fp-finder-latex
Run bounded independent review and revision for a mathematical prototype, filled presentation or LaTeX manuscript package. Fix the stage criterion, preserve exact versions, findings and usage, and report unresolved outcomes honestly. Simulated review, not guaranteed acceptance.
- License 许可
- GPL-3.0
- 最近更新
- 2026-11-13
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- fp-finder-latex
- description
- Run bounded independent review and revision for a mathematical prototype, filled presentation or LaTeX manuscript package. Fix the stage criterion, preserve exact versions, findings and usage, and report unresolved outcomes honestly. Simulated review, not guaranteed acceptance.
02 资源文件 1
- 资产文件assets/scenarios.json4.5 KB
03 SKILL.md 正文
fp-finder-latex
Parent: LaTeX workflows.
You are the coordinator. Use separate native agents as referees and revisors; sequential dispatch is sufficient. Judge arguments, not author reputation. The result is a simulated review with stated coverage, never human approval or formal verification. Keep a scientifically useful WORKING revision even when further findings remain; only independent review can make an exact submission ACCEPTED under its declared criterion.
Stage criterion
When called by the research-file synthesis workflow, record one of these distinct tracks before dispatch:
- Outline: a prototype/spec, including Markdown when the project forbids scratch TeX. Judge audience, exact targets, dependency order, source coverage, negative findings and explicitly mapped gaps. Open proof obligations are allowed only as named research gaps, outside the outline's proof-verification scope. Missing or misleading gap accounting remains a blocking review finding; outline acceptance never certifies the unproved claims.
- Presentation: the filled candidate and its selected source/evidence map. Judge the promised self-containedness, proofs, imports and narrative. A Markdown candidate with LaTeX excerpts may receive source-level approval; build/PDF checks remain NOT RUN until the integrated artifact exists.
- Integrated manuscript: the declared TeX root and input closure after insertion. Check the actual build/PDF and affected mathematical interfaces; source-level approval alone cannot accept this track.
The coordinator passes the existing selection, audience, plan and authorization. Do not restart intake or force a prototype to meet a finished-paper criterion. Use the same independent referee/revisor loop below, adapted to the recorded stage: inventory planned items for outlines and actual claims for presentations. Freeze all reviewed inputs, including the selection/spec; altered inputs require review of affected dependencies. Label any acceptance with its stage and hash. Ordinary manuscript review continues to require final build/PDF checks.
Preflight and contract
Read the nearest project instructions and canonical-file contracts. Load using-latexpowers, using-mathpowers, referee-report and the relevant shared writer mechanics. Use their scoped proof/source/writing leaves in workers, without restarting the whole intake in every worker. Keep mathematical claim states in the existing LEDGER contract; this skill owns review history, not a competing proof-status system.
Before expensive work, verify inputs and primary sources, distinct-agent dispatch and any proof backend required by this track. Probe TeX/PDF tools for TeX targets; Markdown outline/source review does not require them. Record actual capability and model provenance. If a required dependency is absent, say what is missing, the concrete setup action and which work it blocks; continue only independent work. Do not invent a backend or call inline self-review independent. An explicitly authorized inline mode must retain that limitation and cannot satisfy the independent-review acceptance condition.
Create one run root in the project's permitted scratch area. Record:
- User goal, canonical manuscripts, active-document boundaries, local input closures and cross-paper dependencies; preserve existing work and archives.
- Revision authorization and limits. Existing autonomous authorization uses the manual referee-response route with writer mechanics and an item-to-tag map. Use revise only for its actual accepted-item-list contract. A review-only request permits reports, not manuscript mutation; obtain missing authorization at that boundary, without asking again for authorization already granted. Real received reports retain their chosen adjudication route.
- A fixed criterion and track ID: genre, contribution bar, reader background, correctness/source standards, required coverage and package objective. Do not infer a research-only venue from “high quality.” State necessary assumptions; clarify a consequential unresolved choice. A later narrower genre/claim is a separate track, never retroactive closure of the original.
- Referee/revisor identities and relevant prior authorship. A referee never writes the manuscript revision it judges. Disclose same-task/model limits.
- A positive round budget, default 8, checked before every new dispatch. Record any user-specified time/token limit. A limit is not a verdict.
For multiple papers, first assess their division and dependency graph. Combining or splitting is allowed within authorization, with a claim and finding relocation map, stable aliases and review of affected interfaces. Do not improve apparent acceptance by hiding a difficult claim elsewhere.
One round
-
Freeze. Snapshot sources, bibliography/figures/preambles and available PDFs, with hashes and build/tool evidence. Never overwrite a reviewed snapshot. Record an initial failed build or absent PDF explicitly; source review may proceed, with PDF checking owed for manuscript acceptance.
-
Referee. Dispatch a read-only reviewer using referee-report at proof level. Read the source before weighing the author response. Inventory active claims/proofs and catalogue entries; mark checked, imported and unchecked coverage precisely. Test hypotheses, edge cases, counterexamples, source editions and actual source-to-manuscript maps. Separate correctness, contribution, presentation and the stronger package objective. A citation, compilation or statement count is not proof. Imported theorems can be legitimate journal inputs without making a package self-contained.
Each report gives ACCEPT, MINOR_REVISION, MAJOR_REVISION or REJECT under the fixed criterion. Every finding has a stable ID, original objection and locator, severity/type, evidence, concrete closure requirement and reviewing authority. New objections get new IDs. Preserve counterexamples and source evidence. Reviewers must not manufacture objections for tone, or approve from fatigue, pressure or iteration count.
Before issuing a verdict, the reviewer runs a report-completeness audit. Every material objection discovered in the source, a prior report, a source comparison, or an adversarial check appears in the current finding table with a stable ID, exact locator, closure test, and one destination:
RESOLVED-IN-REPORT,OPEN-TODO,IMPORTED-OBLIGATION,OUT-OF-SCOPE, orINVALIDATED. Findings may not remain only in worker scratch or an ancillary calculation. The report links each TODO to the manuscript response map.RESOLVED-IN-REPORTrequires the exact resolving passage and closure test;OPEN-TODOrequires an owner, evidence requirement, next source or computation, and closure test;IMPORTED-OBLIGATIONrequires the precise source, hypotheses, import boundary, and explicit permission under the fixed criterion;OUT-OF-SCOPErequires the criterion and exclusion reason;INVALIDATEDrequires the defeating evidence. ACCEPT is forbidden while any material finding isOPEN-TODO, or while an imported row lacks those source, hypothesis, boundary and criterion-permission fields.Apply an example admissibility gate to every calculation used as evidence. It must be source-native with exact hypotheses and source interfaces, or interpretation-compatible with an explicit typed map into the interpretation catalogue that preserves the relevant operations. The workflow does not use the labels "toy model", "toy example", or "naive calculation". An illustrative construction that fails the gate is excluded from the verdict and recorded only as a rejected diagnostic, never as a counterexample to the manuscript's source-level theorem.
-
Adjudicate and revise. Check findings against evidence; a referee is not automatically correct. Keep disputed/conflicting items open with reasons until resolved by mathematical evidence or an independent focused adjudicator; seek the user only for a genuinely missing choice. Within existing authorization, a different agent makes tagged, bounded revisions and maps every finding to an edit/evidence locator or an explicit unresolved response. The revisor proposes closure, never awards it. Preserve human comments and dormant material under project rules. Prove new claims or mark their unresolved status; don't insert a plausible argument as proved. The response map must cover the referee's complete finding table and every linked investigation TODO. An imported obligation may remain open only when the fixed criterion expressly permits that import and the report gives its precise source, hypotheses, and boundary.
-
Validate a working candidate. Check the coherent diff and complete item accounting. For TeX targets, build changed roots with isolated auxiliaries and inspect logs, references/citations and compiled prose/formulas. TeX comments can swallow text without a build error. Keep required AI/software statements accurate and at the project's specified introduction location. Record regressions; repair/revert them or leave them explicitly unresolved. A supported partial revision may be retained as WORKING, with unchecked claims visible. It is not an accepted article. Freeze its next submission.
-
Return to the referee. Give the original referee the exact new snapshot, complete response map and earlier reports. Verify each item as CLOSED, PARTIALLY_ADDRESSED, OPEN or RETRACTED, with reasons and revised locators. A retraction retains the original finding and evidence. Check all changed arguments and affected dependencies, including new mistakes. Retain limits for unchanged material; do not claim a fresh complete proof audit from a diff check. A changed hash creates an unreviewed version and invalidates affected dependent claims, without erasing historical acceptance of the old version. No acceptance silently transfers between versions or tracks.
The returned report repeats the completeness table, including findings discovered during revision. A clean build or polished exposition cannot close a mathematical item whose closure test was not met. If an example used in the prior round fails the admissibility gate, remove it from the evidentiary chain and preserve the correction in the response history.
-
Account and decide. Update the per-round records by delegating: [[update-ai-usage]] for provenance (call/agent IDs, requested versus observed models and harness, tool and source use, failed and discarded attempts, source-check limits) and [[update-tokens]] for the priced counters. Both consolidate upward into the repository's master
ai-usage.mdandtokens.md; root summaries reference those records. Their contract — never invent totals or costs, never treat unavailable as zero, label every estimate — isAI-POLICY.md(AI15, AI16) and is not restated here. The per-roundusage.mdindex stays this skill's own artifact: it links the round'sai-usage.mdandtokens.mdrather than duplicating them, and neither delegate writes or consolidates it. Round-local duties that remain this skill's: do not double-count mirrors and do not add overlapping intervals.Accept only when all in-scope papers and affected package interfaces meet the fixed criterion and required coverage on the exact current tuple, with no mandatory unresolved items or unchecked load-bearing proofs under that track. Outline and source-level acceptance carry their stated limits; integrated manuscript acceptance requires final build/PDF checks. Record permitted imported prerequisites. A larger self-containedness goal needs its own dependency closure; article approval does not imply it. Further substantive repairs start another round within budget, not a repeat of the same report to solicit a nicer verdict.
Non-success and completion
After two successive rounds with the same substantive open obligations and no new resolving evidence, reassess strategy. Continue only with a recorded materially different approach within the remaining budget. A blocker needing new research or unavailable evidence can stop its track while independent work finishes. On exhaustion, unresolved disagreement or inability to progress, report NOT CONVERGED with exact blockers and the best working snapshot; never manufacture an acceptance or weaken the criterion silently.
Return manuscript paths, exact reviewed versions, per-track verdicts, closed and remaining findings, package dependencies and accounting links. Preserve all reports and rejected candidates. Clean up only task-created worktrees, branches and disposable files after preserving evidence; no implicit commit, push, submission or publication. Simulated acceptance means only that the recorded reviewers found no remaining mandatory objection under that scope.
This is a thin manuscript adapter, not the generic META-FP formula running unchanged. That formula already permits LaTeX growth; its APPROVING quality floor differs from retaining a partial WORKING revision. To stabilize this skill's own text, use fp-finder-skill; its per-file shrink rule does not govern manuscripts.
Example and validation
Example prompt: “Use fp-finder-latex on these two canonical papers as specialist exposition. Revise autonomously for at most four rounds; preserve the separate goal of a self-contained package.” Paths come from the caller.
Behavioral fixtures are portable inputs for bounded agent dry runs. For the final skill hash, give a separate agent each input, request its decision and next actions under this skill, and save the prompt, response and independent comparison with the expected observable outcome. The inputs are fictional tests, not real manuscript states. Distinguish executed cases from supplied-only cases. Hash/accounting checks certify only those mechanics; semantic trials do not certify general mathematical judgment. After any promoted change, rerun affected cases on the final tuple.
Negative results travel with the positive ones (LX11, LX12)
This skill renders previous work into a new artifact, so the negative-result
floor of ../../POLICY.md applies. Before finishing, enumerate the inputs'
refutations, counterexamples, ill-posed proposals, and failed expectations, and
carry each into the output as a numbered statement (or a labelled question when
unproved), naming the expectation it corrects: "it is natural to expect $X$; in
fact $Y$". Report the count carried and, for anything deliberately left out,
the scope reason. Dropping a refuted claim together with its refutation is a
revise-level violation: the refutation is the surviving result.
04 同仓库更多 Skills
writing-skills
Use when creating new skills, editing existing skills, or verifying skills work before deployment
writing-plans
Use when you have a spec or requirements for a multi-step task, before touching code
verification-before-completion
Use when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always
using-superpowers
Use when starting any conversation - establishes how to find and use skills, requiring skill invocation before ANY response including clarifying questions
using-git-worktrees
Use when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback
test-driven-development
Use when implementing any feature or bugfix, before writing implementation code