agent-tempo 1.7.0-beta.3 → 1.7.0-beta.5

package/dashboard/package.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-tempo-dashboard",
   "private": true,
-  "version": "1.7.0-beta.3",
+  "version": "1.7.0-beta.5",
   "type": "module",
   "description": "Web dashboard for agent-tempo. Bundled into the npm package; served by the daemon at /dashboard/*.",
   "scripts": {

package/dist/pi/headless.d.ts

@@ -38,42 +38,37 @@ export interface RunHeadlessPiOptions {
     continueSessionId?: string;
 }
 /**
- * Build the `DefaultResourceLoader` options for a headless Pi player.
+ * #715 — compute the registration-level `excludeTools` denylist for
+ * `createAgentSession`. Excluded tools are never registered → ABSENT from the
+ * model's toolset AND system prompt: the LLM cannot request what it never sees.
  *
- * SECURITY — S2 (MD-C deny-list soundness). The `restricted` tool gate is a
- * DENY-LIST over shell/exec tool *names* (tool-capability.ts EXEC_TOOLS, via
- * `classify(name) === 'exec'` — F1 replaced extension.ts's former local set). That
- * guarantee — "restricted = no host execution" — holds ONLY IF no third-party
- * extension can register an un-blacklisted execution tool (e.g. a custom
- * `python` / `npm` / `run` tool). It therefore depends on a hard structural
- * fact: which extensions Pi loads.
+ * This is a registration-level FLOOR beneath the call-time MD-C handler + #712
+ * gate. It is tamper-RESISTANT, NOT tamper-PROOF: it defends a PROMPT-INJECTED
+ * agent (and holds even if the call-time gate had a bug — the tool simply isn't
+ * there), but it does NOT defend against PROCESS COMPROMISE — a tampered /
+ * modified extension can re-register or un-exclude tools (this is OUR code
+ * passing a denylist; an attacker who modifies the code/process bypasses it).
+ * That residual is OS-sandbox + supply-chain integrity, tracked as #724.
  *
- * Verified against the installed Pi SDK 0.78 source (NOT assumed):
- *   - `DefaultResourceLoader.reload()` (resource-loader.js:271-276) builds
- *     `extensionPaths = noExtensions ? cliEnabledExtensions
- *                                     : merge(cliEnabledExtensions, enabledExtensions)`
- *     where `enabledExtensions` (line 229) are the DISK/package extensions from
- *     `packageManager.resolve()` (`~/.pi/agent/extensions/`, `<cwd>/.pi/extensions/`,
- *     installed packages). `loadExtensions(extensionPaths)` then loads them and
- *     MERGES with our inline factories (lines 274-276).
- *   - `noExtensions` defaults to `false` (constructor, line 132) — so the naive
- *     loader DOES load disk extensions. That is the S2 gap.
+ * `excludeTools` is matched by NAME against BOTH Pi built-ins AND
+ * extension-registered tools (incl. agent-tempo's MCP tools via `renderToPi`), so
+ * this list contains ONLY Pi-built-in / exec names — never agent-tempo tool names
+ * (`cue`/`report`/`recruit`/…). Posture:
+ *   - `toolAccess === 'restricted'` → exclude {@link EXEC_TOOLS} (exec/bash
+ *     registration-absent; a strict upgrade of the prior call-time block, and the
+ *     headless default → the model never even sees exec).
+ *   - `guardrailPolicy === 'observe-only'` → also exclude the Pi built-in act
+ *     tools ({@link PI_BUILTIN_ACT_TOOLS}); read/grep/glob stay. The agent-tempo
+ *     MCP act tools (recruit/destroy/…) stay covered by the client-side no-act
+ *     handler (commit 5) — excludeTools handles the Pi built-ins only.
+ *   - `monitored` / `supervised` / `autonomous` → NO exec exclusion: those tools
+ *     stay REGISTERED so they can be gated/approved per-use (#712). (`supervised`
+ *     = approve-and-run, NOT exec-absent.)
  *
- * Fix (= security's "exclude the extensions dir", done structurally):
- *   - `noExtensions: true` → `extensionPaths` collapses to `cliEnabledExtensions`,
- *     which is empty because we pass NO `additionalExtensionPaths`. So
- *     `loadExtensions([])` registers nothing from disk/packages.
- *   - Inline `extensionFactories` load UNCONDITIONALLY (reload() line 275 is not
- *     gated by `noExtensions`), so our agent-tempo extension still attaches.
- * Net: the ONLY tools present are Pi's built-ins (bash/read/edit/write/grep —
- * all covered by the deny-list) + our agent-tempo MCP tools (no exec). No
- * third-party tool can slip past the deny-list. Skills/prompts/themes cannot
- * register tools, so they are not a vector and are left at defaults.
- *
- * Kept as a pure, exported helper so the `noExtensions: true` invariant has a
- * unit regression test (test/pi-headless-loader.test.ts) without needing the Pi
- * SDK installed.
+ * Pure + exported so the registration-absence invariant has a unit regression
+ * test without the Pi SDK (mirrors {@link buildPiResourceLoaderOptions}).
  */
+export declare function computeExcludeTools(toolAccess: PiToolAccess, guardrailPolicy: GuardrailPolicy | undefined): string[];
 export declare function buildPiResourceLoaderOptions(params: {
     cwd: string;
     agentDir: string;

package/dist/pi/headless.js

@@ -1,5 +1,6 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.computeExcludeTools = computeExcludeTools;
 exports.buildPiResourceLoaderOptions = buildPiResourceLoaderOptions;
 exports.runHeadlessPi = runHeadlessPi;
 /**
@@ -31,6 +32,7 @@ exports.runHeadlessPi = runHeadlessPi;
 const config_1 = require("../config");
 const sdk_probe_1 = require("../utils/sdk-probe");
 const extension_1 = require("./extension");
+const tool_capability_1 = require("../security/tool-capability");
 const probe_1 = require("./probe");
 const session_seed_1 = require("./session-seed");
 const log = (...args) => {
@@ -114,6 +116,61 @@ async function resolveModel(modelStr) {
  * unit regression test (test/pi-headless-loader.test.ts) without needing the Pi
  * SDK installed.
  */
+/**
+ * Pi BUILT-IN mutating ("act") tool names, excluded at registration for the
+ * `observe-only` no-act posture (#715). Pi's default built-ins are
+ * read/bash/edit/write; `multiedit` is listed defensively (a no-op if Pi doesn't
+ * register it). `bash`/exec are covered by {@link EXEC_TOOLS}. We keep the READ
+ * built-ins (read/grep/glob/ls). These are Pi BUILT-IN names — deliberately NOT
+ * agent-tempo MCP tool names, so excluding them never removes an agent-tempo
+ * coordination tool (those stay handler-gated, commit 5).
+ */
+const PI_BUILTIN_ACT_TOOLS = ['write', 'edit', 'multiedit'];
+/**
+ * #715 — compute the registration-level `excludeTools` denylist for
+ * `createAgentSession`. Excluded tools are never registered → ABSENT from the
+ * model's toolset AND system prompt: the LLM cannot request what it never sees.
+ *
+ * This is a registration-level FLOOR beneath the call-time MD-C handler + #712
+ * gate. It is tamper-RESISTANT, NOT tamper-PROOF: it defends a PROMPT-INJECTED
+ * agent (and holds even if the call-time gate had a bug — the tool simply isn't
+ * there), but it does NOT defend against PROCESS COMPROMISE — a tampered /
+ * modified extension can re-register or un-exclude tools (this is OUR code
+ * passing a denylist; an attacker who modifies the code/process bypasses it).
+ * That residual is OS-sandbox + supply-chain integrity, tracked as #724.
+ *
+ * `excludeTools` is matched by NAME against BOTH Pi built-ins AND
+ * extension-registered tools (incl. agent-tempo's MCP tools via `renderToPi`), so
+ * this list contains ONLY Pi-built-in / exec names — never agent-tempo tool names
+ * (`cue`/`report`/`recruit`/…). Posture:
+ *   - `toolAccess === 'restricted'` → exclude {@link EXEC_TOOLS} (exec/bash
+ *     registration-absent; a strict upgrade of the prior call-time block, and the
+ *     headless default → the model never even sees exec).
+ *   - `guardrailPolicy === 'observe-only'` → also exclude the Pi built-in act
+ *     tools ({@link PI_BUILTIN_ACT_TOOLS}); read/grep/glob stay. The agent-tempo
+ *     MCP act tools (recruit/destroy/…) stay covered by the client-side no-act
+ *     handler (commit 5) — excludeTools handles the Pi built-ins only.
+ *   - `monitored` / `supervised` / `autonomous` → NO exec exclusion: those tools
+ *     stay REGISTERED so they can be gated/approved per-use (#712). (`supervised`
+ *     = approve-and-run, NOT exec-absent.)
+ *
+ * Pure + exported so the registration-absence invariant has a unit regression
+ * test without the Pi SDK (mirrors {@link buildPiResourceLoaderOptions}).
+ */
+function computeExcludeTools(toolAccess, guardrailPolicy) {
+    const exclude = new Set();
+    if (toolAccess === 'restricted') {
+        for (const t of tool_capability_1.EXEC_TOOLS)
+            exclude.add(t);
+    }
+    if (guardrailPolicy === 'observe-only') {
+        for (const t of tool_capability_1.EXEC_TOOLS)
+            exclude.add(t); // no-act ⊇ no-exec
+        for (const t of PI_BUILTIN_ACT_TOOLS)
+            exclude.add(t);
+    }
+    return [...exclude];
+}
 function buildPiResourceLoaderOptions(params) {
     return {
         cwd: params.cwd,
@@ -192,10 +249,20 @@ async function runHeadlessPi(opts = {}) {
     // heartbeat). The SDK's own doc comment (sdk.js:74-83) prescribes this exact
     // construct → reload() → pass-as-resourceLoader sequence.
     await resourceLoader.reload();
+    // #715 — registration-level exec/act exclusion (the true "agent physically
+    // lacks the tools" boundary; see computeExcludeTools). Excluded tools are never
+    // registered, so they're absent from the model's toolset + system prompt — a
+    // hard layer beyond the call-time MD-C handler + #712 gate (kept as
+    // belt-and-suspenders). Empty for monitored/supervised/autonomous+standard.
+    const excludeTools = computeExcludeTools(toolAccess, opts.guardrailPolicy);
+    if (excludeTools.length > 0) {
+        log(`#715: excluding ${excludeTools.length} tool(s) at registration (toolAccess=${toolAccess}, guardrailPolicy=${opts.guardrailPolicy ?? 'autonomous'}): ${excludeTools.join(', ')}`);
+    }
     const { session } = await createAgentSession({
         cwd: process.cwd(),
         agentDir,
         ...(model ? { model } : {}),
+        ...(excludeTools.length > 0 ? { excludeTools } : {}),
         resourceLoader,
         // H1 (#645): in-memory session (seeded above via the session-seed chokepoint).
         // H2 will seed it from agent-tempo durable state (ENV.PI_CONTINUE_SESSION

package/dist/types.d.ts

@@ -238,24 +238,39 @@ export interface SessionMetadata {
      * the real posture on EVERY attach (across restart / migrate / re-attach), so
      * a previously-`supervised` agent stays supervised. (tempo-architect ruling.)
      *
-     * **★ Enforcement scope (#712 — claims EXACTLY this, no more).** The daemon
-     * enforces fail-closed for a `supervised` player's gate ENGAGEMENTS: the gate's
-     * `open()` cross-check derives the `failMode` from THIS durable policy (read by
-     * the daemon, populated at spawn + lazily resolved on a cache-miss), so the
-     * `failMode` an engaging agent stamps on its `gate_pending` frame can't
-     * self-downgrade a supervised player out of fail-closed (a `'open'` claim against
-     * a `supervised` policy is overridden to `closed` and a `failmode-override`
-     * audit record is written). This is **NOT a hard boundary** against an agent
-     * that SKIPS gate-engagement entirely (runs the tool without emitting a
-     * `gate_pending` frame) or bypasses the client-side MD-C exec-block — a true
-     * boundary requires spawn-time tool-restriction (the agent physically lacks the
-     * tools), tracked separately in **#715**.
+     * **★ Enforcement scope (#712/#715).** `supervised` is the daemon-enforced
+     * approval boundary for the realistic threat: a prompt-injected agent. A
+     * manipulated LLM can only *emit* tool-call requests — Pi routes every one to
+     * agent-tempo's `tool_call` handler, which engages the gate (non-`low-risk`;
+     * #712 daemon-computes `failMode` from this durable policy, falling `closed` on
+     * any lookup failure — no-fail-open, so an engaging agent can't self-downgrade).
+     * The agent **cannot** skip the gate — it doesn't control the hook.
+     *
+     * **#715 adds a registration-level floor.** For `toolAccess: 'restricted'` (and
+     * `observe-only`'s act tools) the exec/act tools are EXCLUDED at
+     * `createAgentSession` (`excludeTools`) → **absent** from the model's toolset and
+     * system prompt entirely; the LLM cannot request what it never sees. That is
+     * stronger than a call-time block — it holds even if the call-time gate had a bug
+     * (the tool simply isn't there). `supervised` with exec present keeps exec
+     * **present + gated** (approve-per-use), so this floor applies to the exec/no-act
+     * postures, not to a `supervised`+`standard` player.
+     *
+     * **Residual (all postures): process compromise** — code execution *inside* the
+     * Pi process (in-process syscalls; host RCE bypassing the handler), OR a
+     * tampered / modified extension that un-excludes or re-registers tools.
+     * `excludeTools` is OUR code passing a denylist; an attacker who modifies that
+     * code or the process bypasses it. The only defense is OS-level sandboxing +
+     * supply-chain integrity, a separate future `'sandboxed'` posture (#724). So:
+     * **tamper-RESISTANT** vs prompt-injection + an honest gate bug; **NOT
+     * tamper-PROOF** vs a compromised process. Against prompt-injection — the
+     * realistic threat — it **is** a real enforcement boundary; #724 is not a gap in
+     * that scope.
      *
      * **Post-restart window:** on daemon restart the in-memory ingest tokens are
      * invalidated, so existing players' gate engagements are rejected (403) until a
      * re-spawn re-mints. In that window a `supervised` player's gate-client
-     * fail-closes on its own derived deadline (client-side safety holds), but the
-     * gate is NOT daemon-mediated — the #715 client-cooperative residual.
+     * fail-closes on its own derived deadline (client-side safety holds, not
+     * daemon-mediated) — same process-compromise residual, not a distinct gap.
      */
     guardrailPolicy?: GuardrailPolicy;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-tempo",
-  "version": "1.7.0-beta.3",
+  "version": "1.7.0-beta.5",
   "description": "Many agents, one tempo. Durable coordination for multi-agent work via Temporal.",
   "keywords": [
     "mcp",
@@ -72,12 +72,12 @@
     "copilot-bridge": "ts-node src/adapters/copilot/adapter.ts",
     "clean:test": "node -e \"require('fs').rmSync('dist-test',{recursive:true,force:true})\"",
     "build:test": "npm run clean:test && tsc -p test/tsconfig.json",
-    "pretest": "npm run build:test",
+    "pretest": "node scripts/check-bundle-present.js && npm run build:test",
     "test:tui": "vitest run",
     "test:conformance": "npm run build:test && mocha --config .mocharc.conformance.yml",
-    "pretest:shard-1": "npm run build:test && npm run build:scripts",
+    "pretest:shard-1": "node scripts/check-bundle-present.js && npm run build:test && npm run build:scripts",
     "test:shard-1": "node dist/scripts/run-shard.js 1",
-    "pretest:shard-2": "npm run build:test && npm run build:scripts",
+    "pretest:shard-2": "node scripts/check-bundle-present.js && npm run build:test && npm run build:scripts",
     "test:shard-2": "node dist/scripts/run-shard.js 2",
     "test": "mocha && vitest run",
     "lint:surface-drift": "node scripts/check-surface-drift.js",