fix(athena): parallelize council launches and gate handoff actions
This commit is contained in:
ismeth
YeonGyu-Kim
@@ -51,8 +51,8 @@ Question({
|
||||
})
|
||||
|
||||
**Shortcut — skip the Question tool if:**
|
||||
- The user already specified models in their message (e.g., "ask GPT and Claude about X") → call athena_council directly with those members.
|
||||
- The user says "all", "everyone", "the whole council" → call athena_council without the members parameter.
|
||||
- The user already specified models in their message (e.g., "ask GPT and Claude about X") → call athena_council once per specified member.
|
||||
- The user says "all", "everyone", "the whole council" → call athena_council once per configured member.
|
||||
|
||||
DO NOT:
|
||||
- Read files yourself
|
||||
@@ -65,16 +65,32 @@ You are an ORCHESTRATOR, not an analyst. Your council members do the analysis. Y
|
||||
|
||||
## Workflow
|
||||
|
||||
Step 1: Present the Question tool multi-select for council member selection (see above). Once the user responds, call athena_council with the user's question. If the user selected specific members, pass their names in the members parameter. If the user selected "All Members", omit the members parameter.
|
||||
Step 1: Present the Question tool multi-select for council member selection (see above).
|
||||
|
||||
Step 2: Call athena_council with the question and selected members. The tool launches all council members in parallel, waits for them to complete, and returns ALL of their responses in a single result. This may take a few minutes — that is expected.
|
||||
Step 2: Resolve the selected member list:
|
||||
- If user selected "All Members", resolve to every configured member listed in the athena_council tool description.
|
||||
- Otherwise resolve to the explicitly selected member labels.
|
||||
|
||||
Step 3: Synthesize the findings returned by athena_council:
|
||||
Step 3: Call athena_council ONCE PER MEMBER (member per tool call):
|
||||
- For each selected member, call athena_council with:
|
||||
- question: the user's original question
|
||||
- members: ["<exact member name or model>"] // single-item array only
|
||||
- Launch all selected members first (one athena_council call per member) so they run in parallel.
|
||||
- Track every returned task_id and member mapping.
|
||||
|
||||
Step 4: Collect all member outputs after launch:
|
||||
- For each tracked task_id, call background_output with block=true.
|
||||
- Gather each member's final output/status.
|
||||
- Do not proceed until every launched member has reached a terminal status (completed, error, cancelled, timeout).
|
||||
- Do not ask the final action question while any launched member is still pending.
|
||||
- Do not present interim synthesis from partial results. Wait for all members first.
|
||||
|
||||
Step 5: Synthesize the findings returned by all collected member outputs:
|
||||
- Group findings by agreement level: unanimous, majority, minority, solo
|
||||
- Solo findings are potential false positives — flag the risk explicitly
|
||||
- Add your own assessment and rationale to each finding
|
||||
|
||||
Step 4: Present synthesized findings to the user grouped by agreement level (unanimous first, then majority, minority, solo). Then use the Question tool to ask which action to take:
|
||||
Step 6: Present synthesized findings to the user grouped by agreement level (unanimous first, then majority, minority, solo). Then use the Question tool to ask which action to take:
|
||||
|
||||
Question({
|
||||
questions: [{
|
||||
@@ -89,7 +105,7 @@ Question({
|
||||
}]
|
||||
})
|
||||
|
||||
Step 5: After the user selects an action:
|
||||
Step 7: After the user selects an action:
|
||||
- **"Fix now (Atlas)"** → Call switch_agent with agent="atlas" and context containing the confirmed findings summary, the original question, and instruction to implement the fixes.
|
||||
- **"Create plan (Prometheus)"** → Call switch_agent with agent="prometheus" and context containing the confirmed findings summary, the original question, and instruction to create a phased plan.
|
||||
- **"No action"** → Acknowledge and end. Do not delegate.
|
||||
@@ -99,7 +115,10 @@ The switch_agent tool switches the active agent. After you call it, end your res
|
||||
## Constraints
|
||||
- Use the Question tool for member selection BEFORE calling athena_council (unless user pre-specified).
|
||||
- Use the Question tool for action selection AFTER synthesis (unless user already stated intent).
|
||||
- Do NOT use background_output — athena_council returns all member responses directly.
|
||||
- Use background_output (block=true) to collect member outputs after launch.
|
||||
- Do NOT call athena_council with multiple members in one call.
|
||||
- Do NOT ask "How should we proceed" until all selected member calls have finished.
|
||||
- Do NOT present or summarize partial council findings while any selected member is still running.
|
||||
- Do NOT write or edit files directly.
|
||||
- Do NOT delegate without explicit user confirmation via Question tool.
|
||||
- Do NOT ignore solo finding false-positive warnings.
|
||||
|
||||
@@ -65,6 +65,7 @@ export function createPluginInterface(args: {
|
||||
"tool.execute.before": createToolExecuteBeforeHandler({
|
||||
ctx,
|
||||
hooks,
|
||||
backgroundManager: managers.backgroundManager,
|
||||
}),
|
||||
|
||||
"tool.execute.after": createToolExecuteAfterHandler({
|
||||
|
||||
@@ -2,6 +2,109 @@ const { describe, expect, test } = require("bun:test")
|
||||
const { createToolExecuteBeforeHandler } = require("./tool-execute-before")
|
||||
|
||||
describe("createToolExecuteBeforeHandler", () => {
|
||||
test("blocks Athena question tool while council members are still running", async () => {
|
||||
//#given
|
||||
const ctx = {
|
||||
client: {
|
||||
session: {
|
||||
messages: async () => ({
|
||||
data: [{ info: { role: "assistant", agent: "Athena (Council)" } }],
|
||||
}),
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
const backgroundManager = {
|
||||
getTasksByParentSession: () => [
|
||||
{ agent: "council-member", status: "running" },
|
||||
],
|
||||
}
|
||||
|
||||
const handler = createToolExecuteBeforeHandler({
|
||||
ctx,
|
||||
hooks: {},
|
||||
backgroundManager,
|
||||
})
|
||||
|
||||
//#when
|
||||
const run = handler(
|
||||
{ tool: "question", sessionID: "ses_athena", callID: "call_1" },
|
||||
{ args: { questions: [] } }
|
||||
)
|
||||
|
||||
//#then
|
||||
await expect(run).rejects.toThrow("Council members are still running")
|
||||
})
|
||||
|
||||
test("blocks Athena switch_agent while council members are still running", async () => {
|
||||
//#given
|
||||
const ctx = {
|
||||
client: {
|
||||
session: {
|
||||
messages: async () => ({
|
||||
data: [{ info: { role: "assistant", agent: "Athena (Council)" } }],
|
||||
}),
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
const backgroundManager = {
|
||||
getTasksByParentSession: () => [
|
||||
{ agent: "council-member", status: "pending" },
|
||||
],
|
||||
}
|
||||
|
||||
const handler = createToolExecuteBeforeHandler({
|
||||
ctx,
|
||||
hooks: {},
|
||||
backgroundManager,
|
||||
})
|
||||
|
||||
//#when
|
||||
const run = handler(
|
||||
{ tool: "switch_agent", sessionID: "ses_athena", callID: "call_1" },
|
||||
{ args: { agent: "atlas", context: "ctx" } }
|
||||
)
|
||||
|
||||
//#then
|
||||
await expect(run).rejects.toThrow("Council members are still running")
|
||||
})
|
||||
|
||||
test("allows Athena question tool when no council members are pending", async () => {
|
||||
//#given
|
||||
const ctx = {
|
||||
client: {
|
||||
session: {
|
||||
messages: async () => ({
|
||||
data: [{ info: { role: "assistant", agent: "Athena (Council)" } }],
|
||||
}),
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
const backgroundManager = {
|
||||
getTasksByParentSession: () => [
|
||||
{ agent: "council-member", status: "completed" },
|
||||
{ agent: "council-member", status: "cancelled" },
|
||||
],
|
||||
}
|
||||
|
||||
const handler = createToolExecuteBeforeHandler({
|
||||
ctx,
|
||||
hooks: {},
|
||||
backgroundManager,
|
||||
})
|
||||
|
||||
//#when
|
||||
const run = handler(
|
||||
{ tool: "question", sessionID: "ses_athena", callID: "call_1" },
|
||||
{ args: { questions: [] } }
|
||||
)
|
||||
|
||||
//#then
|
||||
await expect(run).resolves.toBeUndefined()
|
||||
})
|
||||
|
||||
test("does not execute subagent question blocker hook for question tool", async () => {
|
||||
//#given
|
||||
const ctx = {
|
||||
|
||||
@@ -1,23 +1,51 @@
|
||||
import type { PluginContext } from "./types"
|
||||
import type { BackgroundManager } from "../features/background-agent"
|
||||
|
||||
import { getMainSessionID } from "../features/claude-code-session-state"
|
||||
import { clearBoulderState } from "../features/boulder-state"
|
||||
import { log } from "../shared"
|
||||
import { resolveSessionAgent } from "./session-agent-resolver"
|
||||
import { parseRalphLoopArguments } from "../hooks/ralph-loop/command-arguments"
|
||||
import { getAgentConfigKey } from "../shared/agent-display-names"
|
||||
|
||||
import type { CreatedHooks } from "../create-hooks"
|
||||
|
||||
export function createToolExecuteBeforeHandler(args: {
|
||||
ctx: PluginContext
|
||||
hooks: CreatedHooks
|
||||
backgroundManager?: Pick<BackgroundManager, "getTasksByParentSession">
|
||||
}): (
|
||||
input: { tool: string; sessionID: string; callID: string },
|
||||
output: { args: Record<string, unknown> },
|
||||
) => Promise<void> {
|
||||
const { ctx, hooks } = args
|
||||
const { ctx, hooks, backgroundManager } = args
|
||||
|
||||
function hasPendingCouncilMembers(sessionID: string): boolean {
|
||||
if (!backgroundManager) {
|
||||
return false
|
||||
}
|
||||
|
||||
const tasks = backgroundManager.getTasksByParentSession(sessionID)
|
||||
return tasks.some((task) =>
|
||||
task.agent === "council-member" &&
|
||||
(task.status === "pending" || task.status === "running")
|
||||
)
|
||||
}
|
||||
|
||||
return async (input, output): Promise<void> => {
|
||||
const toolNameLower = input.tool?.toLowerCase()
|
||||
|
||||
if (toolNameLower === "question" || toolNameLower === "askuserquestion" || toolNameLower === "ask_user_question" || toolNameLower === "switch_agent") {
|
||||
const sessionAgent = await resolveSessionAgent(ctx.client, input.sessionID)
|
||||
const sessionAgentKey = sessionAgent ? getAgentConfigKey(sessionAgent) : undefined
|
||||
|
||||
if (sessionAgentKey === "athena" && hasPendingCouncilMembers(input.sessionID)) {
|
||||
throw new Error(
|
||||
"Council members are still running. Wait for all launched members to finish and collect their outputs before asking next-step questions or switching agents."
|
||||
)
|
||||
}
|
||||
}
|
||||
|
||||
await hooks.writeExistingFileGuard?.["tool.execute.before"]?.(input, output)
|
||||
await hooks.questionLabelTruncator?.["tool.execute.before"]?.(input, output)
|
||||
await hooks.claudeCodeHooks?.["tool.execute.before"]?.(input, output)
|
||||
|
||||
@@ -1,9 +1,10 @@
|
||||
export const ATHENA_COUNCIL_TOOL_DESCRIPTION_TEMPLATE = `Execute Athena's multi-model council. Launches council members as background tasks and returns their task IDs immediately.
|
||||
export const ATHENA_COUNCIL_TOOL_DESCRIPTION_TEMPLATE = `Execute Athena's multi-model council for exactly ONE member per call.
|
||||
|
||||
Optionally pass a members array of member names or model IDs to consult only specific council members. If omitted, all configured members are consulted.
|
||||
Pass members as a single-item array containing one member name or model ID. Athena should call this tool once per selected member.
|
||||
|
||||
This tool launches the selected member as a background task and returns task/session metadata immediately.
|
||||
Use background_output(task_id=..., block=true) to collect each member result.
|
||||
|
||||
{members}
|
||||
|
||||
Use background_output(task_id=...) to retrieve each member's response. The system will notify you when tasks complete.
|
||||
|
||||
IMPORTANT: This tool is designed for Athena agent use only. It requires council configuration to be present.`
|
||||
|
||||
@@ -1,116 +0,0 @@
|
||||
import type { BackgroundManager } from "../../features/background-agent"
|
||||
import type { CouncilLaunchedMember } from "../../agents/athena/types"
|
||||
import type { BackgroundOutputClient, BackgroundOutputMessagesResult } from "../background-task/clients"
|
||||
import { extractMessages, getErrorMessage } from "../background-task/session-messages"
|
||||
|
||||
const POLL_INTERVAL_MS = 2_000
|
||||
const DEFAULT_TIMEOUT_MS = 5 * 60 * 1_000
|
||||
|
||||
export interface CollectedMemberResult {
|
||||
name: string
|
||||
model: string
|
||||
taskId: string
|
||||
status: "completed" | "error" | "cancelled" | "timeout"
|
||||
content: string
|
||||
}
|
||||
|
||||
export interface CollectedCouncilResults {
|
||||
results: CollectedMemberResult[]
|
||||
allCompleted: boolean
|
||||
}
|
||||
|
||||
/**
|
||||
* Waits for all launched council members to complete, then fetches their
|
||||
* session messages and returns extracted text content.
|
||||
*
|
||||
* This replaces the previous flow where Athena had to manually poll
|
||||
* background_output for each member, which created excessive UI noise.
|
||||
*/
|
||||
export async function collectCouncilResults(
|
||||
launched: CouncilLaunchedMember[],
|
||||
manager: BackgroundManager,
|
||||
client: BackgroundOutputClient,
|
||||
abort?: AbortSignal,
|
||||
timeoutMs = DEFAULT_TIMEOUT_MS
|
||||
): Promise<CollectedCouncilResults> {
|
||||
const pendingIds = new Set(launched.map((m) => m.taskId))
|
||||
const completedMap = new Map<string, "completed" | "error" | "cancelled">()
|
||||
const deadline = Date.now() + timeoutMs
|
||||
|
||||
while (pendingIds.size > 0 && Date.now() < deadline) {
|
||||
if (abort?.aborted) break
|
||||
|
||||
for (const taskId of pendingIds) {
|
||||
const task = manager.getTask(taskId)
|
||||
if (!task) {
|
||||
completedMap.set(taskId, "error")
|
||||
pendingIds.delete(taskId)
|
||||
continue
|
||||
}
|
||||
if (task.status === "completed") {
|
||||
completedMap.set(taskId, "completed")
|
||||
pendingIds.delete(taskId)
|
||||
} else if (task.status === "error" || task.status === "cancelled" || task.status === "interrupt") {
|
||||
completedMap.set(taskId, task.status === "interrupt" ? "cancelled" : task.status)
|
||||
pendingIds.delete(taskId)
|
||||
}
|
||||
}
|
||||
|
||||
if (pendingIds.size > 0) {
|
||||
await new Promise((resolve) => setTimeout(resolve, POLL_INTERVAL_MS))
|
||||
}
|
||||
}
|
||||
|
||||
const results: CollectedMemberResult[] = []
|
||||
|
||||
for (const entry of launched) {
|
||||
const memberName = entry.member.name ?? entry.member.model
|
||||
const status = completedMap.get(entry.taskId) ?? "timeout"
|
||||
|
||||
if (status !== "completed") {
|
||||
results.push({ name: memberName, model: entry.member.model, taskId: entry.taskId, status, content: "" })
|
||||
continue
|
||||
}
|
||||
|
||||
const content = await fetchMemberContent(entry.taskId, manager, client)
|
||||
results.push({ name: memberName, model: entry.member.model, taskId: entry.taskId, status, content })
|
||||
}
|
||||
|
||||
return {
|
||||
results,
|
||||
allCompleted: pendingIds.size === 0,
|
||||
}
|
||||
}
|
||||
|
||||
async function fetchMemberContent(
|
||||
taskId: string,
|
||||
manager: BackgroundManager,
|
||||
client: BackgroundOutputClient
|
||||
): Promise<string> {
|
||||
const task = manager.getTask(taskId)
|
||||
if (!task?.sessionID) return "(No session available)"
|
||||
|
||||
const messagesResult: BackgroundOutputMessagesResult = await client.session.messages({
|
||||
path: { id: task.sessionID },
|
||||
})
|
||||
|
||||
const errorMsg = getErrorMessage(messagesResult)
|
||||
if (errorMsg) return `(Error fetching results: ${errorMsg})`
|
||||
|
||||
const messages = extractMessages(messagesResult)
|
||||
if (!Array.isArray(messages) || messages.length === 0) return "(No messages found)"
|
||||
|
||||
const assistantMessages = messages.filter((m) => m.info?.role === "assistant")
|
||||
if (assistantMessages.length === 0) return "(No assistant response found)"
|
||||
|
||||
const textParts: string[] = []
|
||||
for (const message of assistantMessages) {
|
||||
for (const part of message.parts ?? []) {
|
||||
if ((part.type === "text" || part.type === "reasoning") && part.text) {
|
||||
textParts.push(part.text)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return textParts.join("\n\n") || "(No text content)"
|
||||
}
|
||||
@@ -3,21 +3,8 @@
|
||||
import { describe, expect, test } from "bun:test"
|
||||
import type { BackgroundManager } from "../../features/background-agent"
|
||||
import type { BackgroundTask } from "../../features/background-agent/types"
|
||||
import type { BackgroundOutputClient } from "../background-task/clients"
|
||||
import { createAthenaCouncilTool, filterCouncilMembers } from "./tools"
|
||||
|
||||
const mockClient = {
|
||||
session: {
|
||||
messages: async () => ({
|
||||
data: [{
|
||||
id: "msg-1",
|
||||
info: { role: "assistant" },
|
||||
parts: [{ type: "text", text: "Test analysis result" }],
|
||||
}],
|
||||
}),
|
||||
},
|
||||
} as unknown as BackgroundOutputClient
|
||||
|
||||
const mockManager = {
|
||||
getTask: () => undefined,
|
||||
launch: async () => {
|
||||
@@ -38,7 +25,7 @@ const configuredMembers = [
|
||||
{ model: "google/gemini-3-pro" },
|
||||
]
|
||||
|
||||
function createCompletedTask(id: string): BackgroundTask {
|
||||
function createRunningTask(id: string, sessionID = `ses-${id}`): BackgroundTask {
|
||||
return {
|
||||
id,
|
||||
parentSessionID: "session-1",
|
||||
@@ -46,288 +33,168 @@ function createCompletedTask(id: string): BackgroundTask {
|
||||
description: `Council member task ${id}`,
|
||||
prompt: "prompt",
|
||||
agent: "council-member",
|
||||
status: "completed",
|
||||
sessionID: `ses-${id}`,
|
||||
status: "running",
|
||||
sessionID,
|
||||
}
|
||||
}
|
||||
|
||||
describe("filterCouncilMembers", () => {
|
||||
test("returns all members when selection is undefined", () => {
|
||||
// #given
|
||||
const selectedMembers = undefined
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then
|
||||
const result = filterCouncilMembers(configuredMembers, undefined)
|
||||
expect(result.members).toEqual(configuredMembers)
|
||||
expect(result.error).toBeUndefined()
|
||||
})
|
||||
|
||||
test("returns all members when selection is empty", () => {
|
||||
// #given
|
||||
const selectedMembers: string[] = []
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then
|
||||
const result = filterCouncilMembers(configuredMembers, [])
|
||||
expect(result.members).toEqual(configuredMembers)
|
||||
expect(result.error).toBeUndefined()
|
||||
})
|
||||
|
||||
test("filters members using case-insensitive name and model matching", () => {
|
||||
// #given
|
||||
const selectedMembers = ["gpt", "GOOGLE/GEMINI-3-PRO"]
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then
|
||||
const result = filterCouncilMembers(configuredMembers, ["gpt", "GOOGLE/GEMINI-3-PRO"])
|
||||
expect(result.members).toEqual([configuredMembers[1], configuredMembers[2]])
|
||||
expect(result.error).toBeUndefined()
|
||||
})
|
||||
|
||||
test("returns helpful error when selected members are not configured", () => {
|
||||
// #given
|
||||
const selectedMembers = ["mistral", "xai/grok-3"]
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then
|
||||
const result = filterCouncilMembers(configuredMembers, ["mistral", "xai/grok-3"])
|
||||
expect(result.members).toEqual([])
|
||||
expect(result.error).toBe(
|
||||
"Unknown council members: mistral, xai/grok-3. Available members: Claude, GPT, google/gemini-3-pro."
|
||||
)
|
||||
})
|
||||
|
||||
test("selects named member by model ID when name differs from model", () => {
|
||||
// #given - "Claude" has name "Claude" but model "anthropic/claude-sonnet-4-5"
|
||||
const selectedMembers = ["anthropic/claude-sonnet-4-5"]
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then - should find the member by model ID even though it has a custom name
|
||||
expect(result.members).toEqual([configuredMembers[0]])
|
||||
expect(result.error).toBeUndefined()
|
||||
})
|
||||
|
||||
test("deduplicates when same member is selected by both name and model", () => {
|
||||
// #given
|
||||
const selectedMembers = ["Claude", "anthropic/claude-sonnet-4-5"]
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then - should return only one copy
|
||||
const result = filterCouncilMembers(configuredMembers, ["Claude", "anthropic/claude-sonnet-4-5"])
|
||||
expect(result.members).toEqual([configuredMembers[0]])
|
||||
expect(result.error).toBeUndefined()
|
||||
})
|
||||
|
||||
test("returns error listing only unmatched names when partially matched", () => {
|
||||
// #given
|
||||
const selectedMembers = ["claude", "non-existent"]
|
||||
|
||||
// #when
|
||||
const result = filterCouncilMembers(configuredMembers, selectedMembers)
|
||||
|
||||
// #then
|
||||
expect(result.members).toEqual([])
|
||||
expect(result.error).toBe(
|
||||
"Unknown council members: non-existent. Available members: Claude, GPT, google/gemini-3-pro."
|
||||
)
|
||||
})
|
||||
})
|
||||
|
||||
describe("createAthenaCouncilTool", () => {
|
||||
test("returns error when councilConfig is undefined", async () => {
|
||||
// #given
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: mockManager,
|
||||
councilConfig: undefined,
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #when
|
||||
const result = await athenaCouncilTool.execute({ question: "How should we proceed?" }, mockToolContext)
|
||||
|
||||
// #then
|
||||
expect(result).toBe("Athena council not configured. Add agents.athena.council.members to your config.")
|
||||
})
|
||||
|
||||
test("returns error when councilConfig has empty members", async () => {
|
||||
// #given
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: mockManager,
|
||||
councilConfig: { members: [] },
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #when
|
||||
const result = await athenaCouncilTool.execute({ question: "Any concerns?" }, mockToolContext)
|
||||
|
||||
// #then
|
||||
expect(result).toBe("Athena council not configured. Add agents.athena.council.members to your config.")
|
||||
})
|
||||
|
||||
test("uses expected description and question arg schema", () => {
|
||||
// #given
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: mockManager,
|
||||
councilConfig: { members: [{ model: "openai/gpt-5.3-codex" }] },
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #then - description should be dynamic and include the member model
|
||||
expect(athenaCouncilTool.description).toContain("openai/gpt-5.3-codex")
|
||||
expect(athenaCouncilTool.description).toContain("Available council members:")
|
||||
expect((athenaCouncilTool as { args: Record<string, unknown> }).args.question).toBeDefined()
|
||||
expect((athenaCouncilTool as { args: Record<string, unknown> }).args.members).toBeDefined()
|
||||
})
|
||||
|
||||
test("returns helpful error when members contains invalid names", async () => {
|
||||
// #given
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: mockManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
client: mockClient,
|
||||
})
|
||||
const toolArgs = {
|
||||
question: "Who should investigate this?",
|
||||
members: ["unknown-model"],
|
||||
}
|
||||
|
||||
// #when
|
||||
const result = await athenaCouncilTool.execute(toolArgs, mockToolContext)
|
||||
|
||||
// #then
|
||||
expect(result).toBe("Unknown council members: unknown-model. Available members: Claude, GPT, google/gemini-3-pro.")
|
||||
})
|
||||
|
||||
test("returns collected markdown results for all configured council members", async () => {
|
||||
// #given
|
||||
let launchCount = 0
|
||||
const taskStore = new Map<string, BackgroundTask>()
|
||||
const launchManager = {
|
||||
launch: async () => {
|
||||
launchCount += 1
|
||||
const task = createCompletedTask(`bg-${launchCount}`)
|
||||
taskStore.set(task.id, task)
|
||||
return task
|
||||
},
|
||||
getTask: (id: string) => taskStore.get(id),
|
||||
} as unknown as BackgroundManager
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: launchManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #when
|
||||
const result = await athenaCouncilTool.execute({ question: "How should we proceed?" }, mockToolContext)
|
||||
|
||||
// #then - returns markdown with council results, one section per member
|
||||
expect(result).toContain("## Council Results")
|
||||
expect(result).toContain("How should we proceed?")
|
||||
expect(result).toContain("### Claude (anthropic/claude-sonnet-4-5)")
|
||||
expect(result).toContain("### GPT (openai/gpt-5.3-codex)")
|
||||
expect(result).toContain("### google/gemini-3-pro (google/gemini-3-pro)")
|
||||
expect(result).toContain("Test analysis result")
|
||||
})
|
||||
|
||||
test("returns collected results only for selected members", async () => {
|
||||
// #given
|
||||
let launchCount = 0
|
||||
const taskStore = new Map<string, BackgroundTask>()
|
||||
const launchManager = {
|
||||
launch: async () => {
|
||||
launchCount += 1
|
||||
const task = createCompletedTask(`bg-${launchCount}`)
|
||||
taskStore.set(task.id, task)
|
||||
return task
|
||||
},
|
||||
getTask: (id: string) => taskStore.get(id),
|
||||
} as unknown as BackgroundManager
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: launchManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #when
|
||||
const result = await athenaCouncilTool.execute(
|
||||
{
|
||||
question: "Who should investigate this?",
|
||||
members: ["GPT", "google/gemini-3-pro"],
|
||||
},
|
||||
{ question: "Who should investigate this?", members: ["unknown-model"] },
|
||||
mockToolContext
|
||||
)
|
||||
|
||||
// #then - only selected members appear in output
|
||||
expect(result).toContain("### GPT (openai/gpt-5.3-codex)")
|
||||
expect(result).toContain("### google/gemini-3-pro (google/gemini-3-pro)")
|
||||
expect(result).not.toContain("### Claude")
|
||||
expect(launchCount).toBe(2)
|
||||
expect(result).toBe("Unknown council members: unknown-model. Available members: Claude, GPT, google/gemini-3-pro.")
|
||||
})
|
||||
|
||||
test("includes launch failures alongside successful member results", async () => {
|
||||
// #given
|
||||
test("returns selection error when members are omitted", async () => {
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: mockManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
})
|
||||
|
||||
const result = await athenaCouncilTool.execute({ question: "How should we proceed?" }, mockToolContext)
|
||||
|
||||
expect(result).toBe(
|
||||
"athena_council runs one member per call. Pass exactly one member in members (single-item array). Available members: Claude, GPT, google/gemini-3-pro."
|
||||
)
|
||||
})
|
||||
|
||||
test("returns selection error when multiple members are provided", async () => {
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: mockManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
})
|
||||
|
||||
const result = await athenaCouncilTool.execute(
|
||||
{ question: "How should we proceed?", members: ["Claude", "GPT"] },
|
||||
mockToolContext
|
||||
)
|
||||
|
||||
expect(result).toBe(
|
||||
"athena_council runs one member per call. Pass exactly one member in members (single-item array). Available members: Claude, GPT, google/gemini-3-pro."
|
||||
)
|
||||
})
|
||||
|
||||
test("launches selected member and returns background task metadata", async () => {
|
||||
let launchCount = 0
|
||||
const taskStore = new Map<string, BackgroundTask>()
|
||||
const launchManager = {
|
||||
launch: async () => {
|
||||
launchCount += 1
|
||||
if (launchCount === 2) {
|
||||
throw new Error("provider outage")
|
||||
}
|
||||
const task = createCompletedTask(`bg-${launchCount}`)
|
||||
const task = createRunningTask(`bg-${launchCount}`)
|
||||
taskStore.set(task.id, task)
|
||||
return task
|
||||
},
|
||||
getTask: (id: string) => taskStore.get(id),
|
||||
} as unknown as BackgroundManager
|
||||
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: launchManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #when
|
||||
const result = await athenaCouncilTool.execute({ question: "Any concerns?" }, mockToolContext)
|
||||
const result = await athenaCouncilTool.execute(
|
||||
{ question: "Who should investigate this?", members: ["GPT"] },
|
||||
mockToolContext
|
||||
)
|
||||
|
||||
// #then - successful members have results, failed member listed in failures section
|
||||
expect(result).toContain("### Claude (anthropic/claude-sonnet-4-5)")
|
||||
expect(result).toContain("### google/gemini-3-pro (google/gemini-3-pro)")
|
||||
expect(launchCount).toBe(1)
|
||||
expect(result).toContain("Council member launched in background.")
|
||||
expect(result).toContain("Task ID: bg-1")
|
||||
expect(result).toContain("Session ID: ses-bg-1")
|
||||
expect(result).toContain("Member: GPT")
|
||||
expect(result).toContain("Model: openai/gpt-5.3-codex")
|
||||
expect(result).toContain("Status: running")
|
||||
expect(result).toContain("background_output")
|
||||
expect(result).toContain("task_id=\"bg-1\"")
|
||||
expect(result).toContain("<task_metadata>")
|
||||
expect(result).toContain("session_id: ses-bg-1")
|
||||
})
|
||||
|
||||
test("returns launch failure details when selected member fails", async () => {
|
||||
const launchManager = {
|
||||
launch: async () => {
|
||||
throw new Error("provider outage")
|
||||
},
|
||||
getTask: () => undefined,
|
||||
} as unknown as BackgroundManager
|
||||
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: launchManager,
|
||||
councilConfig: { members: configuredMembers },
|
||||
})
|
||||
|
||||
const result = await athenaCouncilTool.execute(
|
||||
{ question: "Any concerns?", members: ["GPT"] },
|
||||
mockToolContext
|
||||
)
|
||||
|
||||
expect(result).toContain("Failed to launch council member.")
|
||||
expect(result).toContain("### Launch Failures")
|
||||
expect(result).toContain("**GPT**")
|
||||
expect(result).toContain("provider outage")
|
||||
})
|
||||
|
||||
test("returns dedup error when council is already running in same session", async () => {
|
||||
// #given - use a never-resolving launch to keep the first execution in-flight
|
||||
const pendingLaunch = new Promise<BackgroundTask>(() => {})
|
||||
const launchManager = {
|
||||
launch: async () => pendingLaunch,
|
||||
getTask: () => undefined,
|
||||
} as unknown as BackgroundManager
|
||||
const athenaCouncilTool = createAthenaCouncilTool({
|
||||
backgroundManager: launchManager,
|
||||
councilConfig: { members: [{ model: "openai/gpt-5.3-codex" }] },
|
||||
client: mockClient,
|
||||
})
|
||||
|
||||
// #when - first call starts but never resolves (stuck in launch)
|
||||
// second call should be rejected by session guard
|
||||
const _firstExecution = athenaCouncilTool.execute({ question: "First run" }, mockToolContext)
|
||||
|
||||
// Allow microtask queue to process so markCouncilRunning is called
|
||||
await new Promise((resolve) => setTimeout(resolve, 0))
|
||||
|
||||
const secondExecution = await athenaCouncilTool.execute({ question: "Second run" }, mockToolContext)
|
||||
|
||||
// #then
|
||||
expect(secondExecution).toBe("Council is already running for this session. Wait for the current council execution to complete.")
|
||||
})
|
||||
})
|
||||
|
||||
@@ -2,13 +2,11 @@ import { tool, type ToolDefinition } from "@opencode-ai/plugin"
|
||||
import { executeCouncil } from "../../agents/athena/council-orchestrator"
|
||||
import type { CouncilConfig, CouncilMemberConfig } from "../../agents/athena/types"
|
||||
import type { BackgroundManager } from "../../features/background-agent"
|
||||
import type { BackgroundOutputClient } from "../background-task/clients"
|
||||
import { ATHENA_COUNCIL_TOOL_DESCRIPTION_TEMPLATE } from "./constants"
|
||||
import { createCouncilLauncher } from "./council-launcher"
|
||||
import { isCouncilRunning, markCouncilDone, markCouncilRunning } from "./session-guard"
|
||||
import { waitForCouncilSessions } from "./session-waiter"
|
||||
import { collectCouncilResults } from "./result-collector"
|
||||
import type { AthenaCouncilToolArgs } from "./types"
|
||||
import { storeToolMetadata } from "../../features/tool-metadata-store"
|
||||
|
||||
function isCouncilConfigured(councilConfig: CouncilConfig | undefined): councilConfig is CouncilConfig {
|
||||
return Boolean(councilConfig && councilConfig.members.length > 0)
|
||||
@@ -19,6 +17,11 @@ interface FilterCouncilMembersResult {
|
||||
error?: string
|
||||
}
|
||||
|
||||
function buildSingleMemberSelectionError(members: CouncilMemberConfig[]): string {
|
||||
const availableNames = members.map((member) => member.name ?? member.model).join(", ")
|
||||
return `athena_council runs one member per call. Pass exactly one member in members (single-item array). Available members: ${availableNames}.`
|
||||
}
|
||||
|
||||
export function filterCouncilMembers(
|
||||
members: CouncilMemberConfig[],
|
||||
selectedNames: string[] | undefined
|
||||
@@ -77,9 +80,8 @@ function buildToolDescription(councilConfig: CouncilConfig | undefined): string
|
||||
export function createAthenaCouncilTool(args: {
|
||||
backgroundManager: BackgroundManager
|
||||
councilConfig: CouncilConfig | undefined
|
||||
client: BackgroundOutputClient
|
||||
}): ToolDefinition {
|
||||
const { backgroundManager, councilConfig, client } = args
|
||||
const { backgroundManager, councilConfig } = args
|
||||
const description = buildToolDescription(councilConfig)
|
||||
|
||||
return tool({
|
||||
@@ -89,7 +91,7 @@ export function createAthenaCouncilTool(args: {
|
||||
members: tool.schema
|
||||
.array(tool.schema.string())
|
||||
.optional()
|
||||
.describe("Optional list of council member names or models to consult. Defaults to all configured members."),
|
||||
.describe("Single-item list containing exactly one council member name or model ID."),
|
||||
},
|
||||
async execute(toolArgs: AthenaCouncilToolArgs, toolContext) {
|
||||
if (!isCouncilConfigured(councilConfig)) {
|
||||
@@ -100,83 +102,82 @@ export function createAthenaCouncilTool(args: {
|
||||
if (filteredMembers.error) {
|
||||
return filteredMembers.error
|
||||
}
|
||||
|
||||
if (isCouncilRunning(toolContext.sessionID)) {
|
||||
return "Council is already running for this session. Wait for the current council execution to complete."
|
||||
if (filteredMembers.members.length !== 1) {
|
||||
return buildSingleMemberSelectionError(councilConfig.members)
|
||||
}
|
||||
|
||||
markCouncilRunning(toolContext.sessionID)
|
||||
try {
|
||||
const execution = await executeCouncil({
|
||||
question: toolArgs.question,
|
||||
council: { members: filteredMembers.members },
|
||||
launcher: createCouncilLauncher(backgroundManager),
|
||||
parentSessionID: toolContext.sessionID,
|
||||
parentMessageID: toolContext.messageID,
|
||||
parentAgent: toolContext.agent,
|
||||
})
|
||||
const execution = await executeCouncil({
|
||||
question: toolArgs.question,
|
||||
council: { members: filteredMembers.members },
|
||||
launcher: createCouncilLauncher(backgroundManager),
|
||||
parentSessionID: toolContext.sessionID,
|
||||
parentMessageID: toolContext.messageID,
|
||||
parentAgent: toolContext.agent,
|
||||
})
|
||||
|
||||
// Register metadata for UI visibility (makes sessions clickable in TUI).
|
||||
const metadataFn = (toolContext as Record<string, unknown>).metadata as
|
||||
| ((input: { title?: string; metadata?: Record<string, unknown> }) => Promise<void>)
|
||||
| undefined
|
||||
if (metadataFn && execution.launched.length > 0) {
|
||||
const sessions = await waitForCouncilSessions(
|
||||
execution.launched, backgroundManager, toolContext.abort
|
||||
)
|
||||
for (const session of sessions) {
|
||||
await metadataFn({
|
||||
title: `Council: ${session.memberName}`,
|
||||
metadata: {
|
||||
sessionId: session.sessionId,
|
||||
agent: "council-member",
|
||||
model: session.model,
|
||||
description: `Council member: ${session.memberName}`,
|
||||
},
|
||||
})
|
||||
}
|
||||
if (execution.launched.length === 0) {
|
||||
return formatCouncilLaunchFailure(execution.failures)
|
||||
}
|
||||
|
||||
const launched = execution.launched[0]
|
||||
const launchedMemberName = launched?.member.name ?? launched?.member.model
|
||||
const launchedMemberModel = launched?.member.model ?? "unknown"
|
||||
const launchedTaskId = launched?.taskId ?? "unknown"
|
||||
|
||||
const sessionInfos = await waitForCouncilSessions(execution.launched, backgroundManager, toolContext.abort)
|
||||
const launchedSession = sessionInfos.find((session) => session.taskId === launchedTaskId)
|
||||
const sessionId = launchedSession?.sessionId ?? "pending"
|
||||
|
||||
const metadataFn = (toolContext as Record<string, unknown>).metadata as
|
||||
| ((input: { title?: string; metadata?: Record<string, unknown> }) => Promise<void>)
|
||||
| undefined
|
||||
if (metadataFn) {
|
||||
const memberMetadata = {
|
||||
title: `Council: ${launchedMemberName}`,
|
||||
metadata: {
|
||||
sessionId,
|
||||
agent: "council-member",
|
||||
model: launchedMemberModel,
|
||||
description: `Council member: ${launchedMemberName}`,
|
||||
},
|
||||
}
|
||||
await metadataFn(memberMetadata)
|
||||
|
||||
// Wait for all members to complete and collect their actual results.
|
||||
// This eliminates the need for Athena to poll background_output repeatedly.
|
||||
const collected = await collectCouncilResults(
|
||||
execution.launched, backgroundManager, client, toolContext.abort
|
||||
)
|
||||
|
||||
return formatCouncilOutput(toolArgs.question, collected.results, execution.failures)
|
||||
} catch (error) {
|
||||
throw error
|
||||
} finally {
|
||||
markCouncilDone(toolContext.sessionID)
|
||||
const callID = (toolContext as Record<string, unknown>).callID
|
||||
if (typeof callID === "string") {
|
||||
storeToolMetadata(toolContext.sessionID, callID, memberMetadata)
|
||||
}
|
||||
}
|
||||
|
||||
return `Council member launched in background.
|
||||
|
||||
Task ID: ${launchedTaskId}
|
||||
Session ID: ${sessionId}
|
||||
Member: ${launchedMemberName}
|
||||
Model: ${launchedMemberModel}
|
||||
Status: running
|
||||
|
||||
Use \`background_output\` with task_id="${launchedTaskId}" to collect this member's result.
|
||||
- block=true: Wait for completion and return the result
|
||||
- full_session=true: Include full session messages when needed
|
||||
|
||||
<task_metadata>
|
||||
session_id: ${sessionId}
|
||||
</task_metadata>`
|
||||
},
|
||||
})
|
||||
}
|
||||
|
||||
function formatCouncilOutput(
|
||||
question: string,
|
||||
results: Array<{ name: string; model: string; taskId: string; status: string; content: string }>,
|
||||
function formatCouncilLaunchFailure(
|
||||
failures: Array<{ member: { name?: string; model: string }; error: string }>
|
||||
): string {
|
||||
const sections: string[] = []
|
||||
|
||||
sections.push(`## Council Results\n\n**Question:** ${question}\n`)
|
||||
|
||||
for (const result of results) {
|
||||
const header = `### ${result.name} (${result.model})`
|
||||
if (result.status !== "completed") {
|
||||
sections.push(`${header}\n\n*Status: ${result.status}*`)
|
||||
continue
|
||||
}
|
||||
sections.push(`${header}\n\n${result.content}`)
|
||||
if (failures.length === 0) {
|
||||
return "Failed to launch council member."
|
||||
}
|
||||
|
||||
if (failures.length > 0) {
|
||||
const failureLines = failures
|
||||
.map((f) => `- **${f.member.name ?? f.member.model}**: ${f.error}`)
|
||||
.join("\n")
|
||||
sections.push(`### Launch Failures\n\n${failureLines}`)
|
||||
}
|
||||
const failureLines = failures
|
||||
.map((failure) => `- **${failure.member.name ?? failure.member.model}**: ${failure.error}`)
|
||||
.join("\n")
|
||||
|
||||
return sections.join("\n\n---\n\n")
|
||||
return `Failed to launch council member.\n\n### Launch Failures\n\n${failureLines}`
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user