feat(scan): add live progress and recon bootstrap (#81)

* Add live scan progress view

* Make local-agent recon report real tool progress

* Feed recon bootstrap into local agent ReAct loop

* Remove stale local-agent JSON tool parser

* Label backend-seeded scan progress

---------

Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
adrianwandCodex authored and GitHub committed 2026-07-15 19:42:20 -04:00
1 parent 8959a288e5
commit 089591b18f
6 files changed
+476 -11

No files matched your search

+70 -4
View File
@@ -71,8 +71,8 @@ export interface AgentResult {
export interface AgentEvents {
'agent:step': AgentStep;
'agent:tool_call': { name: string; args: Record<string, unknown> };
'agent:tool_result': { name: string; result: ToolResult };
'agent:tool_call': { name: string; args: Record<string, unknown>; source?: 'agent' | 'backend_seeded' };
'agent:tool_result': { name: string; result: ToolResult; source?: 'agent' | 'backend_seeded' };
'agent:thinking': { content: string };
'agent:complete': AgentResult;
'agent:error': { error: Error; step: number };
@@ -133,6 +133,39 @@ export class AgentLoop extends EventEmitter<AgentEvents> {
{ role: 'user', content: this.buildTaskPrompt(task, target, toolDefs, sourceContext, sharedContext) },
];
const bootstrapCall = this.getLocalReconBootstrap(task, target, toolDefs);
if (bootstrapCall) {
const toolStep = await this.executeTool(bootstrapCall, target, -1);
steps.push(toolStep);
seenCalls.set(
`${bootstrapCall.name}:${JSON.stringify(bootstrapCall.arguments || {})}`,
String(toolStep.toolResult?.output || toolStep.toolResult?.error || 'no output').replace(/\s+/g, ' ').slice(0, 160)
);
if (toolStep.toolResult?.findings) {
const out = String(toolStep.toolResult.output ?? '').slice(0, 4000);
for (const tf of toolStep.toolResult.findings) {
allFindings.push({ ...tf, provenance: 'tool', toolName: bootstrapCall.name, toolOutput: out || tf.details });
}
}
messages.push({
role: 'assistant',
content: 'Running an initial low-noise recon check before planning deeper actions.',
toolCalls: [bootstrapCall],
});
messages.push({
role: 'tool',
content: this.formatToolResult(toolStep.toolResult),
toolCallId: bootstrapCall.id,
name: bootstrapCall.name,
});
messages.push({
role: 'user',
content: 'Use the bootstrap observation above as your first tool observation. Continue through the normal ACTION CONTRACT: request another listed tool only if it will materially improve the evidence; otherwise provide the final recon debrief.',
});
}
for (let i = 0; i < this.options.maxIterations; i++) {
try {
// Ask the LLM what to do next
@@ -298,6 +331,38 @@ export class AgentLoop extends EventEmitter<AgentEvents> {
return result;
}
private getLocalReconBootstrap(
task: Task,
target: Target | undefined,
tools: LLMToolDefinition[]
): LLMToolCall | undefined {
if (this.llm.getProvider() !== 'local-agent') return undefined;
if (!target?.address) return undefined;
if (task.phase !== 'reconnaissance' && task.operatorType !== 'recon') return undefined;
const hasTool = (name: string) => tools.some((tool) => tool.name === name);
const taskName = task.name.toLowerCase();
const id = `local-agent-bootstrap-${Date.now()}`;
if (taskName.includes('dns') && hasTool('dns_lookup')) {
return { id, name: 'dns_lookup', arguments: { domain: target.address, type: 'A' } };
}
if (taskName.includes('port') && hasTool('port_scan')) {
return { id, name: 'port_scan', arguments: { target: target.address, ports: '80,443', timeout: 2000 } };
}
if ((taskName.includes('web') || taskName.includes('fingerprint')) && hasTool('header_analysis')) {
return { id, name: 'header_analysis', arguments: { url: `https://${target.address}` } };
}
if ((taskName.includes('content') || taskName.includes('discover')) && hasTool('robots_txt_fetch')) {
return { id, name: 'robots_txt_fetch', arguments: { url: `https://${target.address}` } };
}
return undefined;
}
/**
* Execute a single tool call via the Arsenal
*/
@@ -306,7 +371,8 @@ export class AgentLoop extends EventEmitter<AgentEvents> {
target: Target | undefined,
iteration: number
): Promise<AgentStep> {
this.emit('agent:tool_call', { name: toolCall.name, args: toolCall.arguments });
const source = iteration < 0 ? 'backend_seeded' : 'agent';
this.emit('agent:tool_call', { name: toolCall.name, args: toolCall.arguments, source });
let toolResult: ToolResult;
try {
@@ -330,7 +396,7 @@ export class AgentLoop extends EventEmitter<AgentEvents> {
};
}
this.emit('agent:tool_result', { name: toolCall.name, result: toolResult });
this.emit('agent:tool_result', { name: toolCall.name, result: toolResult, source });
return {
iteration,
+100
View File
@@ -212,6 +212,8 @@ import type {
OperatorArchetype,
CommandEvents,
Finding,
ScanProgressEvent,
Task,
} from './types/index.js';
// Re-export commonly used types
@@ -271,6 +273,7 @@ import {
const DEFAULT_AGENT_MAX_ITERATIONS = 15;
const LOCAL_AGENT_MAX_ITERATIONS = Number(process.env.T3MP3ST_LOCAL_AGENT_MAX_ITERATIONS || 30);
const MAX_PROGRESS_EVENTS = 300;
/**
* TEMPEST Command - Main orchestration controller
@@ -544,6 +547,40 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
* Setup event forwarding for an operator
*/
private setupOperatorEvents(operator: OperatorAgent): void {
operator.on('task:started', ({ task }) => {
this.recordProgress('task_started', operator, task, `Started ${task.name}`);
});
operator.on('task:completed', ({ task, result }) => {
const findings = result.findings?.length ? ` Findings: ${result.findings.join(', ')}` : '';
this.recordProgress(
'task_completed',
operator,
task,
`${result.success === false ? 'Finished unsuccessfully' : 'Completed'} ${task.name}.${findings}`,
{ success: result.success !== false }
);
});
operator.on('task:failed', ({ task, error }) => {
this.recordProgress('task_failed', operator, task, error, { success: false });
});
operator.on('agent:thinking', ({ task, content }) => {
this.recordProgress('thinking', operator, task, content);
});
operator.on('agent:tool_call', ({ task, name, args, source }) => {
this.recordProgress('tool_call', operator, task, `${name} ${JSON.stringify(args || {})}`, { toolName: name, source });
});
operator.on('agent:tool_result', ({ task, name, result, source }) => {
const detail = result.success
? (result.output || 'Tool completed')
: (result.error || result.output || 'Tool failed');
this.recordProgress('tool_result', operator, task, detail, { toolName: name, success: result.success, source });
});
operator.on('finding:discovered', ({ finding }) => {
this.vault.addFinding(finding);
this.emit('finding:discovered', { finding, operatorId: operator.id });
@@ -628,6 +665,34 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
});
}
private recordProgress(
kind: ScanProgressEvent['kind'],
operator: OperatorAgent,
task: Task | undefined,
detail: string,
extra: Partial<Pick<ScanProgressEvent, 'toolName' | 'success' | 'source'>> = {}
): void {
const compact = String(detail || '').replace(/\s+/g, ' ').trim().slice(0, 2000);
const event: ScanProgressEvent = {
id: `progress-${Date.now()}-${Math.random().toString(16).slice(2)}`,
timestamp: Date.now(),
kind,
operatorId: operator.id,
callsign: operator.callsign,
archetype: operator.archetype,
taskId: task?.id,
taskName: task?.name,
detail: compact,
...extra,
};
this.progressEvents.push(event);
if (this.progressEvents.length > MAX_PROGRESS_EVENTS) {
this.progressEvents.splice(0, this.progressEvents.length - MAX_PROGRESS_EVENTS);
}
this.emit('scan:progress', event);
}
/**
* Parse a finding and update the target's services/vulnerabilities.
* This is the intelligence pipeline that feeds data between phases.
@@ -826,6 +891,8 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
* CLI work; it only fires on truly-hung dispatches. Override via
* T3MP3ST_TASK_TIMEOUT_MS.
*/
private progressEvents: ScanProgressEvent[] = [];
/**
* Resolve the dispatch timeout from the environment, falling back to the
* provider-specific default. Guards against a non-numeric / non-positive override.
@@ -951,6 +1018,7 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
// Auto-spawn an operator if none exists for this archetype
if (!operator) {
if (this.llm.getProvider() === 'local-agent') continue;
const allOps = this.cell.getAllOperators();
const archetypeCount = allOps.filter(op => op.archetype === task.operatorType).length;
// Spawn up to 3 operators per archetype for parallelism
@@ -1099,6 +1167,9 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
* Auto-spawn operators needed for a given kill chain phase
*/
private autoSpawnForPhase(phase: KillChainPhase): void {
if (this.llm.getProvider() === 'local-agent') {
return;
}
const phaseOperators: Record<string, OperatorArchetype[]> = {
[KillChainPhase.RECON]: ['recon'],
[KillChainPhase.WEAPONIZE]: ['scanner'],
@@ -1139,6 +1210,7 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
this.on('detection:triggered', (data) => broadcast('detection', data));
this.on('mission:phase_changed', (data) => broadcast('phase_changed', data));
this.on('approval:decision', (data) => broadcast('arsenal.approval', data));
this.on('scan:progress', (data) => broadcast('scan:progress', data));
this.on('tick', (count) => {
// Broadcast status every 5 ticks to avoid flooding
if (typeof count === 'number' && count % 5 === 0) {
@@ -1232,8 +1304,19 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
opsec: ReturnType<OpsecController['getStats']>;
activeMission: string | null;
stallReason: string | null;
progress: ScanProgressEvent[];
tasks: Array<{
id: string;
name: string;
phase: string;
status: string;
operatorType: string;
assignedTo?: string;
result?: { success: boolean; output?: string; error?: string; findings?: string[] };
}>;
} {
const activeMission = this.mission.getActiveMission();
const taskQueue = this.mission.getTaskQueue();
return {
name: this.name,
@@ -1246,6 +1329,23 @@ export class TempestCommand extends EventEmitter<CommandEvents> {
opsec: this.opsec.getStats(),
activeMission: activeMission?.id || null,
stallReason: this.stallReason,
progress: [...this.progressEvents],
tasks: activeMission
? taskQueue.getForMission(activeMission.id).map(task => ({
id: task.id,
name: task.name,
phase: task.phase,
status: task.status,
operatorType: task.operatorType,
assignedTo: task.assignedTo,
result: task.result ? {
success: task.result.success,
output: task.result.output,
error: task.result.error,
findings: task.result.findings,
} : undefined,
}))
: [],
};
}
+36 -5
View File
@@ -22,6 +22,7 @@ import {
import type { LLMBackbone } from '../llm/index.js';
import type { Arsenal } from '../arsenal/index.js';
import type { AgentLoop } from '../agent/index.js';
import type { ToolResult } from '../types/index.js';
import type { Target } from '../types/index.js';
import { OPERATOR_SYSTEM_PROMPTS } from '../prompts/index.js';
import { gateLiveFinding } from '../evidence/gate.js';
@@ -36,6 +37,9 @@ export interface OperatorEvents {
'task:completed': { task: Task; result: TaskResult };
'task:failed': { task: Task; error: string };
'task:decomposed': { parent: Task; subtasks: Task[]; reason: string };
'agent:thinking': { task: Task; content: string };
'agent:tool_call': { task: Task; name: string; args: Record<string, unknown>; source?: 'agent' | 'backend_seeded' };
'agent:tool_result': { task: Task; name: string; result: ToolResult; source?: 'agent' | 'backend_seeded' };
'finding:discovered': { finding: Finding };
'finding:gate-blocked': { finding: Finding; reasons: string[] };
'credential:harvested': { credential: Credential };
@@ -422,11 +426,38 @@ export class OperatorAgent extends EventEmitter<OperatorEvents> {
// Pass the white-box source excerpt (if any) so the model sees the target's
// real source alongside the task; empty string keeps black-box behavior.
if (this.agentLoop && this.llm) {
// [Phase-2] Register as live on the board and pull the shared situation report so this
// operator sees teammates' verified leads/claims — it builds on them instead of running blind.
this.board?.heartbeat(this.id, 'hunting', task.name);
const sharedContext = this.board?.situationReport(this.id);
const result = await this.agentLoop.run(task, this.profile.systemPrompt, target, this.whiteboxSource, sharedContext);
const onThinking = ({ content }: { content: string }): void => {
this.emit('agent:thinking', { task, content });
};
const onToolCall = ({ name, args, source }: { name: string; args: Record<string, unknown>; source?: 'agent' | 'backend_seeded' }): void => {
this.emit('agent:tool_call', { task, name, args, source });
};
const onToolResult = ({ name, result, source }: { name: string; result: ToolResult; source?: 'agent' | 'backend_seeded' }): void => {
this.emit('agent:tool_result', { task, name, result, source });
};
const canForwardAgentEvents =
typeof this.agentLoop.on === 'function' &&
typeof this.agentLoop.off === 'function';
if (canForwardAgentEvents) {
this.agentLoop.on('agent:thinking', onThinking);
this.agentLoop.on('agent:tool_call', onToolCall);
this.agentLoop.on('agent:tool_result', onToolResult);
}
let result;
try {
// [Phase-2] Register as live on the board and pull the shared situation report so this
// operator sees teammates' verified leads/claims — it builds on them instead of running blind.
this.board?.heartbeat(this.id, 'hunting', task.name);
const sharedContext = this.board?.situationReport(this.id);
result = await this.agentLoop.run(task, this.profile.systemPrompt, target, this.whiteboxSource, sharedContext);
} finally {
if (canForwardAgentEvents) {
this.agentLoop.off('agent:thinking', onThinking);
this.agentLoop.off('agent:tool_call', onToolCall);
this.agentLoop.off('agent:tool_result', onToolResult);
}
}
// Convert agent findings to operator findings.
// PROVENANCE-HONEST: only a tool-backed finding gets tool-output evidence (the raw
+3 -1
View File
@@ -6293,7 +6293,7 @@ app.post('/api/mission/resume', (_req: Request, res: Response) => {
app.get('/api/mission/status', (_req: Request, res: Response) => {
const cmd = getTempestCommand();
if (!cmd) {
res.json({ active: false });
res.json({ active: false, progress: [], tasks: [] });
return;
}
@@ -6323,6 +6323,8 @@ app.get('/api/mission/status', (_req: Request, res: Response) => {
targets: status.targets,
vault: status.vault,
opsec: status.opsec,
tasks: status.tasks,
progress: status.progress,
findings: findings.map(f => ({
id: f.id,
title: f.title,
+16
View File
@@ -534,6 +534,21 @@ export interface Appendix {
// EVENT TYPES
// =============================================================================
export interface ScanProgressEvent {
id: string;
timestamp: number;
kind: 'task_started' | 'thinking' | 'tool_call' | 'tool_result' | 'task_completed' | 'task_failed';
operatorId: string;
callsign: string;
archetype: OperatorArchetype;
taskId?: string;
taskName?: string;
toolName?: string;
source?: 'agent' | 'backend_seeded';
detail: string;
success?: boolean;
}
export interface CommandEvents {
'command:started': void;
'command:stopped': void;
@@ -547,6 +562,7 @@ export interface CommandEvents {
'target:owned': { target: Target; operatorId: string };
'detection:triggered': DetectionEvent;
'mission:phase_changed': { missionId: string; phase: KillChainPhase };
'scan:progress': ScanProgressEvent;
'abort:recommended': string;
/** A capability-approval gate decision (allowed/denied) on an intrusive/dangerous tool — bridged to
* the dashboard's live approval/audit feed. Structural match for arsenal/approval.ts ApprovalRecord. */