Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for

[santoshkumarradha]

Skills installed for Claude Code and Codex now reach codeaf directly, with memory on or off, and the model picks the one a request fits by what the skill is for.

What changed

• Claude Code plugin skills are read, the way Claude Code decides what is live: the plugin is named in ~/.claude/plugins/installed_plugins.json, enabledPlugins says true in the layered settings (home, project, project local), and a project-scoped install counts only in its own project. A plugin reads the skill folders its manifest or its marketplace catalog entry names, and only those; a plugin that names none is read from skills/. The plugins folder is never walked, so marketplace clones and older cached versions are left alone. A plugin skill keeps its bare folder name, ranks below every hand-kept skill and above Codex's bundled ones; the decision is written up in internal/skills/plugins.go.
• Codex's bundled skills in .codex/skills/.system are read, below everything else in their scope.
• Linked skill folders are followed, which is how npx skills and similar installers reach every tool from one copy.
• Memory off no longer turns skills off. With memory.enabled off the process builds its shelf in a temporary store from the skill folders and removes it on close. Nothing is written to the memory store the person turned off. use_skill, the catalog and the per-message carry all work. The Dropping a folder shows Add but does not attach it #1380 "switched off" line is gone.
• The catalog lists every skill, by what it is for. It is in the stable prefix in a fixed order: scope match first, then by name. The recency ordering is gone so the order cannot move between requests. Descriptions are clipped to 160 characters, and the described lines stop at 12 KiB. The names that did not fit follow on one line of up to 2 KiB, and anything past that is counted. The model opens the skill a request fits with use_skill. The word-match carry still runs first on every message.
• The carried-skills line stays visible. On the chat surface the line reads skills · <names>. Once the answer landed, the ▸ worked chip swallowed it, and even an opened chip did not show it. It now stays under the message, with the chip starting below it.
• The /skill picker and the attachment work on the ordinary launch. The picker reads the conversation's own shelf. Attach, detach, list and clear cross the session-host connection (remote.Agent methods, Welcome.Skills, serveSkills), where they had been missing. skillAttacher leaves the surface-door ledger, and the ratchet is now 21 after the merge with the base's own removal.
• The attachment is a fact and not a call. session.Facts.Skills carries the attachment down unasked, and the engine announces it after every attach, detach or clear. The chip reads the replica, and the picker asks the shelf and the attachment doors off the update loop (the base's new off-loop laws).
• The shelf holds 400 skills (was 100).
• Manual: skills-from-other-tools.md (new), use-skill.md, putting-a-skill-in-front.md, skills-a-turn-used.md, what-i-remember.md; six new retrieval probes. Change entry docs/changes/unreleased/1396-skills-from-other-tools.md.

Catalog cost: each described line is - name: description, with the description clipped to 160 characters. That is about 120 to 230 bytes per skill, most often 150 to 180. It is paid once, in the cached prefix. The whole section is bounded at 12 KiB of described lines plus 2 KiB of names.

Evidence (remote test box; every go test with -count=1)

• Red first. Each new unit test was run against the unfixed implementation with the new tests in place. skills: a compile red, then six assertion reds. resident: two assertion reds. session, tui3, remote and cmd/codeaf: compile reds. The plugin manifest test failed on the earlier discovery code. The carried-line test failed on the earlier feed and fold code.
• End-to-end, the real binary in tmux, fresh home, both memory settings, on the head: TestTUIE2E/foreign_skills_reach_the_conversation PASS for memory_on and memory_off. Three skills are installed: one in .claude/skills, one in .codex/skills, and one in an installed and enabled plugin. It asserts:
  • (a) automatic: skills · tide-almanac with QUILLON-TIDE-7431, then skills · lantern-ledger with LANTERN-LEDGER-2958;
  • (b) by hand: /skill orchard lists the plugin skill with no refusal, Enter attaches it, and a question sharing no words carries skills · orchard-census and answers with ORCHARD-CENSUS-6612;
  • (c) the transcript records use_skill with "mode":"list" and "mode":"get".
• Real installs in a fresh home. Installed with npx skills add (one skill global for Claude Code, one project-level for Codex into .agents/skills), claude plugin marketplace add plus claude plugin install example-skills@… (Claude Code 2.1.278), claude plugin init (a skills-dir plugin), and a hand-written ~/.codex/skills skill. use_skill list returned all 15 skills, memory off and memory on. The 12 plugin skills are exactly the set claude plugin details lists; the 7 other folders in the plugin's repository, which belong to other plugins, are not offered. A p5.js request opened the plugin's algorithmic-art skill, and a harbour-log request followed the Codex skill.
• Relevance eval (TestSkillRelevanceEval, deepseek/deepseek-v4-flash, 12 paraphrases sharing almost no words with their skill plus 4 needing no skill, two runs per arm, each request in its own folder). Before is the base build 3f46952:

arm	paraphrases reaching their skill	requests left alone
before, memory on	12/24	8/8
before, memory off	0/24	8/8
after, memory on	16/24	8/8
after, memory off	15/24	8/8

• Gate on the head 2f87cbc. Build, vet (untagged and e2e) and make changelog-check are clean, and the touched light packages (skills, resident, remote, manual, e2e, store, cmd/codeaf) pass.
  • Full make check exits 2 and is not green. It failed in internal/run (TestNotesBeyondTheBoundWaitForTheNextBoundary: step cap), internal/session (TestHandingAYourCallToTheModelAndItsResolveAreOneRoad) and internal/telemetry (two spool tests).
  • None of these packages' code paths is touched here, and none of these tests failed in the previous full run on this branch. They have not been confirmed against the base.
  • internal/tui3 passes in that run, including the off-loop laws.

On hold since 2026-09-23: #1409 carries the approved skills design, so this branch stops here until Santosh decides.

Not reached yet

• Claude Code's disable-model-invocation and user-invocable frontmatter are not honoured; every skill is offered both ways.
• A SKILL.md deeper than a direct child of a skills folder is not read. A skills-dir plugin with only a nested skills/ folder is not read either.
• CLAUDE_CONFIG_DIR relocation is not followed.
• With CODEAF_HOME set, the user-scope folders are read under that folder, not $HOME. That is by design (home.Login).
• The picker discovers from the local disk even over --host to another machine.
• Task workers with memory off do not get the shelf through the executor's own store path.

[santoshkumarradha]

On hold and will close unmerged: skills land on dev through #1409 in the design settled in #1277. This branch stays as the reference for its discovery, the memory-off path and the evidence (real installs, the reworded-request check, the tmux end-to-end). See #1409 for the plan.