@spardutti/claude-skills 1.28.0 → 1.29.1

package/README.md

@@ -101,18 +101,22 @@ The CLI will:
 
 After installing skills, the CLI asks if you want to set up automatic skill evaluation. If you say yes, it will:
 
-- **Create a PreToolUse gate hook** at `.claude/hooks/skill-gate.sh`
+- **Install two hooks** in `.claude/hooks/`:
+  - `skill-gate.sh` — PreToolUse gate on `Write|Edit|MultiEdit`
+  - `skill-gate-automark.sh` — PostToolUse on `Skill` that auto-clears the gate
 - **Update your `CLAUDE.md`** with the skill-evaluation rule
 
-The gate hard-blocks `Write`, `Edit`, and `MultiEdit` tool calls until Claude has evaluated the available skills and emitted the literal token `[skills-checked]` in its response. Once emitted, every subsequent edit in that turn passes through. The next user prompt resets the gate.
+The gate hard-blocks `Write`, `Edit`, and `MultiEdit` until a per-session marker file exists at `/tmp/claude-skill-gate-<SESSION_ID>`. The marker is created automatically the first time Claude invokes any `Skill()` in the session — so the normal flow is: Claude lists skills as ACTIVATE/SKIP, calls `Skill()` for the ACTIVATE ones, and the gate clears for the rest of the session. If every skill is SKIP, Claude clears the gate by running `touch /tmp/claude-skill-gate-<SESSION_ID>`.
+
+The marker is **per-session, not per-turn** — short follow-ups like "yes" don't re-lock the gate after evaluation has already happened.
 
 Unlike a soft reminder injected into context (which Claude can ignore), the gate denies the tool call outright — so the only path forward is to actually evaluate skills.
 
-The gate auto-passes when the project has no skills installed, so it's safe to leave on globally.
+The gate auto-passes when the project has no `.claude/skills/*/SKILL.md` files, so it's safe to leave on globally.
 
 ### What gets created
 
-**`.claude/settings.json`** — Registers the gate on file-writing tools:
+**`.claude/settings.json`** — Registers both hooks:
 
 ```json
 {
@@ -127,12 +131,23 @@ The gate auto-passes when the project has no skills installed, so it's safe to l
           }
         ]
       }
+    ],
+    "PostToolUse": [
+      {
+        "matcher": "Skill",
+        "hooks": [
+          {
+            "type": "command",
+            "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/skill-gate-automark.sh"
+          }
+        ]
+      }
     ]
   }
 }
 ```
 
-**`CLAUDE.md`** — Appends the evaluation rule, including the `[skills-checked]` sentinel that the gate looks for.
+**`CLAUDE.md`** — Appends the skill-evaluation rule that tells Claude to enumerate skills as ACTIVATE/SKIP and call `Skill()` for ACTIVATE entries before writing code.
 
 ## Manual Install
 

package/lib/setup-hook.mjs

@@ -2,25 +2,15 @@ import { mkdir, writeFile, readFile, chmod } from "node:fs/promises";
 import { join, resolve } from "node:path";
 
 // PreToolUse gate on Write|Edit|MultiEdit. Blocks the tool call unless
-// a per-prompt marker file exists at /tmp/claude-skill-gate-<UUID>,
-// where <UUID> is the uuid of the most recent typed user prompt. The
-// assistant creates the marker via Bash; Bash output is flushed
-// synchronously, so the marker is race-free against the message-buffering
-// behavior that broke the earlier text-sentinel approach (Claude Code
-// writes assistant content blocks to JSONL only after the turn completes,
-// so [skills-checked] text emitted in the same message as a tool_use is
-// invisible to PreToolUse). /tmp keeps the marker out of the project
-// directory so it can't leak into commits.
-//
-// LAST_PROMPT_UUID is detected by matching user-role lines whose content
-// is a JSON string ("role":"user","content":"..."), which excludes
-// tool_results, skill loads, task notifications, and slash-command
-// payloads (all of which use array content).
+// a per-session marker file exists at /tmp/claude-skill-gate-<SESSION_ID>.
+// Per-session (not per-prompt) so simple confirmations like "yes" don't
+// re-lock the gate after evaluation has already happened in the session.
+// The PostToolUse hook on Skill creates the marker automatically; for
+// all-SKIP cases the model can `touch` the path manually.
 //
 // Pass-through cases:
 //   - project has no .claude/skills/*/SKILL.md files
-//   - transcript_path missing or unreadable
-//   - no typed user prompt found in transcript
+//   - session_id missing from hook input
 const GATE_SCRIPT = `#!/bin/bash
 # PreToolUse gate: forces skill evaluation before file-writing tools run.
 
@@ -33,48 +23,34 @@ if ! find "$PROJECT_DIR" -path '*/.claude/skills/*/SKILL.md' 2>/dev/null | grep
   exit 0
 fi
 
-TRANSCRIPT=$(printf '%s' "$INPUT" | grep -o '"transcript_path":"[^"]*"' | head -1 | sed 's/"transcript_path":"//; s/"$//')
-if [ -z "$TRANSCRIPT" ] || [ ! -f "$TRANSCRIPT" ]; then
-  exit 0
-fi
-
-LAST_LINE=$(grep -E '"role":"user","content":"' "$TRANSCRIPT" 2>/dev/null | tail -1)
-LAST_PROMPT_UUID=$(printf '%s' "$LAST_LINE" | grep -o '"uuid":"[^"]*"' | head -1 | sed 's/"uuid":"//;s/"$//')
-if [ -z "$LAST_PROMPT_UUID" ]; then
+SESSION_ID=$(printf '%s' "$INPUT" | grep -o '"session_id":"[^"]*"' | head -1 | sed 's/"session_id":"//; s/"$//')
+if [ -z "$SESSION_ID" ]; then
   exit 0
 fi
 
-MARKER="/tmp/claude-skill-gate-$LAST_PROMPT_UUID"
+MARKER="/tmp/claude-skill-gate-$SESSION_ID"
 if [ -f "$MARKER" ]; then
-  find /tmp -maxdepth 1 -name 'claude-skill-gate-*' ! -name "claude-skill-gate-$LAST_PROMPT_UUID" -delete 2>/dev/null
   exit 0
 fi
 
 cat <<EOF
-{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny","permissionDecisionReason":"Skill evaluation required before file edits. (1) List each available skill as ACTIVATE or SKIP with a one-line reason. (2) Call Skill() for any ACTIVATE entries — this auto-clears the gate. If all skills are SKIP, run this exact Bash command instead: touch /tmp/claude-skill-gate-$LAST_PROMPT_UUID  (3) Then retry the file edit. The marker is unique to this user prompt and is auto-cleaned on the next prompt."}}
+{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny","permissionDecisionReason":"BLOCKED: skill evaluation required before file edits in this session.\\n\\nStep 1 — evaluate every available skill as ACTIVATE or SKIP with a one-line reason.\\n\\nStep 2 — you MUST take EXACTLY ONE of these tool actions to clear the gate. Listing skills in text is NOT enough; retrying the edit without doing one of these will be denied again:\\n  (a) If any skill is ACTIVATE → call Skill(name) for it. This auto-clears the gate.\\n  (b) If ALL skills are SKIP → run this Bash tool call: touch /tmp/claude-skill-gate-$SESSION_ID\\n\\nStep 3 — only after Step 2 completes, retry the file edit."}}
 EOF
 exit 0
 `;
 
-// PostToolUse on Skill: auto-creates the gate marker so the model doesn't
-// need an explicit touch. Uses the same UUID-detection logic as the gate.
+// PostToolUse on Skill: auto-creates the per-session gate marker.
 const AUTO_MARK_SCRIPT = `#!/bin/bash
-# PostToolUse on Skill: auto-marks the skill-gate as satisfied.
+# PostToolUse on Skill: auto-marks the skill-gate as satisfied for the session.
 
 INPUT=$(cat)
 
-TRANSCRIPT=$(printf '%s' "$INPUT" | grep -o '"transcript_path":"[^"]*"' | head -1 | sed 's/"transcript_path":"//; s/"$//')
-if [ -z "$TRANSCRIPT" ] || [ ! -f "$TRANSCRIPT" ]; then
-  exit 0
-fi
-
-LAST_LINE=$(grep -E '"role":"user","content":"' "$TRANSCRIPT" 2>/dev/null | tail -1)
-LAST_PROMPT_UUID=$(printf '%s' "$LAST_LINE" | grep -o '"uuid":"[^"]*"' | head -1 | sed 's/"uuid":"//;s/"$//')
-if [ -z "$LAST_PROMPT_UUID" ]; then
+SESSION_ID=$(printf '%s' "$INPUT" | grep -o '"session_id":"[^"]*"' | head -1 | sed 's/"session_id":"//; s/"$//')
+if [ -z "$SESSION_ID" ]; then
   exit 0
 fi
 
-touch "/tmp/claude-skill-gate-$LAST_PROMPT_UUID"
+touch "/tmp/claude-skill-gate-$SESSION_ID"
 exit 0
 `;
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@spardutti/claude-skills",
-  "version": "1.28.0",
+  "version": "1.29.1",
   "description": "CLI to install Claude Code skills from the claude-skills collection",
   "type": "module",
   "bin": {