editor-test-harvester

description	Mine external editor repositories for portable editor-behavior tests with ClawSweeper-style discipline: multi-pass exhaustive inventory, confidence scoring, framework-specific skip reasons, Slate/Plate coverage mapping, license-aware invariant extraction, and copy/refactor/create decisions.
argument-hint	[<repo-path-or-owner/repo> [--apply] [surface/tag/filter]]
disable-model-invocation	true
name	editor-test-harvester
metadata	{"skiller":{"source":".agents/rules/editor-test-harvester.mdc"}}

Editor Test Harvester

Use this skill when mining another editor repository for tests worth harvesting: Lexical, ProseMirror, CodeMirror, Tiptap, Monaco, Quill, or any local clone under ...

The job is not to clone their framework. The job is to extract portable editor behavior proof and route it to the right owner: raw Slate v2 substrate or Plate packages, kits, examples, and docs.

This follows the ClawSweeper method and the goal-plan confidence model: source-first, exhaustive inventory, explicit skip reasons, evidence rows, scored passes, narrow claims, license-aware invariant extraction, then implementation only when asked.

Source Of Truth

Read these first:

1. The target repo license evidence: LICENSE, LICENCE, COPYING, NOTICE, package metadata, or workspace metadata.
2. The target editor repo, preferably a local checkout like ../lexical.
3. .tmp/slate-v2 current test files and package scripts.
4. Plate package/docs/example owners in this checkout when the behavior is a plugin, kit, UI, React integration, or product policy.
5. docs/solutions/ for prior browser, IME, selection, and mobile proof lessons.
6. .agents/skills/clawsweeper/SKILL.md for issue/PR provenance discipline when a test exists because of a known upstream bug.

Never browse GitHub files. If the target is owner/repo and the checkout is missing, clone it to ../repo-name.

License Gate And Output Mode

Before harvesting behavior, classify the target repo license from local source files only: LICENSE, LICENCE, COPYING, NOTICE, package metadata, or workspace metadata.

Use one of these modes:

• permissive: MIT, BSD-2-Clause, BSD-3-Clause, Apache-2.0, ISC, CC0, or clearly equivalent permissive licensing.
• behavior-only: GPL, LGPL, AGPL, MPL, EPL, CDDL, proprietary, commercial, source-available, mixed/unclear licensing, missing license evidence, or any repo whose tests are not clearly permissively reusable.

Output location:

• permissive: write harvest artifacts under docs/editor-test-harvester/<repo>/.
• behavior-only: write harvest artifacts under .tmp/editor-test-harvester/<repo>/.

.tmp/editor-test-harvester/<repo>/ is unversioned scratch space. It may contain source-adjacent notes, detailed provenance, raw local observations, and working material needed to understand the source repo.

The strict invariant-only rule applies when producing versioned or durable project output from a behavior-only source:

• Do not copy upstream test code, fixtures, helpers, snapshots, expected output blobs, or expressive test prose into Plate, Slate, docs, examples, commits, or PRs.
• Do not translate upstream code line-by-line into versioned local tests.
• Do not preserve upstream fixture shape in versioned output unless the shape is a generic factual editor state required to describe the behavior.
• Use source paths, minimal line references, and short paraphrased test labels as provenance in versioned output.
• Extract the behavior invariant, then write a fresh local proof using local Slate/Plate fixtures, names, helpers, and assertions.
• When in doubt about license mode, treat the source as behavior-only.

Rule of thumb: study the wound, don’t transplant the skin.

behavior-only does not block learning from non-permissive, source-available, or unclear-license projects. It blocks copying into versioned Plate/Slate output. The local proof should show that the implementation was independently expressed from the invariant, not pasted or mechanically ported from upstream source.

Core Rules

• License-gate the target before choosing the report directory.
• For permissive targets, write normal durable artifacts under docs/editor-test-harvester/<repo>/.
• For behavior-only targets, write unversioned scratch artifacts under .tmp/editor-test-harvester/<repo>/.
• For behavior-only targets, copying source material inside .tmp is allowed as scratch provenance, but copying it into versioned Plate, Slate, docs, examples, commits, or PRs is forbidden.
• For behavior-only targets, every versioned behavior row, implementation, docs note, and test must be rewritten as a fresh invariant in local language, with only source path/line provenance.
• Never claim an upstream test was copied from a behavior-only repo into versioned output. The valid versioned actions are invariant extraction, fresh local proof, defer, or skip.
• Build a complete target test-file inventory before reading favorite files.
• Store harvest artifacts in one stable repo folder chosen by license mode: docs/editor-test-harvester/<repo>/ for permissive targets, or .tmp/editor-test-harvester/<repo>/ for behavior-only targets.
• For behavior-only targets, inventory.md and test-index.md also stay under .tmp/editor-test-harvester/<repo>/. Never create or refresh docs/editor-test-harvester/<repo>/inventory.md.
• Do not create dated report files under docs/plans/ for normal harvest output.
• Reruns are idempotent. Read the existing repo folder first, rerun the source inventory, update the same files in place, and record new/removed source rows instead of creating a duplicate report.
• Every target test file ends in exactly one category: portable, portable-mixed, plate-owned, skip, harness, product-shell, or uncertain.
• Every skip gets a reason. "Looks internal" is too vague.
• Preserve the behavior invariant, not the upstream API shape.
• Do not paste upstream test code verbatim into versioned output. For behavior-only targets, do not paste, mechanically translate, or preserve upstream fixture/helper shape in versioned output. Write the strongest fresh local proof for the invariant, with source path references only.
• Prefer DRY Slate coverage: if a related test exists, strengthen or split it instead of creating a duplicate.
• Use plate-owned when the useful behavior should fit Plate, not raw Slate. Examples: link/autolink grammar, list/checklist policy, markdown transformer UX, mention/hashtag/emoji/date-time plugins, media/product decorators, toolbar/menu state, React plugin hosts, NodeView/PluginView-style authoring, and rich product examples.

Goal And Report State

Use an agent-native goal for comprehensive harvests, long-running reruns, and apply runs. Always call get_goal first. Call create_goal only when no active matching goal exists. There can be only one active goal per thread.

Create the harvest goal plan from the project template:

node .agents/skills/autogoal/scripts/create-goal-scratchpad.mjs \
  --template editor-test-harvester \
  --title "<Repo> Editor Test Harvest"

The generated docs/plans file is the durable pass-state ledger, scratch log, decision record, verification ledger, and completion checklist. Do not create hook state.

The harvest report itself still lives in the license-selected report directory:

• permissive: docs/editor-test-harvester/<repo>/report.md
• behavior-only: .tmp/editor-test-harvester/<repo>/report.md

Record this state in the active goal plan:

target_repo: <path>
repo_key: <repo>
license_mode: permissive|behavior-only
output_mode: durable|scratch
report_path: <docs/.../report.md or .tmp/.../report.md>
inventory_path: <report_dir>/inventory.md
test_index_path: <report_dir>/test-index.md
current_pass: <pass-name>
current_pass_status: in_progress
current_pass_skill: .agents/skills/editor-test-harvester/SKILL.md
next_pass: <next-pass>

Set done only when:

• total score is >= 0.92;
• no dimension is below 0.85;
• inventory count equals classified count;
• every runnable test file is either test-name indexed, read, or explicitly skipped as non-behavior with a reason;
• every portable and portable-mixed source has at least one behavior row or a recorded reason why its rows were rejected;
• every covered, refactor-existing, create-new, copy-now, defer, and plate-owned action has a raw Slate owner, Plate owner, or explicit gap;
• every create/copy/refactor/plate-owned row has an implementation target and verification command or a documented Plate backlog owner;
• the report has a full inventory appendix or a linked inventory file.

If any gate fails but more work is possible, keep pending and write the active goal plan with the next harvest pass. Use blocked only when the target repo, Slate v2 checkout, required browser/device tooling, or a user decision is missing and no useful autonomous pass remains.

Confidence Score

Score each harvest pass from 0.00 to 1.00.

Dimension	Weight
Inventory completeness	0.20
Behavior extraction depth	0.20
Skip precision and negative controls	0.15
Slate/Plate coverage mapping accuracy	0.20
Actionability of copy/refactor/create plan	0.15
Provenance and reproducibility	0.10

Score caps:

• Inventory completeness cannot exceed 0.75 unless the report records the exact inventory command, total count, classified count, and unresolved count.
• Inventory completeness cannot exceed 0.90 unless every test file path appears in a full inventory appendix or linked inventory file.
• Behavior extraction depth cannot exceed 0.75 if portable files were routed only by file name.
• Behavior extraction depth cannot exceed 0.85 unless all portable and portable-mixed runnable files have test-name extraction with line pointers.
• Behavior extraction depth cannot exceed 0.85 for behavior-only targets unless the report separates scratch provenance from versioned-output invariants.
• Skip precision cannot exceed 0.80 unless every skipped file or skip family has a concrete reason and at least one negative-control example was read.
• Slate/Plate coverage mapping cannot exceed 0.80 unless the report records the .tmp/slate-v2 search commands for raw rows and the current Plate owner or explicit Plate gap for plate-owned rows.
• Actionability cannot exceed 0.80 unless every create/refactor/copy row names a target test file, proof kind, and focused verification command, and every plate-owned row names the likely Plate package, kit, example, docs, or backlog owner.
• Provenance cannot exceed 0.80 when issue/PR-linked upstream regression tests are used without ClawSweeper-style exact thread or local source rationale.
• Provenance cannot exceed 0.80 for behavior-only targets unless the report records license evidence, output directory, and versioned-output copy policy.
• The total score cannot exceed 0.88 if the report has no pass-state ledger.
• The total score cannot exceed 0.90 if the full inventory appendix is absent.

Completion threshold:

• total score >= 0.92;
• no dimension below 0.85;
• no uncertain test files remain;
• no portable-mixed file remains unexamined;
• no create/refactor/copy/plate-owned row lacks a target owner;
• no browser/IME/mobile claim is based only on jsdom or synthetic model tests;
• pass-state ledger proves every pass completed before closure;
• for behavior-only targets, no versioned report, test, docs, fixture, snapshot, or implementation output contains copied or mechanically translated upstream source material.

Below threshold, the report can still be useful, but it must say first-pass, partial, or pending and name the next owner.

Pass Schedule

Run harvests as passes, not one giant skim:

1. Intake and boundary pass:
   • target repo path or clone target;
   • license evidence and license mode;
   • report directory selected from license mode;
   • surfaces in scope and explicit non-goals;
   • report-only vs apply;
   • browser/device availability.
2. Inventory pass:
   • find package manager and test harnesses;
   • build the full test inventory;
   • split actual runnable tests from fixtures/support files;
   • write every inventory row to the appendix with an initial category.
3. Test-name extraction pass:
   • extract describe / it / test names for every runnable portable, portable-mixed, and uncertain file;
   • for huge files, index test names first, then read targeted ranges;
   • no portable-mixed file may stay summarized only by file name;
   • for behavior-only targets, record raw/source-adjacent extraction only in .tmp; convert anything that leaves .tmp into short paraphrased provenance.
4. Classification pressure pass:
   • challenge skips and product-shell routing;
   • read negative-control examples from each large skip family;
   • route every file to portable, portable-mixed, plate-owned, skip, harness, product-shell, or defer.
5. Behavior extraction pass:
   • turn source tests into behavior invariants;
   • attach source file, line, test name, tag, proof kind, browser/device requirements, and issue/PR rationale when present;
   • for behavior-only targets, scratch extraction may live in .tmp, but any versioned row or implementation must be rewritten as a fresh local invariant.
6. Slate/Plate coverage mapping pass:
   • search .tmp/slate-v2 by behavior keywords and adjacent concepts;
   • search current Plate packages, docs, examples, and kits for plugin/product behavior;
   • map to exact current Slate tests, browser stress rows, Plate owners, docs, or explicit gaps;
   • do not use old plans as current coverage proof.
7. Action planning pass:
   • assign covered, refactor-existing, create-new, copy-now, defer, plate-owned, or skip;

Pass-state ledger rows must include:

• pass name;
• status: pending, in_progress, or complete;
• evidence added;
• report delta;
• open issues;
• next owner.

Discovery Workflow

1. Resolve the target repo:

   test -d ../lexical || git clone https://github.com/facebook/lexical.git ../lexical
   

2. Classify license mode before creating report artifacts:

   target="../lexical"
   
   mapfile -t license_files < <(
     {
       find "$target" -maxdepth 2 -type f \
         \( -iname 'LICENSE*' -o -iname 'LICENCE*' -o -iname 'COPYING*' -o -iname 'NOTICE*' \)
       find "$target" -maxdepth 3 -type f \
         \( -name 'package.json' -o -name 'pnpm-workspace.yaml' -o -name 'package.yaml' \)
     } | sort -u
   )
   
   mkdir -p .tmp/editor-test-harvest-license
   printf '%s\n' "${license_files[@]}" > .tmp/editor-test-harvest-license/files.txt
   
   if [ "${#license_files[@]}" -gt 0 ] && \
      rg -i 'MIT License|Apache License|BSD 2-Clause|BSD 3-Clause|ISC License|CC0|SPDX-License-Identifier:\s*(MIT|Apache-2.0|BSD-2-Clause|BSD-3-Clause|ISC|CC0-1.0)|"license"\s*:\s*"(MIT|Apache-2.0|BSD-2-Clause|BSD-3-Clause|ISC|CC0-1.0)"' \
        "${license_files[@]}" >/dev/null && \
      ! rg -i 'GPL|AGPL|LGPL|MPL|EPL|CDDL|commercial|proprietary|source-available|Business Source|Elastic License|Server Side Public License|SSPL' \
        "${license_files[@]}" >/dev/null; then
     license_mode="permissive"
   else
     license_mode="behavior-only"
   fi
   
   echo 
    .tmp/editor-test-harvest-license/files.txt
   

• covered: current Slate v2 or Plate test fully proves the invariant.
• refactor-existing: related coverage exists but needs split, rename, stronger assertions, or browser proof.
• create-new: no adequate coverage; propose the target test location.
• copy-now: user asked to apply, the invariant is safe to implement now, and the target is permissive. For behavior-only targets, this means "write a fresh local proof from the invariant" for versioned output, never copy upstream code, fixtures, snapshots, helpers, or expressive prose into Plate, Slate, docs, examples, commits, or PRs.
• defer: requires unavailable raw device/browser/tooling proof.
• plate-owned: route to Plate package, kit, example, docs, or backlog owner.
• skip: after reading, it is not portable enough.

Portable Test Taxonomy

Use these tags consistently:

• ime-composition
• beforeinput-input
• composition-selection-repair
• selection-dom-mapping
• clipboard-paste
• drag-drop
• history-undo-redo
• normalization-schema
• delete-backspace
• insert-fragment
• marks-inline
• void-atom
• tables-grid
• decorations-overlays
• collaboration-remote
• performance-large-doc
• accessibility-keyboard
• serialization-parsing
• browser-engine
• mobile-device
• focus-blur
• shadow-dom
• pagination-layout
• structured-blocks
• markdown-richtext-roundtrip

Portable examples:

• IME composition through marked text, decorations, void boundaries, and undo.
• Selection mapping around zero-width text, voids, inline boundaries, and tables.
• Browser-specific selection behavior such as Firefox multi-range selections.
• Native paste, drag, drop, beforeinput delete, and compositionend ordering.
• Undo granularity across composition, paste, delete, and fragment insertions.
• Fragment insertion through mixed inline/block/void boundaries.
• Collaborative position mapping across concurrent insert/delete and undo.
• Pagination and deterministic layout behavior when the source proves editor semantics rather than renderer-only appearance.

Skip examples:

• React/Vue/Svelte integration internals.
• Lexical node-class lifecycle or ProseMirror plugin object mechanics.
• Command registry wiring with no observable editor behavior.
• Test utility tests that prove only the upstream harness.
• Product demo chrome, menus, shortcuts, or app-specific UX policy.
• Snapshot-only styling assertions with no editor behavior invariant.
• Versioned-output candidates from behavior-only repos that cannot be reduced to a fresh invariant without preserving expressive source material.

Plate-owned examples:

• Link/autolink grammar, link-splitting policy, and no-nested-link decisions.
• List/checklist indentation, ordered metadata, list-exit, and ARIA policy.
• Markdown transformer UX, MDX/custom transformers, and serializer choices.
• Mention, hashtag, keyword, emoji, date-time, equation, and media plugins.
• React plugin hosts, menus, typeahead, toolbar state, and example app behavior.
• ProseMirror NodeView, MarkView, PluginView, and plugin prop lifecycle ideas when they improve Plate authoring but should not become raw Slate API.
• Tiptap-style extension ergonomics, command ergonomics, and schema packaging when the behavior is developer-product surface rather than raw editor truth.

Output Shape

For report-only runs, choose the output directory by license mode:

# permissive
docs/editor-test-harvester/<repo>/report.md

# behavior-only
.tmp/editor-test-harvester/<repo>/report.md

behavior-only reports are unversioned scratch artifacts. They may contain detailed local research notes. They must not be copied into durable project docs without rewriting into fresh invariant language.

Use stable companion files in the same directory:

<report_dir>/inventory.md
<report_dir>/test-index.md

For behavior-only targets, <report_dir> means .tmp/editor-test-harvester/<repo>. Do not write inventory or test-index files to docs/editor-test-harvester/<repo>/.

Do not include the date in file names. Put run dates, inventory commands, source checkout path, source revision/provenance, license mode, and license evidence inside the report instead.

If a previous harvest exists in the chosen license-mode report directory, rewrite these files in place and add a rerun/update note that names newly discovered and removed target test files.

Use this shape:

# Editor Test Harvest: <repo>

status: pending|first-pass done|done
score: 0.xx
license_mode: permissive|behavior-only
license_evidence: `<path/to/LICENSE or package metadata>`
output_mode: durable|scratch
versioned_copy_policy: normal|fresh-invariant-only

## Inventory

- target: `<path>`
- test files found: N
- portable: N
- portable-mixed: N
- plate-owned: N
- skipped: N
- harness/product/uncertain: N

## License Gate

| Field                 | Value                              |
| --------------------- | ---------------------------------- |
| License mode          | `permissive` or `behavior-only`    |
| Evidence files        | `<local paths>`                    |
| Output directory      | `docs/...` or `.tmp/...`           |
| Output mode           | `durable` or `scratch`             |
| Versioned copy policy | `normal` or `fresh-invariant-only` |

For `behavior-only` targets, this report is scratch material. Anything promoted
from it into versioned Plate/Slate output must be rewritten as a fresh invariant
and local proof.

## Confidence Score

| Dimension | Score | Evidence | Cap hit |
| --------- | ----: | -------- | ------- |

## Pass-State Ledger

| Pass | Status | Evidence added | Report delta | Open issues | Next owner |
| ---- | ------ | -------------- | ------------ | ----------- | ---------- |

## Matrix

| Source ref       | Test ref   | Tag               | Behavior invariant | Proof kind              | Owner coverage                            | Action                                                 |
| ---------------- | ---------- | ----------------- | ------------------ | ----------------------- | ----------------------------------------- | ------------------------------------------------------ |
| `../lexical/...` | `test ref` | `ime-composition` | ...                | browser/unit/raw-device | `.tmp/slate-v2/...`, Plate owner, or none | covered/refactor-existing/create-new/defer/plate-owned |

For `behavior-only` targets, `Source ref` and `Test ref` are scratch provenance.
Any versioned output derived from `Behavior invariant` must be rewritten locally.

## Skips

| Source           | Reason                                                       |
| ---------------- | ------------------------------------------------------------ |
| `../lexical/...` | Framework node-class invariant, no portable editor behavior. |

## Next Slice

1. Refactor existing ...
2. Create new ...
3. Defer raw-device ...

## Full Inventory Appendix

| Source | Runnable | Category | Reason | Test-name extraction |
| ------ | -------- | -------- | ------ | -------------------- |

If the inventory appendix is too large, write it to:

<report_dir>/inventory.md

Then link it from the main report. Do not omit it.

Write the extracted portable/portable-mixed test-name index to:

<report_dir>/test-index.md

For apply runs, edit .tmp/slate-v2 or Plate tests only after the matrix exists. Keep the implementation slice small enough to verify.

Source comments are useful only when they explain the behavior being protected. Do not make tests read like upstream changelogs. For behavior-only targets, source comments in versioned output must not quote or closely paraphrase upstream test bodies, fixtures, snapshots, expected output blobs, or expressive prose.

Verification

Report-only verification:

# Set this from the license gate.
# permissive:
report_dir="docs/editor-test-harvester/<repo>"

# behavior-only:
report_dir=".tmp/editor-test-harvester/<repo>"

rg --files <target> | rg '<test inventory pattern>' | wc -l
rg -n "License Gate|Confidence Score|Pass-State Ledger|Matrix|Skips|Next Slice|Full Inventory Appendix" "$report_dir/report.md"
test -f "$report_dir/inventory.md"
test -f "$report_dir/test-index.md"
node .agents/skills/autogoal/scripts/check-complete.mjs docs/plans/<goal-plan>.md

Versioned-output hygiene check for behavior-only sources:

test ! -f "docs/editor-test-harvester/<repo>/inventory.md"
test ! -f "docs/editor-test-harvester/<repo>/test-index.md"
rg -n "copied from|verbatim|fixture copied|ported from" \
  docs .tmp/slate-v2 packages apps examples 2>/dev/null && \
  echo "Review versioned output for unsafe behavior-only wording" || true

Implementation verification in .tmp/slate-v2 or Plate:

• Run the focused package or browser test that owns the new proof.
• Run bun check before claiming the applied slice is closed.
• For release-quality browser claims, use bun check:full.
• Use raw mobile proof only when real device/Appium artifacts exist; semantic mobile handles and Playwright mobile viewports are not raw-device proof.
• For behavior-only sources, verify that versioned local tests use local fixtures, helpers, names, and assertions, not upstream test code or mechanically ported fixture shape.

ClawSweeper Use

Use ClawSweeper as the model for provenance work. If a test points at an upstream issue, PR, browser bug, or closed-thread rationale, apply the ClawSweeper bar: exact thread, exact behavior, no speculative closure claim, and no broad claim without current source proof.

When local issue archaeology is useful, use the gitcrawl command surface through that ClawSweeper discipline:

gitcrawl doctor --json
gitcrawl search issues "<behavior phrase>" -R <owner/repo> --state all --json number,title,state,url --limit 20
gitcrawl sync <owner/repo> --numbers <issue-or-pr-number> --include-comments --with pr-details --json

Treat gitcrawl as provenance and candidate generation only. The target repo test inventory still owns coverage accounting.

For behavior-only targets, issue/PR provenance may identify why a behavior matters, but it still must not justify copying upstream code, fixtures, snapshots, expected output blobs, or expressive test prose into versioned output. Extract the invariant. Write the local proof fresh.