feat(plugins): plugins, MCP, and skills management on the provider-plugin model

[lucas77778]

Builds the client-facing management surface for provider plugins, user-owned custom MCP servers, and skills on top of the provider-plugin model that landed with CODE-23/403/432 (which was backend-only: no wire contract, no UI, managementCapabilities hardcoded all-false).

Supersedes the closed #261 / CODE-382~387 batch — that design invented a parallel "capability unit + connector" concept in the same paths. Design doc: Plugins / MCP / Skills Management.

What ships

Wire (protocol 67; minimum compatible 67)

• plugin.list.get/result carrying Plugin[], standalone skills, and providerStatus[] — so the UI can tell "this agent has no plugins" from "its CLI failed".
• plugin.set-enabled/plugin.updated and skill.set-enabled/skill.updated, both replying with the re-read entity so clients patch one cache entry instead of re-running discovery.
• config.get/config.set gain custom MCP servers: masked reads (envKeys/headerKeys, never values) and per-key patch ops.
• session.started gains optional mcpWarnings.

Provider management, honestly capability-gated

• Claude plugin enable/disable via claude plugin enable|disable -s <scope>; Codex enable/disable via config/value/write on plugins."<id>".enabled.
• Codex plugin install/uninstall via app-server plugin/install and plugin/uninstall; Claude installation remains unimplemented by this adapter.
• Per-skill toggles: Claude writes skillOverrides in settings.json (what the TUI /skills dialog writes; read-modify-write preserving every other key and any finer name-only/user-invocable-only tier the user set), Codex calls skills/config/write.
• Both providers blind-write — a nonexistent plugin/skill still reports success — so every write is followed by a re-read, and that readback is the only success check. Covered by tests.

Custom (BYO) MCP servers

• Persisted in ~/.linkcode/config.json (0600, per-entry tolerant parse), injected into MCP-capable sessions' StartOptions.mcpServers alongside the existing simulator endpoint, with agent-unsupported/name-conflict warnings instead of silent drops.
• Writes are patch ops with per-key secret semantics (blank = keep, typed = replace, removed = delete). A whole-array replace would clobber real secrets with mask placeholders, so it is structurally impossible here. buildCustomMcpPatch is the highest-value unit test in the batch.

Settings page — one "Plugins & Skills" category on desktop and webview, four tabs: Plugins (installed only, provider-grouped, capability-gated switches), Market (uninstalled marketplace listings), MCP (custom servers + read-only plugin-provided ones), Skills (plugin + standalone, per-skill switches). Manual refresh only — discovery is a real CLI shell-out.

Verification

• pnpm check:ci + pnpm test green on the current head (2341 tests).
• Dev-mock Playwright pass, 17/17: card rendering, toggle round-trip, secret never reaching the DOM, blank-save preserving stored secrets, add/edit/remove, empty states.
• Live against a real daemon before the latest rebase (then wire 64): 469 real plugins + 70 skills discovered; claude plugin disable/enable round-trips on the real CLI; a real skill toggle writes skillOverrides and restoring it removes the key; ~/.claude/settings.json diffed byte-identical to a pre-test backup afterwards.

Two real bugs were found by that live pass and are fixed here rather than left for later:

• Discovery was broken for both providers on any current install (CODE-505). claude's available[].source became a union (bare path string in 52 of 275 entries, object with its own source discriminator in the rest) and its version is null for 262 of 275; codex omits version entirely while the schema required the key present. One bad entry failed the whole array, so the page was empty everywhere. Regression fixtures now cover every observed shape.
• Two secret-row inputs inside one base-ui Field both rendered name="secrets.0.value" (Field owns a single control), silently dropping the key field from the form.

Deliberately out of scope

Managed/HQ connectors and OAuth (CODE-94/96/340 untouched); plugin update and Claude install/uninstall (Codex install/uninstall ships here); Claude's name-only/user-invocable-only skill tiers have no wire representation yet (preserved, not coarsened); retrofitting the same masking onto accounts, which still round-trips plaintext.

The Market list renders a bounded 60 entries per provider with a visible "showing 60 of 273 — narrow it down with search" note rather than truncating silently; virtualizing it is a follow-up if browsing the full catalog matters.

Closes CODE-487, CODE-488, CODE-490, CODE-491, CODE-492, CODE-493, CODE-494, CODE-495, CODE-496, CODE-497, CODE-502, CODE-503, CODE-504, CODE-505.

[linear-code]

CODE-487

CODE-488

CODE-490

CODE-491

CODE-492

CODE-493

CODE-494

CODE-495

CODE-496

CODE-497

CODE-502

CODE-503

CODE-504

CODE-505

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Zerlight]

I found three high-priority safety issues and seven correctness or UX issues. The most serious problems are that custom MCP credentials bypass the existing secret vault, while Claude skill toggles target the wrong settings layer and can weaken settings-file permissions. Please address the inline comments before merging.

Verification: 136 focused tests passed. In the full suite, 2,312 tests passed and one unrelated timeout passed immediately when rerun in isolation. The wire-version change from 66/66 to 67/67 is consistent with the new required config.get.result.customMcpServers field.

[Copilot]

🟡 Human review recommended

It introduces a wide cross-layer wire/schema + daemon persistence/secret-handling + client/UI surface area change where small mismatches can cause hard-to-debug compatibility and data-loss issues.

Review details

Suppressed comments (2)

packages/foundation/schema/src/wire/message.ts:16

• The wire-version bump here (66 → 67) doesn’t match the PR description (“Wire (62 → 64)”). Either the PR description needs updating (likely after rebases), or these constants need to reflect the intended version bump so reviewers/users aren’t misled about required lockstep rebuilds.
  packages/presentation/ui/src/shell/plugins/plugins-shell.tsx:20
• This doc comment says “three tabs”, but the component renders four tabs (Plugins/Market/MCP/Skills).

• Files reviewed: 79/79 changed files
• Comments generated: 0 new
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[Zerlight]

Thanks for addressing the first review. The earlier UI, schema, and persistence issues are largely resolved, and the focused verification is green. Three correctness issues remain: custom MCP persistence is not crash-consistent across the vault/config pair, Codex install and uninstall have no shutdown cancellation path, and the Claude settings rewrite still narrows some existing file modes through the process umask. Please address the inline comments before merging.

[Copilot]

🟡 Not ready to approve

There are user-visible text/grammar and documentation mismatches (and an apparent PR-description wire-version mismatch) that should be corrected before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Review details

Suppressed comments (5)

packages/presentation/ui/src/shell/plugins/plugins-shell.tsx:20

• The doc comment says this page has "three tabs", but the component renders four (plugins/market/mcp/skills). This can confuse future edits and reviewers.
  packages/presentation/ui/src/shell/plugins/skills-tab.tsx:22
• This doc comment says standalone skills are "display-only" and that skills are grouped by plugin, but the UI renders standalone skills with toggles (when supported) and does not actually group by plugin in the list.
  packages/presentation/ui/src/shell/plugins/custom-server-list.tsx:21
• The comment says "LinkCode-owned" custom MCP servers, but these are user-owned (BYO) servers; LinkCode-owned servers are the simulator/browser tool endpoints and should stay reserved internally. The current wording conflicts with the UI copy about "servers you add here are yours".
  packages/presentation/i18n/src/locales/en.ts:757
• The English strings use plural component kinds ("skills", "commands", …) with a naive "{count} {kind}" template, which will render ungrammatical text for singular counts (e.g. "1 skills"). Consider switching to singular kind labels + ICU pluralization.
  packages/foundation/schema/src/wire/message.ts:16
• PR description claims a wire bump "62 → 64", but this diff sets WIRE_PROTOCOL_VERSION / MIN_COMPATIBLE_WIRE_VERSION to 67. Please align the PR description with the actual wire version being shipped (or adjust the constants if 67 is unintended), since this affects upgrade sequencing across daemon/clients.

• Files reviewed: 79/79 changed files
• Comments generated: 0 new
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[Copilot]

🟡 Not ready to approve

It introduces a few user-visible correctness issues (missing i18n keys referenced by settings search, and a semantic change that forces mcpServers: [] even when no MCP config was provided) that should be resolved before merging.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Review details

Suppressed comments (3)

packages/client/workbench/src/settings/search.ts:102

• Settings search keywords reference plugins.componentKind.*, but those message keys don’t exist in the i18n locales added in this PR, so the settings sidebar search will show raw keys or throw depending on use-intl config. Use existing, defined strings (tab titles / section titles) for plugin-related keywords, or add the missing messages to every locale.
  packages/host/engine/src/session/start-options-resolver.ts:120
• withCustomMcpServers always writes mcpServers: servers even when options.mcpServers was undefined and nothing was actually injected (e.g. all custom servers skipped due to conflicts/unsupported). This changes the semantic from “no MCP config” to an explicit empty list and can leak into provider adapters; preserve undefined when the merged list is empty and the caller didn’t provide one.
  packages/presentation/ui/src/shell/plugins/custom-server-list.tsx:47
• The custom MCP section shows the “empty” state while rows is still undefined (initial load). This conflates loading with a true empty list and can flash misleading UI; treat rows === undefined as a loading state (render nothing/skeletons) and only show empty when rows is defined and length is 0.

• Files reviewed: 79/79 changed files
• Comments generated: 0 new
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[lucas77778]

Addressed all findings from Copilot’s two completed reviews:

• 2f84c5e7: corrected stale comments/copy, ICU component plurals, and the PR wire-version description.
• 2bdf2100: restored the singular componentKind locale keys used by settings search while retaining plural-aware counts.
• 9320b3e8: preserves an absent mcpServers value when every custom server is skipped, and renders a skeleton rather than the empty state while custom MCP data is loading. Added regression coverage for both semantics.

Validation: pnpm check:ci (the final lint pass used the equivalent serial pnpm lint:ci after an auto-concurrency worker OOM), pnpm test (2342 passed), and the webview production/mock browser smoke all pass.

The latest requested Copilot re-review did not run because the requester had reached the Copilot review quota.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.