菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| 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. |
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.
sqlcmd -S "(localdb)\MSSQLLocalDB" -d StreetSamurai (Windows Auth, read-only for audits).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.audit-outlines-<today>/logic/<CODE>.md. Fix files → .../fixes/<CODE>/.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.DELETE from Nodes/Beats/NodeBeats. Soft-disable = UPDATE NodeBeats SET IsEnabled=0.
Snapshot SortKeys to a backup file before re-seating.powershell -File tools/codex.ps1 digest then doctor — must PASS.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.
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.
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.
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.
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).
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.
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.
Build the all-universe Story Completeness dashboard as an HTML artifact — per-story coverage (bible↔blueprint↔beat), chapter count, blueprint granularity, open gaps, node score, standard deviation, and ballot count, across every universe. Usage /completeness [--fresh]. --fresh re-runs coordination first (no LLM); default reads the latest reports/coordination/*.json.
Stage, commit, push, and print the commit hash. Use /commit for auto-message or /commit "Message" for custom.
Economical node evaluation in one call — a sampled panel casts cheap score-ballots (overall + per-beat + one gripe), clusters them into a Pareto/contested/seam study, upgrades the most informative few to full prose, and prints one report with the node score + 95% CI. Usage /review-node [slug-or-id]; defaults to the most-recently-edited node.
Launch the StreetSamurai Writer host and open it in Chrome at localhost. No arguments needed.
Discard all uncommitted changes, reverting to the last commit. Usage /discard
Revert a specific commit by hash. Usage /revert <commit-hash>