ocsmarttools 0.1.8 → 0.1.10

package/CHANGELOG.md

@@ -2,6 +2,25 @@
 
 All notable changes to `ocsmarttools` are documented here.
 
+## [0.1.10] - 2026-02-22
+
+### Changed
+- Expanded auto-routing trigger intents in managed `AGENTS.md` policy to cover:
+  - research/evidence
+  - reporting/writing
+  - coding/engineering
+  - data/extraction
+  - risk/compliance
+- Expanded built-in `tool_search` keyword coverage across runtime, filesystem, web, browser, and session tools for broader intent matching.
+- Improved `tool_search` scoring tokenization so punctuation-heavy terms (for example `CI/CD`) match more reliably.
+
+## [0.1.9] - 2026-02-22
+
+### Changed
+- Strengthened auto-managed routing policy so research/report prompts with sources/benchmarks must run retrieval tools before final output.
+- Added explicit anti-fabrication guidance in routing policy: do not output citations that were not retrieved in the current run.
+- Added failure behavior guidance: return clear retrieval gaps/partial result instead of fabricated evidence.
+
 ## [0.1.8] - 2026-02-22
 
 ### Changed

package/README.md

@@ -1,194 +1,139 @@
-# OCSmartTools for OpenClaw
+# OCSmartTools
 
-Provider-agnostic tool orchestration for lower latency and lower token cost.
+Provider-agnostic tool orchestration plugin for OpenClaw.
 
-`ocsmarttools` is designed to work immediately after install, with safe defaults and minimal setup.
+OCSmartTools helps reduce latency and token cost by routing multi-step work through deterministic tool flows (`tool_dispatch` / `tool_batch`) and shaping large outputs.
 
-## What It Adds
+## Highlights
 
-- `tool_search`: finds relevant tools quickly
-- `tool_dispatch`: runs one tool call through normal OpenClaw policy checks
-- `tool_batch`: runs bounded multi-step workflows (`call` + `foreach`)
-- `tool_result_get`: retrieves large stored outputs by handle
-
-Skills compatibility:
-- Works with OpenClaw skills because skills are instructions layered on top of tools.
-- For blocked websites, skill workflows that use `browser`/Playwright-style flows are compatible with `tool_dispatch` and `tool_batch`.
-
-## Why It Helps
-
-```mermaid
-flowchart LR
-  A["Traditional"] --> B["Model -> tool -> model -> tool"] --> C["More turns + repeated context"]
-  D["ocsmarttools"] --> E["Batch + shaped results"] --> F["Fewer turns + smaller context"]
-  C --> G["Higher cost/latency"]
-  F --> H["Lower cost/latency"]
-```
+- Works with existing OpenClaw installs (`openclaw >= 2026.2.19`)
+- Zero-touch startup (auto-bootstrap on install)
+- Tools added:
+  - `tool_search`
+  - `tool_dispatch`
+  - `tool_batch`
+  - `tool_result_get`
+- Built-in metrics for success/failure/timeout, latency, and estimated savings
+- Skill-compatible (including browser/Playwright-style workflows)
+- Auto-managed routing guidance for research/report/coding/data tasks (no per-prompt guardrails needed)
 
 ## Install
 
-Compatibility:
-- OpenClaw: `>=2026.2.19`
-- Works with existing installed instances (no core patch required)
-
-### npm
-
 ```bash
 openclaw plugins install ocsmarttools --pin
 openclaw plugins enable ocsmarttools
 openclaw gateway restart
 ```
 
-### Local folder
-
-```bash
-openclaw plugins install /absolute/path/to/OpenClaw-SmartToolCalls
-openclaw plugins enable ocsmarttools
-openclaw gateway restart
-```
-
 ## Quick Start
 
-1. Install + enable + restart.
-2. Done. The plugin auto-bootstraps and starts working in background.
-3. It also auto-manages an OCSmartTools routing block in `AGENTS.md` (unless disabled).
-4. Enable strict plugin-managed routing (optional):
-
 ```text
+/ocsmarttools status
 /ocsmarttools strict on
+/ocsmarttools stats
 ```
 
-5. Optional check:
+What these do:
+- `status`: shows current mode, strict-routing state, and key limits
+- `strict on`: enables strict plugin-managed routing guidance
+- `stats`: shows usage, failures, latency, and savings
+
+## Core Commands (Chat)
+
+- `/ocsmarttools help` -> simple command guide
+- `/ocsmarttools status` -> current state and limits
+- `/ocsmarttools strict <on|off|status>` -> strict routing control
+- `/ocsmarttools sync` -> re-apply managed routing block in `AGENTS.md`
+- `/ocsmarttools stats` -> metrics summary
+- `/ocsmarttools stats reset` -> reset metrics window
+- `/ocsmarttools setup [safe|standard]` -> apply recommended defaults
+- `/ocsmarttools mode <safe|standard>` -> change mode only
+- `/ocsmarttools config` -> show effective config
+- `/ocsmarttools config keys` -> list editable keys
+- `/ocsmarttools config set <key> <value>` -> update one key
+- `/ocsmarttools config reset [key]` -> reset one key or all
+- `/ocsmarttools version` -> installed plugin version
+
+## Core Commands (CLI)
+
+- `openclaw ocsmarttools help`
+- `openclaw ocsmarttools status`
+- `openclaw ocsmarttools strict <on|off>`
+- `openclaw ocsmarttools sync`
+- `openclaw ocsmarttools stats`
+- `openclaw ocsmarttools stats-reset`
+- `openclaw ocsmarttools setup [safe|standard]`
+- `openclaw ocsmarttools mode <safe|standard>`
+- `openclaw ocsmarttools config`
+- `openclaw ocsmarttools config-keys`
+- `openclaw ocsmarttools config-set <key> <value>`
+- `openclaw ocsmarttools config-reset [key]`
+- `openclaw ocsmarttools version`
+
+## Config
 
-```text
-/ocsmarttools status
-```
+Config path:
+- `plugins.entries.ocsmarttools.config`
 
-Model note:
-- No model setup is required in this plugin.
-- It automatically uses the model OpenClaw is already using.
-
-## Commands
-
-### Chat Commands
-
-| Command | What it does |
-|---|---|
-| `/ocsmarttools help` | Shows all commands in plain language |
-| `/ocsmarttools status` | Shows mode and safety settings |
-| `/ocsmarttools strict <on\|off\|status>` | Turns strict routing on/off or checks status |
-| `/ocsmarttools sync` | Rewrites the managed routing block in `AGENTS.md` |
-| `/ocsmarttools stats` | Shows calls, errors, timeouts, latency, and estimated savings |
-| `/ocsmarttools stats reset` | Resets plugin stats |
-| `/ocsmarttools setup [safe\|standard]` | Applies recommended defaults |
-| `/ocsmarttools mode <safe\|standard>` | Changes mode only |
-| `/ocsmarttools config` | Shows plugin config |
-| `/ocsmarttools config keys` | Lists editable config keys |
-| `/ocsmarttools config set <key> <value>` | Changes one config value |
-| `/ocsmarttools config reset [key]` | Resets one key or all keys |
-| `/ocsmarttools version` | Shows installed plugin version |
-
-### CLI Commands
-
-| Command | What it does |
-|---|---|
-| `openclaw ocsmarttools help` | Shows all commands in plain language |
-| `openclaw ocsmarttools status` | Shows mode and safety settings |
-| `openclaw ocsmarttools strict <on\|off>` | Turns strict routing on/off |
-| `openclaw ocsmarttools sync` | Rewrites the managed routing block in `AGENTS.md` |
-| `openclaw ocsmarttools stats` | Shows calls, errors, timeouts, latency, and estimated savings |
-| `openclaw ocsmarttools stats-reset` | Resets plugin stats |
-| `openclaw ocsmarttools setup [safe\|standard]` | Applies recommended defaults |
-| `openclaw ocsmarttools mode <safe\|standard>` | Changes mode only |
-| `openclaw ocsmarttools config` | Shows plugin config |
-| `openclaw ocsmarttools config-keys` | Lists editable config keys |
-| `openclaw ocsmarttools config-set <key> <value>` | Changes one config value |
-| `openclaw ocsmarttools config-reset [key]` | Resets one key or all keys |
-| `openclaw ocsmarttools version` | Shows installed plugin version |
-
-## Common Config Actions
+High-impact keys:
+- `strictRouting` (`true|false`)
+- `maxResultChars` (result shaping threshold)
+- `maxSteps`, `maxForEach` (batch limits)
+- `invokeTimeoutMs` (per-dispatch timeout)
+- `toolSearch.useLiveRegistry` (`true|false`)
+
+Example:
 
 ```text
-/ocsmarttools config
-/ocsmarttools config keys
-/ocsmarttools config set maxResultChars 120000
-/ocsmarttools config set storeLargeResults true
-/ocsmarttools config set toolSearch.useLiveRegistry true
-/ocsmarttools config set toolSearch.liveTimeoutMs 1500
 /ocsmarttools config set strictRouting true
-/ocsmarttools config set autoInjectRoutingGuide true
-/ocsmarttools config set autoInjectRoutingGuide false
-/ocsmarttools config reset maxResultChars
-/ocsmarttools stats
+/ocsmarttools config set maxResultChars 120000
+/ocsmarttools config set invokeTimeoutMs 30000
 ```
 
-Config path:
-- `plugins.entries.ocsmarttools.config`
-
 ## Modes
 
-- `standard` (default): zero-touch mode, no sandbox requirement, control-plane dispatch still blocked
-- `safe`: requires sandboxed execution and blocks control-plane dispatch (`gateway`, `cron`)
-- `strictRouting=true`: strict plugin-managed routing policy is auto-synced with no manual workspace edits.
+- `standard` (default): balanced, zero-touch
+- `safe`: requires sandbox and blocks control-plane dispatch
+- `strictRouting=true`: enforces plugin-managed routing guidance and keeps routing block synced
 
-Setup default:
-- `/ocsmarttools setup` and `openclaw ocsmarttools setup` default to `standard`.
+## Auto Trigger Intents
 
-## Recommended Defaults
+When tools are needed, routing is auto-biased toward `tool_dispatch` / `tool_batch` for these intent families:
+- research/evidence (sources, benchmarks, case studies, latest/news)
+- reporting/writing (reports, briefs, proposals, plans, emails, docs/specs)
+- coding/engineering (build, debug, fix, test, refactor, deploy, migration)
+- data/extraction (scrape/fetch/parse/aggregate/rank/CSV/JSON/ETL)
+- risk/compliance (security/privacy/legal/audit requests)
 
-For mixed daily usage (research + writing + coding):
+## Skills and Blocked Sites
 
-- `maxResultChars`: `120000`
-- `storeLargeResults`: `true`
-- `resultStoreTtlSec`: `3600`
-- `resultSampleItems`: `10`
+Skills in OpenClaw are instruction layers, not separate execution engines. OCSmartTools can still dispatch/batch the underlying tools used by skills.
 
-Example:
+For websites that block plain fetch, use browser-based flows via installed skills/plugins; OCSmartTools can orchestrate those tool calls with the same metrics and shaping behavior.
 
-```text
-/ocsmarttools config set maxResultChars 120000
-/ocsmarttools config set resultStoreTtlSec 3600
-/ocsmarttools config set resultSampleItems 10
-```
-
-## Strict Policy Note
+## Safety Notes
 
-If your instance uses strict `tools.allow`, include:
-
-```json5
-{
-  tools: {
-    allow: ["tool_search", "tool_dispatch", "tool_batch", "tool_result_get"]
-  }
-}
-```
+- OCSmartTools does not bypass OpenClaw policy.
+- Control-plane tools (`gateway`, `cron`) are blocked by default in dispatch/batch.
+- Large result handles are in-memory and expire by TTL.
 
-## Safety and Limits
+## Troubleshooting
 
-- `ocsmarttools` does not bypass OpenClaw tool policy.
-- Routing policy is auto-injected into `AGENTS.md` with managed markers; it can be re-synced via `/ocsmarttools sync`.
-- `strictRouting=true` forces `autoInjectRoutingGuide=true`.
-- `tool_batch` is intentionally bounded (`maxSteps`, `maxForEach`).
-- Large-result handles are in-memory and expire by TTL.
-- `tool_result_get` works only while handle is still valid.
-- `tool_search` tries live registry endpoints first, then automatically falls back to policy/static catalog.
+- If plugin commands do not appear, restart gateway:
+  - `openclaw gateway restart`
+- If strict policy blocks tools, verify `tools.allow`/`tools.deny` in OpenClaw config.
+- If discovery seems stale, run:
+  - `/ocsmarttools sync`
+  - `/ocsmarttools stats`
 
 ## Development
 
 ```bash
 npm install
 npm run typecheck
-```
-
-## Versioning
-
-```bash
-npm run version:show
-npm run version:bump:patch
-npm run version:bump:minor
-npm run version:bump:major
 npm run release:check
 ```
 
-- Version source of truth: `package.json`.
-- Release notes: `CHANGELOG.md`.
+Version and release notes:
+- Version source of truth: `package.json`
+- Release notes: `CHANGELOG.md`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ocsmarttools",
-  "version": "0.1.8",
+  "version": "0.1.10",
   "description": "Provider-agnostic advanced tool orchestration plugin for OpenClaw with search, dispatch, and batching",
   "type": "module",
   "scripts": {

package/src/lib/routing-guide.ts

@@ -6,6 +6,15 @@ import type { OpenClawConfig, OpenClawPluginApi } from "openclaw/plugin-sdk";
 const BLOCK_START = "<!-- OCSMARTTOOLS_ROUTING_START -->";
 const BLOCK_END = "<!-- OCSMARTTOOLS_ROUTING_END -->";
 
+const AUTO_TRIGGER_LINES = [
+  "Auto-route trigger intents (use routing tools by default when tools are needed):",
+  "- Research and evidence: research, investigate, analyze, benchmark, compare, evaluate, market intelligence, trends, latest, news, sources, citations, references, case study, due diligence.",
+  "- Reporting and writing: report, summary, brief, memo, presentation, plan, proposal, recommendation, email, RFC, PRD, specification, documentation, SOP.",
+  "- Coding and engineering: code, implement, build, website, app, script, debug, fix, bug, stack trace, test, lint, refactor, optimize, compile, deploy, CI/CD, migration, release notes.",
+  "- Data and extraction: scrape, crawl, fetch, parse, transform, aggregate, rank, classify, table, CSV, JSON, dataset, ETL.",
+  "- Risk and compliance: risk register, security review, privacy, legal, policy, SOC 2, GDPR, HIPAA, audit.",
+];
+
 function buildRoutingBlock(strictRouting: boolean): string {
   const body = strictRouting
     ? [
@@ -16,11 +25,17 @@ function buildRoutingBlock(strictRouting: boolean): string {
         "Mandatory rules in strict mode:",
         "1. Route tool usage through `tool_dispatch` or `tool_batch`.",
         "2. Use `tool_batch` for 2+ related calls, iterative steps, or map-style tasks.",
-        "3. Use `tool_search` only when tool choice is unclear.",
-        "4. Use `tool_result_get` only when a stored handle needs more detail.",
-        "5. Avoid direct native tool calls unless routing tools are unavailable due to runtime/tool-policy constraints.",
+        "3. For research/report/benchmark/source requests, always run live tool retrieval before final output (typically `web_search` + `web_fetch`).",
+        "4. Use `tool_search` only when tool choice is unclear.",
+        "5. Use `tool_result_get` only when a stored handle needs more detail.",
+        "6. Avoid direct native tool calls unless routing tools are unavailable due to runtime/tool-policy constraints.",
+        "7. Do not present citations/sources unless they came from successful tool calls in the current run.",
+        "8. If required retrieval fails, return a clear partial/failure report instead of fabricated evidence.",
+        "9. Treat trigger intents below as routing-required unless the request is clearly a no-tool task.",
         "",
         "Common large/noisy tools: `web_fetch`, `read` (large files), `exec`, `process`, `browser`, `nodes`.",
+        "",
+        ...AUTO_TRIGGER_LINES,
       ].join("\n")
     : [
         "## OCSmartTools Routing Policy (Auto-Managed)",
@@ -29,11 +44,17 @@ function buildRoutingBlock(strictRouting: boolean): string {
         "",
         "1. If tool usage is needed and result size is uncertain, use `tool_dispatch`.",
         "2. If the task needs 2+ related tool calls, use `tool_batch`.",
-        "3. Use `tool_search` only when tool choice is unclear.",
-        "4. Prefer compact/tool-shaped outputs; use `tool_result_get` only when more detail is required.",
-        "5. Use direct native tool calls only for simple one-shot small-output actions.",
+        "3. For research/report requests with sources or benchmarks, run retrieval tools first before final output.",
+        "4. Use `tool_search` only when tool choice is unclear.",
+        "5. Prefer compact/tool-shaped outputs; use `tool_result_get` only when more detail is required.",
+        "6. Use direct native tool calls only for simple one-shot small-output actions.",
+        "7. Do not include citations/sources that were not retrieved by tools in the current run.",
+        "8. If retrieval fails, explicitly report the gap.",
+        "9. Treat trigger intents below as strong routing signals unless the task is clearly no-tool.",
         "",
         "Common large/noisy tools: `web_fetch`, `read` (large files), `exec`, `process`, `browser`, `nodes`.",
+        "",
+        ...AUTO_TRIGGER_LINES,
       ].join("\n");
 
   return `${BLOCK_START}

package/src/lib/tool-catalog.ts

@@ -9,34 +9,184 @@ export type ToolEntry = {
   source: "builtin" | "policy" | "live";
 };
 
+const RESEARCH_KEYWORDS = [
+  "research",
+  "investigate",
+  "analyze",
+  "analysis",
+  "benchmark",
+  "compare",
+  "evaluate",
+  "market intelligence",
+  "trend",
+  "latest",
+  "news",
+  "evidence",
+  "sources",
+  "citations",
+  "references",
+  "case study",
+  "due diligence",
+];
+
+const REPORT_KEYWORDS = [
+  "report",
+  "summary",
+  "brief",
+  "memo",
+  "recommendation",
+  "proposal",
+  "plan",
+  "writeup",
+  "presentation",
+  "email",
+  "documentation",
+  "spec",
+  "rfc",
+  "prd",
+];
+
+const CODING_KEYWORDS = [
+  "code",
+  "implement",
+  "build",
+  "website",
+  "app",
+  "script",
+  "debug",
+  "fix",
+  "bug",
+  "stack trace",
+  "test",
+  "lint",
+  "refactor",
+  "optimize",
+  "compile",
+  "deploy",
+  "ci",
+  "cd",
+  "migration",
+];
+
+const DATA_KEYWORDS = [
+  "extract",
+  "scrape",
+  "crawl",
+  "parse",
+  "transform",
+  "aggregate",
+  "rank",
+  "classify",
+  "csv",
+  "json",
+  "dataset",
+  "etl",
+  "table",
+];
+
+const BROWSER_KEYWORDS = [
+  "playwright",
+  "stealth",
+  "anti-bot",
+  "captcha",
+  "dynamic site",
+  "javascript page",
+  "login wall",
+  "rendered page",
+  "automation",
+];
+
 const BUILTIN: ToolEntry[] = [
   {
     name: "exec",
     group: "runtime",
     description: "Run shell commands.",
     paramsHint: "{ command }",
-    keywords: ["script", "python", "node", "playwright", "stealth", "scrape"],
+    keywords: [
+      "shell",
+      "terminal",
+      "command line",
+      "python",
+      "node",
+      "npm",
+      "git",
+      "build",
+      "run tests",
+      "start server",
+      ...CODING_KEYWORDS,
+      ...DATA_KEYWORDS,
+      ...BROWSER_KEYWORDS,
+    ],
+    source: "builtin",
+  },
+  {
+    name: "bash",
+    group: "runtime",
+    description: "Alias for exec in many contexts.",
+    paramsHint: "{ command }",
+    keywords: ["shell", "terminal", "script", ...CODING_KEYWORDS],
+    source: "builtin",
+  },
+  {
+    name: "process",
+    group: "runtime",
+    description: "Manage background command sessions.",
+    paramsHint: "{ action, sessionId? }",
+    keywords: ["long running", "background job", "watch logs", "daemon", "worker", "server process", "job status"],
+    source: "builtin",
+  },
+  {
+    name: "read",
+    group: "fs",
+    description: "Read files.",
+    paramsHint: "{ path }",
+    keywords: ["open file", "inspect file", "view logs", "read config", "read docs", "audit code", ...REPORT_KEYWORDS],
+    source: "builtin",
+  },
+  {
+    name: "write",
+    group: "fs",
+    description: "Write files.",
+    paramsHint: "{ path, content }",
+    keywords: ["create file", "save output", "generate report", "draft", "write docs", "export", ...REPORT_KEYWORDS],
+    source: "builtin",
+  },
+  {
+    name: "edit",
+    group: "fs",
+    description: "Patch file text by replacement.",
+    paramsHint: "{ path, oldText, newText }",
+    keywords: ["modify file", "update code", "refactor file", "fix snippet", ...CODING_KEYWORDS],
+    source: "builtin",
+  },
+  {
+    name: "apply_patch",
+    group: "fs",
+    description: "Apply unified-style file patch.",
+    paramsHint: "{ input }",
+    keywords: ["multi-file patch", "diff", "code patch", "batch edits", "surgical edit", ...CODING_KEYWORDS],
+    source: "builtin",
+  },
+  { name: "sessions_list", group: "sessions", description: "List sessions.", paramsHint: "{ ... }", keywords: ["session", "agents", "workers"], source: "builtin" },
+  { name: "sessions_history", group: "sessions", description: "Read session history.", paramsHint: "{ sessionKey? }", keywords: ["conversation history", "chat log", "audit trail"], source: "builtin" },
+  { name: "sessions_send", group: "sessions", description: "Send to another session.", paramsHint: "{ ... }", keywords: ["delegate", "handoff", "subtask"], source: "builtin" },
+  { name: "sessions_spawn", group: "sessions", description: "Spawn subagent session.", paramsHint: "{ prompt, ... }", keywords: ["spawn agent", "parallel agent", "subagent", "delegate"], source: "builtin" },
+  { name: "session_status", group: "sessions", description: "Show current session status.", paramsHint: "{}", keywords: ["status", "health", "progress"], source: "builtin" },
+  {
+    name: "memory_search",
+    group: "memory",
+    description: "Search memory snippets.",
+    paramsHint: "{ query }",
+    keywords: ["recall", "memory", "previous notes", "past decisions", "context lookup"],
     source: "builtin",
   },
-  { name: "bash", group: "runtime", description: "Alias for exec in many contexts.", paramsHint: "{ command }", source: "builtin" },
-  { name: "process", group: "runtime", description: "Manage background command sessions.", paramsHint: "{ action, sessionId? }", source: "builtin" },
-  { name: "read", group: "fs", description: "Read files.", paramsHint: "{ path }", source: "builtin" },
-  { name: "write", group: "fs", description: "Write files.", paramsHint: "{ path, content }", source: "builtin" },
-  { name: "edit", group: "fs", description: "Patch file text by replacement.", paramsHint: "{ path, oldText, newText }", source: "builtin" },
-  { name: "apply_patch", group: "fs", description: "Apply unified-style file patch.", paramsHint: "{ input }", source: "builtin" },
-  { name: "sessions_list", group: "sessions", description: "List sessions.", paramsHint: "{ ... }", source: "builtin" },
-  { name: "sessions_history", group: "sessions", description: "Read session history.", paramsHint: "{ sessionKey? }", source: "builtin" },
-  { name: "sessions_send", group: "sessions", description: "Send to another session.", paramsHint: "{ ... }", source: "builtin" },
-  { name: "sessions_spawn", group: "sessions", description: "Spawn subagent session.", paramsHint: "{ prompt, ... }", source: "builtin" },
-  { name: "session_status", group: "sessions", description: "Show current session status.", paramsHint: "{}", source: "builtin" },
-  { name: "memory_search", group: "memory", description: "Search memory snippets.", paramsHint: "{ query }", source: "builtin" },
-  { name: "memory_get", group: "memory", description: "Read memory entries.", paramsHint: "{ key }", source: "builtin" },
+  { name: "memory_get", group: "memory", description: "Read memory entries.", paramsHint: "{ key }", keywords: ["memory key", "stored context", "knowledge"], source: "builtin" },
   {
     name: "web_search",
     group: "web",
     description: "Search the web.",
     paramsHint: "{ query }",
-    keywords: ["research", "discover", "find"],
+    keywords: ["discover", "find", "lookup", ...RESEARCH_KEYWORDS, ...REPORT_KEYWORDS],
     source: "builtin",
   },
   {
@@ -44,7 +194,7 @@ const BUILTIN: ToolEntry[] = [
     group: "web",
     description: "Fetch and extract web page content.",
     paramsHint: "{ url }",
-    keywords: ["read page", "extract", "html"],
+    keywords: ["read page", "fetch url", "extract", "html", "article text", "content extraction", ...DATA_KEYWORDS, ...RESEARCH_KEYWORDS],
     source: "builtin",
   },
   {
@@ -52,14 +202,14 @@ const BUILTIN: ToolEntry[] = [
     group: "ui",
     description: "Browser automation.",
     paramsHint: "{ action, ... }",
-    keywords: ["playwright", "stealth", "anti-bot", "captcha", "dynamic site", "javascript page"],
+    keywords: [...BROWSER_KEYWORDS, "web navigation", "click", "form fill", "screenshot", "scrape dynamic"],
     source: "builtin",
   },
-  { name: "canvas", group: "ui", description: "Canvas/artifact rendering actions.", paramsHint: "{ action, ... }", source: "builtin" },
-  { name: "cron", group: "automation", description: "Manage scheduled jobs.", paramsHint: "{ action, ... }", source: "builtin" },
-  { name: "gateway", group: "automation", description: "Gateway control-plane actions.", paramsHint: "{ action, ... }", source: "builtin" },
-  { name: "message", group: "messaging", description: "Message actions across channels.", paramsHint: "{ action, ... }", source: "builtin" },
-  { name: "nodes", group: "nodes", description: "Invoke node-side capabilities.", paramsHint: "{ action, ... }", source: "builtin" },
+  { name: "canvas", group: "ui", description: "Canvas/artifact rendering actions.", paramsHint: "{ action, ... }", keywords: ["diagram", "visual", "chart", "artifact"], source: "builtin" },
+  { name: "cron", group: "automation", description: "Manage scheduled jobs.", paramsHint: "{ action, ... }", keywords: ["schedule", "recurring", "automation", "job timer"], source: "builtin" },
+  { name: "gateway", group: "automation", description: "Gateway control-plane actions.", paramsHint: "{ action, ... }", keywords: ["gateway control", "restart", "config patch", "service"], source: "builtin" },
+  { name: "message", group: "messaging", description: "Message actions across channels.", paramsHint: "{ action, ... }", keywords: ["notify", "chat", "post update", "send message"], source: "builtin" },
+  { name: "nodes", group: "nodes", description: "Invoke node-side capabilities.", paramsHint: "{ action, ... }", keywords: ["device", "camera", "screen", "location", "node action"], source: "builtin" },
 ];
 
 const GROUP_NAMES = new Set([
@@ -328,6 +478,10 @@ function score(query: string, entry: ToolEntry): number {
   const d = entry.description.toLowerCase();
   const g = entry.group.toLowerCase();
   const keywords = (entry.keywords ?? []).map((v) => v.toLowerCase());
+  const tokens = q
+    .split(/[^a-z0-9]+/)
+    .map((t) => t.trim())
+    .filter(Boolean);
 
   let s = 0;
   if (n === q) s += 100;
@@ -337,7 +491,7 @@ function score(query: string, entry: ToolEntry): number {
   if (d.includes(q)) s += 10;
   if (keywords.includes(q)) s += 25;
 
-  for (const token of q.split(/\s+/).filter(Boolean)) {
+  for (const token of tokens) {
     if (n.includes(token)) s += 12;
     if (d.includes(token)) s += 4;
     if (g.includes(token)) s += 4;