick now cites the rule that is already there. Case, spacing and a trailing full stop do not make words different. Agreeing to DIFFERENT words still adds a `~` line, as before.
|
|
83
|
+
- **A fact the look tied to a counting `~` rule ("exactly 3 bullets") now stays in the plan beside the pick, and a pick follows the line's fact on the card (check block).** The tied fact used to be swallowed; the card drew picks first. Test-proven; not seen live.
|
|
84
|
+
- **A word the draft gate would refuse now makes that one check Not checked, in the gate's own words, and the draft goes on (check block).** A label the look dropped or never kept, a unit or currency sign not on the page, words found fewer times than the line's result count, an input the job does not have, or no input named when the job has two used to stop the whole draft. Test-proven; the real run mwdhma1l's "12 results returned." counted once for 3 results is one of the cases; not seen live.
|
|
85
|
+
- **A file job's "I can't check this" question is fwd's again, and a proposal comes before it in rule order (check block).** The look's questions come first, then the pick questions in rule order; the Drop it button on a can't-check question drops the rule. All questions go to the person whole, so Agree shows on a proposal. The picks record is now written on every draft (a Revise that asks nothing still carries the picks), and a rule a broken pick pair rides is not asked on its own. Test-proven in real panel sessions; not seen live.
|
|
86
|
+
- **A rule the person left reads "You chose to leave it Not checked, so you judge it at Accept." and a rule they answered in their own words reads "Your own words are on the line; the machine has no check for them, so you judge it at Accept." (check block).** The words the old chain used, kept. A job line no check could be built for is listed under Not checked as a line, never as a draft stop. Test-proven.
|
|
87
|
+
- **A version or a date a tied `~` rule does not name is no longer signed as fixed words (check block, hamr 2026-11-11 one owner).** It is placed by a place rule or stays Not checked; a version and a date in one phrase are two facts named "<name> version" and "<name> date". Test-proven; not seen live.
|
|
88
|
+
|
|
8
89
|
## [0.55.0] - 2026-11-10
|
|
9
90
|
|
|
10
91
|
### Added
|
|
11
92
|
|
|
93
|
+
- **A rubric job whose declaration ends with no real check is refused before signing, whatever the seed gate says (R10 end check, hamr ruling 2026-11-10).** Test-proven; not seen live. Paid round 6 draft 17 (session smwag7ncubmip) ended signed-ready with a rubric-step holding no part plus files-changed: the earlier stop fires only when the seed gate reads green, and an empty rubric-step reads red at seed. Now, after the declaration is settled and before the signing gates, both doors (the panel `finish`, the author CLI) refuse such a close with exactly the same words (`noCheckStopText`, one owner in `src/authorreadout.js`); a part of class none or guard no longer counts as a check. No prompt text changed.
|
|
12
94
|
- **A rubric / page job whose close has no check that can fail now stops with words that say so (R10, hamr ruling 2026-11-10).** Seen live as draft 18 of paid round 5 (session smwabidzpqvgt: the only stage was files-changed, so signing refused with the repository sentence); the new words not seen live. Test-proven. When signing is refused because nothing fails on the seed and the close is a text-lang close holding no check part (only files-changed, or rubric-step stages with no parts), the stop reads "This job has no check that can fail..." with a to-do to add a `~` rule or answer the question. Repo jobs and rubric closes with a real check keep the old words. This stop fires only when nothing in the close fails on the seed. A plan can still end with zero checks (after a drop, or when every part goes to Not checked), seen live in paid round 6 drafts 2 and 17; an end check that closes this is the next release's first fix (F309). `isNoCheckRubricClose` and the `closeDecl` argument of `signingStopText` are in `src/authorreadout.js`; the CLI (`src/authorrun.js`) and the panel (`src/panel/authorsession.js`) both pass the close.
|
|
13
95
|
- **A rubric draft whose last try still has a broken AI check now drops that check to Not checked instead of dying (R10, PRD v1.131 tick 4).** Seen live in paid round 6 (3 drafts); round 5 had 0 firings. After the LAST try failed the plan check, and only when every red points at one part the AI wrote (not a pick, not a guard, not a check code built from the person's picks), code drops those parts: a pure extra becomes a `line N: "<words>"` note, a part that carried a `~` rule no other part carries becomes that rule's Not checked row, both with `The AI couldn't build a working check for this, so you judge it at Accept.`; the plan is then checked once more. Any red on a pick or on no part leaves the plan refused as before. A $0 replay of the 7 saved dead drafts of paid rounds 1-4b rescued the 3 that were still red (round 4 d8, round 4b d1 and d4) with both agreed picks and the picks record identical, and 21 passing real declarations came out unchanged. Rubric jobs only. The retention record counts each drop as `dropped`. `dropExtras` is in `src/checkpick.js`, `dropAfterLastTry` in `src/authorflow.js`. Caveat (F309): when the drop removes every AI part of a line the rubric-step is left with `parts: []` and the plan is still accepted, so it can sign with zero checks (seen live, paid round 6 draft 2); the next release's first fix is an end check for this.
|
|
14
96
|
|
package/README.md
CHANGED
|
@@ -25,7 +25,7 @@ bareloop finds that path for you. Write your problem in plain language, as steps
|
|
|
25
25
|
The harness (budget, time, what it may touch) and the judge both sit **outside** the agent loop. The agent can't judge itself. Every run is recorded, so you can audit it and replay it at $0 and see exactly what happened.
|
|
26
26
|
|
|
27
27
|
<p align="center">
|
|
28
|
-
<img src="https://raw.githubusercontent.com/hamr0/bareloop/main/docs/media/panel-walkthrough.gif" alt="The bareloop panel: write a job, the runs list, a rubric run
|
|
28
|
+
<img src="https://raw.githubusercontent.com/hamr0/bareloop/main/docs/media/panel-walkthrough.gif" alt="The bareloop panel: write a job, the runs list, a rubric run with a fixed try, its Result, the Audit checks fold and the signed Job with its check rows, then a repo run ready to merge" width="900">
|
|
29
29
|
</p>
|
|
30
30
|
|
|
31
31
|
## The machine that makes machines
|
|
@@ -36,7 +36,7 @@ Like a CNC machine, bareloop builds the thing that does the work, then lets a ga
|
|
|
36
36
|
|
|
37
37
|
**Guide rails, not just guardrails.** Guardrails stop the agent going off the road: a money cap, a time cap, and a read/write fence it can't move. Guide rails keep it on the road: each new worker hears what earlier tries read, searched and changed, diffs included; two misses redraw the plan; past lessons go to the planner; an unclear job gets up to 3 questions, not a guess.
|
|
38
38
|
|
|
39
|
-
**Verification.** The check sits outside the agent's loop: code
|
|
39
|
+
**Verification.** The check sits outside the agent's loop: code decides. Anything code can't check is listed as Not checked, and you judge it at Accept. A red goes back with the real error to be fixed, until it's green or out of budget.
|
|
40
40
|
|
|
41
41
|
**Evidence.** Every call, every allowed or blocked read and write, every check result and every cent is recorded. You can audit it in the panel and replay it at $0. Research facts are quoted from saved sources.
|
|
42
42
|
|
|
@@ -45,7 +45,7 @@ Like a CNC machine, bareloop builds the thing that does the work, then lets a ga
|
|
|
45
45
|
**Deterministic (code)**: code decides. Tests pass, the linter is quiet, the build exits clean. If no check exists yet, say what you need and the agent builds that check first.
|
|
46
46
|
> *"Fix the lint errors."* · *"Add doc comments to every function."* · *"Split this 3,000-line file without breaking a test."*
|
|
47
47
|
|
|
48
|
-
**Rubric (research)**: it searches, collects, compares and hands you a report where every fact is quoted from a saved copy of its source.
|
|
48
|
+
**Rubric (research)**: it searches, collects, compares and hands you a report where every fact is quoted from a saved copy of its source. An AI only points at the page's own words, and fixed rules in code decide.
|
|
49
49
|
> *"Find flights LAX→SFO under $400, nonstop, no red-eye."* · *"Compare these 5 laptops on price, weight and warranty."*
|
|
50
50
|
|
|
51
51
|
## What it won't do
|