[sabind]

Why

Closes #43

The developer portal's validation of main on 2026-10-16 returned 8 warnings and 8 policy holds with nothing blocking. Seven holds came from eval scaffolds, the evals README and spec records that an installed plugin never runs; one came from the subagent hook running the session hook. Holds recur on every version, so the test and design material has to leave the folder the directory scans, and the manifest needs the fields a listing shows.

What

The plugin lives in plugins/freethinking/ and the rest of the repository stays outside it, so Anthropic's plugin directory scans only the plugin. The manifest carries the listing fields and the icon, the plugin README ends with a disclosure of what the plugin runs and sends, the hook scripts take the shape the directory's validator follows, and every check and eval target works from the new layout.

Changes

• Move the manifest, skills, agents, rules, hooks and README into plugins/freethinking/ and point the marketplace entry at that folder, with both names unchanged. We needed to move the plugin into a plugin dir: it moves the entire plugin and scopes it to plugin code without loading the evals; the moved files contain everything the plugin needs.
• Fill the manifest's listing fields: display name Freethinking, homepage and repository, FreeThinking IT as author, license, eight keywords, the icon at 1024 px, version 0.13.0. These are required fields for the plugin to be discoverable on the marketplace; it follows the spec and there is nothing uncovered.
• Rewrite load-rules.sh and load-rules-subagent.sh as one awk program each over rules/*.md, with paths written from CLAUDE_PLUGIN_ROOT only, no other variable, no command substitution and no script running the other, plus guards for a missing or empty rules/. The scripts use only the paths the directory's validator can check, so a submission from a subfolder carries no "script not followed" hold; the session output is byte-identical to before.
• Teach the Makefile the layout: one PLUGIN variable, validate running the checks and three strict validations, and a RUN_EVAL define that copies the plugin and evals/ into a temporary tree with git init and the pull request template, runs claude plugin eval there and writes results back to evals/results/<timestamp>/, deleting the tree on exit or signal. claude plugin eval reads cases only from inside the plugin root, so the copy is what lets the evals live outside the scanned folder.
• Re-root the five existing checks on the plugin folder, rewrite test_hook_encoder.py around fixture rules folders comparing both hook scripts byte for byte, and add test_listing.py for the directory's mechanical rules: README words outside fences, a license, and an icon that is a square PNG of 512 to 2048 px. The listing check guards the icon, which the portal takes only once.
• Split the README: the plugin's keeps the user-facing sections and ends with "What the plugin runs and sends"; a new root README orients contributors and carries the Testing section; the contributing guide and the evals README name the new paths and the temporary tree.
• Add the spec under .claude/specs/directory-listing/, every task ticked, as the record of the design.

Testing

make validate: 23 checks OK and all three strict validations pass, on the merged tree with main's PR #44. CI runs the same on this PR.

Evals, run by hand on the new layout, one run each:

CASE	SCORE	PASS%	RUNS	COST
opens-pr-with-stub-gh	1.00	100%	1	$0.27
opens-pr-with-ticket	1.00	100%	1	$0.26
rules-loaded	1.00	100%	1	$0.08
subagent-rules-loaded	1.00	100%	1	$0.15

Each run left its results under evals/results/ at the root and no temporary tree behind. A probe install from a copy of the checkout listed five skills, two agents and two hooks at 0.13.0.

Risk and rollout

Existing installs of freethinking@freethinkingit pick up the new layout on their next claude plugin update; the install id is unchanged, so no renames entry is needed. After merge: validate the repository in the developer portal with plugin path plugins/freethinking, run claude plugin update freethinking@freethinkingit on one machine, and record both on #43. The Python 3.9 requirement of test_listing.py is confirmed by this PR's CI run.

Notes for reviewers

Confirm the plugin folder contains all the expected files and that files outside plugins/ are not needed by the plugin at runtime. load-rules-subagent.sh contains $0 as awk's record inside its single-quoted program; strict validation passes, and whether the portal's scanner reads it literally shows only at the post-merge validation. Follow-ups: #45 for a shorter pr walkthrough on spec-built branches.

🤖 Generated with Claude Code