@softspark/ai-toolkit 4.32.2 → 4.32.3

package/CHANGELOG.md

@@ -7,6 +7,26 @@ Versioning follows [Semantic Versioning](https://semver.org/).
 
 ---
 
+## v4.32.3 — Search-first stops firing on things nobody asked (2026-09-04)
+
+### Fixed
+
+- **A background task finishing no longer demands a KB search.**
+  `user-prompt-submit.sh` set the search-required flag on any stdin longer
+  than 30 characters. The harness delivers background-task notifications, CI
+  events and replayed slash-command output on the same channel as a prompt, so
+  a task completing blocked `Stop` and asked for a search over text nobody
+  wrote. Harness event envelopes no longer set the flag.
+- **A pasted credential no longer becomes a search query.** A 43-character API
+  key on its own line cleared the length gate, and the block message names the
+  prompt as the thing to search for — so the enforcement path asked for a live
+  secret to be sent to a retrieval service. Single-token prompts no longer set
+  the flag; a genuine one-word prompt is under the length gate anyway, so
+  nothing that was enforced before stops being enforced. Prose that merely
+  mentions a task notification still sets it, and that case is now a test.
+
+---
+
 ## v4.32.2 - Plugin exports land where you run the command (2026-09-04)
 
 ### Fixed

package/README.md

@@ -6,25 +6,30 @@
 [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)
 [![Skills](https://img.shields.io/badge/skills-114-brightgreen)](app/skills/)
 [![Agents](https://img.shields.io/badge/agents-44-blue)](app/agents/)
-[![Tests](https://img.shields.io/badge/tests-1974%20passing-success)](tests/)
-
-## What's New in v4.32.2
-
-**v4.32.2** makes plugin exports land where you run the command:
-
-- `ai-toolkit claude-app export`, `codex-plugin export`, and
-  `antigravity-plugin export` write their ZIP (and the sibling
-  `*-global-instructions.md`) to the directory you typed the command in. They
-  used to appear inside `node_modules/@softspark/ai-toolkit` because the CLI
-  runs them with its own working directory; the wrapper now hands the shell's
-  directory to the exporters. Absolute `--output` paths are unchanged.
-- Root-level `*.zip` and `*-global-instructions.md` are gitignored so an
-  export run from the repository is never committed.
-- Still in this train: `plugin remove` leaves nothing behind, every skill
-  script answers `--help` (v4.32.1); path-scoped common rules, project-scoped
-  language skills, `doctor` context-budget and permission checks, one strict
-  frontmatter parser, `git-team` for `--profile strict` (v4.32.0). Test
-  count: 1972 -> 1974 bats + 348 pytest.
+[![Tests](https://img.shields.io/badge/tests-1978%20passing-success)](tests/)
+
+## What's New in v4.32.3
+
+**v4.32.3** stops search-first enforcement firing on things nobody asked:
+
+- **A background task finishing no longer blocks `Stop`.** The search-required
+  flag was set on any stdin over 30 characters, and the harness delivers
+  background-task notifications, CI events and replayed slash-command output
+  on the same channel as a prompt. Event envelopes no longer set it.
+- **A pasted credential no longer becomes a search query.** A 43-character API
+  key on its own line cleared the length gate, and the block message names the
+  prompt as the thing to search for — so the enforcement path asked for a live
+  secret to be sent to a retrieval service. Single-token prompts no longer set
+  the flag.
+- Nothing that was enforced before stops being enforced: a genuine one-word
+  prompt is under the length gate anyway, and prose that merely mentions a task
+  notification still sets the flag. That case is now a test.
+- Still in this train: exports land in the directory you ran the command in
+  (v4.32.2); `plugin remove` leaves nothing behind, every skill script answers
+  `--help` (v4.32.1); path-scoped common rules, project-scoped language skills,
+  `doctor` context-budget and permission checks, one strict frontmatter parser,
+  `git-team` for `--profile strict` (v4.32.0). Test count: 1974 -> 1978 bats +
+  348 pytest.
 
 See [CHANGELOG.md](CHANGELOG.md) for full history.
 

package/app/.claude-plugin/plugin.json

@@ -3,7 +3,7 @@
   "name": "ai-toolkit",
   "displayName": "AI Toolkit",
   "description": "Professional-grade engineering skills, agents, rules, and lifecycle guardrails for Claude Code, Claude Chat, and Cowork.",
-  "version": "4.32.2",
+  "version": "4.32.3",
   "author": {
     "name": "SoftSpark",
     "url": "https://github.com/softspark"

package/app/hooks/user-prompt-submit.sh

@@ -29,7 +29,28 @@ PROMPT_LEN=${#PROMPT_TEXT}
 STATE_DIR="$HOME/.softspark/ai-toolkit/state"
 FLAG="$STATE_DIR/search-required-$(hook_session_id).flag"
 mkdir -p "$STATE_DIR" 2>/dev/null
+
+# The flag means "somebody asked a question the KB might answer". Length alone
+# does not establish that. The harness delivers background-task notifications,
+# CI events and replayed slash-command output on the same channel as a prompt,
+# and a 43-character API key pasted on its own line clears the length gate too.
+# Both were observed in one session: the first demanded a KB search for text
+# nobody asked about, the second would have sent a live credential to a
+# retrieval service as the query. Neither is a prompt.
+PROMPT_HEAD="${PROMPT_TEXT#"${PROMPT_TEXT%%[![:space:]]*}"}"
+PROMPT_IS_QUESTION=1
+case "$PROMPT_HEAD" in
+    '<task-notification'*|'<ci-monitor-event'*|'<system-reminder'*|\
+    '<local-command-'*|'<command-name'*|'<command-message'*) PROMPT_IS_QUESTION=0 ;;
+esac
+# A single opaque token — API key, hash, URL, path — carries no question. A
+# genuine one-word prompt is under the length gate anyway, so nothing is lost.
+if [ "$(printf '%s' "$PROMPT_TEXT" | wc -w | tr -d ' ')" -le 1 ]; then
+    PROMPT_IS_QUESTION=0
+fi
+
 if [ "$PROMPT_LEN" -gt 30 ] && \
+   [ "$PROMPT_IS_QUESTION" -eq 1 ] && \
    [ "${CLAUDE_SKIP_SEARCH_FIRST:-0}" != "1" ] && \
    ai_toolkit_has_search_provider; then
     { printf '%s\n%s\n' "$(date -u +%s)" "$PROMPT_TEXT" > "$FLAG"; } 2>/dev/null

package/app/surface.json

@@ -16,6 +16,7 @@
     "chaos",
     "ci",
     "ci-cd-patterns",
+    "claude-toolkit-rules",
     "clean-code",
     "command-creator",
     "commit",
@@ -36,12 +37,14 @@
     "docs",
     "documentation-standards",
     "ecommerce-patterns",
+    "edit-discipline",
     "evaluate",
     "evolve",
     "explain",
     "explore",
     "fix",
     "flutter-patterns",
+    "git-conventions",
     "git-mastery",
     "golang-rules",
     "grill-me",
@@ -68,6 +71,7 @@
     "observability-patterns",
     "onboard",
     "orchestrate",
+    "output-mode",
     "panic",
     "performance-profiling",
     "persona",
@@ -81,6 +85,7 @@
     "prompt-caching-patterns",
     "python-rules",
     "qa-session",
+    "quality-gates",
     "rag-patterns",
     "refactor",
     "refactor-plan",

package/manifest.json

@@ -1,5 +1,5 @@
 {
-  "version": "4.32.2",
+  "version": "4.32.3",
   "components": {
     "agents": {
       "description": "44 specialized agents (orchestrator, backend, frontend, security, devops, etc.)",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@softspark/ai-toolkit",
-  "version": "4.32.2",
+  "version": "4.32.3",
   "description": "AI coding toolkit: 109 skills, 44 agents, 14 developer-tool integrations, recoverable native tool-output filtering, Claude Chat/Cowork export, safety constitution, SARIF audit, and signed npm provenance.",
   "keywords": [
     "claude",