Repository navigation
feat: port upstream skills and Tools settings (#17691) - #39
Merged
Merged
Conversation
Adds a Tools settings page with Skills and MCP servers tabs, scoped to environments and projects like other settings. - MCP servers added in T3 are injected into every provider session next to the t3-code server (Claude, Codex, Cursor, OpenCode 1/2, Grok and ACP agents). Sensitive env/header values live in the secret store. - Skills can be switched off by name; Claude, Codex and OpenCode 1.x hide them from the session, and the composer drops them from its menu. - Projects override servers and skill switches per name instead of replacing the environment's list. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Main moved per-thread MCP config into the McpProviderSessions service and the ACP adapter into packages/provider-acp. The user's tools now ride on that service's config, and the adapters read them from the session they are handed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
claudeMcpQueryOverrides built the skillOverrides settings, but the turn never handed them to makeClaudeQueryOptions, so a skill switched off in Settings → Tools still loaded in Claude sessions. Found by running a real Claude turn; the new test opens a turn and checks the query options. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Renaming an MCP server or variable keeps its stored secret: the client sends where it was saved and the server moves it. A sensitive value hand-edited into settings.json moves into the store on the next save. - Turning a secret off and on again with nothing typed keeps it. - Editing a server at All environments no longer copies it, without its secrets, onto environments that never had it; adding one checks names across every environment. - A settings-only client can change a project's skills or switches when the project has its own servers: resending a server unchanged is not a provider change. - Environment server switches need provider management, like the server. - OpenCode 2 always denies other threads' user servers and retries ones that failed to register. OpenCode 1 and 2 give an external server the user's servers, and OpenCode 1 reapplies skill rules on resume and rollback. - ACP agents get an absolute command for bare names like npx. - The composer hides disabled skills from slash commands too, a project switch for a server named like an object property is ignored, and a search for MCP servers opens that tab. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The SDK passes mcpServers to the CLI as an inline --mcp-config argument,
which any local user can read. A sensitive env value or header from
Settings → Tools now goes into the child's environment and --mcp-config
carries a ${VAR} reference, the same as T3's own token since pingdotgg#17408.
Checked against claude 2.1.292: a stdio server received the expanded value.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…skills by scope - Secret copies read every source before anything is written, so swapping two variable names no longer leaves both with one value. - A repository skill under a folder named plugins stays in Repository; the plugins path only classifies skills a provider reports without a scope. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two sensitive entries with the same name made the second reference point at the first one's placeholder. Each name now gets one reference holding its last value, matching how the variable record resolves duplicates. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s that ask The user's servers were added to Claude's allowedTools, which skips the permission prompt in approval-required and auto-accept sessions. Full access already runs them without asking, so the allow rule only bypassed the modes that ask. Their tools now go through the session's approval. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds a hidden `t3 skills-cli` subcommand that runs the pinned skills CLI (vercel-labs/skills 1.7.1, telemetry off) and a SkillLibrary service with preview, install, update, remove and inspect, exposed as skills.* RPCs. Installs land where `npx skills` puts them, with links for agents that read only their own folder, chosen from a per-agent folder table. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ge Tools Settings → Tools → Skills gets Add skills (preview a source's skills, then install them at the current scope), groups installed skills under the source they came from with one switch, marks conflicting copies of a name, and adds View files, Update and Remove to each skill's menu. Agents get t3_tools_read, t3_tools_update, t3_skills_install and t3_skills_remove. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- remove: the skills CLI exits 0 when it can't delete a skill, so check the folder and lock entry afterwards and report the failure. - update: follow the CLI's own update source rules, using a GitLab or git lock's recorded URL and full-depth discovery when a source can't take a folder. - Show a per-skill update failure, and lock the source field while finding. - t3_tools_update builds its patch inside the settings write lock (updateSettingsWith), so a concurrent change isn't overwritten. - Project skill switches match names ignoring case, like the environment's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ling - install, update and remove share a lock: each skills CLI run rewrites a whole lock file, so two at once could drop each other's entries. - A project skill switch is written with the environment's spelling of the name. - The CLI child gets the host environment it resolved (HostProcessEnvironment) rather than the process's own, so tests run it against their temp home. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ny skill name - A global install, update or remove now rescans every project an agent has listed, since a project's skill list includes the global skills. - Project skill switches use a null-prototype map, so a skill named __proto__ gets a switch like any other. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- After a skill change, agents that list the same folder rescan it one after another: a fresh scan drops the others' snapshots of that folder, and the registry discards a scan whose starting snapshot changed. - mcpServerEnabledPatch reads only own keys, so `constructor` isn't taken for an inherited server. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ports the Skills UI from n0mahd's skills series (pingdotgg#17513) onto the Tools page: - dense one-line rows, sections with counts and a hint ('This project · 2 · Lives in this repo'), ✦ when every agent loads a skill; - installed packs as collapsible 'From owner/repo' groups with one switch and the first three skills shown; - a Needs attention filter (conflicts, and project or global skills an enabled agent doesn't load); - a skill page replacing the View files dialog: Used by chips, the source, Includes scripts with the script files, Copy path, Update and Remove, a file tree with script labels, SKILL.md rendered or as source, each copy of a conflicting name, and Escape back to the list. The server marks each file as a script, so the list and the tree agree. After an install, one agent rescans a project fresh and the rest follow without dropping its result, so every agent's list shows the new skill. Co-Authored-By: n0mahd <39080654+n0mahd@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The CLI matches --skill against a skill's SKILL.md name and keys its lock files by it, but installs into that name sanitized as a folder (`Git Review` -> `git-review`). Preview now returns the CLI's name with the folder, and inspect, update and remove find a lock entry by either form. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lts whole - The Add skills dialog ticks at most 200 skills, the most one install takes, says so for a bigger source, and the agent tool shares the limit. - The skills CLI prints its --json result and exits at once, so a result over 64 KiB was cut short in the pipe and the preview failed. The subcommand now exits only after stdout drains. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cking Deferring process.exit until stdout drained let the CLI's code run on past an exit it relied on: `add --list` returned into the install path. A blocking stdout pipe writes everything before the next statement, so process.exit stays immediate and nothing is cut. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The skill page drew its own Skills / section crumb under the scope picker. The open skill now lives in the URL (?skill=), so the settings header reads Settings / Tools / <skill>, with Tools leading back to the list. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…p pack rows At All projects, Codex's own skill list (read in the folder T3 was launched from) filled a This project section that belonged to no picked project. Project skills now come only from a scan of the picked checkout. Skills inside a From <source> pack, and its N more row, now start where the pack's name does instead of under its icon. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A pack's skills now line up with every other row. Its From <source> heading is a slim tinted band whose switch and chevron sit in the rows' columns. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The tabs, search and buttons took three rows above the list. Each tab now puts the tabs at the start of its own toolbar: search, the attention filter, rescan and Add skills on Skills; Add server on MCP servers. Needs attention becomes a warning icon with its count, shown only when something needs a look, and its tooltip says what that means. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Not available to Claude gave nothing to act on: the agent chips already show who loads a skill, and nothing in T3 makes another agent load it. Only a name used by different skills needs a look now, and its message says what to do. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ilter Each skill in the Add skills picker took three lines, so a big source showed three at a time. Rows now match the Skills list: name and one-line description, a script mark, and the file count on the right. A header box selects or clears every skill shown, and a source with many skills gets a filter. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hovering a description that ends in an ellipsis, in the Skills list or the Add skills picker, shows the full text. The fit is checked as the tooltip opens, so a line that fits shows none and the lists measure nothing until hovered. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Remove only reached skills the skills CLI had installed, so most of a skill folder could be switched off but not deleted. A skill's menu now has Remove for any copy an agent finds directly in its own skill folders, at home or in the project. One the CLI installed still goes through the CLI; another is deleted from each folder an agent reports for it, links first, along with the folder a link points to when that is one of the user's own. Plugin and built-in skills stay switch-only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copy path sat in the skill's menu, away from the path it copies. It is now an icon button after the folder path, for whichever copy is shown, and turns into a check once copied. The menu keeps Update and Remove. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With Copy path beside the path, the skill's menu held one or two items. They are now buttons on the skill's page: Update for a skill installed from a source, and Remove. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every agent session gets the t3-code server, but the MCP servers tab only listed servers added there, so it looked like T3's tools weren't given at all. It now heads the list as a built-in server with a switch that is on and locked, since nothing turns it off. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With several environments in scope, the Skills tab read only the primary environment, so a skill on one machine but not another looked the same. It now reads every environment in scope and merges their skills by name. A row shows the machines that have the skill, or one fleet mark when they all do; its page lists them and can copy the skill to the ones missing it. A copy goes through two new RPCs. skills.export packs the folder an agent reports for the skill (bounded at 4 MiB and 500 files). skills.import reinstalls from the skill's remote source when the CLI recorded one, so the copy can be updated, and otherwise writes the files into .agents/skills with links for agents that read their own folder. Update and Remove on a skill's page now reach every environment in scope that has it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings → Tools listed MCP servers but couldn't show what each one offers. `mcpServers.listTools` starts the named server the way an agent session would, with the project's overrides and the stored secrets, asks it for `tools/list` (following cursors), and stops it again, within 30 seconds. Command servers run in the project's checkout with the host environment under their own variables; URL servers fall back to SSE on a 404/405. Failures come back as a typed McpServerToolsError with a plain message. T3's own `t3-code` server is described from the toolkits `/mcp` registers, without starting anything; a registration test keeps that list in step. Listing runs the server's command with its secrets, so it needs `providers:manage`, the scope that gates adding that command. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clicking a server on the MCP servers tab, including the built-in t3-code row, opens its page in place of the list, the way a skill opens: the header reads Settings / Tools / <server>, and the Tools crumb or Escape goes back. The page shows the transport and command or URL, the switch, Edit and Remove as buttons, then the server's tools in dense rows that open to their parameters (name, type, required, description). Tools load through `mcpServers.listTools`, with a delayed skeleton, an error with Retry, and a refresh; an edit while the page is open lists them again. A server that is switched off can still be inspected. The row's switch, Reset and menu keep their clicks to themselves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bring in pingdotgg#17691 (Settings → Tools: skills install, update, remove, copy to environments, the tools MCP toolkit) and keep the fork's MCP hub authoritative for MCP servers: the Tools page's MCP servers tab opens Settings → MCP Servers, hub and Tools servers both reach every provider adapter, and the fork's hub contract types are renamed McpHub* to avoid clashing with upstream's agentTools. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
The fork has no way to see, install, update or remove skills across projects and providers. Upstream's open PR pingdotgg#17691 adds that as Settings → Tools, but it also brings its own MCP servers store and injection, which overlaps the fork's MCP hub (Settings → MCP Servers).
What was merged
upstream/t3code/skills-installer(36 commits) is merged as-is (no squash, no rewrite), so the daily upstream sync stays clean once pingdotgg#17691 lands. It adds:owner/repo/ git URL / folder (bundledskills@1.7.1CLI), Update, Remove, Copy to other environments, skill file viewer.toolsMCP toolkit for agents:t3_tools_read,t3_tools_update,t3_skills_install,t3_skills_remove.skills.inspect|preview|install|update|remove|export|importandmcpServers.listTools.How conflicts were resolved (28 files)
RpcAuthorization.layer(scopes, currentSettings, onAuthorized)keeps the fork's analytics hook and upstream's settings-aware scope check;projectsMutatescoping kept.FORK_RPC_METHODS),projectsListIgnoredFiles, the override sweep, Docs/AI PM/MCP hub toolkits all kept; skills handlers,SkillLibraryandMcpServerToolslayers added.T3_MCP_TOOLKITSalso lists the fork's Docs, AI PM and MCP servers toolkits. The tools toolkit getsThreadCommandExecutorlike the environment toolkit.mapSettingsSecretsstructure;updateSettingsWithadded.providerInstancesplus upstream'smcpServers/disabledSkillsproject overrides. The fork hub'sMcpServerName/Transport/Configwere renamedMcpHub*because upstream'sagentTools.tsexports the same names.upstreamServers(fork hub, via T3's forwarding route, withheld in read-only sandboxes) andtools(upstream servers + disabled skills). Claude adopts upstream's${T3_CODE_MCP_AUTHORIZATION}env reference, now also used for hub servers, so no bearer token sits in process args. Codex keeps the fork's dottedmcp_servers.<name>keys. OpenCode keepst3-mcp-*hub registrations and addst3u-*Tools registrations with matching deny/allow rules.useSettings/useScopedSettings; composer skill filtering written without manual memo (React Compiler file).@t3tools/shared/hostProcess→HostProcess,EnvironmentServerConfig,shouldShowInstanceBadgeneedingenabled) was adapted.apps/server/src/mcp/toolkits/worktree/registration.test.tsstays deleted (upstream main deleted it in test: delete 11k low-signal tests so CI runs faster and flakes less pingdotgg/t3code#22557).MCP overlap
The two systems are separate stores (
mcp-servers.jsonhub vssettings.mcpServers) with separate injection paths, so they don't double-inject unless the same server is added to both. To keep one place to manage servers, the fork hub stays the authority:#tools-mcp-serverssearch link redirect to Settings → MCP Servers (route-level,settings.tools.tsx). The duplicate search entry is removed.t3_tools_read/t3_tools_updatedescriptions point agents tomcp_servers_list/mcp_server_set_enabledfor hub servers.settings.mcpServersplumbing stays in place (empty in practice) so future syncs don't conflict.docs/user/tools.mdpoints to the MCP Servers guide.Verification
SkillLibrary,toolsSettings.logic,agentTools,tools/handlers,McpServerTools: 42/42 pass.origin/main.ChatView.tsx:1908, which also fails on a cleanorigin/main.origin/main: 6OpenCode2OrchestratorV2.integrationtests (timeouts) andMcpUpstreamRoute"runs a stdio server with its secret environment".no-manual-memo,prefer-effect-fn,no-test-timeout), and the fork's mcpHub files have their own existing lint errors. Upstream PR code was left untouched to keep the sync clean.Done by Claude Opus 5.5 in Claude Code (T3 Code).
🤖 Generated with Claude Code