Skip to content

yjj3019/proposal-skills

v1.3.0MIT

Korean IT proposal workflows for RFP analysis, editable slide decks, evidence review, and submission checks.

Changelog

Versions describe the local candidate in plugin.json; a version here does not imply a GitHub release or an official-directory publication (see references/marketplace-management.md).

Unreleased

Added

  • references/editorial-deck-design.md (nine-step type scale, one content grid, decluttering, seven at-a-glance patterns, render loop and the DRM caveat) and scripts/tidy_audit.py (font sizes, near-miss alignments, thin rules, characters per slide).
  • references/ai-vendor-comparison-playbook.md: how to compare AI vendors from their own documentation, a Korea data-location checklist and a dated 2026-11-12 snapshot; assets/examples/ai-model-comparison/make_deck.py builds the 17-slide example.
  • mapping_check.py (requirements matrix vs REQ-IDs in the document, both directions) and package_inspect.py (PPTX/DOCX package hints: author metadata, notes, comments, hidden slides, external links, embeddings, macros); a DRM-wrapped file is NOT INSPECTED (exit 2).
  • Optional Dockerfile and runtime_check.py (python-pptx, Pillow, LibreOffice, Poppler) that follow requirements.txt.
  • writing-style.md section 7 (translationese and mechanical phrasing) and a matching review-checklist item.

Fixed (found by an independent review of the review loop)

  • review_ledger.py: slide hashes now reach the theme (slide, layout, master, theme) and include the slide size and the table styles, so a theme colour or font change or a resize resets every verdict; before, an earlier verdict survived a visible change.
  • render_delta.py: more than 12 changed slides split the sheet, and the output and delta.json (changed_sheets, sheets) now name every file.
  • render_delta.py deletes only what it created: the output folder must be new, empty or its own (render_state.json), otherwise it is refused before anything is written; it removes only the sheets and page images it recorded, never a folder and never another file that matches a name (before, files named like sheet*.png and the whole pages folder could be deleted).
  • render_delta.py records the hash of each page image it writes and refuses (exit 2, nothing changed) to overwrite or delete a recorded page the user has since edited or replaced.
  • render_delta.py also records the hash of each sheet it writes and refuses (nothing changed) to delete a recorded sheet that was edited, a state written before hashes existed (no recorded hash), or a symbolic link as the output or pages folder.
  • review_ledger.py set --viewed sheet now needs --sheet <file> and records that overview's hash (checked by check); the page image is no longer passed off as the overview. Viewing again a slide whose old verdict has no image hash clears the four rungs.
  • render_delta.py also compares each rendered image's own hash and the render environment (LibreOffice build read from the install folder, the PDF converter): a slide whose content is unchanged but whose picture differs after a font or engine change is reported in render_changed, and a changed environment re-checks every slide. An unknown part of the environment is not treated as a change.
  • review_ledger.py ties a verdict to the image that was looked at (its SHA-256 is recorded when viewing is recorded): check refuses an image that changed since, and sync does not carry a verdict onto a slide whose picture differs.
  • render_delta.py no longer deletes a page image by its name: it renders in a temporary folder, installs only after checking, removes only the page images and sheets recorded as its own, and refuses (exit 2, nothing changed) when a file it is about to write exists and is not recorded (a page-99.png or sheet.png the user added); the PDF is no longer written into the output folder.
  • review_ledger.py sync carries a verdict onto a rebuilt slide only when it can prove the picture is the same (the old and the new image hash are both recorded and equal); a missing hash (an older ledger) resets the slide. set --image with another image, or viewing again after the image changed, clears the viewing and the four rungs. sheet viewing (cover, closing) is held to the same image proof as full. Open Critical or Major findings of a slide that no longer exists are kept in orphaned_findings and block check.

Added (cheap review loop)

  • scripts/render_delta.py: renders every page, compares each slide's content hash with the last run and builds a numbered contact sheet of only the changed slides (changed_sheet.png, delta.json); exit 2 without LibreOffice or a PDF-to-PNG converter.
  • review_ledger.py: per-slide content hashes and sync (keeps the verdicts of unchanged slides, resets the changed or new ones, never carries a verdict onto a slide whose hash differs); check refuses a ledger whose slide hashes no longer match the deck.
  • contact_sheet.py: tile labels can carry real slide numbers; the sheet is no wider than its tiles.
  • deck-intake-and-self-review.md: the cheap iteration loop, and the conditions an HTML preview would have to meet to be allowed (one source, identical geometry, a read-back parity check, a text-fit margin, the final PPTX still reviewed).

Added (editorial theme and review ledger)

  • build_deck.py: meta.theme: "editorial" (opt-in; default standard unchanged): paper background, serif title, light lead, no title rule, hairline tables, flat cards with a top rule; role names kept, so deck_check.py and the contrast check still apply. meta.title_font sets the title face; an unknown theme is rejected.
  • scripts/review_ledger.py (init, set, check): records per page that it was viewed, judged on the four rungs and closed out; refuses completion for an unseen page, a chart seen only on a sheet, an image older than the deck, an open Critical or Major finding, no revision, or a deck that changed after the review. It checks the review's honesty, not the quality.

Fixed (found while adding the theme)

  • build_deck.py: every run now carries the East Asian and complex-script typeface (a:ea, a:cs) next to the Latin one; before, Korean text fell back to the theme font and looked different on each machine.

Added (self-review without user feedback)

  • references/deck-intake-and-self-review.md: up to five questions with defaults before building, one direction checkpoint, a four-rung self-review (plain, busy, not at a glance, unfinished) with a report format, and the rule to write assumptions down when nobody can answer.
  • scripts/tidy_audit.py also measures body fill per slide and the share of text-only content slides; scripts/contact_sheet.py tiles rendered pages into one view (exit 2 for an unreadable or DRM-wrapped image).
  • AGENTS.md: "Look before you deliver, ask before you guess".

Fixed (third simulation round, closed later)

  • hook_guard_bash.py: also denies a download piped into python/node/perl/ruby/php with no script (curl ... | python -m json.tool stays allowed), PowerShell iwr ... | iex and -EncodedCommand, git filter-branch, git reflog expire, git gc --prune=now, find -exec rm without a filter or from a broad start, find -exec shred, tee into a system folder and rm -rf $PWD.

Added (simulation catalog)

  • references/harness-simulation-catalog.md (how to run a round, what three rounds found, documented limits) and a testing-pitfalls section in AGENTS.md (hygiene tests read tracked files only, Windows 8.3 temp paths, exit codes behind pipes, patch-script discipline).

Added (slide types)

  • build_deck.py: three at-a-glance slide types inside the standard frame: scoreboard (head-to-head with a winner track and a required verdict rule), triage (which option first, per situation) and weights (one stacked bar that must add to 100).

Fixed (found while building the AI model comparison deck)

  • sensitive_scan.py: an Office package embedded one level deep (a native chart's value workbook) is now read, so a deck with charts is no longer INCOMPLETE for that reason; nested parts are named outer.xlsx!inner/part.xml. Images, OLE objects, an embedded file that is not a ZIP and nesting deeper than one level are still reported UNSCANNED.
  • build_deck.py: the generated file no longer carries python-pptx's template author or its "generated using python-pptx" comment; the author is empty unless meta.author is set and the title comes from meta.title.
  • package_inspect.py: hidden slides are read from the slide part (PowerPoint stores show="0" there, not in presentation.xml); chart value workbooks are reported as information, not as an attachment; tracked changes in a DOCX block.
  • mapping_check.py: REQ-IDs that PowerPoint or Word split across text runs (REQ- and 001) are found, five- and six-digit IDs are no longer dropped, and DOCX headers, footers and footnotes are read.

Fixed (found by a 10-scenario simulation of the skills and harness)

  • hook_guard_bash.py: also denies commands wrapped in bash -c/sh -c/eval, bash <(curl ...) and eval $(curl ...), git checkout ., git stash drop|clear, git branch -D, git push --delete/:branch, unfiltered find -delete, rm -rf ./ and .*, dd of=/dev/..., mkfs, shred, xargs rm -r, recursive chmod/chown/mv on a broad target, redirects into system folders, a fork bomb, format C: and PowerShell Remove-Item -Recurse on a broad target.
  • task_contract.py: write_boundary must name one specific folder (not ., .., a drive or filesystem root, ~, *).
  • plain_language_check.py: a DRM-wrapped or non-ZIP Office file is reported as NOT INSPECTED (exit 2), not a traceback.
  • build_deck.py: a slides.json whose top level is not an object, or a damaged python-pptx install, exits 2 with a message.
  • install_skill.py: test files are no longer copied into the installed skill folder.
  • evidence_chain.py verify: --expect-head / --expect-count compare the log with a value written down elsewhere, so a truncated or rebuilt log is caught.

Fixed (second simulation round)

  • hook_guard_bash.py: judges the command inside wrappers (nice, timeout, exec, busybox, sudo -u, env -i) and groupings ( )/{ }; treats // and /tmp/.. as the root; also denies git push --mirror, kill -9 -1, and uploading a local file with curl/wget (-d @file, -F x=@file, -T).
  • skill_audit.py: folds fullwidth letters and invisible format characters before matching, and recognises Korean "ignore the previous instructions" and "print the system prompt" phrasings.
  • unified_gate.py --bundle: a file in the submission folder that attachments[] does not list (and that is not the checked document or audit) is reported as a warning, and an unlisted executable or script (.exe, .bat, .ps1, .js ...) blocks.
  • build_deck.py: when python-pptx's default template has been encrypted by an enterprise DRM agent (header <##), the error says so and suggests a virtual environment outside the DRM-managed paths.

1.3.0

Added

  • Repository tools: gate_mutation_probe.py (weaken a gate, require a test to fail), skill_audit.py (injection patterns and license declarations), merge_guard.py (advisory PR readiness), hook_guard_bash.py (opt-in destructive-command hook).
  • sensitive_scan.py: credentials and personal data (including resident-registration numbers with date and checksum notes) in PPTX/DOCX/XLSX/HWPX and text; unreadable files are reported as UNSCANNED, and an incomplete scan exits 2.
  • HWPX (Hancom, ZIP of XML) reading in sensitive_scan.py, quality_gate.py, check_numbers.py and plain_language_check.py: paragraphs, table cells, headers/footers and metadata; embedded BinData/ is reported as not inspected. The older binary .hwp is still reported as UNSCANNED (save it as HWPX first).
  • Proposal checks: edit_preservation.py, deck_story_check.py, plain_language_check.py, deck_stage_gate.py, verification_report.py, retry_guard.py, evidence_chain.py, handoff_lint.py.
  • cross_format_numbers.py (one numbers ledger over every delivered format), xlsx_check.py (error cells, #REF!, never-calculated formulas, --recalc through a throwaway LibreOffice profile that forces recalculation), task_contract.py (limits and states for a long job; Verified needs evidence), eval_metrics.py (paired baseline/candidate metrics from recorded runs), skill_audit.py --inventory/--deny (what each skill's scripts can do).
  • evidence_chain.py: rerun lists the checks whose artifacts changed; optional --seconds and --cost are totaled.
  • References: AI-tell review, visual critique (mandatory defects, non-numeric trend), plain-writing guide (question coverage, strongest objection), verdict vocabulary (PASS / FAIL / BLOCKED / NOT VERIFIED); operating principles in AGENTS.md.
  • THIRD-PARTY-NOTICES.md and license-hygiene tests; .gitattributes pins line endings to LF.

Changed

  • skill_audit.py: a negation word exempts a match only when it precedes it in the same clause.
  • sensitive_scan.py: UTF-16/32 XML and DOCTYPE/ENTITY anywhere in a part are refused; an Office-named file that is not a ZIP (for example one wrapped by an enterprise DRM agent) is reported as encrypted or DRM-protected, not silently skipped.

1.2.5

Previous local candidate. See git history.