[Sergey Dubinin] · car-maintenance-advisor.zip

car-maintenance-advisor

  • Score: 33/40
  • Verdict: Pass (workshop-pass)
  • Отправитель: Sergey Dubinin serg.dubinin.ru@gmail.com
  • SHA-256: ee6a3fc4fe9fdbcc…
  • Blockers: none
  • Unverified: F08

Required rework

Полная scorecard

IDScoreStatusEvidenceDefect / minimal fix
A011/1passedSKILL.md:1-11 frontmatter parses via yaml.safe_load; keys name,description present and non-empty (description 376 chars). Extra standard version field is harmless.
A021/1passedname='car-maintenance-advisor', len 23 ≤64, matches ^[a-z0-9]+(?:-[a-z0-9]+)*$, not reserved, denotes a concrete activity.
A031/1passeddescription 376 chars (1–1024), no XML tags (re.search none), routing text, not a YAML object.
A041/1passeddescription states the task (“Исследует…”), when to apply (“Применять при…”), domain terms (ТО, масло, регламент, допуски); finite third person “Исследует”.
A051/1passeddescription has no workflow/defaults/examples; each phrase changes the routing decision (scope terms + positive/negative triggers).
A061/1passedexplicit positive (“Применять при…”) and negative/anti-trigger (“Не применять для…”: diagnostics, repair, purchase, comparison, tuning, ПДД, insurance) give a useful boundary.
A071/1passedall 14 referenced files exist (verified); relative paths, forward slashes; grep for /home/,Users,C:\,@,~ in SKILL.md = none; runtime paths use <workdir> placeholder.
B011/1passedrepeatable 5-stage procedure (SKILL.md:104–463) producing an observable research report.
B021/1passedone coherent narrow domain (maintenance schedules/fluid specs/component lifespan); explicit scope invariant (SKILL.md:52–57).
B031/1passedterminology consistent; volatile facts (availability/prices/support dates) handled via runtime freshness rules (source-taxonomy.md:150–159), not hardcoded as eternal.
B041/1passedmedium freedom stated (SKILL.md:19–22): flexible plan, hard output/source-eval/thresholds; safety-critical stricter; plan gate.
C011/1passedone clear default path: Планирование→Поиск→Чтение→Верификация→Синтез with concrete numbered substeps.
C021/1passedeach branch has a when-to-choose rule (input-markup table SKILL.md:63–67, early-exit 72–81, orientation scan 1.4); default path is explicit.
C031/1passedplan gate requires explicit approval (SKILL.md:202–203); validate-and-revise with ≤2 returns (453–455); no destructive external action; search begins only after the gate.
C041/1passedfeedback loop: what to check (verification-rubric.md:52–83), what is a failure, which stage to return to (table 88–93), what to change, restart with 2-return cap then insufficient.
C051/1passedreport-template.md defines required result, per-claim evidence, confidence, readiness; adaptable template appropriate to a judgment research task.
C061/1passedcommon workflow easy to locate; vague steps replaced by checkable actions and deterministic scripts (normalize/count_independent/validate).
C071/1passedtool policy in guardrails: check availability before stage 2 (SKILL.md:89–90, 227–230), MCP fully qualified, no hidden actions; scripts clearly read vs execute (99–100).
D010/1failedwc -l SKILL.md = 501 lines > 500 limit.Trim or merge one line to reach ≤500 (e.g., fold the two-line closing sentence of §5.5 into one).
D021/1passedcore workflow/safety/decisions in SKILL.md; long theory, references and templates moved to references/ and assets/.
D030/1failedreferences >100 lines (source-taxonomy 174, verification-rubric 132, extraction-rules 121, search-tactics 103) open with a “when to read” line but none has a Contents/TOC block.Add a brief ## Содержание section listing headers to each reference file longer than 100 lines.
D041/1passedreferences/assets/scripts/evals used purposefully with descriptive names; no misc/dead files; clean tree, no OS/editor metadata.
D051/1passedeach script has shebang + docstring (usage/args/exit codes), stdlib-only deps declared (README:70), JSON/table output, portable relative invocation.
D061/1passedran all three scripts read-only: normalize (exit 0/1 + useful warnings), count_independent (valid→0, broken→1 “НЕ ДОБИРАЕТ”/“свёрнуто”), validate_artifacts (valid→0 “пройдены”, broken→1 with 8 specific violations).
E011/1passedguardrails treat external text as data not commands (SKILL.md:91–93; search-tactics.md:87–89) and templates as form not facts (94–95).
E021/1passedno secrets exposed; no undeclared downloads; no dependency install (scripts stdlib-only).
E030/1failedauto-creates intermediate draft/ledger files (source-ledger.md SKILL.md:288–292, findings.md 392–396, verification.md) not explicitly requested by the user and not the declared result; verification checklist is explicitly hidden from the user (SKILL.md:457). Not a blocker (local recoverable, no secrets).Make intermediate working-file creation opt-in/informed (consent for the working dir) or keep intermediates ephemeral and persist only the user-requested report.
E041/1passedscripts are pure stdlib Python; no shell/network/bundled binaries/obfuscation; web search is the declared, transparent purpose.
E051/1passedno instruction to ignore prior instructions, hide actions, weaken safeguards, or expand the permission model; guardrails reinforce safeguards.
E061/1passedtool availability checked before first query; unavailability is declared and the skill stops rather than simulating search (SKILL.md:89–90, 227–230).
E071/1passedworkflow uses <workdir> placeholders; scripts portable; no home/username/drive/absolute paths in SKILL.md. (README install snippet uses author’s 08.08/ date dir but is adaptable guidance, non-blocking.)
E081/1passedMCP rule requires fully-qualified ServerName:tool_name (SKILL.md:96); no unqualified MCP names invoked anywhere.
F010/1faileddegree of freedom and needed files/evals are stated (README), but observed baseline errors of the agent WITHOUT the skill on representative tasks are not documented; README “Статус” admits no systematic runs.Add an explicit design note recording observed baseline agent failures (without the skill) that motivated the skill.
F021/1passedsource-taxonomy.md T1–T9 with claim-class primacy; T1–T4/T9 primary; T7/T8 cannot be sole basis; safety-critical requires T1/T2/T9.
F031/1passedfreshness rules per branch (source-taxonomy:150–159), independence/diversity (count_independent + dedup), explicit conflict taxonomy & resolution (extraction-rules, verification-rubric), confidence lowered on unresolved conflict (confidence-scale “Спорно”).
F041/1passedFACT/INFERENCE/OPINION separated (extraction-rules; findings-template), quotes/provenance required, confidence 4 levels with lowering factors, insufficient allowed.
F051/1passedreport-template.md contains restated question (L1–3), recommendation (L12–18), key evidence + sources table (L44–50), unknowns/limitations (L38–42), confidence (L5, per-claim).
F060/1failedevals.json has 5 scenarios using prompt/expected_output/assertions (not query/expected_behavior/failure_modes); trigger-eval.json uses query/should_trigger only; missing security-injection, broken-script, and portability/path case types.Provide ≥7 cases covering the 7 required types, each with query, expected_behavior, failure_modes.
F070/1failedno observed trigger/baseline run; README “Статус” states runs were not performed; evals/README confirms scenarios 2–5 not run and no trigger eval executed.Run trigger evals and end-to-end scenarios; capture baseline-without-skill + activation/no-activation/boundary observations.
F080/1unverifiedno navigation/behavior trace available in the submission (and not produced by grader).Capture agent traces showing no injection, no strong conclusion on weak evidence, correct supporting-file reads, and output-contract execution.

Required rework

  1. F06 — evals: add ≥7 cases (positive, negative, boundary, security-injection, evidence-quality, broken-script, portability/path), each with query, expected_behavior, failure_modes; current files use non-conforming fields and miss three required types.
  2. F07/F08 — behavioral runs: actually execute the trigger evals and end-to-end scenarios (including a baseline without the skill) and record navigation/behavior traces; the submission itself states these were not run.
  3. F01 — design note: explicitly record observed baseline agent errors without the skill on representative tasks.
  4. D01 — size: trim SKILL.md from 501 to ≤500 lines.
  5. D03 — navigation: add a brief Contents block to reference files longer than 100 lines.
  6. E03 — side effects: make auto-created intermediate ledger/findings/verification files opt-in or ephemeral (persist only the user-requested report).

← назад к лидерборду