[sambitcreate]

🤖 Generated with Claude Code

https://claude.ai/code/session_018VA6RFoM21JGeMe45njSXW

---

Update 2026-10-31 (backlog cleanup): the owner chose to keep only the desktop half. The CLI part (--add-dir skill folders and CLI plugin skills) is removed, because the CLI is becoming a headless server (#675). main is merged in (README clash fixed), the CLI-only discoverPluginSkillCandidates export and its test calls are gone, and docs/skills.md now says the CLI does not read installed plugins. Local checks on 890e3df: root typecheck, CI policy, skills-discovery + onboarding-flow (52) and skills-settings (3) tests pass.

---

Desktop skill discovery now includes enabled installed Claude Code plugins. Plugin skills keep namespaced identities (plugin:name), bounded reads, existing precedence, and workspace permission checks.

Explicit manifest skills: "." and "./" collections scan recursively under existing bounds; implicit root fallback remains direct-only.

Limits: default Claude configuration paths, exact canonical workspace matching, and no parent/worktree installation inheritance. Existing native catalog contracts are reused (phones only see the longer plugin:name display name; their tests were updated). Physical-device acceptance was not performed.

Flaky test record (2026-10-31): head ca575b6 hit tests/e2e/chat-message-queue.spec.ts:719 "messages sent during compaction queue behind it and survive a failed compaction" in Electron E2E (3/3). The chat showed "This chat could not be loaded" and Queue message stayed disabled. It passed on retry, but --fail-on-flaky-tests fails the job. This is the known chat-reload flake family, and this PR does not touch chats or the queue. Failed job. A main merge (d04b0a9) replaced that run before the rerun.

Head d04b0a9 hit main/services/bot-runtime/live-projection.test.ts "while classifying fails, a cached routine stays hidden…" ("retries are throttled, saw 3 calls") in macOS Desktop tests (runtime-subagents). This PR doesn't touch Bot runtime code. The root cause is a real retry-timer bug, fixed in #683. Failed job. A review fix (33ceb3e) replaced that run before the rerun.

---

Generated by Claude Code

[pullfrog]

Important

Two reproduced edge cases need attention: root-declared skill collections are missed, and CLI invocation-policy filtering can expose a shadowed plugin implementation.

Reviewed changes Local installed-plugin discovery and additional CLI skill directories, including their permission, precedence, and invocation-policy integration.

• Plugin inventory: Bounded local metadata reads, layered enablement, exact workspace matching, namespaced identities, and integration with desktop discovery budgets.
• CLI integration: Trusted plugin publication, repeated launch-relative --add-dir options, automatic-discovery gates, and native/explicit skill precedence.
• User-facing contracts: Settings and onboarding copy, discovery documentation, CLI help, and namespaced-name fixtures for both native clients.
• Validation: CLI build, desktop and CLI typechecks, and 110 focused tests passed; independent filesystem and built-CLI RPC probes reproduced both inline findings.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed f7785ca against the previously reviewed ba4b53e, with the full PR diff checked for context.

• Restored root collections: Allowed explicitly declared skills: "." and "./" collections to scan recursively while retaining the direct-only implicit fallback and existing containment limits.
• Preserved policy precedence: Claimed the winning CLI plugin identity before invocation-policy filtering, preventing a shadowed implementation from becoming available.
• Extended behavioral coverage: Added discovery and built-CLI regression assertions for both fixes and documented their behavior. Both earlier review findings are addressed; their threads were already resolved.
• Verified the delta: Passed 57 discovery/registry tests and 25 built CLI runtime/path tests, plus the CLI build and typecheck. No repository changes were made.

  ^{｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}

[very-hermes-bot]

Hermes Review Bot

Confidence: 4

Engine: opencode/opencode-go/deepseek-v4.1-flash
Review mode: medium
Head: 86c44ba343f70447197e1e413ef16ed80573d1ee
Generated: 2026-10-31T16:18:53+00:00
Reviews: 1

Summary

Desktop skill discovery now also reads enabled, installed Claude Code plugins: main/services/claude-plugin-skill-roots.ts resolves ~/.claude/plugins/installed_plugins.json, merges enabledPlugins from user and active-workspace settings, selects the first applicable install record per plugin key, and returns bounded search roots. main/services/skills-discovery.ts scans those roots and prefixes each authored name with the plugin namespace, with plugin skills sorted below ordinary folders inside the same workspace/global source so authored skills still win. The manifest skills paths are containment-checked, symlinks are rejected, and named SKILL.md files are read at depth 0 while the implicit root fallback stays direct-only. Docs, onboarding, settings copy, and the iOS/Android catalog fixtures were updated to reflect the longer plugin:name display name. A maintainer should double-check the untested manifest path shapes (a manifest that explicitly names a collection) and the fact that managed/synced Claude inventory variants beyond the documented four scopes are not exercised by the new tests.

Confidence Score: 4/5

Traced discovery through SkillRegistry, Bot inventory, Telegram, and the CLI's separate discovery path; verified phone codecs accept the colon name and that the new tests' temp homes actually exercise the roots. Remaining uncertainty is limited to Claude inventory variants and physical-device acceptance, which the PR explicitly leaves out.

📁 Important Files Changed

• main/services/claude-plugin-skill-roots.ts (new) — reads plugin inventory/settings/manifests, applies workspace matching and namespace/scope rules, and produces bounded search configs.
• main/services/skills-discovery.ts — adds plugin roots to the authoritative scan, namespaces plugin names, enforces the 80-character identity bound, and adds the baseSkill depth-0 rule.
• main/services/skills-discovery.test.ts — new behavioral coverage for enablement, scope/permission, manifest containment, symlink rejection, root fallback, and malformed metadata.
• docs/skills.md, README.md, PRODUCT.md — document plugin discovery, precedence, limits, and the desktop-only scope.
• renderer/components/settings/skills-settings.tsx, renderer/components/onboarding-feature-gallery.tsx, renderer/components/onboarding-flow.test.tsx — user-facing copy and its rendering assertion.
• ios/AidenOnTheGoTests/AidenChatTests.swift, android/.../AidenSkillTest.kt — show the existing contract accepts a namespaced display name; no production mobile code changed.

Findings

No findings.

📊 Sequence Diagram

sequenceDiagram
    participant C as Chat/composer
    participant R as SkillRegistry
    participant D as skills-discovery
    participant P as claudePluginSkillRoots
    participant FS as Filesystem
    C->>R: snapshot(workspaceId)
    R->>D: discoverSkillCandidates(workspaceRoot or undefined)
    D->>P: claudePluginSkillRoots(home, workspaceRoot)
    P->>FS: read installed_plugins.json + settings*.json + plugin.json
    P-->>D: bounded plugin roots + namespace + scope
    D->>FS: scan SKILL.md under ordinary and plugin roots
    D-->>R: candidates (namespaced, prioritized)
    R-->>C: catalog/available skills after collision resolution

Loading

Machine-Readable Findings

[]

---

Last reviewed commit: 86c44ba343f7
Reviews (1) · Comment /hermes review to trigger a new review · /hermes review full for full re-review

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed the changes since f7785ca, focusing on the desktop-only scope after the merge of main.

• Removed CLI additions: Removed installed-plugin discovery, additional-directory options, and their CLI integration and tests while preserving existing CLI skill discovery.
• Retained desktop coverage: Removed the CLI-only discovery export and updated tests to keep exercising desktop root collections, deduplication, containment, and malformed metadata handling.
• Clarified supported scope: Updated the skill guide to explicitly distinguish desktop plugin discovery from standalone CLI behavior.
• Validated the delta: Passed 86 focused desktop tests and 26 built-CLI runtime/path tests, both typechecks, and the CLI build. Electron and native mobile runtime suites were not rerun; their behavior is unchanged by this delta.

  ^{｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}

[very-hermes-bot]

Inline review — 1 finding(s) anchored to the diff. See the pinned summary comment for the overview.

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed aa4d673 against the previously reviewed 890e3df, focusing on the asynchronous plugin-metadata discovery path.

• Moved metadata I/O off the main thread: Converted metadata opens, descriptor reads, cleanup, and canonical-path resolution to fs/promises, with awaited results throughout the discovery call chain.
• Preserved discovery semantics: Retained bounded reads, no-follow/nonblocking open flags, ordered installation selection, layered enablement, and exact canonical workspace matching; resolved home/workspace paths once per call.
• Validated the integration: Passed all 57 focused discovery and registry tests and the root TypeScript check. Electron responsiveness benchmarks and native mobile runtime suites were not run; this delta changes no mobile or wire contracts.

  ^{｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}

[very-hermes-bot]

Inline review — 1 finding(s) anchored to the diff. See the pinned summary comment for the overview.

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed ca575b6 against the previously reviewed aa4d673, focusing on the plugin skill name-length fix.

• Omitted overlong plugin identities: Added a discovery-time check for composed plugin:name values exceeding 80 code points after NFKC, preventing those candidates from appearing as unavailable “Invalid skill” entries while leaving ordinary skill validation unchanged.
• Extended coverage and documentation: Added an overlong plugin-name fixture that verifies valid neighboring skills remain discoverable, and documented the omission rule in docs/skills.md.
• Validated the integration: Passed all 64 focused discovery, registry, and catalog-validation tests and the root TypeScript check. Electron and native mobile runtime suites were not rerun; this delta changes no wire contracts or client implementations.

  ^{｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}

[very-hermes-bot]

Inline review — 1 finding(s) anchored to the diff. See the pinned summary comment for the overview.

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed the manifest-file traversal fix in 33ceb3e against d04b0a9; the intervening main merge had already been reviewed separately.

• Restricted named-file discovery: Limited explicit SKILL.md paths to a depth-zero search, retaining the named skill without importing nested skills while preserving recursive directory and root-collection declarations.
• Extended behavioral coverage and documentation: Added an undeclared nested-skill fixture that verifies file-only discovery and documented that distinction in docs/skills.md.
• Validated the integration: Passed all 65 focused discovery, registry, and catalog-validation tests and the root TypeScript check. Electron and native mobile runtime suites were not rerun; this delta changes no UI or wire contracts.

  ^{｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}

[very-hermes-bot]

Inline review — 1 finding(s) anchored to the diff. See the pinned summary comment for the overview.

[pullfrog]

✅ No new issues found.

Reviewed changes Reviewed 86c44ba against the previously reviewed 33ceb3e, focusing on collection-root skill eligibility.

• Separated collection and skill bases: Replaced namespace-wide base-file eligibility with a per-search baseSkill flag, excluding the standard skills/SKILL.md while retaining explicit manifest paths and the root fallback.
• Extended behavioral coverage: Added a collection-level SKILL.md fixture whose exact catalog assertion verifies exclusion while retaining ordinary nested skills, custom paths, and the root skill.
• Validated the integration: Passed all 65 focused discovery, registry, and catalog-validation tests and the root TypeScript check. Electron and native mobile runtime suites were not rerun; this delta changes no UI or wire contracts.

  ^{｜ View workflow run ｜ Using gpt-6.1-sol ｜ 𝕏}