token-pilot 0.24.2 → 0.26.6

package/dist/server/tool-profiles.js

@@ -0,0 +1,81 @@
+/**
+ * v0.26.3 — tool profiles.
+ *
+ * Idea lifted honestly from Token Savior's TOKEN_SAVIOR_PROFILE. When an
+ * MCP server advertises 22 tools, every tools/list response costs the
+ * agent ~3 k tokens before it does anything. Most sessions don't need
+ * every tool — a code-review agent uses smart_read + find_usages +
+ * outline and nothing else. A profile lets the user ship a narrower
+ * tools/list while keeping the handlers live (so a subagent or another
+ * user in the same server can still reach the full set if they know
+ * the name).
+ *
+ * Three profiles:
+ *   - full  (default): everything, same as pre-v0.26.3.
+ *   - nav  : read-only exploration. smart_read, outline, find_usages,
+ *            read_symbol, project_overview, module_info, related_files,
+ *            explore_area, smart_log, smart_diff.
+ *   - edit : nav + batch reads + everything Edit needs to hit a symbol
+ *            precisely. Adds read_symbols, read_range, read_section,
+ *            read_diff, read_for_edit, smart_read_many.
+ *
+ * Selection: TOKEN_PILOT_PROFILE=nav|edit|full env var. Unknown values
+ * fall back to full with a stderr warning. Silent on missing env.
+ */
+export const PROFILE_NAMES = [
+    "full",
+    "nav",
+    "edit",
+];
+/** Minimum nav profile — exploration only, no editing support. */
+export const NAV_TOOLS = new Set([
+    "smart_read",
+    "read_symbol",
+    "outline",
+    "find_usages",
+    "project_overview",
+    "module_info",
+    "related_files",
+    "explore_area",
+    "smart_log",
+    "smart_diff",
+]);
+/** Edit profile adds batch reads + edit-preparation tools. */
+export const EDIT_EXTRAS = new Set([
+    "read_symbols",
+    "read_range",
+    "read_section",
+    "read_diff",
+    "read_for_edit",
+    "smart_read_many",
+]);
+/**
+ * Decide which tools the LLM sees in tools/list given a profile.
+ * Pure — safe to unit-test without spinning up the server.
+ *
+ * Tool names NOT matched by any profile rule (e.g. future additions)
+ * fall into 'full' only, to stay conservative by default.
+ */
+export function filterToolsByProfile(tools, profile) {
+    if (profile === "full")
+        return [...tools];
+    if (profile === "nav")
+        return tools.filter((t) => NAV_TOOLS.has(t.name));
+    // edit = nav + extras
+    return tools.filter((t) => NAV_TOOLS.has(t.name) || EDIT_EXTRAS.has(t.name));
+}
+/**
+ * Parse the TOKEN_PILOT_PROFILE env value. Unknown values get a warning
+ * and fall back to full — we never silently apply a guess.
+ */
+export function parseProfileEnv(envValue, warn = () => { }) {
+    if (!envValue)
+        return "full";
+    const lower = envValue.trim().toLowerCase();
+    if (lower === "full" || lower === "nav" || lower === "edit") {
+        return lower;
+    }
+    warn(`[token-pilot] Unknown TOKEN_PILOT_PROFILE="${envValue}". Expected full|nav|edit. Falling back to full.`);
+    return "full";
+}
+//# sourceMappingURL=tool-profiles.js.map

package/dist/server.js

@@ -44,7 +44,9 @@ import { handleReadSection } from "./handlers/read-section.js";
 import { detectContextMode } from "./integration/context-mode-detector.js";
 import { estimateTokens } from "./core/token-estimator.js";
 import { checkPolicy, isFullReadTool } from "./core/policy-engine.js";
+import { appendToolCall } from "./core/tool-call-log.js";
 import { MCP_INSTRUCTIONS, TOOL_DEFINITIONS, } from "./server/tool-definitions.js";
+import { filterToolsByProfile, parseProfileEnv, } from "./server/tool-profiles.js";
 import { createTokenEstimates } from "./server/token-estimates.js";
 import { validateSmartReadArgs, validateReadSymbolArgs, validateReadSymbolsArgs, validateReadRangeArgs, validateReadDiffArgs, validateFindUsagesArgs, validateSmartReadManyArgs, validateReadForEditArgs, validateRelatedFilesArgs, validateOutlineArgs, validateFindUnusedArgs, validateCodeAuditArgs, validateProjectOverviewArgs, validateModuleInfoArgs, validateSmartDiffArgs, validateExploreAreaArgs, validateSmartLogArgs, validateTestSummaryArgs, validateReadSectionArgs, } from "./core/validation.js";
 export async function createServer(projectRoot, options) {
@@ -240,14 +242,33 @@ export async function createServer(projectRoot, options) {
         capabilities: { tools: {} },
         instructions: MCP_INSTRUCTIONS,
     });
+    // v0.26.3 — tool profiles. TOKEN_PILOT_PROFILE=nav|edit|full (default
+    // full) trims the advertised tools/list payload. Handlers stay live,
+    // so a subagent that explicitly names a filtered-out tool still gets
+    // a response — we just don't brag about every tool upfront.
+    const activeProfile = parseProfileEnv(process.env.TOKEN_PILOT_PROFILE, (m) => process.stderr.write(m + "\n"));
+    const advertisedTools = filterToolsByProfile(TOOL_DEFINITIONS, activeProfile);
+    if (activeProfile !== "full") {
+        process.stderr.write(`[token-pilot] Profile: ${activeProfile} — advertising ${advertisedTools.length}/${TOOL_DEFINITIONS.length} tools. Unset TOKEN_PILOT_PROFILE for the full set.\n`);
+    }
     server.setRequestHandler(ListToolsRequestSchema, () => ({
-        tools: TOOL_DEFINITIONS,
+        tools: advertisedTools,
     }));
     // Token estimation functions (extracted to server/token-estimates.ts)
     const { fullFileTokens, estimateProjectOverviewWorkflowTokens, estimateOutlineWorkflowTokens, estimateRelatedFilesWorkflowTokens, estimateFindUsagesWorkflowTokens, estimateExploreAreaWorkflowTokens, detectSavingsCategory, } = createTokenEstimates(() => projectRoot, fileCache);
     /** Record analytics with intent classification and decision trace. Returns policy advisory if any. */
     function recordWithTrace(call) {
         const { absPath, args, recentlyEdited, ...rest } = call;
+        // v0.26.1 — honest accounting. When a handler signals 'none' as
+        // the savings category (e.g. smart_read small-file pass-through),
+        // we weren't compressing anything — the caller got the file back
+        // verbatim plus a tiny header. Claiming wouldBe = fullFile here
+        // produced the -2% "negative savings" line on Opus 4.7's
+        // session_analytics. Zero the delta: 0% savings claimed, no ghost
+        // overhead.
+        if (rest.savingsCategory === "none") {
+            rest.tokensWouldBe = rest.tokensReturned;
+        }
         analytics.record({
             ...rest,
             intent: classifyIntent(rest.tool),
@@ -279,6 +300,22 @@ export async function createServer(projectRoot, options) {
             totalCallCount,
             totalTokensReturned,
         });
+        // v0.26.2 — persist for cumulative tool-audit. Fire-and-forget;
+        // disk failures must not block the tool-response path. The audit
+        // CLI reads all archives + current to build a per-tool savings
+        // distribution across sessions, which is the foundation for any
+        // future prune/fix decision.
+        void appendToolCall(projectRoot, {
+            ts: rest.timestamp,
+            session_id: call.sessionId ?? "",
+            tool: rest.tool,
+            path: rest.path,
+            tokensReturned: rest.tokensReturned,
+            tokensWouldBe: rest.tokensWouldBe,
+            savingsCategory: rest.savingsCategory ?? "compression",
+            sessionCacheHit: rest.sessionCacheHit,
+            delegatedToContextMode: rest.delegatedToContextMode,
+        });
         return advisory ? `\n${advisory.message}` : null;
     }
     // Handle tool calls with validated arguments

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "token-pilot",
-  "version": "0.24.2",
+  "version": "0.26.6",
   "description": "Save up to 80% tokens when AI reads code — MCP server for token-efficient code navigation, AST-aware structural reading instead of dumping full files into context window",
   "type": "module",
   "main": "dist/index.js",

package/.mcp.json

@@ -1,8 +0,0 @@
-{
-  "mcpServers": {
-    "token-pilot": {
-      "command": "sh",
-      "args": ["${CLAUDE_PLUGIN_ROOT}/start.sh"]
-    }
-  }
-}