From 53d4c09016999e50b3256b60f1e6c0f3ccc5f62d Mon Sep 17 00:00:00 2001 From: Sejun Park Date: Tue, 6 Oct 2026 13:43:34 +0900 Subject: [PATCH 1/2] fix(registry): install harness-neutral skills for Claude Code too brainstorming, macos-storage-cleanup, and windows-cleanup were limited to the .agents target with no harness dependency behind it. brainstorming inherited the restriction from the old per-harness manifest, where it was added only to the Codex list; the cleanup skills were registered as Codex skills by default. Drop their targetExceptions so sjskills places them in both .agents and .claude. brainstorming's dispatch step named Codex's fork_turns parameter only; it now states the intent (a fresh, non-forked subagent) with the Codex and Claude Code forms. codex-cleanup, delegate-ui-to-claude, and impeccable keep their single target because each is genuinely tied to one harness. Co-Authored-By: Claude Opus 5.5 --- skill-registry.json | 9 ----- skills/brainstorming/SKILL.md | 74 ++++++++++++++++++----------------- 2 files changed, 38 insertions(+), 45 deletions(-) diff --git a/skill-registry.json b/skill-registry.json index f217169..1d04ca3 100644 --- a/skill-registry.json +++ b/skill-registry.json @@ -129,9 +129,6 @@ } }, "targetExceptions": { - "brainstorming": [ - ".agents" - ], "codex-cleanup": [ ".agents" ], @@ -140,12 +137,6 @@ ], "impeccable": [ ".claude" - ], - "macos-storage-cleanup": [ - ".agents" - ], - "windows-cleanup": [ - ".agents" ] }, "skills": [ diff --git a/skills/brainstorming/SKILL.md b/skills/brainstorming/SKILL.md index a92392f..4d87698 100644 --- a/skills/brainstorming/SKILL.md +++ b/skills/brainstorming/SKILL.md @@ -1,36 +1,38 @@ ---- -name: brainstorming -description: "Brainstorm from diverse perspectives through fresh subagents, then synthesize their insights. Explicit request only." ---- - -# Brainstorming - -Explore the invoking request through fresh, independent perspectives, then -synthesize a coherent landscape of ideas. Ground every worker in the topic, -goal, known facts, constraints, and desired outcome. - -## Design and dispatch the perspectives - -Choose at least three pairwise-distinct lenses suited to the topic. Examples -include stakeholder needs, feasibility, failure, assumption inversion, analogy, -second-order effects, and time horizons. Add a lens when it covers consequential -territory the others miss; avoid duplicating perspectives to increase the count. - -Write each prompt independently so its questions and thinking operation embody -its lens. Include enough shared facts to work without conversation history, -limit the assignment to analysis and response, and request concrete insights -with their reasoning. - -Spawn one fresh subagent per prompt with `fork_turns: "none"`. Omit `model` and -`reasoning_effort` so both inherit from the caller. Run independent prompts in -parallel when capacity allows; otherwise dispatch as slots become available. -Do not claim the requested fan-out occurred when fresh subagents are unavailable. - -## Synthesize - -Consider every worker's contribution. Explain the strongest ideas, unique -angles, recurring themes, tensions, and useful combinations. Preserve meaningful -disagreement and use parent judgment instead of treating consensus as proof. -Include open questions when another round would resolve something consequential. -The result should make the useful possibilities and tradeoffs clear without -repeating the workers' reports in full. +--- +name: brainstorming +description: "Brainstorm from diverse perspectives through fresh subagents, then synthesize their insights. Explicit request only." +--- + +# Brainstorming + +Explore the invoking request through fresh, independent perspectives, then +synthesize a coherent landscape of ideas. Ground every worker in the topic, +goal, known facts, constraints, and desired outcome. + +## Design and dispatch the perspectives + +Choose at least three pairwise-distinct lenses suited to the topic. Examples +include stakeholder needs, feasibility, failure, assumption inversion, analogy, +second-order effects, and time horizons. Add a lens when it covers consequential +territory the others miss; avoid duplicating perspectives to increase the count. + +Write each prompt independently so its questions and thinking operation embody +its lens. Include enough shared facts to work without conversation history, +limit the assignment to analysis and response, and request concrete insights +with their reasoning. + +Spawn one fresh subagent per prompt that does not inherit this conversation: +in Codex pass `fork_turns: "none"`; in Claude Code start a new general-purpose +agent rather than a fork. Leave model and reasoning-effort overrides unset so +both inherit from the caller. Run independent prompts in +parallel when capacity allows; otherwise dispatch as slots become available. +Do not claim the requested fan-out occurred when fresh subagents are unavailable. + +## Synthesize + +Consider every worker's contribution. Explain the strongest ideas, unique +angles, recurring themes, tensions, and useful combinations. Preserve meaningful +disagreement and use parent judgment instead of treating consensus as proof. +Include open questions when another round would resolve something consequential. +The result should make the useful possibilities and tradeoffs clear without +repeating the workers' reports in full. From 0bebe6b68e32da602fbff2f61aa1d8310f119281 Mon Sep 17 00:00:00 2001 From: Sejun Park Date: Tue, 6 Oct 2026 13:57:04 +0900 Subject: [PATCH 2/2] Address Codex review on Claude Code placement Placing brainstorming and the cleanup skills in .claude/skills would make them eligible for automatic selection, because Claude Code reads its invocation policy only from frontmatter. Add disable-model-invocation to all three, matching allow_implicit_invocation: false in their OpenAI metadata; no shared instruction routes to them by name. In Claude Code an unset subagent model can resolve to a configured subagent default instead of the caller's, so brainstorming now tells Claude to pass the caller's model explicitly. The Agent tool takes model aliases rather than "inherit", so the instruction names the intent. Co-Authored-By: Claude Opus 5.5 --- skills/brainstorming/SKILL.md | 7 +++++-- skills/macos-storage-cleanup/SKILL.md | 1 + skills/windows-cleanup/SKILL.md | 1 + 3 files changed, 7 insertions(+), 2 deletions(-) diff --git a/skills/brainstorming/SKILL.md b/skills/brainstorming/SKILL.md index 4d87698..d34037a 100644 --- a/skills/brainstorming/SKILL.md +++ b/skills/brainstorming/SKILL.md @@ -1,6 +1,7 @@ --- name: brainstorming description: "Brainstorm from diverse perspectives through fresh subagents, then synthesize their insights. Explicit request only." +disable-model-invocation: true --- # Brainstorming @@ -23,8 +24,10 @@ with their reasoning. Spawn one fresh subagent per prompt that does not inherit this conversation: in Codex pass `fork_turns: "none"`; in Claude Code start a new general-purpose -agent rather than a fork. Leave model and reasoning-effort overrides unset so -both inherit from the caller. Run independent prompts in +agent rather than a fork. Workers use the caller's model and reasoning effort: +in Codex leave `model` and `reasoning_effort` unset; in Claude Code set `model` +to the caller's model, because an unset model can resolve to a configured +subagent default. Run independent prompts in parallel when capacity allows; otherwise dispatch as slots become available. Do not claim the requested fan-out occurred when fresh subagents are unavailable. diff --git a/skills/macos-storage-cleanup/SKILL.md b/skills/macos-storage-cleanup/SKILL.md index 3914421..0387065 100644 --- a/skills/macos-storage-cleanup/SKILL.md +++ b/skills/macos-storage-cleanup/SKILL.md @@ -1,6 +1,7 @@ --- name: macos-storage-cleanup description: "Audit and reclaim macOS storage using measured, authorized cleanup of caches, build artifacts, app data, and local copies. Explicit invocation only; excludes Codex-owned storage, repair, and device reset." +disable-model-invocation: true compatibility: "Requires macOS and read-only filesystem, process, and command inspection; optional cleanup requires the supported owning app or CLI." --- diff --git a/skills/windows-cleanup/SKILL.md b/skills/windows-cleanup/SKILL.md index 5d7ac1b..863d729 100644 --- a/skills/windows-cleanup/SKILL.md +++ b/skills/windows-cleanup/SKILL.md @@ -1,6 +1,7 @@ --- name: windows-cleanup description: "Diagnose Windows slowdown and reclaim storage or reduce startup and background load with supported tools. Explicit invocation only; excludes registry optimization, debloating, incident response, hardware repair, and reset." +disable-model-invocation: true --- # Windows Cleanup