logic-sweep

name	logic-sweep
description	Run the canonical Logic Sweep (docs/LOGIC.md, SS-A44) on one or more stories — six-dimension logic/continuity audit, severity triage, minimal-splice fixes, verification. NO votes, NO panels. Usage /logic-sweep [slug ...]; no argument = every story with beats changed since its last sweep.

/logic-sweep — the canonical QA pass (SS-A44)

The law lives in docs/LOGIC.md. This skill is the runbook. Never launch votes, panels, or reviews from this skill — if the user wants a score, they will say so explicitly.

Fixed facts

• DB: sqlcmd -S "(localdb)\MSSQLLocalDB" -d StreetSamurai (Windows Auth, read-only for audits).
• Live story = enabled beats in reading order: JOIN NodeBeats nb ... WHERE nb.IsEnabled = 1 ORDER BY nb.SortKey. Books: chapter children by Nodes.SortKey, beats within by nb.SortKey. NEVER order by Beats.Number.
• Reports → audit-outlines-<today>/logic/<CODE>.md. Fix files → .../fixes/<CODE>/.
• Text pushes via built exe v3\StreetSamurai.Cli\bin\Release\net10.0\StreetSamurai.Cli.exe --beat update --id <guid> --text - with OS-level < redirection (PowerShell pipelines inject BOM; verify UNICODE(SUBSTRING(Text,1,1)) != 65279 after every push). One CLI invocation at a time (port). dotnet run may silently drop args — prefer the exe.
• NEVER DELETE from Nodes/Beats/NodeBeats. Soft-disable = UPDATE NodeBeats SET IsEnabled=0. Snapshot SortKeys to a backup file before re-seating.
• Doc edits → powershell -File tools/codex.ps1 digest then doctor — must PASS.

Steps

1. Resolve scope. Args = those slugs. No args: find stories whose beats changed since their last sweep (compare Beats.UpdatedAt / NodeBeats changes against the story's most recent audit-outlines-*/logic/<CODE>.md date; when in doubt, include it). Small edit → scoped sweep (touched beats + blast radius); structural change (merge/disable/reorder) → full-story sweep.

2. AUDIT (report-only agents, one per 1–2 stories, Sonnet). Each agent reads the story end-to-end and reports on the six dimensions of docs/LOGIC.md §3: causality chain / knowledge states (incl. names-before-coining) / timeline (respect bible-marked deliberate impossibilities) / plant-payoff ledger incl. arithmetic walked end-to-end / orphan references (grep distinctive phrases of every disabled beat; check IsChapterStart stranded on disabled beats) / bible agreement. Findings: severity (BLOCKER/MODERATE/MINOR) + SortKeys + quoted text + minimal fix proposal. Instruct verbatim: "Do not invent problems — if the logic holds, say so." Bible agreement is checked ACROSS ALTITUDES (docs/LOGIC.md §8): ss --altitude-audit --slug <slug> automates the 10,000↔100 ft comparison (bible/blueprint vs chapter synopses; findings filed as OutlineDrift); agents may read the story's story-synopsis.txt (or NodeChapterSummaries) for cheap chapter-altitude scoping before deep beat reads. Cross-story canon (shared characters, series arcs): one additional cross-read agent over the affected set when a series story is in scope.

3. VERIFY QUOTE GROUNDING (mechanical, before triage — docs/LOGIC.md §SS-LOGIC-4a). Every finding that quotes beat text gets checked against the DB before it's trusted, not after: build the claim list (BeatId + quoted text) from all audit reports and run ss --verify-quotes-batch --json-file <path> (array of {"beatId":"<guid>","quote":"<text>"}) — or ss --verify-quote --id <beatId> --quote "<text>" --claimed-by "<agent>" one at a time. MCP: VerifyQuoteGroundingBatch / VerifyQuoteGrounding. Any Fail = that finding is misattributed or fabricated — drop it from triage and, if it seems worth chasing, re-read the actual beat yourself before deciding. This is cheap and mechanical; run it on every quoted finding, not just the ones that feel off.

4. TRIAGE. Fix all BLOCKERs; MODERATEs almost always; MINORs when one word/clause. If a finding can't be named concretely, drop it. Editor-taste items go to the deferred ledger, not the fix pass.

5. FIX (separate agents, the audit report as input). Minimal-splice discipline per docs/LOGIC.md §4: prefer data fix → clause → passage → rewrite; reassign to established cast, reconcile counts to the load-bearing version; bible wins on facts unless the finding proves the bible stale — then fix the bible in the same change (+ digest/doctor).

6. VERIFY (inside each fix pass). Changed passages re-read with neighbors; old-defect greps = 0; disabled-content greps = 0; repaired arithmetic walked and printed; BOM checks; doctor PASS. Honest reporting: deviations and judgment calls are stated, never silently forced to satisfy a checklist.

7. CLOSE. Write/refresh audit-outlines-<today>/logic/<CODE>.md verdicts and, for multi-story sweeps, a CORPUS-REPORT.md. Summarize per story: verdict, findings by severity, fixes applied, anything deferred.