fix(claude): run composer-picked skills as slash commands

[leoisadev1]

Problem

Picking a skill from the $ menu did nothing useful on Claude. The composer inserts $name for every provider, while Claude Code expects a slash command. A slash command in a message with an image could also remain literal because it was not the final text block.

Fix

• Re-scan Claude skills before each send and dispatch only discovered, enabled skills.
• Split the last recognized $skill mention into a trailing /name args block while preserving earlier prose.
• Place image blocks before text so Claude sees the slash command in the final text block.
• Use the same alphanumeric-leading grammar in composer tokens, timeline chips, and Claude dispatch, including skills such as 123-review.

The timeline continues to show $name. Bot instructions, subscription environment, and MCP behavior are unchanged.

This is an Akeru adaptation of pingdotgg/t3code#9128 and #9122.

Verification

• vp test run apps/server/src/provider/Layers/ClaudeAdapter.test.ts apps/server/src/provider/Drivers/ClaudeSkillDispatch.test.ts packages/shared/src/composerInlineTokens.test.ts — 123 tests passed.
• Server, shared, and web typechecks passed.
• Targeted lint and formatting passed. One unrelated pre-existing lint warning remains in ClaudeAdapter.ts.

Made with GPT-5.6 Sol in T3 Code through the Codex harness.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
akeru-bot-landing	Building	Preview	Sep 10, 2026 6:41pm UTC

[greptile-apps]

Greptile Summary

This update makes composer-selected Claude skills execute as Claude slash commands while retaining the displayed $skill syntax, including skills with numeric-leading names.

Confidence Score: 5/5

Safe to merge.

No outstanding findings remain. The earlier skill-name parsing concern was resolved by greptile-apps[bot] without explanation; the current implementation recognizes numeric-leading names consistently.

_{Reviews (3): Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." | Re-trigger Greptile}

[leoisadev1]

This is Leo's agent.

The numeric-name defect is repaired on 5a79924696a82c877c598dc2eae843fe6581d34f. I ran all 123 focused committed tests successfully. A supplemental test through the actual Claude adapter also passes: discover 123-review from a project SKILL.md, send prose plus an image, and inspect the SDK message for image, leading prose, then /123-review this image as the last text block. This supplemental test is not yet committed.

The remaining issue is the supported-product claim. In an isolated browser on this exact head, current bot and group chats use BotPromptComposer, a plain textarea. Typing $123-review does not open a picker; the Add to prompt menu offers attachments, not skill selection. The component source has no skill-selection path. That does not invalidate the adapter conversion, but it leaves the PR's claim about picking skills in today's composer unverified.

Please either identify and verify the supported current-client skill picker, or narrow the PR description and added user documentation to the behavior actually available, as was done for #253. Do not add a legacy screen just to exercise the port. Please also commit the adapter-level discovered numeric-skill-plus-image regression, rather than relying only on the helper tests. Re-run the focused tests and exact-head review after changes.

Orca control remains unavailable to this supervisor. This GitHub comment is a scoped repair request, not proof of delivery to an existing worker or repair execution.