feat: add a recall skill to the Claude Code plugin

[doronp]

What and why

gitmemory has been a Claude Code plugin since 0.1.0 (gitmemory@gitmemory), but only as the hook shim, which triggers a capture at the compaction boundary. It gave the agent no way to get the captured words back inside a session. You had to paste recall output yourself, or add a line to CLAUDE.md.

This PR adds one skill to the existing plugin: /gitmemory:recall. Claude loads it on its own when you refer to something from before a compaction or an earlier session ("what did we decide about X", "I told you not to…"). You can also run it with a topic: /gitmemory:recall retry loop. The skill:

• runs gitmemory recall "<topic>" and quotes each hit verbatim with its full address (agent/session/gNN@offset). It says so when a hit comes from another agent, session or project;
• runs gitmemory index when the index is missing or older than what's being asked. That derived index is the only thing it writes;
• skips its own echoes: earlier recalls and earlier answers that were captured too;
• when nothing is set up, tells the user the setup steps and stops. It never captures, configures, pushes or opens the dashboard.

Its component comes from the default skills/ directory, so neither manifest needs a path.

Commits

1. feat: add a recall skill to the Claude Code plugin: skills/recall/SKILL.md, plus the plugin and marketplace descriptions.
2. docs: say the plugin ships the recall skill, not only the hook: README, hook/README.md (which said "installs the hook and nothing else"), docs/USAGE.md and CHANGELOG. SECURITY.md gets two additions: the skill joins the in-scope list, and a paragraph says what it reads goes to the model provider past no redaction gate, since the gate stands at push.
3. tests: hold the recall skill to the CLI it tells the agent to run: one drift test in tests/test_docs.py.
   • Every gitmemory … command in the skill must answer --help, including the ones it forbids, and every flag it names must appear in that help.
   • The four CLI messages the skill branches on are produced by running the CLI on a missing store and an empty one: no index / exit 2, not a store, 0 generation(s), and no matches / exit 0.
   • It also pins MAX_TERMS, the skill's name, and /gitmemory:recall in the three docs that mention it.
   • The suite grows by one, so the pinned counts move: 1,027 on a fresh checkout, 1,352 with the corpora.

Heads-up: existing installs won't see it until a version bump

plugin.json stays at 0.1.0, because test_hook.py pins it to the pyproject version. Claude Code keeps a cached copy until the version changes. So claude plugin update on a 0.1.0 install reports it's current, and the skill arrives with the next release. Until then, uninstall and reinstall. hook/README.md and CHANGELOG both say this.

Checklist

• uv run pytest -q and uv run ruff check . pass locally: 1010 passed, 17 skipped, 1 deselected (the opt-in corpus test); ruff clean
• Each behavioural change has a test that fails without it. No fix to runtime code, so there is no tests/mutate_index.py row. The drift test was run against mutations of the skill by hand, and each one failed it; the list is in the commit body
• Pinned numbers that moved are updated (README board, REPRODUCE.md)
• No real transcript, credential, or absolute home-directory path in the diff
• Every commit is signed off (DCO)

Also: claude plugin validate --strict . passes, and a throwaway-CLAUDE_CONFIG_DIR install lists the skill.

Review (Gemini 3.1 Pro + Claude)

• Agreed and fixed: the skill now matches the CLI after feat: let recall choose which of two disagreeing hits comes first #15. Lines are dated, the order is user-first and newest first, and when two lines disagree the skill says the later user word is the current decision, quoting both.
• Rebased onto feat: let recall choose which of two disagreeing hits comes first #15. The README test count is 1,037.

🤖 Generated with Claude Code